jackkuo commited on
Commit
68c213d
·
verified ·
1 Parent(s): 25a1ce4

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. -dE2T4oBgHgl3EQfQgYx/content/tmp_files/2301.03770v1.pdf.txt +3142 -0
  2. -dE2T4oBgHgl3EQfQgYx/content/tmp_files/load_file.txt +0 -0
  3. .gitattributes +37 -0
  4. 19AyT4oBgHgl3EQfbvd_/content/tmp_files/2301.00269v1.pdf.txt +1475 -0
  5. 19AyT4oBgHgl3EQfbvd_/content/tmp_files/load_file.txt +0 -0
  6. 19FAT4oBgHgl3EQfDBzX/content/2301.08414v1.pdf +3 -0
  7. 2NE1T4oBgHgl3EQf5QXz/content/tmp_files/2301.03511v1.pdf.txt +1151 -0
  8. 2NE1T4oBgHgl3EQf5QXz/content/tmp_files/load_file.txt +0 -0
  9. 3tFKT4oBgHgl3EQf8i6_/vector_store/index.pkl +3 -0
  10. 4NA0T4oBgHgl3EQfNf9p/content/tmp_files/2301.02147v1.pdf.txt +641 -0
  11. 4NA0T4oBgHgl3EQfNf9p/content/tmp_files/load_file.txt +0 -0
  12. 5NFAT4oBgHgl3EQfFRzt/vector_store/index.faiss +3 -0
  13. 5NFAT4oBgHgl3EQfFRzt/vector_store/index.pkl +3 -0
  14. 6tE1T4oBgHgl3EQfTQPr/vector_store/index.faiss +3 -0
  15. 7dAzT4oBgHgl3EQf-f5K/vector_store/index.faiss +3 -0
  16. 7dE1T4oBgHgl3EQf7QV5/content/tmp_files/2301.03532v1.pdf.txt +1137 -0
  17. 7dE1T4oBgHgl3EQf7QV5/content/tmp_files/load_file.txt +0 -0
  18. 8dFQT4oBgHgl3EQf4jbF/content/tmp_files/2301.13432v1.pdf.txt +686 -0
  19. 8dFQT4oBgHgl3EQf4jbF/content/tmp_files/load_file.txt +0 -0
  20. ANE0T4oBgHgl3EQfPgBb/content/tmp_files/2301.02179v1.pdf.txt +1237 -0
  21. ANE0T4oBgHgl3EQfPgBb/content/tmp_files/load_file.txt +0 -0
  22. AdAyT4oBgHgl3EQf3_pa/content/tmp_files/2301.00778v1.pdf.txt +2579 -0
  23. AdAyT4oBgHgl3EQf3_pa/content/tmp_files/load_file.txt +0 -0
  24. B9E5T4oBgHgl3EQfTg8R/content/tmp_files/2301.05536v1.pdf.txt +2756 -0
  25. B9E5T4oBgHgl3EQfTg8R/content/tmp_files/load_file.txt +0 -0
  26. CtAyT4oBgHgl3EQfefgG/content/2301.00320v1.pdf +3 -0
  27. CtAyT4oBgHgl3EQfefgG/vector_store/index.faiss +3 -0
  28. CtAyT4oBgHgl3EQfefgG/vector_store/index.pkl +3 -0
  29. D9A0T4oBgHgl3EQfAv_u/content/tmp_files/2301.01968v1.pdf.txt +1499 -0
  30. D9A0T4oBgHgl3EQfAv_u/content/tmp_files/load_file.txt +0 -0
  31. DNE1T4oBgHgl3EQfqAUl/vector_store/index.faiss +3 -0
  32. DNE1T4oBgHgl3EQfqAUl/vector_store/index.pkl +3 -0
  33. ENAyT4oBgHgl3EQfevhN/content/tmp_files/2301.00326v1.pdf.txt +4172 -0
  34. ENAyT4oBgHgl3EQfevhN/content/tmp_files/load_file.txt +0 -0
  35. EtAyT4oBgHgl3EQfevi1/vector_store/index.pkl +3 -0
  36. EtFLT4oBgHgl3EQfFS_K/vector_store/index.faiss +3 -0
  37. FNAzT4oBgHgl3EQfG_ve/content/tmp_files/2301.01039v1.pdf.txt +1143 -0
  38. FNAzT4oBgHgl3EQfG_ve/content/tmp_files/load_file.txt +532 -0
  39. FtE1T4oBgHgl3EQfXARM/content/tmp_files/2301.03121v1.pdf.txt +1863 -0
  40. FtE1T4oBgHgl3EQfXARM/content/tmp_files/load_file.txt +0 -0
  41. FtE3T4oBgHgl3EQfVwq1/content/2301.04463v1.pdf +3 -0
  42. FtE3T4oBgHgl3EQfVwq1/vector_store/index.faiss +3 -0
  43. FtE3T4oBgHgl3EQfVwq1/vector_store/index.pkl +3 -0
  44. GdE2T4oBgHgl3EQfTQfv/content/tmp_files/2301.03802v1.pdf.txt +1384 -0
  45. GdE2T4oBgHgl3EQfTQfv/content/tmp_files/load_file.txt +0 -0
  46. JNFRT4oBgHgl3EQfzTgG/content/2301.13649v1.pdf +3 -0
  47. JNFRT4oBgHgl3EQfzTgG/vector_store/index.faiss +3 -0
  48. JNFRT4oBgHgl3EQfzTgG/vector_store/index.pkl +3 -0
  49. JtAzT4oBgHgl3EQfIPso/content/tmp_files/2301.01057v1.pdf.txt +895 -0
  50. JtAzT4oBgHgl3EQfIPso/content/tmp_files/load_file.txt +0 -0
-dE2T4oBgHgl3EQfQgYx/content/tmp_files/2301.03770v1.pdf.txt ADDED
@@ -0,0 +1,3142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.03770v1 [cs.DB] 10 Jan 2023
2
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
3
+ Junyong Yang
4
+ Wuhan University
5
+ Wuhan, China
6
+ thomasyang@whu.edu.cn
7
+ Ming Zhong∗
8
+ Wuhan University
9
+ Wuhan, China
10
+ clock@whu.edu.cn
11
+ Yuanyuan Zhu
12
+ Wuhan University
13
+ Wuhan, China
14
+ yyzhu@whu.edu.cn
15
+ Tieyun Qian
16
+ Wuhan University
17
+ Wuhan, China
18
+ qty@whu.edu.cn
19
+ Mengchi Liu
20
+ South China Normal University
21
+ Guangzhou, China
22
+ liumengchi@scnu.edu.cn
23
+ Jeffery Xu Yu
24
+ The Chinese University of Hong
25
+ Kong
26
+ Hong Kong, China
27
+ yu@se.cuhk.edu.hk
28
+ ABSTRACT
29
+ Querying cohesive subgraphs on temporal graphs with various
30
+ time constraints has attractedintensive research interests recently.
31
+ In this paper, we study a novel Temporal 푘-Core Query (TCQ)
32
+ problem: given a time interval, find all distinct 푘-cores that exist
33
+ within any subintervals from a temporal graph, which general-
34
+ izes the previous historical 푘-core query. This problem is chal-
35
+ lenging because the number of subintervals increases quadrati-
36
+ cally to the span of time interval. For that, we propose a novel
37
+ Temporal Core Decomposition (TCD) algorithm that decremen-
38
+ tally induces temporal 푘-cores from the previously induced ones
39
+ and thus reduces “intra-core” redundant computationsignificantly.
40
+ Then, we introduce an intuitive concept named Tightest Time
41
+ Interval (TTI) for temporal 푘-core, and design an optimization
42
+ technique with theoretical guarantee that leverages TTI as a key
43
+ to predict which subintervals will induce duplicated푘-cores and
44
+ prunes the subintervals completely in advance, thereby eliminat-
45
+ ing “inter-core” redundant computation. The complexity of op-
46
+ timized TCD (OTCD) algorithm no longer depends on the span
47
+ of query time interval but only the scale of final results, which
48
+ means OTCD algorithm is scalable. Moreover, we propose a com-
49
+ pact in-memory data structure named Temporal Edge List (TEL)
50
+ to implement OTCD algorithm efficiently in physical level with
51
+ bounded memory requirement. TEL organizes temporal edges
52
+ in a “timeline” and can be updated instantly when new edges ar-
53
+ rive, and thus our approach can also deal with dynamic temporal
54
+ graphs. We compare OTCD algorithm with the incremental his-
55
+ torical 푘-core query on several real-world temporal graphs, and
56
+ observe that OTCD algorithm outperforms it by three orders of
57
+ magnitude, even though OTCD algorithm needs none precom-
58
+ puted index.
59
+ 1
60
+ INTRODUCTION
61
+ 1.1
62
+ Motivation
63
+ Discovering communities or cohesive subgraphs from temporal
64
+ graphs has great values in many application scenarios, thereby
65
+ attracting intensive research interests [1, 5, 12, 19, 25, 27, 34,
66
+ ∗The corresponding author.
67
+ This work is licensed under the Creative Commons BY-NC-ND 4.0 International
68
+ License. Visit https://creativecommons.org/licenses/by-nc-nd/4.0/ to view a copy
69
+ of this license. For any use beyond those covered by this license, obtain
70
+ permission by emailing info@vldb.org. Copyright is held by the owner/author(s).
71
+ Publication rights licensed to the VLDB Endowment.
72
+ Proceedings of the VLDB Endowment, Vol. 14, No. 1 ISSN 2150-8097.
73
+ doi:XX.XX/XXX.XX
74
+ v10
75
+ v1
76
+ v2
77
+ v3
78
+ v4
79
+ v5
80
+ v6
81
+ v7
82
+ v8
83
+ v9
84
+ 1
85
+ 1
86
+ 6
87
+ 6
88
+ 6
89
+ 6
90
+ 5
91
+ 5
92
+ 2
93
+ 2
94
+ 2
95
+ 2
96
+ 7
97
+ 7
98
+ 2
99
+ 6
100
+ 5
101
+ 5
102
+ 4
103
+ 3
104
+ 5
105
+ 5
106
+ 5
107
+ 5
108
+ 3
109
+ 2
110
+ 2
111
+ 8
112
+ 4
113
+ 1
114
+ 2-core of time interval [1,8]
115
+ 2-core of time interval [5,6]
116
+ 2-core of time interval [2,4]
117
+ 2-core of time interval [2,6]
118
+ Figure 1: A running example of temporal graph.
119
+ 36] in recent years. Here, a temporal graph refers to an undi-
120
+ rected multigraph in which each edge has a timestamp to indi-
121
+ cate when it occurred, as illustrated in Figure 1. For example,
122
+ consider a graph consisting of bank accounts as vertices and
123
+ fund transfer transactions between accounts as edges with natu-
124
+ ral timestamps. For applications such as anti-money-laundering,
125
+ we would like to search communities like 푘-cores that contain
126
+ a known suspicious account and emerge within a specific time
127
+ interval like the World Cup, and investigate the associated ac-
128
+ counts.
129
+ To address the community query/search problem for a fixed
130
+ time interval, the concept of historical 푘-core [36] is proposed
131
+ recently, which is the 푘-core induced from the subgraph of a
132
+ temporal graph in which all edges occurred out of the time in-
133
+ terval have been excluded and the parallel edges between each
134
+ pair of vertices have been merged. Also, the PHC-Query method
135
+ is proposed to deal with historical 푘-core query/search by using
136
+ a precomputed index efficiently.
137
+ However, we usually do not know the exact time interval of
138
+ targeted historical 푘-core in real-world applications. Actually, if
139
+ we can know the exact time interval, a traditional core decom-
140
+ position on the projected graph over the given time interval is
141
+ efficient enough to address the problem. Thus, it is more reason-
142
+ able to assume that we can only offer a flexible time interval
143
+ and need to induce cores from all its subintervals. For example,
144
+ for detecting money laundering by soccer gambling during the
145
+ World Cup, the 푘-cores emerged over a few of hours around one
146
+ of the matches are more valuable than a large 푘-core emerging
147
+ over the whole month.
148
+ Therefore, we aim to generalize historical 푘-core query by al-
149
+ lowing the result 푘-cores to be induced by any subinterval of a
150
+ given time interval, like “flexible versus fixed”. The historical 푘-
151
+ core query can be seen as a special case of our problem that only
152
+ evaluates the whole interval. Consider the following example.
153
+
154
+ Junyong Yang, Ming Zhong, Yuanyuan Zhu, Tieyun Qian, Mengchi Liu, and Jeffery Xu Yu
155
+ Example 1. As illustrated in Figure 1, given a time interval
156
+ [1,8], historical 푘-core query only returns the largest core marked
157
+ by the grey dashed line. In contrast, our temporal 푘-core query re-
158
+ turns four cores marked by dashed lines with different colors. These
159
+ cores can reveal various insights unseen by the largest one. For ex-
160
+ ample, some cores like red and blue that emerge in bursty periods
161
+ may be caused by special events. Also, some persistent or periodic
162
+ cores may be found. Further, we can analyze the interaction be-
163
+ tween cores and how they evolve over time, such as the small cores
164
+ like red and blue are merged to the large cores like yellow. Lastly,
165
+ some underlying details may be found. During the merge, the ver-
166
+ tex 푣5 may play a vital role because it appears in all the cores.
167
+ Overall, our general and flexible query model can support many
168
+ interesting temporal community analytics applications.
169
+ The general and flexible temporal k-core query we study is
170
+ naturally a generalization of existing query models like histori-
171
+ cal 푘-core and also potentially a common technique for various
172
+ temporal graph mining tasks mentioned in the above example.
173
+ 1.2
174
+ Contribution
175
+ In this paper, we study a novel temporal 푘-core query problem:
176
+ given a time interval, find all distinct 푘-cores that exist within
177
+ any subintervals from a temporal graph. Although the existing
178
+ PHC-Query returns the historical 푘-core of a fixed time inter-
179
+ val efficiently, it cannot be trivially applied to deal with the new
180
+ problem. Because inducing 푘-cores for each subinterval individ-
181
+ ually from scratch is not scalable, since the number of subinter-
182
+ vals increases quadratically with the span of time interval. More-
183
+ over, PHC-Query suffers from two other intrinsic shortcomings.
184
+ Firstly, it relies on a PHC-Index that precomputes the coreness
185
+ of all vertices over all time intervals, thereby incurring heavy of-
186
+ fline time and space overheads. Secondly, due to its sophisticated
187
+ construction, it is unclear if PHC-Index can be updated dynami-
188
+ cally. It is against the dynamic nature of temporal graphs.
189
+ In order to overcome the above challenges, we present a novel
190
+ temporal core decomposition algorithm and auxiliary optimiza-
191
+ tion and implementation techniques. Our contributions can be
192
+ summarized as follows.
193
+ • We formalize a general time-range cohesive subgraph query
194
+ problem on ubiquitous temporal graphs, namely, tempo-
195
+ ral푘-core query. Many previous typical푘-core query mod-
196
+ els on temporal graphs can be equivalently represented
197
+ by temporal 푘-core query with particular constraints.
198
+ • To address temporal 푘-core query, we propose a simple
199
+ and yet efficient algorithm framework based on a novel
200
+ temporal core decomposition operation. By using tempo-
201
+ ral core decomposition, our algorithm always decremen-
202
+ tally induces a temporal k-core from the previous induced
203
+ temporal k-core except the initial one, thereby reducing
204
+ redundant computation significantly.
205
+ • Moreover, we propose an intuitive concept named tight-
206
+ est time interval for temporal k-core. According to the
207
+ properties of tightest time intervals, we design three prun-
208
+ ing rules with theoretical guarantee to directly skip subin-
209
+ tervals that will not induce distinct temporal 푘-core. As
210
+ a result, the optimized algorithm is scalable in terms of
211
+ the span of query time interval, since only the necessary
212
+ subintervals are enumerated.
213
+ • For physical implementation of our algorithm, we pro-
214
+ pose a both space and time efficient data structure named
215
+ temporal edge list to represent a temporal graph in mem-
216
+ ory. It can be manipulated to perform temporal core de-
217
+ composition and tightest time interval based pruning rapidly
218
+ with bounded memory. More importantly, temporal edge
219
+ list can be incrementally updated with evolving temporal
220
+ graphs, so that our approach can support dynamic graph
221
+ applications naturally.
222
+ • Lastly, we evaluate the efficiency and effectiveness of our
223
+ algorithm on real-world datasets. The experimental re-
224
+ sults demonstrate that our algorithm outperforms the im-
225
+ proved PHC-Query by three orders of magnitude.
226
+ The rest of this paper is organized as follows. Section 2 for-
227
+ mally introduces the data model and query model, and also gives
228
+ a baseline algorithm. Sections 3-5 present our algorithm, opti-
229
+ mization and implementation techniques respectively. Section
230
+ 6 briefly discusses some meaningful extension of our approach.
231
+ Section 7 presents the experiments and analyzes the results. Sec-
232
+ tion 8 investigates the related work. Section 9 concludes our
233
+ work.
234
+ 2
235
+ PRELIMINARY
236
+ In this section, we propose a generalized 푘-core query problem
237
+ on temporal graphs, which facilitates various temporal commu-
238
+ nity query/search demands. The previous historical푘-core query [36]
239
+ can be seen as a special case of the proposed problem. Specifi-
240
+ cally, we introduce the data model and query model of the pro-
241
+ posedproblem in Section 2.1 and 2.2 respectively, and then present
242
+ a nontrivial baseline that addresses the proposed problem based
243
+ on the existing PHC-Query.
244
+ 2.1
245
+ Data Model
246
+ A temporal graph is normally an undirected graph G = (V, E)
247
+ with parallel temporal edges. Each temporal edge (푢,푣,푡) ∈ E is
248
+ associated with a timestamp 푡 that indicates when the interac-
249
+ tion happened between the vertices 푢,푣 ∈ V. For example, the
250
+ temporal edges could be transfer transactions between bank ac-
251
+ counts in a finance graph. Without a loss of generality, we use
252
+ continuous integers that start from 1 to denote timestamps. Fig-
253
+ ure 1 illustrates a temporal graph as our running example.
254
+ There are two useful concepts derived from the temporal graph.
255
+ Given a time interval [푡푠,푡푒], we define the projected graph of G
256
+ over [푡푠,푡푒] as G[푡푠,푡푒] = (V[푡푠,푡푒], E[푡푠,푡푒]), where V[푡푠,푡푒] =
257
+ V and E[푡푠,푡푒] = {(푢,푣,푡)|(푢,푣,푡) ∈ E, 푡 ∈ [푡푠,푡푒]}. Moreover,
258
+ we define the detemporalized graph of G[푡푠,푡푒] as a simple graph
259
+ 퐺[푡푠,푡푒] = (푉[푡푠,푡푒], 퐸[푡푠,푡푒]), where 푉[푡푠,푡푒]=V[푡푠,푡푒] and 퐸[푡푠,푡푒]
260
+ = {(푢,푣)|(푢,푣,푡) ∈ E[푡푠,푡푒] }.
261
+ 2.2
262
+ Query Model
263
+ For revealing communities in graphs, the 푘-core query is widely
264
+ adopted. Given an undirected graph 퐺 and an integer 푘, 푘-core
265
+ is the maximal induced subgraph of 퐺 in which all vertices have
266
+ degrees at least 푘, which is denoted by C푘 (퐺). The coreness of
267
+ a vertex 푣 in a graph 퐺 is the largest value of 푘 such that 푣 ∈
268
+ C푘 (퐺).
269
+ For temporal graphs, the Historical 푘-Core Query (HCQ) [36]
270
+ is proposed recently. It aims to find a 푘-core that appears during
271
+ a specific time interval. Formally, a historical 푘-core H푘
272
+ [푡푠,푡푒] (G)
273
+ is a 푘-core in the detemporalized projected graph 퐺[푡푠,푡푒] of G.
274
+ Thus, HCQ can be defined as follows.
275
+
276
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
277
+ Definition 1 (Historical 푘-Core Qery). For a temporal
278
+ graph G, given an integer 푘 and a time interval [푡푠,푡푒], return
279
+ H푘
280
+ [푡푠,푡푒] (G) = C푘 (퐺[푡푠,푡푒]).
281
+ In this paper, we propose a novel query model called Tempo-
282
+ ral 푘-Core Query (TCQ) that generalizes HCQ. The main differ-
283
+ ence is that the query time interval [푇푠,푇푒] of TCQ is a range
284
+ but not fixed query condition like [푡푠,푡푒] of HCQ. In TCQ,푇푠 and
285
+ 푇푒 are the minimum start time and maximum end time of query
286
+ time interval respectively, and thereby the 푘-cores induced by
287
+ each subinterval [푡푠,푡푒] ⊆ [푇푠,푇푒] are all potential results of
288
+ TCQ. Moreover, TCQ directly returns the maximal induced sub-
289
+ graphs of G in which all vertices have degrees (note that, the
290
+ number of neighbor vertices but not neighbor edges) at least 푘
291
+ as results. We call these subgraphs as temporal 푘-cores and de-
292
+ note by T 푘
293
+ [푡푠,푡푒] (G) a temporal 푘-core that appears over [푡푠,푡푒]
294
+ on G. Obviously, a historical 푘-core H푘
295
+ [푡푠,푡푒] (G) is the detempo-
296
+ ralized temporal 푘-core T 푘
297
+ [푡푠,푡푒] (G). Therefore, TCQ can be seen
298
+ as a group of HCQ and HCQ can be seen as a special case of
299
+ TCQ.
300
+ The formal definition of TCQ is as follows.
301
+ Definition 2 (Temporal푘-Core Qery). For a temporalgraph
302
+ G, given an integer 푘 and a time interval [푇푠,푇푒], return all dis-
303
+ tinct T 푘
304
+ [푡푠,푡푒] (G) with [푡푠,푡푒] ⊆ [푇푠,푇푒].
305
+ Note that, TCQ only returns the distinct temporal 푘-cores
306
+ that are not identical to each other, since multiple subintervals
307
+ of [푇푠,푇푒] may induce an identical subgraph of G. For brevity,
308
+ T 푘
309
+ [푡푠,푡푒] (G) is abbreviated as T 푘
310
+ [푡푠,푡푒] if the context is self-evident.
311
+ 2.3
312
+ Baseline Algorithm
313
+ A straightforward solution to TCQ is to enumerate each subin-
314
+ terval [푡푠,푡푒] ⊆ [푇푠,푇푒] and induce T 푘
315
+ [푡푠,푡푒] respectively, which
316
+ takes 푂(|푇푒 −푇푠|2|E|) time. However, the span of query time in-
317
+ terval (namely,푇푒−푇푠) can be extremely large in practice, which
318
+ results in enormous time consumption for inducing all temporal
319
+ 푘-cores from scratch independently. Therefore, we start from a
320
+ non-trivial baseline based on the existing PHC-Query.
321
+ 2.3.1
322
+ A Short Review of PHC-Qery. PHC-Query relies on a heavy-
323
+ weight index called PHC-Index that essentially precomputes the
324
+ coreness of all vertices in the projected graphs over all possible
325
+ time intervals. The index is logically a table that stores a set of
326
+ timestamp pairs for each vertex 푣 ∈ V (column) and each rea-
327
+ sonable coreness 푘 (row). Given a value of 푘, the coreness of a
328
+ vertex 푣 is exactly 푘 in the projected graph over [푡푠,푡푒] for each
329
+ timestamp pair 푡푠 and 푡푒 in the cell (푘, 푣). In particular, due to
330
+ the monotonicity of coreness of a vertex with respect to 푡푒 when
331
+ 푡푠 is fixed, PHC-Index can reduce its space cost significantly by
332
+ only storing the necessary but not all possible timestamp pairs.
333
+ Specifically, for a vertex 푣, a coreness 푘 and a start time 푡푠, only a
334
+ discrete set of core time need to be recorded, since the coreness
335
+ of the vertex over [푡푠,푡푒] will not change with the increase of
336
+ 푡푒 until 푡푒 is a core time. Consequently, given an HCQ instance,
337
+ PHC-Query leverages PHC-Index to directly determine whether
338
+ a vertex has the coreness no less than the required 푘, by compar-
339
+ ing the query time interval with the retrieved timestamp pairs,
340
+ and then induces historical 푘-cores with qualified vertices.
341
+ 2.3.2
342
+ Incremental PHC-Qery Algorithm. The main idea of our
343
+ baseline algorithm is to induce temporal 푘-cores incrementally,
344
+ Algorithm 1: Baseline iPHC-Query algorithm.
345
+ Input: G, 푘, 푇푠, 푇푒
346
+ Output: all distinct T 푘
347
+ [푡푠,푡푒] (G) with [푡푠,푡푒] ⊆ [푇푠,푇푒]
348
+ 1 for 푡푠 ← 푇푠 to 푇푒 do
349
+ 2
350
+ V ← ∅, E ← ∅, H푣 ← ∅, H푒 ← ∅
351
+ 3
352
+ for 푘 and 푡푠, retrieve the core time of each vertex in
353
+ G from PHC-Index and push them into H푣
354
+ 4
355
+ push the temporal edges with timestamps in [푡푠,푇푒]
356
+ in G into H푒
357
+ 5
358
+ for 푡푒 ← 푡푠 to 푇푒 do
359
+ 6
360
+ pop a vertex from H푣 and add it to V, until the
361
+ min core time of H푣 exceeds 푡푒
362
+ 7
363
+ pop an edge from H푒 and add it to E if both
364
+ vertices linked by this edge are in V, until the
365
+ min timestamp of H푒 exceeds 푡푒
366
+ 8
367
+ push all edges that have been popped from H푒
368
+ and are not added to E back to H푒
369
+ 9
370
+ collect T 푘
371
+ [푡푠,푡푒] = (V, E) if it is neither empty nor
372
+ identical to other existing results
373
+ thereby reducing redundant computation. With a temporal 푘-
374
+ core T 푘
375
+ [푡푠,푡푒], we induce T 푘
376
+ [푡푠,푡푒+1] simply by appending new ver-
377
+ tices to T 푘
378
+ [푡푠,푡푒], whose coreness has become no less than푘 due to
379
+ the expand of time interval. Those vertices can be directly iden-
380
+ tified by using core time retrieved from PHC-Index since 푡푠 is
381
+ fixed. The correctness of baseline algorithm is guaranteed while
382
+ the correctness of PHC-Query holds.
383
+ The pseudo code of incremental PHC-Query (iPHC-Query) al-
384
+ gorithm is presented in Algorithm 1. It enumerates all subinter-
385
+ vals of a given [푇푠,푇푒] in a particular order for fulfilling efficient
386
+ incremental temporal 푘-core induction. Specifically, it anchors
387
+ the value of 푡푠 (line 1), and increases the value of 푡푒 from 푡푠 to
388
+ 푇푒 (line 5), so that T 푘
389
+ [푡푠,푡푒+1] can always be incrementally gen-
390
+ erated from an existing T 푘
391
+ [푡푠,푡푒]. For each 푡푠 anchored and the
392
+ input 푘, the algorithm firstly retrieves the core time of all ver-
393
+ tices from PHC-Index, and pushes the vertices into a minimum
394
+ heap H푣 ordered by their core time (line 3). Moreover, all tem-
395
+ poral edges with timestamps in [푡푠,푇푒] are pushed into another
396
+ minimum heap H푒 ordered by their timestamp (line 4). Then, the
397
+ algorithm maintains a vertex set V and an edge set E, which rep-
398
+ resent the vertices and edges of T 푘
399
+ [푡푠,푡푒] respectively, whenever
400
+ 푡푒 is increased by the following steps. It pops remaining vertices
401
+ with core time no greater than 푡푒 from H푣 and adds them to V
402
+ (line 6), since the corenesss of these vertices are no less than
403
+ 푘 according to PHC-Index. Also, it pops remaining edges with
404
+ timestamp no greater than 푡푒 from H푒 and adds them to E if
405
+ both vertices linked by the edges are in V (line 7). Then, it puts
406
+ back the popped edges that are not in E into H푒 (line 8), because
407
+ they could still be contained by other temporal 푘-cores induced
408
+ later. Lastly, a temporal푘-core comprised of V and E that are not
409
+ empty is collected if it has not been induced before (line 9).
410
+ The complexity of baseline mainly depends on the mainte-
411
+ nance of both V and E. For the maintenance of V, each ver-
412
+ tex in T 푘
413
+ [푡푠,푇푒]is added to V from H푣 at most once in the inner
414
+ loop (lines 5-9), which takes logarithmic time for a heap. There-
415
+ fore, the total cost is bounded by �푇푒
416
+ 푡=푇푠 |V[푡,푇푒]| log |V[푡,푇푒]|.
417
+ The case is more complicated for the maintenance of E, since
418
+
419
+ Junyong Yang, Ming Zhong, Yuanyuan Zhu, Tieyun Qian, Mengchi Liu, and Jeffery Xu Yu
420
+ each edge with a timestamp within [푡,푇푒] is likely to be trans-
421
+ ferred between H푒 and E (lines 7-8), until both its endpoints are
422
+ contained by V. In the worst case, the total cost is bounded by
423
+ �푇푒
424
+ 푡=푇푠 |푇푒 − 푡||E[푡,푇 푒]| log |E[푡,푇 푒]|. While, the real cost in prac-
425
+ tice can be much lower since the |푇푒 − 푡| part should be a more
426
+ reasonable value.
427
+ Although the baseline algorithm can achieve incremental in-
428
+ duction of temporal k-core for each start time, PHC-Index incurs
429
+ a huge amount of extra space and time overheads. Moreover, its
430
+ incremental induction only offers a kind of “intra-core” optimiza-
431
+ tion that reduces the redundant computation in each temporal
432
+ 푘-core induction, and lacks of a kind of “inter-core” optimization
433
+ that can directly avoids inducing some temporal 푘-cores. In the
434
+ following sections, we first propose a novel algorithm that can
435
+ outperform baseline algorithm without any precomputation and
436
+ index, and then optimize it significantly to further improve the
437
+ efficiency by at least three orders of magnitude.
438
+ 3
439
+ ALGORITHM
440
+ In this section, we propose a novel efficient algorithm to address
441
+ TCQ. Our algorithm leverages a fundamental operation called
442
+ temporal core decomposition to induce T 푘
443
+ [푡푠,푡푒] from T 푘
444
+ [푡푠,푡푒+1] decre-
445
+ mentally. More importantly, our algorithm does not require any
446
+ precomputation and index space, and can still outperform the
447
+ baseline algorithm. Next, Section 3.1 introduces the temporal
448
+ core decomposition operation, and Section 3.2 presents our al-
449
+ gorithm.
450
+ 3.1
451
+ Temporal Core Decomposition (TCD)
452
+ Firstly, we introduce Temporal Core Decomposition (TCD) as
453
+ a basic operation on temporal graphs, which is derived from
454
+ the traditional core decomposition [2] on ordinary graphs. TCD
455
+ refers to a two-step operation of inducing a temporal 푘-core
456
+ T 푘
457
+ [푡푠,푡푒] of a given time interval [푡푠,푡푒] from a given temporal
458
+ graph G. The first step is truncation: remove temporal edges with
459
+ timestamps not in [푡푠,푡푒] from G, namely, induce the projected
460
+ graph G[푡푠,푡푒]. The second step is decomposition: iteratively peel
461
+ vertices with degree (the number of neighbor vertices but not
462
+ neighbor edges) less than 푘 and the edges linked to them to-
463
+ gether. The correctness of TCD is as intuitive as core decom-
464
+ position.
465
+ An excellent property of TCD operation is that, it can induce
466
+ a temporal푘-core T 푘
467
+ [푡푠,푡푒] from another temporal푘-core T 푘
468
+ [푡푠′,푡푒′]
469
+ with [푡푠,푡푒] ⊂ [푡푠′,푡푒′], so that we can develop a decremental al-
470
+ gorithm based on TCD operation to achieve efficient processing
471
+ of TCQ. To prove the correctness of this property, let us consider
472
+ the following Theorem 1.
473
+ Lemma 1. Given time intervals [푡푠,푡푒] and [푡푠′,푡푒′] such that
474
+ [푡푠,푡푒] ⊂ [푡푠′,푡푒′], we have T 푘
475
+ [푡푠,푡푒] is a subgraph of T 푘
476
+ [푡푠′,푡푒′].
477
+ Proof. For each vertex in T 푘
478
+ [푡푠,푡푒], its coreness in G[푡푠′,푡푒′] is
479
+ certainly no less than in G[푡푠,푡푒] (namely, ⩾ 푘), because G[푡푠,푡푒]
480
+ is a subgraph of G[푡푠′,푡푒′]. Thus, all vertices in T 푘
481
+ [푡푠,푡푒] will be
482
+ contained by T 푘
483
+ [푡푠′,푡푒′] that is a temporal 푘-core of G[푡푠′,푡푒′].
484
+
485
+ Theorem 1. Given a time interval [푡푠, 푡푒] and a temporal 푘-
486
+ core T 푘
487
+ [푡푠′,푡푒′] with [푡푠,푡푒] ⊂ [푡푠′,푡푒′], the subgraph induced by
488
+ using TCD operation from T 푘
489
+ [푡푠′,푡푒′] for [푡푠,푡푒] is T 푘
490
+ [푡푠,푡푒].
491
+ v3
492
+ v4
493
+ v5
494
+ v6
495
+ v7
496
+ v8
497
+ 6
498
+ 6
499
+ 6
500
+ 6
501
+ 5
502
+ 5
503
+ 2
504
+ 2
505
+ 2
506
+ 6
507
+ 5
508
+ 5
509
+ 5
510
+ 5
511
+ 5
512
+ 5
513
+ 3
514
+ 2
515
+ 4
516
+ v3
517
+ v4
518
+ v5
519
+ v6
520
+ 6
521
+ 6
522
+ 6
523
+ 6
524
+ 5
525
+ 5
526
+ 6
527
+ 5
528
+ 5
529
+ 5
530
+ 5
531
+ 5
532
+ 5
533
+ v7
534
+ v8
535
+ v3
536
+ v4
537
+ v5
538
+ v6
539
+ 6
540
+ 6
541
+ 6
542
+ 6
543
+ 5
544
+ 5
545
+ 5
546
+ 5
547
+ 5
548
+ 5
549
+ truncation
550
+ decomposition
551
+ Figure 2: Temporal core decomposition from T 2
552
+ [2,6] to
553
+ T 2
554
+ [5,6].
555
+ Proof. Firstly, we prove for any temporal graph G′ satisfy-
556
+ ing that T 푘
557
+ [푡푠,푡푒] is a subgraph of G′ and G′ is a subgraph of
558
+ G, we can induce T 푘
559
+ [푡푠,푡푒] from G′ by using TCD operation. For
560
+ each vertex in T 푘
561
+ [푡푠,푡푒], its coreness is not less than 푘 in G′ over
562
+ [푡푠,푡푒], because this temporal 푘-core is a subgraph of G′. Mean-
563
+ while, for each vertex in G′ but not in T 푘
564
+ [푡푠,푡푒], its coreness in G′
565
+ is not greater than in G, because G′ is a subgraph of G. Thus,
566
+ its coreness in G′ over [푡푠,푡푒] is less than 푘, because it is not
567
+ in the temporal 푘-core T 푘
568
+ [푡푠,푡푒] of G. As a result, T 푘
569
+ [푡푠,푡푒] is also
570
+ a temporal 푘-core of G′, and thereby can be induced by using
571
+ TCD operation from G′.
572
+ Then, consider two temporal 푘-cores T 푘
573
+ [푡푠,푡푒] and T 푘
574
+ [푡푠′,푡푒′] with
575
+ [푡푠,푡푒] ⊆ [푡푠′,푡푒′]. Due to Lemma 1, we have T 푘
576
+ [푡푠,푡푒] is a sub-
577
+ graph of T 푘
578
+ [푡푠′,푡푒′]. Let G′
579
+ [푡푠,푡푒] be the temporal graph induced by
580
+ the first step of TCD from T 푘
581
+ [푡푠′,푡푒′], which is certainly a subgraph
582
+ of T 푘
583
+ [푡푠′,푡푒′]. Since G′
584
+ [푡푠,푡푒] only removes the temporal edges not
585
+ in [푡푠,푡푒], which means these edges are not contained by T 푘
586
+ [푡푠,푡푒],
587
+ it is obviously T 푘
588
+ [푡푠,푡푒] is a subgraph of G′
589
+ [푡푠,푡푒]. Thus, the correct-
590
+ ness of this theorem holds.
591
+
592
+ For example, Figure 2 illustrates the procedure of TCD from
593
+ T 2
594
+ [2,6] to T 2
595
+ [5,6] on our running example graph in Figure 1. The
596
+ edges with timestamps not in [5, 6] (marked by dashed lines)
597
+ are firstly removed from T 2
598
+ [2,6] by truncation, which results in
599
+ the decrease of degrees of vertices 푣5, 푣7 and 푣8. Then, the ver-
600
+ tices with degree less than 2 (marked by dark circles), namely, 푣7
601
+ and 푣8 are further peeled by decomposition, together with their
602
+ edges. The remaining temporal graph is T 2
603
+ [5,6].
604
+ 3.2
605
+ TCD Algorithm
606
+ We propose a TCD algorithm to address TCQ by using temporal
607
+ core decomposition. In general, given a TCQ instance, the TCD
608
+ algorithm enumerates each subinterval of [푇푠,푇푒] in a particu-
609
+ lar order, so that the temporal 푘-cores of each subinterval are in-
610
+ duced decrementally from previously induced temporal 푘-cores
611
+ except the initial one.
612
+ Specifically, we enumerate a subinterval [푡푠,푡푒] of [푇푠,푇푒] as
613
+ follows. Initially, let 푡푠 = 푇푠 and 푡푒 = 푇푒. It means we induce
614
+ the largest temporal 푘-core T 푘
615
+ [푇푠,푇푒] at the beginning. Then, we
616
+ will anchor the start time 푡푠 = 푇푠 and decrease the end time 푡푒
617
+ from 푇푒 until 푡푠 gradually. As a result, we can always leverage
618
+ TCD to induce the temporal푘-core of current subinterval [푡푠,푡푒]
619
+ from the previously induced temporal 푘-core of [푡푠, 푡푒 + 1] but
620
+ not from G[푡푠,푡푒] or even G. Whenever the value of 푡푒 is de-
621
+ creased to 푡푠, the value of 푡푠 will be increased to 푡푠 + 1 until
622
+ 푡푠 = 푇푒, and the value of 푡푒 will be reset to 푇푒. Then, we in-
623
+ duce T 푘
624
+ [푡푠+1,푡푒] from T 푘
625
+ [푡푠,푡푒], and start over the decremental TCD
626
+
627
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
628
+ Algorithm 2: TCD algorithm.
629
+ Input: G, 푘, [푇푠,푇푒]
630
+ Output: all distinct T 푘
631
+ [푡푠,푡푒] with [푡푠,푡푒] ⊆ [푇푠,푇푒]
632
+ 1 for 푡푠 ← 푇푠 to 푇푒 do
633
+ // anchor a new start time
634
+ 2
635
+ 푡푒 ← 푇푒
636
+ // reset the end time
637
+ 3
638
+ if 푡푠 = 푇푠 then
639
+ 4
640
+ T 푘
641
+ [푡푠,푡푒] ← TCD(G[푇푠,푇푒], 푘, [푡푠,푡푒])
642
+ 5
643
+ else
644
+ 6
645
+ T 푘
646
+ [푡푠,푡푒] ← TCD(T 푘
647
+ [푡푠−1,푡푒], 푘, [푡푠,푡푒])
648
+ 7
649
+ collect T 푘
650
+ [푡푠,푡푒] if it is distinct
651
+ 8
652
+ for 푡푒 ← 푇푒 − 1 to 푡푠 do
653
+ // iteratively
654
+ decremental induction
655
+ 9
656
+ T 푘
657
+ [푡푠,푡푒] ← TCD(T 푘
658
+ [푡푠,푡푒+1], 푘, [푡푠,푡푒])
659
+ 10
660
+ collect T 푘
661
+ [푡푠,푡푒] if it is distinct
662
+ procedure. The pseudo code of TCD algorithm is given in Algo-
663
+ rithm 2. Note that, the details of TCD(G, 푘, [푡푠,푡푒]) function is
664
+ left to Section 5.2, in which we design a specific data structure
665
+ to implement TCD operation efficiently in physical level.
666
+ Figure 3 gives a demonstration of TCD algorithm for finding
667
+ temporal 2-cores of time interval [1,8] on our running example
668
+ graph. The temporal 푘-cores are induced line by line and from
669
+ left to right. Each arrow between temporal 푘-cores represents
670
+ a TCD operation from tail to head. We can see that, compared
671
+ with inducing each temporal 푘-core independently, the TCD al-
672
+ gorithm reduces the computational overhead significantly. For
673
+ most induced temporal 푘-cores, a number of vertices and edges
674
+ have already been excluded while inducing the previous tempo-
675
+ ral 푘-cores. Moreover, with the increase of 푡푠 and the decrease of
676
+ 푡푒 when 푡푠 is fixed, the size of T 푘
677
+ [푡푠,푡푒] will be reduced monotoni-
678
+ cally until no temporal푘-core exists over [푡푠,푡푒], so that the time
679
+ and space costs of TCD operation will also be reduced gradually.
680
+ Lastly, we compare TCD algorithm with Baseline algorithm
681
+ abstractly. When 푡푠 is fixed, Baseline algorithm conducts an in-
682
+ cremental procedure, in which each vertex is popped once and
683
+ each edge may be popped and pushed back many times, and in
684
+ contrast, TCD algorithm conducts a decremental procedure, in
685
+ which each vertex is peeled once and each edge is also removed
686
+ once due to Lemma 1. Therefore, TCD algorithm that is well im-
687
+ plemented in physical level (see Section 5.2) can be even more
688
+ efficient than Baseline algorithm, though it does not need any
689
+ precomputed index.
690
+ 4
691
+ OPTIMIZATION
692
+ In this section, we dive deeply into the procedure of TCD al-
693
+ gorithm and optimize it dramatically by introducing an intu-
694
+ itive concept called tightest time interval for temporal 푘-cores.
695
+ In a nutshell, we directly prune subintervals without inducing
696
+ their temporal 푘-cores if we can predict that the temporal 푘-
697
+ cores are identical to other induced temporal 푘-cores, and tight-
698
+ est time interval is the key to fulfill prediction. In this way, the
699
+ optimized TCD algorithm only performs TCD operations that
700
+ are necessary for returning all distinct answers to a given TCQ
701
+ instance. Conceptually, the new pruning operation of optimized
702
+ algorithm eliminates the “inter-core” redundant computation, and
703
+ v10
704
+ v3
705
+ v4
706
+ v5
707
+ v6
708
+ v7
709
+ v8
710
+ v9
711
+ 6
712
+ 6
713
+ 6
714
+ 6
715
+ 5
716
+ 5
717
+ 2
718
+ 2
719
+ 2
720
+ 2
721
+ 7
722
+ 7
723
+ 6
724
+ 5
725
+ 5
726
+ 5
727
+ 5
728
+ 5
729
+ 5
730
+ 3
731
+ 2
732
+ 2
733
+ 8
734
+ 4
735
+ 1
736
+ v10
737
+ v3
738
+ v4
739
+ v5
740
+ v6
741
+ v7
742
+ v8
743
+ v9
744
+ 6
745
+ 6
746
+ 6
747
+ 6
748
+ 5
749
+ 5
750
+ 2
751
+ 2
752
+ 2
753
+ 2
754
+ 7
755
+ 7
756
+ 6
757
+ 5
758
+ 5
759
+ 5
760
+ 5
761
+ 5
762
+ 5
763
+ 3
764
+ 2
765
+ 4
766
+ 1
767
+ v3
768
+ v4
769
+ v5
770
+ v6
771
+ v7
772
+ v8
773
+ 6
774
+ 6
775
+ 6
776
+ 6
777
+ 5
778
+ 5
779
+ 2
780
+ 2
781
+ 2
782
+ 6
783
+ 5
784
+ 5
785
+ 5
786
+ 5
787
+ 5
788
+ 5
789
+ 3
790
+ 2
791
+ 4
792
+ 2
793
+ v3
794
+ v4
795
+ v5
796
+ v6
797
+ v7
798
+ v8
799
+ 5
800
+ 5
801
+ 2
802
+ 2
803
+ 5
804
+ 5
805
+ 5
806
+ 5
807
+ 5
808
+ 3
809
+ 4
810
+ v5
811
+ v7
812
+ v8
813
+ 2
814
+ 2
815
+ 3
816
+ 4
817
+ v5
818
+ v7
819
+ v8
820
+ 2
821
+ 2
822
+ 2
823
+ 3
824
+ v5
825
+ v7
826
+ v8
827
+ 2
828
+ 2
829
+ 2
830
+ 2
831
+ 2
832
+ 2
833
+ 2
834
+ 2
835
+ 2
836
+ 5
837
+ v3
838
+ v4
839
+ v5
840
+ v6
841
+ 6
842
+ 6
843
+ 6
844
+ 6
845
+ 5
846
+ 5
847
+ 5
848
+ 5
849
+ 5
850
+ 5
851
+ v3
852
+ v4
853
+ v5
854
+ v6
855
+ 6
856
+ 6
857
+ 6
858
+ 5
859
+ 5
860
+ 5
861
+ 5
862
+ 5
863
+ 56
864
+ v3
865
+ v4
866
+ v5
867
+ v6
868
+ 5
869
+ 5
870
+ 5
871
+ 5
872
+ 5
873
+ 5
874
+ v3
875
+ v4
876
+ v5
877
+ v6
878
+ 6
879
+ 6
880
+ 6
881
+ 6
882
+ v3
883
+ v4
884
+ v5
885
+ v6
886
+ 5
887
+ 5
888
+ 5
889
+ 5
890
+ 5
891
+ 5
892
+ v10
893
+ v3
894
+ v4
895
+ v5
896
+ v6
897
+ v7
898
+ v8
899
+ v9
900
+ 6
901
+ 6
902
+ 6
903
+ 6
904
+ 5
905
+ 5
906
+ 2
907
+ 2
908
+ 2
909
+ 7
910
+ 6
911
+ 5
912
+ 5
913
+ 5
914
+ 5
915
+ 5
916
+ 5
917
+ 3
918
+ 2
919
+ 2
920
+ 8
921
+ 4
922
+ 2
923
+ v10
924
+ v3
925
+ v4
926
+ v5
927
+ v6
928
+ v7
929
+ v8
930
+ v9
931
+ 6
932
+ 6
933
+ 6
934
+ 6
935
+ 5
936
+ 5
937
+ 2
938
+ 2
939
+ 2
940
+ 7
941
+ 6
942
+ 5
943
+ 5
944
+ 5
945
+ 5
946
+ 5
947
+ 5
948
+ 3
949
+ 2
950
+ 2
951
+ 4
952
+ 2
953
+ 7
954
+ 7
955
+ v3
956
+ v4
957
+ v5
958
+ v6
959
+ v7
960
+ v8
961
+ 6
962
+ 6
963
+ 6
964
+ 6
965
+ 5
966
+ 5
967
+ 2
968
+ 2
969
+ 2
970
+ 6
971
+ 5
972
+ 5
973
+ 5
974
+ 5
975
+ 5
976
+ 5
977
+ 3
978
+ 2
979
+ 4
980
+ v3
981
+ v4
982
+ v5
983
+ v6
984
+ v7
985
+ v8
986
+ 5
987
+ 5
988
+ 2
989
+ 2
990
+ 5
991
+ 5
992
+ 5
993
+ 5
994
+ 5
995
+ 3
996
+ 4
997
+ v5
998
+ v7
999
+ v8
1000
+ 2
1001
+ 2
1002
+ 3
1003
+ 4
1004
+ v5
1005
+ v7
1006
+ v8
1007
+ 2
1008
+ 2
1009
+ 2
1010
+ 3
1011
+ v5
1012
+ v7
1013
+ v8
1014
+ 2
1015
+ 2
1016
+ 2
1017
+ 2
1018
+ 2
1019
+ 2
1020
+ 2
1021
+ 2
1022
+ 2
1023
+ 5
1024
+ v3
1025
+ v4
1026
+ v5
1027
+ v6
1028
+ v7
1029
+ 6
1030
+ 6
1031
+ 6
1032
+ 6
1033
+ 5
1034
+ 5
1035
+ 6
1036
+ 5
1037
+ 5
1038
+ 5
1039
+ 5
1040
+ 5
1041
+ 5
1042
+ 3
1043
+ v3
1044
+ v4
1045
+ v5
1046
+ v6
1047
+ v7
1048
+ 5
1049
+ 5
1050
+ 5
1051
+ 5
1052
+ 5
1053
+ 5
1054
+ 5
1055
+ 5
1056
+ 3
1057
+ v3
1058
+ v4
1059
+ v5
1060
+ v6
1061
+ v7
1062
+ 6
1063
+ 6
1064
+ 6
1065
+ 6
1066
+ 5
1067
+ 5
1068
+ 6
1069
+ 5
1070
+ 5
1071
+ 5
1072
+ 5
1073
+ 5
1074
+ 5
1075
+ 3
1076
+ v3
1077
+ v4
1078
+ v5
1079
+ v6
1080
+ v7
1081
+ 6
1082
+ 6
1083
+ 6
1084
+ 6
1085
+ 5
1086
+ 5
1087
+ 6
1088
+ 5
1089
+ 5
1090
+ 5
1091
+ 5
1092
+ 5
1093
+ 5
1094
+ 3
1095
+ v3
1096
+ v4
1097
+ v5
1098
+ v6
1099
+ 6
1100
+ 6
1101
+ 6
1102
+ 6
1103
+ 5
1104
+ 5
1105
+ 5
1106
+ 5
1107
+ 5
1108
+ 5
1109
+ v3
1110
+ v4
1111
+ v5
1112
+ v6
1113
+ 6
1114
+ 6
1115
+ 6
1116
+ 6
1117
+ 5
1118
+ 5
1119
+ 5
1120
+ 5
1121
+ 5
1122
+ 5
1123
+ v3
1124
+ v4
1125
+ v5
1126
+ v6
1127
+ 6
1128
+ 6
1129
+ 6
1130
+ 5
1131
+ 5
1132
+ 5
1133
+ 5
1134
+ 5
1135
+ 56
1136
+ v3
1137
+ v4
1138
+ v5
1139
+ v6
1140
+ 6
1141
+ 6
1142
+ 6
1143
+ 5
1144
+ 5
1145
+ 5
1146
+ 5
1147
+ 5
1148
+ 56
1149
+ v3
1150
+ v4
1151
+ v5
1152
+ v6
1153
+ 6
1154
+ 6
1155
+ 6
1156
+ 6
1157
+ v3
1158
+ v4
1159
+ v5
1160
+ v6
1161
+ 6
1162
+ 6
1163
+ 6
1164
+ 6
1165
+ ts
1166
+ te
1167
+ 1
1168
+ 2
1169
+ 3
1170
+ 4
1171
+ 5
1172
+ 6
1173
+ 1
1174
+ 2
1175
+ 3
1176
+ 4
1177
+ 5
1178
+ 6
1179
+ 8
1180
+ 7
1181
+ 7
1182
+ 8
1183
+ Figure 3: A demonstration of TCD algorithm for finding
1184
+ temporal 2-cores of time interval [1,8].
1185
+ the original TCD operation eliminates the “intra-core” redun-
1186
+ dant computation. Thus, the computational complexity of opti-
1187
+ mized algorithm no longer depends on the span of query time in-
1188
+ terval [푇푠,푇푒] like the baseline algorithm and the original TCD
1189
+ algorithm but only depends on the scale of final results.
1190
+ Next, we introduce the concept and properties of tightest time
1191
+ interval in Section 4.1, present three pruning rules based on tight-
1192
+ est time interval for TCD algorithm in Section 4.2, and briefly
1193
+ conclude and discuss the optimized TCD algorithm in Section 4.3.
1194
+ 4.1
1195
+ Tightest Time Interval (TTI)
1196
+ We have such an observation, a temporal 푘-core of [푡푠,푡푒] may
1197
+ only contain edges with timestamps in a subinterval [푡푠′,푡푒′] ⊂
1198
+ [푡푠,푡푒], since the edges in [푡푠, 푡푠′) and (푡푒′,푡푒] have been re-
1199
+ moved by core decomposition. For example, consider a tempo-
1200
+ ral 푘-core T 2
1201
+ [4,8] illustrated in Figure 3. We can see that it does
1202
+ not contain edges with timestamps 4, 7 and 8. As a result, if
1203
+ we continue to induce T 2
1204
+ [4,7] from T 2
1205
+ [4,8] and to induce T 2
1206
+ [4,6]
1207
+ from T 2
1208
+ [4,7], the returned temporal 푘-cores remain unchanged.
1209
+ The sameness of temporal 푘-cores induced by different subinter-
1210
+ vals inspires us to further optimize TCD algorithm by pruning
1211
+ subintervals directly. As illustrated in Figure 3, the subintervals
1212
+ such as [4,7], [4,6], [5,8], [5,7] and [5,6] all induce the identical
1213
+ temporal 푘-cores to [4,8], so that they can be potentially pruned
1214
+ in advance.
1215
+ For that, we propose the concept of Tightest Time Interval
1216
+ (TTI) for temporal 푘-cores. Given a temporal 푘-core of [푡푠,푡푒],
1217
+ its TTI refers to the minimal time interval [푡푠′,푡푒′] that can in-
1218
+ duce an identical temporal 푘-core to T 푘
1219
+ [푡푠,푡푒], namely, there is no
1220
+ subinterval of [푡푠′,푡푒′] that can induce an identical temporal 푘-
1221
+ core to T 푘
1222
+ [푡푠,푡푒]. We formalize the definition of TTI as follows.
1223
+ Definition 3 (Tightest Time Interval). Given a temporal
1224
+ 푘-core T 푘
1225
+ [푡푠,푡푒], its tightest time interval T 푘
1226
+ [푡푠,푡푒].TTI is [푡푠′,푡푒′], if
1227
+ and only if
1228
+ 1) T 푘
1229
+ [푡푠′,푡푒′] is an identical temporal 푘-core to T 푘
1230
+ [푡푠,푡푒];
1231
+ 2) there does not exist [푡푠′′,푡푒′′] ⊂ [푡푠′,푡푒′], such that T 푘
1232
+ [푡푠′′,푡푒′′]
1233
+ is an identical temporal 푘-core to T 푘
1234
+ [푡푠,푡푒].
1235
+
1236
+ Junyong Yang, Ming Zhong, Yuanyuan Zhu, Tieyun Qian, Mengchi Liu, and Jeffery Xu Yu
1237
+ It is easy to prove the TTI of a temporal 푘-core of [푡푠,푡푒] is
1238
+ surely a subinterval of [푡푠,푡푒]. To evaluate the TTI of a given
1239
+ T 푘
1240
+ [푡푠,푡푒], we have the following theorem.
1241
+ Theorem 2. Given a temporal 푘-core T 푘
1242
+ [푡푠,푡푒], T 푘
1243
+ [푡푠,푡푒].TTI =
1244
+ [푡푚푖푛,푡푚푎푥 ], where 푡푚푖푛 and 푡푚푎푥 are the minimum and maxi-
1245
+ mum timestamps in T 푘
1246
+ [푡푠,푡푒] respectively.
1247
+ Proof. On one hand, T 푘
1248
+ [푡푚푖푛,푡푚푎푥 ] is identical to T 푘
1249
+ [푡푠,푡푒]. Be-
1250
+ cause we can induce T 푘
1251
+ [푡푚푖푛,푡푚푎푥 ] by TCD operation from T 푘
1252
+ [푡푠,푡푒]
1253
+ due to [푡푚푖푛,푡푚푎푥 ] ⊆ [푡푠,푡푒]. Meanwhile, during the operation,
1254
+ none edge is actually removed since there is no edge with times-
1255
+ tamp outsides [푡푚푖푛,푡푚푎푥] in T 푘
1256
+ [푡푠,푡푒], and thus the temporal 푘-
1257
+ core T 푘
1258
+ [푡푠,푡푒] will remain unchanged. On the other hand, any time
1259
+ interval [푡푠′,푡푒′] ⊂ [푡푚푖푛,푡푚푎푥 ] cannot induce a temporal푘 core
1260
+ that is identical to T 푘
1261
+ [푡푠,푡푒], since the edges with timestamp either
1262
+ 푡푚푖푛 or 푡푚푎푥 in T 푘
1263
+ [푡푠,푡푒] are excluded at least.
1264
+
1265
+ With Theorem 2, we can evaluate the TTI of a given tempo-
1266
+ ral 푘-core instantly (by 푂(1) time, see Section 5), which guar-
1267
+ antees the following optimization based on TTI will not incur
1268
+ extra overheads.
1269
+ Moreover, there are the following important properties of TTI
1270
+ that support our pruning strategies.
1271
+ Property 1 (Uniqeness). Given a temporal 푘-core T 푘
1272
+ [푡푠,푡푒],
1273
+ there exists no other time interval than T 푘
1274
+ [푡푠,푡푒].TTI evaluated by
1275
+ Theorem 2 that is also a TTI of T 푘
1276
+ [푡푠,푡푒].
1277
+ Proof. Let T 푘
1278
+ [푡푠,푡푒].TTI be [푡푠′,푡푒′], and [푡푠′′,푡푒′′] ≠ [푡푠′,푡푒′]
1279
+ be any other time interval. There are only two possibilities. Firstly,
1280
+ [푡푠′,푡푒′] ⊄ [푡푠′′,푡푒′′]. However, the edges with timestamp 푡푠′
1281
+ and 푡푒′ are contained by T 푘
1282
+ [푡푠,푡푒] according to Theorem 2, and
1283
+ thereby [푡푠′′,푡푒′′] that does not cover [푡푠′,푡푒′] cannot induce
1284
+ T 푘
1285
+ [푡푠,푡푒]. Thus, the first possibility does not satisfy the first con-
1286
+ dition in Definition 3. Secondly, [푡푠′,푡푒′] ⊂ [푡푠′′,푡푒′′]. However,
1287
+ since [푡푠′,푡푒′] can induce T 푘
1288
+ [푡푠,푡푒], [푡푠′′,푡푒′′] is certainly not the
1289
+ tightest even if it can also induce T 푘
1290
+ [푡푠,푡푒]. Thus, the second possi-
1291
+ bility does not satisfy the second condition in Definition 3. Con-
1292
+ sequently, [푡푠′′,푡푒′′] ≠ [푡푠′,푡푒′] is not a TTI of T 푘
1293
+ [푡푠,푡푒].
1294
+
1295
+ Property 2 (Eqivalence). Given two temporal푘-cores T 푘
1296
+ [푡푠,푡푒]
1297
+ and T 푘
1298
+ [푡푠′,푡푒′], they are identical temporal graphs if and only if
1299
+ T 푘
1300
+ [푡푠,푡푒].TTI = T 푘
1301
+ [푡푠′,푡푒′].TTI.
1302
+ Proof. If T 푘
1303
+ [푡푠,푡푒].TTI = T 푘
1304
+ [푡푠′,푡푒′].TTI, T 푘
1305
+ [푡푠,푡푒] and T 푘
1306
+ [푡푠′,푡푒′] are
1307
+ both identical to the temporal 푘-core of the TTI according to
1308
+ Definition 3, and thus are identical to each other. Conversely, if
1309
+ T 푘
1310
+ [푡푠,푡푒] and T 푘
1311
+ [푡푠′,푡푒′] are identical, they must have a same unique
1312
+ TTI according to Theorem 2 and Property 1.
1313
+
1314
+ Property 3 (Inclusion). Given two temporal 푘-cores T 푘
1315
+ [푡푠,푡푒]
1316
+ and T 푘
1317
+ [푡푠′,푡푒′], we have T 푘
1318
+ [푡푠,푡푒].TTI ⊆ T 푘
1319
+ [푡푠′,푡푒′].TTI, if [푡푠,푡푒] ⊆
1320
+ [푡푠′,푡푒′].
1321
+ Proof. Since [푡푠,푡푒] ⊆ [푡푠′,푡푒′], we have T 푘
1322
+ [푡푠,푡푒] is a sub-
1323
+ graph of T 푘
1324
+ [푡푠′,푡푒′] according to Lemma 1. Thus, the minimum
1325
+ timestamp in T 푘
1326
+ [푡푠,푡푒] is certainly no earlier than the the min-
1327
+ imum timestamp in T 푘
1328
+ [푡푠′,푡푒′], and the maximum timestamp in
1329
+ T 푘
1330
+ [푡푠,푡푒] is certainly no later than the the maximum timestamp in
1331
+ T 푘
1332
+ [푡푠′,푡푒′]. Then, according to Theorem 2, we have T 푘
1333
+ [푡푠,푡푒].TTI ⊆
1334
+ T 푘
1335
+ [푡푠′,푡푒′].TTI.
1336
+
1337
+ Figure 4a abstracts Figure 3 as a schedule table of subinter-
1338
+ val enumeration, and TCD algorithm will traverse the cells row
1339
+ by row and from left to right. For example, the cell in row 1
1340
+ and column 6 represents a subinterval [1, 6], in which [2, 6] is
1341
+ the TTI of T 2
1342
+ [1,6]. In particular, the grey cells indicate that the
1343
+ temporal 푘-cores of the corresponding subintervals do not exist.
1344
+ Figure 4a clearly reveals that TCD algorithm suffers from induc-
1345
+ ing a number of identical temporal 푘-cores (with the same TTIs).
1346
+ For example, the TTI [5, 6] repeats six times, which means six
1347
+ cells will induce identical temporal 푘-cores.
1348
+ 4.2
1349
+ Pruning Rules
1350
+ The main idea of optimizing TCD algorithm is to predict the in-
1351
+ duction of identical temporal푘-cores by leveraging TTI, thereby
1352
+ skipping the corresponding subintervals during the enumera-
1353
+ tion. Specifically, whenever a temporal 푘-core of [푡푠,푡푒] is in-
1354
+ duced, we evaluate its TTI [푡푠′,푡푒′]. If 푡푠′ > 푡푠 or/and 푡푒′ < 푡푒,
1355
+ it is triggered that a number of subintervals on the schedule can
1356
+ be pruned in advance. According to different relations between
1357
+ [푡푠,푡푒] and [푡푠′,푡푒′], our pruning technique can be categorized
1358
+ into three rules which are not mutually exclusive. In other words,
1359
+ the three rules may be triggered at the same time, and prune dif-
1360
+ ferent subintervals respectively. Next, we present these pruning
1361
+ rules in Section 4.2.1, Section 4.2.2 and Section 4.2.3, respectively.
1362
+ 4.2.1
1363
+ Rule 1: Pruning-on-the-Right. Consider the schedule illus-
1364
+ trated in Figure 4a. For each row, TCD algorithm traverses the
1365
+ cells (namely, subintervals) from left to right. If the TTI [푡푠′,푡푒′]
1366
+ in the current cell [푡푠,푡푒] meets such a condition, namely, 푡푒′ <
1367
+ 푡푒, a pruning operation will be triggered, and the following cells
1368
+ in this row from [푡푠,푡푒 − 1] until [푡푠,푡푒′] will be skipped be-
1369
+ cause these subintervals will induce identical temporal 푘-cores
1370
+ to T 푘
1371
+ [푡푠,푡푒]. Since the pruned cells are on the right of trigger cell,
1372
+ we call this rule Pruning-On-the-Right (PoR). The pseudo code
1373
+ of PoR is given in lines 2-4 of Algorithm 3. The correctness of
1374
+ PoR is guaranteed by the following lemma.
1375
+ Lemma 2. Given a temporal푘-core T 푘
1376
+ [푡푠,푡푒] whose TTI is [푡푠′,푡푒′],
1377
+ for any time interval [푡푠,푡푒′′] with 푡푒′′ ∈ [푡푒′,푡푒], T 푘
1378
+ [푡푠,푡푒′′].TTI =
1379
+ [푡푠′,푡푒′].
1380
+ Proof. On one hand, since [푡푠,푡푒′′] ⊆ [푡푠,푡푒], T 푘
1381
+ [푡푠,푡푒′′].TTI
1382
+ ⊆ T 푘
1383
+ [푡푠,푡푒].TTI = [푡푠′,푡푒′] according to Inclusion (Property 3).
1384
+ On the other hand, we can prove [푡푠′,푡푒′] ⊆ T 푘
1385
+ [푡푠,푡푒′′].TTI. If
1386
+ we induce T 푘
1387
+ [푡푠′,푡푒′] from T 푘
1388
+ [푡푠,푡푒] by TCD operation, it is easy
1389
+ to know T 푘
1390
+ [푡푠,푡푒] will remain unchanged, because it only con-
1391
+ tains the edges with timestamps in [푡푠′,푡푒′] according to Theo-
1392
+ rem 2. Thus, we have T 푘
1393
+ [푡푠′,푡푒′].TTI = [푡푠′,푡푒′] according to Equiv-
1394
+ alence (Property 2). Also, since [푡푠′,푡푒′] ⊆ [푡푠,푡푒′′], [푡푠′,푡푒′] =
1395
+ T 푘
1396
+ [푡푠′,푡푒′].TTI ⊆ T 푘
1397
+ [푡푠,푡푒′′].TTI according to Inclusion (Property 3).
1398
+
1399
+ With Lemma 2, we can predict that the TTIs in the cells [푡푠,푡푒−
1400
+ 1], · · · , [푡푠,푡푒′] are the same as the trigger cell [푡푠,푡푒], when the
1401
+
1402
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
1403
+ ts te
1404
+ 1
1405
+ 2
1406
+ 3
1407
+ 4
1408
+ 5
1409
+ 6
1410
+ 1
1411
+ 2
1412
+ 3
1413
+ 4
1414
+ 5
1415
+ 6
1416
+ 8
1417
+ 7
1418
+ 7
1419
+ 8
1420
+ [6,6]
1421
+ [5,6]
1422
+ [5,6]
1423
+ [3,6]
1424
+ [2,8]
1425
+ [1,8]
1426
+ [6,6]
1427
+ [5,6]
1428
+ [5,6]
1429
+ [3,6]
1430
+ [2,7]
1431
+ [1,7]
1432
+ [6,6]
1433
+ [5,6]
1434
+ [5,6]
1435
+ [3,6]
1436
+ [2,6]
1437
+ [2,6]
1438
+ [5,5]
1439
+ [5,5]
1440
+ [3,5]
1441
+ [2,5]
1442
+ [2,5]
1443
+ [2,4]
1444
+ [2,4]
1445
+ [2,3]
1446
+ [2,3]
1447
+ [2,2]
1448
+ [2,2]
1449
+ (a) Without pruning.
1450
+ ts te
1451
+ 1
1452
+ 2
1453
+ 3
1454
+ 4
1455
+ 5
1456
+ 6
1457
+ 1
1458
+ 2
1459
+ 3
1460
+ 4
1461
+ 5
1462
+ 6
1463
+ 8
1464
+ 7
1465
+ 7
1466
+ 8
1467
+ [5,6]
1468
+ [3,6]
1469
+ [2,8]
1470
+ [1,8]
1471
+ [2,7]
1472
+ [1,7]
1473
+ [6,6]
1474
+ [2,6]
1475
+ [5,5]
1476
+ [3,5]
1477
+ [2,5]
1478
+ [2,4]
1479
+ [2,3]
1480
+ [2,2]
1481
+ Cell without core induced
1482
+ Pruning-on-the-Right
1483
+ Pruning-on-the-Underside
1484
+ Pruning-on-the-Left
1485
+ [x,y]
1486
+ Cell with core induced, TTI = [x,y]
1487
+ Pruning triggered by cell [1,6]
1488
+ Pruning triggered by cell [3,8]
1489
+ Pruning triggered by cell [4,8]
1490
+ (b) With pruning.
1491
+ Figure 4: Examples of subinterval pruning based on tightest time interval.
1492
+ PoR rule is satisfied. Thus, the temporal푘-cores induced by these
1493
+ subintervals are all identical to the induced T 푘
1494
+ [푡푠,푡푒] according to
1495
+ Equivalence (Property 2).
1496
+ For example, Figure 4b illustrates two instances of PoR (the
1497
+ cells in orange and blue colors with left arrow). When T 2
1498
+ [3,8] has
1499
+ been induced, we evaluate its TTI as [3, 6], and thus PoR is trig-
1500
+ gered. PoR immediately excludes the following two cells [3, 7]
1501
+ and [3, 6] from the schedule. As a proof, we can see the TTIs in
1502
+ these two cells are both [3, 6] in Figure 4a.
1503
+ 4.2.2
1504
+ Rule 2: Pruning-on-the-Underside. We now consider 푡푠′ >
1505
+ 푡푠, which causes pruning in the following rows but not the cur-
1506
+ rent row. So we call this rule Pruning-On-the-Underside (PoU).
1507
+ Specifically, if 푡푠′ > 푡푠, for each row 푟 ∈ [푡푠 + 1,푡푠′], the cells
1508
+ [푟,푡푒], [푟,푡푒 − 1], · · · , [푟,푟] will be skipped. The pseudo code of
1509
+ PoU is given in lines 5-8 of Algorithm 3. The correctness of PoU
1510
+ is guaranteed by the following lemmas.
1511
+ Lemma 3. Given a temporal푘-core T 푘
1512
+ [푡푠,푡푒] whose TTI is [푡푠′,푡푒′],
1513
+ for any time interval [푡푠′′,푡푒] with푡푠′′ ∈ [푡푠,푡푠′], we have the TTI
1514
+ of T 푘
1515
+ [푡푠′′,푡푒] is [푡푠′,푡푒′].
1516
+ Proof. The proof of this lemma is similar to Lemma 2 and
1517
+ thus is omitted.
1518
+
1519
+ Lemma 4. Given a temporal푘-core T 푘
1520
+ [푡푠,푡푒] whose TTI is [푡푠′,푡푒′],
1521
+ for any time interval [푟,푐] with 푟 ∈ [푡푠 + 1,푡푠′] and 푐 ∈ [푡푠,푡푒],
1522
+ we have T 푘
1523
+ [푟,푐] is identical to T 푘
1524
+ [푡푠,푐].
1525
+ Proof. For 푟 ∈ [푡푠 + 1,푡푠′], we have T 푘
1526
+ [푟,푡푒].TTI = [푡푠′,푡푒′]
1527
+ according to Lemma 3. Thus, T 푘
1528
+ [푟,푡푒] is identical to T 푘
1529
+ [푡푠,푡푒] ac-
1530
+ cording to the Equivalence (Property 2). Then, we have T 푘
1531
+ [푟,푐] is
1532
+ identical to T 푘
1533
+ [푡푠,푐] when 푐 = 푡푒 − 1 since them are induced by
1534
+ the same TCD operation from identical temporal graphs, and so
1535
+ on for the rest [푟,푐] with the decrease of 푐 until 푐 = 푡푠.
1536
+
1537
+ Lemma 4 indicates that, PoU safely prunes some cells in the
1538
+ following rows, since these cells contain the same TTIs as their
1539
+ upper cells, which even have not been enumerated yet except
1540
+ the trigger cell. For example, Figure 4b illustrates two PoU in-
1541
+ stances (the cells in yellow and blue colors with up arrow). On
1542
+ enumerating the cell [1, 6], since the contained TTI is [2, 6], the
1543
+ cells [2, 6], · · �� , [2, 2] are pruned by PoU, because the TTIs in
1544
+ these cells are the same as the cells [1, 6], · · · , [1, 2] respectively,
1545
+ though the TTIs of cells [1, 5], · · · , [1, 2] have not been evalu-
1546
+ ated.
1547
+ 4.2.3
1548
+ Rule 3: Pruning-on-the-Lef. Lastly, if both 푡푠′ > 푡푠 and
1549
+ 푡푒′ < 푡푒, for each row 푟 ∈ [푡푠′+1, 푡푒′], the cells [푟,푡푒], [푟,푡푒 −1],
1550
+ · · · , [푟,푡푒′ + 1] will also be skipped, besides the cells pruned by
1551
+ PoR and PoU. Although these cells are in the rows under the
1552
+ current row 푡푠, the temporal 푘-core of each of them is identical
1553
+ to the temporal 푘-core of a cell (namely, [푟,푡푒′]) on the right in
1554
+ the same row but not its upper cell like PoU. So we call this rule
1555
+ Pruning-On-the-Left (PoL). The pseudo code of PoL is given in
1556
+ lines 9-12 of Algorithm 3. The correctness of PoL is guaranteed
1557
+ by the following lemma.
1558
+ Lemma 5. Given a temporal푘-core T 푘
1559
+ [푡푠,푡푒] whose TTI is [푡푠′,푡푒′],
1560
+ for any time interval [푟,푐] with 푟 ∈ [푡푠′ + 1,푡푒′] and 푐 ∈ [푡푒′ +
1561
+ 1,푡푒], we have T 푘
1562
+ [푟,푐] is identical to T 푘
1563
+ [푟,푡푒′].
1564
+ Proof. Assume T 푘
1565
+ [푟,푐].TTI = [푟 ′,푐′]. According to Inclusion
1566
+ (Property 3), we have [푟 ′,푐′] ⊆ [푡푠′,푡푒′] since [푟,푐] ⊆ [푡푠,푡푒].
1567
+ Thus,푐′ ⩽ 푡푒′. Then, according to Lemma 2, we have T 푘
1568
+ [푟,푡푒′].TTI
1569
+ = [푟 ′,푐′] since 푡푒′ ∈ [푐′,푐]. Lastly, according to Equivalence
1570
+ (Property 2), we have T 푘
1571
+ [푟,푐] is identical to T 푘
1572
+ [푟,푡푒′].
1573
+
1574
+ For example, Figure 4b illustrates a PoL instance (the cells
1575
+ in blue color with right arrow). On enumerating the cell [4, 8],
1576
+ PoL is triggered since the contained TTI is [5, 6]. Then, the cells
1577
+ [6, 8] and [6, 7] are pruned by PoL because the TTIs contained
1578
+ in them are the same as the cell [6, 6] on the right of them. PoL
1579
+ is more tricky than PoU because the cells are pruned for contain-
1580
+ ing the same TTIs as other cells that are scheduled to traverse
1581
+ after them by TCD algorithm. Note that, the cell [4, 8] triggers
1582
+ all three kinds of pruning. In fact, a cell may trigger PoL only,
1583
+ PoU only, or all three rules.
1584
+ 4.3
1585
+ Optimized TCD Algorithm
1586
+ Compared with TCD algorithm, the improvement of Optimized
1587
+ TCD (OTCD) algorithm is simply to conduct a pruning opera-
1588
+ tion whenever a temporal 푘-core has been induced. Specifically,
1589
+ we evaluate the TTI of this temporal 푘-core, check each pruning
1590
+ rule to determine if it is triggered, and prune the specific subin-
1591
+ tervals on the schedule in advance. The pseudo code of pruning
1592
+ operation is given in Algorithm 3. Note that, the “prune” in Al-
1593
+ gorithm 3 is a logical concept, and can have different physical
1594
+ implementations.
1595
+
1596
+ Junyong Yang, Ming Zhong, Yuanyuan Zhu, Tieyun Qian, Mengchi Liu, and Jeffery Xu Yu
1597
+ Algorithm 3: Pruning operation.
1598
+ Input: [푡푠, 푡푒] and T 푘
1599
+ [푡푠,푡푒]
1600
+ 1 [푡푠′,푡푒′] ← T 푘
1601
+ [푡푠,푡푒].TTI // Theorem 2
1602
+ 2 if 푡푒′ < 푡푒 then
1603
+ // Rule 1: PoR
1604
+ 3
1605
+ for 푐 ← 푡푒 - 1 to 푡푒′ do
1606
+ 4
1607
+ prune the subinterval [푡푠,푐]
1608
+ 5 if 푡푠′ > 푡푠 then
1609
+ // Rule 2: PoU
1610
+ 6
1611
+ for 푟 ← 푡푠 + 1 to 푡푠′ do
1612
+ 7
1613
+ for 푐 ← te to r do
1614
+ 8
1615
+ prune the subinterval [푟,푐]
1616
+ 9 if 푡푠′ > 푡푠 and 푡푒′ < 푡푒 then
1617
+ // Rule 3: PoL
1618
+ 10
1619
+ for r ← ts’+1 to te’ do
1620
+ 11
1621
+ for c ← te to te’+1 do
1622
+ 12
1623
+ prune the subinterval [푟,푐]
1624
+ As illustrated in Figure 4b, OTCD algorithm completely elim-
1625
+ inates repeated inducing of identical temporal 푘-cores, namely,
1626
+ each distinct temporal 푘-core is induced exactly once during the
1627
+ whole procedure. It means, the real computational complexity
1628
+ of OTCD algorithm is the summation of complexity for induc-
1629
+ ing each distinct temporal 푘-core but not the temporal 푘-core
1630
+ of each subinterval of [푇푠,푇푒]. Therefore, we say OTCD algo-
1631
+ rithm is scalable with respect to the query time interval [푇푠,푇푒].
1632
+ For many real-world datasets, the span of [푇푠,푇푒] could be very
1633
+ large, while there exist only a limited number of distinct tem-
1634
+ poral 푘-cores over this period, so that OTCD algorithm can still
1635
+ process the query efficiently.
1636
+ 5
1637
+ IMPLEMENTATION
1638
+ In this section, we address the physical implementation of pro-
1639
+ posed algorithm. We first introduce a data structure for temporal
1640
+ graph representation in Section 5.1, based on which we explain
1641
+ the details of TCD Operation implementation in Section 5.2.
1642
+ 5.1
1643
+ Temporal Edge List (TEL)
1644
+ We propose a novel data structure called Temporal Edge List
1645
+ (TEL) for representing an arbitrary temporal graph (including
1646
+ temporal 푘-cores that are also temporal graphs), which is both
1647
+ the input and output of TCD operation. Conceptually, TEL(G)
1648
+ preserves a temporal graph G = (V, E) by organizing its edges
1649
+ in a 3-dimension space, each dimension of which is a set of bidi-
1650
+ rectional linked lists, as illustrated in Figure 5. The first dimen-
1651
+ sion is time, namely, all edges in E are grouped by their times-
1652
+ tamps. Each group is stored as a bidirectional linked list called
1653
+ Time List (TL), and TL(푡) denotes the list of edges with a times-
1654
+ tamp 푡. Then, TEL(G) uses a bidirectional linked list, in which
1655
+ each node represents a timestamp in G, as a timeline in ascend-
1656
+ ing order to link all TLs, so that some temporal operations can
1657
+ be facilitated. Moreover, the other two dimensions are source
1658
+ vertex and destination vertex respectively. We use a container
1659
+ to store the Source Lists (SL) or Destination Lists (DL) for each
1660
+ vertex 푣 ∈ V, where SL(푣) or DL(푣) is a bidirectional linked list
1661
+ that links all edges whose source or destination vertex is 푣. Ac-
1662
+ tually, an SL or DL is an adjacency list of the graph, by which
1663
+ we can retrieve the neighbor vertices and edges of a given vertex
1664
+ efficiently. Given a temporal graph G, TEL(G) is built in mem-
1665
+ ory by adding its edges iteratively. For each edge (푢,푣,푡) ∈ E,
1666
+ it is only stored once, and TL(푡), SL(푢) and DL(푣) will append its
1667
+ pointer at the tail respectively.
1668
+ Figure 5 illustrates a partial TEL of our example graph. The
1669
+ SLs and DLs other than SL(푣5) and DL(푣3) are omitted for con-
1670
+ ciseness. Basically, TL, SL and DL offer the functionality of re-
1671
+ trieving edges by timestamp and linked vertex respectively. For
1672
+ example, for removing all neighbor edges of a vertex 푣 with de-
1673
+ gree less than 푘 in TCD operation, we can locate SL(푣) and DL(푣)
1674
+ to retrieve these edges. Moreover, the linked list of TL can offer
1675
+ efficient temporal operations. For example, for truncating G to
1676
+ G[푡푠,푡푒] in TCD operation, we can remove TL(푡) with 푡 < 푡푠 or
1677
+ 푡 > 푡푒 from the linked list of TL conveniently. To get the TTI
1678
+ of a temporal 푘-core, we only need to check the head and tail
1679
+ nodes of the linked list of TL in its TEL to get the minimum
1680
+ and maximum timestamps respectively. The superiority of TEL
1681
+ is summarized as follows.
1682
+ • By TCD operation, a TEL will be trimmed to a smaller
1683
+ TEL, and there is none intermediate TEL produced. Thus,
1684
+ the memory requirement of (O)TCD algorithm only de-
1685
+ pends on the size of initial TEL(G[푇푠,푇푒]).
1686
+ • TEL consumes 푂(|E|) space for storing a temporal graph,
1687
+ which is optimal because at least 푂(|E|) space is required
1688
+ for storing a graph (e.g., adjacency lists). Although there
1689
+ are 6|E|+2|V|+3푛 pointers of TLs, SLs and DLs stored ad-
1690
+ ditionally, TEL is still compact compared with PHC-Index,
1691
+ where 푛 is the number of timestamps in the graph.
1692
+ • TEL supports the basic manipulations listed in Table 1 in
1693
+ constant time, which are cornerstones of implementing
1694
+ our algorithms and optimization techniques.
1695
+ • For dynamic graphs, when a new edge coming, TEL sim-
1696
+ ply appends a new node representing the current time at
1697
+ the end of linked list of TL, and then adds this edge as
1698
+ normal. Thus, TEL can also deal with dynamic graphs.
1699
+ 5.2
1700
+ Implement TCD Operation on TEL
1701
+ Given a TCQ instance, our algorithm starts to work on a copy
1702
+ of TEL(G[푇푠,푇푒]) in memory, which is obtained by truncating
1703
+ TEL(G). Then, our algorithm only needs to maintain an instance
1704
+ of TEL(T 푘
1705
+ [푡푠,푇푒]) and another instance of TEL(T 푘
1706
+ [푡푠,푡푒]) with [푡푠,푡푒]
1707
+ ⊆ [푇푠,푇푒] in memory. The first instance is used to induce the
1708
+ first temporal 푘-core T 푘
1709
+ [푡푠+1,푇푒] by TCD for each row in Figure 3.
1710
+ The second instance is used to induce the following temporal
1711
+ 푘-cores T 푘
1712
+ [푡푠,푡푒−1] by TCD in each row. Each TCD operation is
1713
+ decomposed to a series of TEL manipulations, and trims the in-
1714
+ put TEL without producing any intermediate data.
1715
+ To assist the implementation of TCD operation, our algorithm
1716
+ uses a global data structure H푣 that organizes all vertices in the
1717
+ maintained TEL into a minimum heap ordered by their degree,
1718
+ so that the vertices with less than 푘 neighbors can be retrieved
1719
+ directly. Note that, whenever an edge is deleted from the main-
1720
+ tained TEL, H푣 will also be updated due to the possible decrease
1721
+ of vertex degrees. The trivial details of updating H푣 is omitted.
1722
+ Algortithm 4 gives the implementation of TCD operation on
1723
+ TEL. The algorithm takes as input the TEL of a given graph G,
1724
+ along with the parameters 푘, 푡푠 and 푡푒 specifying the target tem-
1725
+ poral 푘-core T 푘
1726
+ [푡푠,푡푒]. In truncation phase, TEL(G) is projected
1727
+ to TEL(G[푡푠,푡푒]) (lines 1-14). Specifically, the linked list of TL
1728
+ is traversed from the head and tail bidirectionally until meet-
1729
+ ing 푡푠 and 푡푒 respectively. For each node representing the times-
1730
+ tamp 푡 traversed, the edges in TL(푡) are removed from TEL, and
1731
+
1732
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
1733
+ SL(v1)
1734
+ SL(v2)
1735
+ SL(v10)
1736
+ DL(v1)
1737
+ DL(v2)
1738
+ DL(v3)
1739
+ DL(v10)
1740
+
1741
+ (v1,v3)
1742
+ (v1,v3)
1743
+ TL(1)
1744
+ (v2,v3)
1745
+ (v7,v10)
1746
+ (v5,v7)
1747
+ (v5,v8)
1748
+ (v5,v8)
1749
+ (v7,v8)
1750
+ (v7,v9)
1751
+ (v7,v9)
1752
+ (v2,v3)
1753
+ (v5,v7)
1754
+ (v1,v3)
1755
+ (v7,v8)
1756
+ (v3,v4)
1757
+ (v3,v4)
1758
+ (v3,v5)
1759
+ (v3,v6)
1760
+ (v4,v6)
1761
+ (v4,v7)
1762
+ (v4,v7)
1763
+ (v5,v6)
1764
+ (v5,v7)
1765
+ (v3,v4)
1766
+ (v3,v6)
1767
+ (v4,v5)
1768
+ (v4,v7)
1769
+ (v5,v6)
1770
+ (v7,v10)
1771
+ (v9,v10)
1772
+ (v9,v10)
1773
+ SL(v5)
1774
+
1775
+
1776
+ Time Lists
1777
+ Destination
1778
+ Lists
1779
+ Source
1780
+ Lists
1781
+ TL(2)
1782
+ TL(3)
1783
+ TL(4)
1784
+ TL(5)
1785
+ TL(6)
1786
+ TL(7)
1787
+ TL(8)
1788
+ Figure 5: The conceptual illustration of a partial TEL of our running example graph.
1789
+ Table 1: The basic manipulations of TEL.
1790
+ Name
1791
+ Functionality
1792
+ Complexity
1793
+ next_TL(푇퐿) / prev_TL(푇퐿)
1794
+ get the next or previous TL in the linked list of TL
1795
+ 푂 (1)
1796
+ get_SL(푣) / get_DL(푣)
1797
+ get the SL or DL of a given vertex 푣 from a hash map
1798
+ 푂 (1)
1799
+ del_TL(푇퐿)
1800
+ remove the given TL node from the linked list of TL
1801
+ 푂 (1)
1802
+ del_edge(푒)
1803
+ delete a given edge 푒 = (푢, 푣, 푡) and update TL(푡), SL(푢) and DL(푣) respectively
1804
+ 푂 (1)
1805
+ get_TTI()
1806
+ return the timestamps of head and tail nodes of linked list of TL
1807
+ 푂 (1)
1808
+ H푣 is updated for each edge removed. In decomposition phase,
1809
+ TEL(G[푡푠,푡푒]) is further transformed to TEL(T 푘
1810
+ [푡푠,푡푒]) (lines 15-24).
1811
+ Specifically, the algorithm pops the vertex with the least neigh-
1812
+ bors from H푣 iteratively until the remaining vertices all have at
1813
+ least 푘 neighbors or the heap is empty. For each popped vertex
1814
+ 푣, it removes the linked edges of 푣 preserved in SL(푣) and DL(푣)
1815
+ from TEL respectively and updates H푣 accordingly. In particular,
1816
+ a TL will be removed from the linked list of TL after the last edge
1817
+ in it has been removed (lines 19 and 23).
1818
+ To clarify the procedure of Algorithm 4, Figure 6 illustrates an
1819
+ example of inducing T 2
1820
+ [4,5] from T 2
1821
+ [3,6]. The edges are going to be
1822
+ deleted are marked in red color. We can see that, the procedure
1823
+ is actually a stream of edge deletion, while TEL maintains the
1824
+ entries to retrieve the remaining edges.
1825
+ 5.3
1826
+ Complexity
1827
+ TCD.Theoretically, the complexity of TCD algorithm is bounded
1828
+ by �푇푒
1829
+ 푡=푇푠{(|V[푡,푇푒]|+|E[푡,푇 푒]|) log |V[푡,푇푒]|+푚|E[푡,푇푒]|}, where
1830
+ 푚 is a small constant. For each anchored푡, TCD algorithm gradu-
1831
+ ally peels T 푘
1832
+ [푡,푇푒] like an onion by TCD operation until it contains
1833
+ none temporal 푘-core. In the process, there are at most |E[푡,푇 푒]|
1834
+ edges deleted, and deleting each edge takes a small constant time
1835
+ 푂(푚) for TEL updating and at most 푂(log |V[푡,푇푒]|) time for
1836
+ H푣 maintenance. Similarly, there are at most |V[푡,푇푒]| vertices
1837
+ deleted, and deleting each vertex takes 푂(log |V[푡,푇푒]|) time for
1838
+ H푣 maintenance. Therefore, The total time overhead is the sum
1839
+ of edge and vertex deleting costs.
1840
+ Note that, the complexity of TCD algorithm can also be rep-
1841
+ resented by 푂((푇푒 −푇푠)2퐵) according to Algorithm 2, where 퐵
1842
+ is the average time overhead of TCD operation. However, 퐵 can-
1843
+ not be estimated precisely, since each TCD operation may delete
1844
+ zero to |E[푡,푇 푒]| edges. Therefore, we bound the complexity by
1845
+ the maximum deleting cost according to Algorithm 4, which is
1846
+ more reasonable.
1847
+ OTCD. The complexity of OTCD algorithm is simply bounded
1848
+ by �(|푉 ∗| + |퐸∗|) log |푉 ∗| +푚|퐸∗|, where 푉 ∗ and 퐸∗ refer to the
1849
+ sets of vertices and edges that have to be deleted for inducing the
1850
+ result temporal 푘-cores respectively. Due to the pruning rules,
1851
+ there are much less temporal 푘-cores induced by OTCD algo-
1852
+ rithm. Thus, |푉 ∗| and |퐸∗| are orders of magnitude less than the
1853
+ total number of vertices and edges deleted in TCD algorithm,
1854
+ most of which are actually used for inducing identical temporal
1855
+ 푘-cores, though they cannot be really estimated.
1856
+ 6
1857
+ EXTENSION
1858
+ To demonstrate the wide applicability of our approach in prac-
1859
+ tice, we present several typical scenarios that extends the data
1860
+ model or query model of TCQ, and sketch how to address them
1861
+ based on our data structure and algorithm in this section.
1862
+ 6.1
1863
+ Data Model Extension
1864
+ Dynamic Graph. Since most real-world graphs are evolving
1865
+ over time, it is significant to fulfill TCQ on dynamic graphs. Ben-
1866
+ efitted from its design in “timeline” style, our data structure TEL
1867
+ can deal with new edges naturally in memory through two new
1868
+ manipulations add_TL(푡) and add_edge(푢,푣,푡). When a new edge
1869
+ (푢,푣,푡) arrived, we firstly create an empty TL(푡), and append it
1870
+ at the end of the linked list of TL since 푡 is obviously greater
1871
+ than the existing timestamps. Then, we create a new edge node
1872
+ for (푢,푣,푡) and append it to TL(푡), SL(푢) and DL(푣) respectively.
1873
+
1874
+ Junyong Yang, Ming Zhong, Yuanyuan Zhu, Tieyun Qian, Mengchi Liu, and Jeffery Xu Yu
1875
+ DL(v1)
1876
+ DL(v2)
1877
+ DL(v4)
1878
+ SL(v1)
1879
+ SL(v2)
1880
+ SL(v3)
1881
+ SL(v4)
1882
+ DL(v3)
1883
+ TL(3)
1884
+ TL(4)
1885
+ TL(5)
1886
+ TL(6)
1887
+ v2
1888
+ v1
1889
+ v4
1890
+ v3
1891
+ 5
1892
+ 3
1893
+ 6
1894
+ 5
1895
+ 5
1896
+ 4
1897
+ (v1,v4)
1898
+ (v1,v2)
1899
+ (v1,v3)
1900
+ (v3,v4)
1901
+ (v2,v3)
1902
+ (v2,v4)
1903
+ DL(v1)
1904
+ DL(v2)
1905
+ DL(v4)
1906
+ SL(v1)
1907
+ SL(v2)
1908
+ SL(v3)
1909
+ SL(v4)
1910
+ DL(v3)
1911
+ TL(4)
1912
+ TL(5)
1913
+ (v1,v2)
1914
+ (v1,v3)
1915
+ (v2,v3)
1916
+ (v2,v4)
1917
+ DL(v1)
1918
+ DL(v2)
1919
+ DL(v4)
1920
+ SL(v1)
1921
+ SL(v2)
1922
+ SL(v3)
1923
+ SL(v4)
1924
+ DL(v3)
1925
+ TL(4)
1926
+ TL(5)
1927
+ (v1,v2)
1928
+ (v1,v3)
1929
+ (v2,v3)
1930
+ decomposition
1931
+ truncation
1932
+ v2
1933
+ v1
1934
+ v4
1935
+ v3
1936
+ 5
1937
+ 5
1938
+ 5
1939
+ 4
1940
+ v2
1941
+ v1
1942
+ v3
1943
+ 5
1944
+ 5
1945
+ 4
1946
+ Figure 6: An example of TCD operation on TEL.
1947
+ Algorithm 4: TCD operation in Algorithm 2
1948
+ Input: TEL(G), [푡푠,푡푒], 푘
1949
+ Output: TEL(T 푘
1950
+ [푡푠,푡푒])
1951
+ 1 푇퐿 ← the head of linked list of TL in TEL(G)
1952
+ 2 while 푇퐿.timestamp ≠ 푡푠 do
1953
+ 3
1954
+ for edge 푒 in 푇퐿 do
1955
+ 4
1956
+ del_edge(푒)
1957
+ 5
1958
+ udpate H푣
1959
+ 6
1960
+ del_TL(푇퐿)
1961
+ 7
1962
+ 푇퐿 ← next_TL(푇퐿)
1963
+ 8 푇퐿 ← the tail of linked list of TL in TEL(G)
1964
+ 9 while 푇퐿.timestamp ≠ 푡푒 do
1965
+ 10
1966
+ for edge 푒 in 푇퐿 do
1967
+ 11
1968
+ del_edge(푒)
1969
+ 12
1970
+ udpate H푣
1971
+ 13
1972
+ del_TL(푇퐿)
1973
+ 14
1974
+ 푇퐿 ← prev_TL(푇퐿)
1975
+ 15 while H푣 is not empty and H푣.peek < 푘 do
1976
+ 16
1977
+ vertex 푣 ← H푣.pop()
1978
+ 17
1979
+ for edge 푒 in SL(푣) do
1980
+ 18
1981
+ del_edge(푒)
1982
+ 19
1983
+ del_TL(TL(푒.timestamp)) if the TL is empty
1984
+ 20
1985
+ update H푣
1986
+ 21
1987
+ for edge 푒 in DL(푣) do
1988
+ 22
1989
+ del_edge(푒)
1990
+ 23
1991
+ del_TL(TL(푒.timestamp)) if the TL is empty
1992
+ 24
1993
+ update H푣
1994
+ Both manipulations are finished in constant time. The mainte-
1995
+ nance of a dynamic TEL is actually consistent with the construc-
1996
+ tion of a static TEL. Therefore, our (O)TCD algorithm can run
1997
+ on the dynamic TEL anytime.
1998
+ In contrast, updating PHC-Index is a non-trivial process. Al-
1999
+ though there are previous work [20, 29] on coreness updating
2000
+ for dynamic graphs, the update is only valid for the whole life
2001
+ time of graph. While, for an arbitrary start time, it is uncertain
2002
+ whether the coreness of a vertex will be changed by a new edge.
2003
+ 6.2
2004
+ Query Model Extension
2005
+ The existing graph mining tasks regarding 푘-core introduce var-
2006
+ ious constraints. For temporal graphs, we only focus on the tem-
2007
+ poral constraints. In the followings, we present two of them
2008
+ that can be integrated into TCQ model and also be addressed
2009
+ by our algorithm directly, which demonstrate the generality of
2010
+ our model and algorithm.
2011
+ Link Strength Constraint. In the context of temporal graph,
2012
+ link strength usually refers to the number of parallel edges be-
2013
+ tween a pair of linked vertices. Obviously, the minimum link
2014
+ strength in a temporal 푘-core represents some important prop-
2015
+ erties like validity, since noise interaction may appear over time
2016
+ and a pair of vertices with low link strength may only have oc-
2017
+ casional interaction during the time interval. Actually, the previ-
2018
+ ous work [34] has studied this problem without the time interval
2019
+ constraint. Therefore, it is reasonable to extend TCQ to retrieve
2020
+ 푘-cores with a lower bound of link strength during a given time
2021
+ interval. It can be achieved by trivially modifying the TCD Oper-
2022
+ ation. Specifically, the modified TCD Operation will remove the
2023
+ edges between two vertices once the number of parallel edges
2024
+ between them is decreased to be lower than the given lower
2025
+ bound of link strength, while the original TCD operation will
2026
+ do this when the number becomes zero. Thus, the modification
2027
+ brings almost none extra time and space consumption.
2028
+ Time Span Constraint. In many cases, we prefer to retrieve
2029
+ temporal 푘-cores with a short time span (between their earliest
2030
+ and latest timestamps), which is similar to the previous work on
2031
+ density-bursting subgraphs [5]. Because such a kind of short-
2032
+ term cohesive subgraphs tend to represent the occurrence of
2033
+ some special events. TCQ can be conveniently extended for re-
2034
+ solving the problem by specifying a constraint of time span. Since
2035
+ the time span of a temporal푘-core is preserved in its TEL, which
2036
+ is actually the length of its TTI, we can abandon the tempo-
2037
+ ral 푘-cores returned by TCD operation that cannot satisfy the
2038
+ time span constraint on the fly. It brings almost no extra time
2039
+ and space consumption. Moreover, we can also extend TCQ to
2040
+ find the temporal 푘-core with the shortest or top-푛 shortest time
2041
+ span.
2042
+
2043
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
2044
+ Table 2: Datasets.
2045
+ Name
2046
+ |V|
2047
+ |E|
2048
+ Span(days)
2049
+ Youtube
2050
+ 3.2M
2051
+ 9.4M
2052
+ 226
2053
+ DBLP
2054
+ 1.8M
2055
+ 29.5M
2056
+ 17532
2057
+ Flickr
2058
+ 2.3M
2059
+ 33M
2060
+ 198
2061
+ CollegeMsg
2062
+ 1.8K
2063
+ 20K
2064
+ 193
2065
+ email-Eu-core-temporal
2066
+ 0.9K
2067
+ 332K
2068
+ 803
2069
+ sx-mathoverflow
2070
+ 24.8K
2071
+ 506K
2072
+ 2350
2073
+ sx-stackoverflow
2074
+ 2.6M
2075
+ 63.5M
2076
+ 2774
2077
+ Table 3: Selected temporal 푘-core queries.
2078
+ id
2079
+ G
2080
+ 푡푠 (sec)
2081
+ 푡푒 (sec)
2082
+
2083
+ result #
2084
+ 1
2085
+ CollegeMsg
2086
+ 554400
2087
+ 565200
2088
+ 2
2089
+ 61
2090
+ 2
2091
+ CollegeMsg
2092
+ 558000
2093
+ 568800
2094
+ 2
2095
+ 21
2096
+ 3
2097
+ CollegeMsg
2098
+ 561600
2099
+ 572400
2100
+ 2
2101
+ 27
2102
+ 4
2103
+ CollegeMsg
2104
+ 565200
2105
+ 576000
2106
+ 2
2107
+ 26
2108
+ 5
2109
+ CollegeMsg
2110
+ 568800
2111
+ 579600
2112
+ 2
2113
+ 10
2114
+ 6
2115
+ email-Eu-core-temporal
2116
+ 36000
2117
+ 46800
2118
+ 3
2119
+ 2
2120
+ 7
2121
+ email-Eu-core-temporal
2122
+ 39600
2123
+ 50400
2124
+ 3
2125
+ 3
2126
+ 8
2127
+ email-Eu-core-temporal
2128
+ 284400
2129
+ 295200
2130
+ 3
2131
+ 7
2132
+ 9
2133
+ email-Eu-core-temporal
2134
+ 288000
2135
+ 298800
2136
+ 3
2137
+ 25
2138
+ 10
2139
+ email-Eu-core-temporal
2140
+ 291600
2141
+ 302400
2142
+ 3
2143
+ 16
2144
+ 11
2145
+ sx-mathoverflow
2146
+ 864000
2147
+ 867600
2148
+ 2
2149
+ 8
2150
+ 12
2151
+ sx-mathoverflow
2152
+ 1116000
2153
+ 1119600
2154
+ 2
2155
+ 4
2156
+ 13
2157
+ sx-mathoverflow
2158
+ 1389600
2159
+ 1393200
2160
+ 2
2161
+ 5
2162
+ 14
2163
+ sx-mathoverflow
2164
+ 1483200
2165
+ 1486300
2166
+ 2
2167
+ 2
2168
+ 15
2169
+ sx-mathoverflow
2170
+ 1738800
2171
+ 1742400
2172
+ 2
2173
+ 8
2174
+ 16
2175
+ sx-stackoverflow
2176
+ 378000
2177
+ 381600
2178
+ 2
2179
+ 6
2180
+ 17
2181
+ sx-stackoverflow
2182
+ 417600
2183
+ 421200
2184
+ 2
2185
+ 37
2186
+ 18
2187
+ sx-stackoverflow
2188
+ 421200
2189
+ 424800
2190
+ 2
2191
+ 5
2192
+ 19
2193
+ sx-stackoverflow
2194
+ 424800
2195
+ 428400
2196
+ 2
2197
+ 5
2198
+ 20
2199
+ sx-stackoverflow
2200
+ 486000
2201
+ 489600
2202
+ 2
2203
+ 10
2204
+ 7
2205
+ EXPERIMENT
2206
+ In this section, we conduct experiments to verify both efficiency
2207
+ and effectiveness of the proposed algorithm on a Windows ma-
2208
+ chine with Intel Core i7 2.20GHz CPU and 64GB RAM. The al-
2209
+ gorithms are implemented through C++ Standard Template Li-
2210
+ brary. Our source codes are shared on GitHub1.
2211
+ 7.1
2212
+ Dataset
2213
+ We choose seven temporal graphs with different sizes and do-
2214
+ mains for our experiments. The first three graphs are obtained
2215
+ from KONECT Project [16], and the other four graphs are ob-
2216
+ tained from the SNAP [17]. The basic statistics of these graphs
2217
+ are given in Table 2. All timestamps are unified to integers in
2218
+ seconds.
2219
+ 7.2
2220
+ Efficiency
2221
+ To evaluate the efficiency of our algorithm, we firstly manually
2222
+ select twenty temporal푘-core queries from tested random queries
2223
+ with a time span (namely, 푇푒 −푇푠) of 1-3 days, which have been
2224
+ verified to be valid, namely, there is at least one temporal 푘-core
2225
+ returned for each query. The setting of time span is moderate,
2226
+ otherwise other algorithms than OTCD can hardly stop success-
2227
+ fully. Table 3 gives the details of query parameters, so that other
2228
+ 1https://github.com/ThomasYang-algo/Temporal-k-Core-Query-Project
2229
+ Table 4: Effect of pruning rules.
2230
+ id
2231
+ Triggered Times
2232
+ Pruned Cell Percentage (%)
2233
+ PoR
2234
+ PoU
2235
+ PoL
2236
+ PoR
2237
+ PoU
2238
+ PoL
2239
+ Total
2240
+ 1
2241
+ 54
2242
+ 72
2243
+ 2
2244
+ 0.02
2245
+ 72
2246
+ 23.6
2247
+ 95.62
2248
+ 6
2249
+ 2
2250
+ 4
2251
+ 1
2252
+ 0.01
2253
+ 51.8
2254
+ 32.1
2255
+ 83.91
2256
+ 11
2257
+ 8
2258
+ 10
2259
+ 1
2260
+ 0.04
2261
+ 57.1
2262
+ 24.5
2263
+ 81.64
2264
+ 16
2265
+ 5
2266
+ 9
2267
+ 1
2268
+ 0.04
2269
+ 56.9
2270
+ 33.5
2271
+ 90.44
2272
+ 1
2273
+ 2
2274
+ 3
2275
+ 4
2276
+ 5
2277
+ 0.1
2278
+ 1
2279
+ 10
2280
+ 100
2281
+ 1000
2282
+ 0.01
2283
+ 3600
2284
+ Response Time(s)
2285
+ Query Id
2286
+ Baseline
2287
+ TCD
2288
+ OTCD
2289
+ (a) CollegeMsg
2290
+ 6
2291
+ 7
2292
+ 8
2293
+ 9
2294
+ 10
2295
+ 0.1
2296
+ 1
2297
+ 10
2298
+ 100
2299
+ 1000
2300
+ 0.01
2301
+ 3600
2302
+ Response Time(s)
2303
+ Query Id
2304
+ Baseline
2305
+ TCD
2306
+ OTCD
2307
+ (b) email-Eu-core-temporal
2308
+ 11
2309
+ 12
2310
+ 13
2311
+ 14
2312
+ 15
2313
+ 0.1
2314
+ 1
2315
+ 10
2316
+ 100
2317
+ 1000
2318
+ 0.01
2319
+ 3600
2320
+ Response Time(s)
2321
+ Query Id
2322
+ Baseline
2323
+ TCD
2324
+ OTCD
2325
+ (c) sx-mathoverflow
2326
+ 16
2327
+ 17
2328
+ 18
2329
+ 19
2330
+ 20
2331
+ 0.1
2332
+ 1
2333
+ 10
2334
+ 100
2335
+ 1000
2336
+ 0.01
2337
+ 3600
2338
+ Response Time(s)
2339
+ Query Id
2340
+ Baseline
2341
+ TCD
2342
+ OTCD
2343
+ (d) sx-stackoverflow
2344
+ Figure 7: The comparison of response time for selected
2345
+ queries on SNAP graphs.
2346
+ researchers can reverify our experimental results or compare
2347
+ with our approach with the same queries.
2348
+ Figure 7 compares the response time of Baseline (iPHC-Query),
2349
+ TCD and OTCD algorithms for each selected query respectively,
2350
+ which clearly demonstrates the efficiency of ouralgorithm. Firstly,
2351
+ TCD performs better than baseline for all twenty queries due to
2352
+ the physical efficiency of TEL, though they both decrementally
2353
+ or incrementally induce temporal푘-cores. Specifically, TCD spends
2354
+ around 100 sec for each query. In contrast, baseline spends more
2355
+ than 1000 sec on CollegeMsg and even cannot finish within an
2356
+ hour on two other graphs, though it uses a precomputed in-
2357
+ dex. Furthermore, OTCD runs two or three orders of magnitude
2358
+ faster than TCD, and only spends about 0.1-1 sec for each query,
2359
+ which verifies the effectiveness of our pruning method based on
2360
+ TTI.
2361
+ To compare the effect of three pruning rules in OTCD algo-
2362
+ rithm, Table 4 lists their triggered times and the percentage of
2363
+ subintervals pruned by them for several queries respectively. PoR
2364
+ and PoU are triggered frequently because their conditions are
2365
+ more easily to be satisfied. However, PoR actually contributes
2366
+ pruned subintervals much less than the other two. Because it
2367
+ only prunes subintervals in the same row, and in contrast, PoU
2368
+ and PoL can prune an “area” of subintervals. Overall, the three
2369
+ pruning rules can achieve significant optimization effect together
2370
+
2371
+ Junyong Yang, Ming Zhong, Yuanyuan Zhu, Tieyun Qian, Mengchi Liu, and Jeffery Xu Yu
2372
+ TCD
2373
+ OTCD
2374
+ 0.01
2375
+ 0.1
2376
+ 1
2377
+ 10
2378
+ 100
2379
+ 1000
2380
+ Response Time(s)
2381
+ TCD- 25%~75%
2382
+ OTCD- 25%~75%
2383
+ Range within 1.5IQR
2384
+
2385
+ Median Line
2386
+
2387
+ Mean
2388
+ Outliers
2389
+ (a) Youtube
2390
+ TCD
2391
+ OTCD
2392
+ 0.1
2393
+ 1
2394
+ 10
2395
+ 100
2396
+ 1000
2397
+ Response Time(s)
2398
+ TCD- 25%~75%
2399
+ OTCD- 25%~75%
2400
+ Range within 1.5IQR
2401
+ Median Line
2402
+ Mean
2403
+ Outliers
2404
+ (b) Flickr
2405
+ Figure 8: The statistical distribution of response time for
2406
+ random queries on KONECT graphs.
2407
+ Table 5: Memory consumption of (O)TCD algorithm.
2408
+ Dataset
2409
+ Process Memory (GB)
2410
+ CollegeMsg
2411
+ 0.02
2412
+ sx-mathoverflow
2413
+ 0.06
2414
+ Youtube
2415
+ 1.7
2416
+ DBLP
2417
+ 3.1
2418
+ Flickr
2419
+ 3.5
2420
+ sx-stackoverflow
2421
+ 6.5
2422
+ by enabling OTCD algorithm to skip more than 80 percents of
2423
+ subintervals.
2424
+ To evaluate the stability of our approach, we conduct statis-
2425
+ tical analysis of one hundred valid random queries on two new
2426
+ graphs, namely, Youtube and Flickr. We visualize the distribution
2427
+ of response time of TCD and OTCD algorithms for these random
2428
+ queries as boxplots, which are shown by Figure 8. The boxplots
2429
+ demonstrate that the response time of OTCD varies in a very
2430
+ limited range, which indicates that the OTCD indeed performs
2431
+ stable in practice. The outliers represent some queries that have
2432
+ exceptionally more results, which can be seen as a normal phe-
2433
+ nomenon in reality. They may reveal that many communities of
2434
+ the social networks are more active during the period.
2435
+ Moreover, Table 5 reports the process memory consumption
2436
+ for different datasets, which depends on the size of TEL mostly.
2437
+ We can observe that, 1) for the widely-used graphs like Youtube,
2438
+ DBLP, Flickr and stackoverflow, several gigabytes of memory
2439
+ are needed for storing TEL, which is acceptable for the ordinary
2440
+ hardware; and 2) for the very large graphs with billions of edges,
2441
+ the size of TEL is hundreds of gigabytes approximately, which
2442
+ would require the distributed memory cluster like Spark.
2443
+ To verify the scalability of our method with respect to the
2444
+ query parameters, we test the three algorithms with varing min-
2445
+ imum degree 푘 and time span (namely, 푇푒 −푇푠) respectively.
2446
+ Impact of 푘. We select a typical query with span fixed and
2447
+ 푘 ranging from 2 to 6 for different graphs. The response time of
2448
+ tested algorithms are presented in Figure 9, from which we have
2449
+ an important observation against common sense. That is, differ-
2450
+ ent from core decomposition on non-temporal graphs, when the
2451
+ value of 푘 increases, the response time of TCD and OTCD algo-
2452
+ rithms decreases gradually. For OTCD, the behind rationale is
2453
+ clear, namely, its time cost is only bounded by the scale of re-
2454
+ sults, which decreases sharply with the increase of 푘. To sup-
2455
+ port the claim, Figure 10 and Figure 11 show the trend of the
2456
+ amount of result cores and connected components in the result
2457
+ cores changing with 푘. Intuitively, a greater value of k means
2458
+ a stricter constraint and thereby filters out some less cohesive
2459
+ 2
2460
+ 3
2461
+ 4
2462
+ 5
2463
+ 6
2464
+ 0.1
2465
+ 1
2466
+ 10
2467
+ 100
2468
+ 1000
2469
+ 0.01
2470
+ 3600
2471
+ Response Time(s)
2472
+ k
2473
+ Baseline
2474
+ TCD
2475
+ OTCD
2476
+ (a) CollegeMsg
2477
+ 2
2478
+ 3
2479
+ 4
2480
+ 5
2481
+ 6
2482
+ 0.1
2483
+ 1
2484
+ 10
2485
+ 100
2486
+ 1000
2487
+ 0.01
2488
+ 3600
2489
+ Response Time(s)
2490
+ k
2491
+ Baseline
2492
+ TCD
2493
+ OTCD
2494
+ (b) sx-mathoverflow
2495
+ 2
2496
+ 3
2497
+ 4
2498
+ 5
2499
+ 6
2500
+ 0.1
2501
+ 1
2502
+ 10
2503
+ 100
2504
+ 1000
2505
+ 0.01
2506
+ 3600
2507
+ Response Time(s)
2508
+ k
2509
+ Baseline
2510
+ TCD
2511
+ OTCD
2512
+ (c) sx-stackoverflow
2513
+ Figure 9: Trend of response time under a range of 푘.
2514
+ 2
2515
+ 3
2516
+ 4
2517
+ 5
2518
+ 6
2519
+ 10
2520
+ 1
2521
+ 10
2522
+ 2
2523
+ 10
2524
+ 3
2525
+ 10
2526
+ 4
2527
+ 10
2528
+ 5
2529
+ 10
2530
+ 6
2531
+ 10
2532
+ 0
2533
+ Quantity of Core
2534
+ k
2535
+ (a) CollegeMsg
2536
+ 2
2537
+ 3
2538
+ 4
2539
+ 5
2540
+ 6
2541
+ 10
2542
+ 1
2543
+ 10
2544
+ 2
2545
+ 10
2546
+ 3
2547
+ 10
2548
+ 4
2549
+ Quantity of Core
2550
+ k
2551
+ (b) sx-mathoverflow
2552
+ 2
2553
+ 3
2554
+ 4
2555
+ 5
2556
+ 6
2557
+ 10
2558
+ 0
2559
+ 10
2560
+ 1
2561
+ 10
2562
+ 2
2563
+ 10
2564
+ 3
2565
+ 10
2566
+ 4
2567
+ 10
2568
+ 5
2569
+ 10
2570
+ 6
2571
+ Quantity of Core
2572
+ k
2573
+ (c) sx-stackoverflow
2574
+ Figure 10: Trend of amount of distinct temporal 푘-cores
2575
+ under a range of 푘.
2576
+ 2
2577
+ 3
2578
+ 4
2579
+ 5
2580
+ 6
2581
+ 10
2582
+ 0
2583
+ 10
2584
+ 1
2585
+ 10
2586
+ 2
2587
+ 10
2588
+ 3
2589
+ 10
2590
+ 4
2591
+ 10
2592
+ 5
2593
+ 10
2594
+ 6
2595
+ Connected Component
2596
+ k
2597
+ (a) CollegeMsg
2598
+ 2
2599
+ 3
2600
+ 4
2601
+ 5
2602
+ 6
2603
+ 10
2604
+ 1
2605
+ 10
2606
+ 2
2607
+ 10
2608
+ 3
2609
+ 10
2610
+ 4
2611
+ 10
2612
+ 5
2613
+ 10
2614
+ 0
2615
+ 10
2616
+ 6
2617
+ Connected Component
2618
+ k
2619
+ (b) sx-mathoverflow
2620
+ 2
2621
+ 3
2622
+ 4
2623
+ 5
2624
+ 6
2625
+ 10
2626
+ 0
2627
+ 10
2628
+ 1
2629
+ 10
2630
+ 2
2631
+ 10
2632
+ 3
2633
+ 10
2634
+ 4
2635
+ 10
2636
+ 5
2637
+ 10
2638
+ 6
2639
+ Connected Component
2640
+ k
2641
+ (c) sx-stackoverflow
2642
+ Figure 11: Trend of amount of connected components in
2643
+ temporal 푘-cores under a range of 푘.
2644
+ cores. We can see the trend of runtime decrease for OTCD in
2645
+ Figure 9 is almost the same as the trend of core amount decrease
2646
+ in Figure 10, which also confirms the scalability of OTCD algo-
2647
+ rithm. For TCD, the behind rationale is complicated, since it enu-
2648
+ merates all subintervals and each single decomposition is more
2649
+ costly with a greater value of 푘. It is just like peeling an onion
2650
+ layer by layer, which has less layers with a greater value of 푘, so
2651
+ that the maintenance between layers become less.
2652
+ Impact of span. Similarly to the test of 푘, we also evalu-
2653
+ ate the scalability of different algorithms when the query time
2654
+ span increases. The results are presented in Figure 12. Although
2655
+ the number of subintervals increases quadratically, the response
2656
+ time of OTCD still increases moderately following the scale of
2657
+ query results. In contrast, TCD runs dramatically slower when
2658
+ the query time span becomes longer.
2659
+ The above results demonstrate that the efficiency of OTCD
2660
+ is not sensitive to the change of query parameters, so that it is
2661
+ scalable in terms of query time interval.
2662
+ Lastly, for a large graph with a long time span like Youtube,
2663
+ we test OTCD algorithm by querying temporal 10-cores over
2664
+ the whole time span. The result is, to find all 19,146 temporal
2665
+ 10-cores within 226 days, the OTCD algorithm spent about 55
2666
+ minutes, which is acceptable for such a “full graph scan” task.
2667
+
2668
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
2669
+ 24
2670
+ 36
2671
+ 48
2672
+ 60
2673
+ 72
2674
+ 0.1
2675
+ 1
2676
+ 10
2677
+ 100
2678
+ 1000
2679
+ 0.01
2680
+ 3600
2681
+ Response Time(s)
2682
+ Span(h)
2683
+ Baseline
2684
+ TCD
2685
+ OTCD
2686
+ (a) CollegeMsg
2687
+ 24
2688
+ 36
2689
+ 48
2690
+ 60
2691
+ 72
2692
+ 0.1
2693
+ 1
2694
+ 10
2695
+ 100
2696
+ 1000
2697
+ 0.01
2698
+ 3600
2699
+ Response Time(s)
2700
+ Span(h)
2701
+ Baseline
2702
+ TCD
2703
+ OTCD
2704
+ (b) sx-mathoverflow
2705
+ 24
2706
+ 36
2707
+ 48
2708
+ 60
2709
+ 72
2710
+ 0.1
2711
+ 1
2712
+ 10
2713
+ 100
2714
+ 1000
2715
+ 0.01
2716
+ 3600
2717
+ Response Time(s)
2718
+ Span(h)
2719
+ Baseline
2720
+ TCD
2721
+ OTCD
2722
+ (c) sx-stackoverflow
2723
+ Figure 12: Trend of response time under a range of span.
2724
+ 7.3
2725
+ Effectiveness
2726
+ The effectiveness of TCQ is two-fold. Firstly, by given a flexible
2727
+ time interval, we can find many temporal 푘-cores of different
2728
+ subintervals, each of which represents a community emerged in
2729
+ a specific period. Consider the above test on Youtube. Although
2730
+ it is not feasible to exhibit all 19,146 cores, Figure 13 shows their
2731
+ distribution by time span. The number of cores generally de-
2732
+ creases with the increase of time span, which makes sense be-
2733
+ cause there are always a lot of small communities emerged dur-
2734
+ ing short periods and then they will interact with each other and
2735
+ be merged to larger communities within a longer time span.
2736
+ Secondly, we can continue to filter and analyse the result cores
2737
+ to gain insights. For example, we record the date in GMT time
2738
+ for nine of the result cores with a time span less than one day in
2739
+ Youtube, and try to figure out if they emerged for some special
2740
+ reasons. Table 6 lists the date and size of the nine cores. We can
2741
+ see that there is a large core emerged on Dec 10, 2006, which
2742
+ means more than 40,000 accounts had nearly one million inter-
2743
+ actions with each other in just a day. That is definitely caused
2744
+ by a special event. While, most of the rest cores emerged during
2745
+ summer vacation, which may mean people have more interac-
2746
+ tions on Youtube in the period.
2747
+ 0
2748
+ 50
2749
+ 100
2750
+ 150
2751
+ 200
2752
+ 226
2753
+ 0
2754
+ 20
2755
+ 40
2756
+ 60
2757
+ 80
2758
+ 100
2759
+ 120
2760
+ 140
2761
+ 160
2762
+ 180
2763
+ Number of core
2764
+ Time span(days)
2765
+ Figure 13: Distribution of
2766
+ all
2767
+ temporal
2768
+ 10-cores
2769
+ in
2770
+ Youtube by time span.
2771
+ Table 6: The date and size
2772
+ of nine temporal 10-cores
2773
+ emerged within one day in
2774
+ Youtube.
2775
+ Date
2776
+ |V|
2777
+ |E|
2778
+ Dec 10 2006
2779
+ 46499
2780
+ 885128
2781
+ Feb 08 2007
2782
+ 1268
2783
+ 12054
2784
+ Mar 25 2007
2785
+ 21
2786
+ 139
2787
+ Jun 15 2007
2788
+ 98
2789
+ 713
2790
+ Jun 18 2007
2791
+ 20
2792
+ 100
2793
+ Jun 20 2007
2794
+ 124
2795
+ 1012
2796
+ Jun 30 2007
2797
+ 21
2798
+ 110
2799
+ Jul 02 2007
2800
+ 21
2801
+ 110
2802
+ Jul 06 2007
2803
+ 12
2804
+ 66
2805
+ 7.4
2806
+ Case Study
2807
+ For case study, we employ OTCD algorithm to query tempo-
2808
+ ral 10-cores on DBLP. The query interval is set as 2010 to 2018,
2809
+ which spans over 8 years. By statistics, there exist 43 temporal
2810
+ 10-cores during that period, with 39 of them containing the au-
2811
+ thor Jian Pei, for whom we further build an ego network from
2812
+ three selected cores in defferent years. Figure 14 shows the ego
2813
+ network. The authors in the three cores emerged in 2010, 2012
2814
+ and 2014 are shaded by red, yellow and blue respectively. By ob-
2815
+ serving the evolution of ego network over years, we can infer
2816
+ the change of author’s research interests or affiliations.
2817
+ 22 vertices of a 10-core
2818
+ arising in 2010
2819
+ 14 vertices of a 10-core
2820
+ arising in 2012
2821
+ 15 vertices of a 10-core
2822
+ arising in 2014
2823
+ Figure 14: Case Study in DBLP coauthorship network.
2824
+ A friendship community with 32 members arising in 2007
2825
+ 114 newly added members on the first day after
2826
+ 124 newly added members on the second day after
2827
+ Figure 15: Case Study in Youtube friendship network.
2828
+ To further demonstrate the potential of TCQ, we also employ
2829
+ TCQ to find temporal 푘-cores that expand quickly over time.
2830
+ This topic has been addressed in [5]. Since OTCD returns all
2831
+ distinct cores efficiently, we can conveniently achieve the goal
2832
+ by identifying the cores contained by other larger cores within
2833
+ a few of days from the results. Figure 15 shows such a bursting
2834
+ community on Youtube friendship network. The 32 central ver-
2835
+ tices colored in red comprise an initial temporal 10-core within
2836
+ two days. This core is contained by another core about four
2837
+ times larger, while the TTI of the larger core only expands by
2838
+ one day. The new vertices in the larger core are colored in or-
2839
+ ange. Then, the new vertices colored in yellow join them to com-
2840
+ prise a twice larger new core in the next day. Clearly, these three
2841
+ temporal 10-cores together represent a community that grows
2842
+ remarkably fast. In the real world, with more concrete informa-
2843
+ tion of graphs, such usages of TCQ will facilitate applications
2844
+ like recommendation, disease control, etc.
2845
+ 7.5
2846
+ Discussion on the value of 푘
2847
+ TCQ achieves relaxing the constraint on query time interval when
2848
+ composing푘-core queries on temporal graphs. However, the value
2849
+ of 푘 is still needed as an input parameter. We give a simple and
2850
+ rational criteria here for selecting the proper푘 value on different
2851
+ graphs, though many potential factors have different impacts on
2852
+ the selection. The criteria is based on two intuitive facts. Firstly,
2853
+ the number of distinct temporal 푘-cores over a given time in-
2854
+ terval will decrease with the increase of 푘. Secondly, the size
2855
+
2856
+ uCtelg
2857
+ Surya Nepal
2858
+ JianYin
2859
+ EnhongiChen
2860
+ Li Xiong
2861
+ Bin Jiang
2862
+ ShuhuiWang
2863
+ Jian Pei
2864
+ Qingming Huang
2865
+ Jiawei Han
2866
+ Siyuan Liu
2867
+ Ying Zhang
2868
+ Xindong Wu
2869
+ Jie Tang
2870
+ Kai Xu
2871
+ Chang Liu
2872
+ Xiang Wang
2873
+ Rong Jin
2874
+ Yang Wang
2875
+ Jinjun Chen
2876
+ Jeffrey Xu YuJiangchuan Liu
2877
+ Philip S. Yu
2878
+ Feng Zhao
2879
+ Ke Wang
2880
+ Xuemin Lin
2881
+ Jian Chen
2882
+ Hua Wang
2883
+ Kunbiu
2884
+ Wenjie Zhang
2885
+ KeYi
2886
+ XueLi
2887
+ Jin Huang
2888
+ QiangYang
2889
+ Wei Wang
2890
+ Hang Li
2891
+ Yu Yeng
2892
+ lunduoJunyong Yang, Ming Zhong, Yuanyuan Zhu, Tieyun Qian, Mengchi Liu, and Jeffery Xu Yu
2893
+ 3
2894
+ 5
2895
+ 2
2896
+ 4
2897
+ 6
2898
+ 10
2899
+ 1
2900
+ 10
2901
+ 2
2902
+ 10
2903
+ 3
2904
+ 10
2905
+ 4
2906
+ 10
2907
+ 5
2908
+ 10
2909
+ 6
2910
+ 10
2911
+ 0
2912
+ CollegeMsg
2913
+ Quantity of Core
2914
+ k
2915
+ Quantity of Core
2916
+ Average Size
2917
+ 25
2918
+ 50
2919
+ 75
2920
+ 100
2921
+ 125
2922
+ 150
2923
+ Average Core Size
2924
+ Figure 16: A statistical chart for selecting the value of 푘.
2925
+ of returned temporal 푘-cores will shrink with the increase of
2926
+ 푘. Normally, we expect the result cores to be concise and non-
2927
+ overlapping, especially when detecting the suspicious commu-
2928
+ nities that are inherently small and isolated, thereby preferring
2929
+ a greater value of 푘. However, the number of result cores also
2930
+ matters, which requires the value of 푘 not being too great, oth-
2931
+ erwise there could be too few results. Therefore, the selection of
2932
+ 푘 should take both size and number of result cores into account,
2933
+ just like the trade-off between precision and recall.
2934
+ For example, with 푘 ranging from 2 until 6, Figure 16 shows
2935
+ the falling curves of both number and average size of result cores
2936
+ over a specific time interval on CollegeMsg. We can observe that,
2937
+ setting 푘 = 5 should be a good choice, since the core size has
2938
+ declined to a relatively small level while the number of results
2939
+ is still fairly sufficient.
2940
+ 8
2941
+ RELATED WORK
2942
+ Recently, a variety of 푘-core query problems have been stud-
2943
+ ied on temporal graphs, which involve different temporal objec-
2944
+ tives or constraints in addition to cohesiveness. The most rele-
2945
+ vant work to ours is historical 푘-core query [36], which gives a
2946
+ fixed time interval as query condition. In contrast, our tempo-
2947
+ ral 푘-core query flexibly find cores of all subintervals. Moreover,
2948
+ Galimberti et al [12] proposed the span-core query, which also
2949
+ gives a time interval as query condition. However, the span-core
2950
+ requires all edges to appear in every moment within the query
2951
+ interval, which is too strict in practice. Actually, historical푘-core
2952
+ relaxes span-core, and temporal 푘-core further relaxes historical
2953
+ 푘-core.
2954
+ Besides, there are the following related work. Wu et al [34]
2955
+ proposed (푘,ℎ)-core and studied its decomposition algorithm,
2956
+ which gives an additional constraint on the number of parallel
2957
+ edges between each pair of linked vertices in the 푘-core, namely,
2958
+ they should have at least ℎ parallel edges. Li et al [19] proposed
2959
+ the persistent community search problem and gives a compli-
2960
+ cated instance called (휃,휏)-persistent 푘-core, which is a 푘-core
2961
+ persists over a time interval whose span is decided by the pa-
2962
+ rameters. Similarly, Li et al [21] proposed the continual cohe-
2963
+ sive subgraph search problem. Chu et al [5] studied the prob-
2964
+ lem of finding the subgraphs whose density accumulates at the
2965
+ fastest speed, namely, the subgraphs with bursting density. Qin
2966
+ et al [27, 28] proposed the periodic community problem to re-
2967
+ veal frequently happening patterns of social interactions, such
2968
+ as periodic 푘-core. Wen et al [1] relaxed the constraints of (푘,ℎ)-
2969
+ core and proposed quasi-(푘,ℎ)-core for better support of main-
2970
+ tenance. Lastly, Ma et al [25] studied the problem of finding
2971
+ dense subgraph on weighted temporal graph. These works all
2972
+ focus on some specific patterns of cohesive substructure on tem-
2973
+ poral graphs, and propose sophisticated models and methods.
2974
+ Compared with them, our work addresses a fundamental query-
2975
+ ing problem, which finds the most general 푘-cores on temporal
2976
+ graphs with respect to two basic conditions, namely, 푘 and time
2977
+ interval. As discussed in Section 6.2, we can extend TCQ to find
2978
+ the more specific 푘-cores by importing the constraints defined
2979
+ by them, because most of the definitions are special cases of tem-
2980
+ poral 푘-core, but not vice versa.
2981
+ Lastly, many research work on cohesive subgraph query for
2982
+ non-temporal graphs also inspire our work. We categorize them
2983
+ by the types of graphs as follows: undirected graph [3, 9, 13, 23,
2984
+ 35, 37], directed graph [4, 24, 30], labeled graph [6, 18, 31], attrib-
2985
+ uted graph [7, 14, 15, 26], spatial graph [8, 10, 39], heterageneous
2986
+ information network [11]. Besides, many work specific to bipar-
2987
+ tite graph [22, 32, 33, 38] also contain valuable insights.
2988
+ 9
2989
+ CONCLUSION AND FUTURE WORK
2990
+ For querying communities like푘-cores on temporal graphs, spec-
2991
+ ifying a time interval in which the communities emerge is the
2992
+ most fundamental query condition. To the best knowledge we
2993
+ have, we are the first to study a temporal 푘-core query that al-
2994
+ lows the users to give a flexible interval and returns all distinct 푘-
2995
+ cores emerging in any subintervals. Dealing with such a query in
2996
+ brute force is obviously costly due to the possibly large number
2997
+ of subintervals. Thus, we propose a novel decremental 푘-core
2998
+ inducing algorithm and the auxiliary optimization and imple-
2999
+ mentation methods. Our algorithm only enumerates the neces-
3000
+ sary subintervals that can induce a final result and reduces re-
3001
+ dundant computation between subintervals significantly. More-
3002
+ over, the algorithm is physically decomposed to a series of ef-
3003
+ ficient data structure manipulations. As a result, although our
3004
+ algorithm does not use any precomputed index, it still outper-
3005
+ forms an incremental version of the latest index-based approach
3006
+ by a remarkable margin. In conclusion, our algorithm is scalable
3007
+ with respect to the span of given time interval.
3008
+ In the future, we will study how to leverage our algorithm
3009
+ as a framework to integrate various temporal 푘-core analytics.
3010
+ There are a number of related work have considered different
3011
+ temporal constraints of 푘-cores, most of which can be combined
3012
+ with the time interval condition to offer more powerful function-
3013
+ ality. However, their query processing algorithms are essentially
3014
+ diverse. Therefore, we need to bridge the gap based on a general
3015
+ and scalable algorithm like ours.
3016
+ REFERENCES
3017
+ [1] Wen Bai, Yadi Chen, and Di Wu. 2020. Efficient temporal core maintenance
3018
+ of massive graphs. Information Sciences 513 (2020), 324–340.
3019
+ [2] Vladimir Batagelj and Matjaz Zaversnik. 2003. An O (m) algorithm for cores
3020
+ decomposition of networks. arXiv preprint cs/0310049 (2003).
3021
+ [3] Francesco Bonchi, Arijit Khan, and Lorenzo Severini. 2019.
3022
+ Distance-
3023
+ generalized core decomposition. In Proceedings of the 2019 International Con-
3024
+ ference on Management of Data. 1006–1023.
3025
+ [4] Yankai Chen, Jie Zhang, Yixiang Fang, Xin Cao, and Irwin King. 2021. Ef-
3026
+ ficient community search over large directed graphs: An augmented index-
3027
+ based approach. In Proceedings of the Twenty-Ninth International Conference
3028
+ on International Joint Conferences on Artificial Intelligence. 3544–3550.
3029
+ [5] Lingyang Chu, Yanyan Zhang, Yu Yang, Lanjun Wang, and Jian Pei. 2019. On-
3030
+ line density bursting subgraph detection from temporal graphs. Proceedings
3031
+ of the VLDB Endowment 12, 13 (2019), 2353–2365.
3032
+ [6] Zheng Dong, Xin Huang, Guorui Yuan, Hengshu Zhu, and Hui Xiong.
3033
+ 2021. Butterfly-core community search over labeled graphs. arXiv preprint
3034
+
3035
+ Scalable Time-Range 푘-Core Qery on Temporal Graphs
3036
+ arXiv:2105.08628 (2021).
3037
+ [7] Yixiang Fang, Reynold Cheng, Yankai Chen, Siqiang Luo, and Jiafeng Hu.
3038
+ 2017. Effective and efficient attributed community search. The VLDB Journal
3039
+ 26, 6 (2017), 803–828.
3040
+ [8] Yixiang Fang, Reynold Cheng, Xiaodong Li, Siqiang Luo, and Jiafeng Hu. 2017.
3041
+ Effective community searchover large spatialgraphs. Proceedings of the VLDB
3042
+ Endowment 10, 6 (2017), 709–720.
3043
+ [9] Yixiang Fang, Xin Huang, Lu Qin, Ying Zhang, Wenjie Zhang, Reynold Cheng,
3044
+ and Xuemin Lin. 2020. A survey of community search over big graphs. The
3045
+ VLDB Journal 29, 1 (2020), 353–392.
3046
+ [10] Yixiang Fang, Zheng Wang, Reynold Cheng, Xiaodong Li, Siqiang Luo, Ji-
3047
+ afeng Hu, and Xiaojun Chen. 2018. On spatial-awarecommunity search. IEEE
3048
+ Transactions on Knowledge and Data Engineering 31, 4 (2018), 783–798.
3049
+ [11] Yixiang Fang, Yixing Yang, Wenjie Zhang, Xuemin Lin, and Xin Cao. 2020. Ef-
3050
+ fective and efficient community search over large heterogeneous information
3051
+ networks. Proceedings of the VLDB Endowment 13, 6 (2020), 854–867.
3052
+ [12] Edoardo Galimberti, Alain Barrat, Francesco Bonchi, Ciro Cattuto, and
3053
+ Francesco Gullo. 2018. Mining (maximal) span-coresfrom temporal networks.
3054
+ In Proceedings of the 27th ACM international Conference on Information and
3055
+ Knowledge Management. 107–116.
3056
+ [13] Xin Huang, Hong Cheng, Lu Qin, Wentao Tian, and Jeffrey Xu Yu. 2014.
3057
+ Querying k-truss community in large and dynamic graphs. In Proceedings
3058
+ of the 2014 ACM SIGMOD international conference on Management of data.
3059
+ 1311–1322.
3060
+ [14] Xin Huang and Laks VS Lakshmanan. 2017. Attribute-driven community
3061
+ search. Proceedings of the VLDB Endowment 10, 9 (2017), 949–960.
3062
+ [15] Md Saiful Islam, Mohammed Eunus Ali, Yong-Bin Kang, Timos Sellis,
3063
+ Farhana M Choudhury, and Shamik Roy. 2022. Keyword aware influential
3064
+ community search in large attributed graphs. Information Systems 104 (2022),
3065
+ 101914.
3066
+ [16] Jérôme Kunegis. 2013. Konect: the koblenz network collection. In Proceedings
3067
+ of the 22nd international conference on world wide web. 1343–1350.
3068
+ [17] Jure Leskovec and Andrej Krevl. 2014. SNAP Datasets: Stanford Large Net-
3069
+ work Dataset Collection. http://snap.stanford.edu/data.
3070
+ [18] Rong-Hua Li, Lu Qin, Jeffrey Xu Yu, and Rui Mao. 2015. Influential commu-
3071
+ nity search in large networks. Proceedings of the VLDB Endowment 8, 5 (2015),
3072
+ 509–520.
3073
+ [19] Rong-Hua Li, Jiao Su, Lu Qin, Jeffrey Xu Yu, and Qiangqiang Dai. 2018. Persis-
3074
+ tent community search in temporal networks. In 2018 IEEE 34th International
3075
+ Conference on Data Engineering (ICDE). IEEE, 797–808.
3076
+ [20] Rong-Hua Li, Jeffrey Xu Yu, and Rui Mao. 2014. Efficient core maintenance in
3077
+ large dynamic graphs. IEEE Transactions on Knowledge and Data Engineering
3078
+ 26, 10 (2014), 2453–2465.
3079
+ [21] Yuan Li, Jinsheng Liu, Huiqun Zhao, Jing Sun, Yuhai Zhao, and Guoren Wang.
3080
+ 2021. Efficient continual cohesive subgraph search in large temporal graphs.
3081
+ World Wide Web 24, 5 (2021), 1483–1509.
3082
+ [22] Boge Liu, Long Yuan, Xuemin Lin, Lu Qin, Wenjie Zhang, and Jingren Zhou.
3083
+ 2019. Efficient (훼, 훽)-core computation: An index-based approach. In The
3084
+ World Wide Web Conference. 1130–1141.
3085
+ [23] Qing Liu, Xuliang Zhu, Xin Huang, and Jianliang Xu. 2021. Local algorithms
3086
+ for distance-generalized core decomposition over large dynamic graphs. Pro-
3087
+ ceedings of the VLDB Endowment 14, 9 (2021), 1531–1543.
3088
+ [24] Chenhao Ma, Yixiang Fang, Reynold Cheng, Laks VS Lakshmanan, Wenjie
3089
+ Zhang, and Xuemin Lin. 2020. Efficient algorithms for densest subgraph dis-
3090
+ covery on large directed graphs. In Proceedings of the 2020 ACM SIGMOD
3091
+ International Conference on Management of Data. 1051–1066.
3092
+ [25] Shuai Ma, Renjun Hu, Luoshu Wang, Xuelian Lin, and Jinpeng Huai. 2019. An
3093
+ efficient approach to finding dense temporal subgraphs. IEEE Transactions on
3094
+ Knowledge and Data Engineering 32, 4 (2019), 645–658.
3095
+ [26] Shohei Matsugu, Hiroaki Shiokawa, and Hiroyuki Kitagawa. 2019. Flexible
3096
+ community search algorithm on attributed graphs. In Proceedings of the 21st
3097
+ International Conference on Information Integration and Web-based Applica-
3098
+ tions & Services. 103–109.
3099
+ [27] Hongchao Qin, Ronghua Li, Ye Yuan, Guoren Wang, Weihua Yang, and Lu
3100
+ Qin. 2020. Periodic communities mining in temporal networks: Concepts and
3101
+ algorithms. IEEE Transactions on Knowledge and Data Engineering (2020).
3102
+ [28] Hongchao Qin, Rong-Hua Li, Guoren Wang, Lu Qin, Yurong Cheng, and Ye
3103
+ Yuan. 2019. Mining periodic cliques in temporal networks. In 2019 IEEE 35th
3104
+ International Conference on Data Engineering (ICDE). IEEE, 1130–1141.
3105
+ [29] Ahmet Erdem Sarıyüce, Buğra Gedik, Gabriela Jacques-Silva, Kun-Lung Wu,
3106
+ and Ümit V. Çatalyürek. 2016. Incremental k-core decomposition: algorithms
3107
+ and evaluation. The VLDB Journal 25 (2016), 425–447.
3108
+ [30] Mauro Sozio and Aristides Gionis. 2010. The community-search problem
3109
+ and how to plan a successful cocktail party. In Proceedings of the 16th ACM
3110
+ SIGKDD international conference on Knowledge discovery and data mining.
3111
+ 939–948.
3112
+ [31] Renjie Sun, Chen Chen, Xiaoyang Wang, Ying Zhang, and Xun Wang. 2020.
3113
+ Stable community detection in signed social networks. IEEE Transactions on
3114
+ Knowledge and Data Engineering (2020).
3115
+ [32] Kai Wang, Wenjie Zhang, Xuemin Lin, Ying Zhang, Lu Qin, and Yuting Zhang.
3116
+ 2021. Efficient and effective community search on large-scalebipartite graphs.
3117
+ In 2021 IEEE 37th International Conference on Data Engineering (ICDE). IEEE,
3118
+ 85–96.
3119
+ [33] Kai Wang, Wenjie Zhang, Ying Zhang, Lu Qin, and Yuting Zhang. 2021. Dis-
3120
+ covering significant communities on bipartite graphs: An index-based ap-
3121
+ proach. IEEE Transactions on Knowledge and Data Engineering (2021).
3122
+ [34] Huanhuan Wu, James Cheng, Yi Lu, Yiping Ke, Yuzhen Huang, Da Yan, and
3123
+ Hejun Wu. 2015. Core decomposition in large temporal graphs. In 2015 IEEE
3124
+ International Conference on Big Data (Big Data). IEEE, 649–658.
3125
+ [35] Kai Yao and Lijun Chang. 2021. Efficient size-bounded community search
3126
+ over large networks. Proceedings of the VLDB Endowment 14, 8 (2021), 1441–
3127
+ 1453.
3128
+ [36] Michael Yu, Dong Wen, Lu Qin, Ying Zhang, Wenjie Zhang, and Xuemin Lin.
3129
+ 2021. On querying historical k-cores. Proceedings of the VLDB Endowment 14,
3130
+ 11 (2021), 2033–2045.
3131
+ [37] Chen Zhang, Fan Zhang, Wenjie Zhang, Boge Liu, Ying Zhang, Lu Qin, and
3132
+ Xuemin Lin. 2020. Exploring finer granularity within the cores: Efficient (k,
3133
+ p)-core computation. In 2020 IEEE 36th International Conference on Data En-
3134
+ gineering (ICDE). IEEE, 181–192.
3135
+ [38] Yuting Zhang, Kai Wang, Wenjie Zhang, Xuemin Lin, and Ying Zhang. 2021.
3136
+ Pareto-optimal community search on large bipartite graphs. In Proceedings of
3137
+ the 30th ACM International Conference on Information & Knowledge Manage-
3138
+ ment. 2647–2656.
3139
+ [39] Qijun Zhu, Haibo Hu, Cheng Xu, Jianliang Xu, and Wang-Chien Lee. 2017.
3140
+ Geo-social group queries with minimum acquaintance constraints. The VLDB
3141
+ Journal 26, 5 (2017), 709–727.
3142
+
-dE2T4oBgHgl3EQfQgYx/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
.gitattributes CHANGED
@@ -474,3 +474,40 @@ UNAzT4oBgHgl3EQfXvzs/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -tex
474
  EtFLT4oBgHgl3EQfFS_K/content/2301.11987v1.pdf filter=lfs diff=lfs merge=lfs -text
475
  X9FRT4oBgHgl3EQf_DhA/content/2301.13693v1.pdf filter=lfs diff=lfs merge=lfs -text
476
  ItAzT4oBgHgl3EQfx_5k/content/2301.01746v1.pdf filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
474
  EtFLT4oBgHgl3EQfFS_K/content/2301.11987v1.pdf filter=lfs diff=lfs merge=lfs -text
475
  X9FRT4oBgHgl3EQf_DhA/content/2301.13693v1.pdf filter=lfs diff=lfs merge=lfs -text
476
  ItAzT4oBgHgl3EQfx_5k/content/2301.01746v1.pdf filter=lfs diff=lfs merge=lfs -text
477
+ 6tE1T4oBgHgl3EQfTQPr/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
478
+ 7dAzT4oBgHgl3EQf-f5K/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
479
+ X9FRT4oBgHgl3EQf_DhA/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
480
+ ndFQT4oBgHgl3EQfpzbH/content/2301.13378v1.pdf filter=lfs diff=lfs merge=lfs -text
481
+ YNAyT4oBgHgl3EQf9fo9/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
482
+ WdFJT4oBgHgl3EQf4i1g/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
483
+ CtAyT4oBgHgl3EQfefgG/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
484
+ CtAyT4oBgHgl3EQfefgG/content/2301.00320v1.pdf filter=lfs diff=lfs merge=lfs -text
485
+ ONFOT4oBgHgl3EQf3DT4/content/2301.12945v1.pdf filter=lfs diff=lfs merge=lfs -text
486
+ RtE0T4oBgHgl3EQfUQDk/content/2301.02249v1.pdf filter=lfs diff=lfs merge=lfs -text
487
+ kNE4T4oBgHgl3EQfTgwO/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
488
+ DNE1T4oBgHgl3EQfqAUl/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
489
+ ONFOT4oBgHgl3EQf3DT4/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
490
+ RtE0T4oBgHgl3EQfUQDk/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
491
+ 19FAT4oBgHgl3EQfDBzX/content/2301.08414v1.pdf filter=lfs diff=lfs merge=lfs -text
492
+ a9FST4oBgHgl3EQfCDgc/content/2301.13705v1.pdf filter=lfs diff=lfs merge=lfs -text
493
+ FtE3T4oBgHgl3EQfVwq1/content/2301.04463v1.pdf filter=lfs diff=lfs merge=lfs -text
494
+ YNAyT4oBgHgl3EQf9fo9/content/2301.00874v1.pdf filter=lfs diff=lfs merge=lfs -text
495
+ uNAzT4oBgHgl3EQfPvtT/content/2301.01188v1.pdf filter=lfs diff=lfs merge=lfs -text
496
+ ctFJT4oBgHgl3EQfRSzM/content/2301.11495v1.pdf filter=lfs diff=lfs merge=lfs -text
497
+ R9FRT4oBgHgl3EQfLjfi/content/2301.13503v1.pdf filter=lfs diff=lfs merge=lfs -text
498
+ ctFJT4oBgHgl3EQfRSzM/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
499
+ XNE3T4oBgHgl3EQfFwkF/content/2301.04307v1.pdf filter=lfs diff=lfs merge=lfs -text
500
+ EtFLT4oBgHgl3EQfFS_K/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
501
+ YtE2T4oBgHgl3EQfZAd1/content/2301.03860v1.pdf filter=lfs diff=lfs merge=lfs -text
502
+ R9FRT4oBgHgl3EQfLjfi/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
503
+ ndFQT4oBgHgl3EQfpzbH/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
504
+ t9AyT4oBgHgl3EQf0fl1/content/2301.00719v1.pdf filter=lfs diff=lfs merge=lfs -text
505
+ idE1T4oBgHgl3EQffwRJ/content/2301.03221v1.pdf filter=lfs diff=lfs merge=lfs -text
506
+ _NE2T4oBgHgl3EQfQwY9/content/2301.03773v1.pdf filter=lfs diff=lfs merge=lfs -text
507
+ JNFRT4oBgHgl3EQfzTgG/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
508
+ 5NFAT4oBgHgl3EQfFRzt/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
509
+ idE1T4oBgHgl3EQffwRJ/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
510
+ JNFRT4oBgHgl3EQfzTgG/content/2301.13649v1.pdf filter=lfs diff=lfs merge=lfs -text
511
+ K9E5T4oBgHgl3EQfYQ9F/content/2301.05572v1.pdf filter=lfs diff=lfs merge=lfs -text
512
+ FtE3T4oBgHgl3EQfVwq1/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
513
+ K9E5T4oBgHgl3EQfYQ9F/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
19AyT4oBgHgl3EQfbvd_/content/tmp_files/2301.00269v1.pdf.txt ADDED
@@ -0,0 +1,1475 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ WiFi Physical Layer Stays Awake and Responds
2
+ When it Should Not
3
+ Ali Abedi
4
+ Stanford University
5
+ USA
6
+ abedi@stanford.edu
7
+ Haofan Lu
8
+ UCLA
9
+ USA
10
+ haofan@cs.ucla.edu
11
+ Alex Chen
12
+ University of Waterloo
13
+ Canada
14
+ zihanchen.ca@gmail.com
15
+ Charlie Liu
16
+ University of Waterloo
17
+ Canada
18
+ charlie.liu@uwaterloo.ca
19
+ Omid Abari
20
+ UCLA
21
+ USA
22
+ omid@cs.ucla.edu
23
+ ABSTRACT
24
+ WiFi communication should be possible only between devices in-
25
+ side the same network. However, we find that all existing WiFi
26
+ devices send back acknowledgments (ACK) to even fake packets
27
+ received from unauthorized WiFi devices outside of their network.
28
+ Moreover, we find that an unauthorized device can manipulate the
29
+ power-saving mechanism of WiFi radios and keep them continu-
30
+ ously awake by sending specific fake beacon frames to them. Our
31
+ evaluation of over 5,000 devices from 186 vendors confirms that
32
+ these are widespread issues. We believe these loopholes cannot be
33
+ prevented, and hence they create privacy and security concerns.
34
+ Finally, to show the importance of these issues and their conse-
35
+ quences, we implement and demonstrate two attacks where an
36
+ adversary performs battery drain and WiFi sensing attacks just
37
+ using a tiny WiFi module which costs less than ten dollars.
38
+ 1
39
+ INTRODUCITON
40
+ Today’s WiFi networks use advanced authentication and encryption
41
+ mechanisms (such as WPA3) to protect our privacy and security
42
+ by stopping unauthorized devices from accessing our devices and
43
+ data. Despite all these mechanisms, WiFi networks remain vulner-
44
+ able to attacks mainly due to their physical layer behaviors and
45
+ requirements defined by WiFi standards. In this paper, we find two
46
+ loopholes in the IEEE 802.11 standard for the first time and show
47
+ how they can put our privacy and security at risk.
48
+ a) WiFi radios respond when they should not. In a WiFi
49
+ network, when a device sends a packet to another device, the re-
50
+ ceiving device sends an acknowledgment back to the transmitter.
51
+ In particular, upon receiving a frame, the receiver calculates the
52
+ cyclic redundancy check (CRC) of the packet in the physical layer
53
+ to detect possible errors. If it passes CRC, then the receiver sends
54
+ an Acknowledgment (ACK) to the transmitter to notify the correct
55
+ reception of the frame. Surprisingly, we have found that all existing
56
+ WiFi devices send back ACKs to even fake packets received from
57
+ unauthorized WiFi devices outside of their network. Why should a
58
+ WiFi device respond to a fake packet from an unauthorized device?!
59
+ b) WiFi radios stay awake when they should not. WiFi chipsets
60
+ are mostly in sleep mode to save power. However, to make sure
61
+ that they do not miss their incoming packets, they notify their WiFi
62
+ access point before entering sleep mode so that the access point
63
+ buffers any incoming packets for them. Then, WiFi devices wake up
64
+ periodically to receive beacon frames sent by the associated access
65
+ point. In regular operation, only the access point sends beacon
66
+ frames to notify the devices that have buffered packets. When a
67
+ device is notified, it stays awake to receive them. However, these
68
+ beacon frames are not encrypted. Hence, we find that an unautho-
69
+ rized user can forge those beacon frames to keep a specific device
70
+ awake for receiving the (non-existent) buffered frames.
71
+ We examine these behaviors and loopholes in detail over dif-
72
+ ferent WiFi chipsets from different vendors. Our examination of
73
+ over 5,000 WiFi devices from 186 vendors shows that these are
74
+ widespread issues. We then study the root cause of these issues
75
+ and show that, unfortunately, they cannot be fixed by a simple
76
+ solution such as updating WiFi chipsets firmware. Finally, we im-
77
+ plement and demonstrate two attacks based on these loopholes.
78
+ In the first attack, we show that by forcing WiFi devices to stay
79
+ awake and continuously transmit, an adversary can continuously
80
+ analyze the signal and extract personal information such as the
81
+ breathing rate of the WiFi users. In the second attack, we show that
82
+ by forcing WiFi devices to stay awake and continuously transmit,
83
+ the adversary can quickly drain the battery, and hence disable WiFi
84
+ devices such as home and office security sensors. These attacks
85
+ can be performed from outside buildings despite the WiFi network
86
+ and devices being completely secured. All the attacker needs is a
87
+ $10 microcontroller with integrated WiFi (such as ESP32) and a
88
+ battery bank. The attacker device can easily be carried in a pocket
89
+ or hidden somewhere near the target building.
90
+ The main contributions of this work are:
91
+ • We find that WiFi devices respond to fake 802.11 frames with
92
+ ACK, even when they are from unauthorized devices. We
93
+ also find that WiFi radios can be kept awake by sending them
94
+ fake beacon frames indicating they have packets waiting for
95
+ them.
96
+ • We study these loopholes and their root causes in detail, and
97
+ have tested more than 5,000 WiFi access points and client
98
+ devices from more than 186 vendors.
99
+ • We implement two attacks based on these loopholes using
100
+ just a 10-dollar off-the-shelf WiFi module and validate them
101
+ in real-world settings.
102
+ 2
103
+ RELATED WORK
104
+ The loopholes we present in this paper are explored using packet
105
+ injection, in which an attacker sends fake WiFi packets to devices in
106
+ a secured WiFi network. Packet injection has been used in the past
107
+ arXiv:2301.00269v1 [cs.NI] 31 Dec 2022
108
+
109
+ to perform various types of attacks against WiFi networks such as
110
+ denial of service attacks for a particular client device or total dis-
111
+ ruption of the network [14, 15, 17, 41]. These attacks use different
112
+ approaches such as beacon stuffing to send false information to
113
+ WiFi devices [21, 46], or Traffic Indication Map (TIM) forgery to
114
+ prevent clients from receiving data [18, 42]. However, all of these
115
+ attacks focus on spoofing 802.11 MAC-layer management frames
116
+ to interrupt the normal operation of WiFi networks. To provide a
117
+ countermeasure for some of these attacks, the 802.11w standard [7]
118
+ introduces a protected management frame that prevents attack-
119
+ ers from spoofing 802.11 management frames. Instead of spoofing
120
+ 802.11 MAC frames, we exploit properties of the 802.11 physical
121
+ layer to force a device to stay awake and respond when it should
122
+ not. These loopholes open the door to multiple research avenues
123
+ including new security and privacy threats.
124
+ WiFi sensing attack: Over the past decade, there has been a
125
+ significant amount of research on WiFi sensing where WiFi signals
126
+ are used to detect human activities [13, 32, 34–36, 38, 43–45, 48].
127
+ However, these systems target applications with social benefits
128
+ and cannot be easily used by an attacker to create privacy and
129
+ security threats. This is because either these techniques require
130
+ cooperation from the target WiFi device or the attacker needs to be
131
+ very close to the target to use these systems. A recent study shows
132
+ that by capturing WiFi signals coming out of a private building, it
133
+ is possible for an adversary to track user movements inside that
134
+ building [49]. However, this attack has a bootstrapping stage which
135
+ requires the attacker to walk around the target building for a long
136
+ time to find the location of the WiFi devices. Furthermore, since
137
+ this work relies on only the normal intermittent WiFi activities, it
138
+ cannot capture continuous data such as breathing rate.
139
+ Battery draining attack: Battery draining attacks date back to
140
+ 1999 [40] and there have been many studies on such attacks and
141
+ potential defense mechanisms since then [20]. Battery discharge
142
+ models and energy vulnerability due to operating systems have
143
+ been investigated [30, 47]. A more recent study plays multimedia
144
+ files implicitly to increase power consumption during web browsing
145
+ [27, 28]. In terms of defending, a monitoring agent that searches for
146
+ abnormal current draw is discussed in [19]. In contrast, our attack
147
+ exploits the loopholes in the 802.11 physical layer protocol and the
148
+ power-hungry WiFi transmission to quickly drain a target device’s
149
+ battery. We will discuss in Section 3.2 that stopping our proposed
150
+ attack is nearly impossible on today’s WiFi devices.
151
+ This paper is an extension of our previous workshop publica-
152
+ tion [9]. The workshop paper shows preliminary results for our
153
+ finding that WiFi devices respond with ACKs to packets received
154
+ from outside of their network, and provides a brief discussion on
155
+ potential privacy and security concerns of this behavior without
156
+ studying them. We have also explored how the WiFi power saving
157
+ mechanism can be exploited to keep a target device awake in a
158
+ localization attack [12]. In this paper, we provide an in-depth study
159
+ of these previously discovered loopholes. We also design and per-
160
+ form two privacy and security attacks, based on these loopholes.
161
+ Finally, we implement these attacks on off-the-shelve WiFi devices
162
+ and present detailed performance evaluations.
163
+ Figure 1: WiFi devices send an ACK for any frame they re-
164
+ ceive without checking if the frame is valid.
165
+ Figure 2: Frames exchanged between attacker and victim
166
+ 3
167
+ WIFI RESPONDS WHEN IT SHOULD NOT
168
+ Most networks use security protocols to prevent unauthorized de-
169
+ vices from communicating with their devices. Therefore, one may
170
+ assume that a WiFi device only acknowledges frames received from
171
+ the associated access point or other devices in the same network.
172
+ However, we have found that all today’s WiFi devices acknowledge
173
+ even the frames they receive from an unauthorized device from
174
+ outside of their network. In particular, as long as the destination
175
+ address matches their MAC address, their physical layer acknowl-
176
+ edges it, even if the frame has no valid payload. In this section, we
177
+ examine this behavior in more detail, and explain why this problem
178
+ happens and why it is not preventable.
179
+ To better understand this behavior, we run an experiment where
180
+ we use two WiFi devices to act as a victim and an attacker. The
181
+ attacker sends fake WiFi packets to the victim. We monitor the real
182
+ traffic between the attacker and the victim’s device.
183
+ Setup: For the victim, we use a tablet, and for the attacker, we
184
+ use a USB WiFi dongle that has a Realtek RTL8812AU 802.11ac
185
+ chipset. This is a $12 commodity WiFi device. The attacker uses
186
+ this device to send fake frames to the victim’s device. To do so,
187
+ we develop a python program that uses the Scapy library [37] to
188
+ create fake frames. Scapy is a python-based framework that can
189
+ generate arbitrary frames with custom data in the header fields.
190
+ Note, that the only valid information in the frame is the destination
191
+ MAC address (i.e., the victim’s MAC address). The transmitter MAC
192
+ address is set to a fake MAC address (i.e., aa:bb:bb:bb:bb:bb), and
193
+ the frame has no payload (i.e., null frame) and is not encrypted.
194
+ Result: Figure 2 shows the real traffic between the attacker and the
195
+ victim device captured using Wireshark packet sniffer [22]. As can
196
+ be seen, when the attacker sends a fake frame to the victim, the vic-
197
+ tim sends back an ACK to the fake MAC address (aa:bb:bb:bb:bb:bb).
198
+ This experiment confirms that WiFi devices acknowledge frames
199
+ without checking their validity. Finally, to see if this behavior exists
200
+ on other WiFi devices, we have repeated this test with a variety of
201
+ devices (such as laptops, smart thermostats, tablets, smartphones,
202
+ and access points) with different WiFi chipsets from different ven-
203
+ dors, as shown in Table 1. Note, target devices are connected to a
204
+ private network and the attacker does not have their secret key.
205
+ After performing the same experiment as before, we found that all
206
+ 2
207
+
208
+ Private WiFi Network
209
+ Acknowledgement
210
+ Fake 802.11
211
+ Data Frame
212
+ Access Point
213
+ Target
214
+ AttackerSource
215
+ Destination
216
+ Info
217
+ aa:bb:bb:bb:bb:bb
218
+ f2:6e:0b:
219
+ Null function(No data),
220
+ aa:bb:bb:bb:bb:bb...Acknowledqement,Flaqs=.Device
221
+ WiFi module
222
+ Standard
223
+ MSI GE62 laptop
224
+ Intel AC 3160
225
+ 11ac
226
+ Ecobee3 thermostat
227
+ Atheros
228
+ 11n
229
+ Surface Pro 2017
230
+ Marvel 88W8897
231
+ 11ac
232
+ Samsung Galaxy S8
233
+ Murata KM5D18098
234
+ 11ac
235
+ Google Wifi AP
236
+ Qualcomm IPQ 4019
237
+ 11ac
238
+ Table 1: List of tested chipsets/devices
239
+ of these devices also respond to fake packets received from a device
240
+ outside of their network.
241
+ 3.1
242
+ How widespread is this loophole?
243
+ In the previous section, we examined a few different WiFi devices
244
+ and showed that they are all responding to fake frames from unau-
245
+ thorized devices. Here, we examine thousands of devices to see how
246
+ widespread this behavior is. In the following, we explain the setup
247
+ and results of this experiment.
248
+ Setup: To examine thousands of devices, we mounted a WiFi dongle
249
+ on the roof of a vehicle and drove around the city to test all nearby
250
+ devices. For the WiFi dongle, we use the same Realtek RTL8812AU
251
+ USB WiFi dongle, and connect it to a Microsoft Surface, running
252
+ Ubuntu 18.04. We develop a multi-threaded program using the
253
+ Scapy library [37] to discover nearby devices, send fake 802.11
254
+ frames to the discovered devices, and verify that target devices re-
255
+ spond to our fake frames. Specifically, our implementation contains
256
+ three threads. The first thread discovers nearby devices by sniffing
257
+ WiFi traffic and adding the MAC address of unseen devices to a
258
+ target list. The second thread sends fake 802.11 frames to the list of
259
+ target devices. Finally, the third thread checks to verify that target
260
+ devices respond with an ACK.
261
+ Results: We perform this experiment for one hour while driving
262
+ around the city. In total, we discovered 5,328 WiFi nodes from
263
+ 186 vendors. The list includes 1,523 different WiFi client devices
264
+ from 147 vendors and 3,805 access points from 94 vendors. Table 2
265
+ shows the top 20 vendors for WiFi devices and WiFi access points
266
+ in terms of the number of devices discovered in our experiment.
267
+ The list includes devices from major smartphone manufacturers
268
+ (such as Apple, Google, and Samsung) and major IoT vendors (such
269
+ as Nest, Google, Amazon, and Ecobee). We found that all 5,328 WiFi
270
+ Access Points and devices responded to our fake 802.11 frames with
271
+ an acknowledgment, and hence we infer that most probably all
272
+ of today’s WiFi devices and access points respond to fake frames
273
+ when they should not.
274
+ 3.2
275
+ Can this loophole be fixed?
276
+ So far, we have demonstrated that all existing WiFi devices respond
277
+ to fake packets received from unauthorized WiFi devices outside of
278
+ their network. Now, the next question is why this behavior exists,
279
+ and if it can be prevented in future WiFi chipsets.
280
+ In a WiFi device, when the physical layer receives a frame, it
281
+ checks the correctness of the frame using error-checking mech-
282
+ anisms (such as CRC) and transmits an ACK if the frame has no
283
+ error. However, checking the validity of the content of a frame is
284
+ WiFi Client Device
285
+ WiFi Access Point
286
+ Vendor
287
+ # devices
288
+ Vendor
289
+ # devices
290
+ Apple
291
+ 143
292
+ Hitron
293
+ 723
294
+ Google
295
+ 102
296
+ Sagemcom
297
+ 601
298
+ Intel
299
+ 66
300
+ Technicolor
301
+ 410
302
+ Hitron
303
+ 65
304
+ eero
305
+ 195
306
+ HP
307
+ 63
308
+ Extreme N.
309
+ 188
310
+ Samsung
311
+ 56
312
+ Cisco
313
+ 156
314
+ Espressif
315
+ 47
316
+ HP
317
+ 104
318
+ Hon Hai
319
+ 46
320
+ TP-LINK
321
+ 101
322
+ Amazon
323
+ 41
324
+ Google
325
+ 80
326
+ Sagemcom
327
+ 38
328
+ D-Link
329
+ 75
330
+ Liteon
331
+ 33
332
+ NETGEAR
333
+ 69
334
+ AzureWave
335
+ 30
336
+ ASUSTek
337
+ 51
338
+ Sonos
339
+ 30
340
+ Aruba
341
+ 46
342
+ Nest Labs
343
+ 27
344
+ SmartRG,
345
+ 44
346
+ Murata
347
+ 24
348
+ Ubiquiti N.
349
+ 35
350
+ Belkin
351
+ 20
352
+ Zebra
353
+ 35
354
+ TP-LINK
355
+ 20
356
+ Pegatron
357
+ 28
358
+ Cisco
359
+ 16
360
+ Belkin
361
+ 25
362
+ ecobee
363
+ 13
364
+ Mitsumi
365
+ 25
366
+ Microsoft
367
+ 13
368
+ Apple
369
+ 19
370
+ Others
371
+ 630
372
+ Others
373
+ 789
374
+ Total
375
+ 1523
376
+ Total
377
+ 3805
378
+ Table 2: List of WiFi devices and APs that respond to our
379
+ fake 802.11 frames.
380
+ performed by the MAC and higher layers. Unfortunately, this sepa-
381
+ ration of responsibilities and the fact that the physical layer does
382
+ not coordinate with higher layers about sending ACKs seem to be
383
+ the root cause of the behavior. In particular, we have observed that
384
+ when some access points receive fake frames, they start sending
385
+ deauthentication frames to the attacker, requesting it to leave the
386
+ network. These access points detect the attacker as a “malfunc-
387
+ tioning” device and that is why they send deauthentication frames.
388
+ Surprisingly, although the access points have detected that they are
389
+ receiving fake frames from a “malfunctioning” device, we found
390
+ that they still acknowledge the fake frames.
391
+ An example traffic that demonstrates this behavior is shown in
392
+ Figure 3. As can be seen, although the access point has already sent
393
+ three deauthentication frames to the attacker, it still acknowledges
394
+ the attacker’s fake frame. We then manually blocked the attacker’s
395
+ fake MAC address on the access point. Surprisingly, we observed
396
+ that the AP still acknowledges the fake frames. These observations
397
+ verify that sending ACK frames happens automatically in the physi-
398
+ cal layer without any communication with higher layers. Therefore,
399
+ the software running on the access points does not prevent the
400
+ physical layer from sending ACKs to fake frames.
401
+ The next question is why the software running on WiFi devices
402
+ does not prevent this behavior by verifying if the frame is legitimate
403
+ before sending an ACK. Unfortunately, this is not possible due to the
404
+ WiFi standard timing requirements. Specifically, in the IEEE 802.11
405
+ standard, upon receiving a frame, an ACK must be transmitted
406
+ 3
407
+
408
+ Figure 3: The attacked access point detects that something
409
+ strange is happening, however it still ACKs fake frames
410
+ by the end of the Short Interframe Space (SIFS)1 interval which is
411
+ 10 𝜇s and 16 𝜇s for the 2.4 GHz and 5 GHz bands, respectively. If the
412
+ transmitter does not receive an ACK by the end of SIFS, it assumes
413
+ that the frame has been lost and retransmits the frame. Therefore,
414
+ the WiFi device nefeds to verify the validity of the received frame
415
+ in less than 10 𝜇𝑠. This verification must be done by decoding
416
+ the frame using the secret shared key. Unfortunately, decoding a
417
+ frame in such a short period is not possible. In particular, past work
418
+ has shown that the time required to decode a frame is between
419
+ 200 to 700 𝜇𝑠 when the WPA2 security protocol is used [31, 33,
420
+ 39]. This processing time is orders of magnitude longer than SIFS.
421
+ Hence, existing devices cannot verify the validity of the frame
422
+ before sending the ACK, and they acknowledge a frame as long
423
+ as it passes the error detection check. One potential approach to
424
+ solve this loophole is to implement the security decoder in WiFi
425
+ hardware instead of software to significantly speed up its delay.
426
+ Although this may solve the problem in future WiFi chipsets, it will
427
+ not fix the problem in billions of WiFi chipsets which are already
428
+ deployed.
429
+ 4
430
+ WIFI STAYS AWAKE WHEN IT SHOULD
431
+ NOT
432
+ We have also found a loophole that allows an unauthorized device
433
+ to keep a WiFi device awake all the time. One may think that a
434
+ WiFi device can be kept awake by just sending fake back-to-back
435
+ packets to it and forcing it to transmit acknowledgment. However,
436
+ this approach does not work. Most WiFi radios go to sleep mode
437
+ to save energy during inactive states such as screen lock, during
438
+ which the attacker is not able to keep them awake by sending back-
439
+ to-back packets. Figure 4a show the results of an experiment where
440
+ the attacker is continuously transmitting fake packets to a WiFi
441
+ device. In this figure, we plot the amplitude of CSI over time for
442
+ the ACK packets received from the WiFi device. As can be seen,
443
+ the responses are sparse and discontinued even when the attacker
444
+ sends back-to-back packets to the WiFi device. This is because the
445
+ WiFi device goes to sleep mode frequently. However, we have found
446
+ a loophole in the power saving mechanism of WiFi devices which
447
+ can be used by an unauthorized device to keep any WiFi device
448
+ awake all the time.
449
+ 1The SIFS is used in the 802.11 standard to give the receiver time to go through different
450
+ procedures before it is ready to send the ACK. These procedures include Physical-layer
451
+ and MAC-layer header processing, creating the waveform for the ACK, and switching
452
+ the RF circuit from receiving to transmitting mode.
453
+ (a) Without fake beacon frames
454
+ (b) With fake beacon frames
455
+ Figure 4: The CSI amplitude of ACKs responded by the tar-
456
+ get device when an attacker sends back-to-back fake packets
457
+ to it in two scenarios. (a) In this scenario, the attacker is not
458
+ using fake beacon frames. Therefore, the target device goes
459
+ to sleep mode frequently and does not respond to fake pack-
460
+ ets. (b) In this scenario, the attacker infrequently sends fake
461
+ beacon frames to keep the target device awake all the time.
462
+ 4.1
463
+ How does WiFi power saving mechanism
464
+ work?
465
+ Wireless tranceivers are very power-hungry. Therefore, WiFi radios
466
+ spend most of the time in the sleep mode to save power. When a
467
+ WiFi radio is in sleep mode, it cannot send or receive WiFi packets.
468
+ To avoid missing any incoming packets, when a WiFi device wants
469
+ to enter the sleep mode it notifies the WiFi access point so that
470
+ the access point buffers any incoming packets for this device. WiFi
471
+ devices, however, wake up periodically to receive beacon frames
472
+ to find out if packets are waiting for them. In particular, WiFi
473
+ access points broadcast beacon frames periodically which includes a
474
+ Traffic Indication Map (TIM) field that indicates which devices have
475
+ buffered packets on the access point. For example, if the association
476
+ ID of a WiFi device is 𝑥, then the (𝑥 + 1)𝑡ℎ bit of TIM is assigned to
477
+ that device. Finally, when a device is notified that has some buffered
478
+ packets on the access point, it stays awake and replies with a Null-
479
+ function packet with a power management bit set to "0". In this way,
480
+ the WiFi device informs the access point it is awake and ready to
481
+ receive packets.
482
+ 4.2
483
+ How can one manipulate power saving?
484
+ We have found that an unauthorized device can use the power-
485
+ saving mechanism of WiFi devices to force them to stay awake.
486
+ In particular, an attacker can pretend to be the access point and
487
+ broadcasts fake beacon frames indicating that the WiFi device has
488
+ buffered traffic, forcing them to stay awake. However, this requires
489
+ the attacker to know the MAC address and the SSID of the network’s
490
+ access point, as well as the association ID and MAC address of the
491
+ targeted device so that it can set the correct bit in TIM. The access
492
+ point MAC address and SSID can be easily discovered by sniffing
493
+ 4
494
+
495
+ Source
496
+ Destination
497
+ Info
498
+ f2:6e:0b:
499
+ aa:bb:bb:bb:bb:bb
500
+ Deauthentication,
501
+ SN=3275
502
+ f2:6e:0b:
503
+ aa:bb:bb:bb:bb:bb
504
+ Deauthentication,
505
+ SN=3275
506
+ f2:6e:0b:
507
+ aa:bb:bb:bb:bb:bb
508
+ Deauthentication,
509
+ SN=3275
510
+ aa:bb:bb:bb:bb:bb
511
+ f2:6e:0b:
512
+ Null function (No data),
513
+ aa:bb:bb:bb:bb:bb
514
+ Acknowledgement,
515
+ Flags=..
516
+ f2:6e:0b:
517
+ aa:bb:bb:bb:bb:bb
518
+ Deauthentication,
519
+ SN=3281
520
+ f2:6e:0b:
521
+ aa:bb:bb:bb:bb:bb
522
+ Deauthentication,
523
+ SN=328125
524
+ CSI Amplitude
525
+ 20
526
+ 15
527
+ 10
528
+ 5
529
+ 0
530
+ 0
531
+ 5
532
+ 10
533
+ 15
534
+ 20
535
+ 25
536
+ 30
537
+ Time (s)25
538
+ CSI Amplitude
539
+ 20
540
+ 15
541
+ 10
542
+ 5
543
+ 0
544
+ 0
545
+ 5
546
+ 10
547
+ 15
548
+ 20
549
+ 25
550
+ 30
551
+ Time (s)Figure 5: WiFi devices stay awake on hearing a forged bea-
552
+ con frame with TIM flags set up.
553
+ the WiFi traffic using software such as Wireshark since the MAC
554
+ address is never encrypted and all nodes send packets to the access
555
+ point. Note that MAC randomization does not cause any problem
556
+ for this process because the attacker finds the randomized MAC
557
+ address that is currently being used. Next, the attacker pretends to
558
+ be the access point and broadcasts fake beacon frames with TIM set
559
+ to "0xFF", indicating all client devices have buffered traffic. Then, it
560
+ enters the sniffing mode to sniff for the Null-function packets. The
561
+ null-function packets contain the ID and MAC addresses of all WiFi
562
+ devices. To avoid keeping all WiFi devices awake, we find that one
563
+ can send a fake beacon frame as a unicast packet, instead of the
564
+ usual broadcast beacons. This way only the target device receives
565
+ the packet and we do not interfere with the operation of other
566
+ devices. Interestingly, our experiments show that target devices do
567
+ not care if they receive beacons as broadcast or unicast frames.
568
+ To better understand this behavior, we run an experiment where
569
+ we use two WiFi devices to act as a victim and an attacker, re-
570
+ spectively. The attacker sends fake WiFi packets to the victim. We
571
+ monitor the real traffic between the attacker and the victim’s device.
572
+ Setup: Similar to the experiment described in Section 3, we use an
573
+ RTL8812AU USB dongle to inject fake packets to a smartphone held
574
+ by a person who is watching YouTube on the phone. The distance
575
+ between the smartphone and the user is about 60 cm. The attacking
576
+ device and the victim are in two separate rooms. The attacker also
577
+ uses an ESP32 WiFi module to record the Channel State Information
578
+ (CSI) of received ACKs.
579
+ Result: We find that although sending fake beacon frames keeps
580
+ the target device awake, sending them very frequently will cause
581
+ WiFi devices to recognize the suspicious attacker’s behavior and
582
+ disconnect from it. Therefore, to keep the WiFi device awake, in-
583
+ stead of just sending beacon frames back-to-back, the attacker can
584
+ continuously transmit normal fake packets to a WiFi device and
585
+ periodically sends fake beacon frames to keep it awake. Figure 4b
586
+ shows the result of an experiment where the attacker is continu-
587
+ ously transmitting fake packets to a WiFi device and periodically
588
+ sends fake beacon frames. As it can be seen, the target device is
589
+ continuously awake and responding to fake packets with ACKs.
590
+ 5
591
+ PRIVACY IMPLICATION: WIFI SENSING
592
+ ATTACK
593
+ Recently, there has been a significant amount of work on WiFi
594
+ sensing technologies that use WiFi signals to detect events such as
595
+ motion, gesture, and breathing rate. In this section, we show how
596
+ an adversary can combine WiFi sensing techniques with the above
597
+ loopholes to monitor people’s breathing rate whenever she/he
598
+ wants from outside buildings despite the WiFi network and de-
599
+ vices being completely secured. In particular, an adversary can
600
+ force our WiFi devices to stay awake and continuously transmit
601
+ WiFi signals. Then she/he can continuously analyze our signals
602
+ and extract information such as our breathing rate and presents.
603
+ Note, since most of the time, we are close to a WiFi device (such as
604
+ a smartwatch, laptop, or tablet), our body will change the ampli-
605
+ tude and phase of the signals which can be easily extracted by the
606
+ adversary.
607
+ 5.1
608
+ Attack Design, Scenarios and Setup
609
+ 5.1.1
610
+ Attack Design. The attacker sends fake packets to a WiFi
611
+ device in the target property and pushes it to transmit ACK packets.
612
+ In particular, since an adult’s normal breathing rate is around 12 -20
613
+ times per minute (i.e., 0.2- 0.33Hz), receiving several ACK packets
614
+ per second is sufficient for the attacker to estimate the breathing
615
+ rate, without impacting the performance of the target WiFi network.
616
+ The attacker then takes the Fourier transform of the CSI information
617
+ of ACK packets to estimate the breathing rate of the person who
618
+ is nearby the WiFi device. However, due to the random delays
619
+ of the WiFi random access protocol and the operating system’s
620
+ scheduling protocol, the collected data samples are not uniformly
621
+ spaced in time. Hence, the attacker cannot simply use standard
622
+ FFT to estimate the breathing rate. Instead, they need to use a non-
623
+ uniform Fourier transform, and a voting algorithm to extract the
624
+ breathing rate. The Non-Uniform Fast Fourier Transform (NUFFT)
625
+ algorithm 1 used is shown below.
626
+ Algorithm 1: Non-uniform FFT
627
+ Data: Time indices 𝑡, data samples 𝑥 of length 𝑛
628
+ Result: Magnitude of each frequency component
629
+ 𝑑 ← min𝑖 (𝑡𝑖 − 𝑡𝑖−1)
630
+ 𝑖 = 1, 2, ...,𝑛.;
631
+ for 𝑖 ← 1 to 𝑛 − 1 do
632
+ 𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙 ← 𝑡 [𝑖] − 𝑡 [𝑖 − 1];
633
+ if 𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙 > 𝑑 then
634
+ 𝑐𝑜𝑢𝑛𝑡 ← ⌊𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙/𝑑⌋;
635
+ Interpolation(𝑡, 𝑥, 𝑡 [𝑖], 𝑡 [𝑖 − 1], 𝑐𝑜𝑢𝑛𝑡);
636
+ end
637
+ end
638
+ return FFT(𝑡, 𝑥)
639
+ The algorithm first finds the minimum time gap between any two
640
+ adjacent data points 𝑑, then linearly interpolates any interval that
641
+ is larger than the gap with ⌊𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙/𝑑⌋$ samples. Finally, it uses
642
+ a regular FFT algorithm to find the magnitude of each frequency
643
+ component. A low-pass filter is applied before feeding data to the
644
+ FFT analysis to reduce noise (not shown in the algorithm).
645
+ Figure 6(a) and 6(b) show the amplitude of CSI before and after
646
+ interpolation, respectively, when the attacker sends 10 packets per
647
+ second to a WiFi device that is close to the victim. Each figure shows
648
+ both the original data (in blue) and the filtered data (in orange).
649
+ Figure 6(c) shows the frequency spectrum of the same signals when
650
+ a standard FFT or our non-uniform FFT is applied. A prominent
651
+ peak at 0.3Hz is shown in the non-uniform FFT spectrum, indicating
652
+ a breathing rate of 18 bpm.
653
+ 5
654
+
655
+ Private WiFi Network
656
+ Stay Awake
657
+ Fake 802.11
658
+ Beacon
659
+ Access Point
660
+ Target
661
+ Attacker(a) Raw and filtered data before
662
+ interpolation
663
+ (b) Raw and filtered data after
664
+ interpolation
665
+ 0.0
666
+ 0.5
667
+ 1.0
668
+ 1.5
669
+ 2.0
670
+ 2.5
671
+ 3.0
672
+ Frequency (Hz)
673
+ 0.00
674
+ 0.25
675
+ 0.50
676
+ 0.75
677
+ 1.00
678
+ 1.25
679
+ 1.50
680
+ 1.75
681
+ 2.00
682
+ Power
683
+ non_uniform_fft
684
+ standard_fft
685
+ (c) Standard FFT and a non-uniform FFT of
686
+ Data
687
+ Figure 6: Steps to extract breathing rate from the CSI.
688
+ WiFi CSI gives us the amplitude of 52 subcarriers per packet.
689
+ We observed that these subcarriers are not equally sensitive to the
690
+ motion of the chest. Besides, a subcarrier’s sensitivity may vary
691
+ depending on the surrounding environment. For a more reliable
692
+ attack, the attacker should identify the most sensitive subcarriers
693
+ over a sampling window. Previously proposed voting mechanisms
694
+ for coarse-grained motion detection applications [8, 16, 29, 49]
695
+ cannot be directly applied in this situation, as chest motion during
696
+ respiration is at a much smaller scale. Instead, we developed a soft
697
+ voting mechanism, where each subcarrier gives a weighted vote
698
+ to a breathing rate value. The breathing rate that gets the most
699
+ votes is reported. Specifically, We first find the power of the highest
700
+ peak (𝑃𝑝𝑒𝑎𝑘), and then calculate the average power of the rest bins
701
+ (𝑃𝑎𝑣𝑒). The exponent of the Peak-to-Average Ratio (PAR): 𝑒
702
+ 𝑓𝑝𝑒𝑎𝑘
703
+ 𝑓𝑎𝑣𝑒 is
704
+ used as the weight of the corresponding subcarrier. In this way, we
705
+ guarantee the subcarriers with higher SNR have significantly more
706
+ votes than the rest of the subcarriers.
707
+ 5.1.2
708
+ Attack Scenarios. We evaluate the WiFi sensing attack in
709
+ different scenarios, both indoor and outdoor. In the indoor scenario,
710
+ the attacker and the target are placed in the same building but on
711
+ different floors. The height of one floor in the building is around
712
+ 2.8 m. This scenario is similar to when the attacker and the target
713
+ person are in different units of an apartment or townhouse. In the
714
+ outdoor scenario, the attacker is outside the target’s house. For the
715
+ outdoor experiments, We place the attacker in another building
716
+ which is around 20 m away from the target building. In all of the
717
+ experiments, the target WiFi devices are placed 0.5 to 1.4 m away
718
+ from the person’s body. The person is either watching a movie,
719
+ typing on a laptop, or surfing the web using his cell phone. During
720
+ the experiments, other people are walking and living normally in
721
+ the house. Finally, we run the attack and compare the estimated
722
+ breathing rate with the ground truth. To obtain the ground truth,
723
+ we record the target person’s breathing sound by attaching a mi-
724
+ crophone near his/her mouth [23]. We then calculate the FFT on
725
+ the sound signal to measure the breathing frequency. Note that the
726
+ attack does not need this information and this is just to obtain the
727
+ ground truth in our experiments.
728
+ 5.1.3
729
+ Attacker Setup. Hardware Setup: The attacker uses a Linksys
730
+ AE6000 WiFi card and an ESP32 WiFi module [25] as the attacking
731
+ device. Both devices are connected to a ThinkPad laptop via USB.
732
+ The Linksys AE6000 is used to send fake packets and the ESP32
733
+ WiFi module is used to receive acknowledgments (ACK) and extract
734
+ CSI. Although we use two different devices for sending and receiv-
735
+ ing, one can simply use an ESP32 WiFi module for both purposes.
736
+ The use of two separate modules gave us more flexibility in run-
737
+ ning many experiments. As for the target device, we use a One Plus
738
+ 8T smartphone without any software or hardware modifications.
739
+ We have also tested our attack on an unmodified Lenovo laptop, a
740
+ Microsoft Surface Pro 4 laptop, and a USB WiFi card as the target
741
+ device and we obtained similar results. It is worth mentioning that
742
+ any WiFi device can be a target without any software or hardware
743
+ modification.
744
+ Software Setup: We have implemented the CSI collecting script
745
+ on the ESP32 WiFi module, and the breathing rate estimation algo-
746
+ rithm on the laptop. The collected CSI data is fed to the algorithm
747
+ which produces the breathing rate estimation values in real-time.
748
+ To process this data in real time, a sliding window (buffer) is used.
749
+ The size of the window is 30 s and the stride step is 1 s. 30 seconds
750
+ is a large enough window for estimating a stable breathing rate
751
+ value. Note that an adult breathes around 6 times during such a
752
+ window. The window is a queue of data points, and it updates every
753
+ second by including 1 second of new data points to its head and
754
+ removing 1 second of old data points from its tail. The breathing
755
+ rate estimation runs the analysis algorithm on the data points inside
756
+ the window whenever it is updated. The window slides once per
757
+ second. Hence, our software reports an estimation of breathing rate
758
+ every second. Note that there is a 30-second delay at the beginning
759
+ since the window needs to be filled first.
760
+ 5.2
761
+ Results
762
+ We evaluate the effectiveness of the attack in different scenarios
763
+ such as when the attacker and the target are in the same building
764
+ or different buildings.
765
+ 5.2.1
766
+ Accuracy in Detecting Breathing Rate. Same Building Sce-
767
+ nario: First, we evaluate the accuracy of the attack by estimating
768
+ 6
769
+
770
+ 40
771
+ Raw Data
772
+ 35
773
+ Filtered Data
774
+ 30
775
+ 1Amplitude
776
+ 25
777
+ 20
778
+ CSI
779
+ 15
780
+ 10
781
+ 5
782
+ 0
783
+ 0
784
+ 10
785
+ 20
786
+ 30
787
+ Time(s)40
788
+ Raw Data
789
+ 35
790
+ Filtered Data
791
+ 30
792
+ CSI Amplitude
793
+ 25
794
+ 20
795
+ 15
796
+ 10
797
+ 5
798
+ 0
799
+ 0
800
+ 10
801
+ 20
802
+ 30
803
+ Time(s)Figure 7: The average accuracy of the at-
804
+ tack in estimating the target person’s
805
+ breathing rate when he attacker and
806
+ target device are in the same building.
807
+ Figure 8: The CDF of the error in es-
808
+ timating the target person’s breathing
809
+ rate when he attacker and target de-
810
+ vice are in the same building (different
811
+ floor).
812
+ Figure 9: The CDF of the error in es-
813
+ timating the target person’s breathing
814
+ rate when he attacker and target device
815
+ are in different buildings (20m away)
816
+ the breathing rate in an indoor scenario where the target device
817
+ and attacker are in the same building. We evaluate the accuracy
818
+ when the target’s breathing rate is 12, 15, 20, and 30 breaths per
819
+ minute. Note, that the normal breathing rate for an adult is 12-20
820
+ breaths per minute while resting, and higher when exercising. In
821
+ this experiment, the user is watching a video. To make sure the
822
+ target person’s breathing rate is close to our desired numbers, we
823
+ place a timer in front of the person, where they can adjust their
824
+ breathing rate accordingly. This is just to better control the breath-
825
+ ing rate during the experiment and is not a requirement nor an
826
+ assumption in this attack. We run each experiment for two minutes.
827
+ During this time, we collect the estimated breathing rate from both
828
+ ground truth and the attack for different locations of the target
829
+ device. Figure 7 shows the average accuracy in estimating breath-
830
+ ing rate across all experiments. The accuracy is calculated as the
831
+ ratio of the estimated breathing rate reported by the attack over the
832
+ ground truth breathing rate. The figure shows that the accuracy of
833
+ estimating the breathing rate is over 99% in all scenarios. Finally,
834
+ Figure 8 plots the Cumulative Distribution Function (CDF) of the
835
+ error in detecting breathing rate for over 2400 measurements. The
836
+ figure shows that 78% of the estimated results have no error. The
837
+ figure also shows that 99% of measurements have less than one
838
+ breath per minute error which is negligible.
839
+ Different Building Scenario: So far, we have evaluated our at-
840
+ tack where the target and the attacker are in different rooms or
841
+ floors of the same building. Here we push this further and examine
842
+ whether our attack works if the attacker and the target person are
843
+ in a different building. We place the target device in a building on
844
+ a university campus on a weekday with people around. A person
845
+ is sitting around 0.5 m away from the device. We then place the
846
+ attacker in another building which is around 20 m away from the
847
+ target building. Similar to the previous experiment, we run the
848
+ attack and compare the estimated breathing rate with the ground
849
+ truth. Figure 9 shows the CDF of error for 180 measurements in
850
+ this experiment. Our results show that the attacker successfully
851
+ estimates the breathing rate. Note, that the reason that the attack
852
+ works even in such a challenging scenario with other people being
853
+ around is two-fold. First, using an FFT helps to filter out the effect
854
+ Figure 10: The efficacy of estimating the breathing rate when
855
+ there is no target near the WiFi device.
856
+ of most non-periodic movements and focuses on periodic move-
857
+ ments and patterns. Second, wireless channels are more sensitive
858
+ to changes as we get closer to the transmitter [11, 24], and since
859
+ in these scenarios, the target person is very close to the target de-
860
+ vice, their breathing motion has a higher impact on the CSI signal
861
+ compared to the other mobility in the environment.
862
+ 5.2.2
863
+ Human Presence Detection. We next evaluate the efficacy of
864
+ detecting whether there is a target person near the WiFi device or
865
+ not. In this experiment, the target phone is placed on a desk and the
866
+ person stays around the device for 30 seconds, then walks away
867
+ from the device, and then comes back near the device. Note, in our
868
+ algorithm, when there is no majority vote during the voting phase,
869
+ we return −1 to indicate no breathing detected. Figure 10 shows
870
+ the results of this experiment. As illustrated in the figure, we can
871
+ correctly detect the breathing rate when a person is near the device.
872
+ In other words, the algorithm can detect if there is no one near the
873
+ target device and refrain from reporting a random value.
874
+ 5.2.3
875
+ Effect of Distance and Orientation. Next, we evaluate the
876
+ effectiveness of the attack for different orientations of the device
877
+ with respect to the person. We also evaluate its performance for
878
+ different distances between the target device and the target person.
879
+ 7
880
+
881
+ 99.85%
882
+ 99.44%
883
+ 99.71%
884
+ 99.48%
885
+ 100
886
+ Accuracy (%)
887
+ 80
888
+ 60
889
+ 40
890
+ 20
891
+ 0
892
+ 12
893
+ 15
894
+ 20
895
+ 30
896
+ Orientation1.0
897
+ 0.8
898
+ CDF
899
+ 0.6
900
+ 0.4
901
+ 0.2
902
+ 0.8.0
903
+ 0.5
904
+ 1.0
905
+ 1.5
906
+ 2.0
907
+ 2.5
908
+ Error (RR/min)1.0
909
+ 0.8
910
+ DF
911
+ 0.6
912
+ 0.4
913
+ 0.2
914
+ 0.0
915
+ 0
916
+ 1
917
+ 2
918
+ 3
919
+ 4
920
+ 5
921
+ Error (RR/min)20
922
+ Respiration Rate (bpm)
923
+ 15
924
+ 10
925
+ Target person
926
+ Target person
927
+ 5
928
+ leaves
929
+ comes back
930
+ 0
931
+ -5
932
+ 0
933
+ 10
934
+ 20
935
+ 30
936
+ 40
937
+ 50
938
+ 60
939
+ 70
940
+ 80
941
+ Time (s)(a) various orientations
942
+ (b) different distances.
943
+ Figure 11: effectiveness of the attack for different orienta-
944
+ tion and distance of the targeted WiFi device respect to the
945
+ person.
946
+ Orientation: We evaluate the effect of orientation of the target
947
+ person with respect to the target device (laptop). We run the same
948
+ attack as before for different orientations (i.e. sitting in front, back,
949
+ left, and right side of a laptop). The user is 0.5m away from the target
950
+ device in all cases. Figure 11a shows the result of this experiment.
951
+ Each bar shows the average accuracy for 90 measurements. Our
952
+ result shows that regardless of the orientation of the person with
953
+ respect to the device, the attack is effective and detects the breathing
954
+ rate of the person accurately. In particular, even when the person
955
+ was behind the target device, the attack still detects the breathing
956
+ rate with 99% accuracy.
957
+ Distance: Here, we are interested to find out what the maximum
958
+ distance between the target device and the person can be while
959
+ the attacker still detects the person’s breathing rate. To do so, we
960
+ place the attacker device and the target device 5 meters apart in
961
+ two different rooms with a wall in between. We then run different
962
+ experiments in which the target person stays at different distances
963
+ from the target device. In each experiment, we measure the breath-
964
+ ing rate for two minutes and calculate the average breathing rate
965
+ over this time. Finally, we compare the estimated breathing rate to
966
+ the ground truth and calculate the accuracy as mentioned before.
967
+ Figure 11b shows the results of this experiment. The accuracy
968
+ is over 99% when the distance between the target device and the
969
+ target person is less than 60 cm. Note, in reality, people have their
970
+ laptops or cellphone very close to themselves most of the time, and
971
+ 60 cm is representative of these situations. The accuracy drops as
972
+ we increase the distance. However, even when the device is at 1.4 m
973
+ from the person’s body, the attack can still estimate the breathing
974
+ rate with 80% accuracy. Note, this is the accuracy in finding the
975
+ absolute breathing rate and the change in the breathing rate can be
976
+ detected with much higher accuracy. Finally, the figure shows that
977
+ the accuracy suddenly drops to zero for a distance beyond 1.4 m.
978
+ This is due to the fact that at that distance the power of the peak
979
+ at the output of the FFT goes below the noise floor, and hence, the
980
+ peak is not detectable.
981
+ 5.2.4
982
+ Effect of Multiple People. Last, we evaluate if the attack can
983
+ be used to detect the breathing rate of multiple people simultane-
984
+ ously. We test our attack in three different scenarios. In the first
985
+ scenario, two people are near the laptop while one is working on
986
+ the laptop and the other is just sitting next to him, as shown in
987
+ Figure 12a. The attacker targets the laptop and tries to estimate
988
+ their breathing rate. Note, that the attacker has no prior informa-
989
+ tion about how many people are next to the laptop. In the second
990
+ scenario, we repeat the same experiment as the first scenario except
991
+ that the second person is sitting behind the laptop, as shown in
992
+ Figure 12b. In the third scenario, there are two people in the same
993
+ space but each person is next to a different device. The attacker
994
+ targets the laptops and tries to estimate their breathing rates. In
995
+ these experiments, the target device is 0.5-0.7 m away from the
996
+ person.
997
+ Figure 12c shows the results for this evaluation. The blue bars
998
+ show the result for the first person who is working on the laptop,
999
+ and the red bars show the results for the second person. Our results
1000
+ show that the attack effectively detects the breathing rate of both
1001
+ people regardless of their orientation. However, the accuracy in
1002
+ detecting the breathing rate for the second person is a bit lower than
1003
+ the first person for the first and second scenarios. This is because
1004
+ the second person’s distance to the target device is slightly more
1005
+ and hence the accuracy has decreased.
1006
+ 6
1007
+ SECURITY IMPLICATION: BATTERY
1008
+ DRAIN ATTACK
1009
+ In this section, we show how an adversary can drain the battery
1010
+ of our WiFi devices by using the above loopholes and forcing our
1011
+ WiFi devices to stay awake and continuously transmit WiFi signals.
1012
+ 6.1
1013
+ Attack Design and Setup
1014
+ 6.1.1
1015
+ Attack Design. The attacker forces the target device to stay
1016
+ awake and continuously transmit WiFi packets by sending it back-
1017
+ to-back fake frames and some periodic fake beacons. However, to
1018
+ maximize the amount of time the target device spends transmitting,
1019
+ we study a few different types of fake query packets that the attacker
1020
+ can send. Note, that the power consumption of transmission is
1021
+ typically higher than that of reception.2 Hence, to maximize the
1022
+ battery drain, we want to send a short query packet and receive a
1023
+ long response.
1024
+ Table 3 lists some query packets and their corresponding re-
1025
+ sponses. The best choice for a query packet is Block ACK requests
1026
+ since the target will respond with a Block ACK that is larger than
1027
+ other query responses. Another important factor to consider for
1028
+ maximizing the battery drain is the bitrate. When the bitrate of the
1029
+ query packet increases, the bitrate of the response will also increase
1030
+ as specified in the IEEE 802.11 standard. Hence, at first glance, it
1031
+ 2For example, ESP8266 [26] and ESP32 [25] WiFi modules draw 50 and 100 mA when
1032
+ receiving while they draw 170 and 240 mA when transmitting. These low-power WiFi
1033
+ modules are very popular for IoT devices [10].
1034
+ 8
1035
+
1036
+ 99.91%
1037
+ 99.73%
1038
+ 99.72%
1039
+ 99.91%
1040
+ 100
1041
+ Accuracy (%)
1042
+ 80
1043
+ 60
1044
+ 40
1045
+ 20
1046
+ 0
1047
+ Front
1048
+ Back
1049
+ Left
1050
+ Right
1051
+ Orientation100
1052
+ Accuracy (%)
1053
+ 80
1054
+ 60
1055
+ 40
1056
+ 20
1057
+ 0
1058
+ 0
1059
+ 20
1060
+ 40
1061
+ 60
1062
+ 80
1063
+ 100
1064
+ 120
1065
+ 140
1066
+ 160
1067
+ Distance (cm)(a) Scenario 1
1068
+ (b) Scenario 2
1069
+ (c) Breathing Rate Estimation of two persons
1070
+ Figure 12: Accuracy under three different scenarios: Scenario 1: two people sit side-by-side in front of the target device; Scenario
1071
+ 2: one person sits in front of the target device, the other one sits behind the target device; Scenario 3: two people sit in front
1072
+ of two target devices, respectively. Attacker attacks one by one.
1073
+ Query
1074
+ Query size
1075
+ Response
1076
+ Response size
1077
+ Null
1078
+ 28 bytes
1079
+ ACK
1080
+ 14 bytes
1081
+ RTS
1082
+ 20 bytes
1083
+ CTS
1084
+ 14 bytes
1085
+ BAR
1086
+ 24 bytes
1087
+ BA
1088
+ 32 bytes
1089
+ Table 3: Different types of fake queries and their responses.
1090
+ Note, Null is a data packet without any payload. BAR and BA
1091
+ stand for Block ACK Request, and Block ACK, respectivly.
1092
+ may seem that to maximize the battery drain, the attacker must
1093
+ use the fastest bitrate possible to transmit query packets, forcing
1094
+ the target device to transmit as many responses as possible. How-
1095
+ ever, it turns out that this is not the case. The power consumption
1096
+ depends mostly on the amount of time the target device spends
1097
+ transmitting packets. Hence, when a higher rate is used for the
1098
+ query and response packets, the total time the target spends on
1099
+ transmission does not increase. In fact, the total time spent trans-
1100
+ mitting decreases mainly due to overheads such as channel sensing
1101
+ and backoffs. For example, if we increase the bitrate by 6 times (i.e.,
1102
+ from 1 Mbps to 6 Mbps), the number of packets will increase by
1103
+ only 3.3 times. As a result, to maximize the transmission time of the
1104
+ target device, the attacker should use the lowest rate (i.e., 1 Mbps)
1105
+ for the query packet.
1106
+ 6.1.2
1107
+ Attack Setup.
1108
+ Attacking device: Any WiFi card capable of packet injection can
1109
+ be used as the attacking device. We use a USB WiFi card connected
1110
+ to a laptop running Ubuntu 20.04. The WiFi card has an RTL8812AU
1111
+ chipset [5] that supports IEEE 802.11 a/b/g/n/ac standards. We have
1112
+ installed the aircrack-ng/rtl8812au driver [1] for this card which
1113
+ enables robust packet injection. We utilize the Scapy [37] library to
1114
+ inject fake WiFi packets to the target device. Scapy allows defin-
1115
+ ing customized packets and multiple options for packet injection.
1116
+ Since we need to inject many packets in this attack, we use the
1117
+ sendpfast function to inject packets at high rates. sendpfast relies
1118
+ on tcpreplay [6] for high performance packet injection.
1119
+ Target device: Any WiFi-based IoT device can be used as a target.
1120
+ We choose Amazon Ring Spotlight Cam Battery HD Security Cam-
1121
+ era [2] for our battery drain experiments. The camera is powered
1122
+ by a custom 6040 mAh lithium-ion battery. The battery life of this
1123
+ camera is estimated to be between 6 and 12 months under normal
1124
+ usage [3, 4]. We leave the camera settings to their defaults which
1125
+ means most power-consuming options are turned off. This assures
1126
+ that our measurements will be an upper bound on the battery life
1127
+ and hence the attack might drain the battery much faster in the real
1128
+ world. Authors in [41] pointed out the possibility of a battery drain-
1129
+ ing attack by forging beacon frames. However, they did not provide
1130
+ any evaluations to test this idea. Moreover, we show how sending
1131
+ fake packets in addition to fake beacon frames can significantly
1132
+ increase the power consumption on the victim device.
1133
+ 6.2
1134
+ Results
1135
+ We evaluate the effectiveness of the battery drain attack in terms
1136
+ of range and using different payload configuration.
1137
+ 6.2.1
1138
+ Finding the optimal configuration: As discussed in 6.1.1, send-
1139
+ ing block ACK requests at the lowest bitrate (i.e., 1 Mbps) should
1140
+ maximize the power consumption of the target device. To verify
1141
+ this, we have conducted a series of experiments with different types
1142
+ of query packets and transmission bitrates. In each experiment, we
1143
+ continuously transmit query packets to the Ring security camera.
1144
+ In all experiments, we start with a fully charged battery and the
1145
+ attacker injects query packets as fast as possible.
1146
+ Figure 13 (a) shows the maximum number of packets the attacker
1147
+ could transmit to the target device, and the number of responses
1148
+ it receives per second. Figure 13 (b) shows the amount of energy
1149
+ drawn from the battery during one hour of the attack. As expected,
1150
+ sending Block ACK Requests (BAR) drains more energy from the
1151
+ battery since the target device spends more time on transmission
1152
+ than receiving. Moreover, the results verify that although increas-
1153
+ ing the data rate from 1Mbps to 6Mbps (BAR/1 versus BAR/6)
1154
+ increases the number of responses, it decreases the energy drained.
1155
+ As mentioned before, this is because the total time spent transmit-
1156
+ ting decreases mainly due to overheads such as channel sensing
1157
+ 9
1158
+
1159
+ D100%
1160
+ 99.48%
1161
+ 100% 99.07%
1162
+ 100
1163
+ 86.67%
1164
+ 82.05%
1165
+ Accuracy (%)
1166
+ 80
1167
+ 60
1168
+ 40
1169
+ 20
1170
+ 0
1171
+ Scenario 1
1172
+ Scenario 2
1173
+ Scenario 3Battery Type
1174
+ Voltage (V)
1175
+ Full Capacity (Wh)
1176
+ 100% Drain (min)
1177
+ 25% Drain (min)
1178
+ CR2032 coin
1179
+ 3.0
1180
+ 0.68
1181
+ 14
1182
+ 3.5
1183
+ AAA
1184
+ 1.5
1185
+ 1.87
1186
+ 39
1187
+ 10
1188
+ AA
1189
+ 1.5
1190
+ 4.20
1191
+ 90
1192
+ 22
1193
+ Table 4: The time it takes for the attack to drain different types of batteries
1194
+ 0
1195
+ 500
1196
+ 1000
1197
+ 1500
1198
+ 2000
1199
+ 2500
1200
+ 3000
1201
+ 3500
1202
+ Null/1
1203
+ Data/1
1204
+ BAR/1
1205
+ BAR/6
1206
+ Number of Packets
1207
+ Configurations
1208
+ Attacker's packets
1209
+ Target's responses
1210
+ (a)
1211
+ 0
1212
+ 0.5
1213
+ 1
1214
+ 1.5
1215
+ 2
1216
+ 2.5
1217
+ 3
1218
+ Null/1
1219
+ Data/1
1220
+ BAR/1
1221
+ BAR/6
1222
+ Watt Hour
1223
+ Configurations
1224
+ (b)
1225
+ Figure 13: The figure shows (a) Average number of packets
1226
+ sent to and received from the target device. (b) Energy con-
1227
+ sumption in Watt Hour measured under different configu-
1228
+ rations (i.e. packet type / bitrate (Mbps)
1229
+ and backoffs. This result confirms that sending block ACK requests
1230
+ (BAR) with the lowest datarate is the best option to drain the battery
1231
+ of the target device.
1232
+ 6.2.2
1233
+ Battery drain with optimal configurations. We use the best
1234
+ setting which is a block ACK request (BAR) query transmitted at
1235
+ 1 Mbps to fully drain the battery of the Ring security camera. We
1236
+ are able to drain a fully charged battery in 36 hours. Considering
1237
+ the fact that the typical battery life of this camera is 6 to 12 months,
1238
+ our attack reduces the battery life by 120 to 240 times! It is worth
1239
+ mentioning that since a typical user charges the battery every 6-12
1240
+ months, on average the batteries are at 40-60%, and therefore it
1241
+ would take much less for our attack to kill the battery. Moreover, the
1242
+ RING security camera is using a very large battery, most security
1243
+ sensors are using smaller batteries. Table 4 shows the amount of
1244
+ time it takes to drain different batteries. For example, it takes less
1245
+ than 40 mins to kill a fully charged AAA battery which is a common
1246
+ battery in many sensors.
1247
+ 6.2.3
1248
+ Range of WiFi battery draining attack. A key factor in the
1249
+ effectiveness of the battery draining attack is how far the attacker
1250
+ can be from the victim’s device and still be able to carry on the
1251
+ attack. If the attack can be done from far away, it becomes more
1252
+ threatening. To evaluate the range of this attack, we design an
1253
+ experiment in which the attacker transmits packets to the target
1254
+ from different distances and we measure what percentage of the
1255
+ attacker’s packets are responded to by the target device. We use
1256
+ a realistic testbed. The Ring security camera is installed in front
1257
+ of a house, and the attacker is placed in a car, parked at different
1258
+ locations on the street. We test the attack at 10 different locations
1259
+ up to 150 meters away from the target device. Figure 14 shows
1260
+ these locations and our setup. Each yellow circle represents each
1261
+ of the locations tested at. The numbers inside the circles show the
1262
+ percentage of the attacker’s packets responded to by the camera.
1263
+ Each number is an average of over 60 one-second measurements.
1264
+ The closest distance is about 5 meters when we park the car in front
1265
+ of the target house. In this location 97% of the attacker’s packets are
1266
+ responded to. We conducted other experiments within 10 meters
1267
+ of the target (not shown here) and we obtained similar results. Our
1268
+ results show that even within a distance of 100 meters, almost all
1269
+ attacker’s packets are responded to by the victim’s device. In some
1270
+ locations such as the rightmost circle (at 150 meters away), we
1271
+ could still achieve a reply rate as high as 73%, confirming our attack
1272
+ works even at that distance. The reason for achieving such a long
1273
+ range is that the attacker transmits at a 1 Mbps bitrate which uses
1274
+ extremely robust modulation and coding rate (i.e. BPSK modulation
1275
+ and a 1/11 coding rate).
1276
+ 7
1277
+ ETHICAL CONSIDERATIONS
1278
+ We discussed our project and experiments with our institutions’
1279
+ IRB office and they determined that no IRB review nor IRB approval
1280
+ is required. Moreover, the house and WiFi devices used in most
1281
+ experiments are owned and controlled by the authors. Finally, in
1282
+ order to expedite mitigating the attacks presented in this paper,
1283
+ we have started engagements with WiFi access point and chipset
1284
+ manufacturers.
1285
+ 8
1286
+ CONCLUSION
1287
+ In this work, we identify two loopholes in the WiFi protocol and
1288
+ demonstrate their possible privacy and security threats. In partic-
1289
+ ular, we reveal that today’s WiFi radio responds to packets from
1290
+ unauthorized devices outside of the network and it can be easily
1291
+ manipulated to keep awake. These loopholes can be exploited by
1292
+ malicious attackers to jeopardize our daily use of WiFi devices. As
1293
+ examples, we demonstrate how an attacker can take advantage of
1294
+ these loopholes to extract private information such as breathing
1295
+ rate and quickly exhaust the battery of a typical IoT device, leaving
1296
+ the victim’s device in a disabled state.
1297
+ REFERENCES
1298
+ [1] [n. d.]. aircrack-ng/rtl8812au. https://github.com/aircrack-ng/rtl8812au.
1299
+ 10
1300
+
1301
+ 0 m
1302
+ 50 m
1303
+ 100 m
1304
+ 150 m
1305
+ 50 m
1306
+ 100 m
1307
+ 150 m
1308
+ 73
1309
+ 97
1310
+ 90
1311
+ 54
1312
+ 14
1313
+ 84
1314
+ 83
1315
+ 90
1316
+ 70
1317
+ 64
1318
+ Target
1319
+ Attacker
1320
+ Figure 14: Percentage of attacker’s query packets responded by the target device for different attacker’s locations.
1321
+ [2] [n. d.]. Ring Spotlight Cam Battery. https://ring.com/products/spotlight-cam-
1322
+ battery".
1323
+ [3] [n. d.]. Ring Spotlight Cam Battery Review. https://www.security.org/security-
1324
+ cameras/ring/review/spotlight-cam-battery/.
1325
+ [4] [n. d.]. Ring Spotlight Cam Battery Review. https://www.pcmag.com/reviews/
1326
+ ring-spotlight-cam-battery.
1327
+ [5] [n. d.]. RTL8812AU. https://www.realtek.com/en/products/communications-
1328
+ network-ics/item/rtl8812au.
1329
+ [6] [n. d.]. Tcpreplay - Pcap editing and replaying utilities. https://tcpreplay.appneta.
1330
+ com/.
1331
+ [7] 2009. IEEE Standard for Information technology - Telecommunications and
1332
+ information exchange between systems - Local and metropolitan area networks
1333
+ - Specific requirements. Part 11: Wireless LAN Medium Access Control (MAC)
1334
+ and Physical Layer (PHY) Specifications Amendment 4: Protected Management
1335
+ Frames. IEEE Std 802.11w-2009 (Amendment to IEEE Std 802.11-2007 as amended
1336
+ by IEEE Std 802.11k-2008, IEEE Std 802.11r-2008, and IEEE Std 802.11y-2008) (2009),
1337
+ 1–111.
1338
+ [8] Heba Abdelnasser, Khaled Harras, and Moustafa Youssef. 2019. A Ubiquitous WiFi-
1339
+ Based Fine-Grained Gesture Recognition System. IEEE Transactions on Mobile
1340
+ Computing 18, 11 (2019), 2474–2487. https://doi.org/10.1109/TMC.2018.2879075
1341
+ [9] Ali Abedi and Omid Abari. 2020. WiFi Says "Hi!" Back to Strangers!. In Proceedings
1342
+ of the 19th ACM Workshop on Hot Topics in Networks (HotNets. 132–138.
1343
+ [10] Ali Abedi, Omid Abari, and Tim Brecht. 2019. Wi-le: Can wifi replace bluetooth?.
1344
+ In Proceedings of the 18th ACM Workshop on Hot Topics in Networks. 117–124.
1345
+ [11] Ali Abedi, Farzan Dehbashi, Mohammad Hossein Mazaheri, Omid Abari, and Tim
1346
+ Brecht. 2020. Witag: Seamless wifi backscatter communication. In Proceedings
1347
+ of the Annual conference of the ACM Special Interest Group on Data Communi-
1348
+ cation on the applications, technologies, architectures, and protocols for computer
1349
+ communication (SIGCOMM). 240–252.
1350
+ [12] Ali Abedi and Deepak Vasisht. 2022. Non-Cooperative Wi-Fi Localization and Its
1351
+ Privacy Implications. In Proceedings of the 28th Annual International Conference
1352
+ on Mobile Computing And Networking (MobiCom. 570–582.
1353
+ [13] Fadel Adib, Hongzi Mao, Zachary Kabelac, Dina Katabi, and Robert C Miller. 2015.
1354
+ Smart homes that monitor breathing and heart rate. In Proceedings of the 33rd
1355
+ annual ACM conference on human factors in computing systems. 837–846.
1356
+ [14] Mayank Agarwal, Dileep Pasumarthi, Santosh Biswas, and Sukumar Nandi. 2016.
1357
+ Machine learning approach for detection of flooding DoS attacks in 802.11 net-
1358
+ works and attacker localization. International Journal of Machine Learning and
1359
+ Cybernetics 7 (2016), 1035–1051.
1360
+ [15] Bandar Alotaibi and Khaled Elleithy. 2016. Rogue Access Point Detection: Tax-
1361
+ onomy, Challenges, and Future Directions. Wireless Personal Communications 90
1362
+ (10 2016), 5021– 5028. https://doi.org/10.1007/s11277-016-3390-x
1363
+ [16] Sheheryar Arshad, Chunhai Feng, Yonghe Liu, Yupeng Hu, Ruiyun Yu, Siwang
1364
+ Zhou, and Heng Li. 2017. Wi-chase: A WiFi based human activity recognition
1365
+ system for sensorless environments. In 2017 IEEE 18th International Symposium
1366
+ on A World of Wireless, Mobile and Multimedia Networks (WoWMoM). 1–6. https:
1367
+ //doi.org/10.1109/WoWMoM.2017.7974315
1368
+ [17] John Bellardo and Stefan Savage. 2003. 802.11 Denial-of-Service Attacks: Real Vul-
1369
+ nerabilities and Practical Solutions. Proceedings of 12 USENIX Security Symposium
1370
+ (2003).
1371
+ [18] John Bellardo and Stefan Savage. 2003. 802.11 Denial-of-Service Attacks: Real
1372
+ Vulnerabilities and Practical Solutions.. In USENIX security symposium, Vol. 12.
1373
+ Washington DC, 2–2.
1374
+ [19] Timothy K Buennemeyer, Theresa M Nelson, Lee M Clagett, John P Dunning,
1375
+ Randy C Marchany, and Joseph G Tront. 2008. Mobile device profiling and
1376
+ intrusion detection using smart batteries. In Proceedings of the 41st Annual Hawaii
1377
+ International Conference on System Sciences (HICSS 2008). IEEE, 296–296.
1378
+ [20] Luca Caviglione and Alessio Merlo. 2012. The energy impact of security mecha-
1379
+ nisms in modern mobile devices. Network Security 2012, 2 (2012), 11–14.
1380
+ [21] R. Chandra, J. Padhye, L. Ravindranath, and A. Wolman. 2007. Beacon-Stuffing:
1381
+ Wi-Fi without Associations. In Eighth IEEE Workshop on Mobile Computing Sys-
1382
+ tems and Applications. 53–57.
1383
+ [22] Gerald Combs. 2020. Wireshark. https://www.wireshark.org/.
1384
+ [23] Eliran Dafna, Ariel Tarasiuk, and Yaniv Zigel. 2015. Sleep-wake evaluation from
1385
+ whole-night non-contact audio recordings of breathing sounds. PloS one 10, 2
1386
+ (2015), e0117382.
1387
+ [24] Farzan Dehbashi, Ali Abedi, Tim Brecht, and Omid Abari. 2021. Verification: can
1388
+ wifi backscatter replace RFID?. In Proceedings of the 27th Annual International
1389
+ Conference on Mobile Computing and Networking. 97–107.
1390
+ [25] Espressif Systems 2019. ESP32 datasheet. Espressif Systems.
1391
+ https://www.
1392
+ espressif.com/sites/default/files/documentation/\esp32_datasheet_en.pdf.
1393
+ [26] Espressif
1394
+ Systems
1395
+ 2020.
1396
+ ESP8266
1397
+ datasheet.
1398
+ Espressif
1399
+ Sys-
1400
+ tems.
1401
+ https://www.espressif.com/sites/default/files/documentation/0a-
1402
+ esp8266ex_datasheet_en.pdf.
1403
+ [27] Ugo Fiore, Aniello Castiglione, Alfredo De Santis, and Francesco Palmieri. 2017.
1404
+ Exploiting battery-drain vulnerabilities in mobile smart devices. IEEE Transactions
1405
+ on Sustainable Computing 2, 2 (2017), 90–99.
1406
+ [28] Ugo Fiore, Francesco Palmieri, Aniello Castiglione, Vincenzo Loia, and Alfredo
1407
+ De Santis. 2014. Multimedia-based battery drain attacks for android devices. In
1408
+ 2014 IEEE 11th Consumer Communications and Networking Conference (CCNC).
1409
+ IEEE, 145–150.
1410
+ [29] Yu Gu, Jinhai Zhan, Yusheng Ji, Jie Li, Fuji Ren, and Shangbing Gao. 2017.
1411
+ MoSense: An RF-Based Motion Detection System via Off-the-Shelf WiFi De-
1412
+ vices. IEEE Internet of Things Journal 4, 6 (2017), 2326–2341. https://doi.org/10.
1413
+ 1109/JIOT.2017.2754578
1414
+ [30] Abhilash Jindal, Abhinav Pathak, Y Charlie Hu, and Samuel Midkiff. 2013. Hypnos:
1415
+ understanding and treating sleep conflicts in smartphones. In Proceedings of the
1416
+ 8th ACM European Conference on Computer Systems. 253–266.
1417
+ [31] S. S. Kolahi and A. A. Almatrook. 2017. Impact of security on bandwidth and
1418
+ latency in IEEE 802.11ac client-to-server WLAN. In 2017 Ninth International
1419
+ Conference on Ubiquitous and Future Networks (ICUFN). 893–897.
1420
+ [32] Guohao Lan, Mohammadreza F. Imani, Zida Liu, José Manjarrés, Wenjun Hu,
1421
+ Andrew S. Lan, David R. Smith, and Maria Gorlatova. 2021. MetaSense: Boosting
1422
+ RF Sensing Accuracy Using Dynamic Metasurface Antenna. IEEE Internet of
1423
+ Things Journal 8 (2021).
1424
+ [33] P. Li, S. S. Kolahi, M. Safdari, and M. Argawe. 2011. Effect of WPA2 Security
1425
+ on IEEE 802.11n Bandwidth and Round Trip Time in Peer-Peer Wireless Local
1426
+ Area Networks. In 2011 IEEE Workshops of International Conference on Advanced
1427
+ Information Networking and Applications. 777–782.
1428
+ [34] Tianxiang Li, Haofan Lu, Reza Rezvani, Ali Abedi, and Omid Abari. 2022. Bringing
1429
+ Wifi Localization to Any Wifi Devices (HotNets).
1430
+ [35] Jian Liu, Yan Wang, Yingying Chen, Jie Yang, Xu Chen, and Jerry Cheng. 2015.
1431
+ Tracking Vital Signs During Sleep Leveraging Off-the-Shelf WiFi (MobiHoc).
1432
+ [36] Yongsen Ma, Gang Zhou, and Shuangquan Wang. 2019. WiFi Sensing with
1433
+ Channel State Information: A Survey. Comput. Surveys 52, 3 (2019).
1434
+ [37] Philippe Biondi. 2020. Scapy. https://scapy.net/.
1435
+ [38] Qifan Pu, Sidhant Gupta, Shyamnath Gollakota, and Shwetak Patel. 2013. Whole-
1436
+ home gesture recognition using wireless signals. In Proceedings of the 19th annual
1437
+ international conference on Mobile computing & networking. 27–38.
1438
+ [39] Mohammad Saleh, Jaafar Gaber, and Maxim Wack. 2017. Sensor Networks
1439
+ Applications Performance Measures for IEEE802.11n WiFi Security Protocols. In
1440
+ Proceedings of the International Conference on Future Networks and Distributed
1441
+ Systems (ICFNDS ’17). https://doi.org/10.1145/3102304.3102335
1442
+ [40] Frank Stajano and Ross Anderson. 1999. The resurrecting duckling: Security
1443
+ issues for ad-hoc wireless networks. In International workshop on security protocols.
1444
+ Springer, 172–182.
1445
+ [41] Mathy Vanhoef, Prasant Adhikari, and Christina Pöpper. 2020. Protecting wi-fi
1446
+ beacons from outsider forgeries. In Proceedings of the 13th ACM Conference on
1447
+ 11
1448
+
1449
+ Security and Privacy in Wireless and Mobile Networks. 155–160.
1450
+ [42] Mathy Vanhoef, Prasant Adhikari, and Christina Pöpper. 2020. Protecting Wi-Fi
1451
+ Beacons from Outsider Forgeries (WiSec).
1452
+ [43] Raghav H. Venkatnarayan, Griffin Page, and Muhammad Shahzad. 2018. Multi-
1453
+ User Gesture Recognition Using WiFi (MobiSys).
1454
+ [44] Aditya Virmani and Muhammad Shahzad. 2017. Position and Orientation Agnos-
1455
+ tic Gesture Recognition Using WiFi (MobiSys).
1456
+ [45] Hao Wang, Daqing Zhang, Junyi Ma, Yasha Wang, Yuxiang Wang, Dan Wu, Tao
1457
+ Gu, and Bing Xie. 2016. Human Respiration Detection with Commodity Wifi
1458
+ Devices: Do User Location and Body Orientation Matter? (UbiComp). 25–36.
1459
+ [46] S. Zehl, N. Karowski, A. Zubow, and A. Wolisz. 2016. LoWS: A complete Open
1460
+ Source solution for Wi-Fi beacon stuffing based Location-based Services. In 2016
1461
+ 9th IFIP Wireless and Mobile Networking Conference (WMNC). 25–32.
1462
+ [47] Lide Zhang, Birjodh Tiwana, Zhiyun Qian, Zhaoguang Wang, Robert P Dick,
1463
+ Zhuoqing Morley Mao, and Lei Yang. 2010. Accurate online power estimation
1464
+ and automatic battery behavior based power model generation for smartphones.
1465
+ In Proceedings of the eighth IEEE/ACM/IFIP international conference on Hardware/-
1466
+ software codesign and system synthesis. 105–114.
1467
+ [48] Yue Zheng, Yi Zhang, Kun Qian, Guidong Zhang, Yunhao Liu, Chenshu Wu, and
1468
+ Zheng Yang. 2019. Zero-Effort Cross-Domain Gesture Recognition with Wi-Fi
1469
+ (MobiSys). 313–325.
1470
+ [49] Yanzi Zhu, Zhujun Xiao, Yuxin Chen, Zhijing Li, Max Liu, Ben Y Zhao, and Haitao
1471
+ Zheng. 2020. Et Tu Alexa? When Commodity WiFi Devices Turn into Adversarial
1472
+ Motion Sensors. In Network and Distributed Systems Security (NDSS) Symposium
1473
+ 2020.
1474
+ 12
1475
+
19AyT4oBgHgl3EQfbvd_/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
19FAT4oBgHgl3EQfDBzX/content/2301.08414v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fa02405eabf2c0e80855f98f51faaafc3dbcb7250126557055554db9de9772d8
3
+ size 6562901
2NE1T4oBgHgl3EQf5QXz/content/tmp_files/2301.03511v1.pdf.txt ADDED
@@ -0,0 +1,1151 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ The Value of Internal Memory for Population Growth in Varying
2
+ Environments
3
+ Leo Law, BingKan Xue
4
+ Department of Physics, University of Florida, Gainesville, FL 32611, USA
5
+ Abstract
6
+ In varying environments it is beneficial for organisms to utilize available cues to infer the
7
+ conditions they may encounter and express potentially favorable traits. However, external cues
8
+ can be unreliable or too costly to use. We consider an alternative strategy where organisms
9
+ exploit internal sources of information. Even without sensing environmental cues, their internal
10
+ states may become correlated with the environment as a result of selection, which then form a
11
+ memory that helps predict future conditions. To demonstrate the adaptive value of such internal
12
+ memory in varying environments, we revisit the classic example of seed dormancy in annual
13
+ plants. Previous studies have considered the germination fraction of seeds and its dependence
14
+ on environmental cues. In contrast, we consider a model of germination fraction that depends
15
+ on the seed age, which is an internal state that can serve as a memory. We show that, if the
16
+ environmental variation has temporal structure, then age-dependent germination fractions will
17
+ allow the population to have an increased long-term growth rate.
18
+ The more organisms can
19
+ remember through their internal states, the higher growth rate a population can potentially
20
+ achieve. Our results suggest experimental ways to infer internal memory and its benefit for
21
+ adaptation in varying environments.
22
+ 1
23
+ arXiv:2301.03511v1 [q-bio.PE] 9 Jan 2023
24
+
25
+ 1
26
+ Introduction
27
+ Organisms can adapt to a varying environment by diversifying their traits among individuals of
28
+ the same population. A common form of such diversity is dormancy, where some individuals enter
29
+ a dormant state while others remain active [1, 2, 3]. Those that are active will contribute to the
30
+ growth of the population under good environmental conditions, but will be vulnerable to periods of
31
+ harsh conditions. On the other hand, the dormant individuals are often tolerant to environmental
32
+ stress and thus help preserve the population during harsh periods. For example, in a bacterial
33
+ population, while most cells grow and divide normally, some cells randomly switch to a reversible
34
+ dormant state called persister cells, which makes them tolerant to antibiotics when normal cells
35
+ would perish [4, 5, 6]. Other examples include seed dormancy in plants, dauer larva in nematodes,
36
+ diapause in insects, etc.
37
+ [1, 3, 7, 8].
38
+ These are thought to be a strategy known as diversified
39
+ bet-hedging [9, 10], in which organisms express different traits with some probability to create
40
+ diversity in the population, so as to increase the long-term growth rate of the population under
41
+ environmental variations [11, 12, 13].
42
+ In the simplest form, bet-hedging organisms have fixed probabilities of expressing different traits
43
+ [11]. But more generally, organisms can sense cues from the environment that will influence these
44
+ probabilities [14, 15]. Such cues may be indicative of future environmental conditions, so that the
45
+ organisms may bias the probabilities towards traits that are favorable in the likely environment. It
46
+ has been shown that the information contained in the cue about the environment will contribute
47
+ to an increase in the population growth rate [14, 16, 17]. However, sensing and responding to
48
+ environmental cues may come at a cost, as it requires the expression of specific sensors and signaling
49
+ mechanisms [18]. Besides, there may not be enough time for the organisms to respond to the cues
50
+ through phenotypic plasticity, as the environment may have changed by the time the trait is
51
+ developed [19, 20]. Therefore, it is not always beneficial to rely on environmental cues.
52
+ Besides external signals, the behavior of organisms can be influenced by their internal states, such
53
+ as physiological or metabolic states [21]. One example is the reserve level – a starved animal may
54
+ choose to forage more aggressively despite higher predation risk [22, 21]. Another example is the
55
+ age of the organism – it is known that the age of seeds can affect germination in annual plants [23].
56
+ These internal states are not sensors that directly measure the external environment. However,
57
+ they may become correlated with the environment as a result of selection, because certain states
58
+ are associated with higher fitness in past environmental conditions and thus become more common
59
+ in the population. Therefore, the distribution of such internal states among the population can
60
+ potentially provide information about the environment, which may be utilized by the organisms.
61
+ We will study an example of this situation and show that internal states of the organisms can
62
+ indeed serve as internal cues to help them adapt to varying environmental conditions. Such internal
63
+ states effectively provide a memory about the past outcomes of selection, which helps predict the
64
+ future environment. Moreover, we show that a larger memory capacity enables higher gains in the
65
+ 2
66
+
67
+ population growth rate. Our results suggest that internal states that were not developed for sensing
68
+ the environment could nevertheless be co-opted as internal cues for adaptation, which would save
69
+ the cost of sensors and may thus be a more efficient strategy.
70
+ To study adaptation in varying environments, we will use seed dormancy as our main example.
71
+ Seeds of annual plants will either germinate or stay dormant in a given year. While dormancy
72
+ sacrifices the short-term fitness of the seeds, it preserves the population from a catastrophically
73
+ bad year with very low yield, and thus results in higher long-term benefit. This has been studied
74
+ as a classic model of bet-hedging [11, 14], supported by the fact that dormant seeds eventually ger-
75
+ minate under similar environmental conditions [23], and that the germination fraction is negatively
76
+ correlated with local environmental variability [24]. It is known that germination is influenced by
77
+ environmental cues, such as temperature, humidity, and the number density of surrounding seeds
78
+ [15, 25]. Moreover, there is evidence that the probability a seed will germinate also changes with the
79
+ age [26, 27, 23]. However, the adaptive value of such age dependence in germination has not been
80
+ fully studied [28, 3]. It was shown in [28] that the evolutionarily stable probability of germination
81
+ does not depend on seed age if there is no density dependence. Yet, their model did not include
82
+ temporal correlation in the environmental variation, which is crucial for memory to be useful in
83
+ predicting future environments [29, 30, 31]. We will show that, when there is temporal structure
84
+ in the environmental variation, age-dependent germination probabilities can increase the long-term
85
+ growth rate of the seed population.
86
+ 2
87
+ Background
88
+ 2.1
89
+ Cohen’s model of seed dormancy
90
+ Let us first briefly review the idea of bet-hedging and how information emerges as a central quantity
91
+ in determining the long-term growth rate of the population. We will follow the classic model of
92
+ seed dormancy in annual plants by Cohen [11, 14], as illustrated in Fig. 1A. Each year can be
93
+ “good” (denoted as environment ε = 1) or “bad” (ε = 0) for the plant. Seeds that germinate
94
+ (“phenotype” φ = 1) in a good year will be able to grow and produce a large number (Y1) of new
95
+ seeds. However, in a bad year, germinated plants will have a low yield (Y0). We will set Y0 = 0 and
96
+ denote Y1 = Y for simplicity, meaning that germinating in a bad year will result in no offspring. All
97
+ germinated plants perish at the end of the year, regardless of their yield. Seeds that stay dormant
98
+ (φ = 0) will remain viable the next year with probability V . Thus, the fitness of a seed in a given
99
+ environment can be summarized by the matrix fεφ =
100
+ � V 0
101
+ V Y
102
+
103
+ . In addition, we assume that the
104
+ number of consecutive good years follows a geometric distribution, whereas that of bad years has a
105
+ narrow distribution (see Fig. 1B and Appendix A.2). This is meant to describe the scenario where
106
+ good growth conditions are disrupted by random occurrence of disasters that affect growth for a
107
+ characteristic number of years.
108
+ 3
109
+
110
+
111
+
112
+ dormant
113
+ A
114
+ germinate
115
+ year 2
116
+ viability
117
+ dormant
118
+ germinate
119
+ year 1
120
+ low yield
121
+ viability
122
+ bad
123
+ ...
124
+ ...
125
+ C
126
+ B
127
+ seeds
128
+
129
+ s1
130
+ s2
131
+ s0
132
+ good
133
+ high yield
134
+ seeds
135
+ Figure 1: (A) Schematic illustration of Cohen’s model of seed dormancy in annual plants. Each year may
136
+ be good or bad for plant growth. A seed can either germinate to produce a yield Yε that depends on the
137
+ environmental condition ε, or stay dormant with a probability V of still being viable next year. The number
138
+ of seeds at the end of year t is Nt. The parameter values used in our calculations are Y0 = 0, Y1 = 4,
139
+ V = 0.9. (B) The distribution of duration of consecutive good years and bad years. We choose the duration
140
+ of good years to follow a geometric distribution with a mean of 5, and the duration of bad years to have a
141
+ Gaussian distribution with a mean and standard deviation of 5 ± 2 cut off at 0 and 10. (C) A state diagram
142
+ that represents the seed age. Each state sα represents a seed of age α. Blue arrows represent dormancy that
143
+ increases the age by 1; orange arrows represent germination that may produce new seeds of age 0. Weights
144
+ on the arrows represent the probability of germination or dormancy.
145
+ In the simplest case where seeds receive no environmental cues, the fraction of seeds that germinate
146
+ each year is assumed to be a constant, denoted by q. In a good year, the total number of seeds will
147
+ grow by a factor (1 − q)V + qY , whereas in a bad year, the number of seeds will reduce to only a
148
+ fraction (1 − q)V of the previous year. The long-term growth rate of the population will be given
149
+ by (see derivation in Appendix A.1)
150
+ Λ = p log
151
+
152
+ (1 − q)V + qY
153
+
154
+ + (1 − p) log
155
+
156
+ (1 − q)V
157
+
158
+ ,
159
+ (1)
160
+ where p is the frequency of good years and (1 − p) is that for bad years. The germination fraction
161
+ that maximizes the long-term growth rate is
162
+ q∗ = p Y − V
163
+ Y − V
164
+ (2)
165
+ for p > V/Y and 0 otherwise. In the limit of high yield (Y ≫ V ), this leads to the classic result
166
+ q∗ ≈ p, which means the optimal germination fraction should match the frequency of good years
167
+ [11]. The model can be extended to seeds that receive some external cue (ξ) about the environment
168
+ [14].
169
+ In this case, the optimal germination fraction will depend on the cue.
170
+ As a result, the
171
+ population can grow faster than without the cue (see Appendix A.1).
172
+ 4
173
+
174
+ 0.25
175
+ Good years
176
+ Bad years
177
+ 0.20
178
+ distribution
179
+ 0.15
180
+ 0.10 :
181
+ 0.05
182
+ 0.00
183
+ 12345678910
184
+ 12345678910
185
+ duration
186
+ duration
187
+
188
+ growth rate
189
+ perfect information
190
+ external cue
191
+ no cue
192
+ perfect memory
193
+ internal state
194
+ no memory
195
+ A
196
+ B
197
+ Figure 2: The long-term growth rate Λ of populations with different sources of information.
198
+ (A) The
199
+ value of external cues: Λmax is the maximum possible growth rate attainable if the population has perfect
200
+ information about the future environment.
201
+ Λbet is the highest growth rate achievable by a bet-hedging
202
+ population without receiving cues, which is suppressed by the entropy of the environment H(ε). Λcue is the
203
+ growth rate when the population utilizes a cue ξ that has a mutual information I(ε; ξ) with the environment.
204
+ (B) The value of internal memory: Organisms can utilize their internal states as memory, such that their
205
+ behavior depends on which state they are in. Λbet from bet-hedging also represents the case with no memory,
206
+ which corresponds to having only one internal state (L = 1). More states (L > 1) provides larger memory
207
+ capacity and allows a higher growth rate Λint for the population. Λmem is the highest growth rate achievable
208
+ by organisms with a perfect memory (L → ∞) of their lineage history.
209
+ These well-known results are summarized schematically in Fig. 2A. At the top level is the maximum
210
+ possible growth rate Λmax, which is attainable only if individuals have perfect information about
211
+ future environmental conditions and respond accordingly, i.e., germinate if it will be a good year
212
+ and go dormant if it will be bad. On the other hand, if there is no environmental cue, the best
213
+ strategy is bet-hedging with fixed probabilities, which achieves a growth rate Λbet. This is less than
214
+ Λmax by an amount H(ε), which is the Shannon entropy from information theory that quantifies
215
+ the uncertainty of the varying environment (See Appendix A.1). However, if a cue ξ is used to
216
+ help predict the environment, the population can increase the growth rate from Λbet to Λcue, up
217
+ by an amount I(ε; ξ) that is equal to the mutual information between the cue and the environment
218
+ (Appendix A.1). Note that Λcue is still not as high as Λmax unless the cue is fully accurate. The
219
+ relations between these growth rates illustrated here (similar to plots in [16, 32]) show that, in
220
+ order for the population to better adapt to varying environments, it must utilize available sources
221
+ of information about the environment.
222
+ 2.2
223
+ Internal source of information
224
+ Instead of sensing external cues, below we consider another possibility for organisms to use their
225
+ internal states as a source of information. We will use the age of seeds as an example. The state
226
+ diagram representing seed ages are illustrated in Fig. 1C, where a state sα represents a seed of age
227
+ α. A blue arrow represents a seed going into dormancy for one year, so that the age is increased
228
+ by 1.
229
+ An orange arrow represents a seed that germinates and potentially produces new seeds,
230
+ 5
231
+
232
+ which will have age 0.
233
+ The weights on the arrows represent the probability of germination or
234
+ dormancy. For a simple bet-hedging strategy without any cues, the probability of germination will
235
+ be a constant, which equals q∗ from Eq. (2), independent of the seed age. We will study the case
236
+ where the germination fraction can depend on the seed age, and show that the population can
237
+ acquire information from this internal state to achieve a higher growth rate.
238
+ 3
239
+ Results
240
+ 3.1
241
+ Seed age as an internal cue
242
+ We first study whether the seed age as an internal state contains useful information about the
243
+ environment. Let αt−1 be the seed age at the beginning of year t, and εt be the coming environment
244
+ that year. If αt−1 has no information about the environment, then it will be statistically independent
245
+ of εt, i.e., P(εt|αt−1) = P(εt). Therefore, whether seed age is informative about the environment
246
+ can be inferred from the conditional probability P(εt|αt−1).
247
+ To calculate that, we simulate a
248
+ sufficiently long sequence of environments, denoted by εt for each year t. We also simulate a single
249
+ lineage of plants that uses the constant germination fraction q∗. Each year the seed can either
250
+ germinate or stay dormant, and the probability of choosing the phenotype φt is further weighted
251
+ by the fitness f(εt, φt) to account for selection (see procedure in Appendix A.3). The seed age along
252
+ the lineage is recorded as αt. From the sequences of εt and αt, we estimate the joint probability
253
+ distribution P(εt, αt−1), from which the conditional probability P(εt|αt−1) is calculated. As shown
254
+ in Fig. 3, the probability of the environment εt does depend on the seed age αt−1. This means that
255
+ knowing the seed age allows a more accurate prediction of the coming environment. Therefore, it
256
+ is possible for the population to “co-opt” the seed age as an “internal cue” for the environment. In
257
+ analogy to the case of external cues, we expect that such information can be used to increase the
258
+ long-term population growth rate.
259
+ We therefore consider a strategy where the germination fraction depends on the seed age, denoted
260
+ by qα and represented by weights on the arrows in Fig. 1C. To calculate the long-term growth rate,
261
+ let N be a vector that represents the age-structured population, with components Nα being the
262
+ number of seeds of age α. The dynamics of N is described by a matrix M(ε; q) that depends on
263
+ the environment ε and the germination fractions q (with components qα),
264
+ M(ε; q) =
265
+
266
+
267
+
268
+
269
+
270
+
271
+
272
+ q0 Yε
273
+ q0 Yε
274
+ · · ·
275
+ (1−q1)V
276
+ 0
277
+ · · ·
278
+ 0
279
+ (1−q2)V
280
+ ...
281
+ ...
282
+ ...
283
+ ...
284
+
285
+
286
+
287
+
288
+
289
+
290
+
291
+ (3)
292
+ Each year, the population vector is multiplied by the matrix that corresponds to the current
293
+ 6
294
+
295
+ 0
296
+ 1
297
+ 2
298
+ 3
299
+ 4
300
+ 5
301
+ 6
302
+ 7
303
+ 8
304
+ 9
305
+ seed age,
306
+ 0.0
307
+ 0.2
308
+ 0.4
309
+ 0.6
310
+ 0.8
311
+ 1.0
312
+ probability, P( t|
313
+ t
314
+ 1)
315
+ bet-hedging
316
+ age-dependent
317
+ Figure 3: Probability of the coming environment εt conditioned on the seed age αt−1 at the beginning
318
+ of year t, as calculated by simulating a lineage of seeds. Dashed line is the marginal probability of the
319
+ environment, which would indicate that the seed age is uncorrelated with the environment. Blue bars are
320
+ when the population uses a bet-hedging strategy with a constant germination fraction. Orange bars are
321
+ when the germination fraction depends on the seed age to maximize population growth rate. In both cases
322
+ the seed age is correlated with the environment and thus useful as an internal cue.
323
+ environment εt,
324
+ N t = M(εt; q) · N t−1 ,
325
+ (4)
326
+ Here M(εt; q) is a random matrix because εt is a random variable. The temporal sequence of εt
327
+ is randomly drawn according to the distributions of good and bad years. The long-term growth
328
+ rate Λ of the population is then given by the Lyapunov exponent of the product of these random
329
+ matrices [33], which is calculated numerically (see methods in Appendix A.2).
330
+ We vary the age-dependent germination fractions qα to maximize Λ. As expected, this growth rate
331
+ using seed age as an internal cue (Λint) is greater than that of bet-hedging without cues (Λbet), as
332
+ illustrated in Fig. 2B (see also Fig. 6 below). The optimal germination fraction as a function of seed
333
+ age is shown in Fig. 4. An intuitive explanation for the age dependence is that, in this example, the
334
+ bad environment typically lasts a number of years, so it is advantageous for a seed to stay dormant
335
+ for a similar period of time to wait it out. Those that germinate in the wrong phase of the bad
336
+ year cycle will be eliminated by selection, and the remaining individuals tend to be synchronized
337
+ with the environment. In contrast, if there is no temporal structure in the environment, such as
338
+ when the environment is randomly and independently chosen each year, then the seed age will no
339
+ longer be correlated with the environment. In that case, the best strategy is to have a constant
340
+ germination fraction (equal to q∗ in the bet-hedging case, see Fig. 4), as argued in [28].
341
+ Note that the information about the environment is contained in the distribution of seed ages
342
+ within the population, which results from selection in previous years. Compared to the case of an
343
+ external cue that is shared by all individuals, the seed age varies among individuals (which prevents
344
+ 7
345
+
346
+ 0
347
+ 1
348
+ 2
349
+ 3
350
+ 4
351
+ 5
352
+ 6
353
+ 7
354
+ 8
355
+ 9
356
+ seed age,
357
+ 0.0
358
+ 0.2
359
+ 0.4
360
+ 0.6
361
+ 0.8
362
+ 1.0
363
+ germination fraction, q
364
+ temporally structured
365
+ uncorrelated environment
366
+ Figure 4: Dependence of the germination fraction q on the seed age α that maximizes the population growth
367
+ rate. Blue bars are when the environment is temporally structured, as described by the duration of good and
368
+ bad years in Fig. 1B. Orange bars are when the environment is drawn independently each year, for which the
369
+ germination fraction need not depend on seed age and is equal to the bet-hedging solution in Eq. 2 (dashed).
370
+ an analytic expression for Λ). It acts as an individual’s memory of its own lineage history, which
371
+ helps it infer the likely environment in the future. Importantly, the increase in population growth
372
+ rate does not come at any cost associated with sensing external cues. Thus, such an internal source
373
+ of information proves to be beneficial for the population.
374
+ 3.2
375
+ Internal states as memory
376
+ We have shown that internal states of organisms may help them “remember” the past outcomes of
377
+ selection to be able to predict the future environment, leading to an increased population growth
378
+ rate. Intuitively, the more the organisms can remember, the better they may predict and adapt to
379
+ the environment. To test this in our model, we can vary the memory size by changing the number
380
+ of possible internal states. The state diagram in Fig. 1C has potentially an infinite number of
381
+ states. They can be truncated at a finite number L, such that seeds exceeding age (L − 1) will
382
+ remain in the state sL−1 until they germinate or perish (Fig. 5A). This allows us to study how the
383
+ population growth rate depends on the number of states L.
384
+ We first note that having only one internal state (L = 1, Fig. 5B) is effectively having no memory,
385
+ because the system will always be in that same state regardless of the past events. In this case,
386
+ the germination fraction is always equal to q0 associated with the only state s0. Having a constant
387
+ germination fraction means that this case corresponds to the simple bet-hedging strategy. The
388
+ maximum long-term growth rate will just be Λbet achieved at q0 = q∗ found in Eq. (2).
389
+ For two internal states (L = 2, Fig. 5C), the model reduces to “phenotypic switching”, in which
390
+ 8
391
+
392
+
393
+
394
+ sL−1
395
+
396
+ A
397
+ B
398
+ C
399
+ s0
400
+ s1
401
+ s0
402
+ s0
403
+ s1
404
+ Figure 5: State diagrams for age-dependent germination. (A) The germination fraction q depends on the
405
+ seed age α up to α = L−1, beyond which it remains the same. Varying the length L effectively varies
406
+ the memory capacity of the organisms. (B) With only one state (L = 1), the organism effectively has no
407
+ memory, and the germination fraction is a constant, corresponding to simple bet-hedging. (C) The two-state
408
+ case corresponds to a Markov process where the organisms switch back and forth between two phenotypes,
409
+ with transition probabilities P(φ1|φ0) = q1 and P(φ0|φ1) = 1−q0.
410
+ the organisms randomly switch between two phenotypes (germination or dormancy) with fixed
411
+ transition probabilities. Specifically, the probability for a dormant seed to germinate next year is
412
+ q1, and the probability for a new seed (that came from a germinated plant) to go dormant is 1−q0.
413
+ This is a Markov process, for which the transition between phenotypes does not depend on how
414
+ long a phenotype has lasted. It implies that the germination fraction only depends on whether the
415
+ seed is fresh (age 0) or has been dormant (age > 0), but not on how long it has been dormant. As a
416
+ result of being Markovian, the duration of the dormant phenotype will be geometrically distributed.
417
+ A larger L will allow the germination fraction to depend more sensitively on the seed age (L > 2,
418
+ Fig. 5A). The number of states L roughly represents how many dormant years a seed can remember.
419
+ For each number L, we search for the maximum long-term growth rate Λ over the parameters
420
+ {q0, · · · , qL−1} (see methods in Appendix A.2). As shown in Fig. 6, Λ increases monotonically as
421
+ more states are incorporated. Therefore, more memory allows faster population growth and hence
422
+ better adaptation to environmental variation. Note that Λ quickly approaches a limit Λmem when
423
+ L becomes greater than the typical duration of the bad environment (equal to 5 in this example,
424
+ see Fig. 1B). Intuitively, there is no need to remember longer dormancy because there is no benefit
425
+ in staying dormant for longer than the duration of bad years. The relation between the growth
426
+ rate and memory is illustrated schematically in Fig. 2B.
427
+ If we think of seed age as an internal cue for the environment, we can calculate the mutual infor-
428
+ mation I(εt; αt−1) between the environment εt and the seed age αt−1, using the joint probability
429
+ P(εt, αt−1) calculated the same way as in Sec. 2.2. Fig. 6 shows that the mutual information also
430
+ increases with the number of states L, as more memory is available. When plotted against each
431
+ other, the long-term growth rate Λ increases with the mutual information I (Fig. 6 inset), just like
432
+ 9
433
+
434
+ 1
435
+ 2
436
+ 3
437
+ 4
438
+ 5
439
+ 6
440
+ 7
441
+ 8
442
+ 9
443
+ 10
444
+ number of states, L
445
+ 0.04
446
+ 0.10
447
+ 0.16
448
+ 0.22
449
+ 0.28
450
+ long-term growth rate,
451
+ -0.04
452
+ 0.02
453
+ 0.08
454
+ 0.14
455
+ 0.20
456
+ mutual information, I
457
+ -0.04
458
+ 0.08
459
+ 0.20
460
+ info, I
461
+ 0.04
462
+ 0.16
463
+ 0.28
464
+ growth,
465
+ growth rate
466
+ mutual info
467
+ Figure 6: Long-term growth rate Λ of populations that have different memory capacity as measured by the
468
+ number of internal states L. For each L, the age-dependent germination fractions qα are chosen to maximize
469
+ Λ. Also plotted is the mutual information I between the previous seed age αt−1 and the environment εt.
470
+ Both Λ and I increase monotonically with the memory capacity L, approaching their respective limits as
471
+ L ≫ 5 (mean duration of bad years). (Inset) Long-term growth rate Λ increases monotonically with the
472
+ mutual information I. Gray diagonal line represents Cohen’s model with external cues, in which Λ = Λbet+I.
473
+ for an external cue. Note that in Cohen’s model with external cues [14], Λ is simply proportional
474
+ to I (see Eq. (A14) in Appendix A.1). In comparison, for the same amount of information I, the
475
+ population achieves a higher growth rate Λ using seed age as an internal cue (Fig. 6 inset).
476
+ So far we have considered a very specific structure for the state diagrams (Fig. 5A, “age-diagram”).
477
+ It might be possible that, given the number of internal states, there are other diagrams that can
478
+ lead to a high long-term growth rate. Such diagrams could represent other types of internal states
479
+ instead of the age. For example, the reserve level of an organism can be represented by a linear
480
+ diagram, such that the organism moves up one or more states if it succeeds in foraging or moves
481
+ down one state if it fails [21].
482
+ To find which structure of internal states provides the highest
483
+ long-term growth rate for the population, we searched all possible diagrams of a given number of
484
+ states (up to L = 6, beyond which it is computationally difficult), optimizing the weights qα for
485
+ each diagram (see Appendix A.4). It turns out that the age-diagram in Fig. 5A is optimal for the
486
+ temporal structure of the environment that we assumed (Fig. 1B). In general, the state diagram is
487
+ a mathematical representation of memory, known as the “ϵ-machine” of a stochastic process [34];
488
+ a formal treatment and application to population growth in varying environments is given by [30].
489
+ 10
490
+
491
+ 1 2 3 4 5 6 7 8 9 10
492
+ duration
493
+ 0.00
494
+ 0.05
495
+ 0.10
496
+ 0.15
497
+ 0.20
498
+ 0.25
499
+ 0.30
500
+ distribution
501
+ Germination
502
+ A
503
+ 1 2 3 4 5 6 7 8 9 10
504
+ duration
505
+ Dormancy
506
+ B
507
+ L = 2
508
+ L = 5
509
+ L = 10
510
+ Figure 7: The distribution of the duration of consecutive germinations or dormant years along a lineage of
511
+ seeds. Different colors correspond to age-dependent germination fractions qα for different memory capacities
512
+ L. (A) For each L, the duration of germinations matches a geometric distribution with a mean of 1/q0
513
+ (dashed line for L = 2 and solid line for L = 10), meaning that there is no memory of previous germinations.
514
+ (B) The duration of dormancy has a distribution that changes shape depending on the memory capacity L.
515
+ L = 2 (phenotypic switching) results in a geometric distribution with a mean of 1/(1−q1) (dashed line).
516
+ Larger L’s result in deviation from a geometric distribution, which is indicative of having internal memory.
517
+ 4
518
+ Discussion
519
+ 4.1
520
+ Characterization of internal memory
521
+ Memory arising from age-dependent germination fractions can be characterized by the distribution
522
+ of the duration of dormancy. That is, given a large number of fresh seeds, what is the distribution
523
+ of the time that each seed stays dormant before germinating. To calculate this distribution, we
524
+ simulate one lineage of seeds over a long time in the absence of selection (see Appendix A.3), and
525
+ record the sequence of phenotypes, i.e., whether a seed germinated or not each year. Fig. 7 shows
526
+ the distribution of the number of consecutive years that successive seeds germinate or that a seed
527
+ stays dormant. The number of consecutive germination years is geometrically distributed with a
528
+ mean of 1/q0 (Fig. 7A), because every new seed has the same probability q0 of germinating. In
529
+ other words, a new seed has no memory of the age of the plant that it came from. Thus, the
530
+ absence of phenotypic memory is signified by the geometric distribution.
531
+ On the other hand, the distribution of the consecutive dormant years (i.e., the duration of dor-
532
+ mancy) depends on the number of internal states L. For L = 2, as discussed in Sec. 3.2, there is
533
+ no memory of how long a seed has been dormant. Indeed, the distribution of dormancy durations
534
+ is geometric with a mean of 1/(1−q1) (Fig. 7B). But as L increases, the distribution becomes
535
+ more bell-shaped and closer to the distribution of consecutive bad years (Fig. 1B). (In the limit
536
+ where the fitness matrix fεφ is diagonal, the optimal strategy will be such that the duration of each
537
+ phenotype exactly matches the distribution of the corresponding environment; see Appendix A.5).
538
+ The deviation of the distribution from being geometric indicates that the seed has memory of how
539
+ long it has been dormant, which is necessary for the germination fraction to depend on the seed
540
+ 11
541
+
542
+ age. Thus, the shape of the dormancy distribution can be used as an experimental signature of
543
+ internal memory.
544
+ The best demonstration of memory in phenotypic changes is found in experiments on the bacteria
545
+ Bacillus subtilis [35]. During its growth, B. subtilis can switch between two phenotypes, either as
546
+ a free-moving cell by making flagela or as part of an aggregate by producing extracellular matrix
547
+ [36, 37]. It is thought that the aggregate cells have an advantage for colonization and can better
548
+ cope with a harsh environment by sharing resources, whereas the motile cells are better at dispersing
549
+ and searching for nutrients. The durations of these two cell types along continuous cell lineages are
550
+ measured in a constant environmental condition [35]. It was found that the time a lineage stays in
551
+ the motile cell type follows an exponential distribution with a mean of ∼ 81 generations, while the
552
+ aggregate cell type is maintained for a narrowly distributed duration with a mean and standard
553
+ deviation of 7.6 ± 2.1 generations (see Fig. 2(d,f) of [35]). This implies that the motile cell type
554
+ is memoryless while the aggregate cell type has memory. That is, an aggregate cell keeps track
555
+ of how long it has been part of an aggregate, whereas a motile cell turns off motility with a fixed
556
+ probability at every cell division. These two distributions of phenotype durations look similar to
557
+ those found in our model (Fig. 7). Importantly, since the switching of cell types is measured in
558
+ a constant environment, it is evident that the phenotypic changes are influenced by some internal
559
+ states of the cell, rather than external cues. This method of inferring the existence of internal
560
+ memory by measuring the duration of phenotypes can be potentially applied to seeds. It would
561
+ require measuring the duration of seed dormancy by planting seeds in separate pots under the same
562
+ environmental condition and recording how soon they germinate.
563
+ 4.2
564
+ Evidence for age-dependent dormancy
565
+ Our model assumes that the probability of a seed entering or exiting dormancy depends on the
566
+ age.
567
+ If the bad environment typically persists for a number of years, then the model predicts
568
+ that the probability of exiting dormancy should be small initially and increase over a timescale
569
+ that matches the duration of bad years (Fig. 4). Data from past experiments have shown that
570
+ for different species the germination fraction can either increase or decrease between the first and
571
+ second years [23], while data going beyond the second year are scarce. To test the above prediction
572
+ also requires knowing the statistics of bad years. Alternatively, age-dependent germination can be
573
+ tested by measuring the distribution of dormancy durations, as discussed in Sec. 4.1 (Fig. 7B). For
574
+ that purpose, one has to measure the final age of seeds right before they germinate. Studies on
575
+ seed age structure have been done in the past [26, 27], but with the goal of measuring the current
576
+ age of seeds in a population at a given time, even though some seeds will continue to be dormant.
577
+ We are not aware of existing studies that measured the distribution of final seed ages.
578
+ Dormancy in other organisms can also be studied using our model. One example is insect diapause
579
+ [38], which is considered another example of bet-hedging. In many insect species, the larvae can
580
+ 12
581
+
582
+ enter diapause at a certain developmental stage to avoid unfavorable conditions, instead of pro-
583
+ ceeding with normal development to become adults. In a simple model of diapause [39], the larvae
584
+ may undergo multiple years of diapause and have a fixed probability of (re)entering diapause each
585
+ year (see Fig. 1 of [39]), similar to Cohen’s model of seed dormancy [11]. This would correspond
586
+ to our model with L = 1, such that the decision to enter diapause is memoryless. Another model
587
+ assumes that the larvae can only undergo one period of diapause and must exit after that [40]. This
588
+ pattern is a special case of our model with L = 2, where the state s0 would correspond to a new
589
+ larva and s1 to diapause. The larva can either develop to an adult with probability q0 and produce
590
+ offspring (arrow from s0 back to itself), or enter diapause with probability 1 − q0 (arrow to s1).
591
+ However, once it undergoes diapause, it must exit and develop, so there is only one arrow leaving
592
+ s1, which goes to s0 with probability q1 = 1. In this scenario, it was found that diapause is ben-
593
+ eficial in varying environments that are temporally correlated [40], in agreement with our results.
594
+ More generally, one may study situations where diapause can be repeated for a number of times,
595
+ which would correspond to a diagram like Fig. 5A. Our results suggest that which form of diapause
596
+ is evolutionarily favored depends on the complexity of temporal structure in the environmental
597
+ variation, which could potentially be tested in empirical studies.
598
+ 5
599
+ Conclusion
600
+ We have shown that the internal states of organisms can serve as a memory to help the population
601
+ adapt in varying environments. In order for this strategy to be useful, the environment must be
602
+ temporally structured, and the internal states must become correlated with the environment. We
603
+ have demonstrated that such correlation can arise from selection alone, without direct interaction
604
+ with the environment. More generally, some internal states of organisms may be correlated with
605
+ the environment as a result of phenotypic plasticity. For example, seeds produced in a good year
606
+ may be bigger than those produced in a bad year, so seed size could provide a memory of the
607
+ past environment. It is known that seed size can affect germination probability [41], and it will be
608
+ interesting to study if such dependence can benefit population growth in varying environments.
609
+ Organisms are complex systems with a lot of internal degrees of freedom, some of which might
610
+ happen to become correlated with the environment through selection or plasticity. Even though
611
+ these internal states might not have developed as sensors for environmental cues, they could be
612
+ co-opted as information sources to guide the organism’s behavior. To test whether seed age could
613
+ be co-opted to affect germination, one might compare accessions of annual plants in temporally
614
+ structured environments and those in unpredictable environments. Our model predicts that the
615
+ germination fraction would evolve to depend on the seed age in the former case.
616
+ Dormancy has been proposed to cause a “storage effect” that promotes species coexistence in vary-
617
+ ing environments [42]. Our model of age-dependent dormancy may be studied in such community
618
+ 13
619
+
620
+ ecology context. If the presence of other species is viewed as part of the environment for the focal
621
+ species, then internal states such as seed age could potentially provide a memory of past interac-
622
+ tion with those other species. For example, reserve level of the predator may be an indicator of
623
+ past encounters with prey [21]. History-dependent ecological interactions have been experimentally
624
+ indicated in microbial communities [43]. It will be interesting to use our framework to study such
625
+ ecological dynamics of organisms whose phenotypes depend on their memory.
626
+ A
627
+ Methods
628
+ A.1
629
+ Analytic derivation of Cohen’s model
630
+ Consider a population of annual plant seeds, each of which can either germinate (φ = 1) or stay
631
+ dormant (φ = 0) each year. The environment can be either good (ε = 1) or bad (ε = 0). If a seed
632
+ germinates in a good year, it will reproduce and yield Y1 number of seeds; but a seed germinating
633
+ in a bad year will only yield Y0 seeds, with Y1 > Y0 (in the main text we set Y0 to 0 for simplicity).
634
+ If a seed stays dormant, then the probability that it will remain viable is V . For Y1 > V > Y0, it is
635
+ favorable for a seed to germinate in a good year but stay dormant in a bad year. The number of
636
+ seeds at year t is denoted by Nt and obeys the equation:
637
+ Nt = Nt−1
638
+
639
+ (1 − q)V + qYεt
640
+
641
+ ,
642
+ (A1)
643
+ where εt is the environment in that year and q is the fraction of seeds that germinates. The number
644
+ of seeds at year T can be calculated recursively as:
645
+ NT = N0
646
+ T
647
+
648
+ t=1
649
+
650
+ (1 − q)V + qYεt
651
+
652
+ = N0
653
+
654
+ (1 − q)V + qY0
655
+ �T0�
656
+ (1 − q)V + qY1
657
+ �T1,
658
+ (A2)
659
+ where Tε is the total number of years that the environment is ε. The long-term growth rate Λ is
660
+ defined as the asymptotic rate of logarithmic increase:
661
+ Λ ≡ lim
662
+ T→∞
663
+ 1
664
+ T log NT
665
+ N0
666
+ = P0 log
667
+
668
+ (1 − q)V + qY0
669
+
670
+ + P1 log
671
+
672
+ (1 − q)V + qY1
673
+
674
+ ,
675
+ (A3)
676
+ where Pε ≡ lim
677
+ T→∞
678
+
679
+ T is the frequency of environment ε. The germination fraction q∗ that maximizes
680
+ Λ is found by setting the derivative ∂Λ
681
+ ∂q to zero, which gives (assuming q∗ > 0):
682
+ q∗ =
683
+ V P1
684
+ V − Y0
685
+
686
+ V P0
687
+ Y1 − V .
688
+ (A4)
689
+ And the corresponding maximum growth rate Λbet is:
690
+ Λbet = P0 log P0(Y1 − Y0)V
691
+ (Y1 − V )
692
+ + P1 log P1(Y1 − Y0)V
693
+ (V − Y0)
694
+ .
695
+ (A5)
696
+ 14
697
+
698
+ If the seeds have perfect information about the future environment, then they should all germinate in
699
+ good years and stay dormant in bad years. This would result in a total population NT = N0 V T0 Y T1
700
+ 1
701
+ instead of Eq. (A2), which gives the maximum possible growth rate:
702
+ Λmax = P0 log V + P1 log Y1 .
703
+ (A6)
704
+ The difference between Λmax and Λbet is then given by:
705
+ Λmax − Λbet = −P0 log P0(Y1 − Y0)
706
+ (Y1 − V )
707
+ − P1 log P1(Y1 − Y0)V
708
+ (V − Y0)Y1
709
+ .
710
+ (A7)
711
+ In the limit Y0 → 0 and Y1 ≫ V , it simplifies to:
712
+ Λmax − Λbet = −P0 log P0 − P1 log P1 ≡ H(ε) ,
713
+ (A8)
714
+ which is the entropy of the environment.
715
+ The model above can be generalized to include an external cue ξ that is correlated with the
716
+ environment ε. Assume that, given ξ, the seeds will germinate with probability P(φ = 1|ξ) ≡ qξ.
717
+ The total number of seeds then obeys the equation:
718
+ Nt = Nt−1
719
+
720
+ (1 − qξt)V + qξt Yεt
721
+
722
+ ,
723
+ (A9)
724
+ where ξt is the cue received in year t. Repeating the same procedure as above, one finds that the
725
+ population after T years becomes:
726
+ NT = N0
727
+
728
+ ε,ξ
729
+
730
+ (1 − qξ)V + qξYε
731
+ �Tεξ,
732
+ (A10)
733
+ where Tεξ is the number of years that the environment is ε while the cue is ξ. The long-term growth
734
+ rate is then given by:
735
+ Λ =
736
+
737
+ ε,ξ
738
+ Pεξ log
739
+
740
+ (1 − qξ)V + qξYε
741
+
742
+ ,
743
+ (A11)
744
+ where Pεξ = lim
745
+ T→∞
746
+ Tεξ
747
+ T
748
+ is the joint probability of the environment ε and the cue ξ. The optimal
749
+ germination fraction q∗
750
+ ξ that maximizes Eq. (A11) is given by (assuming q∗
751
+ ξ > 0):
752
+ q∗
753
+ ξ = V P1|ξ
754
+ V − Y0
755
+ − V P0|ξ
756
+ Y1 − V ,
757
+ (A12)
758
+ which is the same as Eq. (A4) except that Pε is replaced by the conditional probability Pε|ξ = Pεξ
759
+ Pξ .
760
+ The maximum growth rate achieved by using the external cue is then given by plugging Eq. (A12)
761
+ into Eq. (A11), which gives:
762
+ Λcue =
763
+
764
+ ε,ξ
765
+ Pεξ log Pε|ξ + P0 log (Y1 − Y0)V
766
+ (Y1 − V )
767
+ + P1 log (Y1 − Y0)V
768
+ (V − Y0) .
769
+ (A13)
770
+ The difference between Λcue and Λbet is then:
771
+ Λcue − Λbet =
772
+
773
+ ε,ξ
774
+ Pεξ log Pε|ξ
775
+
776
+ ≡ I(ε; ξ) ,
777
+ (A14)
778
+ which is precisely the mutual information between the environment ε and the cue ξ.
779
+ 15
780
+
781
+ A.2
782
+ Numerical solution for age-dependent germination
783
+ In our model where the germination fraction depends on the seed age, neither the growth rate nor
784
+ the optimal germination fraction has an analytic solution. Here we describe how they are calculated
785
+ numerically. Since the seeds are heterogeneous in age, the population is described by a vector N
786
+ with components Nα that represents the number of seeds of age α. As described in the main text,
787
+ the vector N t at year t obeys the equation:
788
+ N t = M(εt; q) · N t−1 ,
789
+ (A15)
790
+ where the matrix M depends on the current environment εt and the germination fractions qα ≡
791
+ P(φ=1|α), as given in Eq. (3). Thus, the population vector after a long time T is:
792
+ N T =
793
+ � T
794
+
795
+ t=1
796
+ M(εt; q)
797
+
798
+ · N 0 ,
799
+ (A16)
800
+ and the long-term growth rate is formally given by the largest Lyapunov exponent of the product
801
+ of matrices:
802
+ Λ = lim
803
+ T→∞
804
+ 1
805
+ T log
806
+ �����
807
+ T
808
+
809
+ t=1
810
+ M(εt; q)
811
+ ����� ,
812
+ (A17)
813
+ where | · | is the matrix norm, which we choose to define as the largest eigenvalue for non-negative
814
+ matrices. Compared to Cohen’s model, here Λ cannot be calculated analytically because the matrix
815
+ multiplications are non-commutative. To numerically calculate Λ, we simply use the above equation
816
+ with a very large T, as the limit is expected to converge [33].
817
+ We first draw a sequence of T random environments as follows. Define an epoch of time τε as the
818
+ number of consecutive years that the environment remains to be ε until it switches. The good and
819
+ bad epochs are drawn from the distributions:
820
+ P(τ1 =k) = 1
821
+ µ1
822
+
823
+ 1 − 1
824
+ µ1
825
+ �k−1
826
+ ,
827
+ k = 1, 2, · · · , ∞
828
+ (A18)
829
+ P(τ0 =k) = 1
830
+ Z exp
831
+
832
+ − (k − µ0)2
833
+ 2σ2
834
+
835
+ ,
836
+ k = 1, 2, · · · , 2µ0−1.
837
+ (A19)
838
+ Here µε is the mean duration for the epochs, σ characterizes the variability of the bad epochs, and
839
+ Z is a normalization constant. For the example used in the main text (Fig. 1B), µ1 = µ0 = 5 and
840
+ σ = 2. 50000 epochs are drawn for each environment, with a total length T ≈ 500000.
841
+ To calculate Λ, we need to calculate the product �T
842
+ t=1 M(εt; q). For convenience, we define M (s) ≡
843
+ �s
844
+ t=1 M(εt; q). Then M (T) can be calculated recursively by
845
+ M (t) = M(εt; q) · M (t−1),
846
+ (A20)
847
+ We normalize M (t) at every time step by the value of its largest entry, and this normalization
848
+ factor nt is stored. The Lyapunov exponent is then given by Λ = 1
849
+ T
850
+ � �T
851
+ t=1 log nt + log w
852
+
853
+ , where
854
+ 16
855
+
856
+ w is the largest eigenvalue of the normalized M (T) (which does not matter for Λ when T is large,
857
+ but matters for its derivative that we calculate below).
858
+ To find the germination fractions q∗
859
+ α that maximizes Λ, we use the optimization routine L-BFGS-B,
860
+ which allows us to impose the constraint 0 ≤ q∗
861
+ α ≤ 1. Besides the numerical function that calculates
862
+ Λ as described above, we also supply the Jacobian of the function, i.e., the derivative
863
+ ∂Λ
864
+ ∂qα . This
865
+ requires calculating the derivative of M (T) with respect to qα, which can be done using the recursive
866
+ relation
867
+ ∂M (t)
868
+ ∂qα
869
+ = ∂M(εt; q)
870
+ ∂qα
871
+ · M (t−1) + M(εt; q) · ∂M (t−1)
872
+ ∂qα
873
+ ,
874
+ (A21)
875
+ together with that for M (t) in Eq. (A20), from t = 1 all the way to T. We normalize ∂M(t)
876
+ ∂qα
877
+ by the
878
+ same factor nt as for M (t) at every time step. The derivative of Λ is then given by
879
+ ∂Λ
880
+ ∂qα
881
+ = 1
882
+ T
883
+ 1
884
+ |M (T)|
885
+ ∂|M (T)|
886
+ ∂qα
887
+ = 1
888
+ T
889
+ 1
890
+ w
891
+
892
+ u · ∂M (T)
893
+ ∂qα
894
+ · v
895
+
896
+ ,
897
+ (A22)
898
+ where u and v are the left and right eigenvectors of M (T) corresponding to its largest eigenvalue
899
+ w. This derivative is then supplied as the Jacobian to the L-BFGS-B optimization routine to find
900
+ the optimal q∗ that maximizes Λ.
901
+ A.3
902
+ Simulating a lineage
903
+ Simulation of a continuous lineage of seeds is used to estimate the joint probability P(εt, αt−1) of
904
+ the environment εt and the seed age αt−1 in Sec. 2.2, which is then used to calculate their mutual
905
+ information I(εt; αt−1) in Sec. 3.2. For a given set of germination fractions qα, the simulation is
906
+ done as follows. We start from a fresh seed of age 0. The sequence of environments, {ε1, · · · , εT },
907
+ is drawn beforehand as described in Sec. A.2.
908
+ In each year, we decide whether the seed germinates or not using the germination probability that
909
+ corresponds to its age. To account for selection bias, we weight the probabilities by the fitness
910
+ values in the current environment. That is, in year t, the seed along the lineage has probability
911
+ qαt−1Yεt
912
+ qαt−1Yεt + (1 − qαt−1)V
913
+ to germinate and reset the age to 0, and otherwise stays dormant with its age increased from
914
+ αt−1 to αt = αt−1 + 1. We repeat this procedure from t = 1 to T, recording the sequence of αt.
915
+ Afterwards, the number of times that the pair (εt, αt−1) takes a particular combination of values is
916
+ counted, which is then normalized to be the joint probability distribution P(εt, αt−1), from which
917
+ the mutual information I(εt; αt−1) is calculated.
918
+ Lineage simulation is also used to calculate the distribution of dormancy duration in Sec. 4.1, i.e.,
919
+ the distribution of how many consecutive years a seed stays dormant in the absence of environmental
920
+ 17
921
+
922
+ variation. To calculate this distribution, we once again start with a fresh seed of age 0 and use
923
+ the probability q0 to decide if the seed germinates. This time the probability is not weighted by
924
+ the fitness because we are calculating the dormancy durations in the absence of selection. The
925
+ above procedure is repeated for a long period of time T and the sequence of phenotypes at each
926
+ time step is recorded as φt. The duration of germination or dormancy is calculated by parsing the
927
+ sequence of phenotypes {φt} into consecutive epochs of germination or dormancy. The distribution
928
+ of their durations is then calculated by normalizing the histograms of these epochs. Note that these
929
+ distributions can also be calculated using Eq. (A30) in Appendix A.5.
930
+ A.4
931
+ Exhaustive search of state diagrams
932
+ To verify that the age-diagram in Fig. 5A is the optimal topology, we test all possible state diagrams
933
+ for up to 6 internal states. For a diagram with L states, we label the states as s0, s1, · · · , sL−1.
934
+ Each state has two outgoing arrows, corresponding to either dormancy or germination. Each arrow
935
+ can go to any other state or loop back. Therefore, naively, there can be L2L possible diagrams.
936
+ However, many of these diagrams are equivalent in the sense that they are simply permutations
937
+ of the states. To remove the redundant diagrams, we use a “sieve” method as follows. We first
938
+ represent a diagram by a (L × 2) integer matrix, whose entry of the α-th row and φ-th column
939
+ represents which state the system will transition to if it is at age α and expresses phenotype φ.
940
+ The diagrams are then indexed by a number that results from flattening the matrix and treating
941
+ it as a base-L number.
942
+ Then, we enumerate all L2L diagrams starting from the index 0.
943
+ For
944
+ each diagram, we find all its permutations and remove their indices from the list. Furthermore,
945
+ we exclude diagrams that have two or more disjoint parts to keep only connected diagrams. We
946
+ go over the list of diagrams, skipping the indices that have been removed. In the end, the total
947
+ number of non-degenerate diagrams for L = 1, 2, · · · is
948
+ n(L) = 1, 6, 52, 892, 21291, 658885, · · ·
949
+ which is the number of unlabeled, strongly connected, L-state, 2-input automata (Sequence A027835
950
+ from OLEIS). This number grows quickly and we are only able to study diagrams for up to L = 6.
951
+ For each of the diagrams with L states, we numerically find the optimal qα and the maximum Λ
952
+ as in Sec. A.2. This is computationally intensive and is done on a computer cluster. Then, among
953
+ all diagrams of L states, we find the optimal diagram with the largest Λ. For up to L = 6, it turns
954
+ out that the age-diagram is the optimal diagram for our model.
955
+ A.5
956
+ Analytical results for extreme selection
957
+ In the limit of extreme selection, the fitness matrix is diagonal, i.e., fεφ =
958
+ � V 0
959
+ 0 Y
960
+
961
+ . This means,
962
+ hypothetically, that a seed can survive only if it germinates in a good year or stays dormant in a
963
+ 18
964
+
965
+ bad year. In this case, the long-term growth rate Λ and the optimal germination fractions q∗
966
+ α have
967
+ analytical solutions. Indeed, the population becomes homogeneous because, once it encounters
968
+ a good year, only the seeds that germinate will survive, and subsequently the population will
969
+ consist of only fresh seeds. From then on, the seed age will be synchronized with the number of
970
+ consecutive bad years, and will be reset to 0 whenever there is a good year. Let βt−1 be the number
971
+ of consecutive bad years right before year t (which is 0 if the previous year is good). It will be equal
972
+ to the seed age αt−1 of the population at the beginning of year t. Therefore, the seed population
973
+ changes over time according to:
974
+ Nt = Nt−1(1 − qβt−1)V
975
+ or
976
+ Nt−1 qβt−1Y ,
977
+ (A23)
978
+ depending on whether the environment εt = 0 or 1. Over a period of time T, the number of seeds
979
+ will be:
980
+ NT = N0
981
+
982
+ β
983
+
984
+ (1 − qβ)V
985
+ �T0β�
986
+ qβY
987
+ �T1β ,
988
+ (A24)
989
+ where Tεβ is the number of years that the environment is ε while the previous number of consecutive
990
+ bad years is β. This equation has the same form as Eq. (A10), with the external cue ξ replaced by
991
+ β. The long-term growth rate has the expression
992
+ Λ ≡ lim
993
+ T→∞
994
+ 1
995
+ T log NT
996
+ N0
997
+ =
998
+
999
+ β
1000
+ P0β log[(1 − qβ)V ] +
1001
+
1002
+ β
1003
+ P1β log[qβY ]
1004
+ (A25)
1005
+ where Pεβ = lim
1006
+ T→∞
1007
+ Tεβ
1008
+ T
1009
+ is the joint probability of the environment εt and the number of bad years
1010
+ βt−1. Setting the derivative
1011
+ ∂Λ
1012
+ ∂qα = 0, the optimal germination fractions q∗
1013
+ α are found to be
1014
+ q∗
1015
+ α =
1016
+ P1α
1017
+ P0α + P1α
1018
+ ≡ P1|α ≡ P(εt =1|βt−1 =α) .
1019
+ (A26)
1020
+ Here P(εt =1|βt−1 =α) represents the conditional probability that the coming year is good, given
1021
+ that there has been α consecutive bad years. It is related to the duration distribution of bad years,
1022
+ P(τ0) from Eq. (A19), through
1023
+ P(εt =1|βt−1 =α) = P(τ0 =α)
1024
+ P(τ0 ≥α).
1025
+ (A27)
1026
+ An important consequence of this result is that, for the germination fractions q∗
1027
+ α, the dormancy
1028
+ duration of the seeds (as in Fig. 7B) will have the same distribution as the duration of bad years
1029
+ (Fig. 1B). This is because, by definition, qα ≡ P(φt = 1|αt−1 = α). Let δ0 denote the duration of
1030
+ dormancy, then similar to Eq. (A27), we have
1031
+ P(φt =1|αt−1 =α) = P(δ0 =α)
1032
+ P(δ0 ≥α) .
1033
+ (A28)
1034
+ Equating the left-hand sides of Eqs. (A27) and (A28) leads to, as stated above,
1035
+ P(δ0 =α) = P(τ0 =α) .
1036
+ (A29)
1037
+ 19
1038
+
1039
+ Incidentally, for a general qα, it can be shown that
1040
+ P(δ0 =α) = qα
1041
+ α−1
1042
+
1043
+ k=1
1044
+ (1 − qk) .
1045
+ (A30)
1046
+ References
1047
+ [1] Baskin CC, Baskin JM.
1048
+ Seeds: Ecology, Biogeography, and Evolution of Dormancy and
1049
+ Germination. 2nd ed. Elsevier Science; 2014.
1050
+ [2] Lennon JT, Jones SE. Microbial seed banks: The ecological and evolutionary implications of
1051
+ dormancy. Nat Rev Microbiol. 2011;9(2):119-30.
1052
+ [3] Lennon J, den Hollander F, Wilke-Berenguer M, Blath J. Principles of seed banks and the
1053
+ emergence of complexity from dormancy. Nat Commun. 2021;12(4807).
1054
+ [4] Balaban NQ, Merrin J, Chait R, Kowalik L, Leibler S. Bacterial persistence as a phenotypic
1055
+ switch. Science. 2004;12(5690):305.
1056
+ [5] Harms A, Maisonneuve E, Gerdes K. Mechanisms of bacterial persistence during stress and
1057
+ antibiotic exposure. Science. 2016;354(6318):aaf4268.
1058
+ [6] Manuse S, Shan Y, Canas-Duarte SJ, Bakshi S, Sun WS, Mori H, et al. Bacterial persisters
1059
+ are a stochastically formed subpopulation of low-energy cells. PLoS Biology. 2021;19(4).
1060
+ [7] Simons A. Modes of response to environmental change and the elusive empirical evidence for
1061
+ bet hedging. Proc R Soc B. 2011;278(1712):1601-9.
1062
+ [8] Grimbergen AJ, Siebring J, Solopova A, Kuipers OP. Microbial bet-hedging: the power of
1063
+ being different. Curr Opin Microbiol. 2015;25:67-72.
1064
+ [9] Seger J, Brockmann HJ. What is bet-hedging? In: Oxford Surveys in Evolutionary Biology.
1065
+ vol. 4; 1987. p. 182-211.
1066
+ [10] Philippi T, Seger J. Hedging one’s evolutionary bets, revisited. Trends Ecol Evol. 1989;4(2):41-
1067
+ 4.
1068
+ [11] Cohen D.
1069
+ Optimizing Reproduction in a Randomly Varying Environment.
1070
+ J Theor Biol.
1071
+ 1966;12:119-29.
1072
+ [12] Kussell E, Leibler S. Phenotypic diversity, population growth, and information in fluctuating
1073
+ environments. Science. 2005;23(5743):309.
1074
+ [13] Donaldson-Matasci MC, Lachmann M, Bergstrom CT. Phenotypic diversity as an adaptation
1075
+ to environmental uncertainty. Evol Ecol Res. 2008;10(4):493-515.
1076
+ 20
1077
+
1078
+ [14] Cohen D. Optimizing Reproduction in a Randoruly Varying Environment when a Correlation
1079
+ May Exist between the Conditions at the Time a Choice has to be Made and the Subsequent
1080
+ Outcome. J Theor Biol. 1967;16:1-14.
1081
+ [15] Clauss MJ, Venable DL. Seed Germination in Desert Annuals: An Empirical Test of Adaptive
1082
+ Bet Hedging. Am Nat. 2000;155(2):168-86.
1083
+ [16] Donaldson-Matasci M, Bergstrom C, Lachmann M. The fitness value of information. Oikos.
1084
+ 2010;119(2):219-30.
1085
+ [17] Rivoire O, Leibler S. The Value of Information for Populations in Varying Environments. J
1086
+ Stat Phys. 2011;142:1124-66.
1087
+ [18] Auld JR, Agrawal AA, Relyea RA. Re-evaluating the costs and limits of adaptive phenotypic
1088
+ plasticity. Proc R Soc B. 2010;277(1681):503-11.
1089
+ [19] DeWitt TJ, Sih A, Wilson DS. Costs and limits of phenotypic plasticity. Trends Ecol Evol.
1090
+ 1998;13(2):77-81.
1091
+ [20] Murren CJ, Auld JR, Callahan H, Ghalambor CK, Handelsman CA, Heskel MA, et al. Con-
1092
+ straints on the evolution of phenotypic plasticity: Limits and costs of phenotype and plasticity.
1093
+ Heredity. 2015;115(4):293-301.
1094
+ [21] Higginson A, Fawcett T, Houston A, McNamara J. Trust your gut: using physiological states
1095
+ as a source of information is almost as effective as optimal Bayesian learning. Proc R Soc B.
1096
+ 2018;285.
1097
+ [22] McNamara JM, Houston AI. Starvation and predation as factors limiting population size.
1098
+ Ecology. 1987;68(5):1515-9.
1099
+ [23] Philippi T. Bet-Hedging Germination of Desert Annuals: Beyond the First Year. Am Nat.
1100
+ 1993;142(3):474-87.
1101
+ [24] Venable DL. Bet hedging in a guild of desert annuals. Ecology. 2007;88(5):1086-90.
1102
+ [25] Gremer JR, Venable DL. Bet hedging in desert winter annual plants: optimal germination
1103
+ strategies in a variable environment. Ecol Lett. 2014;17(3):380-7.
1104
+ [26] Kalisz S.
1105
+ Experimental Determination of Seed Bank Age Structure in the Winter Annual
1106
+ Collinsia Verna. Ecology. 1991;72(2):575-85.
1107
+ [27] Kalisz S, McPeek M. Demography of an Age-Structured Annual: Resampled Projection Ma-
1108
+ trices, Elasticity Analyses, and Seed Bank Effects. Ecology. 1992;73(3):1082-93.
1109
+ [28] Valleriani A, Tielb¨orger K. Effect of age on germination of dormant seeds. Theor Popul Biol.
1110
+ 2006;70:1-9.
1111
+ 21
1112
+
1113
+ [29] Lambert G, Kussell E.
1114
+ Memory and Fitness Optimization of Bacteria under Fluctuating
1115
+ Environments. PLoS Genet. 2014;10(9):e1004556.
1116
+ [30] Marzen SE, Crutchfield JP. Optimized bacteria are environmental prediction engines. Phys
1117
+ Rev E. 2018;98(1):12408.
1118
+ [31] Rescan M, Grulois D, Ortega-Aboud E, Chevin LM. Phenotypic memory drives population
1119
+ growth and extinction risk in a noisy environment. Nat Ecol Evol. 2020;4(2):193-201.
1120
+ [32] Xue B, Leibler S. Benefits of phenotypic plasticity for population growth in varying environ-
1121
+ ments. Proc Natl Acad Sci. 2018;115(50):12745–12750.
1122
+ [33] Crisanti A, Paladin G, Vulpiani A.
1123
+ Products of Random Matrices: in Statistical Physics.
1124
+ Springer-Verlag; 2012.
1125
+ [34] Shalizi CR, Crutchfield JP. Computational mechanics: Pattern and prediction, structure and
1126
+ simplicity. J Stat Phys. 2001;104(3-4):817-79.
1127
+ [35] Norman T, Lord N, Paulsson J, Losick R. Memory and modularity in cell-fate decision making.
1128
+ Nature. 2013;503:481-6.
1129
+ [36] Kearns D, Losick R. Cell population heterogeneity during growth of Bacillus subtilis. Genes
1130
+ Dev. 2005;19:3083-94.
1131
+ [37] L´opez D, Fischbach M, Chu F, Losick R, Kolter R. Structurally diverse natural products
1132
+ that cause potassium leakage trigger multicellularity in Bacillus subtilis. Proc Natl Acad Sci.
1133
+ 2009;106(1):280-5.
1134
+ [38] Menu F, Desouhant E. Bet-hedging for variability in life cycle duration: Bigger and later-
1135
+ emerging chestnut weevils have increased probability of a prolonged diapause.
1136
+ Oecologia.
1137
+ 2002;132(2):167-74.
1138
+ [39] Rajon E, Desouhant E, Chevalier M, D´ebias F F Menu. The Evolution of Bet Hedging in
1139
+ Response to Local Ecological Conditions. Am Nat. 2014;184(1).
1140
+ [40] Tuljapurkar S, Istock C. Environmental uncertainty and variable diapause. Theor Popul Biol.
1141
+ 1993;43:251-80.
1142
+ [41] Larios E, Burquez A, Becerra J, Venable D. Natural selection on seed size through the life
1143
+ cycle of a desert annual plant. Ecology. 2014;95(11).
1144
+ [42] Chesson P.
1145
+ Mechanisms of maintenance of species diversity.
1146
+ Annu Rev Ecol Syst.
1147
+ 2000;31(1):343-66.
1148
+ [43] Frentz Z, Kuehn S, Leibler S. Strongly deterministic population dynamics in closed microbial
1149
+ communities. Phys Rev X. 2015;5(4):041014.
1150
+ 22
1151
+
2NE1T4oBgHgl3EQf5QXz/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
3tFKT4oBgHgl3EQf8i6_/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6b4e7b55e98a505684a7d30727e83fc7fd7a0c05c9bbd6cf99197c3029ae38d
3
+ size 130341
4NA0T4oBgHgl3EQfNf9p/content/tmp_files/2301.02147v1.pdf.txt ADDED
@@ -0,0 +1,641 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ 1
3
+
4
+
5
+ Physical Realization of a Hyper Unclonable Function
6
+
7
+ Sara Nocentini*1,2, Ulrich Rührmair3,4, Mauro Barni5, Diederik S. Wiersma1,2,6, Francesco Riboli*2,7
8
+ 1 Istituto Nazionale di Ricerca Metrologica, Strada delle Cacce 91, 10135 Torino, Italy; 2 European Laboratory for
9
+ Nonlinear Spectroscopy, Via Nello Carrara 1, 50019 Sesto Fiorentino (FI), Italy; 3 Physics Dept. LMU Munchen,
10
+ Schellingstraße 4/III D-80799 Munchen, Germany; 4Electrical and Computer Engineering (ECE) Dept., University
11
+ of Connecticut, Storrs, CT, USA; 5 Dipartimento di Ingegneria dell’Informazione e Scienze Matematiche, Università
12
+ di Siena, via Roma 56, 53100 Siena; 6 Dipartimento di Fisica, Università di Firenze, Via Sansone 1, 50019, Sesto
13
+ Fiorentino, Italia; 7CNR-INO, Via N. Carrara 1 Sesto Fiorentino, 50019, Italy.
14
+ *nocentini@lens.unifi.it; riboli@lens.unifi.it
15
+
16
+ Disordered photonic structures are promising materials for the realization of physical
17
+ unclonable functions (PUF) – physical objects that can overcome the limitations of
18
+ conventional digital security methods1–6 and that enable cryptographic protocols
19
+ immune against attacks by future quantum computers 7,8. One PUF limitation, so far, has
20
+ been that their physical configuration is either fixed or can only be permanently
21
+ modified9, and hence allowing only one token per device. We show that it is possible to
22
+ overcome this limitation by creating a reconfigurable structure made by light-
23
+ transformable polymers, in which the physical structure of the unclonable function itself
24
+ can be reversibly reconfigured. We term this novel concept Hyper PUF or HPUF in that
25
+ it allows a large number of physical unclonable functions to co-exist simultaneously
26
+ within one and the same device. The physical transformation of the structure is done all-
27
+ optically in a reversible and spatially controlled fashion. Our novel technology provides
28
+ a massive enhancement in security generating more complex keys containing a larger
29
+ amount of information. At the same time, it allows for new applications, for example
30
+ serving multiple clients on a single encryption device and the practical implementation
31
+ of quantum secure authentication of data10.
32
+
33
+ Complex photonic systems11–17 are characterized by a multitude of spatial degrees of freedom that in
34
+ presence of coherent light illumination produce in the far field a complex intensity pattern (called speckle
35
+ pattern) as the result of the interference of a large number of independent transmission channels18. In
36
+ particular, the optical speckle pattern that is generated by disordered materials is extremely sensitive to
37
+ minute changes in the physical structure of the material19,20, to the level that it is nearly impossible to clone
38
+ such disordered structures and obtain the same optical response without resorting to cloning techniques at
39
+ the molecular level. Such structural characteristics make them ideal candidates for cryptographic primitives
40
+ such as physical unclonable functions for authentication and communication purposes 2,3. Among the other
41
+ types of PUFs, electrical Strong PUFs have been examined intensively by the PUF community 4,21,22, but
42
+ most of them have been attacked successfully via various digital and physical techniques over the years23,24.
43
+ Due to their promise of higher three-dimensional complexities and entropy levels, this has put optical PUFs
44
+ back in the focus of recent PUF research.
45
+
46
+
47
+ 2
48
+
49
+ Optical physical unclonable functions have been introduced by Pappu6 with the name of Physical One-Way
50
+ Functions. In this first instantiation, the PUF interrogation and the resulting challenge-response pair (CRP)
51
+ protocol4,13–15 relied on different angles of incidence of the laser and allowed to extract cryptographic keys
52
+ with 230 independent bits (over a total bit string length of 2400 bits). While the optical setup based on
53
+ moveable mechanical components limits the reproducibility of measurements, in later works the employ of
54
+ modulators as challenge generators in the spatial25,26 or spectral27,28 domain provided a significant
55
+ improvement. However, those PUFs rely on a static hardware whose properties cannot be reconfigured in
56
+ case of detected attack. To overcome this limitation, Kursawe et al. showed that permanent modifications
57
+ can be created by melting the polymer aggregates with a net entropy decrease in every new reconfiguration
58
+ 29. Horstmeyer and coauthors showed that it is possible to reconfigure an optical PUF by exploiting
59
+ electrical driven polymer dispersed liquid crystals 25,23and John et al. managed to do this electrically by
60
+ using halide perovskite memristors30. In all these cases, the internal states of the PUF cannot be recovered
61
+ after reconfiguration, and their entropy remains constant25. To increase the information entropy of the PUF,
62
+ it is necessary to provide a reversible transformation among the possible microscopic configurations. A
63
+ preliminary result in this direction was obtained by Gan and coauthors, who reported that the temperature-
64
+ controlled phase transition of Vanadium oxides nanocrystals can be used to create a reversible switching
65
+ among two states (crystalline and amorphous) 31.
66
+ In this context, we introduce a new concept and technology platform that provides interchangeable multi-
67
+ level operation by reversibly transforming the scattering properties of a complex photonic medium based
68
+ on photosensitive polymeric film. The operation principles of this cryptographic primitive – that we term
69
+ Hyper PUF (HPUF) – is illustrated in Fig. 1a. A “standard” PUF (left panel of Fig. 1) is characterized by
70
+ an authentication process via a single challenge CiProbe, while the HPUF (right panel of Fig. 1) is interrogated
71
+ by a challenge Cik = (CiProbe, LkTrans) consisting of two sub-challenges. First, a configuration pattern LkTrans
72
+ (a spatially modulated parametric matrix) transforms the internal configuration of the PUF between
73
+ different levels in an all-optical and reversible manner. The configuration pattern determines the scattering
74
+ potential. Each scattering potential is associated to a different level of the HPUF. Secondly, a standard
75
+ interrogation challenge CiProbe produces a measurable unique optical interference pattern as the PUF
76
+ response Rik(CiProbe, LkTrans). Mathematically, the HPUF can be modelled as a parametric function that maps
77
+ its domain to a larger codomain, whose dimension depends not only on the number of CiProbe but also on
78
+ the number of transformer challenges LkTrans, i.e. f : (CiProbe, LkTrans) → Rik. The same internal configuration
79
+ can be restored by applying the same transformer challenge, allowing back-and-forth switching between
80
+ the PUF’s internal levels. This marks a significant difference between HPUFs and existing reconfigurable
81
+ PUF designs,9,30,32 in which internal changes are permanent and non-reversible.
82
+ The practical usage of physical unclonable functions is governed by a registration and verification protocol
83
+ of the challenge-response-pairs33 that for standard and HPUFs differ in the library dimensionality and the
84
+ type of challenge sent to the claimant. To discriminate between legitimate and fraudulent authentication
85
+ requests, the similarity of two binary keys needs to be evaluated. Among the several metrics (such as
86
+ standard error, Pearson correlation coefficient, and mutual information), our analysis exploits the fractional
87
+ hamming distance (FHD) – i.e. the percentage of bits that differs between two binary strings. This is a
88
+ common choice both in biometrics and in PUF characterization 6,34. The FHD distribution between the
89
+ responses to the same challenge (like FHD) quantifies the stability of the system, while that one between
90
+ the responses to different random challenges (unlike FHD) is used to evaluate the correlation of the
91
+ independent responses. Indeed, following the method introduced by Daugman 27,28, the number N of
92
+ independent bits (the entropy) of the generated keys – i.e. the number of independent degrees of freedom –
93
+ can be estimated by assuming that the unlike FHD can be modeled with an equivalent binomial distribution
94
+ B(N, p) and expressed as a function of the mean value p and standard distribution of the curve σ, 𝑁 =
95
+
96
+
97
+ 3
98
+
99
+ �∗(���)
100
+ ��
101
+ 34,35. We refer to intra-device FHDs when comparing responses from the same PUF or inter-device
102
+ FHDs when comparing responses from different PUFs.
103
+
104
+ Figure 1. Schematic representation of the interrogation process for standard and Hyper PUFs. Working
105
+ mechanism of the deterministic behavior for the challenge response pair generation for standard PUFs (left panel)
106
+ and HPUFs (right panel). For the standard PUF, the challenge CiProbe probes the only possible internal configuration
107
+ of the hardware, producing only one response Ri to a given challenge. In the HPUF, each configuration pattern
108
+ reversibly transforms the PUF level into a new one, producing different responses Rik to a given challenge CiProbe.
109
+ The HPUF is a 3D disordered photonic medium that is responsive to the transformer challenge while
110
+ unperturbed to the probing challenge. It consists of a polymer film where liquid crystal (LC) droplets are
111
+ randomly dispersed via an emulsion process resulting in polymer dispersed and polymer stabilized liquid
112
+ crystals (PD&SLC)36 as shown in Fig. 2a. The response selectivity between the transformer and the stimulus
113
+ challenges is achieved by doping the common liquid crystal 5CB with a blue absorbing dye (dispersed red
114
+ 1, DR1). Blue incoherent light (LkTrans) transforms the internal state of the PUF by absorbing light and
115
+ thereby generating a temperature driven LC phase transition, while red coherent light (CiProbe) probes the
116
+ transformed PUFs. The LC droplets are further stabilized with cross linker molecules (Fig. 2a) that create
117
+ a fixed polymeric network37 to favor the recovery of the LC alignment in the nematic phase after the phase
118
+ transition (Fig. 2b,c). Fig. 2d-f show the polarized optical microscope characterization of the nematic-
119
+ isotropic-nematic phase transition of the LC within the illuminated spot of blue light. The presence of the
120
+ cross linker molecules guarantees an hysteresis-free process – i.e. a reversible switching between the two
121
+ LC phases38. The transformation between different internal configurations is deterministic, stable, and
122
+ repeatable, regardless of the history of the system.
123
+
124
+ Standard PUF
125
+ HyperPUF
126
+ Challenge
127
+ Levels
128
+ C.Probe
129
+ Trans
130
+ Trans
131
+ Trans
132
+ Challenge
133
+ CProbe
134
+ PUF
135
+ PUFL
136
+ Trans
137
+ PUF (L<Trans)
138
+ PUF (L,Trans)
139
+ Ril
140
+ RiK
141
+ RiN
142
+ R
143
+ Responses
144
+ Responses
145
+ 4
146
+
147
+ Figure 2. Polymer dispersed and polymer stabilized liquid crystals. a) Scheme of the polymeric film used as
148
+ disordered photonic medium to realize the HPUF. The LC droplets, whose molecular composition is reported on the
149
+ left, are randomly dispersed into the polymeric matrix (polydimethylsiloxane, PDMS). The polymer stabilized LC
150
+ formulation is made by a mesogen (5CB), a chromophore (Dispersed Red 1, DR1), and a bi-acrylate (cross-linker)
151
+ mesogen that enables a full recovery of the LC alignment. A scanning electron microscope image of the side view of
152
+ the film is reported on the right. b-c) Representative scheme of the molecule arrangement inside the PD&SLC droplets
153
+ for the switchable operation. b) Scheme of the LC molecular arrangement within the droplets in the nematic and c)
154
+ isotropic phase. d-f) Polarized Optical Microscope images of the PD&SLC before (d), during (e) and after (f) the blue
155
+ light illumination indicated by the dashed blue circle. The four dot cross pattern (d) is a signature of the LC radial
156
+ alignment in each droplet39 and it is lost under blue laser illumination. This is the indication that the stabilized LC
157
+ polymer does not prevent full LC disordering to the isotropic phase. Once the blue illumination is removed (e), the
158
+ system evolves in around 10 seconds to the previously aligned configuration with the same four dot feature that was
159
+ present before the transformation (as highlighted by the green circles in f). The scale bars are 10 µm in length.
160
+ The experimental characterization of a HPUF is illustrated in Fig. 3a. The system is illuminated with the
161
+ challenge CiProbe that is generated by modulating the Gaussian wavefront of an He-Ne laser using a digital
162
+ micro-mirror device (DMD) 40,41. Light is then scattered by the HPUF, generating in the far field the
163
+ response Rik (the speckle pattern) – a 2D image whose spatial features depend uniquely on the probe CiProbe
164
+ and transform challenge LkTrans of the system. The response Rik is imaged on a CCD camera, then filtered
165
+ and binarized to generate the key. The raw speckle images (the optical responses Rik) are converted into
166
+ binary keys by using a Gabor filter to remove pixel-scale noise, averaging the undesired intensity variations
167
+ and extracting the independent bits3,27. The parameters of the Gabor filter have been tuned in order to
168
+ maximize the extractable entropy from the PUF response.
169
+ Switching between the levels of the HPUF is triggered by the bright blue profile LkTrans (spatially overlapped
170
+ on the bright red challenge CiProbe), generated by a standard projector.
171
+
172
+ a
173
+ 5CB
174
+ PD&SLC
175
+ CL
176
+ DR1
177
+ 100μm
178
+ Blue lightOFF:nematicphase
179
+ C Blue light ON: isotropic phase
180
+ Phase
181
+ transition
182
+ Full recovery
183
+ 5CB
184
+ 5CB
185
+ CL
186
+ PDMS
187
+ CL
188
+ PDMS
189
+ DR1
190
+ DR1
191
+ 8
192
+ Blue light OFF
193
+ Blue light ON
194
+ Blue light OFF
195
+ 5
196
+
197
+
198
+ Figure 3. Hyper PUF characterization: one and two-level operation. a) Schematics of the HPUF characterization
199
+ setup where a binary challenge CiProbe is incident on the physical hardware. The transmitted intensity profile Rik is
200
+ collected in the far field by a CCD camera and successively converted into a binary post-processed key. A second
201
+ light beam LkTrans, generated by a blue LED and spatially modulated by a DMD (integrated on a projector board), is
202
+ used to reversibly transform the HPUF. Among the RGB colors of the projector, the blue LED was chosen as it better
203
+ matches the dye absorption peak. The two optical images are overlapped on the token. b) Characterization of one-
204
+ single level of the HPUF. The mean value of the like FHD shows that, on the average, the 12% of the bit between two
205
+ keys generated by the same challenges are different. The test has been made with 400 challenges. The unlike FHD
206
+ (orange histogram) is the result of 79800 pairwise comparisons that results from all possible comparisons of the 400
207
+ responses to random challenges. On average, each pair of generated keys differs in the 50% of its bits and the number
208
+ of the independent bits is N1Level =928 bits. c) Temporal evolution of the Pearson correlation coefficient between two
209
+ raw responses (two speckle images) acquired every two seconds by interrogating the sample with the same challenge
210
+ CiProbe while switching the configuration beam between L1Trans and L2 Trans. The three curves correspond to three
211
+ different values of the blue-light intensities (60, 85, 105 mW/cm2), indicating that higher intensities induce a more
212
+ efficient decorrelation as well as faster dynamics. We observe a full recovery of the LC alignment after every LC
213
+ phase transition. d) Like and unlike FHDs of the HPUF for a two-level configuration (intensity: 85mW/cm2). The
214
+ number of the independent bits of the generated key is N2Level =1323 bits. In the inset, we report the responses for the
215
+ two transformer challenges L1Trans and L2Trans to two different challenges, C1Probe and C2Probe.
216
+ The characterization of the HPUF has been performed by evaluating the entropic content of the keys
217
+ generated by an increasing number of levels: from one-single level PUF up to a ten-level HPUF. We firstly
218
+ characterize the one-single level system by interrogating the HPUF with challenges (CiProbe, LTrans=0) with
219
+ i={1,…,100}. Fig. 3b shows that the like and unlike FHDs distributions are well separated, and that the
220
+ authentication threshold can be safely set around 0.35. The number of the independent bits of the generated
221
+ keys is estimated to be N1Level=928 bits.
222
+
223
+ One-levelPUF
224
+ b
225
+ a
226
+ 0.14
227
+ Challenge C,probe
228
+ HPUF
229
+ Response Rik
230
+ 1
231
+ Intra-deviceunlikeFHD
232
+ 1
233
+ 0.12
234
+ Intra-devicelikeFHD
235
+ 1
236
+ He-Ne
237
+ 0
238
+ 0.1
239
+ Laser &
240
+ 1
241
+ DMD
242
+ Hashing
243
+ 0
244
+ Gabor
245
+ 0
246
+ N=928bits
247
+ 1
248
+ 0.06
249
+ 0
250
+ 0.04
251
+ Projector:
252
+ 0
253
+ Blue LED&
254
+ 1
255
+ 0.02
256
+ DMD
257
+ 0
258
+ 0
259
+ 0
260
+ 1
261
+ 0
262
+ 0.1
263
+ 0.2
264
+ 0.3
265
+ 0.4
266
+ 0.5
267
+ 0.6
268
+ FHD
269
+ p
270
+ c
271
+ Two-level HPUF
272
+ 0.2
273
+ 60mW/cm
274
+ UnlikeFHD
275
+ 85mW/cm²
276
+ Trans
277
+ LikeFHD
278
+ 105mW/cm
279
+ 0.8
280
+ C,Probe
281
+ 0.15
282
+ *0
283
+ O口
284
+ Frequency
285
+
286
+ 0.1
287
+ N=1323bits
288
+
289
+
290
+
291
+ 0.05
292
+ Lo
293
+ 0.2
294
+
295
+
296
+ 0
297
+ 0
298
+ 50
299
+ 100
300
+ 150
301
+ 200
302
+ 250
303
+ 300
304
+ 350
305
+ 0
306
+ 0.1
307
+ 0.2
308
+ 0.3
309
+ 0.4
310
+ 0.5
311
+ 0.6
312
+ Time (sec)
313
+ FHD
314
+ 6
315
+
316
+ The next step is to characterize the two-level HPUF. The first level is obtained by completely shading the
317
+ blue light, while the second one is configured by illuminating the PUF with a uniform blue wavefront. The
318
+ responses of the two levels interrogated with the same challenge are well decorrelated and also reproducible
319
+ (see Fig. 3c). The challenge-response characterization of a two-level HPUF is done by illuminating each
320
+ level with the same set of 100 random challenges (Fig. 3d). The unlike FHD distribution is obtained by
321
+ comparing all possible pairs of responses (roughly 2*104 pairwise comparisons), while the like FHD
322
+ distribution is obtained by comparing a defined set of 150 random challenges acquired multiple times for
323
+ each level. The two distributions do not overlap and the authentication threshold can be set around 0.4. We
324
+ also observe a net gain in the independent bits (entropy) of the keys generated by the two-level HPUF with
325
+ respect to a one-level HPUF, from N1Level= 928 bits to N2Levels= 1323 bits (Fig. 3b and Fig. 3d). This is the
326
+ indication that two keys generated by the two-level HPUFs have a greater probability of differing in 50%
327
+ of the bits (optimal situation), compared to two keys of one-level PUFs.
328
+ The natural question that arises is whether, and to which extent, a further increase in the number of levels
329
+ of the HPUF increases the entropy of the generated keys. To investigate this problem, we define a set of 10
330
+ transformer challenges (LkTrans, with k={1,..,10}) by choosing 10 elements of the Walsh-Hadamard binary
331
+ basis – i.e. a subset of a complete 16 orthogonal set of 4x4 macropixel images. Each level is interrogated
332
+ with the same set of randomly selected challenges CiProbe with i={1,.., 100}. Fig. 4a shows the scheme of
333
+ the domain and codomain (Rik, with i={1,.., 100} and k={1,.., 10}) of the HPUF. The whole codomain of
334
+ responses Rik can be compartmentalized by randomly joining the codomains of individual levels. For each
335
+ compartment, we evaluate the entropy per symbol of the keys, i.e. the entropy per bit. Fig. 4b left panel
336
+ shows that the number of independent bits is of around 950 (0.14 bit/bit). This value is almost independent
337
+ on the chosen level (in this case, each compartment is composed by the codomain of a single level). By
338
+ populating the compartments with the codomains of more levels (up to ten), the entropy of the generated
339
+ key increases up to around N10Levels= 1750 bits that correspond to 0.24 bit/bit (Fig. 4b, left panel, red circles).
340
+ The increase of the entropy per symbol evaluated by the Daugman’s analysis is confirmed by modeling the
341
+ extracted keys with equivalent Markov chains, generated via transition matrices whose coefficients
342
+ represent the permanence and transition probabilities of the binary values of the keys. The entropy per
343
+ symbol of the Markov chains is then calculated analytically30. The fact that the experimental data analyzed
344
+ with two models show the same entropic trend suggests that the different levels behave like different
345
+ cryptographic primitives coexisting in the same hardware.
346
+
347
+ To validate this idea, we fabricated ten different cryptographic primitives. We applied the same
348
+ compartmentalization scheme making an analogy among each codomain of the ten different PUFs and each
349
+ codomain of the ten levels of the HPUF. We observe that the entropy of the generated keys as a function of
350
+ the number of PUFs, has absolute values and a dependence qualitatively similar to the HPUF (Fig. 4b, left
351
+ panel, black circles). This is the confirmation that the transformer challenges LkTrans induce different
352
+ microscopic configurations in the same region of the sample, mimicking different PUFs. It is important to
353
+ notice that the entropy increase does not depend on the number of responses of the PUFs but only on the
354
+ number of levels of the HPUF. Increasing the number of responses but considering a single configuration,
355
+ the entropy per symbol of the PUF remains constant.
356
+ The increase in entropy per symbol implies a greater unpredictability of the bit sequence. By analyzing the
357
+ properties of the equivalent Markov chains, we observe that the increase of the number of levels leads to
358
+ permanence and transition (α and 1-α respectively) probabilities of the Markov transition matrix, that tend
359
+ towards a situation of equiprobability (α = 0.5). This implies the reduction of the correlation length in the
360
+ bit sequence (Fig. 4c-d). Indeed, the correlation length of the bit sequence gets shorter and shorter when
361
+
362
+
363
+ 7
364
+
365
+ increasing the number of levels (Fig. 4c), until it reaches an asymptotic value and the entropy per symbol
366
+ saturates. The physical origin of the increase in the entropy per symbol is due to an increases of the
367
+ microscopic configurations of the system probed by the challenge CiProbe, that translates in an increase of
368
+ the variety of speckle patterns that form the codomain of the HPUF. The growing rate of the entropy is
369
+ reduced up to a saturation level when the compartment is populated by roughly 8-10 levels. For a given
370
+ size of the challenge CProbe and LTrans, the entropy per symbol saturates when light probes all the possible
371
+ accessible configurations of the system.
372
+ Figure 4. Hyper PUF characterization. a) Scheme of the domain (CiProbe, LkTrans) and the codomain Rik of the HPUF.
373
+ The whole codomain can be compartmentalized by joining the responses of different levels. b) Entropy per symbol
374
+ for different compartments of the HPUF codomain. The left panel (blue circles) shows the entropy per symbol for
375
+ each single level (the horizontal labels show the Hadamard basis configuration patterns). The right panel shows the
376
+ increase of the entropy per symbol by randomly joining the codomains of individual levels (black circles). Red circles
377
+ refer to the same analysis performed by populating the compartment by joining the responses of different PUFs. The
378
+ blue error bars refer to 10 different characterizations of each single level. The black and red error bars refer to the
379
+ standard deviation calculated over ten different random selections of the PUF or levels of HPUF, respectively. c)
380
+ Autocorrelation of the bit sequences generated by equivalent Markov chain. The correlation length decreases as the
381
+ number of levels increases, because the permanence probability of the equivalent key increases from α = 0.07 to α =
382
+ 0.12. d) Representation of the binary keys generated by the equivalent Markov chains for one level (α = 0.07) and ten
383
+ levels (α = 0.12).
384
+ Conclusions
385
+ We developed new optical cryptographic primitives, named Hyper PUFs or HPUFs, that allow multi-level
386
+ operation thanks to fully reversible switching of their optical properties. The all-optical HPUF of this paper
387
+ is realized with polymer dispersed and stabilized liquid crystals, and the transformation of the levels is
388
+ enabled by a light pattern that can selectively and locally drive the reversible phase transition of the
389
+ embedded liquid crystals. The entropy of the HPUF’s keys has been studied using different methods, both
390
+ confirming its increase with the number of joint levels. These results show that the HPUF is equivalent to
391
+ combining several physical unclonable functions into a single hardware. The overall entropy per bit is
392
+
393
+ a
394
+ b
395
+ 0.35
396
+ o1LevelPUF
397
+ HyperPUF
398
+ Diff.PUFs
399
+ 0.3
400
+ Trans
401
+ R,10
402
+ 0.25
403
+ HPUF
404
+ 0.2
405
+ 1
406
+ Ri,k
407
+ 10 levels
408
+ +++
409
+ klevels
410
+ 1 level
411
+ C
412
+ Trans
413
+ R,1
414
+ 0.1
415
+ domain
416
+ codomain
417
+ 8
418
+ 2
419
+ 4
420
+ 6
421
+ 8
422
+ 10
423
+ Numberof Levels/Numberof PUFs
424
+ d
425
+ α=0.07
426
+ α=0.12
427
+ C
428
+ 1Lev:Q=0.07
429
+ 3Lev:α=0.95
430
+ 0.8
431
+ 10 Lev: α=0.12
432
+ 20
433
+ 20
434
+ 40
435
+ 40
436
+ Bits
437
+ Bits
438
+ 0.4
439
+ 60
440
+ α / # Levels
441
+ 60
442
+ 0.2
443
+ 80
444
+ 80
445
+ ..
446
+ 0
447
+ :
448
+ 100
449
+ 0
450
+ 5
451
+ 10
452
+ 15
453
+ 20
454
+ 20
455
+ 40
456
+ 60
457
+ 80
458
+ 100
459
+ 20
460
+ 40
461
+ 60
462
+ 80
463
+ 100
464
+ Lag (bits)
465
+ Eg.MCkeys
466
+ Eg.MC keys
467
+ 8
468
+
469
+ affected by the unavoidable presence of spatial correlations between the microscopic configurations and
470
+ reaches a saturation levels when the challenge light probes all the microscopic configurations of the system.
471
+ We believe that the concept described in this paper allows for the development of a new generation of all-
472
+ optical security devices. Amongst the advantages is the unique possibility to create multi-level PUFs
473
+ integrated into one and the same material, thus enabling a practical implementation of quantum secure
474
+ authentication of data10. This not only opens up to novel quantum protocols via strong optical PUF but
475
+ significantly increases their security level and also allows to have multi-user key generators and hence
476
+ multiple clients on one device. 22,33
477
+
478
+ Acknowledgements
479
+ The research leading to these results has received funding from Ente Cassa di Risparmio di Firenze
480
+ (2018/1047), AFOSR/RTA2 (A.2.e. Information Assurance and Cybersecurity) project “Highly Secure
481
+ Nonlinear Optical PUFs” (Award No. FA9550-21-1-0039) and Fondo premiale FOE to the project “Volume
482
+ photography:
483
+ measuring
484
+ three
485
+ dimensional
486
+ light
487
+ distributions
488
+ without
489
+ opening
490
+ the
491
+ box”
492
+ (E17G17000300001).
493
+
494
+ Bibliography
495
+ 1.
496
+ McGrath, T., Bagci, I. E., Wang, Z. M., Roedig, U. & Young, R. J. A PUF taxonomy. Appl Phys
497
+ Rev 6, 011303 (2019).
498
+ 2.
499
+ Gao, Y., Al-sarawi, S. F. & Abbott, D. Physical unclonable functions. Nature Electr. 3, 81-91
500
+ (2020).
501
+ 3.
502
+ Rührmair, U. & Holcomb, D. E. PUFs at a glance. Proceedings -Design, Automation and Test in
503
+ Europe, DATE, 1-6 (2014)
504
+ 4.
505
+ Gassend, B., Clarke, D., van Dijk, M. & Devadas, S. Silicon Physical Random Functions
506
+ Proceedings of the 9th ACM Conference on Computer and Communications Security 148-160
507
+ (2002).
508
+ 5.
509
+ Herder, C., Yu, M. D., Koushanfar, F. & Devadas, S. Physical unclonable functions and
510
+ applications: A tutorial. Proceedings of the IEEE vol. 102, 1126–1141 (2014).
511
+ 6.
512
+ Pappu, R., Recht, B., Taylor, J., & Gershenfeld, N. Physical one-way functions. Science, 297,
513
+ 2026-2030 (2002).
514
+ 7.
515
+ Pavanello, F., O’Connor, I., Ruhrmair, U., Foster, A. C. & Syvridis, D. Recent Advances in
516
+ Photonic Physical Unclonable Functions. Proceedings of the European Test Workshop, 1-
517
+ 10(2021).
518
+ 8.
519
+ Cambou, B. et al. Post quantum cryptographic keys generated with physical unclonable functions.
520
+ Applied Sciences 11, 2801 (2021).
521
+ 9.
522
+ Horstmeyer, R., Assawaworrarit, S., Ruhrmair, U. & Yang, C. Physically secure and fully
523
+ reconfigurable data storage using optical scattering. Proceedings of the 2015 IEEE International
524
+ Symposium on Hardware-Oriented Security and Trust, HOST 2015 157–162 (2015).
525
+ 10.
526
+ Škorić, B., Pinkse, P. W. H. & Mosk, A. P. Authenticated communication from quantum readout
527
+ of PUFs. Quantum Inf Process 16, 1-9 (2017).
528
+ 11.
529
+ Vynck, K., Burresi, M., Riboli, F. & Wiersma, D. S. Photon management in two-dimensional
530
+ disordered media. Nat Mater 11, 7–12 (2012).
531
+
532
+
533
+ 9
534
+
535
+ 12.
536
+ Ke, Y. et al. Smart Windows : Electro- , Thermo- , Mechano- , Photochromics , and Beyond.
537
+ Advanced Energy Materials 9, 1902066 (2019).
538
+ 13.
539
+ Kim, H. & Yang, S. Responsive Smart Windows from Nanoparticle – Polymer Composites. Adv
540
+ Funct Mater, 30, 1902597 (2020).
541
+ 14.
542
+ Redding, B., Liew, S. F., Sarma, R. & Cao, H. Compact spectrometer based on a disordered
543
+ photonic chip. Nat Photonics 7, 746-751 (2013).
544
+ 15.
545
+ Boschetti, A. et al. Spectral super-resolution spectroscopy using a random laser. Nat Photonics 14,
546
+ 177–182 (2020).
547
+ 16.
548
+ Wiersma, D. S. Disordered photonics. Nat Photonics 7, 188–196 (2013).
549
+ 17.
550
+ Yu, S., Qiu, C. W., Chong, Y., Torquato, S. & Park, N. Engineered disorder in photonics. Nat Rev
551
+ Mater 6, 226–243 (2021).
552
+ 18.
553
+ Mosk, A. P., Lagendijk, A., Lerosey, G. & Fink, M. Controlling waves in space and time for
554
+ imaging and focusing in complex media. Nature Photonics 6, 283–292 (2012).
555
+ 19.
556
+ Berkovits, R. Sensitivity of the multiple-scattering speckle pattern to the motion of a single
557
+ scatterer. Physical Review B 43, 8638 (1991).
558
+ 20.
559
+ Riboli, F. et al. Tailoring Correlations of the Local Density of States in Disordered Photonic
560
+ Materials. Phys Rev Lett 119, 1–6 (2017).
561
+ 21.
562
+ Csaba, G. et al. Application of mismatched cellular nonlinear networks for physical cryptography.
563
+ 12th International Workshop on Cellular Nanoscale Networks and their Applications, CNNA 2010
564
+ (2010).
565
+ 22.
566
+ Lugli, P. et al. Physical unclonable functions based on crossbar arrays for cryptographic
567
+ applications. International Journal of Circuit Theory and Applications 41, 619–633 (2013).
568
+ 23.
569
+ Rührmair, U. et al. Modeling attacks on physical unclonable functions. Proceedings of the ACM
570
+ Conference on Computer and Communications Security, 237–249 (2010)
571
+ 24.
572
+ Delvaux, J., & Verbauwhede, I. Side channel modeling attacks on 65nm arbiter PUFs exploiting
573
+ CMOS device noise. IEEE International Symposium on Hardware-Oriented Security and Trust
574
+ (HOST), 137-142 (2013)
575
+ 25.
576
+ Horstmeyer, R., Assawaworrarit, S., Ruhrmair, U., & Yang, C., Physically secure and fully
577
+ reconfigurable data storage using optical scattering. IEEE International Symposium on Hardware
578
+ Oriented Security and Trust (HOST), 157-162, (2015).
579
+ 26.
580
+ Horstmeyer, R., Judkewitz, B., Vellekoop, I. M., Assawaworrarit, S. & Yang, C. Physical key-
581
+ protected one-time pad. Sci Rep 3, 1-6 (2013).
582
+ 27.
583
+ Grubel, B. C. et al. Secure communications using nonlinear silicon photonic keys. Opt Express 26,
584
+ 4710 (2018).
585
+ 28.
586
+ Bosworth, B. T. et al. Unclonable photonic keys hardened against machine learning attacks. APL
587
+ Photonics 5, 010803 (2020).
588
+ 29.
589
+ Kursawe, K., Sadeghi, A. R., Schellekens, D., Škorić, B. & Tuyls, P. Reconfigurable physical
590
+ unclonable functions - Enabling technology for tamper-resistant storage. IEEE International
591
+ Workshop on Hardware-Oriented Security and Trust, HOST 2009, 22–29 (2009).
592
+ 30.
593
+ John, R. A. et al. Halide perovskite memristors as flexible and reconfigurable physical unclonable
594
+ functions. Nat Commun 12, 1-11 (2021).
595
+ 31.
596
+ Gan, Z. et al. Reconfigurable Optical Physical Unclonable Functions Enabled by VO2Nanocrystal
597
+ Films. ACS Appl Mater Interfaces 14, 5785–5796 (2022).
598
+
599
+
600
+ 10
601
+
602
+ 32.
603
+ Ruiz De Galarreta, C. et al. Reconfigurable multilevel control of hybrid all-dielectric phase-
604
+ change metasurfaces. Optica 7, 476–484 (2020).
605
+ 33.
606
+ Burr, W. E. et al. Electronic Authentication Guideline - NIST Special Publication 800-63-2. 1–123
607
+ (2013).
608
+ 34.
609
+ Daugman, J. Information theory and the iriscode. IEEE Transactions on Information Forensics
610
+ and Security 11, 400–409 (2016).
611
+ 35.
612
+ Daugman, J. The importance of being random: statistical principles of iris recognition. Pattern
613
+ Recognit 36, 279–291 (2003).
614
+ 36.
615
+ Guo, S. M. et al. Preparation of a thermally light-transmittance-controllable film from a coexistent
616
+ system of polymer-dispersed and polymer-stabilized liquid crystals. ACS Appl Mater Interfaces 9,
617
+ 2942–2947 (2017).
618
+ 37.
619
+ Dierking, I. Polymer network-stabilized liquid crystals. Advanced Materials 12, 167–181 (2000).
620
+ 38.
621
+ Lee, Y.-H., Gou, F., Peng, F. & Wu, S.-T. Hysteresis-free and submillisecond-response polymer
622
+ network liquid crystal. Opt Express 24, 14793 (2016).
623
+ 39.
624
+ Ondris-Crawford, R. et al. Microscope textures of nematic droplets in polymer dispersed liquid
625
+ crystals. J Appl Phys 69, 6380–6386 (1991).
626
+ 40.
627
+ Rührmair, U., Hilgers, C. & Urban, S. Optical PUFs Reloaded. IACR Cryptology (2013).
628
+ 41.
629
+ Uppu, R. et al. Asymmetric cryptography with physical unclonable keys. Quantum Sci Technol 4,
630
+ 1–20 (2019).
631
+
632
+
633
+
634
+ 1
635
+
636
+
637
+
638
+ 2
639
+
640
+
641
+
4NA0T4oBgHgl3EQfNf9p/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
5NFAT4oBgHgl3EQfFRzt/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e24f29bf1537f992add07c4533360c5e309c723d3e4a61ea6fe4643967c9c316
3
+ size 917549
5NFAT4oBgHgl3EQfFRzt/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f846f9316171cd5315265d3217c3fbade2fdcfd0af86648bfd2bc788ea74902e
3
+ size 38558
6tE1T4oBgHgl3EQfTQPr/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:81dde1d0554af5c8238690a09e3e049dce504de7c64ff91ac0c725c251dcde59
3
+ size 3145773
7dAzT4oBgHgl3EQf-f5K/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb441282f8626fc7ba6cdca782d7841d5971807a977c96026b6d3efa85c681e7
3
+ size 4587565
7dE1T4oBgHgl3EQf7QV5/content/tmp_files/2301.03532v1.pdf.txt ADDED
@@ -0,0 +1,1137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ EFFICIENT ATTACK DETECTION IN IOT DEVICES
2
+ USING FEATURE ENGINEERING-LESS MACHINE
3
+ LEARNING
4
+ ARSHIYA KHAN1 AND CHASE COTTON2
5
+ 1University of Delaware, Newark, USA
6
+ arshiyak@udel.edu
7
+ 2University of Delaware, Newark, USA
8
+ ccotton@udel.edu
9
+ ABSTRACT
10
+ Through the generalization of deep learning, the research community has addressed critical challenges in
11
+ the network security domain, like malware identification and anomaly detection. However, they have yet to
12
+ discuss deploying them on Internet of Things (IoT) devices for day-to-day operations. IoT devices are often
13
+ limited in memory and processing power, rendering the compute-intensive deep learning environment
14
+ unusable. This research proposes a way to overcome this barrier by bypassing feature engineering in the
15
+ deep learning pipeline and using raw packet data as input. We introduce a feature engineering-less
16
+ machine learning (ML) process to perform malware detection on IoT devices. Our proposed model,”
17
+ Feature engineering-less ML (FEL-ML),” is a lighter-weight detection algorithm that expends no extra
18
+ computations on “engineered” features. It effectively accelerates the low-powered IoT edge. It is trained
19
+ on unprocessed byte-streams of packets. Aside from providing better results, it is quicker than traditional
20
+ feature-based methods. FEL-ML facilitates resource-sensitive network traffic security with the added
21
+ benefit of eliminating the significant investment by subject matter experts in feature engineering.
22
+ KEYWORDS
23
+ Feature engineering-less, AI-enabled security, 1D-CNN, Internet-of-Things, Botnet Attack
24
+ 1. INTRODUCTION
25
+ Cyber Security experts have found pivotal features in network traffic, including packet captures
26
+ (pcap). Data scientists have used them to fashion impressive models capable of differentiating
27
+ malicious traffic from benign [1]. However, most network traffic is emitted over encrypted
28
+ channels in the current scheme. This security measure has limited experts’ ability to contrive
29
+ meaningful features for machine learning (ML), which can soon become obsolete. This challenge
30
+ has given birth to analyzing raw bytes to detect malicious behavior in internet flows.
31
+ In the Internet of Things (IoT) domain, devices are sensors that interact with the environment.
32
+ They conditionally react to changes in the environment and exchange information over the
33
+ internet about these changes. A typical example of an IoT device is a doorbell camera. We can
34
+ check any activity at our door using the camera from anywhere on earth. It can also alert us of
35
+ any break-ins. For an IoT device to operate, it must continuously communicate with its users,
36
+ other devices, and servers without fail. This communication is done over the internet, where
37
+ information travels in the form of packets. This flow of packets is commonly known as network
38
+ traffic. For an IoT grid to function, the network traffic must be secure from cyber criminals.
39
+ Intrusion Prevention Systems (IPS) (like Cisco Firewall and McAfee), and Intrusion Detection
40
+ Systems (IDS) (like SolarWinds and Snort), can scan the network traffic and determine if they
41
+ are secure or insecure. Historically, packets are validated using a pre-defined rule book. However,
42
+
43
+ since the introduction of ML, many attempts have been made to detect insecure packets using
44
+ ML. Both detection systems display substantial hardware constraints like processing power and
45
+ sizeable memory requirements to store moving traffic and identify anomalous packets. The
46
+ evolving world of smart sensors and IoT devices demands precision and speed simultaneously.
47
+ Researchers have used machine learning to detect traffic anomalies in several studies [2,3]. They
48
+ have followed the standard pipeline of typical ML modeling. By extracting features from raw
49
+ data, they have trained their models. In the network traffic specialization, they have manipulated
50
+ values in the packets to create features. This conventional approach is not only ineffective over
51
+ encrypted channels but is also task-sensitive. Attack scenarios on network traffic have evolved
52
+ rapidly. Now, attackers can hit multiple levels of network architecture [4], rendering the task-
53
+ sensitive approach defenseless. In 2020, a group of researchers published a survey [5] on the use
54
+ of ML to enable security in IoT devices. They divided the survey into five sections dedicated to
55
+ understanding the enormity of security challenges. The survey discussed current ML solutions
56
+ like anomaly detection, attack evasion, and mitigation. However, it is only possible to use these
57
+ methods if we already know the type of attack being launched. The survey also discussed the
58
+ complexity of IoT networks and multi-level attacks [6] suffered by these devices. They can cause
59
+ the IoT infrastructure to fail anywhere from the application layer to the physical layer. Such a
60
+ complicated structure makes it impossible to design a defense model that can guard against all
61
+ possible attack scenarios.
62
+ Therefore, it is imperative to leverage the generalized scope of ML, which can learn even minor
63
+ irregularities from the dataset. Traditional attack-specific feature-based detection models can crop
64
+ or manipulate the dataset in an unhelpful way.
65
+ The remaining paper is organized as follows. Section 2 is dedicated to a detailed review of
66
+ contemporary work. We have discussed the use of deep learning models for network traffic in
67
+ section 2.1 and their feasibility on IoT devices in Section 2.2. Section 3 goes into detail about our
68
+ proposed methodology. It is divided into three parts. 3.1 discusses our proposed approach in
69
+ detail. 3.2 talks about the dataset and preparation of the experiment, while 3.3 talks about our
70
+ deep learning architecture. Section 4 explains the experiment results and performance
71
+ comparison. Section 5 covers concluding remarks and future work.
72
+ 2. BACKGROUND AND RELATED WORK
73
+ 2.1. The advent of 1D-CNN in Malware Detection
74
+ The last decade has seen exemplary use of machine learning to classify network traffic. Studies
75
+ based on these features usually use three approaches: 1) Port-based, 2) Deep packet inspection
76
+ (DPI)-based, and 3) Behavior-based [7].
77
+ A port-based classifier classifies reliable ports as benign and unreliable ports malicious. However,
78
+ this classification technique has been rendered unusable by employing port hiding techniques
79
+ such as port camouflaging. Port randomization is also used to hide ports.
80
+ The port-based approach uses only headers to classify traffic; however, the DPI-based approach
81
+ examines both header and the payload. The header is inspected to collect information about the
82
+ sender application, and the payload signature is examined to ensure they are not invalid or
83
+ blacklisted. Due to high computational costs, DPI-based is an inefficient method. Lastly, in the
84
+ behavior-based approach, flow or session trends are observed. Features to train an ML model are
85
+ fashioned from the statistical inferences of these trends.
86
+ Velan et al. [7] published a comprehensive survey of 26 papers in 2015. It investigated studies
87
+ focused on encrypted traffic classification published between the period 2005 to 2014. Flow
88
+ features were used in 12 of the 26 papers, packet features in 5, a combination of packet and flow
89
+ features in 7, and other features in two.
90
+
91
+ Wang et al. [8] published another survey in 2019, which addressed the rise in raw packet content
92
+ to train deep learning models. As more parts of the packets were encrypted, extracting features
93
+ became difficult. Both headers and payload were concealed, forcing researchers to use raw data
94
+ in the training set. From 2015 to 2018, 8 out of 12 publications used raw data from the packet for
95
+ encrypted traffic classification, and two used both derivative features and raw data. One of the
96
+ remaining two papers solely used packet-based features, while the other used flow-based features.
97
+ Wei Wang et al. [9] used raw bytes collected in groups of flows and sessions. This study employed
98
+ the Representation Learning technique and converted raw bytes into images to create the training
99
+ dataset. It is called USTC-TFC2016 [9], a private dataset pulled from the Stratosphere IPS Project
100
+ Malware Dataset [10]. It used a two-dimensional convolutional neural network (2D-CNN) to
101
+ classify the images. This experiment saved a small amount of computational power as it did not
102
+ require feature extraction. However, it spent a lot on byte-to-image conversion and image training.
103
+ The binary classifier (malware or benign) was 100% accurate, the 10-label classifier was 99.23%
104
+ accurate, and the 20-label classifier was 99.17% accurate. The classifiers achieved an average
105
+ accuracy of 99.41%. For the 10-label classifier, the lowest and the highest precision score were
106
+ 90.7% and 100%, respectively. Similarly, the highest and the lowest recall score were 91.1% and
107
+ 100%, respectively, for the 10-label classifier.
108
+ Wang et al. [11] conducted a similar study in 2017, where they used the ISCX VPN-non-VPN
109
+ [12] dataset to perform encrypted traffic classification. This dataset contains six traffic types:
110
+ chats, emails, file transfer, peer-to-peer (P2P), streaming, and Voice over Internet Protocol (VoIP)
111
+ captured over both Virtual Private Network (VPN) and non-VPN settings. This study performed
112
+ binary classification on VPN and non-VPN traffic and multi-label classification on the six traffic
113
+ types. It used raw bytes of flows and sessions of the ISCX dataset. To train on raw bytes, it used
114
+ a one-dimensional convolutional neural network (1D-CNN). The 2-label (binary) classification
115
+ achieved 100% precision, and the 6-label classification achieved 85.5% accuracy. Six traffic types
116
+ were branched into 12 classes using VPN as a factor (VPN and non-VPN traffic). The 12-label
117
+ classification achieved 85.8% accuracy. This study compared their results with [9], which used
118
+ 2D-CNN. In a series of 4 experiments, 1D-CNN outperformed 2D-CNN by as much as 2.51%.
119
+ DeepPacket [13], introduced in 2017, is a deep learning framework for network traffic
120
+ characterization and application identification. It employed two deep learning techniques: a)
121
+ stacked autoencoder (SAE) neural network and b) 1D-CNN. Both were trained on ISCX VPN-
122
+ non-VPN [12] dataset. The traffic characterization task classified VPN and non-VPN traffic. With
123
+ SAE, the binary classifier's average precision and recall were 92%. However, with 1D-CNN, the
124
+ same classification resulted in average precision of 94% and an average recall of 93%. The
125
+ application identification task classified packets into these six applications: chats, emails, file
126
+ transfer, peer-to-peer (P2P), streaming, and Voice over Internet Protocol (VoIP). With the SAE
127
+ classifier, the average precision of the 6-label classification was 96%, and the average recall was
128
+ 95%. With the 1D-CNN classifier, the average precision and recall were 98%.
129
+ In 2020, Rezaei and Liu [14] employed multi-task learning to classify network traffic. It divided
130
+ the classification task into two tasks: a) bandwidth requirement and b) predicting the duration of
131
+ traffic flow. It used two datasets for training: ISCX VPN-non-VPN and QUIC [15]. The multi-
132
+ task learning model was trained on a 1D-CNN using these three time-series features: i) packet
133
+ length, ii) inter-arrival time, and iii) direction of the traffic. On the ISCX dataset [12], their model
134
+ classified traffic with 80.67% accuracy. The same experiment classified bandwidth requirement
135
+ and prediction of flow duration with 88.67% and 90% accuracies, respectively. On the QUIC
136
+ dataset [15], the model classified traffic with 94.67% accuracy. In this experiment, bandwidth
137
+ requirement and flow duration prediction attained 90.67% and 91.33% accuracies, respectively.
138
+
139
+ Table 1. Related Works of Network Traffic ML models
140
+ Work
141
+ DL
142
+ Technique
143
+ Model
144
+ Input
145
+ Dataset
146
+ ML Task
147
+ Accuracy
148
+ Year
149
+ [9]
150
+ 2D-CNN
151
+ images
152
+ [9]
153
+ Malware detection
154
+ 99.23%
155
+ 2017
156
+ [11]
157
+ 1D-CNN
158
+ bytes
159
+ [12]
160
+ Traffic
161
+ characterization
162
+ 100%
163
+ 2017
164
+ [13]
165
+ 1D-CNN +
166
+ SAE
167
+ bytes
168
+ [12]
169
+ Traffic
170
+ characterization
171
+ 98%
172
+ 2017
173
+ [14]
174
+ 1D-CNN
175
+ time series
176
+ [12, 15]
177
+ Traffic
178
+ characterization
179
+ 94.67%
180
+ 2020
181
+ [17]
182
+ 1D-CNN +
183
+ LSTM
184
+ raw bytes
185
+ [9]
186
+ Malware detection
187
+ 98.6%
188
+ 2020
189
+
190
+ Huang et al. [16] published a survey in 2019. It was based on deep learning use cases in the time-
191
+ series domain. Cyber Security was one of the emergent real-world disciplines in the time-series
192
+ domain, along with health, finance, and transportation. To achieve this conclusion, the paper
193
+ analyzed topics such as traffic classification, anomaly detection, and malware identification.
194
+ Marin, Casas, and Capdehourat [17] published a study in 2020 aiming to remove engineered
195
+ features, thus eliminating the need for domain experts. They reintroduced two types of malware
196
+ detection approaches using raw packet content: raw packets and raw flows. They used raw byte-
197
+ streams from pcaps to train their deep learning model to achieve this goal. Their ML model was
198
+ a combination of a Long short-term memory (LSTM) network and a 1D-CNN. The raw byte-
199
+ stream of flow performed (98.6% accuracy) better than the byte-stream of a packet (77.6%
200
+ accuracy). A comparative experiment was performed between traditional feature-based and raw
201
+ byte-based models. For the traditional model, they used a random forest (RF) and trained it on
202
+ 200 in-flow features. Both raw byte-based models performed better than RF.
203
+ Discussion of the previous works in this section has revealed two significant points:
204
+ 1. They indicate the advent of 1D-CNN in the network traffic characterization and malware
205
+ detection domain. It is also evident in Table 1, which displays several publications of our study.
206
+ The table is arranged chronologically and shows a clear trend of ML-based network traffic
207
+ classifiers preferring 1D-CNN over other techniques. In addition to its superior performance, 1D-
208
+ CNN has the advantage of preserving the time-series nature of network traffic. Inspired by its
209
+ effectiveness, we have used it as the modeling baseline of our study.
210
+ 2. Studies also suggest that in recent years, malware detection models have moved from
211
+ engineered feature ML to non-engineered feature ML using raw data. However, these models may
212
+ only be applicable to some network traffic use cases. It is practically impossible to employ
213
+ complex models on memory-constrained devices like IoT. More extensive models like LSTMs
214
+ and 2D-CNNs need several preprocessing steps to extract raw data and train a model on them. In
215
+ the next section, we will discuss the use of ML practices in the IoT environment.
216
+ 2.2. ML Techniques for Malware Detection in IoT environment
217
+ Traditional appliances, devices, and machines have moved to smart sensor technologies in the
218
+ last few decades. In addition to routers and IPSs, these technologies are also a component of the
219
+ IoT. Uninterrupted internet access makes them susceptible to malicious attacks, which may result
220
+ in malfunction or failure. Several studies have tried to find security solutions for these IoT
221
+ devices.
222
+
223
+ In 2019, Shouran, Ashari, and Priyambodo [18] introduced a straightforward way of detecting
224
+ threats in IoT devices. It classified every device interaction with the internet into Low, Medium,
225
+ and High impact. The classification criteria included compromise in Confidentiality, Integrity,
226
+ and Authenticity (CIA).
227
+ Another research in 2019 [19] introduced an elaborate IDS for IoT devices. However, the IDS
228
+ was feature-based with three layers of ML algorithms.
229
+ Later in 2020, a paper published by Vinayakumar et al. [20] developed a two-level deep learning
230
+ model which discriminates malicious traffic from benign. The first level detected the most
231
+ frequent DNS queries, and the second level used a domain generation algorithm (DGA) to detect
232
+ illegal domains.
233
+ Sriram et al. [21] presented their deep learning ML system, which used network flows to find
234
+ statistical features. Two datasets were used and compared over different ML modeling techniques
235
+ like logistic regression, random forest, and LSTM. [17] also presented a 1D-CNN model for
236
+ botnet detection on IoT devices. Details of their work are discussed in section 2.1.
237
+ Table 2. Related Work in IoT domain
238
+ Work
239
+ IoT Malware detection Technique
240
+ Task
241
+ Year
242
+ [18]
243
+ Non-ML (Rule-based)
244
+ Malware detection
245
+ 2019
246
+ [19]
247
+ ML-based IDS
248
+ Traffic characterization
249
+ 2019
250
+ [20]
251
+ ML based DNS categorization
252
+ Malware detection
253
+ 2020
254
+ [21]
255
+ Ensemble of ML models
256
+ Botnet detection
257
+ 2020
258
+ [17]
259
+ 1D-CNN
260
+ Malware detection
261
+ 2020
262
+
263
+ Table 2 shows a task-based analysis of research works published on detecting malware on IoT
264
+ devices. Most of the methods presented here require multiple steps to perform one task. This
265
+ approach is not suitable for the IoT environment.
266
+ Our contribution is as follows:
267
+ 1. Outright elimination of ‘engineered features’ that require additional computation. We introduce
268
+ a network traffic classification system in favor of more lightweight features which come directly
269
+ from the input data. As a result, it does not require domain-based expertise to perform feature
270
+ engineering.
271
+ 2. Increase the speed of classification by using 1D-CNN to train the deep learning model. The
272
+ model will consume less memory during both the training and testing phases allowing it to be
273
+ deployed on IoT devices.
274
+ 3. METHODOLOGY
275
+ 3.1. Feature engineering-less ML
276
+ IoT devices at the edge, like voice-based virtual assistants (e.g., Amazon Echo), smart appliances,
277
+ and routers, must react to change at a very high speed. Consequently, they have to validate
278
+ incoming traffic in real-time. They also have a limited battery life to support these unremitting
279
+ transactions. In an internet-dependent environment like this, security from cyber attacks is non-
280
+ negotiable; nevertheless, it can become an overhead. It can be in various forms, like malware
281
+ recognition, anomaly detection, or behavior classification. Detecting cyber attacks using ML has
282
+ shown promising results in the past [22, 23]. However, their deployment on IoT devices is
283
+ unrealistic, as they involve computationally extensive feature engineering and require deep
284
+
285
+ classifiers for precision. Feature engineering is a step-by-step process that incorporates: i)
286
+ extracting desired elements from raw data, ii) cleaning and converting them into features, iii)
287
+ standardizing features, and iv) aggregating them to be used in the classifier. For a time sensitive
288
+ IoT environment, it is a complex and time-consuming operation. We cannot rely on traditional
289
+ methods to make ML adequate on IoT devices. In this study, we propose skipping feature
290
+ engineering and developing an IoT-friendly deep learning technique called Feature Engineering-
291
+ less ML.
292
+ Feature engineering-less ML or FEL-ML is a product of the featureless modeling technique.
293
+ Feature engineering-less modeling is machine learning without feature engineering. It is a lighter-
294
+ weight detection algorithm where no extra computations are expended to compute ‘engineered
295
+ features,’ resulting in an adequate acceleration to the low-powered IoT. We eliminate the feature
296
+ extraction and processing step from the ML pipeline in FEL-ML. We store the streams of packets
297
+ that arrive at an IoT device in their raw state. We create our training dataset by converting the raw
298
+ streams to raw byte streams. The rest of the deep learning process remains the same. The
299
+ advantages of FEL-ML are that it conserves the properties of a stream of bytes and saves time
300
+ during the process. Omitting the feature extraction and generation step makes both model training
301
+ and testing efficient.
302
+ Feature engineering-less modeling eliminates two significant IoT overheads: computation cost
303
+ and human cost. Human cost involves using technical expertise to collect and clean traffic data.
304
+ It also employs domain expertise to manipulate them into meaningful representatives of the
305
+ dataset. Computation cost incorporates the computational overhead a device endures toward
306
+ statistical operations at a device’s central processing unit (CPU). Further down the ML pipeline,
307
+ tensor-based computations like convolution, batch processing, and running multiple epochs
308
+ contribute to computational overhead.
309
+ We have represented network traffic in three views: 1) Raw session, 2) Raw flow, and 3) Raw
310
+ packet. Based on these three representations, we have developed three unique ML models. In the
311
+ end, we will compare their performance to determine the most suitable model.
312
+ Raw session: In an IoT network, a session is a bi-directional stream of communication be- tween
313
+ two devices. All packets in a session share these 5-tuple attributes: a) source IP address, b) source
314
+ port, c) destination IP address, d) destination port, and e) transport protocol present in each packet
315
+ header. In this representation, we split the pcap files into unique sessions based on the 5-tuple.
316
+ We used MIT’s pcapsplitter tool [24]. A pcap file records the live traffic stream on a device. Each
317
+ individual session stream forms a unique component in the ground truth.
318
+ Raw flow: A flow is similar to a session, except that they are unidirectional. A flow is a batch of
319
+ packets going from one device to another. We used SplitCap [25] to prepare our dataset in this
320
+ representation. SplitCap splits pcap files into flows. A sample in the ground truth is a byte stream
321
+ of a flow.
322
+ Raw packet: A pcap file consists of one or more packets. In this representation, we used the byte
323
+ stream of each packet. A single packet in byte format is an individual ground truth component.
324
+ In a similar study [17], raw bytes were used to train the classifier. However, their experiment only
325
+ included packets and sessions. Further, they have not discussed any memory or system metrics to
326
+ demonstrate its efficiency after eliminating feature engineering. Their model constituted 1D-CNN
327
+ and LSTM layers, i.e., a deep model. Deeper models are more complex and can result in
328
+ overfitting. As a result, we have used a smaller but effective neural network to overcome this
329
+ obstacle.
330
+ 3.2. Experimental Design
331
+ We have used the Aposemat IoT-23 dataset [26] to perform experiments for this study. It is a
332
+ labeled dataset captured from 2018 to 2019 on IoT devices and published in 2020 by the
333
+
334
+ Stratosphere Research Lab. Several studies have either discussed it in their work or used it in their
335
+ model. Bobrovnikova et al. [27] used this dataset to perform botnet detection in 2020. They
336
+ extracted features to curate numerical features for classification. They attained 98% accuracy with
337
+ support vector machines (SVM) [28]. Blaise et al. [29] have mentioned that this dataset is similar
338
+ to what they required but not pertinent to their experiments. This dataset also contained
339
+ conn.log.labeled file generated by employing Zeek on this dataset. Stoian [30] used numerical
340
+ and categorical features from this file. They trained a Random Forest classifier [31] with this
341
+ dataset and attained 100% accuracy.
342
+ IoT-23 has 23 sets of scenarios, out of which three are benign and 20 are malicious. The benign
343
+ dataset is collected from three IoT devices: i) Amazon Echo, ii) Somfy smart door lock, and iii)
344
+ Philips hue LED lamp. These devices are then infected with an assortment of botnet attacks
345
+ executed in a simulation environment forming the malicious dataset.
346
+
347
+ Figure 1. Byte Distribution
348
+ IoT-23 contains raw pcap files and Zeek files of each scenario. For our experiments, we used the
349
+ pcap files as our training dataset. We integrated the three benign scenarios into one extensive
350
+ dataset. The malicious scenarios are prepared in a simulation environment where the three devices
351
+ are infected by 20 unique botnet attacks. Pcap files of several of these botnets range between
352
+ hundreds of GBs. To demonstrate the fitness of our proposed model on limited-memory devices,
353
+ we used attack scenarios with smaller sizes. We have selected five out of 20 attacks: i) Hide and
354
+ Seek, ii) Muhstik, and iii) Linux.Hajime, iv) Okiru, and v) Mirai. Byte distribution of the five
355
+ infected and one benign traffic data is shown in Figure 1. Using binary classification, we
356
+ distinguished between benign and malicious traffic, while multi-label classification distinguished
357
+ between botnet traffic.
358
+ As mentioned in section 3.1, we represented this dataset in three unique views: session, flow, and
359
+ packet. In order to use raw bytes, we converted all traffic representations into a hexadecimal
360
+ format using tshark [32]. As we take a deep dive into this experiment, we named these three
361
+ experiments as follows:
362
+ ExpS: The experiment used the session representation of the dataset in the hexadecimal format.
363
+ In the training data, each session is represented as one row of a byte stream.
364
+ ExpF: The experiment used the flow representation of the dataset in the hexadecimal format. In
365
+ the training data, each flow is represented as a byte stream.
366
+ ExpP: The experiment used the packet representation of the dataset in the hexadecimal format.
367
+ Each pcap instance is a sample in the labeled training set.
368
+
369
+ 450
370
+ 400
371
+ 350
372
+ 300
373
+ ytes
374
+ 250
375
+ B
376
+ 200
377
+ 150
378
+ 100
379
+ 50
380
+ 0
381
+ Hide and
382
+ Muhstik
383
+ Linux.Hajime
384
+ Okiru
385
+ Mirai
386
+ Benign
387
+ Seek
388
+ Packet
389
+ Session
390
+ Flow
391
+ Figure 2. Training data generation
392
+ Studies [17] conducted in the past on raw byte streams have an additional step to remove ethernet
393
+ and TCP/IP headers from their dataset. We experimented with the use of headers to achieve our
394
+ goal of diminishing computational overhead further. We reformulated the dataset into four unique
395
+ categories for each ExpS, ExpF, and ExpP. The first category included all headers in the packet
396
+ along with the payload. We called this category: All headers. In the next category, we kept the
397
+ ethernet headers and discarded the IP headers; hence we called this category Ethernet headers
398
+ only. Intuitively, in the third category, we removed ethernet headers from the packet and called
399
+ this category: Without ethernet. Ultimately, we dropped both ethernet and IPv4 headers from the
400
+ packet. It was called the No headers category. As a result, all three representations of traffic
401
+ (session, flow, and packet) were split into these four categories, and each category was then
402
+ trained separately. As shown in Figure 2, there were 12 different experiments in our study. All
403
+ categories in each representation had the same size.
404
+ 3.3. DL Architecture
405
+ At a device, packets arrive as instances of data distributed over time which puts our dataset into
406
+ the time-series domain. Until recent years 1D-CNNs have been primarily used in the Natural
407
+ Language Processing (NLP) models [33]. However, they have also successfully classified time
408
+ series data [34]. This study aims to develop a small neural network that can be installed on
409
+ resource constraint devices and detect malicious streams of packets. Since tensor computations in
410
+ 1D-CNN take less space than 2D-CNN and other ML methods, it has motivated us to use this
411
+ lightweight neural network as our machine learning model. A smaller neural network size will
412
+ increase its practicability on IoT devices.
413
+ As shown in Figure 3, our smaller neural network comprises two 1D-CNN layers, a Maxpooling
414
+ layer, a Dropout layer, and a Dense layer. We started with the first convolution layer that
415
+ performed convolution with a kernel size of 64 and a stride of 3 on the input vectorized byte
416
+ stream. Small values of these parameters reduced the complexity of the model resulting in less
417
+ overfitting. The maxpooling layer was sandwiched between the two convolution layers. It
418
+ performed a 5-to-1 pooling operation to reduce the output tensor size from the upper layer without
419
+ losing significant properties. The second convolution layer was placed after the maxpooling layer
420
+ and performed the same job as the first convolution layer. It also had the same environmental
421
+ controls. Next, we used the dropout layer with a drop rate of 0.5 to reduce validation loss. In the
422
+ end, we added a fully connected Dense layer that helped the model learn any non-linear
423
+ relationship between features.
424
+
425
+
426
+ PcapFiles
427
+ PacketView
428
+ Flow View
429
+ SessionView
430
+ Packet
431
+ Flow
432
+ Session
433
+ Bytestream
434
+ Bytestream
435
+ Bytestream
436
+ Legends
437
+ 1:All headers. 2:Ethernetheaders only 3:WithoutEthernet header 4:No headers
438
+ Figure 3. DL Architecture
439
+ The left section of Figure 3 depicts binary classification. It used the binary loss as the loss function
440
+ and the softmax function for activation. On the other hand, multi-label classification, as depicted
441
+ in the right section of the figure, used categorical cross-entropy as a loss function and a sigmoid
442
+ function for activation. The binary classifier performed classification between benign and
443
+ malicious traffic. We trained the multi-label model to classify the five botnet categories
444
+ mentioned in section 3.2. The neural network trained over 60,000 hyperparameters in batches of
445
+ 32 using Keras [35] for 50 epochs. For efficiency, we also used a checkpoint function to store the
446
+ best model whenever encountered during training. It enables TensorFlow [36] to stop training
447
+ when it achieves the best possible value of the evaluation metric, which is “Accuracy,” in this
448
+ case. When the hyperparameters are suboptimal, the resultant model becomes complex, leading
449
+ to overfitting and high validation loss. This will result in more power usage and potentially
450
+ incorrect classification.
451
+ We trained on Nvidia GeForce GTX 1060 GPU with a 12GB Ubuntu 16.04 server on an x86
452
+ architecture.
453
+ 4. EVALUATION
454
+ 4.1. Evaluation Metrics
455
+ Experiments were evaluated on two metrics: accuracy and f-1 score, as shown in Figure 4.
456
+ Accuracy is the percentage of correct results from the total results produced by the model. It is
457
+ calculated on both training and validation data. The f-1 score is the weighted average of precision
458
+ and recall values on the validation data. f-1 score is also suitable for datasets with uneven
459
+ distribution which makes it suitable for our experiments.
460
+
461
+ Figure 4. Evaluation Metrics
462
+
463
+ ByteStreams
464
+ ByteStreams
465
+ 1-DCNN
466
+ 1-DCNN
467
+ Maxpooling
468
+ Maxpooling
469
+ 1-DCNN
470
+ 1-DCNN
471
+ Dropout
472
+ Dropout
473
+ Dese
474
+ Dense
475
+ Malicious
476
+ Benign
477
+ M1
478
+ M2
479
+ M3
480
+ M4
481
+ M54.2. Experimental evaluations
482
+ Our experiments indicate that binary classification achieved better accuracy the multi-label. First,
483
+ we will discuss binary classification outcomes between malicious and benign traffic. Tables 3 and
484
+ 4 display evaluation metrics in each ExpP, ExpS, and ExpF.
485
+ Table 3. Binary Performance on IoT-23 dataset
486
+ Representation
487
+ Header
488
+ Accuracy
489
+ f1-score
490
+ ExpS
491
+ All headers
492
+ 1.00
493
+ 0.97
494
+ Only Eth
495
+ 1.00
496
+ 0.96
497
+ Without Eth
498
+ 1.00
499
+ 0.96
500
+ No headers
501
+ 1.00
502
+ 0.94
503
+ ExpF
504
+ All headers
505
+ 1.00
506
+ 0.97
507
+ Only Eth
508
+ 1.00
509
+ 0.93
510
+ Without Eth
511
+ 0.97
512
+ 0.96
513
+ No headers
514
+ 0.99
515
+ 1.00
516
+ ExpP
517
+ All headers
518
+ 1.00
519
+ 0.96
520
+ Only Eth
521
+ 0.98
522
+ 0.96
523
+ Without Eth
524
+ 0.98
525
+ 0.97
526
+ No headers
527
+ 0.99
528
+ 0.95
529
+
530
+ Table 4. Multi-label Performance on IoT-23 dataset
531
+ Representation
532
+ Header
533
+ Accuracy
534
+ f1-score
535
+ ExpS
536
+ All headers
537
+ 0.99
538
+ 0.96
539
+ Only Eth
540
+ 0.94
541
+ 0.93
542
+ Without Eth
543
+ 0.84
544
+ 0.92
545
+ No headers
546
+ 0.96
547
+ 0.92
548
+ ExpF
549
+ All headers
550
+ 0.93
551
+ 0.92
552
+ Only Eth
553
+ 0.72
554
+ 0.85
555
+ Without Eth
556
+ 0.79
557
+ 0.91
558
+ No headers
559
+ 0.91
560
+ 0.90
561
+ ExpP
562
+ All headers
563
+ 0.97
564
+ 0.93
565
+ Only Eth
566
+ 0.98
567
+ 0.93
568
+ Without Eth
569
+ 0.74
570
+ 0.80
571
+ No headers
572
+ 0.98
573
+ 0.93
574
+
575
+
576
+ In pcap experiment ExpP, the “no-header” category achieved the maximum accuracy of 99%.
577
+ However, in the “all-headers” category, ExpP achieved 97% accuracy, only 0.02% below the
578
+ highest but gave the highest 99% f-1 score.
579
+ In ExpS, binary classification achieved 100% accuracy in every header category. It also achieved
580
+ the highest f-1 score of 97% in the “all-headers” category.
581
+ In ExpF, the “all-headers” category again achieved 100% accuracy alongside the “only-
582
+ ethernet” category. In contrast, ExpF achieved a 100% f-1 score “no-header” category, while the
583
+ “all-headers” category achieved only a 97% f-1 score.
584
+ We now switch our attention to multi-label classification between five botnet attack scenarios.
585
+ Table 4 shows the performance of the evaluation metrics in all categories: ExpP, ExpS, and ExpF.
586
+ Overall, the 5-label classification achieved a maximum accuracy of 99% and an f-1 score of 96%
587
+ in the ExpS session scenario.
588
+ In ExpP, accuracy was 98%, and the f-1 score was 93% f-1 score in the “no header” category.
589
+ However, the “all-headers” category was only 0.01% behind with a 97% accuracy and the same
590
+ 93% f-1 score.
591
+ In ExpS, the “all-headers” category achieved the highest accuracy of 99% with a 96% f-1 score.
592
+ In ExpF, the “all-headers” category again achieved the highest accuracy of 93%, along with the
593
+ highest f-1 score of 92%.
594
+ Table 5. Binary comparison between IoT-23 and ETF-IoT Performance
595
+ Representation
596
+ Header
597
+ IoT-23 f-1
598
+ ETF-IoT f-1
599
+ ExpS
600
+ All headers
601
+ 0.97
602
+ 0.89
603
+ Only Eth
604
+ 0.96
605
+ 0.87
606
+ Without Eth
607
+ 0.96
608
+ 0.87
609
+ No headers
610
+ 0.94
611
+ 0.90
612
+ ExpF
613
+ All headers
614
+ 0.97
615
+ 0.87
616
+ Only Eth
617
+ 0.93
618
+ 0.86
619
+ Without Eth
620
+ 0.96
621
+ 0.87
622
+ No headers
623
+ 1.00
624
+ 0.98
625
+ ExpP
626
+ All headers
627
+ 0.96
628
+ 0.88
629
+ Only Eth
630
+ 0.96
631
+ 0.89
632
+ Without Eth
633
+ 0.97
634
+ 0.89
635
+ No headers
636
+ 0.95
637
+ 0.89
638
+
639
+ Evidently, session representation or ExpS performed better in binary and multi-label
640
+ classifications. Overall, accuracy was highest when the dataset included all headers. Accuracy
641
+ monotonically decreased when either header was removed. However, it recovered when there
642
+ were no headers. The f-1 score was always the highest when all headers were included, with one
643
+ exception in multi-label ExpF.
644
+ This study shows that headers significantly influence the precision of anomaly detection models.
645
+ Unlike the custom of cropping them out of the training set [17] we achieve better results by
646
+ incorporating them into the training set.
647
+
648
+ 4.3. Comparison with another Dataset
649
+ In this section, we compare the results of our model trained on another dataset named ETF IoT
650
+ Botnet [37]. ETF is the newest publicly available botnet dataset. It has 42 malicious botnet attack
651
+ scenarios collected on RaspberryPi devices and two benign scenarios.
652
+ Tables 5 and 6 show lateral comparisons between both datasets. We have shown f-1 score
653
+ comparisons. The results differ case by case depending on the placement of the header.
654
+ Table 6. Multi-label comparison between IoT-23 and ETF-IoT Performance
655
+ Representation
656
+ Header
657
+ IoT-23 f-1
658
+ ETF-IoT f-1
659
+ ExpS
660
+ All headers
661
+ 0.96
662
+ 0.95
663
+ Only Eth
664
+ 0.93
665
+ 0.93
666
+ Without Eth
667
+ 0.92
668
+ 0.94
669
+ No headers
670
+ 0.92
671
+ 0.93
672
+ ExpF
673
+ All headers
674
+ 0.93
675
+ 0.96
676
+ Only Eth
677
+ 0.85
678
+ 0.95
679
+ Without Eth
680
+ 0.91
681
+ 0.96
682
+ No headers
683
+ 0.90
684
+ 0.94
685
+ ExpP
686
+ All headers
687
+ 0.97
688
+ 0.96
689
+ Only Eth
690
+ 0.93
691
+ 0.95
692
+ Without Eth
693
+ 0.80
694
+ 0.96
695
+ No headers
696
+ 0.93
697
+ 0.93
698
+
699
+ Figures 5 and 6 show a comparative analysis of f-1 scores between the two datasets. Both datasets
700
+ show similar results in all categories with a few exceptions. Noticeably, the “all headers”
701
+ category did not show any change in trend. f-1 scores of this model are consistent with IoT-23,
702
+ which strengthens the claim of FEL-ML’s usability for malware detection. As the obvious next
703
+ step, we have tested our model’s feasibility to be deployed on IoT devices.
704
+ We selected five botnet classes using the same technique we used in IoT-23. They are: 1) 666, 2)
705
+ SNOOPY, 3) arm7.idopoc2, 4) z3hir arm7, and 5) arm7l 1. We also used the same tools and
706
+ scripts to extract session, flow, and packet representation from pcap files. We trained on the same
707
+ GPU setting. All parameters of the 1D-CNN were also the same.
708
+
709
+
710
+
711
+
712
+ Figure 5 f-1 score comparison of Binary Classification (a) All headers (b) No headers (c) With
713
+ Ethernet headers (d) Without Ethernet headers
714
+
715
+ Figure 6 f-1 score comparison of Multi Classification (a) All headers (b) No headers (c) With
716
+ Ethernet headers (d) Without Ethernet headers
717
+
718
+
719
+ 1.00
720
+ 0.95
721
+ 0.90
722
+ 0.85
723
+ 0.80
724
+ 0.70
725
+ 0.65
726
+ 0.60
727
+ 0.55
728
+ 0.50
729
+ ExpS
730
+ ExpF
731
+ ExpP
732
+ loT-23
733
+ ■ETF-loT1.00
734
+ 0.95
735
+ 0.90
736
+ 0.85
737
+ 0.80
738
+ 0.75
739
+ f-1
740
+ 0.70
741
+ 0.65
742
+ 0.60
743
+ 0.55
744
+ 0.50
745
+ ExpS
746
+ ExpF
747
+ ExpP
748
+ 1oT-23
749
+ ETF-loT1.00
750
+ 0.95
751
+ 0.90
752
+ 0.85
753
+ 0.80
754
+ 0.70
755
+ 0.65
756
+ 0.60
757
+ 0.55
758
+ 0.50
759
+ ExpS
760
+ ExpF
761
+ ExpP
762
+ loT-23
763
+ ETF-loT1.00
764
+ 0.95
765
+ 0.90
766
+ 0.85
767
+ score
768
+ 0.80
769
+ 0.75
770
+ 0.70
771
+ 0.65
772
+ 0.60
773
+ 0.55
774
+ 0.50
775
+ ExpS
776
+ ExpF
777
+ ExpP
778
+ 1oT-23
779
+ ETF-IoT1.00
780
+ 0.95
781
+ 0.90
782
+ 0.85
783
+ 0.80
784
+ 0.70
785
+ 0.65
786
+ 0.60
787
+ 0.55
788
+ 0.50
789
+ ExpS
790
+ ExpF
791
+ ExpP
792
+ loT-23
793
+ ETF-loT1.00
794
+ 0.95
795
+ 0.90
796
+ 0.85
797
+ 0.80
798
+ 0.70
799
+ 0.65
800
+ 0.60
801
+ 0.55
802
+ 0.50
803
+ ExpS
804
+ ExpF
805
+ ExpP
806
+ ■loT-23
807
+ ETF-loT1.00
808
+ 0.95
809
+ 0.90
810
+
811
+ 0.85
812
+ 0.80
813
+ 0.70
814
+ 0.65
815
+ 0.60
816
+ 0.55
817
+ 0.50
818
+ ExpS
819
+ ExpF
820
+ ExpP
821
+ loT-23
822
+ ETF-loT1.00
823
+ 0.95
824
+ 0.90
825
+ 0.85
826
+ score
827
+ 0.80
828
+ 0.75
829
+ f-1
830
+ 0.70
831
+ 0.65
832
+ 0.60
833
+ 0.55
834
+ 0.50
835
+ ExpS
836
+ ExpF
837
+ ExpP
838
+ loT-23
839
+ ETF-IoT4.4. Applicability in IoT scenario
840
+ These devices are cost-sensitive, resulting in lower-performing CPUs and less memory,
841
+ demanding much lower-cost detection schemes. We couldn’t find a study that addresses this issue.
842
+ It prompted us to develop a faster ML model than the existing standards. In this section, we
843
+ compare our FEL-ML’s performance with an existing feature-based ML model. Both models
844
+ were trained and tested on the same GPU-enabled architecture.
845
+ Since our focus is on botnet attack detection, for this experiment, we selected an ML model called
846
+ n-BaIoT [2], which detects botnet attacks on IoT devices. We used the [38] GitHub repository to
847
+ reproduce their model and train on the dataset used in the original study. We extracted 115
848
+ numerical features from the dataset, as mentioned in the paper. We performed binary
849
+ classification between benign and malicious traffic to compare model performances. We used
850
+ three performance metrics in this experiment:
851
+ 1. Time consumed on testing the dataset, measured in seconds,
852
+ 2. System time,
853
+ 3. CPU utilization.
854
+ We measured 1 using the “time” function in python. We recorded 2 using the Linux time function.
855
+ We measured 3 using the “perf” tool on Ubuntu [39]. perf is a profiling tool that provides kernel-
856
+ level information about a program when it executes.
857
+ Table 6. Performance applicability for IoT devices
858
+ Model
859
+ Accuracy
860
+ Time elapsed
861
+ (sec)
862
+ System time
863
+ (sec)
864
+ CPU utilization
865
+ (max:2)
866
+ Binary ExpS
867
+ 1.00
868
+ 2.813
869
+ 0.71
870
+ 1.171
871
+ Binary ExpF
872
+ 1.00
873
+ 7.269
874
+ 0.82
875
+ 0.513
876
+ Binary ExpP
877
+ 1.00
878
+ 29.626
879
+ 2.42
880
+ 1.350
881
+ Binary n-BaIoT
882
+ 0.99
883
+ 22.877
884
+ 1.30
885
+ 1.238
886
+
887
+ All the binary classifications of our model used the “all-headers” category of the dataset. As
888
+ displayed in Table 7, the n-BaIoT model trained on 115 features took 22.877 seconds to perform
889
+ testing. However, its accuracy remained at 99.96%. Our featureless model performed better in
890
+ session and flow (ExpS and ExpF) representations, where it used less Testing Time compared to
891
+ the feature-based model. Testing time of only ExpP pcap representation took 6.746 seconds more
892
+ than the n-BaIoT model. Similar trends were seen in system time. Similarly, session and flow
893
+ utilize less CPU compared to feature engineered models.
894
+ 5. CONCLUSION
895
+ Contrary to traditional ML methodologies, FEL-ML does not require the complex processing
896
+ power deemed necessary. With the ease of implementation, more industrial domains can now
897
+ include it in their day-to-day operations. Security of IoT devices is one such domain. With IoT
898
+ devices running on battery power, more accurate results can be achieved with less computational
899
+ overhead by training raw bytes on 1D-CNN. As well as identifying anomalies in traffic with 100%
900
+ accuracy, this methodology is able to identify their types with 99% accuracy. Previous works on
901
+ this topic have scraped headers from their training set. However, our experiment compares models
902
+ trained with and without headers. This extensive experiment reinforces our argument that feature
903
+ engineering and removing headers from packets is a step in the traffic classification process that
904
+ is unnecessary.
905
+
906
+ As evident from Table 7, one challenge faced by our algorithm is that pcap representation could
907
+ perform better than flow and session. Flow and session require additional overhead in
908
+ consolidation before feature engineering. The next step of this research will attempt to discover
909
+ simpler ML systems that are efficient for direct packet captures. Speed of detection and accuracy
910
+ are of utmost importance to performing more granular detection of malware on IoT devices.
911
+ REFERENCES
912
+ [1]
913
+ De Lucia, Michael, and Chase Cotton. 2018. “Importance of Features in Adversarial Machine
914
+ Learning for Cyber Security.” 11.
915
+ [2]
916
+ Meidan, Yair, Michael Bohadana, Yael Mathov, Yisroel Mirsky, Asaf Shabtai, Dominik
917
+ Breitenbacher, and Yuval Elovici. 2018. “N-BaIoT—Network-Based Detection of IoT Botnet
918
+ Attacks Using Deep Autoencoders.” IEEE Pervasive Computing 17 (3): 12–22.
919
+ http://dx.doi.org/10.1109/MPRV. 2018.03367731.
920
+ [3]
921
+ Junior, Gilberto, Joel Rodrigues, Luiz Carvalho, Jalal Al-Muhtadi, and Mario Proenca. 2018. “A
922
+ comprehensive survey on network anomaly detection.” Telecommunication Systems 70.
923
+ [4]
924
+ Bertino, Elisa, and Nayeem Islam. 2017. “Botnets and Internet of Things Security.” Computer 50
925
+ (2): 76–79.
926
+ [5]
927
+ Hussain, Fatima, Rasheed Hussain, Syed Ali Hassan, and Ekram Hossain. 2020. “Machine
928
+ Learning in IoT Security: Current Solutions and Future Challenges.” IEEE Communications
929
+ Surveys Tutorials 22 (3): 1686–1721.
930
+ [6]
931
+ Dacier, Marc C., Hartmut Konig, Radoslaw Cwalinski, Frank Kargl, and Sven Dietrich.2017.
932
+ “Security Challenges and Opportunities of Software-Defined Networking.” IEEE Security Privacy
933
+ 15 (2): 96– 100.
934
+ [7]
935
+ Velan, Milan, Petr an d Cermak, Pavel Celeda, and Martin Drasar. 2015. “A Survey of Methods
936
+ for Encrypted Traffic Classification and Analysis.” Netw. 25 (5): 355–374.
937
+ [8]
938
+ Wang, Pan, Xuejiao Chen, Feng Ye, and Sun Zhixin. 2019. “A Survey of Techniques for Mobile
939
+ Service Encrypted Traffic Classification Using Deep Learning.” IEEE Access PP: 1–1.
940
+ [9]
941
+ Wei Wang, Ming Zhu, Xuewen Zeng, Xiaozhou Ye, and Yiqiang Sheng. 2017. “Malware traffic
942
+ classification using convolutional neural network for representation learning.” In 2017
943
+ International Conference on Information Networking (ICOIN), 712–717.
944
+ [10]
945
+ Stratosphere. 2015. “Stratosphere Laboratory Datasets.” Retrieved March 13, 2020, from
946
+ https://www.stratosphereips.org/datasets-overview.
947
+ [11]
948
+ Wang, W., M. Zhu, J. Wang, X. Zeng, and Z. Yang. 2017. “End-to-end encrypted traffic
949
+ classification with one-dimensional convolution neural networks.” In 2017 IEEE International
950
+ Conference on Intelligence and Security Informatics (ISI), 43–48.
951
+ [12]
952
+ UNB, Canadian Institute for Cybersecurity. 2016. “ICSX VPN-nonVPN Dataset.”
953
+ https://www.unb.ca/cic/datasets/vpn.html.
954
+ [13]
955
+ Lotfollahi, Mohammad, Ramin Shirali hossein zade, Mahdi Jafari Siavoshani, and
956
+ Mohammadsadegh Saberian. 2017. “Deep Packet: A Novel Approach For Encrypted Traffic
957
+ Classification Using Deep Learning.” Soft Computing 24.
958
+ [14]
959
+ Rezaei, S., and X. Liu. 2020. “Multitask Learning for Network Traffic Classification.” In 2020
960
+ 29th International Conference on Computer Communications and Networks (ICCCN), 1–9.
961
+ [15]
962
+ Tong, V., H. A. Tran, S. Souihi, and A. Mellouk. 2018. “A Novel QUIC Traffic Classifier Based
963
+ on Convolutional Neural Networks.” In 2018 IEEE Global Communications Conference
964
+ (GLOBECOM), 1–6.
965
+ [16]
966
+ Huang, X., G. C. Fox, S. Serebryakov, A. Mohan, P. Morkisz, and D. Dutta. 2019. “Benchmarking
967
+ Deep Learning for Time Series: Challenges and Directions.” In 2019 IEEE International
968
+ Conference on Big Data (Big Data), 5679–5682.
969
+
970
+ [17]
971
+ Marin, Gonzalo, Pedro Casas, and German Capdehourat. 2020. “DeepMAL – Deep Learning
972
+ Models for Malware Traffic Detection and Classification.”.
973
+ [18]
974
+ Shouran, Zaied, Ahmad Ashari, and Tri Priyambodo. 2019. “Internet of things (IoT) of smart
975
+ home: privacy and security.” International Journal of Computer Applications 182 (39): 3–8.
976
+ [19]
977
+ Anthi, Eirini, Lowri Williams, Małgorzata Słowinska, George Theodorakopoulos, and Pete
978
+ Burnap. 2019. “A Supervised Intrusion Detection System for Smart Home IoT Devices.” IEEE
979
+ Internet of Things Journal 6 (5): 9042–9053.
980
+ [20]
981
+ Vinayakumar, R., Mamoun Alazab, Sriram Srinivasan, Quoc-Viet Pham, Soman Kotti Padannayil,
982
+ and K. Simran. 2020. “A Visualized Botnet Detection System Based Deep Learning for the
983
+ Internet of Things Networks of Smart Cities.” IEEE Transactions on Industry Applications 56 (4):
984
+ 4436–4456.
985
+ [21]
986
+ Sriram, S., R. Vinayakumar, Mamoun Alazab, and Soman KP. 2020. “Network Flow based IoT
987
+ Botnet Attack Detection using Deep Learning.” In IEEE INFOCOM 2020 - IEEE Conference on
988
+ Computer Communications Workshops (INFOCOM WKSHPS), 189–194.
989
+ [22]
990
+ Limthong, K., and T. Tawsook. 2012. “Network traffic anomaly detection using machine learning
991
+ approaches.” In 2012 IEEE Network Operations and Management Symposium, 542-545.
992
+ [23]
993
+ Zenati, H., M. Romain, C. Foo, B. Lecouat, and V. Chandrasekhar. 2018. “Adversarially Learned
994
+ Anomaly Detection.” In 2018 IEEE International Conference on Data Mining (ICDM), 727–736.
995
+ [24]
996
+ PcapPlusPlus, Santiago Hernandez Ramos. 2019. “PcapSplitter.” https://github.com/shramos/
997
+ pcap-splitter.
998
+ [25]
999
+ Netresec. 2010. “SplitCap.” Used in 2020 https://www.netresec.com/.
1000
+ [26]
1001
+ Agustin Parmisano, Maria Jose Erquiaga, Sebastian Garcia. 2020. “Stratosphere Laboratory
1002
+ Aposemat Iot-23.” https://www.stratosphereips.org/datasets-iot23.
1003
+ [27]
1004
+ Bobrovnikova, Kira, Sergii Lysenko, Piotr Gaj, Valeriy Martynyuk, and Dmytro Denysiuk. 2020.
1005
+ “Technique for IoT Cyberattacks Detection Based on DNS Traffic Analysis.” http://ceur-
1006
+ ws.org/Vol-2623/paper19.pdf.
1007
+ [28]
1008
+ Cortes, C., and V. Vapnik. 1995. “Support-vector networks.” Machine Learning 20: 273–297.
1009
+ [29]
1010
+ Blaise, Agathe, Mathieu Bouet, Vania Conan, and Stefano Secci. 2020. “Botnet Fingerprinting: A
1011
+ Frequency Distributions Scheme for Lightweight Bot Detection.” IEEE Transactions on Network
1012
+ and Service Management PP.
1013
+ [30]
1014
+ Stoian, N.A. 2020. “Machine Learning for anomaly detection in IoT networks: Malware analysis
1015
+ on the IoT-23 data set.” July. http://essay.utwente.nl/81979/.
1016
+ [31]
1017
+ Ho, Tin Kam. 1995. “Random Decision Forests.” In Proceedings of the Third International
1018
+ Conference on Document Analysis and Recognition (Volume 1) - Volume 1, USA, 278. IEEE
1019
+ Computer Society.
1020
+ [32]
1021
+ Wireshark. 2006. “tshark.” Used in 2020 https://www.wireshark.org/docs/man-pages/tshark.html.
1022
+ [33]
1023
+ Kim, Yoon. 2014. “Convolutional Neural Networks for Sentence Classification.” In Proceedings
1024
+ of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), Doha,
1025
+ Qatar,
1026
+ Oct.,
1027
+ 1746–1751.
1028
+ Association
1029
+ for
1030
+ Computational
1031
+ Linguistics.
1032
+ https://www.aclweb.org/anthology/D14-1181.
1033
+ [34]
1034
+ Anguita, Davide, Alessandro Ghio, Luca Oneto, Xavier Parra, and Jorge L. Reyes-Ortiz. 2012.
1035
+ “Human Activity Recognition on Smartphones Using a Multiclass Hardware-Friendly Support
1036
+ Vector Machine.” In Ambient Assisted Living and Home Care, edited by José Bravo, Ramón
1037
+ Hervás, and Marcela Rodriguez, Berlin, Heidelberg, 216–223. Springer Berlin Heidelberg.
1038
+ [35]
1039
+ Chollet, Francois, et al. 2015. “Keras.” https://keras.io.
1040
+ [36]
1041
+ Abadi, Martín, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg
1042
+ S. Corrado, et al. 2015. “TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems.”
1043
+ Software available from tensorflow.org, https://www.tensorflow.org/.
1044
+
1045
+ [37]
1046
+ Jovanovic, Pavle, Dorde; Vuletic. 2021. “ETF IoT Botnet Dataset.” https://data.mendeley.com/
1047
+ datasets/nbs66kvx6n/1, doi=10.17632/nbs66kvx6n.1.
1048
+ [38]
1049
+ Tsimbalist, Sergei. 2019. “Botnet Traffic Analysis.” https://github.com/sergts/ botnet-traffic-
1050
+ analysis/tree/master/classification.
1051
+ [39]
1052
+ www.kernel.org.
1053
+ 2009.
1054
+ “Linux
1055
+ profiling
1056
+ with
1057
+ performance
1058
+ counters.”
1059
+ https://perf.wiki.kernel.org/index.php/Main_Page.
1060
+ Authors
1061
+ Arshiya Khan is currently pursuing her Ph.D. in
1062
+ Electrical and Computer Engineering
1063
+ (Cybersecurity) at the University of
1064
+ Delaware, Newark, DE, USA. Her areas of
1065
+ interest include network security, artificial
1066
+ general intelligence, and fair machine
1067
+ learning. She wrote her M.S. thesis on
1068
+ feature taxonomy of network traffic for
1069
+ machine learning algorithms.
1070
+ Over the past 35 years, Chase Cotton (Ph.D. EE,
1071
+ UD, 1984; BS ME, UT Austin, 1975, CISSP) has
1072
+ held a variety of research, development, and
1073
+ engineering roles, mostly in telecommunications. In
1074
+ both the corporate and academic worlds, he has been
1075
+ involved in computer, communications, and security
1076
+ research in roles including communication carrier
1077
+ executive, product manager, consultant, and
1078
+ educator for the technologies used in Internet and
1079
+ data services.
1080
+
1081
+ Beginning in the mid-1980 Dr. Cotton's
1082
+ communications research in Bellcore's Applied
1083
+ Research Area involved creating new algorithms
1084
+ and methods in bridging, multicast, many forms of
1085
+ packet-based applications including voice & video,
1086
+ traffic monitoring, transport protocols, custom VLSI
1087
+ for communications (protocol engines and Content
1088
+ Addressable Memories), and Gigabit networking.
1089
+
1090
+ In the mid-1990s, as the commercial Internet began
1091
+ to blossom, he transitioned to assist carriers
1092
+ worldwide as they started their Internet businesses,
1093
+ including Internet Service Providers (ISPs), hosting
1094
+ and web services, and the first large scale
1095
+ commercial deployment of Digital Subscriber Line
1096
+ (DSL) for consumer broadband services. In 2000,
1097
+ Dr. Cotton assumed research, planning, and
1098
+ engineering for Sprint's global Tier 1 Internet
1099
+ provider, SprintLink, expanding and evolving the
1100
+ network significantly during his 8-year tenure. At
1101
+ Sprint, his activities include leading a team that
1102
+ enabled infrastructure for the first large-scale
1103
+ collection and analysis of Tier 1 backbone traffic
1104
+ and twice set the Internet 2 Land Speed World
1105
+ Record on a commercial production network.
1106
+
1107
+
1108
+
1109
+
1110
+
1111
+
1112
+
1113
+
1114
+
1115
+ Since 2008, Dr. Cotton has been at the University of
1116
+ Delaware in the Department of Electrical and
1117
+ Computer Engineering, initially as a visiting
1118
+ scholar, and later as a Senior Scientist, Professor of
1119
+ Practice, and Director of Delaware's Center for
1120
+ Intelligent CyberSecurity (CICS). His research
1121
+ interests include cybersecurity and high-availability
1122
+ software systems with funding drawn from the NSF,
1123
+ ARL, U.S. Army C5ISR, JPMorgan Chase, and
1124
+ other industrial sponsors. As Director, Cybersecurity
1125
+ Minor & MS Programs, he currently is involved in
1126
+ the ongoing development of a multi-faceted
1127
+ educational initiative at UD, where he is developing
1128
+ new security courses and degree programs,
1129
+ including a minor, campus and online graduate
1130
+ Master's degrees, and Graduate Certificates in
1131
+ Cybersecurity.
1132
+
1133
+ Dr. Cotton currently consults on communications
1134
+ and Internet architectures, software, and security
1135
+ issues for many carriers and equipment vendors
1136
+ worldwide.
1137
+
7dE1T4oBgHgl3EQf7QV5/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
8dFQT4oBgHgl3EQf4jbF/content/tmp_files/2301.13432v1.pdf.txt ADDED
@@ -0,0 +1,686 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Manipulation of polarization topology using a
2
+ Fabry-Pérot fiber cavity with a higher-order
3
+ mode optical nanofiber
4
+ MAKI MAEDA,1,* JAMEESH KELOTH,1 AND SÍLE NIC CHORMAIC1
5
+ 1Okinawa Institute of Science and Technology Graduate University, Onna, Okinawa 904-0495, Japan
6
+ *maki.maeda@oist.jp
7
+ Abstract: Optical nanofiber cavity research has mainly focused on the fundamental mode. Here,
8
+ a Fabry-Pérot fiber cavity with an optical nanofiber supporting the higher-order modes, TE01,
9
+ TM01, HE𝑜
10
+ 21, and HE𝑒
11
+ 21, is demonstrated. Using cavity spectroscopy, with mode imaging and
12
+ analysis, we observe cavity resonances that exhibit complex, inhomogeneous states of polarization
13
+ with topological features containing Stokes singularities such as C-points, Poincaré vortices, and
14
+ L-lines. In situ tuning of the intracavity birefringence enables the desired profile and polarization
15
+ of the cavity mode to be obtained. These findings open new research possibilities for cold atom
16
+ manipulation and multimode cavity quantum electrodynamics using the evanescent fields of
17
+ higher-order mode optical nanofibers.
18
+ 1.
19
+ Introduction
20
+ Novel phenomena that can be revealed in non-paraxial light, such as transverse spin and spin-orbit
21
+ coupling, have led to increasing interest in the tightly confined light observed in nano-optical
22
+ devices [1]. Optical nanofibers (ONFs), where the waist is subwavelength in size, are useful
23
+ in this context because they provide very tight radial confinement of the electric field and
24
+ facilitate diffraction-free propagation over several centimeters [2]. Most ONF research focuses
25
+ on single-mode ONFs (SM-ONFs) that only support the fundamental mode, HE11. In contrast,
26
+ higher-order mode ONFs (HOM-ONFs), fabricated from a few-mode optical fiber, can guide
27
+ HOMs, such as TE01, TM01, HE𝑒
28
+ 21, and HE𝑜
29
+ 21 [3]. In the weakly guided regime, which is generally
30
+ used to describe light propagation in standard optical fiber, this group of modes can be viewed
31
+ to form the linearly polarized mode, LP11. To date, there has been a lot more attention paid to
32
+ HOM-ONFs in theoretical work [4–10] than experimental work due to the difficulty in precisely
33
+ controlling the fiber waist size and obtaining selective mode excitation at the waist [3,11,12].
34
+ In principle, there are many interesting phenomena which can be explored with a HOM-ONF.
35
+ For example, it has been proposed that the relationship between spin angular momentum (SAM)
36
+ and orbital angular momentum (OAM) can be studied [5,10,13,14]. Additionally, it was proposed
37
+ that a HOM-ONF could be used to trap and manipulate cold atoms [4, 15, 16]. Fabrication
38
+ of an ONF that supports the HOMs was achieved [3,17,18] and subsequently shown to more
39
+ efficiently manipulate dielectric microbeads in the evanescent field than SM-ONFs [19, 20].
40
+ Other experimental work has shown that when cold atoms also interact with HOMs, detected
41
+ signals are stronger than when one uses a SM-ONF only [21].
42
+ Introducing a cavity system to the ONF could further increase light-matter interactions due
43
+ to cavity quantum electrodynamics (cQED) effects [22–24]. To date, numerous types of SM-
44
+ ONF-based cavities have been proposed [25–30] and the interactions of their resonance modes
45
+ with various quantum emitters have been studied [31–33]. Strong light-atom coupling using
46
+ SM-ONF-based Fabry-Pérot and ring resonators has already been achieved [34,35]. Superstrong
47
+ coupling of cold atoms and multiple longitudinal modes of a long fiber-ring resonator consisting
48
+ of a SM-ONF section was demonstrated [36].
49
+ Utilizing multiple degenerate higher-order
50
+ transverse modes in free-space has shown to exhibit strong coupling [37,38], further illustrating
51
+ the importance of realizing a HOM-ONF-based cavity system at this point. The advantages are
52
+ arXiv:2301.13432v1 [physics.optics] 31 Jan 2023
53
+
54
+ not only for enhanced interactions via cQED effects, but also for a better overall understanding of
55
+ the behavior of the modes in such a cavity.
56
+ Studying the behavior of the HOM-ONF cavity spectrum and the cavity mode profiles gives
57
+ additional insight into the nature of the HOMs themselves, as well as how they interfere with each
58
+ other and interact with the external environment. The generation of TE01 and TM01 modes in a
59
+ laser cavity consisting of a microfiber directional coupler-based mode converter was demonstrated
60
+ previously [39]. However, earlier attempts to realize a passive HOM optical microfiber cavity
61
+ did not yield any resonant peaks in the cavity spectrum apart from the fundamental modes; in
62
+ other words, the typical donut- or lobe-shaped intensity profiles associated with HOMs were not
63
+ observed [40], primarily due to challenges when engineering the taper profile to minimize losses
64
+ at the taper transitions.
65
+ The inhomogeneous polarization structure of HOMs needs to be taken into account when
66
+ studying a fiber cavity system with a HOM-ONF. In recent years, complex polarization dis-
67
+ tributions and the generation of polarization singularities have been investigated using various
68
+ methods, giving rise to the relatively new field of singular optics [41]. Polarization singularities
69
+ are a subset of Stokes singularities, i.e., phase singularity points in Stokes phases [42,43]. In
70
+ fact, higher-order fiber eigenmodes are vector optical fields with a polarization singularity called
71
+ a V-point, where the state of polarization (SOP), i.e., how the polarization is distributed in the
72
+ cross-section of a given mode, is undefined [41]. Other types of Stokes singularities can be
73
+ formed in elliptical optical fields, such as the polarization singularity of C-points, where the
74
+ polarization orientation is undefined [41, 42], and Poincaré vortices, where the polarization
75
+ handedness is undefined [43–45]. Moreover, points of linear polarization can form continuous
76
+ lines, which are classified as L-lines [41].
77
+ The generation of all Stokes singularities within a single beam has been demonstrated using a
78
+ free-space interferometer [43,46]. Modal interference in a birefringent crystal can facilitate the
79
+ creation of polarization singularities [47,48]. As a result, the SOP can significantly vary along
80
+ the propagation length, with C-points and L-lines propagating as C-lines, i.e., continuous lines of
81
+ circular polarization, and L-surfaces, i.e., surfaces of linear polarization, respectively [47–49].
82
+ Moreover, polarization singularities can appear, move or disappear from a given cross-sectional
83
+ region with a smooth and continuous change of birefringence [50]. Birefringent media were used
84
+ to create laser cavity modes containing a polarization singularity [51,52]. These experiments were
85
+ limited to the generation of low-order V-points due to a lack of control in the amplitude, phase,
86
+ and SOP, all of which would be required to create other types of polarization singularities [41].
87
+ A few-mode optical fiber cavity has the potential to generate complex laser modes by its highly
88
+ variable degree of birefringence.
89
+ Interference and birefringence are generally inseparable properties in fibers. The modal
90
+ interference pattern in a fiber changes continually with a periodicity of 2𝜋 when the relative phase
91
+ between modes is changed between 0 to 2𝜋 as the eigenmodes propagate along the fiber [53]. This
92
+ effect was used in a few-mode optical fiber to generate ellipse fields containing a C-point [54,55].
93
+ Due to the increasing complexities of modal interference in few-mode fibers, filtering for the
94
+ desired set of HOMs, and selectively exciting them to generate and manipulate polarization
95
+ singularities, are necessary. Realizing a fiber cavity containing an ONF should enable both
96
+ spatial and frequency filtering for selective excitation of HOMs, as well as enhancement of the
97
+ resonant mode coupling effect [56,57].
98
+ In this paper, we experimentally demonstrate a HOM-ONF-based Fabry-Pérot fiber cavity.
99
+ The transverse polarization topology of any given resonant mode is determined by selecting
100
+ modes from the cavity spectra and analyzing the images of the transmitted mode profile. We also
101
+ demonstrate in situ intracavity manipulation of the modal birefringence to change the amplitude,
102
+ frequency position, and the SOP of the modes. This work is a significant step towards gaining
103
+ full control of the evanescent field at the HOM-ONF waist and extends the range of applications
104
+
105
+ Fig. 1. (a) Sketch of tapered optical fiber with trilinear shape, d𝑤𝑎𝑖𝑠𝑡: waist diameter.
106
+ (b) Schematic of experimental setup. L: lens, HWP: half-wave plate, PBS: polarizing
107
+ beam splitter, M: mirror, M𝐶: cavity mirror, IPC: in-line polarization controller, BS:
108
+ beam splitter, QWP: quarter-wave plate, which was inserted to calculate S3, LP: linear
109
+ polarizer, CCD: camera, MMF: multimode fiber, PD: photodiode.
110
+ for which such nanodevices could be used.
111
+ 2.
112
+ Methods
113
+ 2.1.
114
+ Experiments
115
+ For the HOMs described in Section 1 to propagate throughout the cavity with a HOM-ONF,
116
+ the nanofiber must be low loss for the entire LP11 set of modes. Tapered fibers were drawn
117
+ from SM1250 (9/80) fiber (Fibercore) using an oxy-hydrogen flame pulling rig. The untapered
118
+ fiber supports the LP01, LP11, LP21, and LP02 modes at a wavelength, 𝜆 = 776 nm. The modes
119
+ supported by the tapered fiber depend on the tapering profile and the waist diameter. We used
120
+ two different tapered fibers with waist diameters of (i) ∼ 450 nm for SM behavior (HE𝑜
121
+ 11 and
122
+ HE𝑒
123
+ 11) and (ii) ∼ 840 nm for the HOM-ONF, which supports HE𝑜
124
+ 11, HE𝑒
125
+ 11, TE01, TM01, HE𝑜
126
+ 21,
127
+ and HE𝑒
128
+ 21. The shape of the tapered fibers was chosen to be trilinear, see Fig. 1(a), with angles
129
+ of Ω1 = 2 mrad, Ω2 = 0.5 mrad and Ω3 = 1 mrad in order to be adiabatic for the LP11 and LP01
130
+ modes. Fiber transmission following the tapering process was >95% for the fundamental mode.
131
+ A sketch of the experimental setup is given in Fig. 1(b). The cavity was fabricated by splicing
132
+ each pigtail of the tapered fiber to a commercial fiber Bragg grating (FBG) mirror (Omega
133
+ Optical). The two FBG mirrors consisted of stacked dielectric mirrors coated on the end faces
134
+ of fiber patchcords (SM1250 (9/80), Fibercore) and had a reflectivity of 97% at 𝜆 = 776 nm.
135
+ Both mirrors had almost the same reflectivity over all input polarization angles (< 1% variation).
136
+ The cavity also contained an in-line polarization controller (IPC, see Fig.1(b)) to manipulate the
137
+ birefringence inside the cavity. Moving the paddles of the IPC induced stress and strain in the
138
+ fiber, thereby changing the effective cavity length. A typical cavity length was ∼ 2 m, which was
139
+ physically measured and estimated from the cavity free-spectral range (FSR).
140
+
141
+ DigiLockA linearly polarized Gaussian beam from a laser at 𝜆 = 776 nm (Toptica DL100 pro) was
142
+ launched into the fiber cavity. The laser frequency was either scanned or locked to a mode of
143
+ interest using a Pound-Drever-Hall locking module (Toptica Digilock110). The cavity output
144
+ beam was split into three paths: one for the laser feedback controller to observe the cavity spectra
145
+ and to lock to specific modes, one for imaging the spatial profile of the modes with a CCD
146
+ camera, and one for analyzing the transverse SOP of each mode using a removable quarter wave
147
+ plate (QWP), a rotating linear polarizer, and a CCD camera, see Fig. 1(b). Six intensity profile
148
+ images were taken in total for each mode. Four images were taken without the QWP and with the
149
+ linear polarizer angle set to 0◦ (I𝐻), 45◦ (I𝐷), 90◦ (I𝑉 ), and 135◦ (I𝐴), and two images were taken
150
+ by inserting the QWP set to 90◦ while the polarizer was set to 45◦ (I𝑅) and 135◦ (I𝐿). The SOPs
151
+ were determined by analyzing the six profile images using Stokes polarimetry. Furthermore, the
152
+ Stokes phase and Stokes index were determined [41], see Section 2 2.3.
153
+ 2.2.
154
+ Simulations
155
+ Each mode experiences arbitrary birefringence as it propagates along the fiber. The total field
156
+ in the fiber at any point is the sum of the propagating modes with a corresponding phase shift.
157
+ The addition of FBG mirrors to the fiber induces an additional birefringence [56, 57], which
158
+ can be incorporated in a single birefringence matrix. Note, this model does not include cavity
159
+ boundary conditions since we only aim to simulate the spatial profiles of the fiber modes. We can
160
+ calculate an arbitrary fiber field, E, due to interference and birefringence by taking a summation
161
+ over different fiber modes, such that
162
+ E =
163
+ 𝑛
164
+ ∑︁
165
+ 𝑀=1
166
+ 𝐽𝑀 𝐴𝑀E𝑀𝑒𝑖𝜙𝑀 ,
167
+ (1)
168
+ where n is the number of eigenmodes to be interfered, E𝑀 is the electric field of a fiber eigenmode
169
+ M ∈ TE0,𝑚, TM0,𝑚, HEℓ,𝑚 and EHℓ,𝑚, with ℓ ∈ Z+ being the azimuthal mode order, which
170
+ defines the helical phase front and the associated phase gradient in the fiber transverse plane.
171
+ m ∈ Z+ is the radial mode order, which indicates the m𝑡ℎ solution of the corresponding eigenvalue
172
+ equation [5]. A𝑀 is the amplitude, 𝜙𝑀 is the phase between modes, and J𝑀 represents the
173
+ arbitrary birefringence Jones matrix of each eigenmode E𝑀, such that
174
+ 𝐽𝑀 = 𝑒𝑖𝜂𝑀/2 ��
175
+
176
+ 𝑐𝑜𝑠2𝜃𝑀 + 𝑒𝑖𝜂𝑀 𝑠𝑖𝑛2𝜃𝑀
177
+ (1 − 𝑒𝑖𝜂𝑀 )𝑐𝑜𝑠𝜃𝑀 𝑠𝑖𝑛𝜃𝑀
178
+ (1 − 𝑒𝑖𝜂𝑀 )𝑐𝑜𝑠𝜃𝑀 𝑠𝑖𝑛𝜃𝑀
179
+ 𝑠𝑖𝑛2𝜃𝑀 + 𝑒𝑖𝜂𝑀 𝑐𝑜𝑠2𝜃𝑀
180
+ ��
181
+
182
+ ,
183
+ (2)
184
+ where 𝜂𝑀 is the relative phase retardation induced between the fast axis and the slow axis, and
185
+ 𝜃𝑀 is the orientation of the fast axis with respect to the horizontal-axis, i.e., perpendicular to
186
+ mode propagation.
187
+ Let us now consider the system with an ONF supporting HE𝑜
188
+ 11, HE𝑒
189
+ 11, TE01, TM01, HE𝑜
190
+ 21 and
191
+ HE𝑒
192
+ 21, so that the number of modes that can be interfered is n ≤ 6. The cross-sectional profiles
193
+ and SOPs of TE01 and HE𝑒
194
+ 21 are shown in Fig. 2(a, b), respectively. The TM01 and HE𝑜
195
+ 21 modes
196
+ are not shown here but their vector fields are orthogonal to the TE01 and HE𝑒
197
+ 21 at every point,
198
+ respectively. These modes have donut-shape mode profiles with linearly polarized vector fields
199
+ at any point in the mode cross-section. As an example of possible fiber modes using Eq. 1, Fig.
200
+ 2(c) illustrates in-phase interference of the TE01 and HE𝑒
201
+ 21 modes with equal amplitudes. The
202
+ resulting mode has a lobe-shape intensity pattern with scalar fields. Fig. 2(d) is an example of
203
+ a mode resulting from the interference of the circularly polarized HE11 and an out-of-phase (a
204
+ 𝜋/2 phase difference) TE01 and TM01 with equal amplitudes. The SOP, which is overlapped on
205
+ the intensity profile images, are marked as red and blue ellipse, corresponding to right and left
206
+ handed orientation, respectively. This mode is the co-called lemon [55], which contains not only
207
+ linear polarization but also elliptical and circular polarization components in one mode.
208
+
209
+ Fig. 2. Simulations of (a) TE01, (b) HE𝑒
210
+ 21, (c) TE01 + HE𝑒
211
+ 21 and (d) lemon. The red
212
+ and blue SOPs indicate right-handed and left-handed ellipticities, respectively. The
213
+ scale bars show the normalized intensity (from 0 to 1) and the Stokes phase (from 0 to
214
+ 2𝜋). Stokes singularity points of 𝜎12, 𝜎23, and 𝜎31 are indicated as pink, orange, and
215
+ blue dots, respectively. An L-line is indicated in green.
216
+
217
+ (a)
218
+ (b)
219
+ (c)
220
+ (d)
221
+ Φ12
222
+ D.
223
+ DWhen using Eq. 1 to simulate mode profiles, a number of eigenmodes with similar intensity
224
+ patterns and SOPs to an experimentally observed cavity mode were selected as the initial
225
+ conditions. Next, the variables A𝑀, 𝜙𝑀, 𝜂𝑀, and 𝜃𝑀 were tuned to match the experimentally
226
+ observed cavity mode intensities, SOPs, and Stokes phases. Polarization topological defects in
227
+ the simulated modes were then identified, using the method described in the following Section 2
228
+ 2.3.
229
+ 2.3.
230
+ Analysis
231
+ The polarization gradient was calculated in order to identify Stokes singularities in the cross-
232
+ section of the mode. The gradient map is known as the Stokes phase, 𝜙𝑖 𝑗, which is given
233
+ by [42,45]
234
+ 𝜙𝑖 𝑗 = 𝐴𝑟𝑔(𝑆𝑖 + 𝑖𝑆 𝑗),
235
+ (3)
236
+ where 𝑆𝑖 and 𝑆 𝑗 are Stokes parameters with {i, j} ∈ {1, 2, 3} in order, and i ≠ j. The phase
237
+ uncertainty points, i.e., Stokes singularities, were identified by obtaining the Stokes indices, 𝜎𝑖 𝑗,
238
+ which are defined as [42,45]
239
+ 𝜎𝑖 𝑗 = 1
240
+ 2𝜋
241
+
242
+ 𝑐
243
+ 𝜙𝑖 𝑗 · 𝑑𝑐,
244
+ (4)
245
+ where
246
+
247
+ 𝑐 𝜙𝑖 𝑗 · 𝑑𝑐 = Δ 𝜙𝑖 𝑗 is the counterclockwise azimuthal change of the Stokes phase around the
248
+ Stokes singularity. Singularities of 𝜎12 are known as V-points and C-points, in vector and ellipse
249
+ fields, respectively [42]. Singularities of 𝜎23 and 𝜎31 are known as Poincaré vortices [43–45].
250
+ L-lines are located where 𝜙23 = {0, 𝜋, 2𝜋}. Table 1 is a summary of the classification of the Stokes
251
+ singularity types in terms of the Stokes phases and singularity indices with the corresponding
252
+ polarizations in the vector and ellipse fields [43,45,46,58].
253
+ Table 1. List of Stokes singularities in vector fields (v) and ellipse fields (e) by the
254
+ singularity index, 𝜎𝑖 𝑗, using the Stokes phase, 𝜙𝑖 𝑗, with {i, j} ∈ {1, 2, 3} in order.
255
+ Stokes
256
+ Stokes phase
257
+ Stokes index/
258
+ Polarization
259
+ singularity
260
+ Phase values
261
+ V-point (v)
262
+ 𝜙12
263
+ 𝜎12
264
+ Null
265
+ C-point (e)
266
+ 𝜙12
267
+ 𝜎12
268
+ R/L
269
+ Poincaré
270
+ 𝜙23
271
+ 𝜎23
272
+ H/V
273
+ vortex (e)
274
+ 𝜙31
275
+ 𝜎31
276
+ D/A
277
+ L-line (e)
278
+ 𝜙23
279
+ 0, 𝜋, 2𝜋
280
+ Linear
281
+ The Stokes singularity points and L-lines were found from the Stokes phases, then superimposed
282
+ and marked on the mode profiles. As examples, from Figs. 2(a, b), the center of the mode profiles
283
+ for both TE01 and HE𝑒
284
+ 21 contain a V-point, with 𝜎12 = -2 and +2 (pink dot), respectively. These
285
+ points were found from their Stokes phases 𝜙12 (lower panels in Figs. 2(a, b)). In contrast, the
286
+ lemon mode in Fig. 2(d) has a closed loop representing an L-line (green) and all three types of
287
+ Stokes singularities: a C-point with 𝜎12 = -1 (pink dot), Poincaré vortices with 𝜎23 = -1 and +1
288
+ (orange dots), and 𝜎31 = -1 and +1 (blue dots) were found from 𝜙12, 𝜙23, and 𝜙31, respectively.
289
+ The lobe-shaped scalar mode in Fig. 2(c) does not have a 2𝜋 gradient in any associated Stoke
290
+ phases, since topological defects can only exist in non-scalar fields [41].
291
+
292
+ 3.
293
+ Results and discussion
294
+ 3.1.
295
+ Cavity with a single-mode optical nanofiber
296
+ As an initial experimental test, the spectrum for a HOM cavity containing an ONF of waist
297
+ diameter ∼ 450 nm was obtained, see Fig. 3(a). This ONF waist can only support the fundamental
298
+ modes. The IPC paddle angles were set so that two distinct, well-separated modes with minimal
299
+ spectral overlap were observed. The finesses of Modes 1 and 2 in Fig. 3(a) were 12 and 15,
300
+ respectively. The laser was locked to each of these two cavity modes consecutively and the
301
+ mode profiles were observed at the output end face of the fiber cavity. The corresponding mode
302
+ intensity profiles, SOPs, and Stokes phases are shown in Figs. 3(b)(i, ii). The intensity profiles for
303
+ both Modes 1 and 2 were slightly skewed Gaussian shapes. The HE11 eigenmode intensity shape
304
+ is Gaussian, so the slight deviation from the expected shape may be attributed to aberrations in
305
+ the optical beam path. In terms of polarization distribution, the Stokes phases of Modes 1 and 2
306
+ were uniform; in other words, their SOPs were scalar fields, regardless of the IPC paddle angles
307
+ chosen, as expected for the HE11 mode.
308
+ Although the pretapered fiber supported the full set of eigenmodes in LP11, LP02, and LP21,
309
+ when the ONF with a diameter ∼ 450 nm was inserted between the two sets of mirrors, only one
310
+ or two modes with quasi-Gaussian profiles were observed, no matter which IPC paddle angles
311
+ were chosen. The HOMs were filtered out due to the tapered fiber waist being SM, analogous to
312
+ an intracavity pinhole spatial filter. Mode filtering as a function of the ONF waist diameter was
313
+ observed experimentally [17]. However, here, we could additionally observe the mode filtering
314
+ effect on the cavity spectrum and SOP of each mode.
315
+ In an ideal SM-ONF cavity with no birefringence, there are two degenerate orthogonal modes.
316
+ However, due to random birefringence of the fiber and the cavity mirrors, the two modes
317
+ become non-degenerate, i.e., separated in frequency, leading to coupling between the modes [59].
318
+ Mode coupling of orthogonal modes can occur in a birefringent medium and this effect can
319
+ increase in a cavity configuration [60]. Mode coupling in an ONF cavity due to asymmetrical
320
+ mirrors has been discussed previously [56] and experimental evidence of mode coupling due to
321
+ intrinsic birefringence in a SM-ONF cavity has already been reported [57]. In our experiments,
322
+ non-orthogonal combinations of SOPs were observed, as seen in Figs. 3(b)(i, ii). Mode 1 was
323
+ horizontally polarized (red/blue lines in Fig. 3(b)(i)), while Mode 2 was left elliptically polarized
324
+ (blue ellipse in Fig. 3(b)(ii)). By adjusting the IPC angles, it was possible to change the phase
325
+ relationship and coupling between the HE𝑜
326
+ 11 and HE𝑒
327
+ 11 modes, and shift between orthogonal and
328
+ non-orthogonal combinations of SOPs.
329
+ 3.2.
330
+ Cavity with a higher-order mode optical nanofiber
331
+ Next, the spectrum for a HOM cavity containing an ONF of waist diameter ∼ 840 nm was
332
+ obtained, see Fig. 4(a). This ONF can support the HE11, TE01, TM01, HE𝑜
333
+ 21, and HE𝑒
334
+ 21 modes.
335
+ The IPC paddle angles were set to obtain the maximum number of well-resolved modes in a
336
+ single FSR, see Fig. 4(a). One can clearly see five distinct peaks indicating that the HOM-ONF
337
+ does not degrade the modes in the cavity and the finesses of the cavity modes are high enough to
338
+ resolve them. The finesses of Modes 1 to 5 were 12, 16, 13, 22, and 13, respectively. The mode
339
+ finesse values of the cavity with a HOM-ONF were in the same range as those for the cavity
340
+ with a SM-ONF (Fig. 3(a)), implying that the HOM-ONF was adiabatic for the LP11 group of
341
+ modes. The laser was locked to each of the cavity modes consecutively and the mode profiles
342
+ were observed at the output of the fiber cavity. The corresponding mode intensity profiles, SOPs,
343
+ and Stokes phases are shown in Figs. 4(b)(i-iv). In the spectrum shown in Fig. 4(a), there were
344
+ five distinctive modes, but locking to Mode 3 was not possible because of its close proximity to
345
+ the dominant Mode 4.
346
+ Two flat-top intensity profiles were observed in Modes 1 and 4, Figs. 4(b)(i, iii) respectively.
347
+
348
+ Fig. 3. (a) A typical spectrum for a HOM cavity with a SM-ONF as the laser is scanned
349
+ over 150 MHz. The spectrum over a single FSR is indicated by the red box. (b) Mode
350
+ intensity profiles showing the SOPs (top) and corresponding Stokes phases (bottom)
351
+ for (i) Mode 1 and (ii) Mode 2. The red and blue SOPs indicate right-handed and
352
+ left-handed ellipticities, respectively. The scale bars show the normalized intensity
353
+ (from 0 to 1) and the Stokes phase (from 0 to 2𝜋).
354
+
355
+ Laser scan frequency (MHz)
356
+ (i)
357
+ (ii)
358
+ Φ12Fig. 4. (a) A typical spectrum for a cavity with a HOM-ONF as the laser is scanned
359
+ over 150 MHz. The spectrum over a single FSR is indicated by the red box. (b) Mode
360
+ intensity profiles showing the SOP (top) and the corresponding Stokes phases (bottom)
361
+ for (i) Mode 1, (ii) Mode 2, (iii) Mode 4, and (iv) Mode 5. The red and blue SOPs
362
+ indicate right-handed and left-handed ellipticities, respectively. The scale bars show
363
+ the normalized intensity (from 0 to 1) and the Stokes phase (from 0 to 2𝜋). Stokes
364
+ singularity points of 𝜎12, 𝜎23, and 𝜎31 are indicated as pink, orange, and blue dots,
365
+ respectively. L-lines are indicated in green. (c) Corresponding simulated results.
366
+
367
+ Laser scan frequency (MHz)
368
+ (i)
369
+ (ii)
370
+ (i)
371
+ (iv
372
+ D
373
+ D
374
+ (i)
375
+ (iv)
376
+ Φ12
377
+ Φ23
378
+ Φ31
379
+ Φ23
380
+ Φ31
381
+ D
382
+
383
+ 23
384
+ D
385
+ 3The SOPs of these modes are markedly different to those for the Gaussian-type modes in Figs.
386
+ 3(b)(i, ii), which have simple scalar SOPs. Modes 1 and 4 were inhomogeneously polarized
387
+ ellipse fields, showing regions of left and right circular polarizations divided by an L-line (Figs.
388
+ 4(b)(i, iii)). The center of these two modes exhibited diagonal and anti-diagonal polarizations,
389
+ respectively, i.e., the SOPs at the center of the modes were orthogonal to each other. Going
390
+ towards the edges of the modes, the polarization changes from linear to circular, with opposite
391
+ handedness either side of the L-lines. Notice also in Fig. 4(a) that Modes 1 and 4 are not well
392
+ frequency separated from neighboring modes. This suggests that the mode profiles and SOPs of
393
+ these modes were not only affected by birefringence and degenerate modal interference, but also
394
+ some non-degenerate modal interference with neighboring cavity modes [60]. Additionally, for
395
+ Mode 4, we identified two C-points (𝜎12 = -1), indicated by the pink dots in Fig. 4(b)(iii), where
396
+ the value of 𝜙12 changed by 2𝜋 (see Table 1). Interference of HE11 with modes from the LP11
397
+ group can generate C-points in a few-mode fiber [55], see Fig. 2(d).
398
+ We performed basic simulations to determine if combinations of HE11 and some mode(s) in the
399
+ LP11 family could generate similar mode profiles and SOP structures as those in Figs. 4(b)(i, iii).
400
+ The simulated results are shown in Figs. 4(c)(i, iii). The HE11 and TM01 modes were selected
401
+ as possible contributors and their amplitudes, phase, and birefringence fitting parameters were
402
+ tuned to match the experimental results. Modes 1 and 4, see Figs. 4(b)(i, iii), could have been
403
+ formed from different mode combinations rather than our assumed HE11 and TM01; however,
404
+ these modes were very likely formed by interference between HE11 and some mode(s) of the
405
+ LP11 group, resulting in their inhomogeneous SOPs and flat-top shapes.
406
+ We also observed two distorted lobe-shaped modes, Modes 2 and 5, see Figs. 4(b)(ii, iv).
407
+ The lobe-shaped pattern also arises from modal interference between modes in the LP11 family
408
+ (as an example, see Fig. 2(c)). With reference to Table 1, Mode 2, Fig. 4(b)(ii), showed
409
+ all three types of Stokes singularities, indicated by pink dots for C-points (𝜎12 = +1) and
410
+ orange/blue dots for Poincaré vortices (𝜎23 = -1 /𝜎31 = +1), as presented in 𝜙12, 𝜙23, and 𝜙31,
411
+ respectively. A single mode containing all Stokes singularities has been demonstrated using
412
+ free-space interferometers [43,46]; here, we generated them within a single mode using a fiber
413
+ cavity system. Mode 5, Fig. 4(b)(iv), also had two C-points (𝜎12 = +1) and a Poincaré vortex
414
+ (𝜎23 = +1), as seen in 𝜙12, and 𝜙23, respectively. Fig. 4(a) shows that Modes 2 and 5 are not well
415
+ frequency separated from Modes 1 and 4, respectively. Therefore, there is a likely contribution
416
+ from the HE11 mode resulting in distortion of the lobe shape.
417
+ To simulate Mode 2 in Fig. 4(b)(ii), we combined TE01, HE𝑒
418
+ 21, and HE11, and to simulate
419
+ Mode 5 in Fig. 4(b)(iv), we used TM01, HE𝑒
420
+ 21, and HE11. The amplitude of each mode, phase
421
+ shift, and birefringence parameters were adjusted to achieve a close fit. The simulated results
422
+ are shown in Figs. 4(c)(ii, iv). These plots are not exact replications of the experimental results
423
+ since the parameter space is large and the exact initial conditions are not known; nevertheless,
424
+ the match is reasonably close.
425
+ Interestingly, many of the cavity modes obtained in different sets of spectra, which were
426
+ generated using different IPC angles, exhibited Stokes singularities. Polarization singularities are
427
+ known to propagate through a birefringent medium as C-lines and L-surfaces and their evolution
428
+ is affected by the homogeneity of the birefringence along the propagation path [47–49]. This
429
+ phenomenon is due to the conservation of the topological charge [49,58,61], and the Stokes index
430
+ value, 𝜎𝑖 𝑗, remains constant [58]. However, our cavity is an inhomogeneous birefringent medium
431
+ as it contains a number of different birefringent elements such as the FBG mirrors and the IPC, as
432
+ such, the degree of birefringence varies along the propagation direction. Therefore, the presence
433
+ of Stokes singularities in the imaged field at the cavity output does not necessarily guarantee the
434
+ existence of such topological defects in the ONF region. Nonetheless, singularity points can
435
+ enter, move and exit with a smooth and continuous variation of birefringence [50]. Therefore,
436
+ the SOP is expected to evolve along the length of the cavity, with singularity points shifting and
437
+
438
+ making numerous entries and exits in the cross-section profile of the modes. However, since the
439
+ ONF waist is relatively straight and uniform, the birefringence variation at the waist should be
440
+ minimal [62] and topological features appearing at the start of the waist should be preserved
441
+ every 2𝜋 along the waist.
442
+ Theoretically, the HOM-ONF can support a total of six eigenmodes as mentioned earlier.
443
+ Therefore, one might expect that the spectrum should show six distinct modes. However, we
444
+ typically observed three to five distinct peaks in a single FSR depending on the IPC paddle angles.
445
+ This could be explained by the lack of sufficient finesse to resolve all modes, some of which are
446
+ closely overlapped [60]. However, it may be feasible to increase the mode finesses by increasing
447
+ the mirror reflectivity and using an ONF with lower transmission loss than the one used (the
448
+ estimated loss of Mode 4, the highest finesse in Fig. 4(a), was ∼ 20%). Nonetheless, the finesse
449
+ values of our ∼ 2 m long cavity with a HOM-ONF should be sufficient for cQED experiments
450
+ with narrow line-width emitters such as cold atoms.
451
+ Fig. 5. (a) Mode intensity profiles for quasi-donut-shaped cavity modes from the cavity
452
+ containing a HOM-ONF with their SOPs (top) and Stokes phases (bottom) similar to
453
+ the fiber eigenmodes of (i) HE𝑒
454
+ 21, (ii) HE𝑜
455
+ 21, (iii) TE01, and (iv) TM01. The red and
456
+ blue SOPs indicate right-handed and left-handed ellipticities, respectively. Scale bars
457
+ show intensity (from 0 to 1) and Stokes phase (from 0 to 2𝜋). Stokes singularities of
458
+ 𝜎12, 𝜎23, and 𝜎31 are indicated as pink, orange, and blue dots, respectively. L-lines are
459
+ illustrated as green lines. (b) Corresponding simulated results.
460
+ 3.3.
461
+ In situ higher-order cavity mode tuning
462
+ A key feature of this setup is the ability to tune the spectrum and SOP to create the desired mode
463
+ in the cavity. We aimed to observe modes with donut-shaped intensity patterns and SOPs similar
464
+ to the fiber eigenmodes TE01 (Fig. 2(a)), TM01, HE𝑜
465
+ 21, and HE𝑒
466
+ 21 (Fig. 2(b)). To achieve this, the
467
+ laser was locked to a well-resolved lobe-shaped mode. The paddle angles of the IPC were then
468
+ adjusted, and the mode shape was monitored with a CCD camera until a donut mode profile was
469
+ observed. Unlocking and scanning the laser revealed a new spectrum with each mode containing
470
+
471
+ (i)
472
+ (ii)
473
+ (iii)
474
+ (iv)
475
+ (D)
476
+ (iv)
477
+ D
478
+ Da new profile. The IPC was adjusted again to maximize another mode and the laser was locked to
479
+ this new mode. The IPC paddle angles were tuned to once more convert the mode profile to a
480
+ donut shape. This procedure was repeated for four different modes, see Figs. 5(a)(i-iv), and these
481
+ modes look similar to the true fiber eigenmodes of HE𝑒
482
+ 11 (Fig. 2(b)), HE𝑜
483
+ 11, TE01 (Fig. 2(a)), and
484
+ TM01, respectively. There was a slight deformation from a perfect donut shape and their SOPs
485
+ were not vector fields, but rather ellipse fields with alternating regions of opposite handiness.
486
+ While the donut eigenmodes possessed a V-point at the center as indicated by pink dots in Figs.
487
+ 2(a, b), the observed quasi-donut modes in Figs. 5(a)(i-iv) had some nominal intensity at the
488
+ center. These modes had two C-points of 𝜎12 = -1 or +1 near the center (see pink dots in Figs.
489
+ 5 (a)(i-iv)), as opposed to a single point of 𝜎12 = -2 or +2 in the true eigenmodes (Figs. 2(a,
490
+ b)). Indeed, perturbation of vector field polarization singularities can occur when scalar linearly
491
+ polarized beams are interfered [63].
492
+ These donut-shaped cavity modes were also simulated, as shown in Figs. 5(b)(i-iv). To
493
+ obtain a good fit for the experimentally observed intensities, SOPs, and Stokes phases in Figs.
494
+ 5(a)(i-iv), the simulated modes included a slight deformation of the donut shape by adding some
495
+ components of the HE11 mode to modes in the LP11 group. Moreover, the simulated results
496
+ show that the Stokes phases are very similar to those obtained experimentally. The number of
497
+ possible combinations of modal interference with varying birefringence is large and this leads
498
+ to discrepancies between the experiment and simulation. However, these findings indicate that
499
+ the experimentally observed quasi-donut modes are likely the result of residual interference
500
+ between the HE11 mode and modes in the LP11 group. Degeneracy of multiple modes may be
501
+ avoided by increasing the cavity mode finesses so that each mode can be well separated. The
502
+ system demonstrated here shows that, even in a complex system, the HOMs and their SOPs can
503
+ be controlled to create exotic topological states.
504
+ 4.
505
+ Conclusion
506
+ We have experimentally demonstrated a Fabry-Pérot fiber cavity with a HOM-ONF and performed
507
+ cavity spectroscopy. The cavity mode profiles and transverse polarization topology were also
508
+ determined by imaging and analyzing the individual cavity modes at the output. These modes
509
+ had inhomogeneous polarization distributions with a number of Stokes singularities. We also
510
+ simulated the fiber modes which closely match those observed at the output of the cavity.
511
+ Moreover, in situ intracavity manipulation of the modal birefringence and interference to select
512
+ a specific mode of interest was demonstrated. This indicates that the evanescent field of an
513
+ HON-ONF could be tuned by adjusting the IPC paddle angles.
514
+ These findings are a step toward investigating the interactions between SAM and OAM of
515
+ a HOM-ONF. Research into the interference of HOMs at the waist of an ONF is an exciting
516
+ opportunity to uncover the nature of light-matter interactions in tightly confining geometries
517
+ with topological singularities. Additionally, the realization of a (de)multiplexing system using
518
+ degenerate HOMs in an ONF-based cavity may be possible by improving the tunability of the
519
+ modal birefringence and interference. Such a system is attractive for future quantum information
520
+ platforms as efficient and secure storage.
521
+ The interference of higher-order cavity modes with fixed ratios in the evanescent field of an
522
+ ONF may also be used to trap and manipulate cold atoms. Adjusting the overlap and SOP of
523
+ the HOMs should result in movement of the trapping sites relative to each other, enabling some
524
+ trap dynamics to be studied [4,15,16]. This cavity could be also used with quantum emitters
525
+ to study multimode cQED effects using degenerate HOMs. The HOM cavity studied here had
526
+ moderate finesse to enter the cQED experiments for interactions with cold atoms. In free-space
527
+ optics, strong coupling of multiple transverse HOMs with atoms has been achieved [38], whereas
528
+ this has not been achieved using an ONF-type cavity. Our work is a significant step towards this
529
+ realization.
530
+
531
+ Moreover, the ability of our cavity to generate all three types of Stokes singularities may be
532
+ useful to realize not only a C-point laser but also an all-Stokes singularity laser using a few-mode
533
+ fiber. The combinations of fiber modes that we used in the simulations were found via manual
534
+ trial-and-error estimates to obtain a visual match with the experimentally observed modes. More
535
+ accurate control could be achieved by using machine learning techniques to fully cover the
536
+ parameter space of permitted modes in the cavity. This may enable us to determine the correct
537
+ combination of modes that lead to the observed cavity outputs and facilitate feedback to optimize
538
+ the input to the system to generate desired modes in the cavity.
539
+ Funding.
540
+ Okinawa Institute of Science and Technology Graduate University.
541
+ Acknowledgments.
542
+ The authors acknowledge F. Le Kien, L. Ruks, V. G. Truong, and J. M. Ward for
543
+ discussions and K. Karlsson for technical assistance.
544
+ Disclosures.
545
+ The authors declare no conflicts of interest.
546
+ Data availability.
547
+ Data underlying the results presented in this paper are not publicly available at this
548
+ time but may be obtained from the authors upon reasonable request.
549
+ References
550
+ 1.
551
+ K. Y. Bliokh and F. Nori, “Transverse and longitudinal angular momenta of light,” Phys. Reports 592, 1–38 (2015).
552
+ 2.
553
+ P. Solano, J. A. Grover, J. E. Hoffman, S. Ravets, F. K. Fatemi, L. A. Orozco, and S. L. Rolston, “Chapter seven -
554
+ optical nanofibers: A new platform for quantum optics,” in Advances In Atomic, Molecular, and Optical Physics,
555
+ vol. 66 E. Arimondo, C. C. Lin, and S. F. Yelin, eds. (Academic Press, 2017), pp. 439–505.
556
+ 3.
557
+ M. C. Frawley, A. Petcu-Colan, V. G. Truong, and S. Nic Chormaic, “Higher order mode propagation in an optical
558
+ nanofiber,” Opt. Commun. 285, 4648–4654 (2012).
559
+ 4.
560
+ C. Phelan, T. Hennessy, and T. Busch, “Shaping the evanescent field of optical nanofibers for cold atom trapping,”
561
+ Opt. Express 21, 27093 (2013).
562
+ 5.
563
+ F. L. Kien, T. Busch, V. G. Truong, and S. Nic Chormaic, “Higher-order modes of vacuum-clad ultrathin optical
564
+ fibers,” Phys. Rev. A 96, 023835 (2017).
565
+ 6.
566
+ F. L. Kien, S. S. S. Hejazi, T. Busch, V. G. Truong, and S. Nic Chormaic, “Channeling of spontaneous emission from
567
+ an atom into the fundamental and higher-order modes of a vacuum-clad ultrathin optical fiber,” Phys. Rev. A 96,
568
+ 043859 (2017).
569
+ 7.
570
+ F. L. Kien, S. S. S. Hejazi, V. G. Truong, S. Nic Chormaic, and T. Busch, “Chiral force of guided light on an atom,”
571
+ Phys. Rev. A 97, 063849 (2018).
572
+ 8.
573
+ F. L. Kien, D. F. Kornovan, S. S. S. Hejazi, V. G. Truong, M. I. Petrov, S. Nic Chormaic, and T. Busch, “Force of
574
+ light on a two-level atom near an ultrathin optical fiber,” New J. Phys. 20, 093031 (2018).
575
+ 9.
576
+ E. Stourm, M. Lepers, J. Robert, S. Nic Chormaic, K. Mølmer, and E. Brion, “Spontaneous emission and energy
577
+ shifts of a Rydberg rubidium atom close to an optical nanofiber,” Phys. Rev. A 101, 052508 (2020).
578
+ 10. F. L. Kien, S. Nic Chormaic, and T. Busch, “Transfer of angular momentum of guided light to an atom with an
579
+ electric quadrupole transition near an optical nanofiber,” Phys. Rev. A 106, 013712 (2022).
580
+ 11. J. E. Hoffman, F. K. Fatemi, G. Beadie, S. L. Rolston, and L. A. Orozco, “Rayleigh scattering in an optical nanofiber
581
+ as a probe of higher-order mode propagation,” Optica 2, 416 (2015).
582
+ 12. F. K. Fatemi, J. E. Hoffman, P. Solano, E. F. Fenton, G. Beadie, S. L. Rolston, and L. A. Orozco, “Modal interference
583
+ in optical nanofibers for sub-angstrom radius sensitivity,” Optica 4, 157 (2017).
584
+ 13. M. F. Picardi, K. Y. Bliokh, F. J. Rodríguez-Fortuño, F. Alpeggiani, and F. Nori, “Angular momenta, helicity, and
585
+ other properties of dielectric-fiber and metallic-wire modes,” Optica 5, 1016 (2018).
586
+ 14. F. L. Kien and T. Busch, “Torque of guided light on an atom near an optical nanofiber,” Opt. Express 27, 15046
587
+ (2019).
588
+ 15. G. Sagué, A. Baade, and A. Rauschenbeutel, “Blue-detuned evanescent field surface traps for neutral atoms based on
589
+ mode interference in ultrathin optical fibres,” New J. Phys. 10 (2008).
590
+ 16. M. Sadgrove, S. Wimberger, and S. Nic Chormaic, “Quantum coherent tractor beam effect for atoms trapped near a
591
+ nanowaveguide,” Sci. Reports 6, 28905 (2016).
592
+ 17. A. Petcu-Colan, M. Frawey, and S. Nic Chormaic, “Tapered few-mode fibers: Mode evolution during fabrication and
593
+ adiabaticity,” J. Nonlinear Opt. Phys. & Mater. 20, 293–307 (2011).
594
+ 18. J. M. Ward, A. Maimaiti, V. H. Le, and S. N. Chormaic, “Contributed review: Optical micro- and nanofiber pulling
595
+ rig,” Rev. Sci. Instruments 85, 111501 (2014).
596
+ 19. A. Maimaiti, V. G. Truong, M. Sergides, I. Gusachenko, and S. Nic Chormaic, “Higher order microfibre modes for
597
+ dielectric particle trapping and propulsion,” Sci. Reports 5, 9077 (2015).
598
+ 20. A. Maimaiti, D. Holzmann, V. G. Truong, H. Ritsch, and S. Nic Chormaic, “Nonlinear force dependence on optically
599
+ bound micro-particle arrays in the evanescent fields of fundamental and higher order microfibre modes,” Sci. Reports
600
+ 6, 30131 (2016).
601
+
602
+ 21. R. Kumar, V. Gokhroo, K. Deasy, A. Maimaiti, M. C. Frawley, C. Phelan, and S. Nic Chormaic, “Interaction of
603
+ laser-cooled 87Rb atoms with higher order modes of an optical nanofibre,” New J. Phys. 17, 013026 (2015).
604
+ 22. F. L. Kien and K. Hakuta, “Cavity-enhanced channeling of emission from an atom into a nanofiber,” Phys. Rev. A 80,
605
+ 053826 (2009).
606
+ 23. K. P. Nayak, M. Sadgrove, R. Yalla, F. L. Kien, and K. Hakuta, “Nanofiber quantum photonics,” J. Opt. 20, 073001
607
+ (2018).
608
+ 24. P. Romagnoli, M. Maeda, J. M. Ward, V. G. Truong, and S. Nic Chormaic, “Fabrication of optical nanofibre-based
609
+ cavities using focussed ion-beam milling: a review,” Appl. Phys. B 126, 111 (2020).
610
+ 25. J. Keloth, K. P. Nayak, and K. Hakuta, “Fabrication of a centimeter-long cavity on a nanofiber for cavity QED,” Opt.
611
+ Lett. 42, 1003–1006 (2017).
612
+ 26. W. Li, J. Du, V. G. Truong, and S. Nic Chormaic, “Optical nanofiber-based cavity induced by periodic air-nanohole
613
+ arrays,” Appl. Phys. Lett. 110, 253102 (2017).
614
+ 27. W. Li, J. Du, and S. Nic Chormaic, “Tailoring a nanofiber for enhanced photon emission and coupling efficiency
615
+ from single quantum emitters,” Opt. Lett. 43, 1674–1677 (2018).
616
+ 28. T. Tashima, H. Takashima, and S. Takeuchi, “Direct optical excitation of an NV center via a nanofiber Bragg-cavity:
617
+ a theoretical simulation,” Opt. Express 27, 27009 (2019).
618
+ 29. S. K. Ruddell, K. E. Webb, M. Takahata, S. Kato, and T. Aoki, “Ultra-low-loss nanofiber fabry–perot cavities
619
+ optimized for cavity quantum electrodynamics,” Opt. Lett. 45, 4875–4878 (2020).
620
+ 30. Z. Li, X. Li, and X. Zhong, “Strong photon blockade in an all-fiber emitter-cavity quantum electrodynamics system,”
621
+ Phys. Rev. A 103, 043724 (2021).
622
+ 31. R. Yalla, M. Sadgrove, K. P. Nayak, and K. Hakuta, “Cavity quantum electrodynamics on a nanofiber using a
623
+ composite photonic crystal cavity,” Phys. Rev. Lett. 113, 143601 (2014).
624
+ 32. D. H. White, S. Kato, N. Német, S. Parkins, and T. Aoki, “Cavity dark mode of distant coupled atom-cavity systems,”
625
+ Phys. Rev. Lett. 122, 253603 (2019).
626
+ 33. T. Tashima, H. Takashima, A. W. Schell, T. T. Tran, I. Aharonovich, and S. Takeuchi, “Hybrid device of hexagonal
627
+ boron nitride nanoflakes with defect centres and a nano-fibre Bragg cavity,” Sci. Reports 12, 96 (2022).
628
+ 34. S. Kato and T. Aoki, “Strong coupling between a trapped single atom and an all-fiber cavity,” Phys. Rev. Lett. 115,
629
+ 093603 (2015).
630
+ 35. S. K. Ruddell, K. E. Webb, I. Herrera, A. S. Parkins, and M. D. Hoogerland, “Collective strong coupling of cold
631
+ atoms to an all-fiber ring cavity,” Optica 4, 576 (2017).
632
+ 36. A. Johnson, M. Blaha, A. E. Ulanov, A. Rauschenbeutel, P. Schneeweiss, and J. Volz, “Observation of collective
633
+ superstrong coupling of cold atoms to a 30-m long optical resonator,” Phys. Rev. Lett. 123, 243602 (2019).
634
+ 37. T. Salzburger, P. Domokos, and H. Ritsch, “Enhanced atom capturing in a high-Q cavity by help of several transverse
635
+ modes,” Opt. Express 10, 1204 (2002).
636
+ 38. A. Wickenbrock, M. Hemmerling, G. R. M. Robb, C. Emary, and F. Renzoni, “Collective strong coupling in
637
+ multimode cavity QED,” Phys. Rev. A 87, 043817 (2013).
638
+ 39. D. Mao, Z. He, H. Lu, M. Li, W. Zhang, X. Cui, B. Jiang, and J. Zhao, “All-fiber radially/azimuthally polarized lasers
639
+ based on mode coupling of tapered fibers,” Opt. Lett. 43, 1590–1593 (2018).
640
+ 40. A. Jöckel, “Glasfaser-basierte Fabry-Pérot-Resonatoren mit integrierten ultradünnen Passagen,” M.S. thesis, Institut
641
+ für Physik der Johannes Gutenberg-Universität Mainz (2009).
642
+ 41. Q. Wang, C.-H. Tu, Y.-N. Li, and H.-T. Wang, “Polarization singularities: Progress, fundamental physics, and
643
+ prospects,” APL Photonics 6, 040901 (2021).
644
+ 42. I. Freund, “Polarization singularity indices in gaussian laser beams,” Opt. Commun. 201, 251–270 (2002).
645
+ 43. G. Arora and P. Senthilkumaran, “Generation of Stokes singularities using polarization lateral shear interferometer,”
646
+ Opt. Express 30, 27583 (2022).
647
+ 44. I. Freund, “Poincaré vortices,” Opt. Lett. 26, 1996 (2001).
648
+ 45. I. Freund, A. I. Mokhun, M. S. Soskin, O. V. Angelsky, and I. I. Mokhun, “Stokes singularity relations,” Opt. Lett.
649
+ 27, 545 (2002).
650
+ 46. G. Arora, Ruchi, and P. Senthilkumaran, “Full Poincaré beam with all the Stokes vortices,” Opt. Lett. 44, 5638 (2019).
651
+ 47. F. Flossmann, U. T. Schwarz, M. Maier, and M. R. Dennis, “Polarization singularities from unfolding an optical
652
+ vortex through a birefringent crystal,” Phys. Rev. Lett. 95, 253901 (2005).
653
+ 48. F. Flossmann, U. T. Schwarz, M. Maier, and M. R. Dennis, “Stokes parameters in the unfolding of an optical vortex
654
+ through a birefringent crystal,” Opt. Express 14, 11402–11411 (2006).
655
+ 49. K. Y. Bliokh, A. Niv, V. Kleiner, and E. Hasman, “Singular polarimetry: Evolution of polarization singularities in
656
+ electromagnetic waves propagating in weakly anisotropic medium,” Opt. Express 16, 695–709 (2008).
657
+ 50. S. K. Pal, Ruchi, and P. Senthilkumaran, “Polarization singularity index sign inversion by a half-wave plate,” Appl.
658
+ Opt. 56, 6181 (2017).
659
+ 51. D. Pohl, “Operation of a ruby laser in the purely transverse electric mode TE01,” Appl. Phys. Lett. 20, 266–267
660
+ (1971).
661
+ 52. K. Yonezawa, Y. Kozawa, and S. Sato, “Generation of a radially polarized laser beam by use of the birefringence of a
662
+ c-cut Nd:YVO4 crystal,” Opt. Lett. 31, 2151 (2006).
663
+ 53. Y. Jiang, G. Ren, Y. Shen, Y. Xu, W. Jin, Y. Wu, W. Jian, and S. Jian, “Two-dimensional tunable orbital angular
664
+ momentum generation using a vortex fiber,” Opt. Lett. 42, 5014 (2017).
665
+
666
+ 54. Y. V. Jayasurya, V. V. G. K. Inavalli, and N. K. Viswanathan, “Polarization singularities in the two-mode optical fiber
667
+ output,” Appl. Opt. 50, E131 (2011).
668
+ 55. C. H. Krishna and S. Roy, “Polarization singular patterns in modal fields of few-mode optical fiber,” J. Opt. Soc. Am.
669
+ B 37, 2688 (2020).
670
+ 56. F. L. Kien, K. Nayak, and K. Hakuta, “Nanofibers with Bragg gratings from equidistant holes,” J. Mod. Opt. 59,
671
+ 274–286 (2012).
672
+ 57. J. Keloth, K. P. Nayak, J. Wang, M. Morinaga, and K. Hakuta, “Coherent interaction of orthogonal polarization
673
+ modes in a photonic crystal nanofiber cavity,” Opt. Express 27, 1453 (2019).
674
+ 58. E. Otte, C. Alpmann, and C. Denz, “Polarization singularity explosions in tailored light fields,” Laser & Photonics
675
+ Rev. 12, 1700200 (2018).
676
+ 59. J. P. Gordon and H. Kogelnik, “PMD fundamentals: Polarization mode dispersion in optical fibers,” PNAS 97,
677
+ 4541–4550 (2000).
678
+ 60. K. Kolluru, S. Saha, and S. D. Gupta, “Cavity enhanced interference of orthogonal modes in a birefringent medium,”
679
+ Opt. Commun. 410, 836–840 (2018).
680
+ 61. S. Vyas, Y. Kozawa, and S. Sato, “Polarization singularities in superposition of vector beams,” Opt. Express 21, 8972
681
+ (2013).
682
+ 62. F. Lei, G. Tkachenko, J. M. Ward, and S. Nic Chormaic, “Complete polarization control for a nanofiber waveguide
683
+ using directional coupling,” Phys. Rev. Appl. 11, 064041 (2019).
684
+ 63. G. Arora, S. Joshi, H. Singh, V. Haridas, and P. Senthilkumaran, “Perturbation of V-point polarization singular vector
685
+ beams,” Opt. & Laser Technol. 158, 108842 (2023).
686
+
8dFQT4oBgHgl3EQf4jbF/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
ANE0T4oBgHgl3EQfPgBb/content/tmp_files/2301.02179v1.pdf.txt ADDED
@@ -0,0 +1,1237 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Draft version January 6, 2023
2
+ Typeset using LATEX twocolumn style in AASTeX63
3
+ The GLASS-JWST Early Release Science Program. II. Stage I release of NIRCam imaging and
4
+ catalogs in the Abell 2744 region.
5
+ Diego Paris
6
+ ,1 Emiliano Merlin
7
+ ,1 Adriano Fontana
8
+ ,1 Andrea Bonchi
9
+ ,2, 1 Gabriel Brammer
10
+ ,3, 4
11
+ Matteo Correnti,2, 1 Tommaso Treu
12
+ ,5 Kristan Boyett
13
+ ,6, 7 Antonello Calabr`o
14
+ ,1 Marco Castellano
15
+ ,1
16
+ Wenlei Chen
17
+ ,8 Lilan Yang
18
+ ,9 K. Glazebrook
19
+ ,10 Patrick Kelly
20
+ ,8 Anton M. Koekemoer
21
+ ,11
22
+ Nicha Leethochawalit
23
+ ,12 Sara Mascia
24
+ ,1 Charlotte Mason
25
+ ,3, 4 Takahiro Morishita
26
+ ,13
27
+ Mario Nonino
28
+ ,14 Laura Pentericci
29
+ ,1 Gianluca Polenta
30
+ ,2 Guido Roberts-Borsani
31
+ ,5 Paola Santini
32
+ ,1
33
+ Michele Trenti
34
+ ,6, 7 Eros Vanzella
35
+ ,15 Benedetta Vulcani
36
+ ,16 Rogier A. Windhorst
37
+ ,17
38
+ Themiya Nanayakkara
39
+ ,10 and Xin Wang
40
+ 18, 19, 20
41
+ 1INAF Osservatorio Astronomico di Roma, Via Frascati 33, 00078 Monteporzio Catone, Rome, Italy
42
+ 2Space Science Data Center, Italian Space Agency, via del Politecnico, 00133, Roma, Italy
43
+ 3Cosmic Dawn Center (DAWN), Denmark
44
+ 4Niels Bohr Institute, University of Copenhagen, Jagtvej 128, DK-2200 Copenhagen N, Denmark
45
+ 5Department of Physics and Astronomy, University of California, Los Angeles, 430 Portola Plaza, Los Angeles, CA 90095, USA
46
+ 6School of Physics, University of Melbourne, Parkville 3010, VIC, Australia
47
+ 7ARC Centre of Excellence for All Sky Astrophysics in 3 Dimensions (ASTRO 3D), Australia
48
+ 8Minnesota Institute for Astrophysics, University of Minnesota, 116 Church Street SE, Minneapolis, MN 55455, USA
49
+ 9Kavli Institute for the Physics and Mathematics of the Universe, The University of Tokyo, Kashiwa, Japan 277-8583
50
+ 10Centre for Astrophysics and Supercomputing, Swinburne University of Technology, PO Box 218, Hawthorn, VIC 3122, Australia
51
+ 11Space Telescope Science Institute, 3700 San Martin Dr., Baltimore, MD 21218, USA
52
+ 12National Astronomical Research Institute of Thailand (NARIT), Mae Rim, Chiang Mai, 50180, Thailand
53
+ 13IPAC, California Institute of Technology, MC 314-6, 1200 E. California Boulevard, Pasadena, CA 91125, USA
54
+ 14(INAF - Osservatorio Astronomico di Trieste, Via Tiepolo 11, I-34131 Trieste, Italy)
55
+ 15INAF – OAS, Osservatorio di Astrofisica e Scienza dello Spazio di Bologna, via Gobetti 93/3, I-40129 Bologna, Italy
56
+ 16INAF Osservatorio Astronomico di Padova, vicolo dell’Osservatorio 5, 35122 Padova, Italy
57
+ 17School of Earth and Space Exploration, Arizona State University, Tempe, AZ 85287-1404, USA
58
+ 18School of Astronomy and Space Science, University of Chinese Academy of Sciences (UCAS), Beijing 100049, China
59
+ 19National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100101, China
60
+ 20Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing 102206, China
61
+ ABSTRACT
62
+ We present images and a multi–wavelength photometric catalog based on all of the JWST NIRCam
63
+ observations obtained to date in the region of the Abell 2744 galaxy cluster. These data come from
64
+ three different programs, namely the GLASS-JWST Early Release Science Program, UNCOVER, and
65
+ Director’s Discretionary Time program 2756. The observed area in the NIRCam wide-band filters -
66
+ covering the central and extended regions of the cluster, as well as new parallel fields - is 46.5 arcmin2 in
67
+ total. All images in eight bands (F090W, F115W, F150W, F200W, F277W, F356W, F410M, F444W)
68
+ have been reduced adopting the latest calibration and references files available to date. Data reduction
69
+ has been performed using an augmented version of the official JWST pipeline, with improvements
70
+ aimed at removing or mitigating defects in the raw images and improve the background subtraction
71
+ and photometric accuracy. We obtain a F444W-detected multi–band catalog including all NIRCam
72
+ data and available HST data, adopting forced aperture photometry on PSF-matched images. The
73
+ catalog is intended to enable early scientific investigations, and is optimized for the study of faint
74
+ galaxies; it contains 24389 sources, with a 5σ limiting magnitude in the F444W band ranging from
75
+ 28.5 AB to 30.5 AB, as a result of the varying exposure times of the surveys that observed the field. We
76
+ Corresponding author: Diego Paris
77
+ diego.paris@inaf.it
78
+ arXiv:2301.02179v1 [astro-ph.GA] 5 Jan 2023
79
+
80
+ ID2
81
+ Paris et al.
82
+ publicly release the reduced NIRCam images, associated multi-wavelength catalog, and code adopted
83
+ for 1/f noise removal with the aim of aiding users to familiarize themselves with JWST NIRCam data
84
+ and identify suitable targets for follow-up observations.
85
+ Keywords: galaxies: high-redshift, galaxies: photometry
86
+ 1. INTRODUCTION
87
+ In just a few months of observations, JWST has
88
+ demonstrated its revolutionary scientific capabilites.
89
+ Early observations have shown that its performance is
90
+ equal or better than expected, with image quality and
91
+ overall efficiency that match or surpass pre-launch esti-
92
+ mates (Rigby et al. 2022). Publicly available datasets
93
+ obtained by the Early Release Observations and Early
94
+ Release Science programs have already enabled a large
95
+ number of publications based on JWST data, ranging
96
+ from exoplanets to the distant Universe.
97
+ In particular, a number of works exploited the power
98
+ of NIRCAM to gather the first sizeable sample of can-
99
+ didates at z ≥ 10 (e.g., Castellano et al. 2022a; Don-
100
+ nan et al. 2023; Finkelstein et al. 2022; Morishita & Sti-
101
+ avelli 2022; Naidu et al. 2022; Yan et al. 2022; Roberts-
102
+ Borsani et al. 2022; Robertson et al. 2022; Castellano
103
+ et al. 2022b; Bouwens et al. 2022), showing the power of
104
+ JWST in exploring the Universe during the re-ionization
105
+ epoch.
106
+ In this paper we present the full data set obtained
107
+ with NIRCam in the region of of the z = 0.308 cluster
108
+ Abell 2744 that will significantly expands the available
109
+ area for deep extragalactic observations.
110
+ The central
111
+ region of the cluster, with the assistance of lensing mag-
112
+ nification, allows an insight into the distant Universe
113
+ at depth and resolution superior of those of NIRCam
114
+ in blank fields.
115
+ The data set analyzed here are ob-
116
+ tained through three public programs: i) GLASS-JWST
117
+ ERS (Treu et al. 2022), ii) UNCOVER (Bezanson et al.
118
+ 2022), and iii) Director’s Discretionary Time Program
119
+ 2756, aimed at following up a Supernova discovered in
120
+ GLASS-JWST NIRISS imaging. We have analyzed and
121
+ combined the imaging data of all these programs and
122
+ obtained a multi-wavelength catalog of the objects de-
123
+ tected in the F444W band.
124
+ In order to facilitate exploitation of these data, we
125
+ release reduced images and associated catalog on our
126
+ website and through the Mikulski Archives for Space
127
+ Telescopes (MAST). This release fulfills and exceeds the
128
+ requirements of the Stage I data release planned as part
129
+ of the GLASS-JWST program.
130
+ It is anticipated that
131
+ a final (Stage II) release will follow in approximately
132
+ a year, combining additional images scheduled in 2023,
133
+ and taking advantage of future improvements in data
134
+ processing and calibrations.
135
+ This paper is organized as follows. In Section 2 we
136
+ present the data-set and discuss the image processing
137
+ pipeline. In Section 3 the methods applied for the de-
138
+ tection of the sources and the photometric techniques
139
+ used to compute the fluxes are presented.
140
+ Finally in
141
+ Section 4 we summarize the results.
142
+ Throughout the
143
+ paper we adopt AB magnitudes (Oke & Gunn 1983).
144
+ 2. DATA REDUCTION
145
+ 2.1. Data Set
146
+ The NIRCam data analyzed in this paper are taken
147
+ from three programs that targeted the z = 0.308 clus-
148
+ ter Abell 2744 (A2744 hereafter) and its surroundings.
149
+ The first set of NIRCam images were taken as part of
150
+ the GLASS-JWST survey (Treu et al. 2022, hereafter
151
+ T22), in parallel to primary NIRISS observations on
152
+ June 28–29 2022 and to NIRSpec observations on Nov.
153
+ 10–11, 2022. We refer to these data sets as GLASS1 and
154
+ GLASS2, or collectively as GLASS, both of which con-
155
+ sist of imaging in seven broad-band filters from F090W
156
+ to F444W (see Treu et al. 2022 for details). We note that
157
+ the final pointing is different from the scheduled one pre-
158
+ sented by Treu et al. (2022) due to the adoption of an
159
+ alternate position angle (PA) during the NIRSpec spec-
160
+ troscopic observations.
161
+ As the primary spectroscopic
162
+ target was the A2744 cluster, these parallel images are
163
+ offset to the North-West. By virtue of the long exposure
164
+ times, these images are the deepest presented here.
165
+ The second set of NIRCam observations considered
166
+ here were taken as part of the UNCOVER program
167
+ (Bezanson et al. 2022), which targets the center of the
168
+ A2744 cluster and the immediate surroundings. These
169
+ images are composed of four pointings and result in a
170
+ relatively homogeneous depth, as discussed below. They
171
+ were taken on November 2-4-7 and 15, and adopt the
172
+ same filter set as GLASS-JWST, except for the adop-
173
+ tion of the F410M filter instead of F090W.
174
+ Finally, NIRCam imaging of the A2744 center was
175
+ also obtained as part of DDT program 2756 (PI W.
176
+ Chen, DDT hereafter) on October 20 and December
177
+ 6 2022 (UT). These two data sets are dubbed DDT1
178
+ and DDT2 hereafter. The DDT filter set is the same
179
+ as GLASS-JWST with the exception of the F090W fil-
180
+ ter, and overall shorter exposure times. One of the two
181
+ NIRCam modules overlaps with UNCOVER.
182
+
183
+ GLASS-JWST: Abell 2744 NIRCam photometric catalog
184
+ 3
185
+ Figure 1. Full view of the F444W mosaics. Colored boxes show the position of the three different data sets used here: GLASS
186
+ (green), UNCOVER (blue) and DDT (red). The entire image (including the empty space) is approximately 12.7 × 10.9 arc
187
+ minutes wide.
188
+ In Table 1 we list the exposure times adopted in the
189
+ various filters for each of the aforementioned programs,
190
+ while the footprints of the fields are illustrated in Fig-
191
+ ure 1.
192
+ As a result of the overlap between programs and of
193
+ their different observation strategies, the resulting ex-
194
+ posure map is complex and inhomogenous across bands
195
+ and area. An analysis of the depth resulting from this
196
+ exposure map is reported below.
197
+ 2.2. Data reduction
198
+ 2.2.1. Pre-reduction steps
199
+ Image pre-reduction was executed using the official
200
+ JWST calibration pipeline, provided by the Space Tele-
201
+ scope Science Institute (STScI) as a Python software
202
+ suite1. We adopted Version 1.8.2 of the pipeline and Ver-
203
+ sions between cjwst 1014.pmap and cjwst 1019.pmap
204
+ of the CRDS files (the only changes between these
205
+ versions is the astrometric calibration, that is dealt
206
+ with as described below).
207
+ We executed the first two
208
+ stages of the pipeline (i.e.
209
+ calwebb detector1 and
210
+ 1 https://jwst-pipeline.readthedocs.io/en/latest/jwst/
211
+ introduction.html
212
+ Table 1. NIRCam Exposure time
213
+ Filter
214
+ GLASS1
215
+ GLASS2
216
+ DDT1/2
217
+ UNCOVER
218
+ F090W
219
+ 11520
220
+ 16492
221
+ -
222
+ -
223
+ F115W
224
+ 11520
225
+ 16492
226
+ 2104
227
+ 10822
228
+ F150W
229
+ 6120
230
+ 8246
231
+ 2104
232
+ 10822
233
+ F200W
234
+ 5400
235
+ 8246
236
+ 2104
237
+ 6700
238
+ F277W
239
+ 5400
240
+ 8246
241
+ 2104
242
+ 6700
243
+ F356W
244
+ 6120
245
+ 8246
246
+ 2104
247
+ 6700
248
+ F410M
249
+ -
250
+ -
251
+ -
252
+ 6700
253
+ F444W
254
+ 23400
255
+ 32983
256
+ 2104
257
+ 8246
258
+ Note—Exposure time (in seconds) for each pointing of the
259
+ three programs considered here.
260
+ calwebb image2), adopting the optimized parameters
261
+ for the NIRCam imaging mode, that convert single de-
262
+ tector raw images into photometric calibrated images.
263
+ Using the first pipeline stage calwebb detector1 we
264
+ processed the raw uncalibrated data (uncal.fits) in
265
+ order to apply detector-level corrections performed on
266
+ a group-by-group basis, as dark subtractions, reference
267
+ pixels corrections, non-linearity corrections and jump
268
+
269
+ 4
270
+ Paris et al.
271
+ detection that allows to identify cosmic rays (CR) events
272
+ on the single groups. The last step of this pipeline stage
273
+ allows us to derive the mean count rate, in units of
274
+ counts per second, for each pixel by performing a lin-
275
+ ear fit to the data in the input image (the so-called
276
+ ramp-fitting) excluding the group masked due to the
277
+ identification of a cosmic ray jump.
278
+ The output files of the previous steps (rate.fits)
279
+ are
280
+ processed
281
+ through
282
+ the
283
+ second
284
+ pipeline
285
+ stage
286
+ calwebb image2,
287
+ which
288
+ consists
289
+ of
290
+ additional
291
+ instrument-level and observing-mode corrections and
292
+ calibrations, as the geometric-distortion correction, the
293
+ flat-fielding, and the photometric calibrations that con-
294
+ verting the data from units of countrate to surface
295
+ brightness (i.e.
296
+ MJy per steradian) generates a fully
297
+ calibrated exposure (cal.fits).
298
+ The cal.fits file contains also an RMS layer, which
299
+ combines the contribution of all pixel noise sources, and
300
+ a DQ mask where the first bit (DO NOT USE) identifies pix-
301
+ els that should not be used during the resampling phase.
302
+ We then applied a number of custom procedures to
303
+ remove instrumental defects that are not dealt with
304
+ the STScI pipeline. Some of them have already been
305
+ adopted in (Merlin et al. 2022, hereafter M22) and
306
+ described there:
307
+ we illustrate below only the major
308
+ changes to the STScI pipeline in default configuration
309
+ and/or to the procedure adopted in M22.
310
+ • “Snowballs”, i.e. circular artifacts observed in the
311
+ in-flight data caused by a large cosmic ray impacts.
312
+ Those hits leave a bright ring-shaped defect in the
313
+ image since the affected pixels are just partially
314
+ identified and masked. In M22, we developed a
315
+ technique to fully mask out these features, which
316
+ was not necessary here. Indeed, version 1.8.1 of
317
+ the JWST pipeline introduced the option to iden-
318
+ tify snowball events, expanding the typical mask-
319
+ ing area to include all the pixels affected.
320
+ This
321
+ new implementation provides the opportunity to
322
+ correct these artifacts directly at the ramp fitting
323
+ stage, at the cost of a larger noise on the corre-
324
+ sponding pixels. We activated this non-default op-
325
+ tion, and fine tuned the corresponding parameters
326
+ to completely mask all the observed snowballs and,
327
+ at the same time, minimize the size of high noise
328
+ areas.
329
+ • “NL Mask”, on cal images of the NIRCam Mod-
330
+ ule B Long Wavelength detector are visible bright
331
+ groups of pixels not well corrected during prere-
332
+ duction. These pixels are more evident on deeper
333
+ pointing and are identified as “well not defined”
334
+ pixels2 in the Non Linearity Calibration file 3.
335
+ We selected those pixels and masked them as
336
+ DO NOT USE to not to be used during stacking
337
+ phase.
338
+ • 1/f noise, which introduces random vertical and
339
+ horizontal stripes into the images (see Schlawin
340
+ et al. 2020). We remove this by subtracting the
341
+ median value from each line/column, after mask-
342
+ ing out all objects and bad pixels.
343
+ The masks
344
+ were obtained by running SExtractor (Version
345
+ 2.25.0) (Bertin & Arnouts 1996) and then dilat-
346
+ ing the resulting segmentation image, applying a
347
+ differential procedure to dilate objects depending
348
+ on their ISOAREA: the segmentation of objects
349
+ with ISOAREA<5000 pixels was dilated using a
350
+ 3 × 3 convolution kernel and a dilation of 15 pix-
351
+ els, while for the segmentation of objects with
352
+ ISOAREA⩾5000 pixels a 9 × 9 convolution ker-
353
+ nel and a dilation of 4 × 15 pixels was used. The
354
+ procedure was executed separately for each am-
355
+ plifier in the SW detectors (i.e. 4 times for each
356
+ individual image) with the exception of the denser
357
+ areas corresponding to the centers of the clusters
358
+ and the brightest field star, where objects are sig-
359
+ nificantly larger than the amplifier width (500 pix-
360
+ els, corresponding to about 30”) and could not be
361
+ masked efficiently. In this case we removed the 1/f
362
+ noise over the entire row. As this extension of the
363
+ STScI pipeline could be useful for other programs,
364
+ we publicly release the code adopted for this step.
365
+ • Scattered light:
366
+ we identify additive features in
367
+ the F115W, F150W and F200W images.
368
+ These
369
+ low-surface brightness features have already been
370
+ revealed by commissioning data (see Rigby et al.
371
+ 2022) and are due to scattered light entering into
372
+ optical path. These anomalies have been dubbed
373
+ wisps or claws, depending on their origin and mor-
374
+ phology.
375
+ Wisps have a nearly constant shape
376
+ and a template pattern is available for subtraction
377
+ from the images. We removed these features by
378
+ extracting their 2D profile from the available tem-
379
+ plate (we do not use the entire template image to
380
+ avoid subtracting its empty but noisy regions) and
381
+ then normalizing the residual template to match
382
+ the feature intensity in each image. Claws have
383
+ been first identified and singled out in images.
384
+ 2 https://www.stsci.edu/files/live/sites/www/files/home/jwst/
385
+ documentation/technical-documents/ documents/JWST-STScI-
386
+ 004714.pdf
387
+ 3 https://jwst-crds.stsci.edu/browse/jwst nircam linearity 0011.rmap
388
+
389
+ GLASS-JWST: Abell 2744 NIRCam photometric catalog
390
+ 5
391
+ Figure 2. Examples of custom procedures to remove resid-
392
+ ual instrumental defects, not dealt with the current STScI
393
+ pipeline. Top: 1/f stripes removal on a GLASS F200W single
394
+ exposure. Bottom: A portion of the GLASS F150W mosaic
395
+ before and after the claws treatment.
396
+ Their shape on each image has been reconstructed
397
+ by interpolating a 2D mesh with box size 32 pix-
398
+ els and then eventually subtracted from the same
399
+ image. We find that these procedures efficiently
400
+ remove most of these features, as shown in Fig-
401
+ ure 2.
402
+ Other defects were found in the F090W image, and
403
+ to a lesser extent in the F115W one, which are
404
+ due to a so-called “wing-tilt event” that happened
405
+ during the observations. These defects have been
406
+ masked as in M22.
407
+ We then re-scaled the single exposures to units of
408
+ µJy/pixel, using the conversion factors output by the
409
+ pipeline.
410
+ 2.2.2. Astrometry
411
+ The astrometric calibration was performed using
412
+ SCAMP (Bertin 2006), with 3rd order distortion correc-
413
+ tions (PV coefficients up to j = 10). At variance with the
414
+ procedure we adopted in M22, we started from the dis-
415
+ tortion coefficient computed by the STScI pipeline and
416
+ stored in the cal images, and refine the astrometric so-
417
+ lution by running scamp in cal mode, which optimizes
418
+ the solution with limited variations from the starting so-
419
+ lution. We have found this procedure both accurate and
420
+ reliable, as described below. We first obtained a global
421
+ astrometric solution for the F444W image, which is usu-
422
+ ally the deepest, tied to a ground-based catalog obtained
423
+ in the i-band with the Magellan telescope in good see-
424
+ ing condition (see T22 for details) of the same region,
425
+ which had been previously aligned to GAIA-DR3 stars
426
+ (Gaia Collaboration et al. 2016, 2022 in prep.). We then
427
+ took the resulting high-resolution catalog in F444W as
428
+ reference for the other JWST bands, using compact, iso-
429
+ lated sources detected at high signal-to-noise at all wave-
430
+ lengths. Each NIRCam detector has been analysed inde-
431
+ pendently, in order to simplify the treatment of distor-
432
+ tions and minimise the offsets of the sources in different
433
+ exposures. Finally, we used SWarp (Bertin et al. 2002)
434
+ to combine the single exposures into mosaics projected
435
+ onto a common aligned grid of pixels, and SExtractor
436
+ to further clean the images by subtracting the residual
437
+ sky background. The pixel scale of all the images was
438
+ set to 0.031′′ (the approximate native value of the short
439
+ wavelength bands), to allow for simple processing with
440
+ photometric algorithms.
441
+ The final image, computed as a weighted stack of all
442
+ the images from the three programs, has a size of 24397×
443
+ 21040 pixels, corresponding to 12.6 × 10.87 arcmin2. In
444
+ this frame, the area covered by the wide-band NIRCam
445
+ images (F115W, F150W, F200W, F277W, F356W and
446
+ F444W) is of exactly 46.5 arcmin2. The F444W image
447
+ is shown in Figure 1.
448
+ Given the especially deep and sharp nature of the
449
+ JWST images, where most of the faint objects have sizes
450
+ below 0.5′′, the requirements on the final astrometric ac-
451
+ curacy are extremely tight, to avoid errors in the multi-
452
+ band photometry (where a displacement of as little as
453
+ 0.1′′ can bias color estimates). These requirements must
454
+ be met also in the overlapping regions of the various
455
+ surveys, which have often been observed with different
456
+ detectors.
457
+ To verify the final astrometric solution we conducted
458
+ a number of validation tests, where we compare the
459
+ positions of cross-matched objects in catalogues ex-
460
+ tracted from different images.
461
+ For each of these cat-
462
+ alogues we used SExtractor in single image mode
463
+ and adopted the XWIN and YWIN estimators of
464
+ the object center, which are more accurate than other
465
+ choices.
466
+ At the unprecedented image quality of NIR-
467
+ Cam, the accurate center of extra–galactic objects with
468
+ complex morphology may be difficult to estimate with
469
+ high accuracy, especially when observed across a large
470
+ wavelength interval.
471
+ To minimize errors, we limited
472
+ the comparison to objects with well defined positions,
473
+ using the ∆X, ∆Y
474
+ =ERRAWIN WORLD, ERRB-
475
+ WIN WORLD estimators of the error and limiting the
476
+ analysis to objects with (∆X2 + ∆Y 2)1/2 ≤ 0.018”.
477
+ From these catalogues we estimated both the average
478
+ offset of the object centers ∆α and ∆δ, and the median
479
+ average deviation madα and madδ, which measure the
480
+
481
+ 6
482
+ Paris et al.
483
+ Figure 3. Validation tests on the astrometric registration. Left: scatter diagram reporting the displacement δRA and δDEC
484
+ of sources between the Magellan i–band catalog registered to Gaia DR3 used as global reference for calibration and the final
485
+ F444W NIRCam catalog. Middle left: As above, applied to the scatter between the AstroDeep catalog and the final F444W
486
+ NIRCam catalog obtained on the central region of the A2744 cluster, as obtained in the context of the Frontier Fields initiative
487
+ (Merlin et al. 2016a). Middle right: Offset between the position of sources in the F444W and the F115W images. Right:
488
+ Positional offset between the objects detected in the UNCOVER–only images and those in the GLASS and DDT samples, on
489
+ two overlapping regions. In all diagrams the average value ∆α and ∆δ and the median average deviation mad∆α and mad∆δ
490
+ are reported.
491
+ intrinsic scatter in the alignment. In Figure 3 we report
492
+ the main outcome of these tests:
493
+ • (Left) We first compared the positions of objects
494
+ in the original Magellan i-band and the resulting
495
+ F444W of the entire mosaic. We find an essentially
496
+ zero offset and madα ≃ madδ ≃ 0.02”, which is 2/3
497
+ of a pixel.
498
+ • (Middle left) We compared the F444W catalog
499
+ with the AstroDeep H160 catalog obtained on the
500
+ central region of the A2744 cluster, as obtained in
501
+ the context of the Frontier Fields initiative (Mer-
502
+ lin et al. 2016a). While the intrinsic scatter is still
503
+ good (madα ≃ madδ ≃ 0.02”), we find a system-
504
+ atic offset by about 1 pixel in RA and 2.5 pixels in
505
+ DEC, which is most likely due to different choices
506
+ in the absolute calibration of the ACS/WFC3 data
507
+ released within the Frontier Fields.
508
+ • (Middle right) We compare here the relative cali-
509
+ bration of filters at the two extremes of the spec-
510
+ tral range, F444W and F115W, where morphologi-
511
+ cal variations and color terms may change the cen-
512
+ ter position and affect the astrometric procedure.
513
+ We find again very good alignment with negligible
514
+ offset and small madα. ≃ madδ ≃ 0.01”
515
+ • (Right) Finally, we compare the astrometric solu-
516
+ tions on the overlapping areas by summing inde-
517
+ pendently the data of the three different programs
518
+ and checking the accuracy in the overlapping area.
519
+ Again we find very good alignment with negligible
520
+ offset and small madα ≃ madδ ≃ 0.01”.
521
+ We therefore conclude that the astrometric procedure
522
+ is accurate and adequate to the goals of this Stage I
523
+ release. In the future we plan to further explore and
524
+ validate other options for astrometric registration and
525
+ also release images with a smaller pixel scale, to better
526
+ exploit the unprecedented image quality of the JWST
527
+ data. We note that the GLASS-JWST data have a very
528
+ limited dithering pattern (which was driven by spectro-
529
+ scopic requirements) and so may benefit only marginally
530
+ from moving to smaller pixels.
531
+ 2.3. Estimating the Final Depth
532
+ The final coaddition of the different images is weighted
533
+ according to their depth, as estimated by the RMS im-
534
+ age produced by the pipeline. We therefore obtain an
535
+ optimally averaged image with the resulting RMS im-
536
+ age. We a posteriori verified whether the noise estimate
537
+ encoded in the RMS effectively reproduced the photo-
538
+ metric noise.
539
+ To do this, we injected artificial point
540
+ sources of known magnitude in empty regions of the im-
541
+ age, and measured their fluxes and uncertainties with
542
+ a-phot (Merlin et al. 2019), using apertures of radius
543
+ 0.1′′. To take into account the fact that the mosaics are
544
+ the result of a complex pattern of different exposures,
545
+ we divided the maps into regions of similar total expo-
546
+ sure time, and performed this analysis separately in each
547
+ region.
548
+ In general, we find that the RMS of the resulting flux
549
+ distribution is 1.1× larger than the value we would ex-
550
+
551
+ Magellan i vs F444w (1810 obj.)
552
+ △α = 0.000", mad^α = 0.02
553
+ A6 = -0.002", madAs = 0.01
554
+ 0.09
555
+ 0.06
556
+ 0.03
557
+ 0.00
558
+ 2-0.03
559
+ -0.06
560
+ -0.09
561
+ △α (")H160 vs F444W (728 0bi.
562
+ △α = 0.033", madAα = 0.02
563
+ A6 = -0.072", madAs = 0.02
564
+ 0.18
565
+ 0.12
566
+ 0.06
567
+ 0.00
568
+ 1-0.06
569
+ -0.12
570
+ -0.18
571
+ △α (")F115W vs F444W (3649 0bj.)
572
+ △α = 0.000", madAα= 0.01
573
+ A6 = -0.002", madas = 0.01
574
+ 0.09
575
+ 0.06
576
+ 0.03
577
+ 0.00
578
+ 2-0.03
579
+ -0.06
580
+ -0.09
581
+ △α (")F444w overlap regions (1529 obj.
582
+ Aα = -0.002",mad^α = 0.01
583
+ △6 = 0.002", madAs = 0.01
584
+ 0.09
585
+ 0.06
586
+ 0.03
587
+ 0.00
588
+ 2-0.03
589
+ -0.06
590
+ -0.09
591
+ △α (")GLASS-JWST: Abell 2744 NIRCam photometric catalog
592
+ 7
593
+ Figure 4. Depth of the full mosaic F444W image, as pro-
594
+ duced by our pipeline on the basis of the variance image
595
+ of each exposure and with the re-normalization described in
596
+ the text. Each pixel has been converted into 5σ limiting flux
597
+ computed on a circular aperture of 0.2”.
598
+ pect from the SExtractor errors, which are computed
599
+ from the RMS image. A larger difference (1.4×) is found
600
+ for the F444W GLASS image, which is affected by a
601
+ residual pattern due to poor flat–fielding with the cur-
602
+ rent calibration data. We therefore re-scaled the RMS
603
+ maps produced by the pipeline according to these fac-
604
+ tors.
605
+ The resulting depth of this procedure is shown in Fig-
606
+ ure 4. The RMS image is converted into a 5σ limiting
607
+ flux computed on a circular aperture with a diameter
608
+ of 0.2′′, that is the size adopted to estimate colors of
609
+ faint sources. The depth ranges from ≃ 28.6 AB on the
610
+ DDT2 footprint (in particular the area not overlapping
611
+ with DDT1) to ≃ 30.2 AB in the area where GLASS1
612
+ and GLASS2 overlap, arguably one of the deepest im-
613
+ ages obtained so far by JWST.
614
+ A more quantitative assessment of the depth in the
615
+ various filters is reported in Figure 5, where we show the
616
+ distribution of the limiting magnitudes in each image
617
+ resulting from the different strategies adopted by the
618
+ surveys,computed as described above. A clear pattern is
619
+ seen, illustrating the large, mid–depth area obtained by
620
+ UNCOVER and the shallower and deeper parts obtained
621
+ by DDT and GLASS respectively.
622
+ 2.4. HST Imaging
623
+ We have also used the existing images obtained with
624
+ HST in previous programs, namely with the F435W,
625
+ F606W, F775W and F814W bands with ACS and the
626
+ Figure 5. Distribution of the limiting magnitude for each
627
+ band, as shown in the legend. Limiting magnitudes per pixel
628
+ have been computed as for Figure 4.
629
+ F105W, F125W, F140W and F160W bands with WFC3
630
+ - other HST data are available from MAST but are ei-
631
+ ther too shallow and/or limited in area and are not con-
632
+ sidered here. Among these data are included also the
633
+ images that we obtained with DDT Program HST-GO-
634
+ 17231 (PI: Treu), which was specifically aimed at obtain-
635
+ ing ACS coverage for the majority of the GLASS1 and
636
+ GLASS 2 fields. We have used calibrated stacked image
637
+ and weights (G. Brammer, private communication) that
638
+ we have realigned (after checking that the astrometric
639
+ solution is consistent) onto our reference grid to allow a
640
+ straightforward computation of colors.
641
+ 3. PHOTOMETRIC CATALOG
642
+ 3.1. Detection
643
+
644
+ 2.5
645
+ F090W
646
+ F115W
647
+ 2.0
648
+ ou
649
+ 1.5
650
+ (el,
651
+ 1.0
652
+ N
653
+ 0.5
654
+ 9.9
655
+ F150W
656
+ F200W
657
+ 2.0
658
+ QU
659
+ 1.5
660
+ el,
661
+ 1.0
662
+ N
663
+ 0.5
664
+ 2:9
665
+ F277W
666
+ F356W
667
+ 2.0
668
+ xel, norm
669
+ 1.5
670
+ 1.0
671
+ 0.5
672
+ 2:9
673
+ F41QM
674
+ F444W
675
+ 2.0
676
+ ou
677
+ 1.5
678
+ xel,
679
+ 1.0
680
+ N
681
+ 0.5
682
+ 0.0
683
+ 28
684
+ 29
685
+ 30
686
+ 31
687
+ 28
688
+ 29
689
+ 30
690
+ 31
691
+ magim
692
+ maglim26.5
693
+ 27
694
+ 27.5
695
+ 28
696
+ 28.5
697
+ 29
698
+ 29.5
699
+ 30
700
+ 30.58
701
+ Paris et al.
702
+ We follow here the same prescriptions adopted by M22
703
+ and Castellano et al. (2022a,b). We performed source
704
+ detections on the F444W band, since it is generally the
705
+ deepest or among the deepest image for each data set,
706
+ and because high-redshift sources (which are the main
707
+ focus of these observations) are typically brighter at
708
+ longer wavelengths. This approach has the advantage of
709
+ delivering a clear-cut criterion for the object detections,
710
+ that can easily be translated into a cut of rest-frame
711
+ properties for high redshift sources.
712
+ We used SExtractor, adopting a double–pass ob-
713
+ ject detection as applied for the HST-CANDELS cam-
714
+ paign (see Galametz et al. 2013), to detect the objects,
715
+ following the recipes and parameters described in M22.
716
+ We note in particular that we adopt a detection thresh-
717
+ old corresponding to a signal-to-noise ratio (SNR) of 2.
718
+ This is based on simulations, as discussed in M22. The
719
+ other SExtractor parameters used are listed in M22.
720
+ The final SExtractor catalogue on the entire A2744
721
+ area contains 24389 objects.
722
+ Estimating the completeness and purity in a patchy
723
+ (in terms of area and exposure) mosaic derived from
724
+ the large number of observations adopted here, is in-
725
+ trinsically ambiguous. As shown in Figure 5 the depth
726
+ of these images spans approximately 2 magnitudes, and
727
+ the completeness is therefore inhomogenoues - not to
728
+ mention the existence of the cluster that complicates
729
+ both the detection and the estimate of the foreground
730
+ volume (C22b). For these reasons, we do not attempt
731
+ the traditional estimate of the completeness and refer to
732
+ Figure 4 and to Figure 5 for an evaluation of the depth.
733
+ For a proper analysis of the completeness we refer the
734
+ reader to the methodology adopted by C22b were we
735
+ estimate the completeness separately on the individual
736
+ mosaics of the three data sets, which were processed in-
737
+ dependently. We make the three mosaics available upon
738
+ request for this purpose.
739
+ 3.2. Photometry
740
+ We have compiled a multi-wavelength photometric
741
+ catalog following again the prescriptions of M22, which
742
+ in turn is based on previous experience with Hubble
743
+ Space Telescope (HST) images in CANDELS (see e.g.
744
+ Galametz et al. 2013) and in AstroDeep (Merlin et al.
745
+ 2016b, 2021). The catalog is based on a detection per-
746
+ formed on the F444W image described above, and PSF–
747
+ matched aperture photometry of all the sources.
748
+ We
749
+ include all the NIRCam images presented here and ex-
750
+ isting images obtained with HST in previous programs,
751
+ namely with the F435W, F606W, F775W and F814W
752
+ bands with ACS and the F105W, F125W, F140W and
753
+ F160W bands with WFC3.
754
+ The images considered here have PSFs that range
755
+ from 0.035” to 0.2”. Considering that most of the ob-
756
+ jects have small sizes, with half–light–radii less than
757
+ 0.2”, it is necessary to apply a PSF homogenization to
758
+ avoid bias in the derivation of color across the spectral
759
+ range.
760
+ 3.2.1. PSF matching
761
+ Since the detection band is the one with the coars-
762
+ est resolution, we PSF-matched all the other NIRCam
763
+ images to it for color fidelity. We created convolution
764
+ kernels using the WebbPSF models publicly provided
765
+ by STScI4, combining them with a Wiener filtering al-
766
+ gorithm based on the one described in Boucaud et al.
767
+ (2016); and we used a customised version of the con-
768
+ volution module in t-phot (Merlin et al. 2015, 2016a),
769
+ which uses FFTW3 libraries, to smooth the images. This
770
+ approach delivers consistent results with those obtained
771
+ using the software Galight (Ding et al. 2020).
772
+ We
773
+ note that this approach is inevitably approximated. The
774
+ JWST PSF is time– and position–dependent (Nardiello
775
+ et al. 2022), and our dataset is the inhomogeneous com-
776
+ bination of data obtained at different times and with
777
+ different PA, so that the PSF definitely changes over
778
+ the field. For this version of the catalog we used the
779
+ Uncover PSF models as average PSFs, and we plan to
780
+ improve our PSF estimation in the future versions of the
781
+ catalog that will be released in Stage II.
782
+ Similarly, concerning the HST images, we note that all
783
+ of them have too few stars to obtain a robust estimate
784
+ of the PSF directly from the images, so that we adopt
785
+ in all cases existing HST PSFs, taken from CANDELS.
786
+ This approximation may introduce small biases in the
787
+ final catalog. ACS images have been PSF-matched to
788
+ F444W, while for the WFC3 F105W, F125W, F140W
789
+ and F160W images, which have a PSF larger than the
790
+ F444W one, we have done the inverse - smoothed the
791
+ F444W image and the WFC3 F105W, F125W, F140W
792
+ to the F160W and followed a slightly different procedure
793
+ that we describe below.
794
+ 3.2.2. Flux estimate
795
+ The total flux is measured with a-phot on the detec-
796
+ tion image F444W by means of a Kron elliptical aper-
797
+ ture (Kron 1980). As we have shown in M22, simula-
798
+ tions suggest that Kron fluxes measured with a-phot
799
+ are somewhat less affected by systematic errors, while
800
+ being slightly more noisy.
801
+ 4 https://jwst-docs.stsci.edu/jwst-near-infrared-camera/
802
+ nircam-predicted-performance/nircam-point-spread-functions
803
+
804
+ GLASS-JWST: Abell 2744 NIRCam photometric catalog
805
+ 9
806
+ Then, we used a-phot to measure the fluxes at the po-
807
+ sitions of the detected sources on the PSF-matched im-
808
+ ages, masking neighboring objects using the SExtrac-
809
+ tor segmentation map. Given the wide range of magni-
810
+ tudes and sizes of the target galaxies we have measured
811
+ the flux in a range of apertures: the segmentation area
812
+ (the images being on the same grid and PSF-matched)
813
+ and five circular apertures with diameters that are inte-
814
+ ger multiples (2×, 3×, 8×, 16×, ) of the FWHM in the
815
+ F444W band, that correspond to 0.28′′, 0.42′′, 1.12′′ and
816
+ 2.24′′ diameters. For the four WFC3 images (which have
817
+ a PSF larger than F444W) we first filtered the F444W
818
+ to their FWHM and then measured colors between the
819
+ filtered F444W and the WFC3 images. To minimize bi-
820
+ ases when these colors are combined with those of the
821
+ other bands, we use in this case apertures the same mul-
822
+ tiples of the WFC3 PSF adopted for the other bands.
823
+ We remark that this procedure is only approximate, and
824
+ delivers a first order correction of the systematic effects
825
+ due to different PSFs. In a future release we plan to
826
+ adopt more sophisticated approaches to optimize pho-
827
+ tometry, including but not limited to the improvement
828
+ of the PSF estimate and applying T-PHOT on WFC3
829
+ images that have a larger PSF.
830
+ Total fluxes are obtained in the other bands by
831
+ normalizing the colors in a given aperture to the
832
+ F444W total flux,
833
+ i.e.
834
+ by computing fm,total
835
+ =
836
+ fm,aper/fF 444W,aper × fF 444W,total, as described in M22.
837
+ We release the five catalogues described above (one
838
+ computed on segmentation and four on the different
839
+ apertures) and we leave the user to choose which is the
840
+ most suitable for a given science application. In general
841
+ small-aperture catalogues are more appropriate for faint
842
+ sources as they match their small sizes and minimize de-
843
+ belending. Larger apertures may be more appropriate
844
+ for brighter sources and especially cluster members.
845
+ 3.2.3. Validation tests
846
+ We have performed a few validation tests to verify pri-
847
+ marily the flux calibration, that has been the subject of
848
+ many revisions in these first months, and to a lesser ex-
849
+ tent of the procedure adopted to derive the photometric
850
+ catalog.
851
+ The overlap between GLASS1 and GLASS2 southern
852
+ quadrants offers us a nice opportunity to test the NIR-
853
+ Cam flux calibration. Indeed, the two GLASS observa-
854
+ tions have been observed in two epochs (July and Octo-
855
+ ber 2022) with a PA difference of nearly 150 degrees. As
856
+ a result, the southern quadrant of GLASS1 and GLASS2
857
+ are largely overlapping but have been observed with
858
+ modules B and A, respectively. We have therefore ob-
859
+ tained stacked images of the two epochs separately, built
860
+ Figure 6. Stability of the photometric calibration between
861
+ different detectors, as measured by comparing the photom-
862
+ etry of high S/N objects (S/N > 25) detected in the two
863
+ epochs of observations in the SE quadrant of GLASS (lower
864
+ leftmost green square in Figure 1). Objects in this area have
865
+ been observed in two epochs (July and October 2022) and
866
+ with modules B and A, respectively. For each filter difference
867
+ in magnitude ∆M = M1 − M2 for objects between epoch1
868
+ and epoch2 as a function of M1 is reported. Red dashed lines
869
+ represent the median offsets, namely we found: ∆M ≈ 0.06
870
+ with mad ≈ 0.05 for F090W, ∆M ≈ 0.05 with mad ≈ 0.04
871
+ for F115W, ∆M ≈ 0.04 with mad ≈ 0.04 for F150W,
872
+ ∆M ≈ 0.02 with mad ≈ 0.04 for F200W, ∆M ≈ 0.05 with
873
+ mad ≈ 0.04 for F277W, and negligible in F356W and F444W
874
+ with mad ≈ 0.03 and mad ≈ 0.02 respectively. We have vi-
875
+ sually inspected the bright objects with |∆M| > 0.05 and
876
+ verified that they mostly originate from saturated stars or
877
+ objects with incomplete coverage.
878
+ a photometric catalog with the same recipes and checked
879
+ the magnitude difference between objects observed with
880
+ different detectors. The result of this exercise, that has
881
+ been done on all bands, is reported in Figure 6. We note
882
+ that in the short bands the two modules are made of 4
883
+ detectors, each with an independent calibration, that we
884
+ plot all together in Figure 6. The comparison, that is
885
+ limited to objects observed with high S/N > 25, shows
886
+ that the average magnitude difference between the two
887
+
888
+ GLASS 1 vS. 2
889
+ 0.2
890
+ F090W
891
+ 0.0
892
+ -0.2
893
+ 0.2
894
+ F115W
895
+ 0.0
896
+ -0.2
897
+ 0.2
898
+ F150W
899
+ 0.0
900
+ -0.2
901
+ M
902
+ 0.2
903
+ F200W
904
+ 0.0
905
+ 0.2
906
+ F277W
907
+ 0.0
908
+ -0.2
909
+ 0.2
910
+ F356W
911
+ 0.0
912
+ -0.2
913
+ 0.2
914
+ F444W
915
+ 0.0
916
+ -0.2
917
+ 20
918
+ 21
919
+ 22
920
+ 23
921
+ 24
922
+ 25
923
+ 26
924
+ 27
925
+ M110
926
+ Paris et al.
927
+ modules is in general quite small, in all cases below 0.05
928
+ mags (see Figure 6 and its captions for details). This
929
+ confirms that the flux calibration between the different
930
+ modules is reasonably stable at this stage.
931
+ As a further check to validate the photometric
932
+ pipeline, we have compared the m606 and m150 mag-
933
+ nitudes for the sources in the core of the A2744 cluster
934
+ with those measured in the same F606W and in the
935
+ nearby F160W bands measured on HST images, that
936
+ we published within the AstroDeep project (Merlin et al.
937
+ 2016b; Castellano et al. 2016). This comparison is shown
938
+ in Figure 7. Magnitudes in the NIRCam F150W band
939
+ have been shifted by ≃ 0.05 in order to correct for the
940
+ small bandpass difference: the term was estimated using
941
+ theoretical SEDs from a simulated photometric catalog,
942
+ created using Egg (Schreiber et al. 2017). The com-
943
+ parison shows that - when the same approach is used to
944
+ estimate colours, i.e. isophotal magnitudes are adopted
945
+ - the agreement between the two catalogues is excel-
946
+ lent. When we use instead relatively smaller aperture
947
+ in 8×FWHM for the NIRCam photometry we tend to
948
+ underestimate the F150W and - even more - the F606W
949
+ flux of the brightest sources, which are considerably
950
+ more extended than 8×FWHM. We ascribe this effect to
951
+ the existence of colour gradients in bright objects, such
952
+ that small-sized apertures tend to sample the central,
953
+ redder part of the galaxies.
954
+ From this comparison we conclude that - quite reas-
955
+ suringly - the overall photometric chain is consistent be-
956
+ tween the well established Frontier Fields data and these
957
+ new data. At the same time, we remark that the choice
958
+ of which aperture is optimal depends on the size and
959
+ kind of objects under study.
960
+ For faint sources, small
961
+ apertures tend to have higher S/N and should be pre-
962
+ ferred.
963
+ For brightest sources, larger apertures should
964
+ be preferred. It is also possible to estimate rough color
965
+ gradients by comparing the various apertures that we re-
966
+ lease. We also tested that applying the same technique
967
+ without PSF matching introduces an offset of the order
968
+ of ∼0.2 mags in the final colors, which would clearly af-
969
+ fect the derived photometric redshifts and SED fitting
970
+ results.
971
+ Finally, in an effort to cross-validate our results prior
972
+ to release, in the lead up to this paper we compared
973
+ our catalogs to those under development by the UN-
974
+ COVER team (Weaver et al. 2023, in prep) based on
975
+ the same raw datasets. The image processing and pho-
976
+ tometric procedures adopted by the two teams have sig-
977
+ nificant differences.
978
+ The main are: i) image coaddi-
979
+ tion (UNCOVER team adopts grizli, while we use a
980
+ custom pipeline which uses scamp and swarp; ii) ob-
981
+ ject detection (UNCOVER uses an optimally stacked
982
+ Figure 7.
983
+ Comparison of photometry between this work
984
+ and AstroDeep HST catalogs in the core of the A2744 clus-
985
+ ter. Upper: Difference between the magnitude in the F160W
986
+ WFC3 band in AstroDeep and the F150W NIRCam of this
987
+ work for objects in common between the two catalogues. The
988
+ F150W magnitude has been corrected for the ≃ 0.05 mag-
989
+ nitude shift between the two bands. Filled point represent
990
+ the difference between magnitudes computed in isophotal ar-
991
+ eas in both catalogues. Empty points represent the magni-
992
+ tude difference adopting the F150W magnitude computed in
993
+ 8×FWHM. Bottom: As above, for the F606W band. The
994
+ systematic bias between isophotal and 8×FWHM colours is
995
+ due to color gradients in the center of bright sources.
996
+ F277W+F356W+F444W image after removing the intr-
997
+ acluster light, while we use F444W); iii) techniques and
998
+ tools for PSF matching and photometry. For these rea-
999
+ sons, we expect some differences between the catalogs,
1000
+ especially for faint sources at the detection limit. How-
1001
+ ever, our comparison of working versions of the catalogs
1002
+ produced by the two teams shows overall a good agree-
1003
+ ment in the colors and magnitudes of the vast majority
1004
+ of objects, with no evidence of significant bias beyond
1005
+ what can be explained by the different choices. We defer
1006
+ a detailed comparison to future versions of the catalog
1007
+ (Stage II).
1008
+ 4. SUMMARY
1009
+ We present in this paper the data obtained by three
1010
+ NIRCam programs on the A2744 cluster: the GLASS-
1011
+ JWST Early Release Science Program, UNCOVER, and
1012
+ Directory Discretionary Time 2756. All the data, taken
1013
+ with eight different filters (F090W, F115W, F150W,
1014
+ F200W, F277W, F356W, F410M, F444W), have been
1015
+ reduced with an updated pipelines that builds upon the
1016
+ official STScI pipeline but includes a number of improve-
1017
+ ment to better remove some instrumental signature and
1018
+ streamline the process.
1019
+ All frames have been aligned onto a common frame
1020
+ with 0.031” pixel scale, approximately matching the na-
1021
+
1022
+ 1.0
1023
+ F160WAstrodeep
1024
+ -F150Wuncover,corr
1025
+ 0.5
1026
+ 0.0
1027
+ 88
1028
+ D
1029
+ 0.5
1030
+ -1.0
1031
+ 1.0
1032
+ 0.5
1033
+ 0.0
1034
+ 0.5
1035
+ O
1036
+ O
1037
+ 1.0
1038
+ 18
1039
+ 19
1040
+ 20
1041
+ 21
1042
+ 22
1043
+ 23
1044
+ 24
1045
+ F160WAstrodeepGLASS-JWST: Abell 2744 NIRCam photometric catalog
1046
+ 11
1047
+ tive pixel scale of the short wavelength data. The final
1048
+ images on the whole A2744 region cover an area of 46.5
1049
+ arcmin2 with PSF ranging from 0.035” (for the F090W
1050
+ image) to 0.14” (F444W), and reach astonishingly deep
1051
+ 5σ magnitude limits from 28.5 to 30.5, depending on
1052
+ location and filter.
1053
+ We exploit also other HST publicly available programs
1054
+ which have targeted the area, including also the avail-
1055
+ able HST ACS and WFC3 data in the F435W, F606W,
1056
+ F775W and F814W (ACS) and F105W, F125W, F140W
1057
+ and F160W (WFC3) bands, to expand the coverage of
1058
+ the visible-to-IR wavelength range.
1059
+ On these data we derive a photometric catalog by
1060
+ detecting objects in the F444W image and comput-
1061
+ ing PSF-matched forced photometry on the remaining
1062
+ bands.
1063
+ We made a number of tests to validate the photometric
1064
+ calibrations, either internal, based on overlapping parts
1065
+ observed in different epochs with different modules, and
1066
+ external, based on cross-correlation with the AstroDeep
1067
+ catalog of the cluster region. They both confirm that
1068
+ photometric offset are limited to at most 0.05 mags or
1069
+ less. Slightly larger (0.1 mags) systematic biases, espe-
1070
+ cially when HST bands are concerned, could be due to
1071
+ the simplified PSF matching that we adopt in this first
1072
+ release.
1073
+ As we do not explicitly remove the intra-cluster light,
1074
+ photometry of faint sources in the cluster core might
1075
+ also be affected by poor background subtraction.
1076
+ We publicly release the entire mosaic of the NIRCam
1077
+ images. The three individual images of each program,
1078
+ which are more homogeneous in terms of PSF orienta-
1079
+ tion and coverage/depth, and potentially more suitable
1080
+ for accurate photometry and for accurate estimate of
1081
+ incompleteness, are also available upon request.
1082
+ We also publicly release the multi-wavelength cata-
1083
+ logue on the entire A2744 area, which includes 24389
1084
+ objects. We release 5 independent catalogues, based on
1085
+ a different aperture (2×, 3×, 8×, 16× the PSF) and in
1086
+ the isophotal area. This catalog is optimized for high
1087
+ redshift galaxies, and in general for faint extragalactic
1088
+ sources, and aimed at allowing a first look at the data
1089
+ and the selection of targets for Cycle 2 proposals. In
1090
+ future releases we plan to include updated calibrations
1091
+ and procedures for the image processing and to optimize
1092
+ the photometry with more sophisticated approaches for
1093
+ PSF matching.
1094
+ Finally we also release the code developed to remove
1095
+ the 1/f noise from the NIRCam images, that improves
1096
+ upon the current implementation in the STScI pipeline
1097
+ with a more effective masking of sources in the image.
1098
+ Images, catalogues and software are immediately
1099
+ available for download from the GLASS-ERS collabora-
1100
+ tion website5 and from the AstroDeep website6. They
1101
+ will also be made available at the MAST archive upon
1102
+ acceptance of the paper.
1103
+ All the JWST data used in this paper can be found in
1104
+ MAST: 10.17909/fqaq-p393.
1105
+ ACKNOWLEDGEMENT
1106
+ We warmly thank J. Weaver, K. Withaker, I. Labb`e
1107
+ and R. Bezanson for sharing their data with us prior
1108
+ to publication, which made it possible to compare the
1109
+ two processes for data analysis.
1110
+ This work is based
1111
+ on observations made with the NASA/ESA/CSA James
1112
+ Webb Space Telescope, and with the NASA/ESA Hub-
1113
+ ble Space Telescope.
1114
+ The data were obtained from
1115
+ the Mikulski Archive for Space Telescopes at the Space
1116
+ Telescope Science Institute, which is operated by the
1117
+ Association of Universities for Research in Astronomy,
1118
+ Inc., under NASA contract NAS 5-03127 for JWST
1119
+ and NAS 5–26555 for HST. These observations are as-
1120
+ sociated with program JWST-ERS-1324, JWST-DDT-
1121
+ 2756, and JWST-GO-2561, and several HST programs.
1122
+ We acknowledge financial support from NASA through
1123
+ grant JWST-ERS-1324.
1124
+ This research is supported
1125
+ in part by the Australian Research Council Centre of
1126
+ Excellence for All Sky Astrophysics in 3 Dimensions
1127
+ (ASTRO 3D), through project number CE170100013.
1128
+ KG and TN acknowledge support from Australian Re-
1129
+ search Council Laureate Fellowship FL180100060. MB
1130
+ acknowledges support from the Slovenian national re-
1131
+ search agency ARRS through grant N1-0238.
1132
+ We
1133
+ acknowledge financial support through grants PRIN-
1134
+ MIUR 2017WSCC32 and 2020SKSTHZ. We acknowl-
1135
+ edge support from the INAF Large Grant 2022 “Ex-
1136
+ tragalactic Surveys with JWST” (PI Pentericci). CM
1137
+ acknowledges support by the VILLUM FONDEN under
1138
+ grant 37459. RAW acknowledges support from NASA
1139
+ JWST Interdisciplinary Scientist grants NAG5-12460,
1140
+ NNX14AN10G and 80NSSC18K0200 from GSFC. The
1141
+ Cosmic Dawn Center (DAWN) is funded by the Danish
1142
+ National Research Foundation under grant DNRF140.
1143
+ This work has made use of data from the Euro-
1144
+ pean Space Agency (ESA) mission Gaia (https://www.
1145
+ cosmos.esa.int/gaia), processed by the Gaia Data Pro-
1146
+ cessing and Analysis Consortium (DPAC, https://www.
1147
+ cosmos.esa.int/web/gaia/dpac/consortium).
1148
+ Funding
1149
+ for the DPAC has been provided by national institu-
1150
+ tions, in particular the institutions participating in the
1151
+ 5 https://glass.astro.ucla.edu
1152
+ 6 http://www.astrodeep.eu
1153
+
1154
+ 12
1155
+ Paris et al.
1156
+ Gaia Multilateral Agreement. The authors thank Paola
1157
+ Marrese and Silvia Marinoni (Space Science Data Cen-
1158
+ ter, Italian Space Agency) for their contribution to the
1159
+ work.
1160
+ REFERENCES
1161
+ Bertin, E. 2006, in Astronomical Society of the Pacific
1162
+ Conference Series, Vol. 351, Astronomical Data Analysis
1163
+ Software and Systems XV, ed. C. Gabriel, C. Arviset,
1164
+ D. Ponz, & E. Solano, 112
1165
+ Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393
1166
+ Bertin, E., Mellier, Y., Radovich, M., et al. 2002, in
1167
+ Astronomical Society of the Pacific Conference Series,
1168
+ Vol. 281, Astronomical Data Analysis Software and
1169
+ Systems XI, ed. D. A. Bohlender, D. Durand, & T. H.
1170
+ Handley, 228
1171
+ Bezanson, R., Labbe, I., Whitaker, K. E., et al. 2022, arXiv
1172
+ e-prints, arXiv:2212.04026.
1173
+ https://arxiv.org/abs/2212.04026
1174
+ Boucaud, A., Bocchio, M., Abergel, A., et al. 2016, A&A,
1175
+ 596, A63, doi: 10.1051/0004-6361/201629080
1176
+ Bouwens, R., Illingworth, G., Oesch, P., et al. 2022, arXiv
1177
+ e-prints, arXiv:2212.06683.
1178
+ https://arxiv.org/abs/2212.06683
1179
+ Castellano, M., Amor´ın, R., Merlin, E., et al. 2016, A&A,
1180
+ 590, A31, doi: 10.1051/0004-6361/201527514
1181
+ Castellano, M., Fontana, A., Treu, T., et al. 2022a, arXiv
1182
+ e-prints, arXiv:2207.09436.
1183
+ https://arxiv.org/abs/2207.09436
1184
+ —. 2022b, arXiv e-prints, arXiv:2212.06666.
1185
+ https://arxiv.org/abs/2212.06666
1186
+ Ding, X., Silverman, J., Treu, T., et al. 2020, ApJ, 888, 37,
1187
+ doi: 10.3847/1538-4357/ab5b90
1188
+ Donnan, C. T., McLeod, D. J., Dunlop, J. S., et al. 2023,
1189
+ MNRAS, 518, 6011, doi: 10.1093/mnras/stac3472
1190
+ Finkelstein, S. L., Bagley, M. B., Arrabal Haro, P., et al.
1191
+ 2022, arXiv e-prints, arXiv:2207.12474.
1192
+ https://arxiv.org/abs/2207.12474
1193
+ Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al.
1194
+ 2016, A&A, 595, A1, doi: 10.1051/0004-6361/201629272
1195
+ Galametz, A., Grazian, A., Fontana, A., et al. 2013, ApJS,
1196
+ 206, 10, doi: 10.1088/0067-0049/206/2/10
1197
+ Kron, R. G. 1980, ApJS, 43, 305, doi: 10.1086/190669
1198
+ Merlin, E., Fontana, A., Ferguson, H. C., et al. 2015, A&A,
1199
+ 582, A15, doi: 10.1051/0004-6361/201526471
1200
+ Merlin, E., Bourne, N., Castellano, M., et al. 2016a, A&A,
1201
+ 595, A97, doi: 10.1051/0004-6361/201628751
1202
+ Merlin, E., Amor´ın, R., Castellano, M., et al. 2016b, A&A,
1203
+ 590, A30, doi: 10.1051/0004-6361/201527513
1204
+ Merlin, E., Fortuni, F., Torelli, M., et al. 2019, MNRAS,
1205
+ 490, 3309, doi: 10.1093/mnras/stz2615
1206
+ Merlin, E., Castellano, M., Santini, P., et al. 2021, A&A,
1207
+ 649, A22, doi: 10.1051/0004-6361/202140310
1208
+ Merlin, E., Bonchi, A., Paris, D., et al. 2022, A&A, 938,
1209
+ L14, doi: 10.3847/2041-8213/ac8f93
1210
+ Morishita, T., & Stiavelli, M. 2022, arXiv e-prints,
1211
+ arXiv:2207.11671. https://arxiv.org/abs/2207.11671
1212
+ Naidu, R. P., Oesch, P. A., van Dokkum, P., et al. 2022,
1213
+ arXiv e-prints, arXiv:2207.09434.
1214
+ https://arxiv.org/abs/2207.09434
1215
+ Nardiello, D., Bedin, L. R., Burgasser, A., et al. 2022,
1216
+ MNRAS, 517, 484, doi: 10.1093/mnras/stac2659
1217
+ Oke, J. B., & Gunn, J. E. 1983, ApJ, 266, 713,
1218
+ doi: 10.1086/160817
1219
+ Rigby, J., Perrin, M., McElwain, M., et al. 2022, arXiv
1220
+ e-prints, arXiv:2207.05632.
1221
+ https://arxiv.org/abs/2207.05632
1222
+ Roberts-Borsani, G., Treu, T., Chen, W., et al. 2022, arXiv
1223
+ e-prints, arXiv:2210.15639.
1224
+ https://arxiv.org/abs/2210.15639
1225
+ Robertson, B. E., Tacchella, S., Johnson, B. D., et al. 2022,
1226
+ arXiv e-prints, arXiv:2212.04480.
1227
+ https://arxiv.org/abs/2212.04480
1228
+ Schlawin, E., Leisenring, J., Misselt, K., et al. 2020, AJ,
1229
+ 160, 231, doi: 10.3847/1538-3881/abb811
1230
+ Schreiber, C., Pannella, M., Leiton, R., et al. 2017, A&A,
1231
+ 599, A134, doi: 10.1051/0004-6361/201629155
1232
+ Treu, T., Roberts-Borsani, G., Bradac, M., et al. 2022,
1233
+ ApJ, 935, 110, doi: 10.3847/1538-4357/ac8158
1234
+ Yan, H., Cohen, S. H., Windhorst, R. A., et al. 2022, arXiv
1235
+ e-prints, arXiv:2209.04092.
1236
+ https://arxiv.org/abs/2209.04092
1237
+
ANE0T4oBgHgl3EQfPgBb/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
AdAyT4oBgHgl3EQf3_pa/content/tmp_files/2301.00778v1.pdf.txt ADDED
@@ -0,0 +1,2579 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.00778v1 [math.PR] 2 Jan 2023
2
+ LECTURE NOTES ON TREE-FREE REGULARITY
3
+ STRUCTURES
4
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
5
+ Abstract. These lecture notes are intended as reader’s digest of
6
+ recent work on a diagram-free approach to the renormalized cen-
7
+ tered model in Hairer’s regularity structures. More precisely, it is
8
+ about the stochastic estimates of the centered model, based on Malli-
9
+ avin calculus and a spectral gap assumption. We focus on a specific
10
+ parabolic partial differential equation in quasi-linear form driven by
11
+ (white) noise.
12
+ We follow a natural renormalization strategy based on preserving
13
+ symmetries, and carefully introduce Hairer’s notion of a centered
14
+ model, which provides the coefficients in a formal series expansion
15
+ of a general solution. We explain how the Malliavin derivative in
16
+ conjunction with Hairer’s re-expansion map allows to reformulate
17
+ this definition in a way that is stable under removing the small-scale
18
+ regularization.
19
+ A few exemplary proofs are provided, both of analytic and of alge-
20
+ braic character. The working horse of the analytic arguments is an
21
+ “annealed” Schauder estimate and related Liouville principle, which
22
+ is provided. The algebra of formal power series, in variables that
23
+ play the role of coordinates of the solution manifold, and its algebra
24
+ morphisms are the key algebraic objects.
25
+ Keywords: Singular SPDE, Regularity Structures, BPHZ renor-
26
+ malization, Malliavin calculus, quasi-linear PDE.
27
+ MSC 2020: 60H17, 60L30, 60H07, 81T16, 35K59.
28
+ Contents
29
+ 1.
30
+ A singular quasi-linear SPDE
31
+ 3
32
+ 2.
33
+ Annealed Schauder theory
34
+ 5
35
+ 3.
36
+ Symmetry-motivated postulates on the form of the counter
37
+ terms
38
+ 8
39
+ 4.
40
+ Algebrizing the counter term
41
+ 10
42
+ 5.
43
+ Algebrizing the solution manifold: The centered model
44
+ 12
45
+ 6.
46
+ The main result: A stochastic estimate of the centered
47
+ model
48
+ 17
49
+ 7.
50
+ Malliavin derivative and Spectral gap (SG)
51
+ 19
52
+ 8.
53
+ The structure group and the re-expansion map
54
+ 26
55
+ 1
56
+
57
+ 2
58
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
59
+ References
60
+ 33
61
+ The theory of regularity structures by Hairer provides a systematic way
62
+ to treat the small-scale divergences in singular semi-linear stochastic
63
+ PDEs. Quintessential models of mathematical physics like the dynam-
64
+ ical Φ4
65
+ 3 model or the KPZ equation have been treated. Inspired by
66
+ Lyon’s theory of rough paths, this theory separates probabilistic and
67
+ analytical aspects:
68
+ • Centered model. In a first probabilistic step, the coefficients of
69
+ a local formal power series representation of a general solution
70
+ of the renormalized PDE are constructed and estimated; the co-
71
+ efficients are indexed by (decorated) trees, and their stochastic
72
+ estimate follows the diagrammatic approach to renormalization
73
+ of quantum field theories.
74
+ • Modelled distribution. In a second analytical step, inspired by
75
+ Gubinelli’s controlled rough path, the solution of a specific ini-
76
+ tial value problem is found as a fixed point based on modulating
77
+ and truncating the formal power series . This step is purely de-
78
+ terministic.
79
+ This automated two-pronged approach relies on an understanding of
80
+ the algebraic nature of the re-expansion maps that allow to pass from
81
+ one base-point to another in the local power series representation, in
82
+ form of the “structure group”. The main progress of regularity struc-
83
+ tures over the term-by-term treatment in the mathematical physics
84
+ literature is that thanks to centering and re-expansion, the second step
85
+ yields a rigorous (small data) well-posedness result. As an introductory
86
+ text to the theory of regularity structures we recommend [9].
87
+ In [17], motivated by the extension to a quasi-linear setting featuring
88
+ a general non-linearity a(u), an alternative realization of Hairer’s reg-
89
+ ularity structures was proposed; it replaces trees with a more greedy
90
+ index set. This index set of multi-indices naturally comes up when
91
+ writing a general solution u as a functional of a, or rather as a func-
92
+ tion of the coefficients of a in its power law expansion. In [17] it was
93
+ established that any solution of the renormalized PDE can be locally
94
+ approximated by a modelled distribution. This a-priori estimate was
95
+ obtained under the assumption that the natural stochastic estimates
96
+ on the centered model are available.
97
+ In [15] this program was continued: Based on scaling and other symme-
98
+ tries, a canonical renormalization of the PDE and its centered model
99
+ was proposed, and the centered model was stochastically constructed
100
+ and estimated. These notes present selected aspects of [15], providing
101
+ additional motivation. For a simpler setting where no renormalization
102
+ and thus only purely deterministic estimates are needed, we recommend
103
+
104
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
105
+ 3
106
+ to also have a look at1 [13]. The algebraic aspects of the multi-index
107
+ based regularity structures are worked out in [14], where in line with
108
+ Hairer’s postulates the underlying Hopf-algebraic nature of the struc-
109
+ ture group was uncovered. In fact, the Hopf algebra arises from a Lie
110
+ algebra generated by natural actions on the space of non-linearities a
111
+ and solutions u.
112
+ Other approaches to singular SPDEs include the theory of paracon-
113
+ trolled distributions by Gubinelli, Imkeller, and Perkowski, we rec-
114
+ ommend [8] for a first reading, and the renormalization group flow
115
+ approach introduced by Kupiainen and generalized by Duch; we rec-
116
+ ommend [12] and [6] for an introduction. The para-controlled calculus
117
+ provides an alternative to the separation into model and modelled dis-
118
+ tribution, replacing localization in physical space-time by localization
119
+ on the Fourier side; it is (typically) also indexed by trees. The flow
120
+ approach blends the stochastic and the deterministic step of regularity
121
+ structures, and has an index set closer to multi-indices. While these
122
+ alternative approaches might be more efficient in specific situations,
123
+ they presumably lack the full flexibility of the two-pronged approach
124
+ of regularity structures with its conceptual clarity.
125
+ 1. A singular quasi-linear SPDE
126
+ We are interested in nonlinear elliptic or parabolic equations with a
127
+ random and thus typically rough right hand side ξ. Our approach is
128
+ guided by moving beyond the well-studied semi-linear case. We con-
129
+ sider a mildly quasi-linear case where the coefficients of the leading-
130
+ order derivatives depend on the solution u itself. To fix ideas, we focus
131
+ on the parabolic case in a single space dimension; since we treat the
132
+ parabolic equation in the whole space-time like an anisotropic ellip-
133
+ tic equation, we denote by x1 the space-like and by x2 the time-like
134
+ variable. Hence we propose to consider
135
+ (∂2 − ∂2
136
+ 1)u = a(u)∂2
137
+ 1u + ξ,
138
+ (1)
139
+ where we think of the values of a(u) to be such that the equation
140
+ is parabolic.
141
+ We are interested in laws / ensembles of ξ where the
142
+ solutions v to the linear equation
143
+ (∂2 − ∂2
144
+ 1)v = ξ
145
+ (2)
146
+ 1however, the setting in [13] is different in the sense that it imposes an artificial
147
+ space-time periodicity: on the one hand, this allows to separate construction from
148
+ estimation, on the other hand, it obfuscates the quintessential scaling
149
+
150
+ 4
151
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
152
+ are (almost surely) H¨older continuous, where it will turn out to be
153
+ convenient to express this in the “annealed” form2 of
154
+ sup
155
+ x̸=y
156
+ 1
157
+ |y − x|αE
158
+ 1
159
+ 2|v(y) − v(x)|2 < ∞
160
+ (3)
161
+ for some exponent α ∈ (0, 1).
162
+ In view of the anisotropic nature of
163
+ ∂2 − ∂2
164
+ 1 and its invariance under the rescaling x1 = sˆx1 and x2 = s2ˆx2,
165
+ H¨older continuity in (3) is measured w. r. t. the Carnot-Carath´eodory
166
+ distance
167
+ “|y − x|” :=
168
+ 4�
169
+ (y1 − x1)4 + (y2 − x2)2 ∼ |y1 − x1| + |y2 − x2|
170
+ 1
171
+ 2.
172
+ (4)
173
+ By Schauder theory for ∂2 − ∂2
174
+ 1, on which we shall expand on in Sub-
175
+ section 2, this is the case for white noise ξ with α = 1
176
+ 2. The rationale
177
+ is that white noise has order of regularity −D
178
+ 2 , where D is the effective
179
+ dimension, which in case of (2) is D = 1 + 2 = 3 since in view of (4)
180
+ the time-like variable x2 counts twice, and that (∂2 − ∂2
181
+ 1)−1 increases
182
+ regularity by two, leading to −D
183
+ 2 + 2 = 1
184
+ 2.
185
+ In the range of α ∈ (0, 1), the SPDE (1) is what is called “singular”:
186
+ We cannot expect that the order of regularity of u and thus a(u) is
187
+ better than the one of v, which is α, and hence the order of regularity
188
+ of ∂2
189
+ 1u is no better than α − 2. Since α + (α − 2) < 0 for α < 1, the
190
+ product a(u)∂2
191
+ 1u cannot be classically/deterministically defined3. As
192
+ discussed at the end of Section 2, a renormalization is needed4.
193
+ The same feature occurs for the (semi-linear) multiplicative heat equa-
194
+ tion (∂2 −∂2
195
+ 1)u = a(u)ξ; in fact, our approach also applies to this semi-
196
+ linear case, which already has been treated by (standard) regularity
197
+ structures in [10]. A singular product is already present in the case
198
+ when the x1-dependence is suppressed, so that the above semi-linear
199
+ equation turns into the SDE du
200
+ dx2 = a(u)ξ with white noise ξ in the time-
201
+ like variable x2. In this case, the analogue of v from (2) is Brownian
202
+ motion, which is characterized by E(v(y2)−v(x2))2 = |y2−x2| and thus
203
+ annealed H¨older exponent 1
204
+ 2 in x2, which in view of (4) corresponds to
205
+ the border-line setting α = 1. Ito’s integral and, more recently, Lyons’
206
+ rough paths [16] and Gubinelli’s controlled rough paths [7] have been
207
+ devised to tackle the issue in this SDE setting.
208
+ 2Think of Brownian motion which satisfies E
209
+ 1
210
+ 2 (B(s) − B(t))2 = |s − t|
211
+ 1
212
+ 2 while
213
+ not being H¨older continuous of exponent 1
214
+ 2 almost surely. Following the jargon an-
215
+ nealed/quenched from statistical mechanics models (which itself is borrowed from
216
+ metallurgy), we speak of annealed norms when the inner norm is an Lp-norm
217
+ w. r. t. probability E and the outer norm is a space-time one.
218
+ 3It is a classical result that the multiplication extends naturally from Cα × Cβ
219
+ into D′ if and only if α + β > 0, see [1, Section 2.6].
220
+ 4The range α > 1, while still subtle for α < 2, does not require a renormalization,
221
+ see [13].
222
+
223
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
224
+ 5
225
+ 2. Annealed Schauder theory
226
+ This section provides the main (linear) PDE ingredient for our result.
227
+ At the same time, it will allow us to discuss (2).
228
+ In view of (2), we are interested in the fundamental solution of the
229
+ differential operator A := ∂2 − ∂2
230
+ 1. It turns out to be convenient to use
231
+ the more symmetric5 fundamental solution of the non-negative A∗A
232
+ = (−∂2 −∂2
233
+ 1)(∂2 −∂2
234
+ 1) = ∂4
235
+ 1 −∂2
236
+ 2. Moreover, it will be more transparent
237
+ to “disintegrate” the latter fundamental solution, by which we mean
238
+ writing it as
239
+ ´ ∞
240
+ 0 dtψt(z), where {ψt}t>0 are the kernels of the semi-
241
+ group exp(−tA∗A). Clearly, the Fourier transform is given by
242
+ Fψt(q) = exp(−t(q4
243
+ 1 + q2
244
+ 2))
245
+ (4)
246
+ = exp(−t|q|4).
247
+ (5)
248
+ In particular, ψt is a Schwartz function. For a Schwartz distribution f
249
+ like realizations of white noise, we thus define ft(y) as the pairing of f
250
+ with ψt(y − ·); ft is a smooth function. On the level of these kernels,
251
+ the semi-group property translates into
252
+ ψs ∗ ψt = ψs+t
253
+ and
254
+ ˆ
255
+ ψt = 1.
256
+ (6)
257
+ By construction, {ψt}t satisfies the PDE
258
+ ∂tψt + (∂4
259
+ 1 − ∂2
260
+ 2)ψt = 0.
261
+ (7)
262
+ By scale invariance of (7) under x1 = sˆx1, x2 = s2ˆx2, and t = s4ˆt, we
263
+ have
264
+ ψt(x1, x2) =
265
+ 1
266
+ (
267
+ 4√
268
+ t)D=3 ψ1( x1
269
+ 4√
270
+ t,
271
+ x2
272
+ (
273
+ 4√
274
+ t)2).
275
+ (8)
276
+ Lemma 1. Let 0 < α ≤ η < ∞ with η ̸∈ Z, p < ∞, and x ∈ R2 be
277
+ given. For a random Schwartz distribution f with
278
+ E
279
+ 1
280
+ p|ft(y)|p ≤ (
281
+ 4√
282
+ t)α−2(
283
+ 4√
284
+ t + |y − x|)η−α
285
+ for all t > 0, y ∈ R2,
286
+ (9)
287
+ there exists a unique random function u of the class
288
+ sup
289
+ y∈R2
290
+ 1
291
+ |y − x|η E
292
+ 1
293
+ p|u(y)|p < ∞
294
+ (10)
295
+ satisfying (distributionally in R2)
296
+ (∂2 − ∂2
297
+ 1)u = f + (polynomial of degree ≤ η − 2).
298
+ (11)
299
+ It actually satisfies (11) without the polynomial. Moreover, the l. h. s. of
300
+ (10) is bounded by a constant only depending on α and η.
301
+ 5It is symmetric under reflection not just in space but also in time
302
+
303
+ 6
304
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
305
+ Now white noise ξ is an example of such a random Schwartz distri-
306
+ bution:
307
+ Since ξt(y) is a centered Gaussian, we have E
308
+ 1
309
+ p|ξt(y)|p ≲p
310
+ E
311
+ 1
312
+ 2(ξt(y))2. By using the characterizing property of white noise in terms
313
+ of its pairing with a test function ζ
314
+ E(ξ, ζ)2 =
315
+ ˆ
316
+ ζ2,
317
+ (12)
318
+ we have E
319
+ 1
320
+ 2(ξt(y))2 =
321
+ � ´
322
+ ψ2
323
+ t (y − ·)
324
+ � 1
325
+ 2, which by scaling (8) is equal
326
+ to (
327
+ 4√
328
+ t)− D
329
+ 2 (
330
+ ´
331
+ ψ2
332
+ 1)
333
+ 1
334
+ 2 ∼ (
335
+ 4√
336
+ t)− D
337
+ 2 . This specifies the sense in which white
338
+ noise ξ has order of regularity −D
339
+ 2 .
340
+ Fixing a “base point” x, Lemma 1 thus constructs the solution of (2)
341
+ distinguished by v(x) = 0. Note that the output (10) takes the form of
342
+ E
343
+ 1
344
+ p|v(y)−v(x)|p ≲p |y−x|
345
+ 1
346
+ 2, which extends (3) from p = 2 to general p.
347
+ Hence Lemma 1 provides an annealed version of a Schauder estimate,
348
+ alongside a Liouville-type uniqueness result.
349
+ Proof of Lemma 1. By construction,
350
+ ´ ∞
351
+ 0 dt(−∂2 − ∂2
352
+ 1)ψt is the funda-
353
+ mental solution of ∂2 − ∂2
354
+ 1, so that we take the convolution of it with
355
+ f. However, in order to obtain a convergent expression for t ↑ ∞, we
356
+ need to pass to a Taylor remainder:
357
+ u =
358
+ ˆ ∞
359
+ 0
360
+ dt(id − Tη
361
+ x)(−∂2 − ∂2
362
+ 1)ft,
363
+ (13)
364
+ where Tη
365
+ x is the operation of taking the Taylor polynomial of order ≤ η;
366
+ as we shall argue the additional Taylor polynomial does not affect the
367
+ PDE.
368
+ We claim that (13) is well-defined and estimated as
369
+ E
370
+ 1
371
+ p|u(y)|p ≲ |y − x|η.
372
+ To this purpose, we first note that
373
+ E
374
+ 1
375
+ p|∂nft(y)|p ≲ (
376
+ 4√
377
+ t)α−2−|n|(
378
+ 4√
379
+ t + |y − x|)η−α,
380
+ (14)
381
+ where
382
+ ∂nf := ∂n1
383
+ 1 ∂n2
384
+ 2 f
385
+ and
386
+ |n| = n1 + 2n2.
387
+ (15)
388
+ Indeed, by the semi-group property (6) we may write ∂nft(y) =
389
+ ´
390
+ dz
391
+ ∂nψ t
392
+ 2(y−z) f t
393
+ 2(z), so that E
394
+ 1
395
+ p|∂nft(y)|p ≤
396
+ ´
397
+ dz|∂nψ t
398
+ 2(y−z)|E
399
+ 1
400
+ p|f t
401
+ 2(z)|p.
402
+ Hence by (9), (14) follows from the kernel bound
403
+ ´
404
+ dz |∂nψ t
405
+ 2(y − z)|
406
+ (
407
+ 4√
408
+ t + |y − x|)η−α ≲ (
409
+ 4√
410
+ t)−|n|(
411
+ 4√
412
+ t + |y − x|)η−α, which itself is a conse-
413
+ quence of the scaling (8) and the fact that ψ 1
414
+ 2 is a Schwartz function.
415
+ Equipped with (14), we now derive two estimates for the integrand
416
+ of (13), namely for
417
+ 4√
418
+ t ≥ |y − x| (“far field”) and for
419
+ 4√
420
+ t ≤ |y − x|
421
+ (“near field”). We write the Taylor remainder (id − Tη
422
+ x)(∂2 + ∂2
423
+ 1)ft(y)
424
+
425
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
426
+ 7
427
+ as a linear combination of6 (y − x)n∂n(∂2 + ∂2
428
+ 1)ft(z) with |n| > η and
429
+ at some point z intermediate to y and x.
430
+ By (14) such a term is
431
+ estimated by |y − x||n|(
432
+ 4√
433
+ t)α−4−|n|(
434
+ 4√
435
+ t + |y − x|)η−α, which in the far
436
+ field is ∼ |y − x||n|(
437
+ 4√
438
+ t)η−4−|n|. Since the exponent on t is < −1, we
439
+ obtain as desired
440
+ E
441
+ 1
442
+ p|
443
+ ˆ ∞
444
+ |y−x|4 dt(id − Tη
445
+ x)(∂2 + ∂2
446
+ 1)ft(y)|p ≲ |y − x|η.
447
+ For the near-field term, i. e. for
448
+ 4√
449
+ t ≤ |y − x|, we proceed as follows:
450
+ E
451
+ 1
452
+ p |(id − Tη
453
+ x)(∂2 + ∂2
454
+ 1)ft(y)|p
455
+ ≤ E
456
+ 1
457
+ p|(∂2 + ��2
458
+ 1)ft(y)|p +
459
+
460
+ |n|≤η
461
+ |y − x||n|E
462
+ 1
463
+ p|∂n(∂2 + ∂2
464
+ 1)ft(x)|p
465
+ (14)
466
+ ≲ (
467
+ 4√
468
+ t)α−4|y − x|η−α +
469
+
470
+ |n|≤η
471
+ |y − x||n|(
472
+ 4√
473
+ t)η−4−|n|.
474
+ Since η is not an integer, the sum restricts to |n| < η, so that all
475
+ exponents on t are > −1. Hence we obtain as desired
476
+ E
477
+ 1
478
+ p|
479
+ ˆ |y−x|4
480
+ 0
481
+ dt(id − Tη
482
+ x)(∂2 + ∂2
483
+ 1)ft(y)|p ≲ |y − x|η.
484
+ It can be easily checked that (13) is indeed a solution of (11), even
485
+ without a polynomial. For a detailed proof we refer to [15, Proposi-
486
+ tion 4.3].
487
+ We turn to the uniqueness of u in the class (10) satisfying (11). Given
488
+ two such solutions u1, u2, we observe that ¯u := u1 − u2 satisfies (10)
489
+ and (11) with f = 0. In particular ∂n(∂2 − ∂2
490
+ 1)¯u = 0 for |n| > η − 2,
491
+ and thus from (7) we obtain ∂t∂n¯ut = 0 provided |n| > η − 4. Thus,
492
+ ∂n¯ut is independent of t > 0. Moreover, (10) implies that E|∂n¯ut| → 0
493
+ as t → ∞ for |n| > η. Hence we learn from t → 0 that ∂n¯u = 0
494
+ for |n| > η, i.e. ¯u is a polynomial of degree ≤ η. Since η ̸∈ Z this
495
+ strengthens to ¯u is a polynomial of degree < η, and by (10) it vanishes
496
+ at x to order η which yields the desired ¯u = 0.
497
+
498
+ We return to the discussion of the singular product a(u)∂2
499
+ 1u, in its
500
+ simplest form of
501
+ v∂2
502
+ 1v = ∂2
503
+ 1
504
+ 1
505
+ 2v2 − (∂1v)2.
506
+ While in view of Lemma 1 the first r. h. s. term is well-defined as
507
+ a random Schwartz distribution, we now argue that the second term
508
+ 6where xn := xn1
509
+ 1 xn2
510
+ 2
511
+
512
+ 8
513
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
514
+ diverges. Indeed, applying ∂1 to the representation formula (13), so
515
+ that the constant Taylor term drops out, we have
516
+ ∂1v =
517
+ ˆ ∞
518
+ 0
519
+ dt∂1(−∂2 − ∂2
520
+ 1)ξt.
521
+ (16)
522
+ Hence for space-time white noise
523
+ E(∂1v(x))2
524
+ (16)
525
+ =
526
+ ˆ ∞
527
+ 0
528
+ dt
529
+ ˆ ∞
530
+ 0
531
+ ds E
532
+
533
+ ∂1(−∂2 − ∂2
534
+ 1)ξt(x)∂1(−∂2 − ∂2
535
+ 1)ξs(x)
536
+
537
+ (12)
538
+ =
539
+ ˆ ∞
540
+ 0
541
+ dt
542
+ ˆ ∞
543
+ 0
544
+ ds
545
+ ˆ
546
+ R2 dy ∂1(−∂2 − ∂2
547
+ 1)ψt(x − y)∂1(−∂2 − ∂2
548
+ 1)ψs(x − y)
549
+ (6)
550
+ =
551
+ ˆ ∞
552
+ 0
553
+ dt
554
+ ˆ ∞
555
+ 0
556
+ ds ∂2
557
+ 1(∂2
558
+ 2 − ∂4
559
+ 1)ψs+t(0)
560
+ (8)
561
+
562
+ ˆ ∞
563
+ 0
564
+ dt
565
+ ˆ ∞
566
+ 0
567
+ ds
568
+ 4√
569
+ t + s
570
+ −D−6.
571
+ Note that since 1
572
+ 4(−D−6) < −2 for D = 3, the double integral diverges.
573
+ This divergence arises from t ↓ 0 and s ↓ 0, that is, from small space-
574
+ time scales, and thus is called an ultra-violet (UV) divergence. A quick
575
+ fix is to introduce an UV cut-off, which for instance can be implemented
576
+ by mollifying ξ. Using the semi-group convolution ξτ specifies the UV
577
+ cut-off scale to be of the order of
578
+ 4√τ. It is easy to check that in this
579
+ case
580
+ E(∂1v(x))2 ∼
581
+ ˆ ∞
582
+ τ
583
+ dt
584
+ ˆ ∞
585
+ τ
586
+ ds
587
+ 4√
588
+ t + s
589
+ −D−6 ∼ (
590
+ 4√τ)−1.
591
+ The goal is to modify the equation (1) by “counter terms” such that
592
+ • the solution manifold stays under control as the ultra-violet
593
+ cut-off τ ↓ 0,
594
+ • invariances of the solution manifold are preserved i.e. the solu-
595
+ tion manifold keeps as many symmetries as possible.
596
+ In view of the above discussion, we expect the coefficients of the counter
597
+ terms to diverge as the cut-off tends to zero.
598
+ 3. Symmetry-motivated postulates on the form of the
599
+ counter terms
600
+ In view of α ∈ (0, 1), u is a function while we think of all derivatives
601
+ ∂nu as being only Schwartz distributions. Hence it is natural to start
602
+ from the very general Ansatz that the counter term is a polynomial in
603
+ {∂nu}n̸=0 with coefficients that are general (local) functions in u:
604
+ (∂2 − ∂2
605
+ 1)u +
606
+
607
+ β
608
+ hβ(u)
609
+
610
+ n̸=0
611
+ ( 1
612
+ n!∂nu)β(n) = a(u)∂2
613
+ 1u + ξ,
614
+ (17)
615
+
616
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
617
+ 9
618
+ where β runs over all multi-indices7 in n ̸= 0 and n! := (n1!)(n2!). For
619
+ simplicity of this heuristic discussion, we drop the regularization on ξ
620
+ and don’t index the counter term with τ.
621
+ Only counter terms that have an order strictly below the order of the
622
+ leading ∂2 − ∂2
623
+ 1 are desirable, so that one postulates that the sum in
624
+ (17) restricts to those multi-indices for which
625
+ |β|p :=
626
+
627
+ n̸=0
628
+ |n|β(n) < 2
629
+ where
630
+ |n| := n1 + 2n2.
631
+ (18)
632
+ This leaves only β = 0 and β = e(1,0), where the latter means β(n) =
633
+ δ(1,0)
634
+ n
635
+ , so that (17) collapses to
636
+ (∂2 − ∂2
637
+ 1)u + h(u) + h′(u)∂1u = a(u)∂2
638
+ 1u + ξ.
639
+ (19)
640
+ One also postulates that h and h′ depend on the noise ξ only through
641
+ its law / distribution / ensemble, hence are deterministic. Since we
642
+ assume that the law is invariant under space-time translation, i. e. is
643
+ stationary, it was natural to postulate that h and h′ do not explicitly
644
+ depend on x, hence are homogeneous.
645
+ Reflection symmetry.
646
+ Let us now assume that the law of ξ is
647
+ invariant under
648
+ space-time translation y �→ y + x,
649
+ space reflection y �→ (−y1, y2).
650
+ (20)
651
+ We now argue that under this assumption, it is natural to postulate
652
+ that the term h′(u)∂1u in (19) is not present, so that we are left with
653
+ (∂2 − ∂2
654
+ 1)u + h(u) = a(u)∂2
655
+ 1u + ξ.
656
+ (21)
657
+ To this purpose, let x ∈ R2 be arbitrary yet fixed, and consider the
658
+ reflection at the line {y1 = x1} given by Ry = (2x1 − y1, y2), which
659
+ by pull back acts on functions as ˜u(y) := u(Ry). Since (1) features no
660
+ explicit y-dependence, and only involves even powers of ∂1, which like
661
+ ∂2 commute with R, we have
662
+ (u, ξ) satisfies (1)
663
+ =⇒
664
+ (u(R·), ξ(R·)) satisfies (1).
665
+ (22)
666
+ Since we postulated that h and h′ depend on ξ only via its law, and
667
+ since in view of the assumption (20), ˜ξ = ξ(R·) has the same law as ξ,
668
+ it is natural to postulate that the symmetry (22) extends from (1) to
669
+ (19). Spelled out, this means that (19) implies
670
+ (∂2 − ∂2
671
+ 1)˜u + h(˜u) + h′(˜u)∂1˜u = a(˜u)∂2˜u + ˜ξ.
672
+ Evaluating both identities at y = x, and taking the difference, we get
673
+ for any solution of (19) that h′(u(x))∂1u(x) = h′(u(x))(−∂1u(x)), and
674
+ thus h′(u(x))∂1u(x) = 0, as desired.
675
+ 7which associate to every index n a β(n) ∈ N0 such that β(n) vanishes for all
676
+ but finitely many n’s
677
+
678
+ 10
679
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
680
+ Covariance under u-shift. We now come to our most crucial pos-
681
+ tulate, which restricts how the nonlinearity h depends on the non-
682
+ linearity / constitutive law a. Hence we no longer think of a single
683
+ nonlinearity a, but consider all non-linearities at once, in the spirit
684
+ of rough paths. This point of view reveals another invariance of (1),
685
+ namely for any shift v ∈ R
686
+ (u, a) satisfies (1)
687
+ =⇒
688
+ (u − v, a(· + v)) satisfies (1).
689
+ (23)
690
+ A priori, h is a function of the u-variable that has a functional de-
691
+ pendence on a, as denoted by h = h[a](u).
692
+ We postulate that the
693
+ symmetry (23) extends from (1) to (21). This is the case provided we
694
+ have the following shift-covariance property
695
+ h[a](u + v) = h[a(· + v)](u)
696
+ for all u ∈ R.
697
+ (24)
698
+ This property can also be paraphrased as: Whatever algorithm one
699
+ uses to construct h from a, it should not depend on the choice of origin
700
+ in what is just an affine space R ∋ u. Property (24) implies that the
701
+ counter term is determined by a functional c = c[a] on the space of
702
+ nonlinearities a:
703
+ h[a](v) = c[a(· + v)].
704
+ (25)
705
+ Renormalization now amounts to choosing c such that the solution
706
+ manifold stays under control as the UV regularization of ξ fades away.
707
+ 4. Algebrizing the counter term
708
+ In this section, we algebrize the relationship between a and the counter
709
+ term h given by a functional c as in (25). To this purpose, we introduce
710
+ the following coordinates8 on the space of analytic functions a of the
711
+ variable u:
712
+ zk[a] := 1
713
+ k!
714
+ dka
715
+ duk (0)
716
+ for k ≥ 0.
717
+ (26)
718
+ These are made such that by Taylor’s
719
+ a(u) =
720
+
721
+ k≥0
722
+ ukzk[a]
723
+ for a ∈ R[u],
724
+ (27)
725
+ where R[u] denotes the algebra of polynomials in the single variable u
726
+ with coefficients in R.
727
+ We momentarily specify to functionals c on the space of analytic a’s
728
+ that can be represented as polynomials in the (infinitely many) vari-
729
+ ables {zk}k≥0. This leads us to consider the algebra R[zk] of polynomials
730
+ in the variables zk with coefficients in R. The monomials
731
+ zβ :=
732
+
733
+ k≥0
734
+ zβ(k)
735
+ k
736
+ (28)
737
+ 8where here and in the sequel k ≥ 0 stands short for k ∈ N0
738
+
739
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
740
+ 11
741
+ form a basis of this (infinite dimensional) linear space, where β runs
742
+ over all multi-indices9. Hence as a linear space, R[zk] can be seen as
743
+ the direct sum over the index set given by all multi-indices β, and we
744
+ think of c as being of the form
745
+ c[a] =
746
+
747
+ β
748
+ cβzβ[a]
749
+ for c ∈ R[zk].
750
+ (29)
751
+ Infinitesimal u-shift. Given a shift v ∈ R, for ˜u := u − v and
752
+ ˜a := a(· + v) we have ˜a(˜u) = a(u). This leads us to study the mapping
753
+ a �→ a(· + v) which provides an action/representation of the group
754
+ R ∋ v on the set R[u] ∋ a. Note that for c ∈ R[zk] and a ∈ R[u], the
755
+ function R ∋ v �→ c[a(·+v)] = �
756
+ β cβ
757
+
758
+ k≥0( 1
759
+ k!
760
+ dka
761
+ du (v))β(k) is polynomial.
762
+ Thus
763
+ (D(0)c)[a] = d
764
+ dv |v=0c[a(· + v)]
765
+ (30)
766
+ is well-defined, linear in c and even a derivation10, meaning that Leib-
767
+ niz’ rule holds
768
+ (D(0)cc′) = (D(0)c)c′ + c(D(0)c′).
769
+ (31)
770
+ The latter implies that D(0) is determined by its value on the co-
771
+ ordinates zk, which by definitions (26) and (30) is given by D(0)zk
772
+ = (k + 1)zk+1. Hence D(0) has to agree with the following derivation
773
+ on the algebra R[zk]
774
+ D(0) =
775
+
776
+ k≥0
777
+ (k + 1)zk+1∂zk,
778
+ (32)
779
+ which is well defined since the sum is effectively finite when applied to
780
+ a monomial.
781
+ Representation of counter term. Iterating (30) we obtain by
782
+ induction in l ≥ 0 for c ∈ R[zk] and a ∈ R[u]
783
+ dl
784
+ dvl |v=0c[a(· + v)] = ((D(0))lc)[a]
785
+ and thus by Taylor’s (recall that v �→ c[a(· + v)] is polynomial)
786
+ c[a(· + v)] =
787
+ � �
788
+ l≥0
789
+ 1
790
+ l!vl(D(0))lc
791
+
792
+ [a].
793
+ (33)
794
+ 9which means they associate a frequency β(k) ∈ N0 to every k ≥ 0 such that all
795
+ but finitely many β(k)’s vanish
796
+ 10the index (0) is not necessary for these lecture notes, since we do not appeal
797
+ to the other derivations {D(n)}n̸=0 from [14, 15], we keep it here for consistency
798
+ with these papers
799
+
800
+ 12
801
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
802
+ We combine (33) with (25) to obtain the representation
803
+ h[a](v) =
804
+ � �
805
+ l≥0
806
+ 1
807
+ l!vl(D(0))lc
808
+
809
+ [a].
810
+ (34)
811
+ Hence our goal is to determine the coefficients {cβ}β in (29), which
812
+ typically will blow up as τ ↓ 0.
813
+ 5. Algebrizing the solution manifold: The centered
814
+ model
815
+ The purpose of this section is to motivate the notion of a centered
816
+ model; the motivation will be in parts formal.
817
+ Parameterization of the solution manifold. If a ≡ 0 it follows
818
+ from (24) that h is a (deterministic) constant. We learned from the
819
+ discussion after Lemma 1 that – given a base point x – there is a
820
+ distinguished solution v (with v(x) = 0). Hence we may canonically
821
+ parameterize a general solution u of (21) via u = v + p, by space-
822
+ time functions p with (∂2 − ∂2
823
+ 1)p = 0. Such p are necessarily analytic.
824
+ Having realized this, it is convenient11 to free oneself from the constraint
825
+ (∂2 − ∂2
826
+ 1)p = 0, which can be done at the expense of relaxing (21) to
827
+ (∂2 − ∂2
828
+ 1)v = ξ + q
829
+ for some analytic space-time function q.
830
+ (35)
831
+ Since we think of ξ as being rough while q is infinitely smooth, this
832
+ relaxation is still constraining v.
833
+ The implicit function theorem suggests that this parameterization (lo-
834
+ cally) persists in the presence of a sufficiently small analytic nonlin-
835
+ earity a: The nonlinear manifold of all space-time functions u that
836
+ satisfy
837
+ (∂2 − ∂2
838
+ 1)u + h(u) = a(u)∂2
839
+ 1u + ξ + q
840
+ for some analytic space-time function q
841
+ (36)
842
+ is still parameterized by space-time analytic functions p. We now return
843
+ to the point of view of Section 3 of considering all nonlinearities a at
844
+ once, meaning that we consider the (still nonlinear) space of all space-
845
+ time functions that satisfy (36) for some analytic nonlinearity a. We
846
+ want to capitalize on the symmetry (23), which extends from (1) to (21)
847
+ and to (36). We do so by considering the above space of u’s modulo
848
+ constants, which we implement by focusing on increments u − u(x).
849
+ Summing up, it is reasonable to expect that the space of all space-time
850
+ functions u, modulo space-time constants, that satisfy (36) for some
851
+ analytic nonlinearity a and space-time function q (but at fixed ξ), is
852
+ parameterized by pairs (a, p) with p(x) = 0.
853
+ 11otherwise, the coordinates z(2,0) and z(0,1) defined in (38) would be redundant
854
+ on p-space
855
+
856
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
857
+ 13
858
+ Formal series representation. In line with the term-by-term ap-
859
+ proach from physics, we write the increment u(y)−u(x) as a (typically
860
+ divergent) power series
861
+ u(y) − u(x)
862
+ =
863
+
864
+ β
865
+ Πxβ(y)
866
+
867
+ k≥0
868
+ � 1
869
+ k!
870
+ dka
871
+ duk (u(x))
872
+ �β(k) �
873
+ n̸=0
874
+ � 1
875
+ n!∂np(x)
876
+ �β(n),
877
+ (37)
878
+ where β runs over all multi-indices in k ≥ 0 and n ̸= 0. Introducing
879
+ coordinates on the space of analytic space-time functions p with p(0) =
880
+ 0 via12
881
+ zn[p] = 1
882
+ n!∂np(0)
883
+ for n ̸= 0,
884
+ (38)
885
+ (37) can be more compactly written as
886
+ u(y) = u(x) +
887
+
888
+ β
889
+ Πxβ(y)zβ[a(· + u(x)), p(· + x) − p(x)].
890
+ (39)
891
+ This is reminiscent of Butcher series in the analysis of ODE discretiza-
892
+ tions.
893
+ Recall from above that for a ≡ 0 we have the explicit parameterization
894
+ u − u(x) = v + p
895
+ (40)
896
+ with the distinguished solution v of the linear equation. Hence from
897
+ setting a ≡ 0 and p ≡ 0 in (37), we learn that Πx0 = v. From keeping
898
+ a ≡ 0 but letting p vary we then deduce that for all multi-indices β ̸= 0
899
+ which satisfy β(k) = 0 for all k ≥ 0 we must have13
900
+ Πxβ(y) =
901
+
902
+ (y − x)n
903
+ provided β = en
904
+ 0
905
+ else
906
+
907
+ .
908
+ (41)
909
+ Hierarchy of linear equations.
910
+ The collection of coefficients
911
+ {Πxβ(y)}β from (39) is an element of the direct product with the same
912
+ index set as the direct sum R[zk, zn]. Hence the direct product inherits
913
+ the multiplication of the polynomial algebra
914
+ (ππ′) ¯β =
915
+
916
+ β+β′= ¯β
917
+ πβπ′
918
+ β′,
919
+ (42)
920
+ and is denoted as the (well-defined) algebra R[[zk, zn]] of formal power
921
+ series; we denote by 1 its unit element. We claim that in terms of (39),
922
+ (36) assumes the form of
923
+ (∂2 − ∂2
924
+ 1)Πx = Π−
925
+ x
926
+ up to space-time analytic functions
927
+ (43)
928
+ 12where here and in the sequel n ̸= 0 stands short for n ∈ N2
929
+ 0 − {(0, 0)}
930
+ 13where we recall that β = en denotes the multi-index with β(m) = δn
931
+ m next to
932
+ β(k) = 0
933
+
934
+ 14
935
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
936
+ where
937
+ Π−
938
+ x :=
939
+
940
+ k≥0
941
+ zkΠk
942
+ x∂2
943
+ 1Πx −
944
+
945
+ l≥0
946
+ 1
947
+ l!Πl
948
+ x(D(0))lc + ξτ1,
949
+ (44)
950
+ as an identity in formal power series in zk, zn with coefficients that
951
+ are continuous space-time functions. We shall argue below that (44)
952
+ is effectively, i. e. componentwise, well-defined despite the two infinite
953
+ sums, and despite extending from c ∈ R[zk] to c ∈ R[[zk]]. Moreover,
954
+ as will become clear by (64), the β-component of (44) contains on the
955
+ r. h. s. only terms Πxβ′ for “preceding” multi-indices β′ – hence (43)
956
+ describes a hierarchy of equations.
957
+ Here comes the formal argument that relates {∂2, ∂2
958
+ 1}u, a(u), and h(u),
959
+ to {∂2, ∂2
960
+ 1}Πx[˜a, ˜p], (�
961
+ k≥0 zkΠk
962
+ x)[˜a, ˜p], and (�
963
+ l≥0
964
+ 1
965
+ l!Πl
966
+ x(D(0))lc)[˜a, ˜p], re-
967
+ spectively. Here we have set for abbreviation ˜a = a(· + u(x)) and ˜p
968
+ = p(· + x) − p(x). It is based on (39), which can be compactly written
969
+ as u(y) = u(x) + Πx[˜a, ˜p](y). Hence the statement on {∂2, ∂2
970
+ 1}u follows
971
+ immediately. Together with a(u(y)) = ˜a(u(y) −u(x)), this also implies
972
+ by (27) the desired
973
+ a(u(y)) =
974
+ � �
975
+ k≥0
976
+ zkΠk
977
+ x(y)
978
+
979
+ [˜a, ˜p].
980
+ Likewise by (24), we have h[a](u(y)) = h[˜a](u(y) − u(x)), so that by
981
+ (34), we obtain the desired
982
+ h[a](u(y)) =
983
+ � �
984
+ l≥0
985
+ 1
986
+ l!Πl
987
+ x(y)(D(0))lc
988
+
989
+ [˜a, ˜p].
990
+ Finiteness properties. The next lemma collects crucial algebraic
991
+ properties.
992
+ Lemma 2. The derivation D(0) extends from R[zk] to R[[zk]].
993
+ Moreover, for π, π′ ∈ R[[zk, zn]], c ∈ R[[zk]], and ξ ∈ R,
994
+ π− :=
995
+
996
+ k≥0
997
+ zkπkπ′ −
998
+
999
+ l≥0
1000
+ 1
1001
+ l!πl(D(0))lc + ξ1 ∈ R[[zk, zn]]
1002
+ (45)
1003
+ is well-defined, in the sense that the two sums are componentwise finite.
1004
+ Finally, for
1005
+ [β] :=
1006
+
1007
+ k≥0
1008
+ kβ(k) −
1009
+
1010
+ n̸=0
1011
+ β(n)
1012
+ (46)
1013
+
1014
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
1015
+ 15
1016
+ we have the implication
1017
+ πβ = π′
1018
+ β = 0
1019
+ unless
1020
+ [β] ≥ 0 or β = en for some n ̸= 0
1021
+ =⇒
1022
+ π−
1023
+ β = 0
1024
+ unless
1025
+
1026
+
1027
+
1028
+ [β] ≥ 0 or
1029
+ β = ek + en1 + · · · + enk+1
1030
+ for some k ≥ 1 and n1, . . . , nk+1 ̸= 0.
1031
+
1032
+
1033
+  .
1034
+ (47)
1035
+ We note that for β as in the second alternative on the r. h. s. of (47),
1036
+ it follows from (41) that Π−
1037
+ xβ is a polynomial. Hence in view of the
1038
+ modulo in (43), we learn from (47) that we may assume
1039
+ Πxβ ≡ 0
1040
+ unless
1041
+ [β] ≥ 0 or β = en for some n ̸= 0.
1042
+ (48)
1043
+ Proof of Lemma 2. We first address the extension of D(0) and note
1044
+ that from (32) we may read off the matrix representation of D(0) ∈
1045
+ End(R[zk]) w. r. t. (28) given by
1046
+ (D(0))γ
1047
+ β = (D(0)zγ)β
1048
+ (32)
1049
+ =
1050
+
1051
+ k≥0
1052
+ (k + 1)
1053
+
1054
+ zk+1∂zkzγ�
1055
+ β
1056
+ (28)
1057
+ =
1058
+
1059
+ k≥0
1060
+ (k + 1)γ(k)
1061
+
1062
+ 1
1063
+ provided γ + ek+1 = β + ek
1064
+ 0
1065
+ otherwise
1066
+
1067
+ .
1068
+ (49)
1069
+ From this we read off that {γ|(D(0))γ
1070
+ β ̸= 0} is finite for every β, which
1071
+ implies that D(0) naturally extends from R[zk] to R[[zk]]. With help
1072
+ of (42) the derivation property (31) can be expressed coordinate-wise,
1073
+ and thus extends to R[[zk]].
1074
+ We now turn to (45), which component-wise reads
1075
+ π−
1076
+ β =
1077
+
1078
+ k≥0
1079
+
1080
+ ek+β1+···+βk+1=β
1081
+ πβ1 · · · πβkπ′
1082
+ βk+1
1083
+
1084
+
1085
+ l≥0
1086
+ 1
1087
+ l!
1088
+
1089
+ β1+···+βl+1=β
1090
+ πβ1 · · · πβl((D(0))lc)βl+1 + ξδ0
1091
+ β,
1092
+ (50)
1093
+ and claim that the two sums are effectively finite. For the first term
1094
+ of the r. h. s. this is obvious since thanks to the presence of14 ek in
1095
+ ek + β1 + · · · + βk+1 = β, for fixed β there are only finitely many k ≥ 0
1096
+ for which this relation can be satisfied.
1097
+ In preparation for the second r. h. s. term of (50) we now establish that
1098
+ ((D(0))l)γ
1099
+ β = 0
1100
+ unless
1101
+ [β]0 = [γ]0 + l,
1102
+ (51)
1103
+ where we introduced the scaled length [γ]0 := �
1104
+ k≥0 kγ(k) ∈ N0. The
1105
+ argument for (51) proceeds by induction in l ≥ 0. It is tautological
1106
+ for the base case l = 0. In order to pass from l to l + 1 we write
1107
+ 14γ = ek denotes the multi-index with γ(l) = δk
1108
+ l next to γ(n) = 0
1109
+
1110
+ 16
1111
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
1112
+ ((D(0))l+1)γ
1113
+ β = �
1114
+ β′((D(0))l)β′
1115
+ β (D(0))γ
1116
+ β′; by induction hypothesis, the first
1117
+ factor vanishes unless [β]0 = [β′]0 + l. We read off (49) that the second
1118
+ factor vanishes unless [β′]0 = [γ]0 + 1, so that the product vanishes
1119
+ unless [β]0 = [γ]0 + (l + 1), as desired.
1120
+ Equipped with (51) we now turn to the second r. h. s. term of (50) and
1121
+ note that ((D(0))lc)βk+1 vanishes unless l ≤ [βk+1]0 ≤ [β]0, which shows
1122
+ that also here, only finitely many l ≥ 0 contribute for fixed β.
1123
+ We turn to the proof of (47). We use (50) and give the proof for every
1124
+ summand separately. For the first term on the r. h. s. of (50) we obtain
1125
+ by additivity of [·] that [β] = k+[β1]+· · ·+[βk+1]. Note that πβi is only
1126
+ non vanishing if [βi] ≥ −1. If at least one of the β1, . . . , βk+1 satisfies
1127
+ [βi] ≥ 0, we obtain therefore [β] ≥ k−k = 0. For the second r. h. s. term
1128
+ in (50) we appeal to (51): Since D(0) doesn’t affect the zn components,
1129
+ (51) extends from [·]0 to [·].
1130
+ Together with c ∈ R[[zk]] this yields
1131
+ [βl+1] ≥ l. Hence as above [β] = [β1]+· · ·+[βl+1] ≥ −l+[βl+1] ≥ 0.
1132
+
1133
+ Homogeneity.
1134
+ We return to a heuristic discussion.
1135
+ Provided we
1136
+ include, like for (23), a into our considerations, the original equation
1137
+ (1) has a scaling symmetry: Considering for s ∈ (0, ∞) the parabolic
1138
+ space-time rescaling Sy = (sy1, s2y2), we have for any exponent α
1139
+ (u, ξ, a) satisfies (1)
1140
+ =⇒
1141
+
1142
+ s−αu(S·), s2−αξ(S·), a(sα·)
1143
+
1144
+ =: (˜u, ˜ξ, ˜a) satisfies (1).
1145
+ (52)
1146
+ Suppose the scaling transformation ξ �→ ˜ξ preserves the law, which for
1147
+ white noise is the case with α − 2 = −D
1148
+ 2 , i. e. α = 1
1149
+ 2. Since in view of
1150
+ Section 3, the counter term only depends on the law, it is natural to
1151
+ postulate, in line with that section, that the solution manifold of the
1152
+ renormalized problem inherits this invariance15.
1153
+ It is also natural to postulate that the parameterization by the p’s
1154
+ (given a base point x) is consistent with (52) in the sense that p trans-
1155
+ forms as u, i. e. we have invariance under
1156
+ (u, ξ, a, x, p) �→ (˜u, ˜ξ, ˜a, ˜x := S−1x, ˜p := s−αp(S·)).
1157
+ We now appeal to the series expansion (37), both as it stands and
1158
+ with (x, y, u, ξ, a, p) replaced by (˜x, ˜y := S−1y, ˜u, ˜ξ, ˜a, ˜p). Because of
1159
+ u(y) − u(x) = sα(˜u(˜y) − ˜u(˜x)), we obtain a relation between the two
1160
+ right-hand sides. It is natural to postulate that the coefficients {Π·,β}β
1161
+ are individually consistent with this invariance, leading to
1162
+ ΠSxβ[ξ](Sy) = s|β|Πxβ[s2−αξ(S·)](y),
1163
+ (53)
1164
+ 15since this scale invariance in law is not consistent with the mollification ξτ this
1165
+ discussion pertains to the limiting solution manifold
1166
+
1167
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
1168
+ 17
1169
+ where the “homogeneity” |β| of the multi-index β is given by
1170
+ |β| := α(1 + [β]) + |β|p,
1171
+ (54)
1172
+ cf. (18) and (46). We note that
1173
+ |en| = |n|
1174
+ (55)
1175
+ so that (54) is consistent with (41).
1176
+ Appealing once more to the invariance in law of ξ under (52), we obtain
1177
+ from (53)
1178
+ the law of s−|β|ΠSx β(Sy) does not depend on s ∈ (0, ∞).
1179
+ (56)
1180
+ By the invariance of the (original) solution manifold under (u, ξ) �→
1181
+ (˜u := u(·+z), ˜ξ := ξ(·+z)), which by our assumption (20) is passed on
1182
+ to the renormalized solution manifold, it is natural to impose that the
1183
+ parameterization is invariant under (u, ξ, x, p) �→ (˜u, ˜ξ, x + z, p(· + z)),
1184
+ and that the coefficients in (39) are individually consistent with this
1185
+ invariance, so that we likewise have
1186
+ the law of Πx+z β(y + z) does not depend on z ∈ R2.
1187
+ (57)
1188
+ Specifying to x = 0, the invariance (56) implies that E
1189
+ 1
1190
+ p|Π0β(y)|p de-
1191
+ pends on y only through
1192
+ y
1193
+ |y|. From the invariance (57) we thus learn
1194
+ that E
1195
+ 1
1196
+ p |Πxβ(y)|p depends on x, y only through
1197
+ y−x
1198
+ |y−x|. Since
1199
+ y−x
1200
+ |y−x| has
1201
+ compact range, this suggest that
1202
+ E
1203
+ 1
1204
+ p |Πxβ(y)|p ≲ |y − x||β|,
1205
+ which is our main result, see (60) in the next section.
1206
+ The scaling invariance (52) also connects to the notion of “subcritical-
1207
+ ity” which is often referred to in the realm of singular SPDEs. Loosely
1208
+ speaking, it means that by zooming in on small scales, the nonlinear
1209
+ term becomes negligible. Indeed, as can be seen from (52), the rescaled
1210
+ nonlinearity ˜a converges to the constant a(0) in the limit s ↓ 0, i. e. the
1211
+ SPDE (1) turns into a linear one. This is true iff α > 0, and provides
1212
+ the reason for restricting to α > 0 in the assumption of Theorem 1,
1213
+ which is the sub-critical regime for (1).
1214
+ 6. The main result: A stochastic estimate of the
1215
+ centered model
1216
+ Theorem 1. Suppose the law of ξ is invariant under (20); suppose that
1217
+ it satisfies a spectral gap inequality (87) with exponent α ∈ (max{0, 1−
1218
+ D
1219
+ 4 }, 1) \ Q.
1220
+ Then given τ > 0, there exists a deterministic c ∈ R[[zk]], and for every
1221
+ x ∈ R2, a random16 Πx ∈ C2[[zk, zn]], and a random Π−
1222
+ x ∈ C0[[zk, zn]]
1223
+ 16by this we mean a formal power series in zk, zn with values in the twice con-
1224
+ tinuously differentiable space-time functions
1225
+
1226
+ 18
1227
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
1228
+ that are related by (44) and
1229
+ (∂2 − ∂2
1230
+ 1)Πxβ = Π−
1231
+ xβ + polynomial of degree ≤ |β| − 2,
1232
+ (58)
1233
+ and that satisfy (41), the population condition (48) and
1234
+ cβ = 0
1235
+ unless
1236
+ |β| < 2.
1237
+ (59)
1238
+ Moreover, we have the estimates
1239
+ E
1240
+ 1
1241
+ p|Πxβ(y)|p ≲β,p |y − x||β|,
1242
+ (60)
1243
+ E
1244
+ 1
1245
+ p|Π−
1246
+ xβt(y)|p ≲β,p (
1247
+ 4√
1248
+ t)α−2(
1249
+ 4√
1250
+ t + |y − x|)|β|−α.
1251
+ (61)
1252
+ The important feature is that the constants in (60) and (61) are uniform
1253
+ in τ ↓ 0.
1254
+ We remark that we may pass from (61) to (60) by Lemma 1. Indeed,
1255
+ because of (48) we may restrict to β with [β] ≥ 0. In this case, by our
1256
+ assumption α ̸∈ Q,
1257
+ [β] ≥ 0
1258
+ (54)
1259
+ =⇒
1260
+ |β| ̸∈ Z,
1261
+ (62)
1262
+ next to |β| ≥ α. Hence we may indeed apply Lemma 1 with η = |β|
1263
+ and (61) as input. The output yields a Πxβ satisfying (58) and (60).
1264
+ Uniqueness and (implicit) BPHZ renormalization. The con-
1265
+ struction of Πx in [15] proceeds by an inductive algorithm in β. The
1266
+ ordering17 on the multi-indices is provided by
1267
+ (63)
1268
+ |β|≺ := |β| + λβ(0)
1269
+ for fixed λ ∈ (0, α),
1270
+ and we will write γ ≺ β for |γ|≺ < |β|≺. As opposed to the ordering
1271
+ provided by the homogeneity, ≺ allows for the triangular structure:
1272
+ (64)
1273
+ Π−
1274
+ xβ − cβ depends on (Πxγ, cγ) only through γ with γ ≺ β,
1275
+ which can be easily checked on the component-wise level (50). More-
1276
+ over, (63), as opposed to the ordering by homogeneity, is coercive: For
1277
+ fixed β there are only finitely many γ with γ ≺ β, see (101), which is
1278
+ important for the estimates.
1279
+ We now argue that within this induction, (c, Πx, Π−
1280
+ x ) is determined.
1281
+ Indeed, the uniqueness statement of Lemma 1 implies that for given β,
1282
+ Πxβ is determined by Π−
1283
+ xβ. According to (64), Π−
1284
+ xβ − cβ is determined
1285
+ by the previous steps. Finally, we note that provided |β| < 2, we have
1286
+ |EΠ−
1287
+ xβt(x)| ≤ E|Π−
1288
+ xβt(x)|
1289
+ (61)
1290
+ ≲ (
1291
+ 4√
1292
+ t)|β|−2 t↑∞
1293
+ → 0,
1294
+ (65)
1295
+ 17this ordering coincides with the one chosen in [13] but it slightly differs from
1296
+ the one in [15], which is imposed by the restricted triangularity of dΓ∗ in Section
1297
+ 7; for simplicity we stick to (63)
1298
+
1299
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
1300
+ 19
1301
+ so that cβ, because it is deterministic18 may be recovered from cβ =
1302
+ − limt↑∞ E(Π−
1303
+ xβ − cβ)t(x).
1304
+ Hence also cβ is determined.
1305
+ Fixing the
1306
+ counter term by making an expectation19 vanish like in (65) corre-
1307
+ sponds to what Hairer assimilates to a BPHZ renormalization. See [3,
1308
+ Theorem 6.18] for the form BPHZ renormalization takes within regu-
1309
+ larity structures.
1310
+ Mission accomplished. Returning to the end of Section 2, we may
1311
+ claim “mission accomplished”:
1312
+ • On the one hand, the form of the counter terms preserve a
1313
+ number of symmetries of the original solution manifold: shift
1314
+ in x, reflection in x1, shift in u, and to some extend are guided
1315
+ by scaling in x.
1316
+ • On the other hand, in a term-by-term sense as encoded by (37),
1317
+ the solution manifold of the renormalized equation stays under
1318
+ control as τ ↓ 0, cf. (60) and (61).
1319
+ Moreover, the constants cβ = cτ
1320
+ β that determine the counter term via
1321
+ (34) are (canonically) determined by the large-scale part of the estimate
1322
+ (61).
1323
+ As discussed in the introduction, the connection between this term-by-
1324
+ term approach to the solution manifold and the solution of an actual
1325
+ initial/boundary value problem is provided by the second part of reg-
1326
+ ularity structures. This second part, a fixed point argument based on
1327
+ a truncation of (37) to a finite sum20, is not addressed in these lecture
1328
+ notes.
1329
+ 7. Malliavin derivative and Spectral gap (SG)
1330
+ In view of the discussion at the end of the statement of Theorem 1,
1331
+ the main issue is the estimate (61) of Π−
1332
+ xβ. Indeed, its definition of
1333
+ (44) still contains the singular product Πk
1334
+ x∂2
1335
+ 1Πx and the collection of
1336
+ deterministic constants c that diverge as the UV regularization fades
1337
+ away. Hence we seek a relation between Π−
1338
+ x and Πx that is more stable
1339
+ than (44); in fact, it will be a relation between the families {Π−
1340
+ x }x and
1341
+ {Πx}x based on symmetries under a change of the base point x. This
1342
+ relation is formulated on the level of the derivative w. r. t. noise ξ,
1343
+ also known as the Malliavin derivative. We start by motivating this
1344
+ approach.
1345
+ Heuristic discussion of a stable relation {Πx}x �→ {Π−
1346
+ x }x. Let
1347
+ δ denote the operation of taking the derivative of an object like Πxβ(y),
1348
+ 18and independent of the base point x
1349
+ 19in our case it is a space-time next to an ensemble average
1350
+ 20by restricting to homogeneities |β| < 2; in our quasi-linear case, the sum stays
1351
+ infinite w. r. t. the z0-variable, but one has analyticity in that variable since 1 + z0
1352
+ plays the role of a constant elliptic coefficient
1353
+
1354
+ 20
1355
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
1356
+ which is a functional of ξ, in direction of an infinitesimal variation δξ
1357
+ of the latter21.
1358
+ Clearly, since cβ is deterministic, we have δcβ = 0.
1359
+ However, applying δ to (a component of) (44) does not eliminate c
1360
+ because of the specific way c enters (44), which is dictated by the
1361
+ fundamental symmetry (25). However, when evaluating (44) at the
1362
+ base point x itself and appealing to the built-in
1363
+ (66)
1364
+ Πx(x) = 0,
1365
+ see (37) or (60), it collapses to
1366
+ Π−
1367
+ x (x) = z0∂2
1368
+ 1Πx(x) − c + ξτ(x)1.
1369
+ (67)
1370
+ This isolates c so that it can be eliminated by applying δ:
1371
+ δΠ−
1372
+ x (x) = z0∂2
1373
+ 1δΠx(x) + δξτ(x)1.
1374
+ (68)
1375
+ Clearly, (68) is impoverished in the sense that the active point coincides
1376
+ with the base point.
1377
+ Instead of attempting to modify the active point, the idea is to modify
1378
+ the base point from x to y. Such a change of base point, which will be
1379
+ rigorously introduced in Section 8, amounts to a change of coordinates
1380
+ in the heuristic representation (39):
1381
+ u =
1382
+ � u(x) + �
1383
+ β Πxβzβ[a(· + u(x)), px],
1384
+ u(y) + �
1385
+ β Πyβzβ[a(· + u(y)), py],
1386
+ (69)
1387
+ for some polynomials px, py vanishing at the origin. The form in which
1388
+ the u-shift appears in (69) suggests that this change of coordinates can
1389
+ be algebrized by an algebra endomorphism22 Γ∗
1390
+ yx of R[[zk, zn]] with the
1391
+ properties
1392
+ Πy = Γ∗
1393
+ yxΠx + Πy(x)
1394
+ and
1395
+ Γ∗
1396
+ yx =
1397
+
1398
+ l≥0
1399
+ 1
1400
+ l!Πl
1401
+ y(x)(D(0))l on R[[zk]],
1402
+ (70)
1403
+ see the discussion of finite u-shifts around (34). Recall that an alge-
1404
+ bra endomorphism Γ∗
1405
+ yx is a linear map from R[[zk, zn]] to R[[zk, zn]]
1406
+ satisfying
1407
+ (71)
1408
+ Γ∗
1409
+ yxππ′ = (Γ∗
1410
+ yxπ)(Γ∗
1411
+ yxπ′)
1412
+ for π, π′ ∈ R[[zk, zn]].
1413
+ We claim that (70) implies
1414
+ Π−
1415
+ y = Γ∗
1416
+ yxΠ−
1417
+ x .
1418
+ (72)
1419
+ 21in the Gaussian case, this would be an element of the Cameron-Martin space
1420
+ 22in a first reading, the star should be seen as mere notation; Γ∗
1421
+ yx is actually the
1422
+ algebraic dual of a linear endomorphism Γyx on the pre-dual space, see Lemma 3;
1423
+ it is Γyx that can be assimilated to the object denoted by the same symbol in
1424
+ regularity structures; for a concise reference see [14, Section 5.3]
1425
+
1426
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
1427
+ 21
1428
+ Indeed, applying Γ∗
1429
+ yx to definition (44) we obtain by (71)
1430
+ Γ∗
1431
+ yxΠ−
1432
+ x
1433
+ =
1434
+
1435
+ k≥0
1436
+ (Γ∗
1437
+ yxzk)(Γ∗
1438
+ yxΠx)k∂2
1439
+ 1Γ∗
1440
+ yxΠx −
1441
+
1442
+ l≥0
1443
+ 1
1444
+ l!(Γ∗
1445
+ yxΠx)lΓ∗
1446
+ yx(D(0))lc + ξτ1.
1447
+ We substitute Γ∗
1448
+ yxΠx according to the first item in (70), substitute
1449
+ Γ∗
1450
+ yxzk = �
1451
+ l≥0
1452
+ �k+l
1453
+ k
1454
+
1455
+ Πl
1456
+ y(x)zk+l and Γ∗
1457
+ yx(D(0))lc according to the second
1458
+ item in (70) and definition (32), and finally appeal to the binomial
1459
+ formula in both ensuing double sums to obtain (44) with x replaced by
1460
+ y, establishing (72).
1461
+ In view of the scaling (56) and the transformation (70) we expect that
1462
+ the laws of s|β|−|γ|(Γ∗
1463
+ yx)γ
1464
+ β and of (Γ∗
1465
+ SySx)γ
1466
+ β to be identical. On the other
1467
+ hand, we expect (Γ∗
1468
+ SySx)γ
1469
+ β to converge to (Γ∗
1470
+ 00)γ
1471
+ β as s ↓ 0, and we expect
1472
+ Γ∗
1473
+ 00 to be the identity. This suggests strict triangularity:
1474
+ (Γ∗
1475
+ yx − id)γ
1476
+ β = 0
1477
+ unless
1478
+ |γ| < |β|.
1479
+ (73)
1480
+ We claim that applying Γ∗
1481
+ yx to (68), we obtain23
1482
+ δΠ−
1483
+ y (x) − (δΓ∗
1484
+ yx)Π−
1485
+ x (x)
1486
+ =
1487
+
1488
+ k≥0
1489
+ zkΠk
1490
+ y(x)∂2
1491
+ 1
1492
+
1493
+ δΠy − δΠy(x) − (δΓ∗
1494
+ yx)Πx
1495
+
1496
+ (x) + δξτ(x)1.
1497
+ (74)
1498
+ Since by (73), δΓ∗
1499
+ yx is strictly triangular, (74) provides an inductive
1500
+ way of determining {Π−
1501
+ x }x (up to expectation) in terms of {Πx}x. Here
1502
+ comes the argument for (74): Applying Γ∗
1503
+ yx to the l. h. s. of (68) and
1504
+ using (72) in conjunction with Leibniz’ rule w. r. t. δ, we obtain the
1505
+ l. h. s. of (74). For the r. h. s. we first use the multiplicativity of Γ∗
1506
+ yx;
1507
+ according to the second item in (70) and (32) we have
1508
+ Γ∗
1509
+ yxz0 =
1510
+
1511
+ l≥0
1512
+ Πl
1513
+ y(x)zl.
1514
+ (75)
1515
+ To rewrite Γ∗
1516
+ yxδΠx, we apply δ to the first identity in (70). This estab-
1517
+ lishes (74).
1518
+ We now argue that from an analytical point of view, (74) is not quite
1519
+ adequate. Clearly, the r. h. s. of (74) still contains a potentially singular
1520
+ product of Πk
1521
+ y and ∂2
1522
+ 1(δΠy −δΠy(x) −(δΓ∗
1523
+ yx)Πx). Here, it is crucial that
1524
+ applying δ to Πy, which is a multi-linear expression in ξ, means replac-
1525
+ ing one of the instances of ξ by δξ. Now as we shall explain in the next
1526
+ subsection, δξ gains24 D
1527
+ 2 orders of regularity over ξ. However, since the
1528
+ other instances of ξ remain, the regularity of δΠy is not at face value
1529
+ better by D
1530
+ 2 orders over Πy, which is just H¨older continuous with expo-
1531
+ nent α. Hence we can only expect that δΠy is locally, i. e. near a base
1532
+ 23of course, the r. h. s. term δΠy(x) is effectively absent due to the derivative ∂2
1533
+ 1
1534
+ 24however on an L2 instead of a uniform scale
1535
+
1536
+ 22
1537
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
1538
+ point x, described – “modelled” in the jargon of regularity structures
1539
+ – to order D
1540
+ 2 + α in terms of Πx. The Taylor-remainder-like expression
1541
+ δΠy − δΠy(x) −(δΓ∗
1542
+ yx)Πx has the potential of expressing this modeled-
1543
+ ness. Hence the product of Πk
1544
+ y and ∂2
1545
+ 1(δΠy − δΠy(x) −(δΓ∗
1546
+ yx)Πx) has a
1547
+ chance of being well-defined provided α + ( D
1548
+ 2 + α − 2) > 0, which gives
1549
+ rise to the lower bound assumption α > 1 − D
1550
+ 4 in Theorem 1, which
1551
+ reduces to25 α >
1552
+ 1
1553
+ 4 for our D = 3. Since D
1554
+ 2 + α > 1, this only has
1555
+ a chance of working provided every β-component of (δΓ∗
1556
+ yx)Πx involves
1557
+ the affine function Πxe(1,0) = (· − x)1. However, this contradicts the
1558
+ (strict) triangularity (73) for |β| ≤ 1. Hence δΓ∗
1559
+ yx is not rich enough to
1560
+ describe all components of δΠy to the desired order near x.
1561
+ In view of the preceding discussion, we are forced to loosen the pop-
1562
+ ulation constraint (73).
1563
+ To this purpose, we replace the directional
1564
+ Malliavin derivative δΓ∗
1565
+ yx by some dΓ∗
1566
+ yx ∈ End(R[[zk, zn]]) in order to
1567
+ achieve
1568
+ δΠy − δΠy(x) − dΓ∗
1569
+ yxΠx = O(| · −x|
1570
+ D
1571
+ 2 +α).
1572
+ (76)
1573
+ In order to preserve the identity (74) in form of
1574
+ δΠ−
1575
+ y (x) − dΓ∗
1576
+ yxΠ−
1577
+ x (x)
1578
+ =
1579
+
1580
+ k≥0
1581
+ zkΠk
1582
+ y(x)∂2
1583
+ 1
1584
+
1585
+ δΠy − δΠy(x) − dΓ∗
1586
+ yxΠx
1587
+
1588
+ (x) + δξτ(x)1,
1589
+ (77)
1590
+ we need dΓ∗
1591
+ yx to inherit the algebraic properties of δΓ∗
1592
+ yx. More precisely,
1593
+ we impose that dΓ∗
1594
+ yx agrees with δΓ∗
1595
+ yx on the sub-algebra R[[zk]],
1596
+ dΓ∗
1597
+ yx = δΓ∗
1598
+ yx on R[[zk]],
1599
+ (78)
1600
+ and that dΓ∗
1601
+ yx is in the tangent space to the manifold of algebra mor-
1602
+ phisms in Γ∗
1603
+ yx, which means that for all π, π′ ∈ R[[zk, zn]]
1604
+ dΓ∗
1605
+ yxππ′ = (dΓ∗
1606
+ yxπ)(Γ∗
1607
+ yxπ′) + (Γ∗
1608
+ yxπ)(dΓ∗
1609
+ yxπ′).
1610
+ (79)
1611
+ Here is the argument on how to pass from (78) & (79) to (77). On the
1612
+ one hand, we apply δ to (44) to the effect of
1613
+ δΠ−
1614
+ y (x) =
1615
+
1616
+ k≥0
1617
+ zkδ
1618
+
1619
+ Πk
1620
+ y(x)
1621
+
1622
+ ∂2
1623
+ 1Πy(x) +
1624
+
1625
+ k≥0
1626
+ zkΠk
1627
+ y(x)∂2
1628
+ 1δΠy(x)
1629
+
1630
+
1631
+ l≥0
1632
+ 1
1633
+ l!δ
1634
+
1635
+ Πl
1636
+ y(x)
1637
+
1638
+ (D(0))lc + δξτ(x)1.
1639
+ (80)
1640
+ 25This is the analogy of rough path construction of fractional Brownian motion.
1641
+ For the case of fractional Brownian motion with Hurst parameter H, a rough path
1642
+ construction can be only implemented for any H > 1
1643
+ 4 by increasing the number
1644
+ of iterated integrals.
1645
+ However, the stochastic analysis to construct the iterated
1646
+ integrals fails for fractional Brownian motion of Hurst parameter H ≤ 1
1647
+ 4. See [5,
1648
+ Theorem 2].
1649
+
1650
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
1651
+ 23
1652
+ On the other hand, we apply dΓ∗
1653
+ yx to (67) to obtain by26 (79)
1654
+ dΓ∗
1655
+ yxΠ−
1656
+ x (x)
1657
+ = (dΓ∗
1658
+ yxz0)∂2
1659
+ 1Γ∗
1660
+ yxΠx(x) + (Γ∗
1661
+ yxz0)∂2
1662
+ 1dΓ∗
1663
+ yxΠx(x) − dΓ∗
1664
+ yxc.
1665
+ (81)
1666
+ We now argue that the first r. h. s. term of (80) is identical to the one
1667
+ in (81); indeed, by the first item in (70) we have ∂2
1668
+ 1Γ∗
1669
+ yxΠx = ∂2
1670
+ 1Πy. On
1671
+ the other hand, by (78) and the second item in (70) we have
1672
+ dΓ∗
1673
+ yx =
1674
+
1675
+ l≥0
1676
+ 1
1677
+ l!δ
1678
+
1679
+ Πl
1680
+ y(x)
1681
+
1682
+ (D(0))l
1683
+ on R[[zk]].
1684
+ (82)
1685
+ so that by (32) dΓ∗
1686
+ yxz0 = �
1687
+ k≥0 δ(Πk
1688
+ y(x))zk. Identity (82) also implies
1689
+ that the third r. h. s. terms of (80) and (81) are identical. The sec-
1690
+ ond r. h. s. terms of (80) and (81) combine as desired by (75). This
1691
+ establishes (77). In order to use (77) inductively to define – or rather
1692
+ estimate – {Π−
1693
+ x }x, [15] had to come up with an ordering on multi-
1694
+ indices β with respect to which dΓ∗
1695
+ yx is strictly triangular, leading to a
1696
+ modification of (63).
1697
+ Definition of the Malliavin derivative and SG. We have seen
1698
+ that the Malliavin derivative, which we now shall rigorously define,
1699
+ allows to give a more robust relation between Πx and Π−
1700
+ x . Via the SG
1701
+ inequality, which will be introduced here, the control of the Malliavin
1702
+ derivative of a random variable F yields control of the variance of F.
1703
+ Consider the Hilbert norm on (a subspace of) the space of Schwartz
1704
+ distributions27
1705
+ ∥δξ∥2 =
1706
+ ˆ
1707
+ R2 dx
1708
+
1709
+ (∂4
1710
+ 1 − ∂2
1711
+ 2)
1712
+ 1
1713
+ 4(α− 1
1714
+ 2)δξ
1715
+ �2 =
1716
+ ˆ
1717
+ R2 dq
1718
+ ��|q|(α− 1
1719
+ 2)Fδξ
1720
+ ��2.
1721
+ (83)
1722
+ Note that we encounter again A∗A = (−∂2 − ∂2
1723
+ 1)(∂2 − ∂2
1724
+ 1) with Fourier
1725
+ symbol |q|4 = q4
1726
+ 1 + q2
1727
+ 2, see (4). Hence this is one of the equivalent ways
1728
+ of defining the homogeneous L2(R2)-based Sobolev norm of fractional
1729
+ order α − 1
1730
+ 2, however of parabolic scaling, which we nevertheless still
1731
+ denote by H := ˙Hα− 1
1732
+ 2(R2).
1733
+ We now consider “cylindrical” (nonlinear) functionals F on the space
1734
+ S′(R2) of Schwartz distributions, by which one means that for some
1735
+ N ∈ N, F is of the form
1736
+ F[ξ] = f
1737
+
1738
+ (ξ, ζ1), · · · , (ξ, ζN)
1739
+
1740
+ with
1741
+ f ∈ C∞(RN) and ζ1, · · · , ζN ∈ S(R2),
1742
+ (84)
1743
+ 26which also implies dΓ∗
1744
+ yx1 = 0
1745
+ 27we denote the argument by δξ since we think of it as an infinitesimal
1746
+ perturbation.
1747
+
1748
+ 24
1749
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
1750
+ where we recall that (ξ, ζn) denotes the natural pairing between ξ ∈
1751
+ S′(R2) and a Schwartz function ζn ∈ S(R2).
1752
+ Clearly, those func-
1753
+ tion(al)s F are Fr´echet differentiable with
1754
+ dF[ξ].δξ = lim
1755
+ s↓0
1756
+ 1
1757
+ s(F[ξ + sδξ] − F[ξ])
1758
+ =
1759
+ N
1760
+
1761
+ n=1
1762
+ ∂nf
1763
+
1764
+ (ξ, ζ1), · · · , (ξ, ζN)
1765
+
1766
+ (δξ, ζn) = (δξ, ∂F
1767
+ ∂ξ [ξ]),
1768
+ (85)
1769
+ where ∂F
1770
+ ∂ξ [ξ] ∈ S(R2) is defined through
1771
+ ∂F
1772
+ ∂ξ [ξ] =
1773
+ N
1774
+
1775
+ n=1
1776
+ ∂nf
1777
+
1778
+ (ξ, ζ1), · · · , (ξ, ζN)
1779
+
1780
+ ζn.
1781
+ We will monitor the dual norm
1782
+ ∥∂F
1783
+ ∂ξ [ξ]∥∗ := sup
1784
+ δξ
1785
+ (δξ, ∂F
1786
+ ∂ξ [ξ])
1787
+ ∥δξ∥
1788
+ = ∥∂F
1789
+ ∂ξ [ξ]∥ ˙H
1790
+ 1
1791
+ 2 −α(R2).
1792
+ (86)
1793
+ Definition 1. An ensemble E of Schwartz distributions28 is said to
1794
+ satisfy a SG inequality provided for all cylindrical F with E|F| < ∞
1795
+ E(F − EF)2 ≤ E∥∂F
1796
+ ∂ξ ∥2
1797
+ ∗.
1798
+ (87)
1799
+ Note that the l. h. s. of (87) is the variance of F.
1800
+ Inequality (87)
1801
+ amounts to an L2-based Poincar´e inequality with mean value zero on
1802
+ the (infinite-dimensional) space of all ξ’s. By a (parabolic) rescaling
1803
+ of x, we may w. l. o. g. assume that the constant in (87) is unity.
1804
+ Implicitly, we also include closability of the linear operator
1805
+ cylindrical function F �→ ∂F
1806
+ ∂ξ ∈ {cylindrical functions} ⊗ S(R2).
1807
+ (88)
1808
+ This means that the closure of the graph of (88) w. r. t. the topology
1809
+ of L2 and L2(H∗) is still a graph. This allows to extend the Fr´echet
1810
+ derivative (88) to the Malliavin derivative
1811
+ L2 ⊃ D( ∂
1812
+ ∂ξ ) ∋ F �→ ∂F
1813
+ ∂ξ ∈ L2(H∗).
1814
+ By the chain rule, we may post-process (87) to its Lp-version
1815
+ E
1816
+ 1
1817
+ p|F − EF|p ≲p E
1818
+ 1
1819
+ p∥∂F
1820
+ ∂ξ ∥p
1821
+ ∗,
1822
+ (89)
1823
+ which is the form we use it in. A concise proof how to obtain (89) from
1824
+ (87) can be found in [11, Step 2 in the proof of Lemma 3.1].
1825
+ 28It does not have to be a Gaussian ensemble.
1826
+
1827
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
1828
+ 25
1829
+ The obvious examples are Gaussian ensembles of Schwartz distributions
1830
+ with
1831
+ ∥ · ∥ ≤ Cameron-Martin norm,
1832
+ (90)
1833
+ where the norm ∥ · ∥ means the Hilbert norm defined in (83), e. g.
1834
+ white noise
1835
+ −D
1836
+ 2
1837
+ = α − 2
1838
+ =⇒ α = 1
1839
+ 2,
1840
+ free field
1841
+ 1 − D
1842
+ 2
1843
+ = α − 2
1844
+ =⇒ α = 3
1845
+ 2.
1846
+ In other words, the SG inequality (87) holds with Gaussian ensembles
1847
+ satisfying (90), see [2, Theorem 5.5.11].
1848
+ For the reader’s convenience, we sketch the simplest application of SG
1849
+ from [15, Section 4.3], namely (61) for β = 0. To this aim we apply
1850
+ (89) to F := (ξ, ψt(y−·)) = Πx0(y), which is of the form of (84), so that
1851
+ according to (85) its Malliavin derivative is given by ∂F
1852
+ ∂ξ = ψt(y − ·).
1853
+ In view of (86), and then appealing to (8) in conjunction with the
1854
+ translation invariance and scaling of the Sobolev norm we have
1855
+ ∥∂F
1856
+ ∂ξ ∥∗ = ∥ψt(y − ·)∥ ˙H
1857
+ 1
1858
+ 2 −α(R2) = (
1859
+ 4√
1860
+ t)− D
1861
+ 2 − 1
1862
+ 2+α∥ψt=1∥ ˙H
1863
+ 1
1864
+ 2 −α(R2).
1865
+ Noting that the exponent is α − 2 and that ψt=1 is a (deterministic)
1866
+ Schwartz function we obtain from (89)
1867
+ E
1868
+ 1
1869
+ p|Πx0(y)|p ≲ (
1870
+ 4√
1871
+ t)α−2.
1872
+ In view of |0| = α, this amounts to the desired (61) for β = 0.
1873
+ We also remark that SG naturally complements the BPHZ-choice of
1874
+ renormalization, see Section 6:
1875
+ • The choice of cβ takes care of the mean EΠ−
1876
+ xβt(y), while
1877
+ • SG takes care of the variance of Π−
1878
+ xβt(y).
1879
+ Hence the main task in [15] is the estimate of E
1880
+ 1
1881
+ p∥ ∂F
1882
+ ∂ξ ∥p
1883
+ ∗, where F :=
1884
+ Π−
1885
+ xβt(y), which we tackle by duality through estimating the directional
1886
+ derivative
1887
+ δF := (δξ, ∂F
1888
+ ∂ξ )
1889
+ given control of E
1890
+ 1
1891
+ q ∥δξ∥q.
1892
+ The inductive estimate is based on (77).
1893
+ Philosophically speaking, our approach is analytic rather than combi-
1894
+ natorial:
1895
+ analytic
1896
+ combinatorial
1897
+ index set:
1898
+ derivatives w.r.t. a and p
1899
+ Picard iteration
1900
+ ⇝ multi-indices on k ≥ 0, n ̸= 0
1901
+ ⇝ trees with decorations
1902
+ Ass. on ξ:
1903
+ spectral gap inequality
1904
+ cumulant bounds
1905
+ Malliavin derivatives w.r.t. ξ
1906
+ trees with paired nodes
1907
+ ⇝ estimates on E∥ ∂
1908
+ ∂ξΠ−
1909
+ xβ t(y)∥2
1910
+
1911
+ ⇝ Feynman diagrams
1912
+
1913
+ 26
1914
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
1915
+ For us, all combinatorics are contained in Leibniz’ rule. We also point
1916
+ out that our approach may be called “top-down” rather than bottom-
1917
+ up in the sense that we postulate the conditions (space-time trans-
1918
+ lation, spatial reflection, shift-covariance, etc) on the counter term h
1919
+ from the beginning.
1920
+ A closing remark for experts in QFT: The absence of c in (77) means
1921
+ that our approach does not suffer from the well-known difficulty of
1922
+ “overlapping sub-divergences” in Quantum Field Theory, which is also
1923
+ an issue in [4]. Our inductive approach has similarities with the one of
1924
+ Epstein-Glaser, see [18, Section 3.1].
1925
+ 8. The structure group and the re-expansion map
1926
+ In this section we construct the endomorphism Γ∗
1927
+ yx of the algebra
1928
+ R[[zk, zn]] that satisfies (70) for given Πx and Πy. In [15], the construc-
1929
+ tions (and estimates) of Γ∗
1930
+ yx and Πx are actually intertwined, however
1931
+ the proof of Lemma 5 has the same elements as [15, Section 5.3]. In
1932
+ line with regularity structures it is convenient to adopt a more ab-
1933
+ stract point of view: We start by introducing what can be assimilated
1934
+ to Hairer’s structure group G, which here is a subgroup of the au-
1935
+ tomorphism group of the linear space R[zk, zn], where R[zk, zn] now
1936
+ plays the role of the29 (algebraic) pre-dual of R[[zk, zn]]; Γ∗
1937
+ yx will be the
1938
+ transpose of a Γyx ∈ G. The elements Γ ∈ G are parameterized by
1939
+ {π(n)}n ⊂ R[[zk, zn]], see Lemma 3; the group property will be estab-
1940
+ lished in Lemma 4. In Lemma 5 we inductively choose {π(n)
1941
+ yx }n such
1942
+ that the associated Γyx satisfies (70). For a discussion of the Hopf-
1943
+ and Lie-algebraic structure underlying G we refer to [14]. As opposed
1944
+ to [14] and [13], we will capitalize on α < 1, which simplifies several
1945
+ arguments.
1946
+ Lemma 3. Given30 {π(n)}n ⊂ R[[zk, zn]] satisfying
1947
+ π(n)
1948
+ β
1949
+ = 0
1950
+ unless
1951
+ |n| < |β|,
1952
+ (91)
1953
+ there exists a unique linear endomorphism Γ of R[zk, zn] such that Γ∗
1954
+ is an algebra endomorphism31 of R[[zk, zn]] that satisfies
1955
+ Γ∗zk =
1956
+
1957
+ l≥0
1958
+ 1
1959
+ l! (π(0))l(D(0))lzk
1960
+ (32)
1961
+ =
1962
+
1963
+ l≥0
1964
+ �k+l
1965
+ k
1966
+
1967
+ (π(0))lzk+l,
1968
+ (92)
1969
+ Γ∗zn = zn + π(n).
1970
+ (93)
1971
+ 29canonical w. r. t. the monomial basis
1972
+ 30which here as opposed to earlier includes the additional (dummy) index n = 0
1973
+ we first encountered in (30)
1974
+ 31i. e. Γ∗ππ′ = (Γ∗π)(Γ∗π′) and Γ∗1 = 1 hold
1975
+
1976
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
1977
+ 27
1978
+ In addition32,
1979
+ (94)
1980
+ (Γ∗ − id)γ
1981
+ β = 0
1982
+ unless
1983
+ |γ| < |β| and γ ≺ β.
1984
+ We remark that the algebra endomorphism property, the mapping
1985
+ property (92), and the first triangularity in (94) mimic desired proper-
1986
+ ties of Γ∗
1987
+ yx, namely (71), the second item of (70), and (73), respectively.
1988
+ Proof of Lemma 3. We recall that the matrix representation {Γβ
1989
+ γ}β,γ of
1990
+ a linear endomorphism Γ of R[zk, zn] w. r. t. the monomial basis {zβ}β
1991
+ is given by
1992
+ (95)
1993
+ Γzβ =
1994
+
1995
+ γ
1996
+ Γβ
1997
+ γzγ.
1998
+ The algebraic dual Γ∗, as a linear endomorphism of R[[zk, zn]], is given
1999
+ by33
2000
+ (Γ∗π)β =
2001
+
2002
+ γ
2003
+ (Γ∗)γ
2004
+ βπγ
2005
+ where
2006
+ (Γ∗)γ
2007
+ β := (Γ∗zγ)β = Γβ
2008
+ γ.
2009
+ Such a Γ∗ is an algebra endomorphism if and only if
2010
+ (Γ∗)γ
2011
+ β =
2012
+
2013
+ β1+···+βk=β
2014
+ (Γ∗)γ1
2015
+ β1 · · · (Γ∗)γk
2016
+ βk
2017
+ for
2018
+ γ = γ1 + · · · + γk.
2019
+ (96)
2020
+ This includes Γ∗1 = 1 in form of
2021
+ (Γ∗)0
2022
+ β = δ0
2023
+ β
2024
+ (97)
2025
+ Since any multi-index γ ̸= 0 can be written as the sum of γj’s of length
2026
+ one, we learn that an endomorphism Γ of R[zk, zn] with multiplica-
2027
+ tive Γ∗ is determined by how Γ∗ acts on the coordinates {zk}k≥0 and
2028
+ {zn}n̸=0. This establishes the uniqueness statement.
2029
+ For the existence, we need to establish that the numbers {(Γ∗)γ
2030
+ β}β,γ
2031
+ defined through (92) & (93) in form of
2032
+ (Γ∗)ek
2033
+ β − δek
2034
+ β =
2035
+
2036
+ l≥1
2037
+ �k+l
2038
+ k
2039
+
2040
+
2041
+ ek+l+β1+···+βl=β
2042
+ π(0)
2043
+ β1 · · ·π(0)
2044
+ βl ,
2045
+ (98)
2046
+ (Γ∗)en
2047
+ β − δen
2048
+ β = π(n)
2049
+ β
2050
+ (99)
2051
+ and extended by (96) & (97) to all γ satisfy (for fixed β)
2052
+ #{γ | (Γ∗)γ
2053
+ β ̸= 0} < ∞.
2054
+ (100)
2055
+ Indeed, this finiteness condition allows to define Γ via (95) with Γβ
2056
+ γ
2057
+ := (Γ∗)γ
2058
+ β. Since thanks to (103) below in conjunction with 0 < λ, α < 1
2059
+ the ordering ≺ is coercive, by which we mean
2060
+ #{γ | γ ≺ β} < ∞,
2061
+ (101)
2062
+ 32we recall that ≺ is defined in (63)
2063
+ 33note that the sum is effectively finite, since there are only finitely many γ such
2064
+ that Γβ
2065
+ γ ̸= 0 since the monomial basis is an algebraic basis
2066
+
2067
+ 28
2068
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
2069
+ (100) follows once we establish the second strict triangularity in (94).
2070
+ Hence, it remains to establish (94) in form of
2071
+ (Γ∗)γ
2072
+ β − δγ
2073
+ β = 0
2074
+ unless
2075
+ |γ|≺ < |β|≺
2076
+ and
2077
+ |γ| < |β|
2078
+ (102)
2079
+ for the numbers {(Γ∗)γ
2080
+ β}β,γ defined through (98) & (99) and then ex-
2081
+ tended by (96). For this purpose, we note that by definition (54) in
2082
+ form of
2083
+ |β| − α =
2084
+
2085
+ k≥0
2086
+ kβ(k) +
2087
+
2088
+ n̸=0
2089
+ (|n| − α)β(n)
2090
+ (103)
2091
+ and since α ≤ 1 ≤ |n|,
2092
+ | · | − α ≥ 0
2093
+ is additive
2094
+ (63)
2095
+ =⇒
2096
+ same for | · |≺ − α.
2097
+ (104)
2098
+ We first restrict to γ’s of length one in (102), and distinguish the cases
2099
+ γ = en and γ = ek.
2100
+ Since by (54) and (63) we have |en|≺ = |en|
2101
+ = |n| and |β| ≤ |β|≺, the former case follows directly via (99) from
2102
+ assumption (91). We now turn to the latter case of γ = ek and to (98).
2103
+ There is a contribution to the r. h. s. sum only when there exists an
2104
+ l ≥ 1 and a decomposition β = ek+l + β1 + · · · + βl; this implies
2105
+ |β| ≥ |ek+l|
2106
+ (54)
2107
+ = |ek| + αl ≥ |ek| + α
2108
+ (63)
2109
+ =⇒
2110
+ |β|≺ ≥ |ek|≺ + (α − λ),
2111
+ which yields the desired (102) because of α > λ, 0.
2112
+ Finally, we need to upgrade (102) from γ’s of length one to those of
2113
+ arbitrary length, which we do by induction in the length. The base
2114
+ case of zero length, i. e. of γ = 0, is dealt with in (97). We carry out
2115
+ the induction step with help of (96), writing a multi-index γ = γ′ + γ′′
2116
+ with γ′, γ′′ of smaller length:
2117
+ (Γ∗)γ
2118
+ β =
2119
+
2120
+ β′+β′′=β
2121
+ (Γ∗)γ′
2122
+ β′(Γ∗)γ′′
2123
+ β′′.
2124
+ (105)
2125
+ We learn from the induction-hypothesis version of (102) that the sum-
2126
+ mand vanishes unless
2127
+ |γ′| + |γ′′| < |β′| + |β′′| and |γ′|≺ + |γ′′|≺ < |β′|≺ + |β′′|≺
2128
+ or
2129
+ γ′ = β′ and γ′′ = β′′;
2130
+ in the latter case the summand is equal to 1. By (104), the first al-
2131
+ ternative implies |γ| < |β| and |γ|≺ < |β|≺. The second alternative
2132
+ implies γ = β and then holds for exactly one summand to the desired
2133
+ effect of (Γ∗)γ
2134
+ β = 1.
2135
+
2136
+ The two triangular properties (94) from Lemma 3 allow us to establish
2137
+ the group property. Furthermore, a triangular dependence (106) of Γ∗
2138
+ on π(n) will play a crucial role when inductively constructing π(n)
2139
+ yx in
2140
+ Lemma 5.
2141
+
2142
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
2143
+ 29
2144
+ Lemma 4. The set G of all Γ as in Lemma 3 defines a subgroup of the
2145
+ automorphism group of R[zk, zn]. Moreover,
2146
+ for [γ] ≥ 0,
2147
+ (Γ∗)γ
2148
+ β is independent of π(n)
2149
+ β′
2150
+ unless
2151
+ β′ ≺ β.
2152
+ (106)
2153
+ Remark 1. The group G is larger than the one constructed in [14],
2154
+ since 1) we do not require that π(n)
2155
+ β
2156
+ = 0 unless β satisfies (48), and
2157
+ 2) we do not specify the space-time shift structure of the (β = em)-
2158
+ components of π(n)
2159
+ β
2160
+ as in [14, Proposition 5.1]. Both conditions however
2161
+ are satisfied for our construction of π(n)
2162
+ yxβ, see (113) and (115).
2163
+ Proof of Lemma 4. We first argue that for Γ, Γ′ ∈ G we have Γ′Γ ∈ G.
2164
+ More precisely, if Γ and Γ′ are associated to {π(n)}n and {π′(n)}n by
2165
+ Lemma 3, respectively, we consider
2166
+ �π(n) := π(n) + Γ∗π′(n).
2167
+ (107)
2168
+ We note that by triangularity (94) of Γ∗ w. r. t. |·|, the population prop-
2169
+ erty (91) propagates from π(n), π′(n) to �π(n). Let �Γ ∈ G be associated
2170
+ to {�π(n)}n; we claim that Γ′Γ = �Γ.
2171
+ To this purpose, we note that (Γ′Γ)∗ = Γ∗Γ′∗ is an algebra morphism,
2172
+ like �Γ∗ is. Hence by the uniqueness statement of Lemma 3, it is suffi-
2173
+ cient to check that Γ∗Γ′∗ and �Γ∗ agree on the two sets of coordinates
2174
+ {zk}k and {zn}n. On the latter this is easy:
2175
+ �Γ∗zn
2176
+ (93)
2177
+ = zn + �π(n) (107)
2178
+ = zn + π(n) + Γ∗π′(n) (93)
2179
+ = Γ∗(zn + π′(n))
2180
+ (93)
2181
+ = Γ∗Γ′∗zn.
2182
+ We now turn to the zk’s, showing that the algebra endomorphisms Γ∗Γ′∗
2183
+ and �Γ∗ agree on the sub-algebra R[zk] ⊂ R[[zk, zn]]; by multiplicativity
2184
+ of Γ∗ we have according to (92) for Γ′
2185
+ Γ∗Γ′∗ =
2186
+
2187
+ l′≥0
2188
+ 1
2189
+ l′!(Γ∗π′(0))l′Γ∗(D(0))l′
2190
+ on R[zk].
2191
+ Since D(0) preserves R[zk], we may apply (92) for Γ and obtain by the
2192
+ binomial formula:
2193
+ Γ∗Γ′∗ =
2194
+
2195
+ l′≥0
2196
+ 1
2197
+ l′!(Γ∗π′(0))l′ �
2198
+ l≥0
2199
+ 1
2200
+ l!(π(0))l(D(0))l′+l
2201
+ (107)
2202
+ =
2203
+
2204
+ ˜l≥0
2205
+ 1
2206
+ ˜l!
2207
+ (�π(0))
2208
+ ˜l(D(0))
2209
+ ˜l
2210
+ on R[zk],
2211
+ which according to (92) agrees with �Γ∗.
2212
+
2213
+ 30
2214
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
2215
+ We come to the inverse of a Γ ∈ G associated to {π(n)}n.
2216
+ By the
2217
+ strict triangularity (94) w. r. t. the coercive ≺, cf. (101), there exists
2218
+ ˜π(n) ∈ R[[zk, zn]] such that
2219
+ Γ∗˜π(n) = −π(n).
2220
+ (108)
2221
+ We now argue by induction in β w. r. t. ≺ that ˜π(n) satisfies (91). For
2222
+ this, we spell (108) out as
2223
+ ˜π(n)
2224
+ β
2225
+ +
2226
+
2227
+ γ
2228
+ (Γ∗ − id)γ
2229
+ β˜π(n)
2230
+ γ
2231
+ = −π(n)
2232
+ β .
2233
+ If |β| ≤ |n|, the r. h. s. vanishes by (91), and by (94) the sum over
2234
+ γ restricts to |γ| ≤ |β| ≤ |n|, and to γ ≺ β, so that the summand
2235
+ vanishes by induction hypothesis. Thus also ˜π(n)
2236
+ β
2237
+ vanishes.
2238
+ This allows us to argue that ˜Γ ∈ G associated to {˜π(n)}n is the inverse
2239
+ of Γ. By the strict upper triangularity of Γ w. r. t. to the coercive ≺,
2240
+ we already know that Γ is invertible, so that it suffices to show ˜ΓΓ = id,
2241
+ which in turn follows from its transpose Γ∗˜Γ∗ = id. By the composition
2242
+ rule (107) established above, Γ∗�Γ∗ is associated to {π(n) +Γ∗�π(n)}n. By
2243
+ (108) we have that π(n) + Γ∗�π(n) = 0, and learn from Lemma 3 that id
2244
+ is associated with 0.
2245
+ We finally turn to the proof of (106). We note that β1 + · · · + βl = β
2246
+ implies the componentwise βj ≤ β, which by (104) implies |βj|≺ ≤ |β|≺.
2247
+ Since every γ with [γ] ≥ 0 can be written as the sum of γ’s of the form
2248
+ γ = ek + en1 + · · · + enj
2249
+ with
2250
+ j ≤ k,
2251
+ (109)
2252
+ we learn from (96) that we may assume that γ is of this form. Once
2253
+ more by (96) we have for these γ’s
2254
+ (Γ∗)γ
2255
+ β =
2256
+
2257
+ β0+···+βj=β
2258
+ (Γ∗)ek
2259
+ β0(Γ∗)
2260
+ en1
2261
+ β1 · · · (Γ∗)
2262
+ enj
2263
+ βj .
2264
+ From (98) & (99) we learn that this (Γ∗)γ
2265
+ β is a linear combination of
2266
+ π(0)
2267
+ β′
2268
+ 1 · · · π(0)
2269
+ β′
2270
+ l (zn1 + π(n1))β1 · · · (znj + π(nj))βj,
2271
+ (110)
2272
+ where the multi-indices satisfy
2273
+ β = ek+l + β′
2274
+ 1 + · · · + β′
2275
+ l + β1 + · · · + βj.
2276
+ (111)
2277
+ We need to show that the product (110) contains only factors π(n)
2278
+ β′ with
2279
+ β′ ≺ β; w. l. o. g. we may assume l + j ≥ 1. To this purpose we apply
2280
+ | · |≺ to (111); by (104) and |ek+l|≺ ≥ |ek+l| = α(1 + k + l) this implies
2281
+ |β|≺ ≥ α(1 + k − j) + |β′
2282
+ 1|≺ + · · · + |β′
2283
+ l|≺ + |β1|≺ + · · · + |βj|≺, which by
2284
+ j ≤ k implies the desired |β′
2285
+ 1|≺, . . . , |β′
2286
+ l|≺, |β1|≺, . . . , |βj|≺ < |β|≺.
2287
+
2288
+ Finally, we show that the group G is large enough to contain the re-
2289
+ expansion maps.
2290
+
2291
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
2292
+ 31
2293
+ Lemma 5. There exists {π(n)
2294
+ yx }n satisfying (91) such that the Γyx ∈ G
2295
+ associated by Lemma 3 satisfies (70).
2296
+ As a consequence of working with a larger group than in [14], see
2297
+ Remark 1, we don’t have uniqueness of {π(n)
2298
+ yx }n and thus of Γyx. We
2299
+ refer the reader to [19] for a uniqueness result when working with the
2300
+ smaller group. An inspection of our construction reveals transitivity in
2301
+ line with [9, Definition 3.3]
2302
+ Γ∗
2303
+ xyΓ∗
2304
+ yz = Γ∗
2305
+ xz
2306
+ and
2307
+ Γ∗
2308
+ xx = id,
2309
+ see [15, Section 5.3] for the argument; it would also be a consequence
2310
+ of uniqueness.
2311
+ Proof of Lemma 5. We start by specifying π(n)
2312
+ yxβ in the special cases of
2313
+ n = 0 and of β = em for some m ̸= 0:
2314
+ π(0)
2315
+ yx := Πy(x),
2316
+ (112)
2317
+ π(n)
2318
+ yxem :=
2319
+ � �m
2320
+ n
2321
+
2322
+ (x − y)m−n
2323
+ provided n < m,
2324
+ 0
2325
+ otherwise
2326
+
2327
+ for n ̸= 0,
2328
+ (113)
2329
+ where n < m means component-wise (non-strict) ordering and n ̸= m.
2330
+ We note that (112) is necessary in order to bring the second item of
2331
+ (70) into agreement with the form (92). We also remark that (113)
2332
+ yields by (93)
2333
+ (Γ∗
2334
+ yx)en
2335
+ em =
2336
+ � �m
2337
+ n
2338
+
2339
+ (x − y)m−n
2340
+ provided n ≤ m,
2341
+ 0
2342
+ otherwise
2343
+
2344
+ .
2345
+ By the second part of (94), which implies (Γ∗
2346
+ yx)γ
2347
+ 0 = 0 unless γ = 0, by
2348
+ (98) in form of (Γ∗
2349
+ yx)ek
2350
+ em = 0, and via (96) this strengthens to
2351
+ (Γ∗
2352
+ yx)γ
2353
+ em =
2354
+ � �m
2355
+ n
2356
+
2357
+ (x − y)m−n
2358
+ if γ = en with n ≤ m,
2359
+ 0
2360
+ otherwise
2361
+
2362
+ .
2363
+ (114)
2364
+ The latter is imposed upon us by taking the (β = em)-component
2365
+ of the first item in (70) and plugging in (41).
2366
+ The second part of
2367
+ (114) implies that Γyx maps the linear span of {zm}m̸=0 into itself;
2368
+ since this linear span can be identified with the space R[x1, x2]/R of
2369
+ space-time polynomials (modulo constants), this can be assimilated to
2370
+ Hairer’s postulate [9, Assumption 3.20]. We note that (112) and (113)
2371
+ satisfy (91) because of | · | ≥ α > 0, cf. (104), and |em| = |m| > |n|,
2372
+ respectively. In line with (48) and [14], we also set
2373
+ π(n)
2374
+ yxβ = 0
2375
+ unless
2376
+ [β] ≥ 0 or β = em for some m ̸= 0.
2377
+ (115)
2378
+ It thus remains to construct π(n)
2379
+ yxβ for n ̸= 0 and [β] ≥ 0, which we will
2380
+ do by induction in β w. r. t. ≺. According to (106), we may consider
2381
+ (Γ∗)γ
2382
+ β as already constructed for [γ] ≥ 0. According to (64) and by the
2383
+
2384
+ 32
2385
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
2386
+ induction hypothesis (70), an inspection of the argument that leads
2387
+ from there to (72) shows that we also have
2388
+ Π−
2389
+ yβ = (Γ∗
2390
+ yxΠ−
2391
+ x )β.
2392
+ (116)
2393
+ The induction step consists in choosing {π(n)
2394
+ yxβ}0<|n|<|β| such that
2395
+ Πyβ = (Γ∗
2396
+ yxΠx)β + Πyβ(x)
2397
+ (112)
2398
+ = (Γ∗
2399
+ yxΠx)β + π(0)
2400
+ yxβ.
2401
+ (117)
2402
+ Denoting by P the projection on multi-indices γ with [γ] ≥ 0, so that
2403
+ by (41) and (48) we have (id − P)Πx = �
2404
+ n̸=0(· − x)nzn and thus by
2405
+ (91) and (93)
2406
+ (Γ∗
2407
+ yx(1 − P)Πx)β =
2408
+
2409
+ 0<|n|<|β|
2410
+ (· − x)nπ(n)
2411
+ yxβ,
2412
+ (118)
2413
+ allows us to make {π(n)
2414
+ yxβ}0<|n|<|β| in (117) explicit:
2415
+ (Πy − Γ∗
2416
+ yxPΠx)β =
2417
+
2418
+ n:|n|<|β|
2419
+ π(n)
2420
+ yxβ(· − x)n.
2421
+ (119)
2422
+ Hence our task reads
2423
+ (Πy − Γ∗
2424
+ yxPΠx)β = polynomial of degree < |β|.
2425
+ (120)
2426
+ According to the PDE (58), to (116), and to (118) we have
2427
+ (∂2 − ∂2
2428
+ 1)(Πy − Γ∗
2429
+ yxPΠx)β = polynomial of degree < |β| − 2.
2430
+ (121)
2431
+ In order to pass from (121) to (120), we will now appeal to the unique-
2432
+ ness/Liouville statement in Lemma 1 with η = |β|, which is ̸∈ Z ac-
2433
+ cording to (62) and ≥ α according to (104), and p = 1 for simplicity.
2434
+ More precisely, we apply Lemma 1 to
2435
+ u = (Πy − Γ∗
2436
+ yxPΠx)β − its Taylor polynomial in x of order < |β|,
2437
+ which makes sense since (121) implies that (Πy − Γ∗
2438
+ yxPΠx)β is smooth,
2439
+ and to f ≡ 0. Hence for the assumption (10) we need to check that
2440
+ (122)
2441
+ lim sup
2442
+ z:|z−x|↑∞
2443
+ 1
2444
+ |z − x||β|E|(Πy − Γ∗
2445
+ yxPΠx)β(z)| < ∞,
2446
+ which forces us to now become semi-quantitative.
2447
+ By the estimate (60) on Π, for (122) it remains to show34
2448
+ E
2449
+ 1
2450
+ p |(Γ∗
2451
+ yx)γ
2452
+ β|p ≲β,γ,p |y − x||β|−|γ|
2453
+ provided
2454
+ [γ] ≥ 0.
2455
+ (123)
2456
+ In line with the language of [15], we split the argument for (123) into
2457
+ an “algebraic argument”, where we derive (123) from
2458
+ (124)
2459
+ E
2460
+ 1
2461
+ p|π(n)
2462
+ yxβ′|p ≲β′,p |x − y||β′|−|n|
2463
+ for β′ ≺ β,
2464
+ 34which coincides with Hairer’s postulate [9, (3.2) in Definition 3.3]
2465
+
2466
+ LECTURE NOTES ON TREE-FREE REGULARITY STRUCTURES
2467
+ 33
2468
+ and a “three-point argument”, where we derive (124) from the estimate
2469
+ (60) on Π.
2470
+ Here comes the argument for (123), which is modelled after the one
2471
+ for (106) in Lemma 4. By H¨older’s inequality in probability and the
2472
+ additivity of |·|−α, cf. (104), we may restrict to γ’s of the form (109).
2473
+ We are thus lead to estimate the product (110), which now takes the
2474
+ form of
2475
+ π(0)
2476
+ yxβ′
2477
+ 1 · · · π(0)
2478
+ yxβ′
2479
+ l(zn1 + π(n1)
2480
+ yx )β1 · · · (znj + π(nj)
2481
+ yx )βj.
2482
+ (125)
2483
+ Once again by H¨older’s inequality, we infer from (124) that the E
2484
+ 1
2485
+ p|·|p-
2486
+ norm of (125) is
2487
+ ≲ |y − x||β′
2488
+ 1| · · · |y − x||β′
2489
+ l||y − x||β1|−|n1| · · · |y − x||βj|−|nj|.
2490
+ By the additivity of |·|−α, the total exponent of |y−x| can be identified
2491
+ with the desired expression:
2492
+ |β′
2493
+ 1| + · · · + |β′
2494
+ l| + (|β1| − |n1|) + · · · + (|βj| − |nj|)
2495
+ (111)
2496
+ = |β| − |ek+l| + (l + j)α − (|n1| + · · · + |nj|)
2497
+ (109)
2498
+ = |β| − |γ|.
2499
+ Finally, we give the “three-point argument” for the estimate (124), for
2500
+ notational simplicity in case of the current multi-index β, so that we
2501
+ now may use (119) and (123). By (60) and (123), the left hand side of
2502
+ (119) can be estimated as follows
2503
+ E
2504
+ 1
2505
+ p |(Πy − Γ∗
2506
+ yxPΠx)β(z)|p ≲β,p (|z − x| + |y − x|)|β|.
2507
+ By the equivalence of norms on the finite-dimensional space of space-
2508
+ time polynomials of degree < |β|, which by a duality argument can
2509
+ be upgraded to the following estimate of annealed norms for random
2510
+ polynomials
2511
+ max
2512
+ n: |n|<|β| |y − x||n| E
2513
+ 1
2514
+ p|π(n)
2515
+ yxβ|p ≲
2516
+
2517
+ |z−x|≤|y−x|
2518
+ dz E
2519
+ 1
2520
+ p��
2521
+
2522
+ n: |n|<|β|
2523
+ (z − x)nπ(n)
2524
+ yxβ
2525
+ ��p,
2526
+ we obtain (124).
2527
+
2528
+ References
2529
+ [1] H. Bahouri, J.-Y. Chemin, and R. Danchin. Fourier analysis and nonlinear
2530
+ partial differential equations, volume 343. Springer, 2011.
2531
+ [2] V. I. Bogachev. Gaussian measures, volume 62 of Mathematical Surveys and
2532
+ Monographs. American Mathematical Society, Providence, RI, 1998.
2533
+ [3] Y. Bruned, M. Hairer, and L. Zambotti. Algebraic renormalisation of regularity
2534
+ structures, Invent. Math. 215(3):1039–1156, 2019.
2535
+ [4] A. Chandra and M. Hairer. An analytic BPHZ theorem for Regularity Struc-
2536
+ tures, arXiv:1612.08138, 2016.
2537
+ [5] L. Coutin and Z. Qian. Stochastic analysis, rough path analysis and fractional
2538
+ Brownian motions, Probability theory and related fields, 122(1):108–140, 2002.
2539
+
2540
+ 34
2541
+ FELIX OTTO, KIHOON SEONG, AND MARKUS TEMPELMAYR
2542
+ [6] P. Duch. Renormalization of singular elliptic stochastic PDEs using flow equa-
2543
+ tion, arXiv:2201.05031 [math.PR].
2544
+ [7] M. Gubinelli. Ramification of rough paths, J. Differential Equations 248 (2010),
2545
+ no. 4, 693–721.
2546
+ [8] M. Gubinelli and N. Perkowski. An introduction to singular SPDEs, Stochastic
2547
+ partial differential equations and related fields, 69–99, Springer Proc. Math.
2548
+ Stat., 229, Springer, Cham, 2018.
2549
+ [9] M. Hairer. Regularity structures and the dynamical φ4
2550
+ 3 model, arXiv:1508.05261
2551
+ [math.PR].
2552
+ [10] M. Hairer and ´E. Pardoux. A Wong-Zakai theorem for stochastic PDEs. J.
2553
+ Math. Soc. Japan, 67(4):1551–1604, 2015.
2554
+ [11] M. Josien and F. Otto. The annealed Calder´on-Zygmund estimate as conve-
2555
+ nient tool in quantitative stochastic homogenization, J. Funct. Anal. 283 (2022),
2556
+ no. 7, 74 pp.
2557
+ [12] A. Kupiainen. Renormalization Group and Stochastic PDEs, Ann. Henri
2558
+ Poincar´e 17, 497–535 (2016).
2559
+ [13] P. Linares and F. Otto. A tree-free approach to regularity structures: the regular
2560
+ case for quasi-linear equations, arXiv:2207.10627 [math.AP].
2561
+ [14] P. Linares, F. Otto and M. Tempelmayr. The structure group for quasi-linear
2562
+ equations via universal enveloping algebras, arXiv:2103.04187 [math-ph].
2563
+ [15] P. Linares, F. Otto, M. Tempelmayr, and P. Tsatsoulis. A diagram-free ap-
2564
+ proach to the stochastic estimates in regularity structures, arXiv:2112.10739
2565
+ [math.PR].
2566
+ [16] T. Lyons, M. Caruana and T. L´evy. Differential equations driven by rough
2567
+ paths, Lecture Notes in Mathematics, 1908. Springer, Berlin, 2007.
2568
+ [17] F. Otto, J. Sauer, S. Smith, and H. Weber. A priori estimates for quasi-linear
2569
+ SPDEs in the full sub-critical regime, arXiv:2103.11039 [math.AP].
2570
+ [18] G. Scharf. Finite Quantum Electrodynamics: The Causal Approach, Second
2571
+ version, Texts and Monographs in Physics, Springer Berlin, 1995.
2572
+ [19] M. Tempelmayr. Characterizing models in regularity structures: a quasi-linear
2573
+ case, to appear.
2574
+ Felix Otto, Kihoon Seong, and Markus Tempelmayr
2575
+ Max–Planck Institute for Mathematics in the Sciences
2576
+ 04103 Leipzig, Germany
2577
+ felix.otto@mis.mpg.de, kihoon.seong@mis.mpg.de,
2578
+ markus.tempelmayr@mis.mpg.de
2579
+
AdAyT4oBgHgl3EQf3_pa/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
B9E5T4oBgHgl3EQfTg8R/content/tmp_files/2301.05536v1.pdf.txt ADDED
@@ -0,0 +1,2756 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1
2
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
3
+
4
+ An Electromagnetic-Information-Theory Based
5
+ Model for Efficient Characterization of MIMO
6
+ Systems in Complex Space
7
+
8
+ Ruifeng Li, Da Li, Member, IEEE, Jinyan Ma, Zhaoyang Feng, Ling Zhang, Member, IEEE, Shurun Tan, Member,
9
+ IEEE, Wei E. I. Sha, Senior Member, IEEE, Hongsheng Chen, Fellow, IEEE, and Er-Ping Li, Fellow, IEEE
10
+ Abstract—It is the pursuit of a multiple-input-multiple-output
11
+ (MIMO) system to approach and even break the limit of channel
12
+ capacity. However, it is always a big challenge to efficiently
13
+ characterize the MIMO systems in complex space and get better
14
+ propagation performance than the conventional MIMO systems
15
+ considering only free space, which is important for guiding the
16
+ power and phase allocation of antenna units. In this manuscript,
17
+ an Electromagnetic-Information-Theory (EMIT) based model is
18
+ developed for efficient characterization of MIMO systems in
19
+ complex space. The group-T-matrix-based multiple scattering fast
20
+ algorithm, the
21
+ mode-decomposition-based
22
+ characterization
23
+ method, and their joint theoretical framework in complex space
24
+ are discussed. Firstly, key informatics parameters in free
25
+ electromagnetic space based on a dyadic Green’s function are
26
+ derived. Next, a novel group-T-matrix-based multiple scattering
27
+ fast algorithm is developed to describe a representative
28
+ inhomogeneous electromagnetic space. All the analytical results
29
+ are validated by simulations. In addition, the complete form of the
30
+ EMIT-based model is proposed to derive the informatics
31
+ parameters frequently used in electromagnetic propagation,
32
+ through integrating the mode analysis method with the dyadic
33
+ Green’s function matrix. Finally, as a proof-or-concept,
34
+ microwave anechoic chamber measurements of a cylindrical array
35
+ is performed, demonstrating the effectiveness of the EMIT-based
36
+ model. Meanwhile, a case of image transmission with limited
37
+ power is presented to illustrate how to use this EMIT-based model
38
+ to guide the power and phase allocation of antenna units for real
39
+ MIMO applications.
40
+
41
+ Index Terms—multiple-input-multiple-output (MIMO) system,
42
+ complex space, group T matrix, mode analysis, electromagnetic
43
+ information theory (EMIT).
44
+ I. INTRODUCTION
45
+ YPICALLY, for antenna design, it is promising to
46
+ maximize the channel capacity via a multiple-input-
47
+ multiple-output (MIMO) system to approach the limit of
48
+ channel capacity during propagation. Thus, Under the demand
49
+ for high accuracy and low latency nowadays, the basic research
50
+ on efficient characterization of MIMO systems is very
51
+ important [1]. On this basis, we can carry out further work such
52
+ as the optimization solutions for the power and phase allocation
53
+ of antenna units.
54
+ Previous works for MIMO characterization can be roughly
55
+ clarified into two categories: electromagnetic (EM) methods
56
+
57
+ This project is supported in part by Natural Science Foundation of China
58
+ (NSFC), Grant No. 62071424, 62201499 and 62027805. (Corresponding
59
+ Author: Da Li, li-da@zju.edu.cn)
60
+ and information theory. The former mainly focuses on the
61
+ radio-frequency (RF) front-end design by solving Maxwell’s
62
+ equations under different boundary conditions, consisting of the
63
+ descriptions of the complex electromagnetic space [2], [3],
64
+ while the latter mainly analyzes the channel properties under
65
+ different probability models by using Shannon information
66
+ theory [4], [5]. The above two frameworks are faced with a
67
+ major challenge in practical application: how to efficiently
68
+ model MIMO systems in complex space to achieve better
69
+ propagation performance than MIMO analysis that only
70
+ consider free space.
71
+ For EM methods, the core step of intelligent designs
72
+ nowadays is reconstructing the MIMO systems’ radiation
73
+ patterns [6], [7], [8]. To consider the effect of the EM
74
+ propagation space, full-wave numerical algorithms have been
75
+ used to incorporate the RF front-end design and environment
76
+ perception into the EM framework, consuming a lot of time [9].
77
+ To greatly reduce the calculation time of modeling EM space,
78
+ some studies have proposed to use approximate methods like
79
+ ray tracing (RT) [10], [11], [12]. However, this is often not
80
+ acceptable due to lack of high accuracy. Besides, the T-matrix
81
+ can be easily used to characterize efficient MIMO in complex
82
+ EM space, via combining multiple scattering equations [13],
83
+ [14]. However, practical wireless communication often focuses
84
+ on some informatic parameters (such as channel capacity),
85
+ while the EM-only framework is incapable of efficiently
86
+ extracting the informatic parameters in the complex EM space.
87
+ For information theory, the common statistic model, such as
88
+ the Rayleigh fading model, is a mathematical tool based on the
89
+ assumption of rich scattering [15]. When it evolves to cluster
90
+ models like geometry-based stochastic models (GBSMs), the
91
+ EM space is equivalent to the clusters with different shapes or
92
+ distributions for convenient characterization [16], [17].
93
+ However, the accuracy of those models will be reduced due to
94
+ the EM properties of the MIMO system. For example, the work
95
+ in [18] complements numerical methods to make up for the
96
+ problem of using only Fresnel approximation in airborne
97
+ antenna design. Moreover, the main idea of the emerging
98
+ intelligent reflective surface (IRS) is to lay out the controllable
99
+
100
+
101
+ The authors are with ZJU-UIUC Institute, Zhejiang Provincial Key
102
+ Laboratory of Advanced Microelectronic Intelligent Systems and Appli-
103
+ cations, and the College of Information Science and Electronic Engineering,
104
+ Zhejiang University, Hangzhou 310027, China.
105
+ T
106
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
107
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
108
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
109
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
110
+
111
+ 2
112
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
113
+
114
+
115
+ Fig. 1. System model of MIMO analysis in a complex space for indoor
116
+ communication.
117
+
118
+ surfaces in free or complex EM space [19], [20], [21], [22],
119
+ [23]. Due to the lack of efficient MIMO characterization, this
120
+ technology is still in the trial stage. This suggests that many
121
+ basic assumptions of the information-only framework need to
122
+ be reconsidered.
123
+ Nowadays, the electromagnetic information theory (EMIT)
124
+ for the MIMO characterization attracts more attentions, which
125
+ is expected to solve the challenges mentioned above [24], [25],
126
+ [26], [27]. Researchers point out that with the wide layout of
127
+ the antenna array (e.g., Internet of vehicles), it is expected to
128
+ eliminate the step of channel estimation with the help of rich
129
+ environmental information [28]. Some works have been done
130
+ from the perspective of EM fields to study the degree of
131
+ freedom of MIMO systems [29], [30]. There are also
132
+ mathematical methods to model the source region and field
133
+ region as two sets of orthogonal bases in Hilbert space, and then
134
+ construct some characteristic parameters of the MIMO system
135
+ [31], [32]. To integrate the RF front-end design in the EMIT
136
+ framework, the surface currents of antenna elements are
137
+ modeled as the point sources with orthogonal bases. For
138
+ example, a model was established to build a channel matrix
139
+ from the angle of coordinate transformation and orthogonal
140
+ decomposition of EM plane wave expansion, applied in
141
+ holographic MIMO system [33], [34]. Additionally, the work
142
+ in [35] contains the idea of deriving the channel limit of a
143
+ MIMO system by the EM field method. Nevertheless, the above
144
+ research works on EMIT mainly focus on free space or
145
+ revealing the parameter mapping between two theories;
146
+ efficient characterization algorithms and clear EM information
147
+ analysis methods for complex EM space are still unexplored.
148
+ In this paper, we develop an EMIT-based model to conduct
149
+ the efficient characterization for MIMO systems in complex
150
+ EM space. The proposed EMIT-based model uses the group T
151
+ matrix algorithm and dyadic Green’s function-based mode
152
+ analysis method, filling the research gap of efficient
153
+ characterization algorithms and clear EM information analyses.
154
+ The main contributions of this paper are described as follows.
155
+ 1) The key parameters of the MIMO systems are extracted
156
+ through the dyadic Green’s function and matrix mode
157
+ analysis. The information characteristics of the MIMO
158
+ systems are described by the EM method, revealing
159
+ some important conclusions and deducing the key
160
+ informatic parameters and valuable conclusions of
161
+ information theory by means of EM methods.
162
+ 2) A fast algorithm based on the group T matrix is
163
+ developed to model the complex EM space. Since the
164
+ algorithm has semi-analytical characteristics and the
165
+ classical T matrix can be stored, which provides a faster
166
+ calculation compared with the traditional full-wave
167
+ algorithm. In contrast to the RT and pilot-based methods
168
+ for channel estimation, our EM algorithm can be easily
169
+ integrated into EMIT due to its higher accuracy and
170
+ efficiency. In other words, the benefit of our proposed
171
+ method is generated from the fast characteristics of the
172
+ group T matrix and the EM analysis of the channel
173
+ matrix (without the help of statistics).
174
+ 3) The efficient EMIT-based model is proposed to
175
+ characterize the MIMO systems in complex space. As a
176
+ proof-of-concept, a microwave anechoic chamber
177
+ measurement of a cylindrical array is taken as an
178
+ example, demonstrating the effectiveness of the EMIT-
179
+ based model for the MIMO mode analysis. Meanwhile,
180
+ a case of image transmission with limited power is
181
+ presented to illustrate how to guide the MIMO feeding
182
+ based on the model, bringing a new insight into
183
+ extracting information parameters using the basis of
184
+ computational electromagnetism.
185
+ This article is organized as follows. The key informatics
186
+ parameters based on the dyadic Green’s function are derived in
187
+ Section II. Then, the proposed EMIT-based model is analyzed
188
+ in Section III. Experimental verification and an image
189
+ transmission case were conducted in Section IV. Finally, the
190
+ conclusion is drawn in Section V.
191
+ II. SYSTEM MODEL AND KEY PARAMETERS
192
+ As shown in Fig. 1, consider a typical MIMO system
193
+ including the transmitting and receiving array for indoor
194
+ communication, whose overall communication performance
195
+ will be affected by the propagation distance and the properties
196
+ of complex space. In this section, the EM propagation space is
197
+ designated as a free space for extracting key parameters of a
198
+ MIMO system. More complex EM space is characterized in the
199
+ next section.
200
+ To combine the coupling operator
201
+ TR
202
+ G
203
+ and the channel
204
+ matrix  , a series of isotropic point sources are placed in the
205
+ transmission volume and the receiving volume, with the
206
+ position vectors
207
+ Tr
208
+ and
209
+ Rr
210
+ respectively. The EM wave
211
+ received is defined as
212
+ outR
213
+ ψ
214
+ , thus the Helmholtz wave equation
215
+ is given by
216
+
217
+ 2
218
+ 0
219
+ 0
220
+ outR
221
+ outR
222
+ incT
223
+ k
224
+ i
225
+ 
226
+
227
+ =
228
+ ψ
229
+ ψ
230
+ J
231
+ ,
232
+ (1)
233
+ where
234
+ incT
235
+ J
236
+ is the transmitted source and k is the wave vector.
237
+ To solve this equation, the dyadic Green’s function G operator
238
+ based on the impulse function idea is introduced:
239
+
240
+
241
+ 2
242
+ exp[
243
+ ]
244
+ (
245
+ ,
246
+ )
247
+ 4
248
+ R
249
+ T
250
+ R
251
+ T
252
+ R
253
+ T
254
+ ik
255
+ k
256
+
257
+
258
+
259
+
260
+ 
261
+ =
262
+ +
263
+
264
+
265
+
266
+
267
+
268
+ r
269
+ r
270
+ G r r
271
+ I
272
+ r
273
+ r
274
+ ,
275
+ (2)
276
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
277
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
278
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
279
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
280
+
281
+ Coupling operatorGr
282
+ Channel matrix H
283
+ Receiving array
284
+ Electromagnetic
285
+ characteristic
286
+ Information
287
+ characteristic
288
+ Complex Space
289
+ Feeding&
290
+ Beamforming3
291
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
292
+
293
+ where I is the unit tensor. Since the dyadic Green’s function
294
+ tensor G contains the scalar Green’s functions:
295
+
296
+
297
+ Gxx
298
+ Gxy
299
+ Gxz
300
+ Gyx
301
+ Gyy
302
+ Gyz
303
+ Gzx
304
+ Gzy
305
+ Gzz
306
+
307
+
308
+
309
+
310
+ = 
311
+
312
+
313
+
314
+
315
+
316
+ G
317
+ .
318
+ (3)
319
+
320
+ When it comes to the two independent single-polarization
321
+ situation at far field, the coupling operator is able to be
322
+ simplified into the scalar Green’s function without loss of
323
+ accuracy. Therefore, the element
324
+ ij
325
+ h in the channel matrix 
326
+ will changed to the following form in the case of a single
327
+ polarized source:
328
+
329
+
330
+ 0( )
331
+ exp[
332
+ ]
333
+ 4
334
+ Ri
335
+ Tj
336
+ ij
337
+ ij
338
+ Ri
339
+ Tj
340
+ ik
341
+ h
342
+ g
343
+
344
+
345
+ =
346
+ =
347
+
348
+ r
349
+ r
350
+ r
351
+ r
352
+ ,
353
+ (4)
354
+ where
355
+ 0
356
+ g is the scalar Green’s function, that is, the special form
357
+ of G in the case of single polarization. It is worth mentioning
358
+ that the channel matrix  at this time is not normalized, so it
359
+ contains the path loss.
360
+ Assume that the number of transmitting source points is
361
+ T
362
+ N ,
363
+ and the number of receiving field points is
364
+ R
365
+ N . Therefore, the
366
+ transmitting source can be expressed as a
367
+ *1
368
+ T
369
+ N
370
+ matrix, the
371
+ receiving electric field as a
372
+ *1
373
+ R
374
+ N
375
+ matrix, and the coupling
376
+ operator G of the EM space as a
377
+ *
378
+ R
379
+ T
380
+ N
381
+ N
382
+ matrix. By
383
+ introducing the Dirac notation, the MIMO propagation relation
384
+ of free space is expressed as:
385
+
386
+
387
+ outR
388
+ incT
389
+ =
390
+ ψ
391
+ G J
392
+ .
393
+ (5)
394
+
395
+ To normalize the channel matrix, we defined the normalized
396
+ coupling operator
397
+ TR
398
+ G
399
+ as
400
+ TR
401
+
402
+ =
403
+ G
404
+ G , where  is a
405
+ normalization factor, making
406
+ 2
407
+ TR
408
+ T
409
+ R
410
+ F
411
+ N N
412
+
413
+
414
+
415
+ =
416
+
417
+
418
+
419
+
420
+ G
421
+ .
422
+  
423
+
424
+ denotes
425
+ the expectation and
426
+ F means the Frobenius norm. The
427
+ physical meaning of this normalization is that every sub-
428
+ channel should have a unity average channel gain.
429
+ According to the Hermitian nature of
430
+
431
+ TR
432
+ TR
433
+ G
434
+ G
435
+ , singular
436
+ value decomposition (SVD) of the coupling operator could
437
+ conduct mode analysis of MIMO EM propagation, where †
438
+ denotes the conjugate transpose:
439
+
440
+
441
+
442
+ TR
443
+ R
444
+ T
445
+ =
446
+ G
447
+ U SV ,
448
+ (6)
449
+
450
+ where
451
+ T
452
+ V (
453
+ R
454
+ U ) is a
455
+ *
456
+ T
457
+ T
458
+ N
459
+ N (
460
+ *
461
+ R
462
+ R
463
+ N
464
+ N
465
+ ) matrix, and each
466
+ column represents the EM eigenvector of the transmitting
467
+ sources (receiving fields ). Because of the unitary nature of the
468
+ SVD eigenmatrix, it is known that each column is strictly
469
+ orthogonal, which is called the EM space mode. The weight of
470
+ each pattern is determined by the corresponding element in the
471
+ diagonal matrix S. The combinations of those orthogonal modes
472
+ form two Hilbert spaces, and therefore the coupling operator
473
+ TR
474
+ G
475
+ builds a mapping between the transmitting Hilbert space
476
+ and the receiving Hilbert space, which is an important property
477
+ in the subsequent discussion.
478
+ As we all know, the upper limit of information transmission
479
+ per bandwidth in MIMO systems is also limited by Shannon’s
480
+ formula [36]:
481
+
482
+
483
+
484
+ 2
485
+ 2
486
+ 2
487
+ 1
488
+ log
489
+ det
490
+ log
491
+ 1
492
+ TR
493
+ TR
494
+ t
495
+ i
496
+ i
497
+ C
498
+ n N
499
+ n
500
+
501
+  
502
+
503
+
504
+
505
+
506
+
507
+
508
+
509
+
510
+ = 
511
+ +
512
+
513
+
514
+
515
+
516
+
517
+
518
+
519
+
520
+
521
+
522
+
523
+
524
+
525
+
526
+
527
+
528
+
529
+
530
+ =
531
+ +
532
+
533
+
534
+
535
+
536
+
537
+ I
538
+ G
539
+ G
540
+ ,
541
+ (7)
542
+ where I is the identity matrix and
543
+ i
544
+  are the singular values
545
+ of(
546
+ )
547
+ 1/
548
+ TR
549
+ T
550
+ N
551
+ G
552
+ . Apparently,
553
+ 2
554
+ i
555
+  is the decisive parameter of
556
+ key information-carrying capacity in the MIMO system at a
557
+ given SNR
558
+ / n
559
+
560
+ . Besides, we can drop the expectation  
561
+
562
+ in
563
+ (7) and no longer need to make a special distinction for large-
564
+ scale and small-scale path loss and fading, because the
565
+ amplitude and phase changes of the electric field have been
566
+ included in the operator
567
+ TR
568
+ G
569
+ .
570
+ It is seen from (6) and (7) that the singular value of EM
571
+ propagation space determines the number and weight of
572
+ independent modes, which establishes a corresponding
573
+ relationship with the number of independently available
574
+ channels and path loss of wireless communication. We give the
575
+ key informatics parameters of a MIMO system by referring to
576
+ the effective rank idea of existing work [21]:
577
+
578
+
579
+ min(
580
+ ,
581
+ )
582
+ 1
583
+ exp(
584
+ ln(
585
+ ))
586
+ R
587
+ T
588
+ N
589
+ N
590
+ eff
591
+ i
592
+ i
593
+ i
594
+ C
595
+
596
+
597
+ =
598
+
599
+
600
+ =
601
+
602
+
603
+ ,
604
+ (8)
605
+
606
+ where
607
+ eff
608
+ C
609
+ represents
610
+ the
611
+ EM
612
+ effective
613
+ capacity,
614
+ / (
615
+ )
616
+ i
617
+ i
618
+ i
619
+
620
+
621
+
622
+  =
623
+
624
+ represents the normalized singular values of
625
+ TR
626
+ G
627
+ . Hence, (8) establishes the mapping relationship between
628
+ the dyadic Green’s function matrix and typical informatics
629
+ parameters, which is an important tool for the MIMO mode
630
+ analysis.
631
+ To
632
+ understand
633
+ how
634
+ this
635
+ approach
636
+ works,
637
+ both
638
+ mathematically and physically, we set up an
639
+ *
640
+ N
641
+ N MIMO
642
+ system with the same EM space properties, as shown in Fig. 1.
643
+ In fact, in practical engineering applications, the mutual
644
+ coupling is concerned not because it affects the EM equivalent
645
+ capacity, but because it affects the radiation efficiency and
646
+ signal-to-noise ratio of the antennas. The transmitting and
647
+ receiving
648
+ antennas
649
+ are
650
+ modeled
651
+ as
652
+ isotropic
653
+ point
654
+ sources/receivers (delta function basis), which is a widely used
655
+ assumption in EM information theory. From the EM
656
+ perspective, the antennas can also be modeled as continuous
657
+ surface (equivalent) currents by
658
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
659
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
660
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
661
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
662
+
663
+ 4
664
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
665
+
666
+
667
+ Fig. 2. EM effective capability with the change of number of sources in four
668
+ communication distances.
669
+
670
+
671
+ Fig. 3. EM effective capability with the change of aperture sizes in four
672
+ communication distances.
673
+
674
+
675
+ Fig. 4. The attenuation of EM effective capability with the change of
676
+ communication distance.
677
+
678
+
679
+ Fig. 5. Schematic diagram of MIMO mode analysis of complex space. The
680
+ material, position, quantity and shape of the scatterers can be set arbitrarily in
681
+ our algorithm.
682
+
683
+ the rooftop or Rao–Wilton–Glisson (RWG) basis, as frequently
684
+ utilized in the methods of moments. Different basis
685
+ representations of the currents, related to different antenna
686
+ designs, will not influence the estimations of the effective
687
+ degree of freedom limit. Since we want to focus the analysis in
688
+ this work on solving dyadic Green's function and extracting
689
+ informatics parameters in a complex space, we choose the
690
+ model carefully to avoid mutual coupling. To construct the
691
+ basic framework of EMIT, three key parameters (the number of
692
+ sources, communication distance, and the size of antenna
693
+ aperture) are considered to illustrate the relationship between
694
+ RF front-end devices’ design and the effective capability of the
695
+ MIMO system.
696
+ In Fig. 2, the relationship between the EM effective capacity
697
+ and the number of sources is presented at a given aperture
698
+ ( 6 *6
699
+
700
+  ), showing clearly that with the increase of N , the EM
701
+ effective capacity under different communication distances will
702
+ increase with the same slope, but it converges to the channel
703
+ capacity. In this case, considering that the change of the total
704
+ power of the transmitting array will lead to different channel
705
+ capacities, we fixed the total transmitting power at
706
+ 0P ,
707
+ satisfying
708
+ 0
709
+ 1
710
+ T
711
+ N
712
+ i
713
+ i
714
+ P
715
+ P
716
+ =
717
+ =
718
+
719
+ .
720
+ In addition, to illustrate the physical nature of the
721
+ convergence, Fig. 3 shows the EM effective capacity
722
+ corresponding to different aperture sizes with enough point
723
+ sources (30*30). Obviously, the size of the aperture plays a
724
+ determinant role in the information capacity of MIMO systems.
725
+ It suggests that the trend of antenna miniaturization is the
726
+ weakening of maximum carrying information, which cannot be
727
+ solved by multi-antenna technology. Besides, in Fig. 4, we plot
728
+ the curve of EM effective capacity changing with the
729
+ communication distance, revealing the characteristics of
730
+ wireless
731
+ communication-energy
732
+ attenuated
733
+ with
734
+ the
735
+ propagation distance from the perspective of dyadic Green’s
736
+ function. Besides, in Fig. 3 and Fig. 4, the variables we focus
737
+ on are the aperture and distance respectively, so the number of
738
+ point sources is a constant, and the total power
739
+ 0P always
740
+ remains a constant.
741
+ It is worth mentioning that, due to the basis function
742
+ decomposition method (such as Rao-Wilton-Glisson (RWG)
743
+ basis in MoM) commonly used in computational
744
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
745
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
746
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
747
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
748
+
749
+ fwith differentnumber of sources
750
+ 15.1
751
+ Distance=1*lambda
752
+ EM effective capability
753
+ Distance=6*lambda
754
+ Distance=11*lambda
755
+ Distance=16*lambda
756
+ 10
757
+ 15
758
+ 20
759
+ 25
760
+ 30
761
+ Number of sourcesCefr with different aperture
762
+ Distance=1*lambda
763
+ 350
764
+ Distance=6*lambda
765
+ Distance=11*lambda
766
+ EM effective capability
767
+ Distance=16*lambda
768
+ 250
769
+ 200
770
+ 150
771
+ 00
772
+ 50
773
+ Size of aperture ()Cofr with different communication distance
774
+ 200
775
+ 180
776
+ EM effective capability
777
+ 40
778
+ Communication distance (2)Transmitting array
779
+ Modeprofiles
780
+ Scatterers
781
+ PML
782
+ Receiving array
783
+ Complexspace5
784
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
785
+
786
+
787
+ Fig. 6. Illustration of the group-T-matrix-based algorithm. The distribution of
788
+ the field around the scatterer is decomposed, and the steady-state coefficient
789
+ matching is carried out based on the cylindrical wave expansion without
790
+ meshwork and time-domain iteration.
791
+
792
+ electromagnetics, the specific RF front-end structure can be
793
+ decomposed into the sum of point sources through grid
794
+ partitioning, and the multi-channel coupling effect will be
795
+ considered in the coefficient term of the operator
796
+ TR
797
+ G
798
+ .
799
+ Therefore, when using the above method to perform theoretical
800
+ modeling of EMIT, the coupling can be characterized by adding
801
+ a coefficient term to the operator
802
+ TR
803
+ G
804
+ , and the specific physical
805
+ dimensions of the RF front-end can be numerically quantified
806
+ by base function equivalence.
807
+ Essentially, changing the RF front-end or the channel will
808
+ affect the value of
809
+ eff
810
+ C
811
+ in (8) by affecting the distribution of
812
+ i
813
+  on the ith channel. In other words,
814
+ eff
815
+ C
816
+ and the distribution
817
+ of
818
+ i
819
+  are the inherent property of the communication system.
820
+ However, the core assumption of this part is based on free
821
+ EM space, and the specific form of coupling operator
822
+ TR
823
+ G
824
+ will
825
+ change when numerous scatterers are introduced. The next
826
+ section will demonstrate the fast algorithms for characterizing
827
+ the complex EM complex space.
828
+ III. PROPOSED EMIT-BASED MODEL FOR EM COMPLEX SPACE
829
+ In some typical wireless communication scenarios, objects in
830
+ complex scattering environments are usually represented by
831
+ some types of scatterers for convenient EM calculations, among
832
+ which one of the commonly-used classical models is the
833
+ cylindrical array, as described in Fig.5. For example, a vehicle-
834
+ to-vehicle channel is equivalent to a scattering cluster in the
835
+ internet of vehicles channel modeling [16]. Due to the poor
836
+ accuracy and long response time of traditional channel
837
+ measurement schemes, this section proposes an EMIT-based
838
+ model for efficient MIMO characterization in this typical
839
+ scattering complex space based on the group T matrix.
840
+ A. Algorithm Description
841
+ N cylindrical scatterers in MIMO EM propagation space are
842
+ considered, which are centered at
843
+ (
844
+ 1,2,...,
845
+ )
846
+ pr
847
+ p
848
+ N
849
+ =
850
+ and are
851
+ with radius
852
+ (
853
+ 1,2,...,
854
+ )
855
+ p
856
+ a
857
+ p
858
+ N
859
+ =
860
+ . These parameters can be easily
861
+ substituted to simulate different distributions and different
862
+ shapes of scatterers. For the description of the RF front-end, we
863
+ use a
864
+ *1
865
+ s
866
+ N
867
+ dipole antenna array, coordinate
868
+ (
869
+ 1,2,...,
870
+ )
871
+ sr s
872
+ N
873
+ =
874
+ ,
875
+ as a convenient MIMO model. The overall algorithm
876
+ framework is shown in Fig. 6. To be clear, we focus on the
877
+ scenarios where the transceivers and receivers are in the same
878
+ horizontal plane (such as vehicle-to-vehicle communication
879
+ and indoor point-to-point communication). In this case, we can
880
+ regard the scatterer as a cluster of cylindrical scatterers, so
881
+ conducting cylindrical wave expansion is reasonable and
882
+ convenient. This benefits the convenience of calculation and the
883
+ simplicity of the model.
884
+ To take the coupling between scatterers into account, we take
885
+ the th
886
+ q
887
+ scatterer as the analysis object and decompose the total
888
+ external field
889
+ ex
890
+ q
891
+
892
+ around it into the sum of the incident field
893
+ inc
894
+ q
895
+
896
+ and the scattering field
897
+ s
898
+ p
899
+  of the rest scatterers:
900
+
901
+
902
+ 1
903
+ .
904
+ N
905
+ ex
906
+ inc
907
+ s
908
+ q
909
+ q
910
+ p
911
+ p
912
+ p q
913
+
914
+
915
+
916
+ =
917
+
918
+ =
919
+ +
920
+
921
+ (9)
922
+
923
+ For solving the scattered fields, the electric field is expanded
924
+ as a vector cylindrical wave harmonic function:
925
+
926
+
927
+ ( )
928
+ ( (
929
+ ))
930
+ ex
931
+ q
932
+ q
933
+ n
934
+ n
935
+ n I
936
+ Rg
937
+ k
938
+
939
+
940
+ =
941
+
942
+
943
+ q
944
+ r
945
+ r
946
+ ,
947
+ (10)
948
+
949
+ where k is the wave vector,
950
+ ( )
951
+ q
952
+ nI
953
+ is the cylindrical wave
954
+ coefficient, which is the unknown core quantity for solving the
955
+ field distribution.
956
+ In addition, the specific mathematical form of cylindrical
957
+ wave expansion in (10) is given as follows:
958
+
959
+
960
+ (1)
961
+ ( (
962
+ ))
963
+ (
964
+ )exp(
965
+ )
966
+ ( (
967
+ ))
968
+ (
969
+ )exp(
970
+ )
971
+ n
972
+ n
973
+ n
974
+ n
975
+ k
976
+ H
977
+ k
978
+ in
979
+ Rg
980
+ k
981
+ J
982
+ k
983
+ in
984
+
985
+
986
+
987
+
988
+
989
+ =
990
+
991
+
992
+ =
993
+
994
+ p
995
+ p
996
+ p
997
+ p
998
+ rr
999
+ p
1000
+ p
1001
+ rr
1002
+ r
1003
+ r
1004
+ r
1005
+ r
1006
+ r
1007
+ r
1008
+ r
1009
+ r
1010
+ ,
1011
+ (11)
1012
+
1013
+ where r represents the coordinate vector of the field point, 
1014
+ p
1015
+ rr
1016
+ represents the angle between the vectors r and
1017
+ pr ,
1018
+ nJ is the
1019
+ Bessel function of order n ,
1020
+ (1)
1021
+ n
1022
+ H
1023
+ is the Hankel function of
1024
+ order n , and Rg means regularization. Later, we will use the
1025
+ symbol i to represent the imaginary unit.
1026
+ For the mode matching, we perform the same cylindrical
1027
+ wave expansion for the incident field
1028
+ inc
1029
+ q
1030
+
1031
+ determined by the
1032
+ MIMO RF front-end (here is the
1033
+ *1
1034
+ s
1035
+ N
1036
+ dipole array),
1037
+ obtaining:
1038
+
1039
+
1040
+ (1)
1041
+ 0
1042
+ 1
1043
+ (
1044
+ ).
1045
+ 4
1046
+ s
1047
+ N
1048
+ inc
1049
+ q
1050
+ s
1051
+ i H
1052
+ k
1053
+
1054
+ =
1055
+ =
1056
+
1057
+
1058
+ s
1059
+ r
1060
+ r
1061
+
1062
+ (12)
1063
+
1064
+ To obtain the same expansion form as (10), the vector
1065
+ addition theorem is used to further expand (12) to obtain:
1066
+
1067
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
1068
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
1069
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
1070
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
1071
+
1072
+ 2
1073
+ y
1074
+ q
1075
+ x
1076
+ b
1077
+ pl
1078
+ TpI
1079
+ TPV
1080
+ yip2
1081
+ AV
1082
+ V
1083
+ pl
1084
+ pN
1085
+ yot
1086
+ rer
1087
+ pl
1088
+ pN
1089
+ x
1090
+ x
1091
+ pl
1092
+ x
1093
+ Z
1094
+ PId
1095
+ p2
1096
+ pN
1097
+ Scatterer 1
1098
+ Scatterer 2
1099
+ Scatterer N
1100
+ Zo
1101
+ +X06
1102
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1103
+
1104
+
1105
+ Fig. 7. Normalized electric field distribution on the validation plane. (a) (c) (e):
1106
+ FDTD solver for scatterers distribution of 1*1, 1*5 and 4*1 respectively. (b)
1107
+ (d) (f): proposed EMIT-based model for scatterers distribution of 1*1, 1*5 and
1108
+ 4*1 respectively.
1109
+
1110
+
1111
+ (
1112
+ )
1113
+ (
1114
+ )
1115
+ (1)
1116
+ 1
1117
+ exp
1118
+ 4
1119
+ ( (
1120
+ )).
1121
+ s
1122
+ q
1123
+ N
1124
+ inc
1125
+ q
1126
+ n
1127
+ n
1128
+ s
1129
+ n
1130
+ i
1131
+ H
1132
+ k
1133
+ in
1134
+ Rg
1135
+ k
1136
+
1137
+
1138
+
1139
+ =
1140
+ =
1141
+
1142
+
1143
+
1144
+
1145
+  
1146
+ s
1147
+ s
1148
+ r r
1149
+ s
1150
+ r
1151
+ r
1152
+ r
1153
+ r
1154
+
1155
+ (13)
1156
+
1157
+ In (13), as the RF front-end information is part of prior
1158
+ knowledge, the expansion coefficient of
1159
+ inc
1160
+ q
1161
+
1162
+ is determined,
1163
+ which is convenient for solving
1164
+ ( )
1165
+ q
1166
+ nI
1167
+ in (10). Next, we write the
1168
+ scattering field
1169
+ s
1170
+ p
1171
+  of the
1172
+ th
1173
+ p
1174
+ scatterer as follows:
1175
+
1176
+
1177
+ (
1178
+ )
1179
+ ( )
1180
+ 0
1181
+ n
1182
+ s
1183
+ TR
1184
+ TR
1185
+ p
1186
+ dS
1187
+ i
1188
+ dS
1189
+
1190
+ 
1191
+
1192
+
1193
+
1194
+
1195
+
1196
+ =
1197
+
1198
+ −
1199
+
1200
+
1201
+
1202
+
1203
+
1204
+
1205
+ p
1206
+ p
1207
+ G
1208
+ J r
1209
+ G
1210
+ M r
1211
+ , (14)
1212
+
1213
+ where
1214
+ TR
1215
+ G
1216
+ is dyadic Green’s function illustrated in (2),  is
1217
+ the angular frequency, ( )
1218
+ p
1219
+ J r
1220
+ and
1221
+ ( )
1222
+ p
1223
+ M r
1224
+ are the current
1225
+ density and magnetic current density at the
1226
+ th
1227
+ p
1228
+ scatterer,
1229
+ respectively. The EM variation in the complex space is
1230
+ described by the action of the coupling operator
1231
+ TR
1232
+ G
1233
+ on ( )
1234
+ p
1235
+ J r
1236
+
1237
+ and
1238
+ ( )
1239
+ p
1240
+ M r
1241
+ . When ( )
1242
+ p
1243
+ J r
1244
+ and
1245
+ ( )
1246
+ p
1247
+ M r
1248
+ do not exist, (9) will
1249
+ then degenerate into the free space case shown in section II.
1250
+ In order to solve the
1251
+ s
1252
+ p
1253
+
1254
+ described in (14), we use the
1255
+ consistent mathematical form of
1256
+ ex
1257
+ q
1258
+
1259
+ on different scatterers to
1260
+ expand
1261
+ s
1262
+ p
1263
+  into cylindrical waveform by using (10), and the
1264
+ transformation relationship is shown in the red curve in Fig. 6.
1265
+ Thus, (14) can be rewritten as follows based on the group-T-
1266
+ matrix:
1267
+
1268
+
1269
+ ( )
1270
+ ( )
1271
+ ( (
1272
+ ))
1273
+ s
1274
+ p
1275
+ p
1276
+ p
1277
+ m
1278
+ m
1279
+ m
1280
+ mT
1281
+ I
1282
+ Rg
1283
+ k
1284
+
1285
+
1286
+ =
1287
+
1288
+
1289
+ s
1290
+ r
1291
+ r
1292
+ ,
1293
+ (15)
1294
+
1295
+ where
1296
+ (
1297
+ )
1298
+ p
1299
+ T
1300
+ is the group-T-matrix representing the relationship
1301
+ between the incident field and scattering field of the
1302
+ th
1303
+ p
1304
+ clustered scatterer, and its characteristics are only related to the
1305
+ shape and material of the current scatterer. Assuming the
1306
+ internal wave vector of the scatterer is
1307
+ p
1308
+ k , the general form of
1309
+ group-T-matrix in the cylindrical coordinate system can be
1310
+ obtained by using analytical methods:
1311
+
1312
+
1313
+ ( )
1314
+ (1)
1315
+ (1)
1316
+ (
1317
+ )
1318
+ (
1319
+ )
1320
+ (
1321
+ )
1322
+ (
1323
+ )
1324
+ (
1325
+ )
1326
+ (
1327
+ )
1328
+ (
1329
+ )
1330
+ p
1331
+ m
1332
+ p
1333
+ p
1334
+ m
1335
+ p
1336
+ p
1337
+ m
1338
+ p
1339
+ p
1340
+ m
1341
+ m
1342
+ p
1343
+ m
1344
+ p
1345
+ p
1346
+ m
1347
+ p
1348
+ p
1349
+ m
1350
+ p
1351
+ p
1352
+ k J
1353
+ k a
1354
+ J
1355
+ k a
1356
+ kJ
1357
+ ka
1358
+ T
1359
+ kH
1360
+ ka
1361
+ J
1362
+ k a
1363
+ H
1364
+ ka
1365
+ k J
1366
+ k a
1367
+
1368
+
1369
+
1370
+ =
1371
+
1372
+
1373
+ . (16)
1374
+
1375
+ The T-matrix of any shape objects can be solved by
1376
+ numerical methods such as the method of moments (MoM)
1377
+ according to (14).
1378
+ The basic purpose of this section is to illustrate the
1379
+ algorithm’s efficiency, and thus we consider the model of
1380
+ dipole array with TM polarized waves incident on a perfect
1381
+ electric conductor (PEC). In this case, (16) evolves into:
1382
+
1383
+
1384
+
1385
+ ( )
1386
+ (1)
1387
+ (
1388
+ ) .
1389
+ (
1390
+ )
1391
+ m
1392
+ p
1393
+ p
1394
+ m
1395
+ m
1396
+ p
1397
+ J
1398
+ ka
1399
+ T
1400
+ H
1401
+ ka
1402
+ = −
1403
+
1404
+ (17)
1405
+
1406
+ Substitute (17) into (15) to obtain the field distribution with
1407
+ ( )
1408
+ p
1409
+ m
1410
+ I
1411
+ as the only variable. The matrix equation of the unknown
1412
+ coefficient
1413
+ ( )
1414
+ p
1415
+ m
1416
+ I
1417
+ can be obtained by combining (9), (10), (13),
1418
+ and (15):
1419
+
1420
+
1421
+  =
1422
+ Z I
1423
+ V .
1424
+ (18)
1425
+
1426
+ Here, in order to solve the coefficient
1427
+ ( )
1428
+ p
1429
+ m
1430
+ I
1431
+ , the equations
1432
+ with different scatterers are written in matrix form, and the
1433
+ order of the Bessel function is truncated with the truncation
1434
+ number
1435
+ max
1436
+ N
1437
+ . Therefore, Z is a square matrix of dimension
1438
+ max
1439
+ (2
1440
+ 1)
1441
+ N
1442
+ N
1443
+ +
1444
+ , while V is a (
1445
+ )
1446
+ max
1447
+ 2
1448
+ 1
1449
+ 1
1450
+ N
1451
+ N
1452
+ +
1453
+
1454
+ vector. The
1455
+ specific form is:
1456
+
1457
+  (
1458
+ )
1459
+ (
1460
+ )
1461
+ (
1462
+ )
1463
+ (
1464
+ )
1465
+ (
1466
+ )
1467
+ 1
1468
+ ,
1469
+ 1
1470
+ ( )
1471
+ (1)
1472
+ 1,
1473
+ exp
1474
+ ,
1475
+ p q
1476
+ q
1477
+ N
1478
+ n
1479
+ p
1480
+ N
1481
+ m
1482
+ p
1483
+ m
1484
+ n m
1485
+ p
1486
+ q
1487
+ r r
1488
+ p
1489
+ q
1490
+ T
1491
+ H
1492
+ k r
1493
+ r
1494
+ i n
1495
+ m
1496
+ p
1497
+ q
1498
+
1499
+ − 
1500
+ +
1501
+ − 
1502
+ +
1503
+
1504
+
1505
+ =
1506
+ 
1507
+ = 
1508
+
1509
+
1510
+
1511
+
1512
+ 
1513
+ Z
1514
+ ,(19)
1515
+
1516
+  (
1517
+ )
1518
+ (
1519
+ )
1520
+ (
1521
+ )
1522
+ (1)
1523
+ 1
1524
+ 1
1525
+ exp
1526
+ 4
1527
+ s
1528
+ s q
1529
+ N
1530
+ n
1531
+ s
1532
+ q
1533
+ r r
1534
+ q
1535
+ N
1536
+ n
1537
+ s
1538
+ i
1539
+ H
1540
+ k r
1541
+ r
1542
+ in
1543
+ − 
1544
+ +
1545
+ =
1546
+ = −
1547
+
1548
+
1549
+
1550
+ V
1551
+ .
1552
+ (20)
1553
+
1554
+
1555
+ TABLE I
1556
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
1557
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
1558
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
1559
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
1560
+
1561
+ 0.5
1562
+ m
1563
+ 0.53
1564
+ 0.5
1565
+ -
1566
+ 0
1567
+ 0
1568
+ 0.5
1569
+ 1
1570
+ 1.5
1571
+ 0
1572
+ 0.5
1573
+ 1
1574
+ 1.5
1575
+ X (m)
1576
+ x (m)
1577
+ (a)
1578
+ (b)
1579
+ )
1580
+ 0.5
1581
+ ()
1582
+ 0.5
1583
+ 0
1584
+ 0
1585
+ 0.5
1586
+ 1
1587
+ 1.5
1588
+ 0
1589
+ 0.5
1590
+ 1
1591
+ 1.5
1592
+ x (m)
1593
+ x (m)
1594
+ (c)
1595
+ (d)
1596
+ 3008
1597
+ (u)
1598
+ 0.5
1599
+ u)
1600
+ 0.5
1601
+ 0.5
1602
+ 0
1603
+ 0.5
1604
+ 1
1605
+ 1.5
1606
+ 0
1607
+ 0.5
1608
+ 1
1609
+ 1.5
1610
+ x(m)
1611
+ X (m)
1612
+ (e)
1613
+ (f)7
1614
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1615
+
1616
+ COMPARISON OF CPU TIME AND RMS ERROR BETWEEN FDTD AND
1617
+ PROPOSED EMIT-BASED MODEL FOR THE CHARACTERISTIC OF COMPLEX
1618
+ SPACE
1619
+
1620
+
1621
+ In (19) and (20), Z is determined by the properties of the
1622
+ complex space, and V is determined by the properties of the
1623
+ RF front-end. A joint solution can semi-analytically describe
1624
+ the evolution of the MIMO coupling operator
1625
+ EIT
1626
+ G
1627
+ .
1628
+ Therefore, due to the change of EM space coupling operator,
1629
+ (5) will be rewritten as:
1630
+
1631
+
1632
+ EIT
1633
+ outR
1634
+ incT
1635
+ =
1636
+ ψ
1637
+ G
1638
+ J
1639
+ .
1640
+ (21)
1641
+
1642
+ The subsequent analysis only needs to be carried out in the
1643
+ same
1644
+ way
1645
+ as
1646
+ (6)-(8)
1647
+ to
1648
+ complete
1649
+ efficient
1650
+ MIMO
1651
+ characterization in complex space. It should be noted that, when
1652
+ facing the time-varying channel scenario, it will be very
1653
+ convenient to rewrite the coupling operator
1654
+ EIT
1655
+ G
1656
+ into the form
1657
+ based on the time-domain Green's function.
1658
+
1659
+ B. Numerical Results
1660
+ To verify the accuracy and efficiency of the proposed EMIT-
1661
+ based model, numerical calculations of some specific scenarios
1662
+ are carried out and compared with full-wave simulation results.
1663
+ Fig. 7 presents the field distributions of three simple arrays.
1664
+ The effectiveness of the EMIT-based model is verified by
1665
+ comparing the full-wave FDTD algorithm (a, c, e) with the
1666
+ proposed algorithm (b, d, f). We consider an EM space of 1
1667
+ m*1.5 m, where the MIMO system is modeled as a 3*1 dipole
1668
+ array with an aperture of 0.75 m. The scatterer array element is
1669
+ modeled as a metal cylinder with a height of 0.25 m and a radius
1670
+ of 0.015 m, and the boundary is set as PEC. Due to the dense
1671
+ mesh division of the full-wave algorithm, its application is
1672
+ severely limited. However, the proposed semi-analytic
1673
+ algorithm based on group-T-matrix is suitable for various
1674
+ frequencies because it does not need mesh division. To obtain
1675
+ the comparison results, we first define the operating frequency
1676
+ at 915 MHz in this section.
1677
+ Fig. 7 illustrates that the proposed EMIT-based model has
1678
+ achieved good results and can accurately describe the complex
1679
+ space. In order to further demonstrate its efficiency, the
1680
+ scatterer distribution was adjusted, and the solving time and
1681
+ error of the EMIT-based model and FDTD were calculated by
1682
+ analyzing the field intensity curve at the RF back-end, as shown
1683
+
1684
+ Fig. 8. The total normalized electric field distribution on the validation plane
1685
+ corresponding to the proposed complex space obtained by EMIT-based model.
1686
+
1687
+ in Table I. It is worth noting that the 10*15 distribution cannot
1688
+ fully explain the difference between the two algorithms,
1689
+ because the large number of FDTD meshes converge extremely
1690
+ fast due to the inability of the electric field to propagate
1691
+ effectively, and the solutions are often mediocre at this time.
1692
+ Therefore, we further consider the case of a random array, that
1693
+ is, randomly removing 60 scatterers from the 10*15 scatterer
1694
+ distribution. Besides, we clarify that the running time of our
1695
+ proposed algorithm mainly depends on the number of scatters.
1696
+ Therefore, the proposed EMIT-based model has higher
1697
+ computational efficiency than full-wave algorithms like FDTD,
1698
+ which provides great convenience for the description of
1699
+ complex space. But generally, the complexity of the real-world
1700
+ environment increases with the communication distance. In this
1701
+ case, an efficient way to leverage the EMIT-based method is to
1702
+ build a common clustering model database. Compared with
1703
+ pilot-based methods, it also has good efficiency under the
1704
+ condition of a complete database. For example, a vehicle-to-
1705
+ vehicle channel is equivalent to a scattering cluster in the
1706
+ internet-of-vehicles channel modeling [16].
1707
+ C. Mode Analysis Step of the EMIT-Based Model
1708
+ After efficient characterization of the complex space is
1709
+ verified, the EMIT-based model performs a mode analysis of
1710
+ the above characterization results to obtain theoretical
1711
+ interpretations to guide the design of wireless communications.
1712
+ Consider an actual information transmission scenario where
1713
+ the RF front-end is a single-polarized dipole antenna array
1714
+ operating at 2.5GHz (equivalent using an ideal line source
1715
+ operating at 2.5Ghz), with a scale of 10*1 (designed to make
1716
+ the MIMO feature more obvious), and the complex space is
1717
+ simplified to a 4*5 metallic cylindrical scatterer cluster.
1718
+ According to the quick algorithm in the previous section, the
1719
+ electric field distribution on the validation plane is shown in
1720
+ Fig. 8. By substituting the solved coupling operator
1721
+ EIT
1722
+ G
1723
+ into
1724
+ (6), the EM effective capacity
1725
+ eff
1726
+ C
1727
+ of this model in wireless
1728
+ communication is known as 5.2, which means that the actual
1729
+ effective number of available channels is 5. However, Fig. 3
1730
+ shows that dyadic Green’s function operators
1731
+ TR
1732
+ G
1733
+ (coupling
1734
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
1735
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
1736
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
1737
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
1738
+
1739
+ (m)
1740
+ 0.5
1741
+ 0.5
1742
+ 0.5
1743
+ 1.5
1744
+ x (m)8
1745
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1746
+
1747
+
1748
+ Fig. 9. Each mode’s normalized electric field diagram on the validation plane
1749
+ obtained by EMIT-based model. (a-e) Available information transfer modes;
1750
+ (d-j) Unavailable higher-order information transfer modes.
1751
+
1752
+ operators in free space) will bring channel gains far beyond 5.2.
1753
+ Therefore, the EMIT-based model provides a convenient tool
1754
+ to quantitatively explain the influence of the complex
1755
+ environment on wireless communication quality. The
1756
+ information at the receiving end in Fig. 8 can help us get the
1757
+ operator
1758
+ EIT
1759
+ G
1760
+ . Through the SVD mentioned above, 10 modes
1761
+ at the transmitting end can be decomposed, and the EM
1762
+ responses of these 10 modes in the complex space are shown in
1763
+ Fig. 9. It is clearly found that the first five modes successfully
1764
+ send signals to the receiver effectively in different coupling
1765
+
1766
+ Fig. 10. The distribution of normalized singular values of different number of
1767
+ sources.
1768
+
1769
+
1770
+ Fig. 11. Simple MIMO propagation system in complex space.
1771
+
1772
+ paths. However, the coupling paths of higher-order modes
1773
+ bypass the receiver’s acceptance range and become unavailable
1774
+ modes in wireless communication.
1775
+ To define the concepts of “available” and “unavailable” more
1776
+ clearly, we show the distribution of modes’ singular values for
1777
+ the different number of channels in Fig. 10. A formal definition
1778
+ is given as follows: if all modes are numbered according to the
1779
+ normalized singular value in a descending order like Fig. 10,
1780
+ then the available modes are defined as those whose index is
1781
+ less than the EM effective capacity
1782
+ eff
1783
+ C
1784
+ , and the other modes
1785
+ are defined as the unavailable modes. Since the power resources
1786
+ in an actual wireless communication system are limited, the
1787
+ mode weight corresponding to each channel number is
1788
+ normalized here. Obviously, for a MIMO system, there will be
1789
+ an evident truncation of the modes’ singular value distribution,
1790
+ and modes below the truncation usually become “unavailable”.
1791
+ The number of “available” modes will be strictly determined by
1792
+ (8) after the coupling operator
1793
+ EIT
1794
+ G
1795
+ is obtained by the EMIT-
1796
+ based model.
1797
+ Obviously, the more channels available, the more
1798
+ information that can be transmitted, and the greater the channel
1799
+ capacity of the corresponding EM space. However,
1800
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
1801
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
1802
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
1803
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
1804
+
1805
+ Model
1806
+ Mode2
1807
+
1808
+ 0.5
1809
+ (m
1810
+ 0.5
1811
+ 0.5
1812
+ A
1813
+ 0
1814
+ 0.5
1815
+ 1
1816
+ 1.5
1817
+ 0
1818
+ 0.5
1819
+ 1.5
1820
+ x(m)
1821
+ x (m)
1822
+ (a)
1823
+ (q)Mode3
1824
+ Mode4
1825
+ 0.5
1826
+ 0.5
1827
+ 0.5
1828
+ U
1829
+ 0.5
1830
+ 1
1831
+ 1.5
1832
+ 0
1833
+ 0.5
1834
+ x (m)
1835
+ 1.5
1836
+ x (m)
1837
+ (c)
1838
+ (d)Mode5
1839
+ Mode6
1840
+ m
1841
+ 0.5
1842
+ 0.5
1843
+ 0.5
1844
+ 1.5
1845
+ 0
1846
+ 0.5
1847
+ 1.5
1848
+ x (m)
1849
+ x (m)
1850
+ (e)
1851
+ (f)Mode7
1852
+ Mode8
1853
+ 0.5
1854
+ 0.5
1855
+ 0.5
1856
+ 1
1857
+ 1.5
1858
+ 0
1859
+ 0.5
1860
+ 1.5
1861
+ x (m)
1862
+ x (m)
1863
+ (g)
1864
+ (h)Mode9
1865
+ Mode10
1866
+ 0.5
1867
+ (m
1868
+ 0.5
1869
+ 0
1870
+ 0
1871
+ 0
1872
+ 0.5
1873
+ 1.5
1874
+ 0
1875
+ 0.5
1876
+ 1
1877
+ 1.5
1878
+ x (m)
1879
+ x (m)
1880
+ (0)
1881
+ ()Singular value distribution
1882
+ 0.9
1883
+ 0.8
1884
+ 0.7
1885
+ 0.5
1886
+ 0.6
1887
+ 0.5
1888
+ 0.4
1889
+ 0.3
1890
+ 0.2
1891
+ 0.1
1892
+ 6
1893
+ N
1894
+ mode index
1895
+ 0
1896
+ numberofsourceTransmitting Array
1897
+ Receiving Array
1898
+ Scattering Region
1899
+ VNA9
1900
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1901
+
1902
+
1903
+ Fig. 12. The normalized amplitude and phase of
1904
+ 21
1905
+ S
1906
+ in 7*7 MIMO system. (a)
1907
+ Simulation results in EMIT-based model; (b) Measurement results; (c) The
1908
+ error of the two above.
1909
+
1910
+ communication resources of the RF front-end are often limited,
1911
+ so it is important to allocate resources properly to achieve better
1912
+ information transmission efficiency. In the next section, power
1913
+ distribution is taken as the background problem to discuss the
1914
+ guidance significance of the EMIT-based model for real
1915
+ wireless communication in a complex environment.
1916
+ IV. EXPERIMENTAL ANALYSIS
1917
+ It is worth noting that the above discussion on the application
1918
+ of the EMIT-based model is carried out by simulation. In order
1919
+ to fully explain the effectiveness of the EMIT-based model and
1920
+ the application method under the background of wireless
1921
+ communication, we carried out experimental exploration with
1922
+ the aid of a 3*1 MIMO system.
1923
+ The experiment construction is shown in Fig. 11, where the
1924
+ system is surrounded by the absorption boundary covered with
1925
+ absorbing materials, and cylindrical metal scatterers with a
1926
+ height of 0.25 m and a radius of 0.015 m are uniformly
1927
+ distributed in the EM space with 4*5 arrays. The transmitting
1928
+ sources and the receiving fields were replaced by dipole
1929
+ antennas with a center frequency of 2.5 GHz and a gain of 2dBi.
1930
+ The vector network analyzer (VNA) is used to measure the
1931
+ 21
1932
+ S
1933
+
1934
+ between transmitting and receiving dipoles through the coaxial
1935
+ feed. Since the measurement of channel matrix elements is
1936
+ concerned with the single excitation properties of MIMO, we
1937
+ replace the actual MIMO system by changing the spatial
1938
+ position of the antenna in the transmitting aperture (shown as
1939
+
1940
+ Fig. 13. The field distribution of three orthogonal modes at the receiver aperture.
1941
+ The locations of the three sources are indicated by black dotted lines on the
1942
+ diagram.
1943
+
1944
+ the dotted white lines). Therefore, in the experimental design,
1945
+ we selected the weak coupling scenario with the antenna
1946
+ spacing as half-wavelength, and then measured the antenna
1947
+ excitation separately to avoid the impact of coupling on the
1948
+ verification of our EMIT-based model.
1949
+ For the same scene, we performed effective characterization
1950
+ with the EMIT-based model, and the characterization results are
1951
+ shown in Fig. 12. Fig. 12 (a) and (b) respectively represent the
1952
+ amplitude and phase comparison results between the simulation
1953
+ results
1954
+ of
1955
+ EMIT-based model
1956
+ and
1957
+ the
1958
+ experimental
1959
+ measurement
1960
+ values.
1961
+ The
1962
+ experimental
1963
+ results
1964
+ fully
1965
+ demonstrate the effectiveness of the EMIT-based model in
1966
+ complex space characterization. The purpose of conducting 7*7
1967
+ channel measurement in our experiment is to verify the EMIT-
1968
+ based model more convincingly. However, the following is
1969
+ mainly to illustrate how the EMIT-based model guides the RF
1970
+ front-end signal transmission. Therefore, to simplify the
1971
+ demonstration process, we select three groups of data evenly
1972
+ spaced to form a new 3*3 MIMO system. In fact, the selection
1973
+ of 3*3 channel positions is arbitrary. However, in this paper, to
1974
+ make the mode orthogonality more significant and avoid the
1975
+ influence of mode crosstalk on the transmitting strategy, three
1976
+ positions with relatively small mode crosstalk are selected, and
1977
+ the crosstalk matrix CT is as follows:
1978
+
1979
+
1980
+ 15
1981
+ 15
1982
+ 15
1983
+ 15
1984
+ 1
1985
+ 0.1857
1986
+ 4.03*10
1987
+ 0.1857
1988
+ 1
1989
+ 6.51*10
1990
+ 4.03*10
1991
+ 6.51*10
1992
+ 1
1993
+
1994
+
1995
+
1996
+
1997
+
1998
+
1999
+
2000
+
2001
+ = 
2002
+
2003
+
2004
+
2005
+
2006
+
2007
+ CT
2008
+ .
2009
+ (22)
2010
+
2011
+ Consider the following problems in an actual wireless
2012
+ communication scenario: 3 * 3 MIMO system needs to conduct
2013
+ data transmission in disorder EM space (with the standard
2014
+ deviation of noise  ), the maximum transmitted power of RF
2015
+ front-end is
2016
+ 0P . Under this constraint, since it is a very
2017
+ important subject to consider the optimal power distribution,
2018
+ which is related to whether the upper bound for capacity can be
2019
+ achieved, the coupling operator
2020
+ EIT
2021
+ G
2022
+ obtained by the EMIT-
2023
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
2024
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
2025
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
2026
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
2027
+
2028
+ Amplitude
2029
+ Phase
2030
+ 180
2031
+ 0.5
2032
+ 0
2033
+ -180
2034
+ (a)
2035
+ Amplitude
2036
+ Phase
2037
+ 180
2038
+ 0.5
2039
+ 0
2040
+ 0
2041
+ -180
2042
+ (b)
2043
+ Error
2044
+ Error
2045
+ 0.5
2046
+ 0
2047
+ 0
2048
+ -180
2049
+ (c)0.16
2050
+ Port 1
2051
+ Port 2
2052
+ Port 3
2053
+ 0.14
2054
+ -
2055
+ Mode 1
2056
+ -
2057
+ -
2058
+ Mode 2
2059
+ 0.12
2060
+ Mode 3
2061
+ 0.1
2062
+ 0.08
2063
+ 0.06
2064
+ 0.04
2065
+ 0.02
2066
+ 0
2067
+ 0
2068
+ 0.1
2069
+ 0.2
2070
+ 0.3
2071
+ 0.4
2072
+ 0.5
2073
+ 0.6
2074
+ 0.7
2075
+ 0.8
2076
+ 0.9
2077
+ Receiving Position (m)10
2078
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2079
+
2080
+
2081
+ Fig. 14. Image transmission case based on EMIT-based model. (a) Original
2082
+ image with 128*128 pixels. (b) Optimized power distribution. (c)-(e)
2083
+ Conducting single-mode transmission using mode 1, mode 2 and mode 3
2084
+ respectively.
2085
+
2086
+ based model is used to endow the shape of the transmitted
2087
+ signal to obtain the best quality of information transmission.
2088
+ Rewrite (5-6) to obtain a new coupling equation based on the
2089
+ scenario we’re considering:
2090
+
2091
+
2092
+ =
2093
+
2094
+
2095
+ EIT
2096
+ EIT
2097
+ U
2098
+ Y
2099
+ SV
2100
+ X ,
2101
+ (23)
2102
+
2103
+ where
2104
+ EIT
2105
+ U
2106
+ and
2107
+ EIT
2108
+ V
2109
+ are determined by
2110
+ EIT
2111
+ G
2112
+ to guide the
2113
+ signal processing of transmitter and receiver, respectively. X
2114
+ and Y represent the signal form of transmitter and receiver,
2115
+ respectively. The singular value matrix S is disassembled to
2116
+ obtain the received signal evaluation function f :
2117
+
2118
+ 3
2119
+ 1
2120
+ m
2121
+ m
2122
+ m
2123
+ f
2124
+ V
2125
+
2126
+
2127
+ =
2128
+ = 
2129
+ .
2130
+ (24)
2131
+
2132
+ In (24), X is decomposed as a bitstream of information X
2133
+ (here we use simple binary phase-shift keying (BPSK)
2134
+ modulation) multiplied by the excitation coefficient  , and
2135
+ the unitary matrix
2136
+ EIT
2137
+ V
2138
+ is decomposed into three-mode vectors
2139
+ (
2140
+ 1,2,3)
2141
+ m
2142
+ V
2143
+ m =
2144
+ . Since there are three sources in our MIMO
2145
+ system, both
2146
+ m
2147
+ V
2148
+ and
2149
+
2150
+ here have three elements.
2151
+ (
2152
+ 1,2,3)
2153
+ m m
2154
+
2155
+ =
2156
+ are the diagonal elements of the singular value
2157
+ matrix S , representing the influence of each mode on the
2158
+ receiving end. To have a clearer understanding of
2159
+ m
2160
+
2161
+ , we
2162
+ depicted the electric field distribution at the receiving end with
2163
+ the help of the EMIT-based model, as shown in Fig. 13. The
2164
+ calculated proportions of the three modes are 67.55%, 23.19%,
2165
+ and 9.25%, respectively. Mode 1 with the strongest proportion
2166
+ just contributes its crest to the receiving end, while mode 3 with
2167
+ the weakest proportion just contributes its trough to the
2168
+ receiving end. This provides a clear perspective for signal
2169
+ waveform design from the EM point of view, and reveals that
2170
+ EM space is not the only factor determining mode contribution,
2171
+ and EM space characteristics and RF front-end characteristics
2172
+ should be considered together.
2173
+ According to the crosstalk matrix calculated in (22), we treat
2174
+ these three modes as orthogonal. Therefore,
2175
+ m
2176
+ V
2177
+ becomes a
2178
+ set of orthogonal basis in a Hilbert space, meeting
2179
+
2180
+ 0
2181
+ m
2182
+ n
2183
+ V
2184
+ V
2185
+ =
2186
+
2187
+ and
2188
+
2189
+ 1
2190
+ m
2191
+ m
2192
+ V
2193
+ V
2194
+ = . Hence,  can be written as an orthogonal
2195
+ basis expansion:
2196
+
2197
+
2198
+ 3
2199
+ 1
2200
+ m
2201
+ m
2202
+ m
2203
+ V
2204
+
2205
+
2206
+ =
2207
+ = 
2208
+ ,
2209
+ (25)
2210
+
2211
+ where
2212
+ m
2213
+
2214
+ represent the corresponding weight of each basis
2215
+ vector, which determines the power distribution on the
2216
+ transmitting source. Therefore, the constraint of constant total
2217
+ power
2218
+ 0P can be equivalent to that the excitation vector is
2219
+ located on a fixed circle in the Hilbert space, and the received
2220
+ signal evaluation function f is the sum of the weighted
2221
+ projections of the excitation vector on the three basis functions:
2222
+
2223
+
2224
+ 0
2225
+ find :
2226
+ (
2227
+ 1,2,3)
2228
+ max :
2229
+ . .:
2230
+ m
2231
+ m
2232
+ m
2233
+ f
2234
+ s t
2235
+ P
2236
+
2237
+
2238
+
2239
+ =
2240
+ 
2241
+
2242
+ =
2243
+
2244
+
2245
+ .
2246
+ (26)
2247
+
2248
+ The optimization problem can be easily solved by using
2249
+ Cauchy inequality. By substituting (8), the information transfer
2250
+ function can reach the maximum value only when
2251
+ /
2252
+ m
2253
+ m
2254
+
2255
+  is a
2256
+ constant for different m . This is similar to the “water-filling”
2257
+ algorithm in channel estimation, while the core difference is
2258
+ that the key informatics parameters in this paper are deduced by
2259
+ an effective EM algorithm. In addition,
2260
+
2261
+ TR
2262
+ TR
2263
+ G
2264
+ G
2265
+ or
2266
+
2267
+ EIT
2268
+ EIT
2269
+ G
2270
+ G
2271
+
2272
+ have the same physical meaning as the transmit signal
2273
+ covariance matrix, and the main difference is that
2274
+
2275
+ TR
2276
+ TR
2277
+ G
2278
+ G
2279
+ or
2280
+
2281
+ EIT
2282
+ EIT
2283
+ G
2284
+ G
2285
+ is calculated by the EM methods based on dyadic
2286
+ Green’s function.
2287
+ It is seen from the mode analysis based on the EMIT-based
2288
+ model that only under the guidance of a specific power
2289
+ allocation strategy, MIMO information transmission can
2290
+ achieve the effect of receiving power equal to transmitting
2291
+ power times path loss. To fully illustrate the guiding
2292
+ significance of the EMIT-based model for power distribution
2293
+ (essentially waveform design), Fig. 14 shows an image
2294
+ transmission case. BPSK is used to discretize every pixel in the
2295
+ picture into an 8-bit data stream for transmission, and noise 
2296
+ is joined to EM space. Obviously, under the premise of not
2297
+ processing channel noise, the RF front-end working strategy
2298
+ based on the EMIT-based model is much better than other
2299
+ transmission modes. Therefore, the EMIT-based model can not
2300
+ only efficiently represent the complex space, but also make
2301
+ more valuable guidance for wireless communication.
2302
+
2303
+ V. CONCLUSION
2304
+ In this article, an EMIT-based model is presented to simulate
2305
+ the performance of MIMO systems in complex EM complex
2306
+ space effectively. Firstly, the EM expression of the information
2307
+ coupling operator is given in the free space, and two key
2308
+ informatics parameters, EM effective capacity and path loss,
2309
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
2310
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
2311
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
2312
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
2313
+
2314
+ (a)
2315
+ (b)
2316
+ (c)
2317
+ (p)
2318
+ (e)11
2319
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2320
+
2321
+ are extracted from the EM perspective. It is proved that the
2322
+ MIMO antennas’ aperture is the critical factor in the EM
2323
+ effective capacity of the MIMO system. Secondly, the basic
2324
+ principle of the EM representation method in complex space is
2325
+ given, and several typical scenarios are analyzed, which proves
2326
+ the accuracy and efficiency of the EMIT-based model proposed.
2327
+ The results show that the EMIT-based model can reliably
2328
+ analyze the electromagnetic space about 10% of the time
2329
+ compared to the full-wave simulation. Finally, the MIMO
2330
+ performance in real propagation scenarios is calculated using
2331
+ the EMIT-based model. The experimental results verify that the
2332
+ channel matrix calculated is in good agreement with the
2333
+ measured ones. Based on this, it is pointed out how the EMIT-
2334
+ based model can effectively guide the MIMO design and
2335
+ feeding in a power distribution question.
2336
+ The ultimate goal of the proposed EMIT-based model is to
2337
+ advance the development of EMIT and demonstrate a new idea
2338
+ of extracting information parameters to the antenna &
2339
+ propagation community using the basis of computational
2340
+ electromagnetism. Currently, it is suitable for cluster models
2341
+ with arbitrary distribution, size, and material, providing an
2342
+ efficient and reliable method for guiding the power and phase
2343
+ allocation of antenna units in scattering complex space. The
2344
+ proposed model can also be easily extended to the guidance of
2345
+ MIMO antenna design in complex spaces by numerical discrete
2346
+ and optimization methods.
2347
+
2348
+ REFERENCES
2349
+ [1] C. Ehrenborg and M. Gustafsson, "Physical Bounds and Radiation Modes
2350
+ for MIMO Antennas," IEEE Transactions on Antennas and Propagation,
2351
+ vol. 68, no. 6, pp. 4302-4311. 2020.
2352
+ [2] D. Li, T. Li, E. Li, and Y. Zhang, "A 2.5-D Angularly Stable Frequency
2353
+ Selective Surface Using Via-Based Structure for 5G EMI Shielding,"
2354
+ IEEE Transactions on Electromagnetic Compatibility, vol. 60, no. 3, pp.
2355
+ 768-775. 2018.
2356
+ [3] D. He, B. Ai, K. Guan, L. Wang, Z. Zhong, and T. Kurner, "The Design
2357
+ and Applications of High-Performance Ray-Tracing Simulation Platform
2358
+ for 5G and Beyond Wireless Communications: A Tutorial," IEEE
2359
+ Communications Surveys & Tutorials, vol. 21, no. 1, pp. 10-27. 2019.
2360
+ [4] H. Gao, K. Xiao, B. Xia, and Z. Chen, "Mutual Information Analysis of
2361
+ Mixed-ADC MIMO Systems Over Rayleigh Channels Based on Random
2362
+ Matrix Theory," IEEE Transactions on Wireless Communications, vol. 19,
2363
+ no. 7, pp. 4894-4906. 2020.
2364
+ [5] M. A. Azam, A. K. Dutta, and A. Mukherjee, "Performance Analysis of
2365
+ Dipole Antenna Based Planar Arrays With Mutual Coupling and Antenna
2366
+ Position Error in mmWave Hybrid System," IEEE Transactions on
2367
+ Vehicular Technology, vol. 70, no. 10, pp. 10209-10221. 2021.
2368
+ [6] S. Ghosal, R. Sinha, A. De, and A. Chakrabarty, "Characteristic Mode
2369
+ Analysis of Mutual Coupling," IEEE Transactions on Antennas and
2370
+ Propagation, vol. 70, no. 2, pp. 1008-1019. 2022.
2371
+ [7] Y. Li and Q. Chu, "Coplanar Dual-Band Base Station Antenna Array Using
2372
+ Concept of Cavity-Backed Antennas," IEEE Transactions on Antennas
2373
+ and Propagation, vol. 69, no. 11, pp. 7343-7354. 2021.
2374
+ [8] H. Sun, C. Ding, H. Zhu, B. Jones, and Y. J. Guo, "Suppression of Cross-
2375
+ Band Scattering in Multiband Antenna Arrays," IEEE Transactions on
2376
+ Antennas and Propagation, vol. 67, no. 4, pp. 2379-2389. 2019.
2377
+ [9] J. Jin, F. Feng, J. Zhang, J. Ma, and Q. Zhang, "Efficient EM Topology
2378
+ Optimization Incorporating Advanced Matrix Padé Via Lanczos and
2379
+ Genetic Algorithm for Microwave Design," IEEE Transactions on
2380
+ Microwave Theory and Techniques, vol. 69, no. 8, pp. 3645-3666. 2021.
2381
+ [10] M. M. Taygur and T. F. Eibert, "A Ray-Tracing Algorithm Based on the
2382
+ Computation of (Exact) Ray Paths With Bidirectional Ray-Tracing,"
2383
+ IEEE Transactions on Antennas and Propagation, vol. 68, no. 8, pp.
2384
+ 6277-6286. 2020.
2385
+ [11] Z. Cui, K. Guan, C. Briso-Rodriguez, B. Ai, and Z. Zhong, "Frequency-
2386
+ Dependent
2387
+ Line-of-Sight
2388
+ Probability
2389
+ Modeling
2390
+ in
2391
+ Built-Up
2392
+ Environments," IEEE Internet of Things Journal, vol. 7, no. 1, pp. 699-
2393
+ 709. 2020.
2394
+ [12] F. Quatresooz, S. Demey, and C. Oestges, "Tracking of Interaction Points
2395
+ for Improved Dynamic Ray Tracing," IEEE Transactions on Vehicular
2396
+ Technology, vol. 70, no. 7, pp. 6291-6301. 2021.
2397
+ [13] K. K. Tse, L. Tsang, C. H. Chan, K. H. Ding, and K. W. Leung, "Multiple
2398
+ scattering of waves by dense random distributions of sticky particles for
2399
+ applications in microwave scattering by terrestrial snow," Radio Science,
2400
+ vol. 42, no. 5, p. n/a-n/a. 2007.
2401
+ [14] H. Huang, L. Tsang, E. G. Njoku, A. Colliander, T. Liao, and K. Ding,
2402
+ "Propagation and Scattering by a Layer of Randomly Distributed
2403
+ Dielectric Cylinders Using Monte Carlo Simulations of 3D Maxwell
2404
+ Equations With Applications in Microwave Interactions With
2405
+ Vegetation," IEEE Access, vol. 5, pp. 11985-12003. 2017.
2406
+ [15] C. Chuah, D. N. C. Tse, J. M. Kahn, and R. A. Valenzuela, "Capacity
2407
+ scaling in MIMO wireless systems under correlated fading," IEEE
2408
+ transactions on information theory, vol. 48, no. 3, pp. 637-650, 2002-01-
2409
+ 01. 2002.
2410
+ [16] L. Bai, Z. Huang, Y. Li, and X. Cheng, "A 3D Cluster-Based Channel
2411
+ Model for 5G and Beyond Vehicle-to-Vehicle Massive MIMO
2412
+ Channels," IEEE Transactions on Vehicular Technology, vol. 70, no. 9,
2413
+ pp. 8401-8414. 2021.
2414
+ [17] C. Wang, J. Bian, J. Sun, W. Zhang, and M. Zhang, "A Survey of 5G
2415
+ Channel Measurements and Models," IEEE Communications Surveys &
2416
+ Tutorials, vol. 20, no. 4, pp. 3142-3168. 2018.
2417
+ [18] Y. Zeng, B. Duan, S. Lou, and S. Zhang, "Modeling and Analysis of
2418
+ Airborne Conformal Arrays Obstructed by Fixed Blockage," IEEE
2419
+ Transactions on Antennas and Propagation, vol. 70, no. 6, pp. 4342-4354.
2420
+ 2022.
2421
+ [19] H. Zhao, Y. Shuang, M. Wei, T. J. Cui, P. D. Hougne, and L. Li,
2422
+ "Metasurface-assisted massive backscatter wireless communication with
2423
+ commodity Wi-Fi signals," Nature Communications, vol. 11, no. 1. 2020.
2424
+ [20] P. Del Hougne, M. Davy, and U. Kuhl, "Optimal Multiplexing of Spatially
2425
+ Encoded Information across Custom-Tailored Configurations of a
2426
+ Metasurface-Tunable Chaotic Cavity," Physical review applied, vol. 13,
2427
+ no. 4, 2020-01-01. 2020.
2428
+ [21] P. Del Hougne, M. Fink, and G. Lerosey, "Optimally diverse
2429
+ communication channels in disordered environments with tuned
2430
+ randomness," Nature Electronics, vol. 2, no. 1, pp. 36-41. 2019.
2431
+ [22] S. Gong, X. Lu, D. T. Hoang, D. Niyato, L. Shu, and D. I. Kim et al.,
2432
+ "Toward Smart Wireless Communications via Intelligent Reflecting
2433
+ Surfaces: A Contemporary Survey," IEEE Communications Surveys &
2434
+ Tutorials, vol. 22, no. 4, pp. 2283-2314. 2020.
2435
+ [23] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M. Alouini, and R. Zhang,
2436
+ "Wireless
2437
+ Communications
2438
+ Through
2439
+ Reconfigurable
2440
+ Intelligent
2441
+ Surfaces," IEEE Access, vol. 7, pp. 116753-116773. 2019.
2442
+ [24] T. Cui, S. Liu, and L. Li, "Information entropy of coding metasurface,"
2443
+ Light: Science & Applications, vol. 5, no. 11, p. e16172. 2016.
2444
+ [25] D. Dardari, "Communicating With Large Intelligent Surfaces:
2445
+ Fundamental Limits and Models," IEEE Journal on Selected Areas in
2446
+ Communications, vol. 38, no. 11, pp. 2526-2537. 2020.
2447
+ [26] A. S. Y. Poon, R. W. Brodersen, and D. N. C. Tse, "Degrees of Freedom
2448
+ in Multiple-Antenna Channels: A Signal Space Approach," IEEE
2449
+ Transactions on Information Theory, vol. 51, no. 2, pp. 523-536. 2005.
2450
+ [27] H. Wu, G. D. Bai, S. Liu, L. Li, X. Wan, and Q. Cheng et al., "Information
2451
+ theory of metasurfaces," National Science Review, vol. 7, no. 3, pp. 561-
2452
+ 571, 2020-03-01. 2020.
2453
+ [28] M. Ke, Z. Gao, Y. Wu, X. Gao, and R. Schober, "Compressive Sensing-
2454
+ Based Adaptive Active User Detection and Channel Estimation: Massive
2455
+ Access Meets Massive MIMO," IEEE Transactions on Signal Processing,
2456
+ vol. 68, pp. 764-779. 2020.
2457
+ [29] F. K. Gruber and E. A. Marengo, "New Aspects of Electromagnetic
2458
+ Information Theory for Wireless and Antenna Systems," IEEE
2459
+ Transactions on Antennas and Propagation, vol. 56, no. 11, pp. 3470-
2460
+ 3484. 2008.
2461
+ [30] J. Xu and R. Janaswamy, "Electromagnetic Degrees of Freedom in 2-D
2462
+ Scattering Environments," IEEE Transactions on Antennas and
2463
+ Propagation, vol. 54, no. 12, pp. 3882-3894. 2006.
2464
+ [31] K. Choutagunta, I. Roberts, D. A. B. Miller, and J. M. Kahn, "Adapting
2465
+ Mach–Zehnder Mesh Equalizers in Direct-Detection Mode-Division-
2466
+ Multiplexed Links," Journal of Lightwave Technology, vol. 38, no. 4, pp.
2467
+ 723-735. 2020.
2468
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
2469
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
2470
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
2471
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
2472
+
2473
+ 12
2474
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2475
+
2476
+ [32] M. Lee, M. A. Neifeld, and A. Ashok, "Capacity of electromagnetic
2477
+ communication modes in a noise-limited optical system," Applied optics
2478
+ (2004), vol. 55, no. 6, p. 1333, 2016-01-01. 2016.
2479
+ [33] S. S. A. Yuan, Z. He, X. Chen, C. Huang, and W. E. I. Sha,
2480
+ "Electromagnetic Effective Degree of Freedom of an MIMO System in
2481
+ Free Space," IEEE Antennas and Wireless Propagation Letters, vol. 21,
2482
+ no. 3, pp. 446-450. 2022.
2483
+ [34] T. K. Sarkar and M. Salazar-Palma, "MIMO: Does It Make Sense From an
2484
+ Electromagnetic Perspective and Illustrated Using Computational
2485
+ Electromagnetics?" IEEE Journal on Multiscale and Multiphysics
2486
+ Computational Techniques, vol. 4, pp. 269-281. 2019.
2487
+ [35] M. Horodynski, M. Kühmayer, C. Ferise, S. Rotter, and M. Davy, "Anti-
2488
+ reflection structure for perfect transmission through complex media,"
2489
+ Nature, vol. 607, no. 7918, pp. 281-286, 2022-07-14. 2022.
2490
+ [36] X. Chen, P. Kildal, J. Carlsson, and J. Yang, "MRC Diversity and MIMO
2491
+ Capacity Evaluations of Multi-Port Antennas Using Reverberation
2492
+ Chamber and Anechoic Chamber," IEEE Transactions on Antennas and
2493
+ Propagation, vol. 61, no. 2, pp. 917-926. 2013.
2494
+
2495
+
2496
+
2497
+
2498
+
2499
+
2500
+
2501
+
2502
+
2503
+
2504
+
2505
+
2506
+
2507
+
2508
+
2509
+
2510
+
2511
+
2512
+
2513
+
2514
+
2515
+
2516
+
2517
+ Ruifeng Li received the B.S. degree in engineering from
2518
+ University of Electronic Science and Technology of
2519
+ China, Chengdu, China, in 2020. He is currently pursuing
2520
+ the Ph.D. degree at the College of Information Science
2521
+ and Electronic Engineering, Zhejiang University.
2522
+ His
2523
+ current
2524
+ research
2525
+ interests
2526
+ include
2527
+ the
2528
+ electromagnetic
2529
+ information
2530
+ theory
2531
+ for
2532
+ wireless
2533
+ communication, and efficient calculation methods
2534
+ applied in MIMO antennas.
2535
+
2536
+
2537
+ Da Li received the B.S. degree in 2014, and the Ph.D.
2538
+ degree in 2019, from Zhejiang University, Hangzhou,
2539
+ China, both in electrical engineering. From 2017 to
2540
+ 2018, he worked at Nanyang Technological University,
2541
+ Singapore, as a Project Researcher. From 2019 to 2021,
2542
+ he joined Science and Technology on Antenna and
2543
+ Microwave Laboratory, Nanjing, China, as a Research
2544
+ Fellow. He is currently an assistant professor at Zhejiang
2545
+ University. His research interests include machine
2546
+ learning, antennas, matesurfaces, and electromagnetic compatibility. Dr. Li
2547
+ has authored or coauthored more than 40 refereed papers and served as
2548
+ Reviewers for 6 technical journals and TPC Members of 3 IEEE conferences.
2549
+ He was also a recipient of the Outstanding Young Scientist Award at 2022
2550
+ Asia-Pacific International Symposium on Electromagnetic Compatibility.
2551
+
2552
+ Jinyan Ma received the B.S. degree in engineering from
2553
+ Zhejiang University, Hangzhou, China, in 2021. He is
2554
+ currently working toward the Ph.D. degree in electronics
2555
+ science and technology with the College of Information
2556
+ Science and Electronic Engineering, Zhejiang University,
2557
+ Hangzhou, China.
2558
+ His
2559
+ current
2560
+ research
2561
+ interests
2562
+ include
2563
+ the
2564
+ electromagnetic information theory and efficient
2565
+ electromagnetic calculation methods.
2566
+
2567
+
2568
+ Zhaoyang Feng received the B.Sc degree from North
2569
+ China Electric Power University, Beijing, China, in 2017.
2570
+ He is currently working toward the Ph.D. degree in the
2571
+ College
2572
+ of
2573
+ Information
2574
+ Science
2575
+ and
2576
+ Electronic
2577
+ Engineering, Zhejiang University, Hangzhou, Zhejiang.
2578
+ His current research interests include electromagnetic
2579
+ compatibility,
2580
+ computational
2581
+ electromagnetics and
2582
+ multiple scattering theory
2583
+
2584
+
2585
+
2586
+ Ling Zhang (Member, IEEE) received the B.S. degree in
2587
+ electrical engineering from Huazhong University of
2588
+ Science and Technology, Wuhan, China, in 2015, and the
2589
+ M.S. and Ph.D. degrees from Missouri S&T, Rolla, MO,
2590
+ USA, in 2017 and 2021, respectively, both in electrical
2591
+ engineering. He was with Cisco as a student intern from
2592
+ Aug. 2016 to Aug. 2017. He joined Zhejiang University,
2593
+ Hangzhou, China as a research fellow in 2021. He has
2594
+ authored and co-authored more than 30 journal and
2595
+ conference papers. His research interests include machine learning, power
2596
+ integrity, electromagnetic interference, radio-frequency interference, and signal
2597
+ integrity.
2598
+ Dr. Zhang was an Organizing Committee, Special Session Chair, Workshop
2599
+ Session Chair, and Poster Session Chair in APEMC 2022. He has given invited
2600
+ presentations at the IBIS Summit at 2021 IEEE Virtual Symposium on
2601
+ EMC+SIPI, and the 2021 Virtual Asian IBIS Summit China. He was the
2602
+ recipient of the Honorable Mention Paper in APEMC 2022, the Best Paper
2603
+ Award in DesignCon 2019, and the Student Paper Finalist Award in ACES
2604
+ Symposium in 2021. He was also the recipient of the Outstanding Young
2605
+ Scientist Reward in APEMC 2022.
2606
+
2607
+ Shurun Tan (S’14-M’17) received the B.E. degree in
2608
+ information
2609
+ engineering
2610
+ and
2611
+ M.Sc.
2612
+ degree
2613
+ in
2614
+ electromagnetic field and microwave techniques from
2615
+ the Southeast University, Nanjing, China, in 2009 and
2616
+ 2012, respectively, and the Ph.D. degree in electrical
2617
+ engineering from the University of Michigan, Ann
2618
+ Arbor, MI, USA, in Dec. 2016.
2619
+ Dr. Tan is an assistant professor in the Zhejiang
2620
+ University / University of Illinois at Urbana-Champaign
2621
+ Institute located at the International Campus of Zhejiang University, Haining,
2622
+ China. He is also affiliated with the State Key Laboratory of Modern Optical
2623
+ Instrumentation, and the College of Information Science and Electronic
2624
+ Engineering, Zhejiang University, Hangzhou, China. He is also an adjunct
2625
+ assistant professor in the Department of Electrical and Computer Engineering,
2626
+ University of Illinois at Urbana-Champaign, Urbana, USA. From Dec. 2010 to
2627
+ Nov. 2011, he was a Visiting Student with the Department of Electrical and
2628
+ Computer Engineering, the University of Houston, Houston, TX, USA. From
2629
+ Sep. 2012 to Dec. 2014, he was a PhD candidate with the Department of
2630
+ Electrical Engineering, the University of Washington, Seattle, WA, USA. From
2631
+ Jan. 2015 to Dec. 2018, he had been affiliated with the Radiation Laboratory,
2632
+ and the Department of Electrical Engineering and Computer Science, the
2633
+ University of Michigan, Ann Arbor, first as a PhD candidate, and then as a
2634
+ postdoctoral research fellow since Jan. 2017.
2635
+ Dr. Tan is working on electromagnetic theory, computational and applied
2636
+ electromagnetics. His research interests include electromagnetic scattering of
2637
+ random media and periodic structures, microwave remote sensing,
2638
+ electromagnetic information systems with electromagnetic wave-functional
2639
+ devices, electromagnetic integrity in high-speed and high-density electronic
2640
+ integration, electromagnetic environment and reliability of complex electronic
2641
+ systems, etc.
2642
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
2643
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
2644
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
2645
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
2646
+
2647
+ 13
2648
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2649
+
2650
+
2651
+
2652
+ Wei E. I. Sha (M’09-SM’17) received the B.S. and Ph.D.
2653
+ degrees in Electronic Engineering at Anhui University,
2654
+ Hefei, China, in 2003 and 2008, respectively. From Jul.
2655
+ 2008 to Jul. 2017, he was a Postdoctoral Research
2656
+ Fellow and then a Research Assistant Professor in the
2657
+ Department of Electrical and Electronic Engineering at
2658
+ the University of Hong Kong, Hong Kong. From Mar.
2659
+ 2018 to Mar. 2019, he worked at University College
2660
+ London as a Marie Skłodowska-Curie Individual Fellow.
2661
+ From Oct. 2017, he joined the College of Information Science & Electronic
2662
+ Engineering at Zhejiang University, Hangzhou, China, where he is currently a
2663
+ tenure-tracked Assistant Professor.
2664
+ Dr. Sha has authored or coauthored 180 refereed journal papers, 150
2665
+ conference publications (including 5 keynote talks and 1 short course), 9 book
2666
+ chapters, and 2 books. His Google Scholar citation is 8193 with h-index of 45.
2667
+ He is a senior member of IEEE and a member of OSA. He served as Reviewers
2668
+ for 60 technical journals and Technical Program Committee Members of 10
2669
+ IEEE conferences. He also served as Associate Editors of IEEE Journal on
2670
+ Multiscale and Multiphysics Computational Techniques, IEEE Open Journal of
2671
+ Antennas and Propagation, and IEEE Access. In 2015, he was awarded Second
2672
+ Prize of Science and Technology from Anhui Province Government, China. In
2673
+ 2007, he was awarded the Thousand Talents Program for Distinguished Young
2674
+ Scholars of China. He was the recipient of ACES Technical Achievement
2675
+ Award 2022 and PIERS Young Scientist Award 2021. Dr. Sha also received 6
2676
+ Best Student Paper Prizes and one Young Scientist Award with his students.
2677
+ His research interests include theoretical and computational research in
2678
+ electromagnetics and optics, focusing on the multiphysics and interdisciplinary
2679
+ research. His research involves fundamental and applied aspects in
2680
+ computational and applied electromagnetics, nonlinear and quantum
2681
+ electromagnetics, micro- and nano-optics, optoelectronic device simulation,
2682
+ and multiphysics modeling.
2683
+
2684
+
2685
+
2686
+ Hongsheng Chen received the B.S. and Ph.D.degrees in
2687
+ electrical engineering from Zhejiang University (ZJU),
2688
+ Hangzhou, China, in 2000 and 2005, respectively.
2689
+ In 2005, he became an Assistant Professor with ZJU,
2690
+ where he was an Associate Professor in 2007 and a Full
2691
+ Professor in 2011.
2692
+ He was a Visiting Scientist from 2006 to 2008 and a
2693
+ Visiting Professor from 2013 to 2014 with the Research
2694
+ Laboratory of Electronics, Massachusetts Institute of
2695
+ Technology, Cambridge, MA, USA. He is currently a Chang Jiang Scholar
2696
+ Distinguished Professor with the Electromagnetics Academy, ZJU. He has
2697
+ coauthored more than 200 international refereed joumal papers. His works have
2698
+ been highlighted by many scientific magazines and public media, including
2699
+ Nature, Scientific American, MIT Technology Review, Physorg, and so on. His
2700
+ current research interests include metamaterials, invisibility cloaking,
2701
+ transformation optics, graphene, and theoretical and numerical methods of
2702
+ electromagnetics.
2703
+ Dr. Chen serves as a Regular Reviewer for many international journals on
2704
+ electromagnetics, physics, optics, and electrical engineering. He serves as a
2705
+ Topical Editor for the Journal of Optics and the Editorial Board for Nature’s
2706
+ Scientific Reports and Progress in Electromagnetics Research. He was a
2707
+ recipient of the National Excellent Doctoral Dissertation Award in China in
2708
+ 2008, the Zhejiang Provincial Outstanding Youth Foundation in 2008, the
2709
+ National Youth Top-Notch Talent Support Program in China in 2012, the New
2710
+ Century Excellent Talents in University of China in 2012, the National Science
2711
+ Foundation for Excellent Young Scholars of China in 2013, and the National
2712
+ Science Foundation for Distinguished Young Scholars of China in 2016. His
2713
+ research work on an invisibility cloak was selected in Science Development
2714
+ Report as one of the representative achievements of Chinese Scientists in 2007.
2715
+
2716
+ Er-Ping Li (S’91, M’92, SM’01, F’08) is currently a
2717
+ Qiushi-Distinguished Professor with Department of
2718
+ Information Science and Electronic Engineering,
2719
+ Zhejiang University, China; served as Founding Dean
2720
+ for Institute of Zhejiang University - University of
2721
+ Illinois at Urbana-Champaign in 2016. From 1993, he has
2722
+ served as a Research Fellow, Associate Professor,
2723
+ Professor and Principal Scientist and Senior Director at
2724
+ the Singapore Research Institute and University. Dr Li
2725
+ authored or co-authored over 400 papers published in the referred international
2726
+ journals, authored two books published by John-Wiley-IEEE Press and
2727
+ Cambridge University Press. He holds and has filed a number of patents at the
2728
+ US patent office. His research interests include electrical modeling and design
2729
+ of micro/nano-scale integrated circuits, 3D electronic package integration.
2730
+ Dr. Li is a Fellow of IEEE, and a Fellow of USA Electromagnetics Academy,
2731
+ a Fellow of Singapore Academy of Engineering. He is the recipient of IEEE
2732
+ EMC Technical Achievement Award in 2006, Singapore IES Prestigious
2733
+ Engineering Achievement Award and Changjiang Chair Professorship Award
2734
+ in 2007, 2015 IEEE Richard Stoddard Award on EMC, 2021 IEEE EMC
2735
+ Laurence G. Cumming Award and Zhejiang Natural Science 1st Class Award.
2736
+ He served as an Associate Editor for the IEEE MICROWAVE AND
2737
+ WIRELESS COMPONENTS LETTERS from 2006-2008 and for IEEE
2738
+ TRANSACTIOSN on EMC from 2006-2021, Guest Editor for 2006 and 2010
2739
+ IEEE TRANSACTIOSN on EMC Special Issues, Guest Editor for 2010 IEEE
2740
+ TRANSACTIONS on MTT APMC Special Issue. He is currently an Associate
2741
+ Editor for the IEEE TRANSACTIONS ON SIGNAL and POWER
2742
+ INTEGRITY and Deputy Editor in Chief of Electromagnetics Science. He has
2743
+ been a General Chair and Technical Chair, for many international conferences.
2744
+ He was the President for 2006 International Zurich Symposium on EMC, the
2745
+ Founding General Chair for Asia-Pacific EMC Symposium, General Chair for
2746
+ 2008, 2012, 2016, 2018, 2022 APEMC, and 2010 IEEE Symposium on
2747
+ Electrical Design for Advanced Packaging Systems. He has been invited to give
2748
+ 120 invited talks and plenary speeches at various international conferences and
2749
+ forums.
2750
+
2751
+
2752
+ This article has been accepted for publication in IEEE Transactions on Antennas and Propagation. This is the author's version which has not been fully edited and
2753
+ content may change prior to final publication. Citation information: DOI 10.1109/TAP.2023.3235015
2754
+ © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.��See https://www.ieee.org/publications/rights/index.html for more information.
2755
+ Authorized licensed use limited to: Zhejiang University. Downloaded on January 13,2023 at 13:17:46 UTC from IEEE Xplore. Restrictions apply.
2756
+
B9E5T4oBgHgl3EQfTg8R/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
CtAyT4oBgHgl3EQfefgG/content/2301.00320v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5632b32308f0d2d630c303a15cf4d9abea15326830c9a16692ac42c41d75def7
3
+ size 661464
CtAyT4oBgHgl3EQfefgG/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:00370d90add7c64c822bea5328bb94f02b27cf0424b9c223a59681426ae0f71e
3
+ size 1048621
CtAyT4oBgHgl3EQfefgG/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:07d7de155a137b020b38916469f7d8eeca299f3063830ef3555d4c49f2899df6
3
+ size 40468
D9A0T4oBgHgl3EQfAv_u/content/tmp_files/2301.01968v1.pdf.txt ADDED
@@ -0,0 +1,1499 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1
2
+ Deep Learning of Force Manifolds from the
3
+ Simulated Physics of Robotic Paper Folding
4
+ Dezhong Tong∗,1, Andrew Choi∗,2, Demetri Terzopoulos2, Jungseock Joo3, and Mohammad Khalid Jawed†,1
5
+ Abstract—Robotic manipulation of slender objects is challeng-
6
+ ing, especially when the induced deformations are large and
7
+ nonlinear. Traditionally, learning-based control approaches, e.g.,
8
+ imitation learning, have been used to tackle deformable material
9
+ manipulation. Such approaches lack generality and often suffer
10
+ critical failure from a simple switch of material, geometric, and/or
11
+ environmental (e.g., friction) properties. In this article, we ad-
12
+ dress a fundamental but difficult step of robotic origami: forming
13
+ a predefined fold in paper with only a single manipulator. A data-
14
+ driven framework combining physically-accurate simulation and
15
+ machine learning is used to train deep neural network models
16
+ capable of predicting the external forces induced on the paper
17
+ given a grasp position. We frame the problem using scaling
18
+ analysis, resulting in a control framework robust against material
19
+ and geometric changes. Path planning is carried out over the
20
+ generated manifold to produce robot manipulation trajectories
21
+ optimized to prevent sliding. Furthermore, the inference speed
22
+ of the trained model enables the incorporation of real-time
23
+ visual feedback to achieve closed-loop sensorimotor control. Real-
24
+ world experiments demonstrate that our framework can greatly
25
+ improve robotic manipulation performance compared against
26
+ natural paper folding strategies, even when manipulating paper
27
+ objects of various materials and shapes.
28
+ Index Terms—robotic manipulation, deformable material ma-
29
+ nipulation, deep neural networks, data-driven models, closed-loop
30
+ sensorimotor control
31
+ I. INTRODUCTION
32
+ From shoelaces to clothes, we encounter flexible slender
33
+ structures throughout our everyday lives. These structures are of-
34
+ ten characterized by their ability to undergo large deformations
35
+ when subjected even to moderate forces, such as gravity. People
36
+ possess an incredible innate understanding of the dynamics of
37
+ such deformable objects; e.g., we can use gravity to perfectly
38
+ manipulate a shirt over our heads. Instilling such intuition
39
+ into robots remains an important research problem and has
40
+ the potential to breed numerous applications with considerable
41
+ economic and humanitarian potential. Some examples include
42
+ preparing deformable products in the food industry [1], [2],
43
+ assisting in the medical field [3]–[5], and providing caregiving
44
+ assistance to elderly and disabled communities, including with
45
+ respect to dressing [6]–[10] and feeding [11], [12]. However, the
46
+ The authors are with the University of California, Los Angeles (UCLA),
47
+ CA 90095, USA.
48
+ 1Dezhong Tong and M. Khalid Jawed are with the UCLA Department of
49
+ Mechanical & Aerospace Engineering (email: tltl960308@g.ucla.edu;
50
+ khalidjm@seas.ucla.edu).
51
+ 2Andrew Choi and Demetri Terzopoulos are with the UCLA Computer
52
+ Science Department (email: asjchoi@cs.ucla.edu; dt@cs.ucla.edu).
53
+ 3Jungseock Joo is with the UCLA Department of Communication and is
54
+ currently working at NVIDIA Corporation (email: jjoo@comm.ucla.edu).
55
+ ∗ Equal contribution.
56
+ † Corresponding author.
57
+ Position &
58
+ material
59
+ parameters
60
+ Learned model
61
+ Force manifold
62
+ Optimization
63
+ algorithm
64
+ Optimal
65
+ path
66
+ Intuitive trajectory
67
+ (circular curve)
68
+ Gripper
69
+ Paper
70
+ Substrate
71
+ Initial free end
72
+ Initial free end
73
+ Obvious sliding
74
+ Minimal sliding
75
+ Initial state
76
+ Folded result
77
+ Intuitive
78
+ manipulation
79
+ Our optimal
80
+ manipulation
81
+ (a)
82
+ (b)
83
+ Fig. 1. Half valley folding for A4 paper with (a) intuitive manipulation and (b)
84
+ our designed optimal manipulation. An intuitive manipulation scheme such as
85
+ tracing a semi-circle experiences significant sliding due to the bending stiffness
86
+ of the paper, resulting in a poor fold. By contrast, our optimal manipulation
87
+ approach achieves an excellent fold by taking into consideration the paper’s
88
+ deformation and thus minimizing sliding.
89
+ robotic manipulation of deformable objects is highly nontrivial
90
+ as a robot must be able to take into account future deformations
91
+ of the manipulated object to complete manipulation tasks
92
+ successfully.
93
+ Prior research has focused primarily on manipulating either
94
+ cloth [13]–[18] or ropes [12], [19]–[25] and as a result, the
95
+ robotic manipulation of many other deformable objects still
96
+ lacks robust solutions. In this article, we address a particularly
97
+ difficult deformable manipulation task — folding paper. Paper
98
+ is similar to cloth but typically has a much larger bending
99
+ stiffness and a slippery surface. Therefore, compared with
100
+ folding garments and fabrics, more delicate and insightful
101
+ manipulations are required for folding sheets of paper.
102
+ A. Our Approach
103
+ We propose a framework that combines physically accurate
104
+ simulation, scaling analysis, and machine learning to generate
105
+ folding trajectories optimized to prevent sliding. With scaling
106
+ analysis, we make the problem non-dimensional, resulting
107
+ in both dimensionality reduction and generality. We then
108
+ train neural networks, whose outputs are referred to as neural
109
+ force manifolds (NFM), to continuously approximate a scaled
110
+ force manifold sampled purely from simulation. Compared to
111
+ numerical models that require the entire geometric configuration
112
+ arXiv:2301.01968v1 [cs.RO] 5 Jan 2023
113
+
114
+ 2
115
+ of the paper, NFMs map the external forces of the paper given
116
+ only the grasp position. Therefore, we can generate trajectories
117
+ optimized to minimize forces (and thus minimize sliding) by
118
+ applying path planning algorithms in near real-time. We show
119
+ that our approach is capable of folding paper on extremely
120
+ slick surfaces with little-to-no sliding (Fig. 1(b)).
121
+ Our main contributions are as follows: (1) we formulate a
122
+ solution to the folding problem in a physically robust manner
123
+ using scaling analysis, resulting in complete generality with
124
+ respect to material, geometric, and environmental properties;
125
+ (2) we train a neural network with non-dimensional simulation
126
+ data forming a fast and accurate model that can generate a
127
+ descriptive force manifold for trajectory optimization; (3) we
128
+ utilize the high inference speed of our trained model with a
129
+ perception system to construct a robust and efficient closed-
130
+ loop sensorimotor control algorithm for the folding task, and
131
+ finally (4) we demonstrate full sim2real realization through
132
+ an extensive robotic case study featuring 210 experiments
133
+ across paper sheets of various materials and shapes. While
134
+ several previous works have trained their policies purely from
135
+ simulation data [7], [19], [26]–[28], these works lacked real
136
+ world validation. To our knowledge, our framework is the first
137
+ to provide optimal folding trajectories with complete generality.
138
+ We release supplementary videos as well as all source code
139
+ and CAD files as open source at https://github.com/
140
+ StructuresComp/deep-robotic-paper-folding.
141
+ B. Overview
142
+ The remainder of the article is organized as follows: We
143
+ begin with a review of related work in Sec. II. A brief
144
+ description of the folding problem is presented in Sec. III.
145
+ The formulation of a reduced-order physics-based model
146
+ is discussed in Sec. IV, where we formulate the folding
147
+ problem using scaling analysis. In Sec. V, we formulate our
148
+ learning framework as well as algorithms for optimal path
149
+ planning. Next, in Sec. VI, we introduce our robotic system
150
+ as well as formulate our closed-loop visual feedback pipeline.
151
+ Experimental results for a robot case study and analysis of the
152
+ results are given in Sec. VII. Finally, we provide concluding
153
+ remarks and discuss the potential of future research avenues
154
+ in Sec. VIII.
155
+ II. RELATED WORK
156
+ The majority of prior works tackling the folding problem
157
+ can be roughly divided into four categories: mechanical
158
+ design-based solutions, vision-based solutions, learning-based
159
+ solutions, and model-based solutions.
160
+ Mechanical design-based approaches typically involve solv-
161
+ ing the folding problem through highly specialized manipulators
162
+ or end effectors. Early approaches involve specialized punches
163
+ and dies for sheet metal bending [29], More recently, highly
164
+ specialized manipulators for robotic origami folding have
165
+ also been developed [30]. Such methods can reliably produce
166
+ repeatable folding but are often limited to a highly specific
167
+ fold, geometry, and/or material.
168
+ Vision-based approaches involve folding deformable mate-
169
+ rials by generating folding motions purely from visual input.
170
+ These approaches are usually common for folding clothes [14],
171
+ [16], [31] as they are extremely soft, which results in the
172
+ easy predictability of their deformation state given a particular
173
+ action. Such approaches can be effective and rather simple to
174
+ implement, but do not transfer well to paper folding as paper
175
+ possesses a much higher stiffness when compared to fabric
176
+ and will attempt to restore its natural, undeformed state if not
177
+ properly handled.
178
+ Learning-based approaches involve the robot learning how
179
+ to fold through training data. The most popular has been to
180
+ learn control policies from human demonstrations, also known
181
+ as learning from demonstrations (LfD). Prior research has
182
+ demonstrated flattening and folding towels [32], [33]. Teleop
183
+ demonstrations are a popular avenue for training policies
184
+ and have been used to learn how to manipulate deformable
185
+ linear objects (DLOs) [34] as well as folding fabric [35].
186
+ To eliminate the need for expensive human-labeled data,
187
+ researchers have also focused on tackling the sim2real problem
188
+ for robotic folding, where reinforcement learning has been
189
+ used to train robots to fold fabrics and cloths completely from
190
+ simulation [26], [28], [36]. More recently, Zheng et al. [37] used
191
+ reinforcement learning to train a robot to flip pages in a binder
192
+ through tactile feedback. Pure learning-based methods have
193
+ shown promising performance, but only for specific tasks whose
194
+ state distribution matches the training data. Such methods tend
195
+ to generalize quite poorly; e.g., when the material or geometric
196
+ properties change drastically.
197
+ Model-based approaches, where the model can either be
198
+ known or learned, often use model predictive control to
199
+ manipulate the deformable object. They involve learning
200
+ the natural dynamics of deformable objects through random
201
+ perturbations [38]. These models are generally fast, but they can
202
+ be inaccurate when experiencing new states. Known models are
203
+ often formulated to be as physically accurate as possible. They
204
+ can be referred to as physics-based (as opposed to simulated).
205
+ Their physical accuracy allows for the direct application of
206
+ their predictive capabilities in the real world. Examples are
207
+ published for rectangular cloth folding [39], strip folding [40],
208
+ and garment folding [41]. Still, known models are usually
209
+ quite expensive to run and must often face a trade-off between
210
+ accuracy and efficiency.
211
+ Despite the large quantity of prior research focusing on
212
+ 2D deformable object manipulation, the majority of these
213
+ efforts have limited their scope to soft materials such as towels
214
+ and cloth. Such materials are highly compliant and often do
215
+ not exhibit complicated nonlinear deformations, thus allowing
216
+ for solutions lacking physical insight. We instead tackle the
217
+ scenario of folding papers of various stiffnesses with a single
218
+ manipulator. Because of its relatively high bending stiffness
219
+ and slippery surface, paper is significantly more difficult to
220
+ manipulate since large deformations will cause sliding of the
221
+ paper on the substrate. Such an example can be observed in
222
+ Fig. 1(a), where intuitive folding trajectories that may work
223
+ on towels and cloth fail for paper due to undesired sliding.
224
+ However, a few works have attempted to solve the paper fold-
225
+ ing problem. For example, Elbrechter et al. [42] demonstrated
226
+ paper folding using visual tracking and real-time physics-based
227
+ modeling with impressive results, but they required expensive
228
+
229
+ 3
230
+ end effectors (two Shadow Dexterous Hands), one end effector
231
+ to hold the paper down while folding at all times, and the
232
+ paper to have AR tags for visual tracking. Similarly, Namiki et
233
+ al. [43] also achieved paper folding through dynamic motion
234
+ primitives and used physics-based simulations to estimate the
235
+ deformation of the paper sheet, also requiring highly specialized
236
+ manipulators and an end effector to hold the paper down
237
+ while folding. By contrast, our method can fold papers reliably
238
+ without any need for holding down the paper during the folding
239
+ operation and requires only an extremely simple 3D printed
240
+ gripper. Other approaches have also attempted to fold with a
241
+ single manipulator while minimizing sliding [36], [40], but
242
+ these methods focused on fabrics whose ends were taped down
243
+ to the substrate.
244
+ III. PROBLEM STATEMENT
245
+ This article studies a simple but challenging task in robotic
246
+ folding: creating a predefined crease on a sheet of paper
247
+ of typical geometry (e.g., rectangular, diamond, etc.) as is
248
+ illustrated in Fig. 2. Only one end of the paper is manipulated
249
+ while the other end is left free. Thus, extra fixtures are
250
+ unnecessary and the folding task can be completed by a single
251
+ manipulator, which simplifies the workspace, but slippage
252
+ of the paper against the substrate must be mitigated during
253
+ manipulation, which is a challenge.
254
+ The task can be divided into two sub-tasks and three states.
255
+ The first sub-task is manipulating one end of the paper from
256
+ the initial flat state (Fig. 2(a)) to the folding state (Fig. 2(b)),
257
+ with the goal that the manipulated edge or point should overlap
258
+ precisely with the crease target line or point C as shown in
259
+ the figure. With the manipulated edge of the paper at the
260
+ origin, the manipulator moves in the x direction. Since the
261
+ manipulated paper usually has relatively high bending stiffness,
262
+ large nonlinear elastic deformations are induced in the folding
263
+ state. In the second sub-task, the paper must be permanently
264
+ deformed to form the desired crease at C/2, thus achieving
265
+ the final folded state (Fig. 2(c)).
266
+ IV. PHYSICS-BASED MODEL AND ANALYSIS
267
+ We next present the numerical framework for studying the
268
+ underlying physics of the paper folding process. First, we
269
+ analyze the main deformations of the manipulated paper and
270
+ prove that a 2D model is sufficient to learn the behaviors of
271
+ the manipulated paper so long as the sheet is symmetrical.
272
+ Second, we briefly introduce a physically accurate numerical
273
+ model based on prior work in computer graphics [44]. Third,
274
+ we formulate a generalized strategy for paper folding using
275
+ scaling analysis.
276
+ A. Reduced-Order Model Representation
277
+ Paper is a unique deformable object. Unlike cloth, its
278
+ surface is developable [45]; i.e., the surface can bend but not
279
+ stretch. Furthermore, shear deformations are not of particular
280
+ importance as the geometry of the manipulated paper is
281
+ symmetrical. Therefore, the primary nonlinear deformation
282
+ when folding paper in our scenario is bending deformation. We
283
+ Initial state
284
+ Folding state
285
+ Folded state
286
+ Manipulated end
287
+ Manipulated
288
+ node
289
+ Desired crease
290
+ Desired crease
291
+ Free end
292
+ Free node
293
+ Target line
294
+ Target
295
+ node
296
+ Rectangular paper
297
+ Symmetrical paper (square)
298
+ (a)
299
+ (b)
300
+ (c)
301
+ x
302
+ z
303
+ y
304
+ o
305
+ 0.5C
306
+ C
307
+ Fig. 2. Folding sheets of paper. The manipulation process involves (a) the
308
+ initial state, where the paper lies flat on the substrate, followed by (b) the
309
+ folding state, where the manipulated end is moved to the “crease target” line
310
+ C, and finally (c) the folded state, which involves forming the desired crease
311
+ on the paper.
312
+ postulate that the nonlinear behaviors of paper arise primarily
313
+ from a balance of bending and gravitational energies: ϵb ∼ ϵg.
314
+ To further understand the energy balance of the manipulated
315
+ paper, we analyze an arbitrary piece in the paper, as shown in
316
+ Fig. 3(b). The bending energy of this piece can be written as
317
+ ϵb = 1
318
+ 2kbκ2l,
319
+ (1)
320
+ where l is its undeformed length of the piece, κ is its curvature,
321
+ and its bending stiffness is
322
+ kb = 1
323
+ 12Ewh3,
324
+ (2)
325
+ where w is its undeformed width, h is its thickness, and E is
326
+ its Young’s modulus. The gravitational energy of the piece is
327
+ ϵg = ρwhlgH,
328
+ (3)
329
+ where ρ is its volume density and H is its vertical height above
330
+ the rigid substrate.
331
+ From the above equations, we obtain a characteristic length
332
+ called the gravito-bending length, which encapsulates the
333
+ influence of bending and gravity:
334
+ Lgb =
335
+ � Eh2
336
+ 24ρg
337
+ � 1
338
+ 3
339
+
340
+ � h
341
+ κ2
342
+ � 1
343
+ 3
344
+ .
345
+ (4)
346
+ The length is in units of meters, and we can observe that
347
+ it scales proportionally to the ratio of thickness to curvature
348
+ squared, which are the key quantities describing the deformed
349
+ configuration of the manipulated paper. Note that the formula-
350
+ tion of Lgb contains only one geometric parameter, the paper
351
+ thickness h, which means that other geometric quantities (i.e.,
352
+ length l and width w) have no influence on the deformed
353
+ configuration.
354
+
355
+ 4
356
+ g
357
+ (a)
358
+ (b)
359
+ (c)
360
+ (d)
361
+ Mesh S
362
+ H
363
+ Rigid
364
+ substrate
365
+ q0
366
+ q1
367
+ qi-1
368
+ qi-1
369
+ qi
370
+ qi
371
+ ti-1
372
+ ti
373
+ qi+1
374
+ qi+1
375
+ qN
376
+ l
377
+ h
378
+ w
379
+ Fig. 3.
380
+ (a) Schematic of a paper during the folding state. (b) Bending
381
+ deformations of a small piece in the paper. (c) Reduced-order discrete model
382
+ (planer rod) representation of our paper. (d) Notations in the discrete model.
383
+ Additionally, due to the symmetrical geometry of the paper,
384
+ curvature κ should be identical for all regions at the same height
385
+ H. Therefore, we can simply use the centerline of the paper,
386
+ as shown in Fig. 3(a), to express the paper’s configuration. We
387
+ model this centerline as a 2D planar rod since deformations are
388
+ limited to the x, z plane. We implement a discrete-differential-
389
+ geometry (DDG)-based numerical simulation to simulate the 2D
390
+ planar rod. We present the details of this numerical framework
391
+ in the next section.
392
+ B. Discrete Differential Geometry Numerical Model
393
+ Following pioneering work on physics-based modeling and
394
+ simulation of deformable curves, surfaces, and solids [46]–
395
+ [48], the computer graphics community has shown impressive
396
+ results using DDG-based simulation frameworks. For example,
397
+ the Discrete Elastic Rods (DER) [44] framework has shown
398
+ efficient and physically accurate simulation of deformable linear
399
+ objects in various scenarios including knot tying [49], helix
400
+ bifurcations [50], coiling of rods [51], and flagella buckling [52].
401
+ Given this success, we use DER to model the centerline of the
402
+ paper as a 2D planar rod undergoing bending deformations.
403
+ As shown in Fig. 3(c), the discrete model is comprised of
404
+ N + 1 nodes, qi (0 ≤ i ≤ N). Each node, qi, represents two
405
+ degrees of freedom (DOF): position along the x and the z axes.
406
+ This results in a 2N + 2-sized DOF vector representing the
407
+ configuration of the sheet, q = [q0, q1, ..., qN]T , where T is
408
+ the transpose operator. Initially, all the nodes of the paper are
409
+ located in a line along the x-axis in the paper’s undeformed
410
+ state. As the robotic manipulator imposes boundary conditions
411
+ on the end node qN, portions of the paper deform against the
412
+ substrate as shown in Fig. 4(a). We compute the DOFs as a
413
+ function of time q(t) by integrating the equations of motion
414
+ (EOM) at each DOF.
415
+ Before describing the EOM, we first outline the elastic
416
+ energies of the rod as a function of q. Kirchhoff’s rod theory
417
+ tells us that the elastic energies of a rod can be divided into
418
+ stretching Es, bending Eb, and twisting Et energies. First, The
419
+ stretching elastic energy is
420
+ Es = 1
421
+ 2ks
422
+ N−1
423
+
424
+ i=0
425
+
426
+ 1 − ∥qi+1 − qi∥
427
+ ∆l
428
+ �2
429
+ ∆l,
430
+ (5)
431
+ where ks = EA is the stretching stiffness; E is Young’s
432
+ modulus; A = wh is the cross-sectional area, and ∆l is the
433
+ undeformed length of each edge (segment between two nodes).
434
+ The bending energy is
435
+ Eb = 1
436
+ 2kb
437
+ N−1
438
+
439
+ i=2
440
+
441
+ 2 tan φi
442
+ 2 − 2 tan φ0
443
+ i
444
+ 2
445
+ �2 1
446
+ ∆l,
447
+ (6)
448
+ where kb = Ewh3
449
+ 12
450
+ is the bending stiffness; w and h are the
451
+ width and thickness respectively; φi is the “turning angle” at a
452
+ node as shown in Fig. 3(d), and φ0
453
+ i is the undeformed turning
454
+ angle (0 for paper). Finally, since we limit our system to a 2D
455
+ plane, we can forgo twisting energies entirely. The total elastic
456
+ energy is then simply Eel = Es + Eb.
457
+ Indeed, a ratio ks/kb ∼ w/h2 >> 1 indicates that stretching
458
+ strains will be minimal which matches our intuition as paper is
459
+ usually easy to bend but not stretch. Therefore, the stretching
460
+ energy item in (5) acts as a constraint to prevent obvious
461
+ stretching for the modeled planar rod.
462
+ We can now construct our EOM as a simple force balance
463
+ P(q) ≡ M¨q + ∂Eel
464
+ ∂q − Fext = 0,
465
+ (7)
466
+ where M is the diagonal lumped mass matrix; ˙( ) represents
467
+ derivatives with respect to time; − ∂Eel
468
+ ∂q
469
+ is the elastic force
470
+ vector, and Fext is the external forces applied on the paper.
471
+ Note that (7) can be solved using Newton’s method, allowing
472
+ for full simulation of the 2D planar rod under manipulation.
473
+ C. Generalized Solution and Scaling Analysis
474
+ As mentioned in Sec. III, the core of the folding task is to
475
+ manipulate the end qN to the target position C starting from
476
+ an initially flat state shown in Fig. 4(a). To do so, we analyze
477
+ the physical system in order to achieve a solution capable of
478
+ minimizing sliding during manipulation.
479
+ We first denote several quantities to describe the deformed
480
+ configuration of the paper. Here, we introduce a point qC,
481
+ which is the node that connects the suspended (z > 0) and
482
+ unsuspended regions (z = 0) of the paper. We focus primarily
483
+ on the suspended region as deformations occur solely in this
484
+ region. An origin o is defined for our 2D plane which is located
485
+ at the initial manipulated end qN as shown in Fig. 4(a). For
486
+ the manipulated end, the robot end-effector imposes a position
487
+ qN = (x, z) and an orientation angle α to control the pose of
488
+ the manipulated end as shown in Fig. 4(a). On the connective
489
+ node qC, the tangent is always along the x-director. Here, we
490
+ impose a constraint that the curvature at the manipulated end is
491
+ always zero so that sharp bending deformations are prevented,
492
+ which is crucial to preventing permanent deformations during
493
+
494
+ 5
495
+ (a)
496
+ (b)
497
+ q0
498
+ qC
499
+ qN
500
+ (qN
501
+ ')
502
+ Norm. x coord, x
503
+ x
504
+ z
505
+ s
506
+ ls=4.10
507
+ Norm. z coord. z
508
+ c
509
+ Fig. 4. (a) Side view of a symmetrical paper during folding with coordinate
510
+ frame and relevant notations. (b) Sampled λ forces for a particular ¯ls of 4.10.
511
+ This showcases one of the sampled “partial” force manifolds that we use train
512
+ our neural network on.
513
+ the folding process. With these definitions, we can now modify
514
+ (7) with the following constraints:
515
+ P(q) = 0,
516
+ s.t.
517
+ qN = (x, z),
518
+ dqC
519
+ ds
520
+ = (−1, 0),
521
+ MN = 0,
522
+ ls ≡
523
+ � qN
524
+ qC
525
+ ds = qC · ˆx,
526
+ (8)
527
+ where MN is the external moment applied on the manipulated
528
+ end; s is the arc length of the paper’s centerline, and ls is the
529
+ arc length of the suspended region (from qC to qN).
530
+ We can solve (8) with the numerical framework presented in
531
+ Sec. IV-B resulting in a unique DOF vector q. Note that when
532
+ q is determined, we can then obtain the external forces from
533
+ the substrate along the paper Fsubstrate = Fx + Fz, orientation
534
+ angle α of the manipulated end, and the suspended length
535
+ ls. Recall that through (4), Young’s modulus E, thickness h,
536
+ and density ρ were determined to be the main material and
537
+ geometric properties of the paper. Therefore, we can outline
538
+ the following physical relationship relating all our quantities:
539
+ λ = ∥Fx∥
540
+ ∥Fz∥,
541
+ (λ, α, ls) = f (E, h, ρ, x, z) ,
542
+ (9)
543
+ where f is an unknown relationship. It is then trivial to see
544
+ that to prevent sliding the relationship
545
+ λ ≤ µs
546
+ (10)
547
+ must be satisfied, where µs is the static friction coefficient
548
+ between the paper and the substrate. Therefore, a trajectory
549
+ that minimizes sliding is one that minimizes λ along its path.
550
+ One glaring problem remains in that the relation f must be
551
+ known to generate any sort of trajectory. In the absence of an
552
+ analytical solution, the numerical framework from Sec. IV-B
553
+ can be used to exhaustively find mappings between the inputs
554
+ and outputs of f. However, generating tuples in this fashion
555
+ requires solving the high-dimensional problem in (8). Such a
556
+ method would be horribly inefficient and would make real-time
557
+ operation infeasible. Instead, we opt to obtain an analytical
558
+ approximation of f by fitting a neural network on simulation
559
+ data. Currently, this approach has several shortcomings. For
560
+ one, directly learning f is difficult given that (9) currently
561
+ depends on five parameters as input, resulting in a high
562
+ dimensional relationship. Furthermore, since the formulation
563
+ directly depends on intrinsic parameters of the paper (E, ρ,
564
+ and h), an enormously exhaustive range of simulations must
565
+ be run to gather enough data to accurately learn f.
566
+ To solve all the aforementioned shortcomings, we reduce
567
+ the dimensionality of the problem by applying scaling analysis.
568
+ According to Buckingham π theorem, we construct five
569
+ dimensionless groups: ¯x = x/Lgb; ¯z = z/Lgb; ¯ls = ls/Lgb;
570
+ α, and λ = Ft/Fn, where Lgb is the gravito-bending length
571
+ from (4). This results in a non-dimensionalized formulation of
572
+ (9) which is expressed as
573
+ (λ, α, ¯ls) = F (¯x, ¯z) .
574
+ (11)
575
+ Note that the mapping relationship F is now irrelevant to
576
+ quantities with units, e.g., material and geometric properties
577
+ of the paper. As the dimensionality of our problem has been
578
+ reduced significantly, we can now express λ as a function of
579
+ just two parameters ¯x, ¯z. Therefore, training a neural network
580
+ to model F is now trivial as non-dimensionalized simulation
581
+ data from a single type of paper can be used. Furthermore,
582
+ the low dimensionality of F allows us easily visualize the
583
+ λ landscape along a non-dimensional 2D-plane. In the next
584
+ section, we will now go over the steps to model F.
585
+ V. DEEP LEARNING AND OPTIMIZATION
586
+ A. Data Generation
587
+ In order to learn the force manifold, we solve (8) for several
588
+ sampled (x, z) points. An example of the partial force manifold
589
+ produced from this sampling can be observed for a single
590
+ suspended length in Fig. 4(b). For a specific (x, z) location,
591
+ we apply incremental rotations along the y-axis and find the
592
+ optimal rotation angle α that results in MN = 0 on the
593
+ manipulated end. For a particular configuration (x, z, α), we
594
+ then record the suspended length ls as well as the tangential
595
+ and normal forces experienced on the clamped end. This
596
+ leads to a training dataset D consisting of six element tuples
597
+ (Ft, Fn, α, ls, x, z). We then non-dimensionalize this dataset
598
+ to the form (λ, α, ¯ls, ¯x, ¯z). A total of 95796 training samples
599
+ were used within a normalized suspended length of ¯ls ≤ 6.84,
600
+ which adequately includes the workspace of most papers.
601
+ B. Learning Force and Optimal Grasp Orientation
602
+ We can now train on our dataset D to obtain a generalized
603
+ neural network modeling F:
604
+ (λ, α, ¯ls) = FNN(¯x, ¯z).
605
+ (12)
606
+ To obtain the above function, a simple fully-connected feed-
607
+ forward nonlinear regression network is trained with 4 hidden
608
+ layers, each containing 392 nodes. Aside from the final output
609
+ layer, each layer is followed by a rectified linear unit (ReLU)
610
+ activation. In addition, we preprocess all inputs through the
611
+ standardization
612
+ x′ = x − ¯xD
613
+ σD
614
+ ,
615
+ (13)
616
+
617
+ 3
618
+ 5
619
+ 4
620
+ 2
621
+ 3
622
+ 2
623
+ 1
624
+ 1
625
+ 0
626
+ 3
627
+ 4
628
+ 5
629
+ 6
630
+ 76
631
+ (a)
632
+ (b)
633
+ (d)
634
+ (c)
635
+ Fig. 5. (a) Visualization of the trained neural network’s non-dimensionalized λ force manifold M and (b) α manifold. An extremely low ¯δ discretization is used
636
+ to showcase smoothness. For the force manifold, we observe two distinctive local minima canyons. Note that regions outside the workspace W are physically
637
+ inaccurate but are of no consequence to us as they are ignored. For the α manifold, we observe continuous smooth interpolation all throughout which is key
638
+ for producing feasible trajectories. Both manifolds showcase the used trajectories in the experiments for folding paper in half for Lgb ∈ [0.048, 0.060, 0.132].
639
+ (c) Showcases the three trajectories in (a) and (b) scaled back to real space. These are the actual trajectories used by the robot. (d) Arbitrary trajectories for
640
+ various Lgb with identical start and goal states are shown to highlight the effect of the material property on our control policy.
641
+ where x is the original input, ¯xD is the mean of the dataset
642
+ D, and σD is the standard deviation of D.
643
+ We use an initial 80-20 train-val split on the dataset D with
644
+ a batch size of 128. Mean absolute error (MAE) is used as
645
+ the error. We alternate between stochastic gradient descent
646
+ (SGD) and Adam whenever training stalls. Furthermore, we
647
+ gradually increase the batch size up to 4096 and train on the
648
+ entire dataset once MAE reaches < 0.001. Using this scheme,
649
+ we achieve an MAE of < 0.0005.
650
+ C. Constructing the Neural Force Manifold
651
+ The neural force manifold (i.e. λ outputs of FNN for the
652
+ workspace set) is discretized into a rectangular grid consisting
653
+ of ¯δ × ¯δ blocks, where ¯δ = δ/Lgb. For each of the blocks,
654
+ we obtain and store a single λ value using the midpoint of
655
+ the block. This results in a discretized neural force manifold
656
+ M represented as a m × n matrix. For the purposes of path
657
+ planning, we add two components to our manifold. First, we
658
+ do not allow exploration into any region not belonging to
659
+ our dataset distribution (¯ls > 6.84). We do so by defining a
660
+ workspace W as all (¯x, ¯z) pairs within the concave hull of
661
+ the input portion of the dataset D. Secondly, we also exclude
662
+ regions within a certain ¯ls threshold. This is done as positions
663
+ with small suspended lengths and large α angles may result
664
+ in high curvatures that could cause collision with our gripper
665
+ and/or plastic deformation, both of which we wish to avoid.
666
+ We denote this region as the penalty region Ls. A visualization
667
+ of M with the workspace W and penalty boundary Ls regions
668
+ can be seen in Fig. 5(a). The α values corresponding to the
669
+ manifold are also shown in Fig. 5(b).
670
+ D. Path Planning over the Neural Force Manifold
671
+ Given the discretized manifold M, we can now generate
672
+ optimal trajectories through traditional path planning algorithms.
673
+ We define an optimal trajectory τ ∗ as one that gets to the goal
674
+ state while minimizing the sum of λ:
675
+ τ ∗ = arg min
676
+ τ∈T
677
+ i=L−1
678
+
679
+ i=0
680
+ λi,
681
+ (14)
682
+ where L is the length of the trajectory and T is the set of all
683
+ valid trajectories from the desired start to goal state. We define
684
+ a valid trajectory as one that is contained within the acceptable
685
+ region
686
+ (xi, zi) ∈ W \ Ls ∀ (xi, zi) ∈ τ,
687
+ and whose consecutive states are adjacent grid locations. Given
688
+ the discretization of the NFM, we can treat M as a graph
689
+ whose edge weights consist of λ. Therefore, we can use uniform
690
+ cost search to obtain τ ∗. The pseudocode of the path planning
691
+ algorithm can be seen in Alg. 1.
692
+ VI. ROBOTIC SYSTEM
693
+ A. Dual Manipulator Setup
694
+ For our experiments, we use two Rethink Robotics’ Sawyer
695
+ manipulators as shown in Fig. 7. One arm has an elongated
696
+ gripper designed for folding, while the other arm has a spring
697
+ compliant roller for creasing and an Intel Realsense D435
698
+ camera for vision feedback. The elongated gripper has rubber
699
+ attached to the insides of the fingers for tight gripping.
700
+
701
+ 6
702
+ 6
703
+ Start state
704
+ 15
705
+ Goal state
706
+ 4
707
+ 5
708
+ 5
709
+ 10
710
+ Trajectory, T
711
+ Workspace, W
712
+ Is penalty, Ls
713
+ 4
714
+ 3
715
+ 0
716
+ I23
717
+ I23
718
+ 2
719
+ 2
720
+ 2
721
+ 1
722
+ 1
723
+ 0 +
724
+ 0
725
+ 0
726
+ 0
727
+ 2
728
+ 10
729
+ 6
730
+ 10
731
+ 0
732
+ 8
733
+ 0
734
+ 0.08
735
+ 0.08
736
+ Lg6=0.065
737
+ Lgb=0.082
738
+ Lgb =0.103
739
+ 0.06
740
+ 三0.06
741
+ Lg6=0.118
742
+ 0.129
743
+ 2
744
+ 2
745
+ 0.04
746
+ 0.04
747
+ 0.02
748
+ 0.00
749
+ 0.05
750
+ 0.10
751
+ 0.15
752
+ 0.20
753
+ 0.25
754
+ 0.30
755
+ 0.00
756
+ 0.05
757
+ 0.10
758
+ 0.15
759
+ 0.20
760
+ 0.25
761
+ α [m]
762
+ α[m]7
763
+ Fig. 6. Example of our perception system with a top down view of the folding procedure. (a) Showcases the the intuitive baseline results while (b) showcases
764
+ our open-loop algorithm for Lgb = 0.048 and C = 0.25m. Similar to Fig. 2, the solid green line indicates the desired end effector position while the
765
+ dashed blue line indicates the crease location. We observe that the intuitive baseline has considerable sliding while our open-loop algorithm has near-perfect
766
+ performance for this case.
767
+ 1
768
+ 2
769
+ 3
770
+ 4
771
+ 5
772
+ 6
773
+ Fig. 7. Experimental apparatus: Two robot manipulators, one for folding (1)
774
+ and the other for creasing (3). An elongated gripper (2) is used for grabbing
775
+ the manipulated end of the folding paper. A roller (5) with compliant springs
776
+ (6) is used for forming the crease. An Intel RealSense D435 camera (4) is
777
+ attached to the creasing arm offer vision feedback during the folding procedure.
778
+ All gripper attachments were 3D printed.
779
+ B. Perception System
780
+ For our perception, we take an eye-in-hand approach by
781
+ attaching an Intel Realsense D435 to the roller arm. We do not
782
+ use the depth component of the camera as we align the camera
783
+ to be pointing down along the world z-axis and the distance
784
+ from the camera to the table is known. To detect the pose of
785
+ the paper, we use simple color detection to segment the paper
786
+ and then use Shi-Tomasi corner detection [53] to obtain the
787
+ position of the bottom edge. An example of the top-down view
788
+ as well as detected poses produced by the camera can be seen
789
+ in Fig. 6.
790
+ C. Vision-feedback Control
791
+ Although we minimize λ with our proposed framework,
792
+ sliding could still happen due to a substrate’s low friction
793
+ surface and/or jittering of the robot’s end-effector. Notice that
794
+ the generated optimal trajectory τ ∗ from Sec. V-D assumes
795
+ that the origin o of our coordinate system shown in Fig. 4(a)
796
+ is fixed. We can define the origin as o = q0 − Lˆx where
797
+ Algorithm 1: Uniform Cost Search
798
+ Input: ¯xs, ¯zs, ¯xg, ¯zg, M
799
+ Output: τ ∗
800
+ 1 Func UCS(¯xs, ¯zs, ¯xg, ¯zg, M):
801
+ 2
802
+ W ← valid workspace of M
803
+ 3
804
+ Ls ← ls penalty region
805
+ 4
806
+ h ← initialize min heap priority queue
807
+ 5
808
+ c ← initialize empty list
809
+ 6
810
+ ns ← node with location (¯xs, ¯zs) and cost 0
811
+ 7
812
+ ng ← node with location (¯xg, ¯zg) and cost 0
813
+ 8
814
+ h.push(ns)
815
+ 9
816
+ while len(h) > 0 do
817
+ 10
818
+ ni ← h.pop()
819
+ 11
820
+ if ni == ng then
821
+ 12
822
+ τ ∗ ← path from start to goal
823
+ 13
824
+ break
825
+ 14
826
+ c.append(ni)
827
+ 15
828
+ for (¯xj, ¯zj) ∈ neighbors of ni do
829
+ 16
830
+ if (¯xj, ¯zj) /∈ W \ Ls then
831
+ 17
832
+ continue
833
+ 18
834
+ nj ← node with location (¯xj, ¯zj) and cost
835
+ λj from M
836
+ 19
837
+ if nj ∈ c then
838
+ 20
839
+ continue
840
+ 21
841
+ if nj ∈ h and cost of nj is higher then
842
+ 22
843
+ continue
844
+ 23
845
+ h.push(nj)
846
+ 24
847
+ τ ∗ ← perform trajectory smoothing on τ ∗
848
+ 25
849
+ return τ ∗
850
+ L is the total length of the paper. Any amount of sliding
851
+ indicates that q0 is moving along the x-axis and therefore, the
852
+ origin o also moves an identical amount. When this occurs, our
853
+ position within the manifold during traversal deviates from the
854
+ optimal trajectory. Furthermore, without adaptive replanning,
855
+ the amount of sliding ∆x will directly result in ∆x amount
856
+ of error when creasing. To circumvent this, we introduce a
857
+ vision-feedback approach that mitigates the effects of sliding.
858
+ We perform vision-feedback at N evenly spaced out intervals
859
+
860
+ (a)8
861
+ Fig. 8. An overview of our folding pipeline. The top row showcases offline
862
+ proponents while the bottom row shows online. On the offline side, we use our
863
+ trained neural network to generate the necessary force manifold for planning.
864
+ Then, given an input tuple (xs, zs, xg, zg, Lgb), we generate an end-to-end
865
+ trajectory using uniform cost search. This end-to-end trajectory is then split
866
+ up into partial trajectories that are carried out by the robot. At the conclusion
867
+ of each partial trajectory, we measure paper sliding and replan the next partial
868
+ trajectory to rectify the error.
869
+ of the trajectory τ ∗ as shown in Fig. 8. To do so, we first split
870
+ up τ ∗ into N partial trajectories. Aside from the first partial
871
+ trajectory τ ∗
872
+ 0 , we extract the start and goal states of the other
873
+ 1 ≤ i ≤ N partial trajectories resulting in a sequence of N
874
+ evenly spaced out states S = {(x1, z1, α1), ..., (xN, zN, αN)}
875
+ when accounting for overlaps. After carrying out τ ∗
876
+ 0 , we detect
877
+ the amount of sliding ∆x and incorporate this error by updating
878
+ the start state and non-dimensionalizing as
879
+ ¯xc
880
+ i = xi − ∆x
881
+ Lgb
882
+ .
883
+ We then replan a partial trajectory τ ∗
884
+ i from the updated start
885
+ state (xc
886
+ i, zi) to the next state (xi+1, zi+1) in the sequence and
887
+ carry out this updated trajectory. This is repeated until reaching
888
+ the goal state. By properly accounting for sliding, we ensure
889
+ that the traversal through the NFM is as accurate as possible.
890
+ We note that this scheme allows us obtain corrected partial
891
+ trajectories in near real time once N becomes sufficiently large
892
+ as each partial trajectory’s goal state becomes increasingly
893
+ close to its start state, allowing for uniform cost search to
894
+ conclude rapidly. We direct the reader to the supplementary
895
+ videos mentioned in Sec. I which showcase the speed of the
896
+ feedback loop.
897
+ Rectifying the sliding ∆x is not the only error we must
898
+ address. Recount that we assume an optimal grasp orientation
899
+ α for each position within the manifold. When the origin of
900
+ our NFM moves, our true position does not match the intended
901
+ position, resulting in also an angular error
902
+ αc
903
+ i = FNN(¯xc
904
+ i, ¯zi),
905
+ ∆α = αi − αc
906
+ i.
907
+ Algorithm 2: Closed-loop Control Pseudocode
908
+ Input: (xs, zs), (xg, zg), Lgb, δ, N, FNN
909
+ 1 M ←DiscretizeManifold (FNN, δ)
910
+ 2 ¯xs, ¯zs, ¯xg, ¯zg ← non-dimensionalize with Lgb
911
+ 3 ¯τ ∗ ← UCS (¯xs, ¯zs, ¯xg, ¯zg, M)
912
+ 4 update ¯τ ∗ with αs using FNN
913
+ 5 τ ∗ ← convert ¯τ ∗ to real space with Lgb
914
+ 6 τ ∗
915
+ 0 , ..., τ ∗
916
+ N−1 ← SplitTrajectory (τ ∗, N)
917
+ 7 S ← extract start and goal states
918
+ 8 carry out τ ∗
919
+ 0 on robot
920
+ 9 for (xi, zi, αi) and (xi+1, zi+1, αi+1) ∈ S do
921
+ 10
922
+ ∆x ← detect sliding of paper
923
+ 11
924
+ xc
925
+ i ← xi − ∆x
926
+ 12
927
+ ¯xc
928
+ i, ¯zi, ¯xi+1, ¯zi+1 ← non-dimensionalize with Lgb
929
+ 13
930
+ αc
931
+ i ← FNN(¯xc
932
+ i, ¯zi)
933
+ 14
934
+ ∆α ← αi − αc
935
+ i
936
+ 15
937
+ ¯τ ∗
938
+ i ← UCS (¯xc
939
+ i, ¯zi, ¯xi+1, ¯zi+1, M)
940
+ 16
941
+ L ← len(¯τ ∗
942
+ i )
943
+ 17
944
+ αi ← obtain αs of ¯τ ∗
945
+ i using FNN
946
+ 18
947
+ αc
948
+ i ← αi + ∆α[1, (L − 1)/L, ..., 1/L, 0]T
949
+ 19
950
+ append ¯τ ∗
951
+ i with αc
952
+ i
953
+ 20
954
+ τ ∗
955
+ i ← convert ¯τ ∗ to real space with Lgb
956
+ 21
957
+ carry out τ ∗
958
+ i on robot
959
+ 22 crease paper with roller
960
+ Simply applying a −∆α update to the first point in a partial
961
+ trajectory results in a large rotational jump that only exacerbate
962
+ the sliding issue. Furthermore, we postulate that so long as
963
+ sliding is not extremely large, the incorrect α at the current
964
+ position within the manifold is still fairly optimal. Therefore,
965
+ the ∆α error is incorporated into the trajectory gradually:
966
+ τ ∗
967
+ i = UCS(¯xc
968
+ i, ¯zi, ¯xi+1, ¯zi+1, M),
969
+ αi = FNN(τ ∗
970
+ i ),
971
+ αc
972
+ i = αi + ∆α[1, (L − 1)/L, ..., 1/L, 0]T ,
973
+ where UCS stands for uniform cost search and L is the length
974
+ of the trajectory τ ∗
975
+ i . This gradual correction ensures that we
976
+ minimize sliding while maintaining smoothness of the trajectory.
977
+ The pseudocode for our full closed-loop algorithm can be seen
978
+ in Alg. 2.
979
+ VII. EXPERIMENTS AND ANALYSIS
980
+ A. Measuring the Material Property of Paper
981
+ To use our framework, we must develop a way to accurately
982
+ measure the parameter Lgb for a particular piece of paper.
983
+ As mentioned previously, Lgb encapsulates the influence of
984
+ bending and gravity. With this in mind, we propose a simple
985
+ way to measure the parameter.
986
+ As shown in Fig. 10(a), when one end of the paper is
987
+ fixed, the paper will deform due to the coupling of bending
988
+
989
+ OfMine
990
+ Obtain force manifold from NN
991
+ Compute optimal end-to-end path
992
+ T* = [(Cs, Zs, Qs), .., (Cg, Zg, ag)
993
+ Path Planner (UCS)
994
+ Compute corrected
995
+ Detect paper slippage
996
+ partial trajectory
997
+ Perception
998
+ Red is the true location
999
+ Ac, Aa
1000
+ obtained from vision
1001
+ feedback
1002
+ Transform trajectory to real space
1003
+ Once goal state is
1004
+ Carry out partial trajectory,
1005
+ reached, crease paper
1006
+ then repeat for next step
1007
+ Motion
1008
+ Planner
1009
+ Online9
1010
+ Fig. 9. Experimental results for all folding scenarios. Each column indicates a folding scenario while the the top row (a) showcases the fold length and bottom
1011
+ row (b) showcases the spin error. Boxplot results are shown color coded for the intuitive baseline, open-loop control, and closed-loop control algorithms.
1012
+ Medians are shown as orange lines, means are shown as turquoise circles, and the desired target value is shown as a light blue horizontal line. We note that
1013
+ both our open-loop and closed-loop algorithms have significant improvements over the intuitive baseline as shown by the broken axis in (a). Our algorithms
1014
+ also have significantly less variance.
1015
+ 0
1016
+ 0.2
1017
+ 0.4
1018
+ 0.6
1019
+ 0.8
1020
+ 1
1021
+ 0
1022
+ 5
1023
+ 10
1024
+ 15
1025
+ 20
1026
+ lh
1027
+ L
1028
+ (a)
1029
+ (b)
1030
+ Norm. paper legnth, L
1031
+ Cardboard paper
1032
+ US Letter paper
1033
+ A4 paper
1034
+ Square origami paper
1035
+ Fig. 10. (a) Schematic of a hanging plate. The manipulation edge is fixed
1036
+ horizontally; (b) Relationship between the ratio ϵ = lh/L and normalized
1037
+ total length of the paper ¯L = L/Lgb.
1038
+ and gravitational energy. Therefore, the following mapping
1039
+ relationship exists:
1040
+ ¯L = L(ϵ),
1041
+ ¯L =
1042
+ L
1043
+ Lgb
1044
+ ,
1045
+ ϵ = lh
1046
+ L ,
1047
+ (15)
1048
+ where lh is the vertical distance from the free end to the fixed
1049
+ end and L is the total length of the paper. We can obtain the
1050
+ mapping relationship L(ϵ) using numerical simulations, which
1051
+ is shown in Fig. 10(b). With this mapping known, simple
1052
+ algebra can be performed to obtain Lgb. First, we measure the
1053
+ ratio ϵ = lh/L for a particular paper to obtain its corresponding
1054
+ normalized total length ¯L. Then, the value of Lgb can be
1055
+ calculated simply by Lgb = L/¯L. Once we obtain Lgb, we can
1056
+ now use the non-dimensionlized mapping relationship in (11)
1057
+ to find the optimal path for manipulating the paper.
1058
+ B. Experimental Setup
1059
+ For our experiments, we tested folding on 4 distinct types
1060
+ of paper:
1061
+ 1) A4 paper, Lgb = 0.048m,
1062
+ 2) US Letter paper, Lgb = 0.060m,
1063
+ 3) Cardboard paper (US Letter dimensions), Lgb = 0.132m,
1064
+ 4) Square origami paper, Lgb = 0.043m.
1065
+ For the rectangular papers (1-3), we do two sets of experiments.
1066
+ The first involves folding the papers to an arbitrary crease
1067
+ location (C = 0.25m for A4 and C = 0.20m for US Letter and
1068
+ cardboard), while the second involves folding the papers in half.
1069
+ For the square origami paper, we choose an arbitrary crease
1070
+ location of C = 0.30m. This results in a total of 7 folding
1071
+ scenarios. For each of the scenarios, we conduct experiments
1072
+ using 3 different algorithms (an intuitive baseline, our open-
1073
+ loop approach, and our closed-loop approach). We complete
1074
+ 10 trials for each of these algorithms, resulting in a total of
1075
+ 210 experiments.
1076
+ C. Baseline Algorithm
1077
+ To showcase the benefits of our folding algorithm, we
1078
+ compare our algorithm to an intuitive baseline. We can think
1079
+ of an intuitive baseline algorithm as one that would work if the
1080
+ opposite end of the paper were fixed to the substrate. Naturally,
1081
+ such a trajectory would be one that grabs the edge of the paper
1082
+ and traces the half perimeter of a circle with radius R = C/2:
1083
+ dθ = π/M,
1084
+ τB = {(R cos(idθ), R sin(idθ), idθ) ∀ i ∈ [0, M]},
1085
+ (16)
1086
+ where M is an arbitrary number of points used as the resolution
1087
+ of trajectory. We choose M = 250 for all experiments.
1088
+
1089
+ Rect, Lgb=0.048
1090
+ Rect, Lgb=0.048
1091
+ Rect, Lgb=0.060
1092
+ Rect, Lgb=0.060
1093
+ Rect, Lgb=0.132
1094
+ Rect, Lgb=0.132
1095
+ Diag, Lgb = 0.043
1096
+ (a)
1097
+ C=0.25 (0.13)
1098
+ C=Half (0.1485)
1099
+ C=0.20 (0.105)
1100
+ C=Half (0.14)
1101
+ C=0.20 (0.105)
1102
+ C=Haif (0.14)
1103
+ C=0.30 (0.155)
1104
+ 0
1105
+ 0.130
1106
+ 0.140
1107
+ 0.105
1108
+ 0.14 -
1109
+ 0.148
1110
+ T
1111
+ 0.15
1112
+ 0.104
1113
+ 0
1114
+ 0
1115
+ 0.100
1116
+ 0.146
1117
+ 0.14
1118
+ 0
1119
+
1120
+ 0
1121
+ 0.095
1122
+ ← 0.102
1123
+ 0.135 -
1124
+ 0.12 -
1125
+ 0.13
1126
+ 0.055
1127
+ 0.10
1128
+ T
1129
+ 0.07
1130
+
1131
+ 0.12 -
1132
+ .
1133
+ T
1134
+ 0.06 -
1135
+
1136
+ 0.12
1137
+ 0.050
1138
+ 0.08
1139
+ 0.11
1140
+ 0
1141
+ 0.06 -
1142
+ 0
1143
+ 0.100
1144
+ 0.11
1145
+ 0.05
1146
+ (b)
1147
+ 2 -
1148
+ 2-
1149
+ 5.0
1150
+ 3 -
1151
+ 4-
1152
+ 2 -
1153
+ 2
1154
+ 2.5 -
1155
+ [deg]
1156
+ Q
1157
+ 2
1158
+ 1 -
1159
+ 1 -
1160
+ 0
1161
+ 2 -
1162
+ T
1163
+ 0
1164
+ 0.0
1165
+ T
1166
+ 0-
1167
+ -0
1168
+
1169
+ -0
1170
+ 0:
1171
+ 2.5
1172
+ 0
1173
+ 0
1174
+ 1
1175
+ 0
1176
+ -1
1177
+ Intuitive Baseline
1178
+ Open-loop Control
1179
+ Closed-loop Control10
1180
+ Fig. 11. Isometric views of different folding scenarios. (a1-2) showcases C = Half folding for Lgb = 0.048 paper with the intuitive baseline and our open-loop
1181
+ algorithm, respectively. (b1-2) showcases C = 0.30m diagonal folding for Lgb = 0.043 with the intuitive baseline our closed-loop algorithm, respectively.
1182
+ D. Metrics
1183
+ The metrics used for the experiments were the average fold
1184
+ length and the spin error. The average fold length was calculated
1185
+ by simply taking the average of the left and right side lengths
1186
+ up until the crease. The spin error was calculated as the angle
1187
+ θerr that results in the difference between the left and right
1188
+ side lengths. For square papers, the fold length was defined
1189
+ as the perpendicular length from the tip to the crease and the
1190
+ spin error was the angular deviation from this line to the true
1191
+ diagonal.
1192
+ E. Parameters
1193
+ The neural force manifold M was discretized using a ¯δ
1194
+ corresponding to δ = 2mm depending on the material as we
1195
+ found this discretization to have good compromise between
1196
+ accuracy and computational speed. All rectangular papers used
1197
+ a penalty region Ls defined by ¯ls < 0.958 while the square
1198
+ paper used one defined by ¯ls < 1.137. This discrepancy is
1199
+ due to the fact that the diagonal paper has a smaller yield
1200
+ strength compared to the the rectangular paper, i.e., to prevent
1201
+ extremely high curvatures, a larger suspended length ¯ls range
1202
+ must be avoided.
1203
+ For closed-loop control, we chose to split all trajectories into
1204
+ N = 5 intervals regardless of trajectory length. Furthermore,
1205
+ we use an extremely slick (i.e. low friction) table to showcase
1206
+ the robustness of our method. Using an empirical method, we
1207
+ measured the static coefficient of friction of our papers and the
1208
+ substrate to be approximately µs = 0.12. For comparison, the
1209
+ static coefficient of friction for steel on steel (both lubricated
1210
+ with castor oil) is µs = 0.15.
1211
+ F. Results and Analysis
1212
+ All experimental results can be seen expressed as box plots
1213
+ where we showcase achieved fold lengths and spin errors in
1214
+ Fig. 9(a) and (b), respectively. When observing the achieved
1215
+ fold lengths, we see significant improvement over the baseline
1216
+ for all folding scenarios. Due to the large gap in performance,
1217
+ broken axes are used to properly display the variance of the
1218
+ recorded data. We note that not only do our algorithms achieve
1219
+ significantly better performance on average, the variance of
1220
+ our approaches is also much lower as shown by the decreased
1221
+ y-axis resolution after the axis break. We attribute the high
1222
+ variance of the baseline method due to the increased influence
1223
+ of friction, which can often cause chaotic, unpredictable results.
1224
+ In other words, truly deterministic folding can only be achieved
1225
+ when sliding is nonexistent.
1226
+ For a vast majority of cases, we observe a clear improvement
1227
+ over the open-loop algorithm when incorporating vision-
1228
+ feedback. Intuitively, we observe a trend where the performance
1229
+ gap between our open-loop and closed-loop algorithms grow
1230
+ as the material stiffness increases for rectangular folding. For
1231
+ softer materials (Lgb = 0.048), the open-loop algorithm has
1232
+ near perfect performance as shown when folding a paper in
1233
+ half in Fig. 11(a2). In comparison, Fig. 11(a1) showcases the
1234
+ baseline algorithm failing with significant sliding.
1235
+ The sliding problem is only exacerbated by increasing the
1236
+ stiffness of the material (Lgb = 0.132) where Fig. 12(a)
1237
+ showcases the baseline algorithm failing to fold the cardboard
1238
+ paper in half by a margin almost as long as the paper itself.
1239
+ In comparison, our open-loop algorithm is capable of folding
1240
+ the cardboard with significantly better results albeit with some
1241
+ visual sliding as shown in Fig. 12(b). As the material stiffness
1242
+ increases, the benefits of the incorporated vision-feedback
1243
+ are more clearly seen as we are able to achieve near perfect
1244
+
1245
+ (al)
1246
+ (a2)
1247
+ (b1)
1248
+ (b2)11
1249
+ Fig. 12. Isometric views for folding C = Half with the stiffest paper (Lgb = 0.132). (a) showcases the intuitive baseline, which fails drastically as the
1250
+ stiffness of the paper causes excessive sliding during the folding process. (b) showcases our open-loop algorithm, which has significant improvements over the
1251
+ baseline with minimal sliding. Finally, (c) showcases our closed-loop algorithm, which improves upon our open-loop results and achieves near perfect folding.
1252
+ folding for cardboard in Fig. 12(c). All of our findings for
1253
+ rectangular folding also match the results of our diagonal
1254
+ folding experiment shown in Fig. 11(b1-b2), where closed-
1255
+ loop once again achieves minimal sliding when compared to
1256
+ the baseline. Overall, the matching findings across all of our
1257
+ experiments showcase the robustness of our formulation against
1258
+ material and geometric factors.
1259
+ We observe one oddity for the folding scenario of Lgb =
1260
+ 0.048 and C = Half where the open-loop algorithm outper-
1261
+ formed our closed-loop variant. Still, we wish to point out that
1262
+ this decrease in performance is only on average 1mm, which
1263
+ can easily be attributed to repetitive discretization error caused
1264
+ by N = 5 replanning. In fact, as we use a discretization of
1265
+ δ = 2mm for the manifold, compounding rounding errors can
1266
+ easily cause 1-2mm errors. With this in mind, our closed-loop
1267
+ method achieves an average fold length performance within a
1268
+ 1-2mm tolerance across all experiments.
1269
+ In terms of spin error, we found that softer materials had
1270
+ the greatest error. As the frictional surface of the table is not
1271
+ perfectly even, any amount of sliding will directly result in
1272
+ uneven spin as shown in Fig. 11(a). As the material stiffness
1273
+ increases, the spin errors became more uniform across the
1274
+ methods as the influence of friction is not enough to deform
1275
+ the paper. Still, we can see that our open and closed-loop
1276
+ algorithms had less sliding than the baseline on average.
1277
+ VIII. CONCLUSION
1278
+ We have introduced a novel control strategy capable of
1279
+ robustly folding sheets of paper of varying materials and
1280
+ geometries with only a single manipulator. Our framework
1281
+ incorporates a combination of techniques spanning several
1282
+ disciplines, including physical simulation, machine learning,
1283
+ scaling analysis, and path planning. The effectiveness of
1284
+ our framework was showcased through extensive real world
1285
+ experiments against an intuitive baseline. Furthermore, an
1286
+ efficient near real-time visual-feedback algorithm was imple-
1287
+ mented that further minimizes folding error. With our closed-
1288
+ loop sensorimotor control algorithm successfully accomplished
1289
+ challenging scenarios such as folding stiff cardboard with
1290
+ repeatable accuracy.
1291
+ For future work, we hope to to tackle the difficult problem
1292
+ of creating arbitrary creases along sheets of paper with non-
1293
+ symmetric centerlines. Such non-symmetric papers can no
1294
+ longer be represented as a reduced-order model of a 2D
1295
+ elastic rod, thus requiring a different formulation. Additionally,
1296
+ folding along regions of paper with preexisting creases will
1297
+ also be a crucial step to achieving elegant folding tasks such
1298
+ as robotic origami. Moving forward, we anticipate exploring
1299
+ solutions to such problems that take advantage of generalized
1300
+ problem formulations with data-driven control schemes such
1301
+ as reinforcement learning.
1302
+ We acknowledge financial support from the National Science
1303
+ Foundation under Grant numbers IIS-1925360, CAREER-
1304
+ 2047663, and OAC-2209782.
1305
+ REFERENCES
1306
+ [1] M. C. Gemici and A. Saxena, “Learning haptic representation for
1307
+ manipulating deformable food objects,” in 2014 IEEE/RSJ International
1308
+ Conference on Intelligent Robots and Systems, pp. 638–645, IEEE, 2014.
1309
+ [2] P. Long, W. Khalil, and P. Martinet, “Force/vision control for robotic
1310
+ cutting of soft materials,” in 2014 IEEE/RSJ International Conference
1311
+ on Intelligent Robots and Systems, pp. 4716–4721, IEEE, 2014.
1312
+ [3] H. Kang and J. T. Wen, “Endobot: a robotic assistant in minimally invasive
1313
+ surgeries,” in Proceedings 2001 ICRA. IEEE International Conference on
1314
+ Robotics and Automation (Cat. No. 01CH37164), vol. 2, pp. 2031–2036,
1315
+ IEEE, 2001.
1316
+ [4] N. Haouchine, W. Kuang, S. Cotin, and M. Yip, “Vision-based force feed-
1317
+ back estimation for robot-assisted surgery using instrument-constrained
1318
+ biomechanical three-dimensional maps,” IEEE Robotics and Automation
1319
+ Letters, vol. 3, no. 3, pp. 2160–2165, 2018.
1320
+ [5] I. Leizea, A. Mendizabal, H. Alvarez, I. Aguinaga, D. Borro, and
1321
+ E. Sanchez, “Real-time visual tracking of deformable objects in robot-
1322
+ assisted surgery,” IEEE computer graphics and applications, vol. 37,
1323
+ no. 1, pp. 56–68, 2015.
1324
+ [6] A. Kapusta, Z. Erickson, H. M. Clever, W. Yu, C. K. Liu, G. Turk, and
1325
+ C. C. Kemp, “Personalized collaborative plans for robot-assisted dressing
1326
+ via optimization and simulation,” Autonomous Robots, vol. 43, no. 8,
1327
+ pp. 2183–2207, 2019.
1328
+ [7] A. Clegg, W. Yu, J. Tan, C. K. Liu, and G. Turk, “Learning to dress:
1329
+ Synthesizing human dressing motion via deep reinforcement learning,”
1330
+ ACM Transactions on Graphics (TOG), vol. 37, no. 6, pp. 1–10, 2018.
1331
+
1332
+ (a)
1333
+ (b)
1334
+ (c)12
1335
+ [8] W. Yu, A. Kapusta, J. Tan, C. C. Kemp, G. Turk, and C. K. Liu,
1336
+ “Haptic simulation for robot-assisted dressing,” in 2017 IEEE international
1337
+ conference on robotics and automation (ICRA), pp. 6044–6051, IEEE,
1338
+ 2017.
1339
+ [9] Z. Erickson, H. M. Clever, G. Turk, C. K. Liu, and C. C. Kemp, “Deep
1340
+ haptic model predictive control for robot-assisted dressing,” in 2018 IEEE
1341
+ international conference on robotics and automation (ICRA), pp. 4437–
1342
+ 4444, IEEE, 2018.
1343
+ [10] E. Pignat and S. Calinon, “Learning adaptive dressing assistance from
1344
+ human demonstration,” Robotics and Autonomous Systems, vol. 93, pp. 61–
1345
+ 75, 2017.
1346
+ [11] T. L. Chen, M. Ciocarlie, S. Cousins, P. M. Grice, K. Hawkins, K. Hsiao,
1347
+ C. C. Kemp, C.-H. King, D. A. Lazewatsky, A. E. Leeper, et al., “Robots
1348
+ for humanity: using assistive robotics to empower people with disabilities,”
1349
+ IEEE Robotics & Automation Magazine, vol. 20, no. 1, pp. 30–39, 2013.
1350
+ [12] T. Bhattacharjee, G. Lee, H. Song, and S. S. Srinivasa, “Towards robotic
1351
+ feeding: Role of haptics in fork-based food manipulation,” IEEE Robotics
1352
+ and Automation Letters, vol. 4, no. 2, pp. 1485–1492, 2019.
1353
+ [13] Y. Kita, F. Kanehiro, T. Ueshiba, and N. Kita, “Clothes handling
1354
+ based on recognition by strategic observation,” in 2011 11th IEEE-
1355
+ RAS International Conference on Humanoid Robots, pp. 53–58, IEEE,
1356
+ 2011.
1357
+ [14] A. Doumanoglou, J. Stria, G. Peleka, I. Mariolis, V. Petrik, A. Kargakos,
1358
+ L. Wagner, V. Hlav´aˇc, T.-K. Kim, and S. Malassiotis, “Folding clothes
1359
+ autonomously: A complete pipeline,” IEEE Transactions on Robotics,
1360
+ vol. 32, no. 6, pp. 1461–1478, 2016.
1361
+ [15] M. Cusumano-Towner, A. Singh, S. Miller, J. F. O’Brien, and P. Abbeel,
1362
+ “Bringing clothing into desired configurations with limited perception,”
1363
+ in 2011 IEEE international conference on robotics and automation,
1364
+ pp. 3893–3900, IEEE, 2011.
1365
+ [16] J. Maitin-Shepard, M. Cusumano-Towner, J. Lei, and P. Abbeel, “Cloth
1366
+ grasp point detection based on multiple-view geometric cues with
1367
+ application to robotic towel folding,” in 2010 IEEE International
1368
+ Conference on Robotics and Automation, pp. 2308–2315, IEEE, 2010.
1369
+ [17] L. Twardon and H. Ritter, “Interaction skills for a coat-check robot:
1370
+ Identifying and handling the boundary components of clothes,” in 2015
1371
+ IEEE International Conference on Robotics and Automation (ICRA),
1372
+ pp. 3682–3688, IEEE, 2015.
1373
+ [18] A. Doumanoglou, A. Kargakos, T.-K. Kim, and S. Malassiotis, “Au-
1374
+ tonomous active recognition and unfolding of clothes using random
1375
+ decision forests and probabilistic planning,” in 2014 IEEE international
1376
+ conference on robotics and automation (ICRA), pp. 987–993, IEEE,
1377
+ 2014.
1378
+ [19] J. Schulman, A. Gupta, S. Venkatesan, M. Tayson-Frederick, and
1379
+ P. Abbeel, “A case study of trajectory transfer through non-rigid
1380
+ registration for a simplified suturing scenario,” in 2013 IEEE/RSJ
1381
+ International Conference on Intelligent Robots and Systems, pp. 4111–
1382
+ 4117, IEEE, 2013.
1383
+ [20] W. H. Lui and A. Saxena, “Tangled: Learning to untangle ropes with rgb-
1384
+ d perception,” in 2013 IEEE/RSJ International Conference on Intelligent
1385
+ Robots and Systems, pp. 837–844, IEEE, 2013.
1386
+ [21] W. Wang, D. Berenson, and D. Balkcom, “An online method for tight-
1387
+ tolerance insertion tasks for string and rope,” in 2015 IEEE International
1388
+ Conference on Robotics and Automation (ICRA), pp. 2488–2495, IEEE,
1389
+ 2015.
1390
+ [22] Y. Yamakawa, A. Namiki, and M. Ishikawa, “Simple model and
1391
+ deformation control of a flexible rope using constant, high-speed motion
1392
+ of a robot arm,” in 2012 IEEE International Conference on Robotics
1393
+ and Automation, pp. 2249–2254, IEEE, 2012.
1394
+ [23] A. Nair, D. Chen, P. Agrawal, P. Isola, P. Abbeel, J. Malik, and S. Levine,
1395
+ “Combining self-supervised learning and imitation for vision-based rope
1396
+ manipulation,” in 2017 IEEE international conference on robotics and
1397
+ automation (ICRA), pp. 2146–2153, IEEE, 2017.
1398
+ [24] S. Kudoh, T. Gomi, R. Katano, T. Tomizawa, and T. Suehiro, “In-air
1399
+ knotting of rope by a dual-arm multi-finger robot,” in 2015 IEEE/RSJ
1400
+ International Conference on Intelligent Robots and Systems (IROS),
1401
+ pp. 6202–6207, IEEE, 2015.
1402
+ [25] Y. Yamakawa, A. Namiki, and M. Ishikawa, “Motion planning for
1403
+ dynamic knotting of a flexible rope with a high-speed robot arm,” in 2010
1404
+ IEEE/RSJ International Conference on Intelligent Robots and Systems,
1405
+ pp. 49–54, IEEE, 2010.
1406
+ [26] J. Matas, S. James, and A. J. Davison, “Sim-to-real reinforcement learning
1407
+ for deformable object manipulation,” in Conference on Robot Learning,
1408
+ pp. 734–743, PMLR, 2018.
1409
+ [27] D. McConachie and D. Berenson, “Estimating model utility for de-
1410
+ formable object manipulation using multiarmed bandit methods,” IEEE
1411
+ Transactions on Automation Science and Engineering, vol. 15, no. 3,
1412
+ pp. 967–979, 2018.
1413
+ [28] X. Lin, Y. Wang, J. Olkin, and D. Held, “Softgym: Benchmarking
1414
+ deep reinforcement learning for deformable object manipulation,” arXiv
1415
+ preprint arXiv:2011.07215, 2020.
1416
+ [29] H. K. H. Kim, D. Bourne, S. Gupta, and S. S. Krishnan, “Automated
1417
+ process planning for robotic sheet metal bending operations,” Journal of
1418
+ Manufacturing Systems, vol. 17, pp. 338 – 360, September 1998.
1419
+ [30] D. J. Balkcom and M. T. Mason, “Robotic origami folding,” The
1420
+ International Journal of Robotics Research, vol. 27, no. 5, pp. 613–
1421
+ 627, 2008.
1422
+ [31] S. Miller, J. van den Berg, M. Fritz, T. Darrell, K. Goldberg, and P. Abbeel,
1423
+ “A geometric approach to robotic laundry folding,” The International
1424
+ Journal of Robotics Research, vol. 31, no. 2, pp. 249–267, 2012.
1425
+ [32] A. X. Lee, H. Lu, A. Gupta, S. Levine, and P. Abbeel, “Learning force-
1426
+ based manipulation of deformable objects from multiple demonstrations,”
1427
+ in 2015 IEEE International Conference on Robotics and Automation
1428
+ (ICRA), pp. 177–184, IEEE, 2015.
1429
+ [33] A. X. Lee, A. Gupta, H. Lu, S. Levine, and P. Abbeel, “Learning from
1430
+ multiple demonstrations using trajectory-aware non-rigid registration
1431
+ with applications to deformable object manipulation,” in 2015 IEEE/RSJ
1432
+ International Conference on Intelligent Robots and Systems (IROS),
1433
+ pp. 5265–5272, 2015.
1434
+ [34] M. Rambow, T. Schauß, M. Buss, and S. Hirche, “Autonomous ma-
1435
+ nipulation of deformable objects based on teleoperated demonstrations,”
1436
+ in 2012 IEEE/RSJ International Conference on Intelligent Robots and
1437
+ Systems, pp. 2809–2814, IEEE, 2012.
1438
+ [35] P.-C. Yang, K. Sasaki, K. Suzuki, K. Kase, S. Sugano, and T. Ogata,
1439
+ “Repeatable folding task by humanoid robot worker using deep learning,”
1440
+ IEEE Robotics and Automation Letters, vol. 2, no. 2, pp. 397–403, 2017.
1441
+ [36] V. Petr ˜Ak and V. Kyrki, “Feedback-based fabric strip folding,” in 2019
1442
+ IEEE/RSJ International Conference on Intelligent Robots and Systems
1443
+ (IROS), pp. 773–778, 2019.
1444
+ [37] Y. Zheng, F. F. Veiga, J. Peters, and V. J. Santos, “Autonomous learning
1445
+ of page flipping movements via tactile feedback,” IEEE Transactions on
1446
+ Robotics, 2022.
1447
+ [38] W. Yan, A. Vangipuram, P. Abbeel, and L. Pinto, “Learning predictive
1448
+ representations for deformable objects using contrastive estimation,” arXiv
1449
+ preprint arXiv:2003.05436, 2020.
1450
+ [39] V. Petr´ık, V. Smutn`y, P. Krsek, V. Hlav´aˇc, “Physics-based model of a
1451
+ rectangular garment for robotic folding,” in 2016 IEEE/RSJ International
1452
+ Conference on Intelligent Robots and Systems (IROS), pp. 951–956, 2016.
1453
+ [40] V. Petr´ık, V. Smutn`y, and V. Kyrki, “Static stability of robotic fabric
1454
+ strip folding,” IEEE/ASME Transactions on Mechatronics, vol. 25, no. 5,
1455
+ pp. 2493–2500, 2020.
1456
+ [41] Y. Li, Y. Yue, D. Xu, E. Grinspun, and P. K. Allen, “Folding deformable
1457
+ objects using predictive simulation and trajectory optimization,” in 2015
1458
+ IEEE/RSJ International Conference on Intelligent Robots and Systems
1459
+ (IROS), pp. 6000–6006, 2015.
1460
+ [42] C. Elbrechter, R. Haschke, and H. Ritter, “Folding paper with anthro-
1461
+ pomorphic robot hands using real-time physics-based modeling,” in
1462
+ 2012 12th IEEE-RAS International Conference on Humanoid Robots
1463
+ (Humanoids 2012), pp. 210–215, IEEE, 2012.
1464
+ [43] A. Namiki and S. Yokosawa, “Robotic origami folding with dynamic
1465
+ motion primitives,” in 2015 IEEE/RSJ International Conference on
1466
+ Intelligent Robots and Systems (IROS), pp. 5623–5628, IEEE, 2015.
1467
+ [44] M. Bergou, M. Wardetzky, S. Robinson, B. Audoly, and E. Grinspun,
1468
+ “Discrete elastic rods,” in ACM SIGGRAPH 2008 Papers, SIGGRAPH
1469
+ ’08, (New York, NY, USA), Association for Computing Machinery, 2008.
1470
+ [45] D. Hilbert and S. Cohn-Vossen, Geometry and the Imagination, vol. 87.
1471
+ American Mathematical Soc., 2021.
1472
+ [46] D. Terzopoulos, J. Platt, A. Barr, and K. Fleischer, “Elastically deformable
1473
+ models,” in Proceedings of the 14th Annual Conference on Computer
1474
+ Graphics and Interactive Techniques (ACM SIGGRAPH 87), pp. 205–214,
1475
+ 1987.
1476
+ [47] D. Terzopoulos and K. Fleischer, “Modeling inelastic deformation:
1477
+ Viscolelasticity, plasticity, fracture,” in Proceedings of the 15th Annual
1478
+ Conference on Computer Graphics and Interactive Techniques (ACM
1479
+ SIGGRAPH 88), pp. 269–278, 1988.
1480
+ [48] D. Terzopoulos and K. Fleischer, “Deformable models,” The Visual
1481
+ Computer, vol. 4, no. 6, pp. 306–331, 1988.
1482
+ [49] A. Choi, D. Tong, M. K. Jawed, and J. Joo, “Implicit contact model
1483
+ for discrete elastic rods in knot tying,” Journal of Applied Mechanics,
1484
+ vol. 88, no. 5, 2021.
1485
+ [50] D. Tong, A. Borum, and M. K. Jawed, “Automated stability testing
1486
+ of elastic rods with helical centerlines using a robotic system,” IEEE
1487
+ Robotics and Automation Letters, vol. 7, no. 2, pp. 1126–1133, 2021.
1488
+
1489
+ 13
1490
+ [51] M. K. Jawed, F. Da, J. Joo, E. Grinspun, and P. M. Reis, “Coiling of
1491
+ elastic rods on rigid substrates,” Proceedings of the National Academy
1492
+ of Sciences, vol. 111, no. 41, pp. 14663–14668, 2014.
1493
+ [52] D. Tong, A. Choi, J. Joo, and M. K. Jawed, “A fully implicit method
1494
+ for robust frictional contact handling in elastic rods,” arXiv preprint
1495
+ arXiv:2205.10309, 2022.
1496
+ [53] J. Shi and Tomasi, “Good features to track,” in 1994 Proceedings of IEEE
1497
+ Conference on Computer Vision and Pattern Recognition, pp. 593–600,
1498
+ 1994.
1499
+
D9A0T4oBgHgl3EQfAv_u/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
DNE1T4oBgHgl3EQfqAUl/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d648566438ca63b54f074225fd78225ba671f7ac6223310b0587429860cc91ca
3
+ size 4456493
DNE1T4oBgHgl3EQfqAUl/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:721a1bfbb40bd1b031653baf891ef2223ca88e38d169cb84abc285776f8eca14
3
+ size 132858
ENAyT4oBgHgl3EQfevhN/content/tmp_files/2301.00326v1.pdf.txt ADDED
@@ -0,0 +1,4172 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ YUILLE-POGGIO’S FLOW AND GLOBAL MINIMIZER OF
2
+ POLYNOMIALS THROUGH CONVEXIFICATION BY
3
+ HEAT EVOLUTION
4
+ QIAO WANG
5
+ Abstract. Finding the global minimizer of polynomials is an impor-
6
+ tant topic in almost all fields in applied mathematics, statistics, and
7
+ engineering, such as signal processing, machine learning, and data sci-
8
+ ence, etc. In this paper, we investigate the possibility of the backward-
9
+ differential-flow-like algorithm which starts from the minimum of con-
10
+ vexification version of the polynomial.
11
+ We apply the heat evolution
12
+ convexification approach through Gaussian filtering p(x, t) = p(x) ∗
13
+ gt(x) with variance t > 0, which is actually an accumulation version
14
+ of Steklov’s regularization. This heat equation plays a multiscale anal-
15
+ ysis framework in mathematics, image processing and computer vision.
16
+ We generalize the fingerprint theory which was proposed in the theory
17
+ of computer vision by A.L. Yuille and T. Poggio in 1980s, in particu-
18
+ lar their fingerprint trajectory equation, to characterize the evolution
19
+ of minimizers across the scale (time) t.
20
+ On the other hand, we pro-
21
+ pose the ”seesaw” polynomials p(x|s) by replacing the coefficient of x of
22
+ p(x) with an arbitrary real parameter s, and we find a seesaw differen-
23
+ tial equation to characterize the evolution of global minimizer x∗(s) of
24
+ p(x|s) while varying s. Essentially, both the fingerprints FP2 and FP3
25
+ of p(x), consisting of the zeros of ∂2p(x,t)
26
+ ∂x2
27
+ and ∂3p(x,t)
28
+ ∂x3
29
+ , respectively, are
30
+ independent of seesaw coefficient s, upon which we define the Confine-
31
+ ment Zone and Escape Zone. Meanwhile, varying s will monotonically
32
+ condition the location of global minimizer of p(x|s), and all these loca-
33
+ tion form the Attainable Zone. Based on these concepts, we prove that
34
+ the global minimizer x∗ of p(x) can be inversely evolved from the global
35
+ minimizer of its convexification polynomial p(x, t0) if and only if x∗
36
+ is included in the Escape Zone. In particular, we give detailed analy-
37
+ sis for quartic and six degree polynomials. For quartic polynomial, we
38
+ proved that the Attainable Zone is completely contained in the Escape
39
+ Zone, thus heat evolution approach must converge to global minimizer,
40
+ and we even find a simpler Euler’s algorithm which must converge to
41
+ the global minimizer, without heat evolution. For six and higher degree
42
+ polynomials, we illustrate that the Attainable Zone might intersect with
43
+ Confinement Zone, which leads to the failure of immediate backward
44
+ differential flow like algorithm. In this case, we show that how to attain
45
+ the global minimizer through our seesaw differential equation.
46
+ Date: December 31, 2022.
47
+ 2010 Mathematics Subject Classification. 35Q90,46N10,35Q93,90C26.
48
+ Key words and phrases. convex optimization, non-convex optimization, heat equation,
49
+ maximum principle, multiscale Gaussian filter, computer vision, quartic polynomial.
50
+ 1
51
+ arXiv:2301.00326v1 [math.OC] 1 Jan 2023
52
+
53
+ 2
54
+ QIAO WANG
55
+ 1. Background and Motivations
56
+ Global optimization of real polynomials is an important non-convex op-
57
+ timization problem (cf.
58
+ [1] and references there in), and produces many
59
+ excellent theories in past decades. Among them, N. Z. Shor [2] first trans-
60
+ formed univariate polynomial optimization to convex problem through qua-
61
+ dratic optimization in 1987, which can offer approximate solution to this
62
+ global optimization. After that, N.Z. Shor further studied its relationship
63
+ with Hilbert’s 17th problem [3]. Also in 1987, V. N. Nefnov [4] proposed
64
+ an algorithm by computing the roots of algebraic equation for finding the
65
+ minimizer.
66
+ In 2014, J. Zhu, S. Zhao and G. Liu [6] proposed a backward differential
67
+ flow formulation, comes from Kuhn-Tucker equation of constrained opti-
68
+ mization, to find out the global minimizer of polynomials. They consider
69
+ the problem for sufficient smooth function p(x),
70
+ min p(x)
71
+ s.t. x ∈ D := {x ∈ Rn| ∥x∥ < a}
72
+ (1.1)
73
+ by introducing a set
74
+ G = {ρ > 0| [∇2p(x) + ρI] > 0, ∀x ∈ D},
75
+ (1.2)
76
+ and an initial pair (�ρ, �x) ∈ G × D satisfying
77
+ ∇p(�x) + �ρ�x = 0.
78
+ (1.3)
79
+ Then they proved that the back differential flow �x(ρ), defined near �ρ,
80
+ d�x
81
+ dρ+[∇2p(�x) + ρI]−1�x = 0,
82
+ �x(�ρ) = �x
83
+ (1.4)
84
+ will lead to the solution of (1.1).
85
+ The above work is under the condition that all global minimizers of this
86
+ polynomial occur only in a known ball, thus the unconstrained optimization
87
+ problem may be reduced to a constrained optimization problem. However,
88
+ O. Arikan, R.S. Burachik and C.Y. Kaya [7] pointed out in 2015 that the
89
+ method in [6] may not converge to global minimizer by a counter-example
90
+ of quartic polynomial
91
+ p(x) = x4 − 8x3 − 18x2 + 56x.
92
+ (1.5)
93
+ Furthermore, they [8] proposed a Steklov regularization and trajectory method
94
+ to this optimization for univariate polynomials in 2019.
95
+ Then in 2020,
96
+ R.S. Burachik and C.Y. Kaya [9] generalized it to the multi-variable case.
97
+ In these works, the quartic polynomial optimization plays an interesting
98
+ role as toy examples. In addition, the six degree polynomials may fail to
99
+ attain the global minimizers, which are illustrated by several examples and
100
+ counter-examples in [8].
101
+
102
+ HEAT EVOLUTION
103
+ 3
104
+ Actually, the Steklov regularization [8]
105
+ µ(x, t) = 1
106
+ 2t
107
+ � x+t
108
+ x−t
109
+ f(τ) dτ
110
+ (1.6)
111
+ is a low-pass filter, in the viewpoint of signal processing, since we may write
112
+ µ(x, t) = 1
113
+ 2t
114
+ � x+t
115
+ x−t
116
+ f(τ) dτ = f(x) ∗ 1[−t,t](x)
117
+ (1.7)
118
+ where
119
+ 1[−t,t](x) =
120
+
121
+ 1
122
+ 2t,
123
+ x ∈ [−t, t]
124
+ 0,
125
+ x /∈ [−t, t]
126
+ (1.8)
127
+ from which one may obtain µx(x, t), µxt(x, t) and µxx(x, t) (where the sub-
128
+ script means partial derivative). However, µt(x, t) is not explicitly in this
129
+ regime, since we can merely represent a differential equation
130
+ 2µt + tµtt = tµx.
131
+ (1.9)
132
+ Obviously it brings some inconvenience in analyzing the evolution of local
133
+ minimizers. Therefore, we require an approach which can balance between
134
+ the simple differential equation and filters, as well as offer convexification
135
+ for polynomials. Fortunately, the heat conduct equation
136
+ ∂p
137
+ ∂t = 1
138
+ 2
139
+ ∂2p
140
+ ∂x2
141
+ (1.10)
142
+ with initial condition
143
+ p(x, 0) = p(x)
144
+ (1.11)
145
+ is a nice framework to implement the convexification and the similar op-
146
+ timization algorithm. In addition, the analysis for evolution of all critical
147
+ points becomes more analytically tractable.
148
+ Remark 1. It should be pointed that the initial problem of heat equation
149
+ (1.10) is equivalent to Gaussian filter, which will be explained in Subsection
150
+ 2.1. But on the other hand, the accumulation of Steklov regularization will
151
+ lead to Gaussian distribution, since that
152
+ 1[−t,t] ∗ 1[−t,t] ∗ · · · ∗ 1[−t,t]
153
+
154
+ ��
155
+
156
+ n
157
+ → N(0, 2nt3
158
+ 3
159
+ ),
160
+ (1.12)
161
+ for n large enough. Thus replacing Steklov regularity with heat evolution,
162
+ i.e., the Gaussian filtering, is very natural in this paper.
163
+ Our interest in this paper is to explore the method of optimizing the even
164
+ degree polynomial
165
+ min
166
+ x p(x) = xn +
167
+ n
168
+
169
+ j=1
170
+ cjxn−j.
171
+ (1.13)
172
+
173
+ 4
174
+ QIAO WANG
175
+ Different from Kuhn-Tucker’s equation based backward differential flow in
176
+ [6], we propose in this paper a constructive way, through evolving the poly-
177
+ nomial by heat conduct (Gaussian filtering) to build a backward-differential-
178
+ flow-like algorithm, in which we can even explicitly express the differential
179
+ equation to attain the minimizer. However, the algorithm converges to the
180
+ global minimizer for quartic polynomial, and partially success for higher
181
+ degree polynomials. This phenomena was actually observed in [8] with ex-
182
+ amples for Steklov regularization.
183
+ In this paper, we explain this convexification derived trajectory algorithm,
184
+ i.e., a backward-differential-flow-like algorithm, by building the convexifica-
185
+ tion of heat evolution to polynomials, and in particular, we build the suffi-
186
+ cient and necessary condition (Theorem 8) under which the algorithm must
187
+ attain the global minimizer. Our analysis is based on the Yuille-Poggio’s
188
+ fingerprints theory and their trajectory differential equation in the theory
189
+ of computer vision [12][13] which were built in 1980s. In addition, to attain
190
+ the global minimizer when the previous algorithm fails, we build a new tra-
191
+ jectory differential equation (Theorem 5) which characterizes the minimizer
192
+ moving from the global minimizer of ”Seesaw”1 polynomial p(x|s) to that
193
+ of original polynomial p(x).
194
+ Before ending this introduction, we slightly sketch the motivation in our
195
+ contributions.
196
+ The elegant framework of multiscale Gaussian filter is equivalent to the
197
+ model of heat conduct equation.
198
+ Applying this theory, any even degree
199
+ monic polynomials p(x) will become convex by p(x, t) = gt(x) ∗ p(x) for t
200
+ large enough2, where gt stands for Gaussian filter with variance t. Moreover,
201
+ for quartic polynomial p(x), the global minimizer xmin will continuously
202
+ evolve along t > 0 such that it remains global minimizer xt
203
+ min of p(x, t) at
204
+ each scale t ≥ 0. Therefore, reversely and continuously evolving from any
205
+ global minimizer xt
206
+ min of p(x, t) to xmin of p(x) is guaranteed.
207
+ A natural question is, whether the global minimizer xmin of a higher de-
208
+ gree polynomial also evolves continuously to global minimizer of scaled ver-
209
+ sion p(x, t), like the quartic polynomial case? Unfortunately this extremely
210
+ expected property doesn’t hold in general for polynomials whose degree is
211
+ more than 4. We will illustrate it by a counter-example on 6-degree polyno-
212
+ mial. Furthermore, we give a condition which is both sufficient and necessary
213
+ for the convergence to global minimizer.
214
+ The multi-scale Gaussian filter and equivalent heat conduct equation is
215
+ a standard content in the theory of PDEs, signal processing and so on.
216
+ In particular in the field of computer vision, it brought us many powerful
217
+ 1Here the ”seesaw” polynomial of a polynomial p(x), say p(x) = x6 −2x4 +3x3 +4x2 +
218
+ 5x + 6, is p(x|s) = x6 − 2x4 + 3x3 + 4x2 + sx + 6, in which 5x + 6 is replaced by sx + 6,
219
+ where s ∈ R can be conditioned such that sx performs like a seesaw.
220
+ 2I definitely believe that this very simple fact should have been already established.
221
+ But I have not gotten any references, limited to my scope of reading.
222
+
223
+ HEAT EVOLUTION
224
+ 5
225
+ Notations
226
+ Definition
227
+ Index
228
+ Zt,k
229
+ real zeros of ∂kp(x,t)
230
+ ∂xk
231
+ (3.1)
232
+ µ(x, t)
233
+ Steklov regularity of p(x)
234
+ (1.7)
235
+ p(x, t)
236
+ heat evolution of p(x)
237
+ (2.1)
238
+ FPk(p)
239
+ k-th fingerprints, (k = 2 is
240
+ Yuille-Poggio’s fingerprint)
241
+ (3.4)
242
+ Yuille-Poggio’s
243
+ fingerprint
244
+ trajectory equation
245
+ (3.8)
246
+ FlowY P (p)
247
+ Yuille-Poggio’s flow
248
+ (3.10)
249
+ Q(p) + S(p)
250
+ Quadric and higher plus See-
251
+ saw decomposition
252
+ (3.15)(3.16)(3.17)
253
+ S(p, s)
254
+ seesaw term
255
+ (3.18)
256
+ seesaw differential equation
257
+ (3.25)
258
+ AZ(p)
259
+ attainable zone
260
+ (3.24)
261
+ Ω(p) and Ωc(p)
262
+ confinement zone and escape
263
+ zone
264
+ Definition 5
265
+ Table 1. List of notations and symbols
266
+ theoretical tools since 1950s (cf.
267
+ [10][11]).
268
+ Among them, the fingerprint
269
+ theory proposed in 1980s (cf. [12][13]) plays a kernel role for many years.
270
+ In this paper, we apply the ideas of fingerprint from computer vision, and
271
+ define three fingerprints of scaled polynomials p(x, t) across scale t. The
272
+ first fingerprint FP1 characterizes all the local extremals of p(x, t) for each
273
+ t, and the second one, FP2, characterizes the stationary points of p(x, t)
274
+ at each t, which indicate the domain of convexity of p(x, t) during the time
275
+ evolution. Furthermore, FP3 indicates the evolution of curves in FP2. All
276
+ these powerful fingerprints tools offer us insightful understandings to the
277
+ evolution of both local and global extremals of the polynomials, from which
278
+ we proposed a sufficient and necessary condition for attaining the global
279
+ minimizer by the backward trajectory algorithm.
280
+ For the sake of simplicity, we list all the symbol and notations in this
281
+ paper as below.
282
+ 2. Heat Evolution and Convexification of polynomials
283
+ 2.1. Heat evolution of p(x). Consider the heat conduct equation (1.10)
284
+ with initial condition (1.11), the general solution of (1.10) is
285
+ p(x, t) = p(x) ∗ gt(x),
286
+ (2.1)
287
+ in which gt(x) stands for the Gaussian filter
288
+ gt(x) =
289
+ 1
290
+
291
+ 2πte− x2
292
+ 2t ,
293
+ t ≥ 0.
294
+ (2.2)
295
+ In signal processing and computer vision, this time variable t is also called
296
+ scale (of Gaussian filtering) or artificial time. Notice that any differential
297
+
298
+ 6
299
+ QIAO WANG
300
+ -8
301
+ -6
302
+ -4
303
+ -2
304
+ 0
305
+ 2
306
+ 4
307
+ 6
308
+ 8
309
+ 10
310
+ 12
311
+ -2000
312
+ 0
313
+ 2000
314
+ 4000
315
+ 6000
316
+ 8000
317
+ 10000
318
+ 12000
319
+ 14000
320
+ Figure 1. The heat evolution of quartic polynomial p(x) =
321
+ x4−8x3−18x2+56x illustrated in x ∈ [−8, 12]. See Example
322
+ 3 for more details.
323
+ -4
324
+ -2
325
+ 0
326
+ 2
327
+ 4
328
+ 6
329
+ 8
330
+ -1000
331
+ -500
332
+ 0
333
+ 500
334
+ 1000
335
+ 1500
336
+ 2000
337
+ 2500
338
+ Figure 2. The partial enlarged view of p(x) = x4 − 8x3 −
339
+ 18x2 + 56x. See Example 3 for more details.
340
+ operator D is commutative with convolution operator ∗, i.e.,
341
+ D(f ∗ g) = f ∗ Dg = Df ∗ g.
342
+ (2.3)
343
+ For polynomials p(x, t), the heat equation (1.10) can be enhanced to
344
+ ∂kp
345
+ ∂tk = 1
346
+ 2k
347
+ ∂2kp
348
+ ∂x2k ,
349
+ (k = 1, 2, · · · )
350
+ (2.4)
351
+
352
+ HEAT EVOLUTION
353
+ 7
354
+ by differentiating both sides of (1.10) w.r.t t, since the smoothness is guar-
355
+ anteed. Then performing Taylor’s expansion for p(x, t) about t will yield
356
+ p(x, t) = p(x, 0) + t · ∂p
357
+ ∂t
358
+ ����
359
+ t=0
360
+ + t2
361
+ 2
362
+ ∂2p
363
+ ∂t2
364
+ ����
365
+ t=0
366
+ + · · · .
367
+ (2.5)
368
+ If using the heat equation derived (2.4), we may rewrite (2.5) as
369
+ p(x, t) = p(x, 0) + t
370
+ 2 · ∂2p
371
+ ∂x2
372
+ ����
373
+ t=0
374
+ + t2
375
+ 8
376
+ ∂4p
377
+ ∂x4
378
+ ����
379
+ t=0
380
+ + · · · .
381
+ (2.6)
382
+ The convexification of even degree polynomials by heat evolution is char-
383
+ acterized by following Theorem3.
384
+ Theorem 1. For each even degree monic polynomial p(x), there exists an
385
+ specified T ∗ = T ∗(p) such that the heat convolution p(x, t) is convex w.r.t x
386
+ at any t > T ∗.
387
+ We require the following basic results existing in many standard text-
388
+ books.
389
+ Lemma 1. The Gaussian density gt(x) defined in (2.2) satisfies the follow-
390
+ ing equations:
391
+ (1) The moment formula
392
+ � +∞
393
+ −∞
394
+ xmgt(x) dx =
395
+
396
+ t
397
+ m
398
+ 2 (m − 1)!!,
399
+ (m even)
400
+ 0,
401
+ (m odd)
402
+ (2.7)
403
+ (2) The convolution formula
404
+ xm ∗ gt(x) = xm + m(m − 1)txm−2 + · · · + rm(x, t),
405
+ (2.8)
406
+ where
407
+ rm(x, t) =
408
+
409
+ m!! t
410
+ m
411
+ 2 ,
412
+ (m even)
413
+ (m − 1)!! t
414
+ m−1
415
+ 2 x,
416
+ (m odd).
417
+ (2.9)
418
+ Proof. The equation (2.7) can be verified immediately, from which we have
419
+ xm ∗ gt(x)
420
+ =
421
+
422
+ (x − y)mgt(y) dy
423
+ =
424
+ m
425
+
426
+ k=0
427
+ �m
428
+ k
429
+
430
+ xk(−1)m−k
431
+
432
+ ym−kgt(y) dy
433
+ =
434
+ m
435
+
436
+ k=0
437
+ m−k is even
438
+ �m
439
+ k
440
+
441
+ (m − k)!! t
442
+ m−k
443
+ 2 xk
444
+ =
445
+ xm + m(m − 1)txm−2 + · · · + rm(x, t),
446
+ (2.10)
447
+ where the last term rm(x, t) is presented at (2.9).
448
+
449
+ Now we prove the Theorem 1.
450
+ 3Once again, I believe that this convexity result must be known in some literature.
451
+
452
+ 8
453
+ QIAO WANG
454
+ Proof of Theorem 1. In what follows, the subscription k in Pk(x) and Qk(x)
455
+ stands for the degree of polynomials.
456
+ Let’s consider even degree monic
457
+ polynomial
458
+ P2m(x) = x2m + P2m−1(x).
459
+ (2.11)
460
+ Observing the expansion
461
+ P2m(x) ∗ gt(x) = x2m ∗ gt(x) + P2m−1(x) ∗ gt(x),
462
+ (2.12)
463
+ according to (2.8) and (2.9), we may write
464
+ P2m(x) ∗ gt(x) = P2m(x) + β(x, t),
465
+ (2.13)
466
+ in which
467
+ β(x, t) = (2m)!! tm +
468
+ m−1
469
+
470
+ k=1
471
+ tm−kQ2k(x).
472
+ (2.14)
473
+ Using the heat evolution, we have
474
+ 1
475
+ 2
476
+ ∂2p(x, t)
477
+ ∂x2
478
+ = ∂p(x, t)
479
+ ∂t
480
+ = ∂β
481
+ ∂t .
482
+ (2.15)
483
+ In our case,
484
+ ∂β
485
+ ∂t = m(2m)!! tm−1 +
486
+ m−1
487
+
488
+ k=1
489
+ (m − k)tm−k−1Q2k(x).
490
+ (2.16)
491
+ Clearly, all these leading terms of Q2n(x) are contributed by x2m(x) ∗
492
+ gt(x) − x2m, and must be positive. In more detail,
493
+ the coefficient of leading term of Q2n(x) =
494
+ �2m
495
+ 2n
496
+
497
+ (2m − 2n)!! > 0 (2.17)
498
+ which implies that there exists bounded constants K, such that
499
+ Q2n(x) > K > −∞,
500
+ (n = 2, 3, · · · , 2m − 2)
501
+ (2.18)
502
+ So that we have
503
+ ∂β
504
+ ∂t > m(2m)!! tm−1 + K(tm−2 + tm−3 + · · · + 1).
505
+ (2.19)
506
+ Therefore, there exists a T ∗ > 0, such that for all t > T ∗, we have ∂β
507
+ ∂t > 0.
508
+ Thus the convexity is guaranteed by heat evolution.
509
+
510
+ 2.2. Comparison principle. The most important mechanism in heat evo-
511
+ lution is the comparison principle, from which we understand that usually
512
+ a local minimizer will merge to a local maximizer during the evolution, like
513
+ the ”annihilation” action between the pair of minimizer and maximizer.
514
+ Theorem 2 (Comparison principle). Assume that x∗ be a critical point of
515
+ p(x, t∗), then for t > t∗, the heat evolution of the critical point satisfies
516
+ p(x∗(t), t) ≥ p(x∗, t∗), if x∗ is local minimum;
517
+ (2.20)
518
+ p(x∗(t), t) ≤ p(x∗, t∗), if x∗ is local maximum.
519
+ (2.21)
520
+
521
+ HEAT EVOLUTION
522
+ 9
523
+ Proof. Without loss of generality, we set t∗ = 0, due to that the heat operator
524
+ U t : f(x) �→ gt(x) ∗ f(x) forms a semi-group (Lie group). Let x = x(t) be
525
+ one of the integral curves of critical points of p(x, t) w.r.t x, then from
526
+ dp(x(t), t)
527
+ dt
528
+ =∂p(x, t)
529
+ ∂x
530
+ ˙x(t) + ∂p(x, t)
531
+ ∂t
532
+ =0 + ∂p(x, t)
533
+ ∂t
534
+ =1
535
+ 2
536
+ ∂2p(x, t)
537
+ ∂x2
538
+ ,
539
+ thus we can get the required result. Notice that the last equality comes from
540
+ heat conduct equation.
541
+
542
+ Remark 2. If we consider the domain (x, t) ∈ [x∗ − ϵ, x∗ + ϵ] × [0, T) near
543
+ each critical point x∗, we can show this result by maximum principle for
544
+ parabolic operator
545
+
546
+ ∂t − 1
547
+ 2
548
+ ∂2
549
+ ∂x2 (cf. [16][17]).
550
+ This comparison principle reveals that the (local) minimizer and (local)
551
+ maximizer might merge pair-wisely during the evolution. Ideally, there ex-
552
+ ists n−1 critical points for a n degree polynomial (here n is even). Thus we
553
+ hope the global minimizer will not merge with any local maximizer during
554
+ the heat evolution. However, it might fail in some cases, and we will analyze
555
+ this mechanism in details.
556
+ 3. Global minimizer and scale space fingerprint
557
+ 3.1. Fingerprints of scale space. The scale space fingerprint was intro-
558
+ duced by A.L. Yuille and T.A. Poggio in 1980s (cf. [12] [13] etc.), which
559
+ plays an important role in computer vision.
560
+ To capture the information of a signal or image p(x), the multi-scale
561
+ version p(x, t) = p(x) ∗ gt(x), which comes from heat conduct equation, is
562
+ applied, in which the variance t ≥ 0 of Gaussian filter is also called artificial
563
+ time.
564
+ Consider that all the polynomials in our situation are of real coefficients,
565
+ for the sake of simplicity, we need to generalize Yuille-Poggio’s definition of
566
+ fingerprints of multi-scale zero-crossings to more general case as below.
567
+ Definition 1. Denote the set of real zeros of k-th derivative of polynomial
568
+ p(x, t) as
569
+ Zt,k(p) :=
570
+
571
+ xi(t) ∈ R;
572
+ ∂kp(xi(t), t)
573
+ ∂xk
574
+ = 0, i = 1, 2, · · · .
575
+
576
+ ,
577
+ (3.1)
578
+ and denote the sets
579
+ FP+
580
+ k (p) :=
581
+
582
+ (x, t); ∂kp(x, t)
583
+ ∂xk
584
+ > 0, t ≥ 0,
585
+
586
+ ,
587
+ FP−
588
+ k (p) :=
589
+
590
+ (x, t); ∂kp(x, t)
591
+ ∂xk
592
+ < 0, t ≥ 0,
593
+
594
+ .
595
+ (3.2)
596
+
597
+ 10
598
+ QIAO WANG
599
+ then the k-th order fingerprints of polynomial p(x) are defined as
600
+ FPk(p) := FP+
601
+ k (p)
602
+
603
+ FP−
604
+ k (p).
605
+ (3.3)
606
+ In above notations S represents the topological closure of set S. In our
607
+ case, this topological closure is very simple thus we may characterize FPk
608
+ by algebraic equations
609
+ FPk(p) =
610
+
611
+ (x, t); ∂kp(x, t)
612
+ ∂xk
613
+ = 0
614
+
615
+ ,
616
+ (3.4)
617
+ due to the sufficient smoothness of all polynomials.
618
+ Remark 3. When k = 2, the fingerprint FP2 of so-called zero-crossings, as
619
+ well as the equation of zero-crossing contour, are introduced by A.L. Yuille
620
+ and T.A. Poggio [13]. Here, we generalize their fingerprints from FP2 to
621
+ more general FPk (k ≥ 2) in this paper. In other words, if we consider
622
+ P(x) whose derivative is P ′(x) = p(x), then FP1(p) = FP2(P). That is to
623
+ say, our framework of FPk is essentially a generalization of Yuille-Poggio’s
624
+ fingerprints in the theory of computer vision.
625
+ According to this notation, FP1 is the fingerprint of extremal values
626
+ (critical points), and FP2 the zero-crossings (convexity)4, of polynomial
627
+ p(x), respectively. Essentially, as in the theory of computer vision, we can
628
+ get more information from FP+
629
+ 2 and FP−
630
+ 2 . In this paper, we generalize the
631
+ classic concept FP1 and FP2 to general FPk, in particular, FP3 is included
632
+ such that our main results can be represented on these three fingerprints.
633
+ We further consider the dynamics of the elements in FP1, i.e., the tra-
634
+ jectories. Our main interest is to obtain the curves x = x(t) which obey the
635
+ equation
636
+ ∂p(x(t), t)
637
+ ∂x
638
+ = 0,
639
+ (3.5)
640
+ as well as initial conditions
641
+ x(0) = xi ∈ Z0,1(p),
642
+ (i = 1, 2, · · · )
643
+ (3.6)
644
+ where xi (i = 1, 2, · · · ) are the critical points of p(x). To solve these curves,
645
+ an ODE by varying t as follows is introduced by A.L. Yuille and T.A. Poggio
646
+ in [13],
647
+ 0 = d
648
+ dt
649
+ �∂p(x(t), t)
650
+ ∂x
651
+
652
+ = ∂2p(x, t)
653
+ ∂x2
654
+ dx(t)
655
+ dt
656
+ + ∂2p(x, t)
657
+ ∂x∂t
658
+ .
659
+ (3.7)
660
+ Therefore, we may characterize the fingerprint which contains all the maxi-
661
+ mums at different t > 0 by rewriting (3.7) as
662
+ dx(t)
663
+ dt
664
+ = −
665
+ ∂2p(x,t)
666
+ ∂x∂t
667
+ ∂2p(x,t)
668
+ ∂x2
669
+ = −
670
+ ∂3p(x,t)
671
+ ∂x3
672
+ 2 · ∂2p(x,t)
673
+ ∂x2
674
+ ,
675
+ (3.8)
676
+ 4Although there exists certain gap between the rigorous meaning and the definition
677
+ here, we omit it in this paper.
678
+
679
+ HEAT EVOLUTION
680
+ 11
681
+ -0.6
682
+ -0.5
683
+ -0.4
684
+ -0.3
685
+ -0.2
686
+ -0.1
687
+ 0
688
+ 0.1
689
+ 0.2
690
+ 0.3
691
+ 0.4
692
+ x
693
+ 0
694
+ 0.01
695
+ 0.02
696
+ 0.03
697
+ 0.04
698
+ 0.05
699
+ 0.06
700
+ t
701
+ (a) FP1
702
+ -0.5
703
+ -0.4
704
+ -0.3
705
+ -0.2
706
+ -0.1
707
+ 0
708
+ 0.1
709
+ 0.2
710
+ 0.3
711
+ x
712
+ 0
713
+ 0.005
714
+ 0.01
715
+ 0.015
716
+ 0.02
717
+ 0.025
718
+ 0.03
719
+ 0.035
720
+ 0.04
721
+ 0.045
722
+ t
723
+ (b) FP2
724
+ -0.3
725
+ -0.2
726
+ -0.1
727
+ 0
728
+ 0.1
729
+ 0.2
730
+ 0.3
731
+ x
732
+ 0
733
+ 0.01
734
+ 0.02
735
+ 0.03
736
+ 0.04
737
+ 0.05
738
+ 0.06
739
+ t
740
+ (c) FP3
741
+ Figure 3. The fingerprints in Example 1, separately illustrated.
742
+
743
+ 12
744
+ QIAO WANG
745
+ Figure 4. The joint illustration of fingerprints FP1, FP2
746
+ and FP3 of previous figures about Example 1.
747
+ as well as suitable initial conditions5
748
+ x(0) ∈ Z0,1(p).
749
+ (3.9)
750
+ In this paper, we call this ODE (3.8) the Yuille-Poggio equation, since it was
751
+ first proposed in (3.3) of A.L. Yuille and T.A. Poggio’s seminal work [13].
752
+ On the other hand, we also call this equation (3.8) the trajectory equa-
753
+ tion, since the reversely evolution algorithm will backward evolute along
754
+ this curve, provided the initial value is given. Given any initial position,
755
+ one may obtain a trajectory by this equation. In particular, when the initial
756
+ condition is located at the critical points of p(x), the trajectories form the
757
+ fingerprint FP1(p).
758
+ We may further generalize FP1(p) to Yuille-Poggio’s flow.
759
+ Definition 2. For any h ∈ R, the integral curve generated by Yuille-Poggio
760
+ equation (3.8) associated with initial value x(0) = h is called a Yuille-
761
+ Poggio’s curve. All these Yuille-Poggio’s curves consist the set
762
+ FlowY P (p) :=
763
+
764
+ (x(t), t);
765
+ dx(t)
766
+ dt
767
+ = −
768
+ ∂2p(x,t)
769
+ ∂x2
770
+ 2∂p3(x,t)
771
+ ∂x3
772
+ ,
773
+ x(0) = h,
774
+ ∀h ∈ R,
775
+
776
+ ,
777
+ (3.10)
778
+ and we call it the Yuille-Poggio’s flow generated by polynomial p(x).
779
+ Clearly, the fingerprint curve in the fingerprint FP1(p) is a special Yuille-
780
+ Poggio’s curve whose initial value x(0) is restricted to Z0,1(p), i.e., satisfies
781
+ p′(x(0)) = 0. Thus we have
782
+ 5If p(x) is n-degree polynomial, there exists at most n − 1 distinct initial conditions.
783
+
784
+ 0.05
785
+ 0.0450.04
786
+ 0.035
787
+ 0.03
788
+
789
+ 0.025
790
+ 0.02
791
+ 0.015
792
+ 0.01
793
+ 0.005
794
+ 0
795
+ -0.6
796
+ -0.4
797
+ -0.2
798
+ 0
799
+ 0.2
800
+ X0.4HEAT EVOLUTION
801
+ 13
802
+ Theorem 3. The fingerprint FP1 can be represented as
803
+ FP1(p) =
804
+
805
+ (x, t);
806
+ dx(t)
807
+ dt
808
+ = −
809
+ ∂2p(x,t)
810
+ ∂x2
811
+ 2 ∂p3(x,t)
812
+ ∂x3
813
+ ,
814
+ x(0) ∈ Z0,1
815
+
816
+ .
817
+ (3.11)
818
+ And
819
+ FP1(p) ⊂ FlowY P (p).
820
+ (3.12)
821
+ Notice that the singularity occurs at which the denominator
822
+ ∂2p(x, t)
823
+ ∂x2
824
+ = 0.
825
+ (3.13)
826
+ Example 1. We illustrate the fingerprints of six degree polynomial
827
+ p(x) = x6 − 0.3726x4 + 0.0574x3 + 0.0306x2 − 0.0084x
828
+ in Fig.3 and Fig.4. We point out that the global minimizer (”*” in Fig.3(a))
829
+ does not evolute to infinity, which means the convex convolution for this p(x)
830
+ will not converge to its global minimizer.
831
+ The differential equation (3.8) characterizes the trajectory of FP1(p), the
832
+ evolution of critical points of p(x) in scale space, which inspires an backward
833
+ differential flow algorithm, which is actually Euler’s algorithm along the
834
+ trajectory described by Yuille-Poggio’s equation. That is, to solve the global
835
+ minimizer of p(x), we first build its convex version p(x, t0) for certain t0 > 0
836
+ large enough.
837
+ According to Theorem 1 at Section 2, this t0 > 0 exists.
838
+ Suppose that x∗(t0) be the global minimizer of convex polynomial p(x, t0),
839
+ we inversely evolute it to x∗(0) according to the trajectory equation (3.8)
840
+ from t = t0 to t = 0. We expect that x∗(0) be the global minimizer of p(x).
841
+ However, this strategy may fail since in some cases the reversely evolution
842
+ might result in a local minimizer of p(x).
843
+ In this paper, we will analyze the mechanism according to Yuille-Poggio’s
844
+ flow and derived zones, and further build a new trajectory differential equa-
845
+ tion to attain the true global minimizer from connected global minimizer of
846
+ its ”Seesaw” polynomial.
847
+ 3.2. Q-S (”Quadric and higher plus Seesaw”) decomposition. As
848
+ we point out, that the heat conduct based backward-differential-flow-like
849
+ algorithm is not guaranteed to converge to theoretically global minimizer.
850
+ This is similar to a six degree polynomial counter-example of Steklov’s regu-
851
+ larization approach in [8]. In this paper, we explain how the convexification
852
+ method converge to global minimizer, and why it may fail in some cases.
853
+ Furthermore, to recover from the failed cases, we propose a ”Quadric plus
854
+ Seesaw” decomposition (Q-S decomposition), then build a new ordinary dif-
855
+ ferential equation that describes the evolution of global minimizer on account
856
+ of varying S(x) according to this Q-S decomposition.
857
+
858
+ 14
859
+ QIAO WANG
860
+ Definition 3 (”Quadric and higher plus Seesaw” decomposition). For any
861
+ polynomial
862
+ p(x) = xn +
863
+ n−1
864
+
865
+ k=0
866
+ ckxk,
867
+ (3.14)
868
+ we define its Q-S decomposition
869
+ p(x) = Q(p) + S(p),
870
+ (3.15)
871
+ in which
872
+ Q(p) = xn +
873
+ n−1
874
+
875
+ k=2
876
+ ckxk
877
+ (3.16)
878
+ stands for the ”Quadric and higher terms”, and
879
+ S(p) = c1x + c0
880
+ (3.17)
881
+ stands for the ”Seesaw terms”. We further define the generalized Seesaw
882
+ term
883
+ S(p, s) = sx + c0,
884
+ s ∈ R.
885
+ (3.18)
886
+ Instead of studying p(x) = Q(p) + S(p), we will consider its ”Seesaw”
887
+ family of polynomials Q(p) + S(p, s). We have
888
+ Lemma 2. Every Seesaw term S(p, s) is invariant under heat evolution,
889
+ i.e.,
890
+ S(p, s) ∗ gt(x) = S(p, s),
891
+ ∀s ∈ R.
892
+ (3.19)
893
+ Proof. Applying Lemma 1 will lead to above result immediately.
894
+
895
+ Actually, this Lemma 2 leads to an insight on multi-scale decomposition
896
+ of p(x) by
897
+ p(x, t) = p(x)∗gt(x) = Q(p)∗gt(x)+S(p)∗gt(x) = Q(P)∗gt(x)+S(p), (3.20)
898
+ upon which we see that the fingerprints FP2 and FP3 of p(x) is essential of
899
+ Q(p) but independent of S(p). Instead, the fingerprint FP1 of p(x) concerns
900
+ both Q(p) and S(p). That is
901
+ Theorem 4. For any polynomial p(x), all of its Seesaw polynomial
902
+ p(x|s) = Q(p) + S(p, s) =
903
+ n
904
+
905
+ k=2
906
+ ckxk + sx + c0
907
+ (3.21)
908
+ satisfy the following equality,
909
+ FPk(p(x|s)) = FPk(p(x)),
910
+ k ≥ 2.
911
+ (3.22)
912
+ Meanwhile,
913
+ FP1(p(x|s)) ̸= FP1(p(x)).
914
+ (3.23)
915
+ For these seesaw polynomials, we define
916
+
917
+ HEAT EVOLUTION
918
+ 15
919
+ Definition 4. For the even degree polynomial p(x), we denote by x∗(s) the
920
+ global minimizer of seesaw polynomials p(x|s) = Q(p) + S(p, s) for each s,
921
+ and call it the seesaw minimizer. For given p(x), the set of global minimizers
922
+ of p(x|s) by varying s ∈ R is called attainable zone given Q(p), i.e.,
923
+ AZ(p) = {x∗ ∈ R; ∃s ∈ R, x∗ is the global minimizer of p(x|s)} .
924
+ (3.24)
925
+ We first focus on those cases that the global minimizer can not be obtained
926
+ from heat evolution from convexificated version p(x, t) of polynomial p(x).
927
+ If case is this, we investigate the global minimizer of Q-S form Q(p)+S(p, s)
928
+ where S(p, s) = sx + c0. Notice that c0 is always a dumb parameter since it
929
+ doesn’t affect the location of the global minimizer.
930
+ Theorem 5. [seesaw differential equation of minimizers moving of seesaw
931
+ polynomials] The global minimizers x∗(s) (and any critical points) of seesaw
932
+ polynomials p(x|s) = Q(p) + S(p, s), i.e., the x∗(s) ∈ AZ(p), must satisfy
933
+ the seesaw differential equation
934
+ dx
935
+ ds = −
936
+ 1
937
+ p′′(x).
938
+ (3.25)
939
+ Proof. For each s ∈ R, the global minimizer of p(x|s) w.r.t. x satisfies
940
+ 0 = p′(x|s) =
941
+
942
+
943
+ n
944
+
945
+ j=2
946
+ cjxj
947
+
948
+
949
+
950
+ + s = 0.
951
+ (3.26)
952
+ Then differentiating both sides w.r.t. s will lead to
953
+ 0 = p′′(x) dx
954
+ ds + 1 = 0,
955
+ (3.27)
956
+ which produces the required result.
957
+
958
+ Corollary 1. The global minimizer x∗(s) of seesaw polynomial p(x|s) is
959
+ monotonically decreasing as s increasing, i.e.,
960
+ s ≥ s′ =⇒ x∗(s) ≤ x∗(s′).
961
+ (3.28)
962
+ Proof. It follows from (3.25) that dx∗(s)
963
+ ds
964
+ < 0 since that p′′(x) > 0 when x∗(s)
965
+ is the global minimizer of p(x∗(s)|s).
966
+
967
+ These results will help us in some situations, may start from the true
968
+ global minimizer of a suitable p(x|s) as initial value, then move it from x∗(s)
969
+ to required location x∗(c1), and finally obtain the true global minimizer of
970
+ p(x|c1).
971
+ The following Theorem explains the ”Seesaw” properties of p(x|s).
972
+ Theorem 6. For any even degree monic polynomial p(x), let x(s) be the
973
+ (global or local) minimizers of seesaw polynomials p(x|s), then they satisfy
974
+ the differential equation
975
+ dp(x(s)|s)
976
+ ds
977
+ = x(s),
978
+ (3.29)
979
+
980
+ 16
981
+ QIAO WANG
982
+ and
983
+ d2p(x(s)|s)
984
+ ds2
985
+ = −
986
+ 1
987
+ p′′(x) < 0.
988
+ (3.30)
989
+ Proof. This differential equation can be verified immediately,
990
+ dp(x(s)|s)
991
+ ds
992
+ = ∂p(x(s)|s)
993
+ ∂x
994
+ · dx(s)
995
+ ds
996
+ + x(s) = x(s).
997
+ (3.31)
998
+ The reminder is a simple application of previous Theorem 5, and the function
999
+ p(x(s)|x) is concave with respect to s.
1000
+
1001
+ Remark 4. The (3.29) doesn’t distinct the global and local minimizers for
1002
+ these x(s).
1003
+ That is, if x(s0) is the global minimizer of seesaw polyno-
1004
+ mial p(x|s0), the connected minimizer x(s1) might be the local minimizer
1005
+ of p(x|s1). Thus we must identify the interval on which x(s) generated from
1006
+ equation (3.29) with initial x(s0) is global or local.
1007
+ 3.3. Confinement zone and escape zone. In our following analysis, we
1008
+ will give basic framework of FP2
1009
+ � FP3, essentially dependent on Q(x), and
1010
+ varying initial condition of trajectory ODE to partition R into Confinement
1011
+ Zone and Escape Zone, as well as varying S(p) to obtain Attainable Zone
1012
+ for given Q(p).
1013
+ It should be stress that in our study, all the fingerprints are about poly-
1014
+ nomials, thus we have some obvious properties.
1015
+ Lemma 3. For any polynomial p(x) and its heat evolution p(x, t), if
1016
+ (x′, t′) ∈ FPi
1017
+
1018
+ FPi+1,
1019
+ (3.32)
1020
+ then x′ must be a real double root of polynomial equation ∂ip(x,t)
1021
+ ∂xi
1022
+ = 0, and a
1023
+ real root of polynomial equation ∂i+1p(x,t)
1024
+ ∂xi+1
1025
+ = 0.
1026
+ Lemma 4. For n-th (n is even) order polynomial p(x), the set FP2
1027
+ � FP3
1028
+ contains at most n
1029
+ 2 − 1 points (xi, ti), where i = 1, 2, · · · , n
1030
+ 2 − 1.
1031
+ Definition 5. Let c ∈ R, if the Yuille-Poggio’s curve from (c, 0) will not
1032
+ intersect with any Yuille-Poggio’s curve from (c′, 0) ̸= (c, 0), we call this c
1033
+ is in Escape Zone. Otherwise, we say it is in the Confinement Zone, which
1034
+ is denoted by Ω. Accordingly, the Escape Zone is denoted by Ωc.
1035
+ Theorem 7 (Characterization of confinement zone and escape zone). The
1036
+ confinement zone Ω is
1037
+ Ω :=
1038
+ n
1039
+ 2 −1
1040
+
1041
+ i=1
1042
+ [XLL
1043
+ i
1044
+ , XRR
1045
+ i
1046
+ ].
1047
+ (3.33)
1048
+ where
1049
+ XLL
1050
+ i
1051
+ =
1052
+ lim
1053
+ L(�xL
1054
+ i ,�t)→(xi,ti)
1055
+ xLL
1056
+ i
1057
+ ,
1058
+ (3.34)
1059
+ XRR
1060
+ i
1061
+ =
1062
+ lim
1063
+ L(�xL
1064
+ i ,�t)→(xi,ti)
1065
+ xRR
1066
+ i
1067
+ ,
1068
+ (3.35)
1069
+
1070
+ HEAT EVOLUTION
1071
+ 17
1072
+ Figure 5. The illustration of Yuille-Poggio’s flow as well as
1073
+ FP2 and FP3.
1074
+
1075
+ 18
1076
+ QIAO WANG
1077
+ in which the limitation means the point (�xL
1078
+ i , �t) (or (�xR
1079
+ i , �t), resp.) moves
1080
+ to the destination (xi, ti) along the local FP2 fingerprint curve fpi
1081
+ 2(L) (or
1082
+ fpi
1083
+ 2(R), resp.). Here (�xL
1084
+ i , �t) (or (�xR
1085
+ i , �t), resp.) is the end of Yuille-Poggio
1086
+ curve connected to (xLL
1087
+ i
1088
+ , 0) (or (xRR
1089
+ i
1090
+ , 0), resp.).
1091
+ Proof. We first prove that
1092
+ Ω :=
1093
+ n
1094
+ 2 −1
1095
+
1096
+ i=1
1097
+
1098
+ [XLL
1099
+ i
1100
+ , XLR
1101
+ i
1102
+ ]
1103
+
1104
+ [XRL
1105
+ i
1106
+ , XRR
1107
+ i
1108
+ ]
1109
+
1110
+ ,
1111
+ (3.36)
1112
+ in which we add two notations,
1113
+ XLR
1114
+ i
1115
+ =
1116
+ lim
1117
+ L(�xL
1118
+ i ,�t)→(xi,ti)
1119
+ xLR
1120
+ i
1121
+ = K,
1122
+ (3.37)
1123
+ XRL
1124
+ i
1125
+ =
1126
+ lim
1127
+ L(�xL
1128
+ i ,�t)→(xi,ti)
1129
+ xRL
1130
+ i
1131
+ = K.
1132
+ (3.38)
1133
+ Here K stands for the intersection point (K, 0) between the curve in FP3(p)
1134
+ and straight line t = 0.
1135
+ Let’s show that the right hand side of (3.36) is well defined. As illustrated
1136
+ at Fig.5, connecting to each (xi, ti) ∈ FP2
1137
+ � FP3, there exists a pair of
1138
+ curves in FP2, corresponding to (xi + 0, ti − 0) and (xi − 0, ti − 0), denoted
1139
+ by fpi
1140
+ 2(R) and fpi
1141
+ 2(L) respectively.
1142
+ For any point (�xR
1143
+ i , �t) ∈ fpi
1144
+ 2(R), when (�xR
1145
+ i , �t) ̸= (xi, ti), there are a pair of
1146
+ trajectories satisfying (3.8) which contains the point (�xR
1147
+ i , �t). We may denote
1148
+ their ends at t = 0 as (xRL
1149
+ i
1150
+ , 0) and (xRR
1151
+ i
1152
+ , 0) respectively. Here we assume
1153
+ that xRL
1154
+ i
1155
+ ≤ xRR
1156
+ i
1157
+ .
1158
+ Similarly, for any point (�xL
1159
+ i , �t) ∈ fpi
1160
+ 2(L), when (�xL
1161
+ i , �t) ̸= (xi, ti), there are
1162
+ a pair of trajectories satisfying (3.8) which contains the point (�xL
1163
+ i , �t). We
1164
+ denote their ends at t = 0 as (xLL
1165
+ i
1166
+ , 0) and (xLR
1167
+ i
1168
+ , 0) respectively. Here we
1169
+ assume that xLL
1170
+ i
1171
+ ≤ xLR
1172
+ i
1173
+ .
1174
+ Now we may write that
1175
+ xLL
1176
+ i
1177
+ ≤ xLR
1178
+ i
1179
+ < K < xRL
1180
+ i
1181
+ ≤ xRR
1182
+ i
1183
+ (3.39)
1184
+ due to that the Yuille-Poggio curve can not intersect with FP3(p) otherwise
1185
+ it will bring singularities, according to the denominator of the right hand
1186
+ side of Yuille-Poggio equation (3.8).
1187
+ Furthermore, the limitation process in (3.34) etc. remains monotonicity.
1188
+ That is, moving from (�xL
1189
+ i , �t) to (�xL+
1190
+ i
1191
+ , �t+) and finally to (xi, ti), we may
1192
+ observe that
1193
+ − ∞ < �xLL+
1194
+ i
1195
+ < �xLL
1196
+ i
1197
+ < �xLR
1198
+ i
1199
+ < �xLR+
1200
+ i
1201
+ < K.
1202
+ (3.40)
1203
+ This implies that all the limitation (3.34) and so on are well-defined. Finally,
1204
+ we note that ∀h ∈ (K − ϵ, K), there must exist a Yuille-Poggio curve starts
1205
+ from (h, 0), and for ∀ϵ > 0, for any point (x′, t′) ∈ fpi
1206
+ 2(L), that satisfy
1207
+ t′ < t, x′ > xi and ∥(x′, t′) − (xi, ti)∥2 < ϵ, there must exist a Yuille-Poggio
1208
+ curve pass the point (x′, t′). That is to say, the Yuille-Poggio curves near
1209
+
1210
+ HEAT EVOLUTION
1211
+ 19
1212
+ the FP3 curve connecting (K, 0) and (xi, ti), are dense. Here for the sake
1213
+ of simplicity, we omit the topology and differential dynamics description.
1214
+ Now the set Ω in (3.33) is well defined. We observe that any Yuille-Poggio
1215
+ curve starting from (h, 0) for h ∈ Ω occurs if and only if there exists another
1216
+ Yuille-Poggio curve, starting from (h′, 0) for some h′ ∈ Ω. In particular,
1217
+ these two curves meet at fpi
1218
+ 2(L) or fpi
1219
+ 2(R). Thus the current Ω in (3.36) is
1220
+ agreed with that in Definition 5.
1221
+
1222
+ Clearly, we further have
1223
+ Theorem 8. Let p(x) be any even order polynomial with positive leading
1224
+ coefficient. Assume that xt0
1225
+ min be the global minimizer of convex polynomial
1226
+ (sufficient scaled version) p(x, t) = p(x) ∗ gt(x) of p(x) at t = t0, and the
1227
+ end of the trajectory by (3.8) at t = 0 is x∗. Then the global minimizer x∗ of
1228
+ p(x) can be inversely involved from the global minimizer of its convexification
1229
+ version p(x, t0), if and only if x∗ is in the Escape Zone Ωc.
1230
+ Proof. If x∗ ∈ Ω, then the maximum of t coordinate of all the corresponding
1231
+ Yuille-Poggio curves is bounded, thus all these Yuille-Poggio curves can not
1232
+ connect to the point in Rx × Rt with large t > 0.
1233
+
1234
+ Remark 5. Although the explicit representation of XLL
1235
+ i
1236
+ , XLR
1237
+ i
1238
+ , XRL
1239
+ i
1240
+ , XRR
1241
+ i
1242
+ is expected, it is not available in algebraic form since that when the degree of
1243
+ polynomial is no less than 6, the curve in FP1 will involve algebraic equation
1244
+ at least 5 degree. Thus we intend to give some numerical methods to give
1245
+ these values.
1246
+ Remark 6. The methodology of analysis declared here for convexification
1247
+ by heat conduct equation, i.e., the Gaussian filtering, still works for the case
1248
+ of Steklov regularization.
1249
+ 4. Case study of Quartic polynomials
1250
+ In what follows, we will get explicit representation for the fingerprints of
1251
+ quartic polynomials, and explain their geometric properties, such that we
1252
+ can build the algorithm for solving the global minimizer of quartic polyno-
1253
+ mials.
1254
+ 4.1. The structure of fingerprints. For the quartic polynomial p(x), we
1255
+ see that
1256
+ p(x, t) = p(x) + (6x2 + 3ax + b) · t + 3t2
1257
+ = x4 + ax3 + (b + 6t)x2 + (c + 3at)x + (d + bt + 3t2).
1258
+ (4.1)
1259
+
1260
+ 20
1261
+ QIAO WANG
1262
+ Continue to differentiate both sides of (4.1) w.r.t x, the information of
1263
+ ∂p(x,t)
1264
+ ∂x
1265
+ across time t may be represented as
1266
+ ∂p(x, t)
1267
+ ∂x
1268
+ = ∂p(x)
1269
+ ∂x
1270
+ + (12x + 3a) · t
1271
+ = (4x3 + 3ax2 + 2bx + c) + (12x + 3a) · t
1272
+ = 4x3 + 3ax2 + (2b + 12t)x + (c + 3at)
1273
+ = 0.
1274
+ (4.2)
1275
+ Similarily, we have
1276
+ ∂2p(x, t)
1277
+ ∂x2
1278
+ = ∂2p(x)
1279
+ ∂x2
1280
+ + 12t = (12x2 + 6ax + 2b) + 12t = 0,
1281
+ (4.3)
1282
+ and
1283
+ ∂3p(x, t)
1284
+ ∂x3
1285
+ = 24x + 6a = 0.
1286
+ (4.4)
1287
+ These form the description of Fingerprints FP1, FP2 and FP3, respectively.
1288
+ 4.1.1. The structure of fingerprint FP1. Based on (4.2), the fingerprint FP1
1289
+ is characterized by following time-varying cubic equation
1290
+ x3 + 3a
1291
+ 4 x2 + b + 6t
1292
+ 2
1293
+ x + c + 3at
1294
+ 4
1295
+ = 0,
1296
+ (4.5)
1297
+ Now, if xi is a real root of (4.5) at t = 0, then it leads to the trajectory
1298
+ described by the differential equation (3.8). For more details, we have
1299
+ Lemma 5. For quartic polynomial p(x), the local extremal values points
1300
+ xt
1301
+ i (i = 1, 2, 3) of p(x, t) w.r.t x at scale t satisfy the trajectory differential
1302
+ equation
1303
+ dx(t)
1304
+ dt
1305
+ = −
1306
+ 12x + 3a
1307
+ 12x2 + 6ax + 2b + 12t,
1308
+ (4.6)
1309
+ with following (at most three) initial conditions,
1310
+ xi(0) = xi,
1311
+ (i = 1, 2, 3).
1312
+ (4.7)
1313
+ Here xi is the local extremal of p(x).
1314
+ Proof. Inserting (4.3) and (4.4) into (3.8) will lead to required results.
1315
+
1316
+ The equation (4.5) possesses (at most) three real roots at t = 0, corre-
1317
+ sponds to (at most) three trajectories, which form the Fingerprint FP1.
1318
+ However, on the viewpoint of differential algebra (see, e.g. [15]), actually
1319
+ the solution of differential equation (4.6) is real algebraic curve, i.e., a poly-
1320
+ nomial F(x, t) about x(t) and t which satisfy F(x, t) = 0.
1321
+ In our case,
1322
+ the polynomial equation (4.5) describes this algebraic curve, thus we may
1323
+ immediately apply the algebraic representation of FP1:
1324
+ FP1 =
1325
+
1326
+ (x, t); t = −4x3 + 3ax2 + 2bx + c
1327
+ 12x + 3a
1328
+ ,
1329
+ x ̸= −a
1330
+ 4, and t ≥ 0
1331
+
1332
+ . (4.8)
1333
+
1334
+ HEAT EVOLUTION
1335
+ 21
1336
+ According to Subsection A.2, to get the information of the roots of (4.5),
1337
+ we need its discriminant,
1338
+ ∆(t) =
1339
+ �a3 − 4ab + 8c
1340
+ 64
1341
+ �2
1342
+ +
1343
+ �−3a2 + 8b
1344
+ 48
1345
+ + t
1346
+ �3
1347
+ ,
1348
+ (4.9)
1349
+ which will be explained in details in (4.12).
1350
+ Lemma 6. The discriminant ∆(t) of equation (4.5) is monotonically in-
1351
+ creasing to infinity. Its unique zero is
1352
+ tu = a2
1353
+ 16 − b
1354
+ 6 − 1
1355
+ 16(a3 − 4ab + 8c)
1356
+ 2
1357
+ 3 .
1358
+ (4.10)
1359
+ Proof. Using (A.5), we write
1360
+ f(t) = b
1361
+ 2 − 3a2
1362
+ 16 + 3t,
1363
+ g(t) = a3
1364
+ 32 − ab
1365
+ 8 + c
1366
+ 4.
1367
+ (4.11)
1368
+ Now the time-varying discriminant
1369
+ ∆(t) =[g(t)]2
1370
+ 4
1371
+ + [f(t)]3
1372
+ 27
1373
+ ,
1374
+ =
1375
+ �a3 − 4ab + 8c
1376
+ 64
1377
+ �2
1378
+ +
1379
+ �−3a2 + 8b
1380
+ 48
1381
+ + t
1382
+ �3
1383
+ ,
1384
+ (4.12)
1385
+ which means that ∆(t) increases monotonically w.r.t. t. Immediately, (4.12)
1386
+ leads to (4.10).
1387
+
1388
+ Theorem 9 (The ”1+2” structure of FP1). Let tu, defined in (4.10), be the
1389
+ zero of discriminant ∆(t). If tu < 0, then FP1 contains only one trajectory
1390
+ x(t) described by equation (4.6), which evolutes as t → +∞. If tu ≥ 0, then
1391
+ during t ∈ [0, tu] the Fingerprint FP1 contains three distinct trajectories
1392
+ described by (4.6), one of which continues to evolute to +∞ during t > tu,
1393
+ and the other two trajectories will start from t = 0 but merge (stop) when
1394
+ t = tu at the point (x(tu), tu). Here,
1395
+ x(tu) =
1396
+ �a3 − 4ab + 8c
1397
+ 64
1398
+ �1/3
1399
+ − a
1400
+ 4.
1401
+ (4.13)
1402
+ Proof. According to Lemma 6, we know that if tu < 0, then ∆(t) > 0 for
1403
+ all t ≥ 0, which means the equation (4.5) has only one root at each t ≥ 0.
1404
+ When tu ≥ 0, then ∆(t) < 0 (= 0, > 0, respectively) while t ∈ [0, tu)
1405
+ (t = tu, t > tu, respectively), and the equation (4.5) has three distinct real
1406
+ roots (one real and a pair of double real roots, or one real root, respectively).
1407
+ In particular, we consider the critical case t = tu at which ∆(t) = 0. If case
1408
+ is this, the equation (4.5) at t = tu possesses one real root and a real double
1409
+
1410
+ 22
1411
+ QIAO WANG
1412
+ root. It follows from (A.8) that the real double root is
1413
+ x(tu) =
1414
+ �g(t)
1415
+ 2
1416
+ � 1
1417
+ 3
1418
+ − 1
1419
+ 3 · 3a
1420
+ 4 .
1421
+ (4.14)
1422
+ Substituting (4.11) into this formula will produces (4.13).
1423
+
1424
+ 4.1.2. The structure of FP2 and FP3. The structure of FP3 is very simple
1425
+ for quartic polynomial, since from (4.4) we may write
1426
+ FP3 =
1427
+
1428
+ (x, t); x = −a
1429
+ 4, t ≥ 0
1430
+
1431
+ .
1432
+ (4.15)
1433
+ To analyze the structure of FP2, we have a Lemma as below.
1434
+ Lemma 7. Denote
1435
+ t∗ = a2
1436
+ 16 − b
1437
+ 6,
1438
+ (4.16)
1439
+ then the polynomial p(x, t) defined in (4.1) is convex about x at each t >
1440
+ max(t∗, 0). Furthermore, this t∗ can not be improved.
1441
+ Proof. Consider the lower bound of (4.3) at t = 0,
1442
+ ∂2p(x)
1443
+ ∂x2
1444
+ = 12x2 + 6ax + 2b
1445
+ = 12
1446
+
1447
+ x + a
1448
+ 4
1449
+ �2
1450
+ − 3a2
1451
+ 4
1452
+ + 2b
1453
+ ≥ −3a2
1454
+ 4
1455
+ + 2b = −12t∗.
1456
+ (4.17)
1457
+ Combining it with (4.3), we would have
1458
+ ∂2p(x, t)
1459
+ ∂x2
1460
+ ≥ −3a2
1461
+ 4
1462
+ + 2b + 12t = 12(t − t∗),
1463
+ (4.18)
1464
+ which implies the required results.
1465
+ On the other hand, at any fixed t′ < t∗, notice that at x = − a
1466
+ 4, we have
1467
+ ∂2p(x, t′)
1468
+ ∂x2
1469
+ = 12x2 + 6ax + 2b + 12t′
1470
+ = 12
1471
+
1472
+ x + a
1473
+ 4
1474
+ �2
1475
+ − 12(t∗ − t′)
1476
+ = −12(t∗ − t′) < 0,
1477
+ (4.19)
1478
+ which is not convex at this x = − a
1479
+ 4, such that t∗ is the optimal, and can not
1480
+ be improved.
1481
+
1482
+ Theorem 10 (The structure of FP2). For the structure of FP2 of quartic
1483
+ polynomial p(x),
1484
+ (a) when t∗ < 0, the fringerprint FP2 is empty.
1485
+
1486
+ HEAT EVOLUTION
1487
+ 23
1488
+ (b) when t∗ = 0, the
1489
+ FP2 =
1490
+
1491
+ (x, t) = (−a
1492
+ 4, 0)
1493
+
1494
+ has only single element;
1495
+ (c) when t∗ > 0, the fingerprint FP2 consists of two curves: the left one
1496
+ is
1497
+ xL(t) = −a
1498
+ 4 −
1499
+
1500
+ t∗ − t,
1501
+ (t∗ ≥ t ≥ 0),
1502
+ (4.20)
1503
+ and the right one is
1504
+ xR(t) = −a
1505
+ 4 +
1506
+
1507
+ t∗ − t,
1508
+ (t∗ ≥ t ≥ 0).
1509
+ (4.21)
1510
+ Specifically, these two curves must meets up at t = t∗, i.e., at the
1511
+ point
1512
+ (xL(t∗), t∗) = (xR(t∗), t∗) =
1513
+
1514
+ −a
1515
+ 4, t∗�
1516
+ .
1517
+ (4.22)
1518
+ Proof. (a) comes from the fact that for every t ≥ 0, all the
1519
+ ∂2p
1520
+ ∂x2 > 0. That
1521
+ is, FP+
1522
+ 2 = {(x, t); x ∈ R, t ∈ [0, +∞)}, but FP−
1523
+ 2 = ∅. (b) is an immediate
1524
+ result, and (c) is from the quadratic equation (4.18).
1525
+
1526
+ 4.1.3. The intersection between fingerprints. According to Lemma 3, we
1527
+ may summarize the intersection of fingerprints.
1528
+ Theorem 11. For the monic quartic polynomials p(x) = x4+ax3+bx2+cx,
1529
+ the two intersection sets
1530
+ FP2
1531
+
1532
+ FP3 =
1533
+
1534
+ (−a
1535
+ 4, t∗)
1536
+
1537
+ ,
1538
+ (4.23)
1539
+ and
1540
+ FP1
1541
+
1542
+ FP2 = {(x(tu), tu)},
1543
+ (4.24)
1544
+ in which t∗ is defined in (4.16), tu and x(tu) are defined in (4.10) and (4.13)
1545
+ respectively.
1546
+ Remark 7 (Three phase of time evolution). In general settings, the evo-
1547
+ lution of polynomial p(x) can be categorized into three phases according to
1548
+ 0 ≤ tu ≤ t∗. At first phase, t evolutes from 0 to tu, and FP1 contains three
1549
+ distinct trajectories. Two of them will merge at t = tu.
1550
+ Then at the second phase, tu < t < t∗, the FP1 contains only one trajec-
1551
+ tory, but p(x, t) is not convex.
1552
+ Finally, at the third phase, t > t∗, the FP1 contains only one trajectory,
1553
+ and p(x, t) is convex.
1554
+ 4.2. Confinement zone. Now we compute the confinement zone of the
1555
+ quartic polynomial p(x). We have
1556
+ Theorem 12. The confinement zone of quartic polynomial p(x) is
1557
+
1558
+ −a
1559
+ 4 −
1560
+
1561
+ 3t∗, −a
1562
+ 4 +
1563
+
1564
+ 3t∗
1565
+
1566
+ ,
1567
+ (4.25)
1568
+ where t∗ is defined in (4.16).
1569
+
1570
+ 24
1571
+ QIAO WANG
1572
+ Proof. Perform Q-S decomposition for quartic polynomial p(x),
1573
+ p(x, t) = Q(x, t) + S(p),
1574
+ (4.26)
1575
+ where S(p) = cx + d. Clearly, we have
1576
+ FPi(p) = FPi(Q),
1577
+ i = 2, 3.
1578
+ (4.27)
1579
+ Thus we may vary W(p), i.e., vary c, to form a pair of trajectories such
1580
+ that they can expand the scope as large as possible in R, which forms the
1581
+ confinement zone. Re-write (4.10) as
1582
+ tu(c) = t∗ − 1
1583
+ 16(a3 − 4ab + 8c)
1584
+ 2
1585
+ 3 .
1586
+ (4.28)
1587
+ We see that we should vary c such that tu(c) = t∗, i.e.,
1588
+ a3 − 4ab + 8c = 0 =⇒ c = ab
1589
+ 2 − a3
1590
+ 8 .
1591
+ (4.29)
1592
+ Substituting this c into the trajectory algebraic curve equation (4.8) and
1593
+ setting t = 0, we get the equation
1594
+ 4x3 + 3ax2 + 2bx + ab
1595
+ 2 − a3
1596
+ 8 = 0.
1597
+ (4.30)
1598
+ The three roots of this equation are
1599
+ x1 = −a
1600
+ 4, x2,3 = −a
1601
+ 4 ±
1602
+
1603
+ 3t∗,
1604
+ (4.31)
1605
+ which produces two pair of trajectories started from t∗ but reversely evolutes
1606
+ to t = 0, whose four destinations form the confinement zone
1607
+
1608
+ −a
1609
+ 4 −
1610
+
1611
+ 3t∗, −a
1612
+ 4
1613
+ � � �
1614
+ −a
1615
+ 4, −a
1616
+ 4 +
1617
+
1618
+ 3t∗
1619
+
1620
+ =
1621
+
1622
+ −a
1623
+ 4 −
1624
+
1625
+ 3t∗, −a
1626
+ 4 +
1627
+
1628
+ 3t∗
1629
+
1630
+ (4.32)
1631
+
1632
+ Remark 8. This confinement zone of p(x) is essentially dependent of Q(x, t)
1633
+ but independent of S(p).
1634
+ 4.3. Differential equation of critical points across scale. Denote p(x)
1635
+ the quartic polynomial as (1.13). Through out this paper, we denote by
1636
+ x1, x2, x3 the roots of cubic equation p′(x) = 0, i.e.,
1637
+ x3 + 3a
1638
+ 4 x2 + b
1639
+ 2x + c
1640
+ 4 = 0.
1641
+ (4.33)
1642
+ Clearly, the global minimizer of p(x) must be one of x1, x2, x3. Comparing
1643
+ this equation to (14), we may represent a, b, c in terms of x1, x2, x3 as
1644
+
1645
+
1646
+
1647
+
1648
+
1649
+
1650
+
1651
+ a = − 4
1652
+ 3(x1 + x2 + x3),
1653
+ b =2(x1x2 + x2x3 + x3x1),
1654
+ c = − 4x1x2x3.
1655
+ (4.34)
1656
+ Now we give the representation of t∗ and tu in terms of roots of ∂p(x,t)
1657
+ ∂x
1658
+ = 0.
1659
+
1660
+ HEAT EVOLUTION
1661
+ 25
1662
+ Lemma 8. Let x1, x2, x3 be the roots of (4.33), t∗ is defined in (4.16), and
1663
+ tu defined in (4.10), then they can be represented as
1664
+ t∗ =
1665
+ �x1 + x2 + x3
1666
+ 3
1667
+ �2
1668
+ − x1x2 + x2x3 + x3x1
1669
+ 3
1670
+ ,
1671
+ (4.35)
1672
+ and
1673
+ tu = t∗ −
1674
+ �32
1675
+ 27(2x1 − x2 − x3)(2x2 − x3 − x1)(2x3 − x1 − x2)
1676
+ � 2
1677
+ 3
1678
+ .
1679
+ (4.36)
1680
+ This can be verified by substituting with (4.34).
1681
+ Theorem 13. The singularity of the equation (4.6) occurs only at
1682
+ xtu = x(tu) =
1683
+ �a3 − 4ab + 8c
1684
+ 64
1685
+ �1/3
1686
+ − a
1687
+ 4,
1688
+ t = tu.
1689
+ (4.37)
1690
+ Proof. The singularity occurs at differential equation (4.6), which describes
1691
+ the FP1, so it must satisfy (4.5). Meanwhile, the denominator of the r.h.s.
1692
+ of (4.6) is actually the fingerprint of FP2, which should satisfy (4.18). Thus
1693
+ we may combine these two algebraic equations to solve (x, t).
1694
+ Multiplying both sides of (4.18) by x + a
1695
+ 4, and subtracted it from (4.5)
1696
+ will produce
1697
+ t = (3a2 − 8b)x − (6c − ab)
1698
+ 48x + 12a
1699
+ ,
1700
+ x ̸= −a
1701
+ 4.
1702
+ (4.38)
1703
+ Substituting it into (4.18) will yield a cubic equation about x,
1704
+ 48x3 + 36ax2 + 9a2x + (3ab − 6c) = 0.
1705
+ (4.39)
1706
+ This cubic equation has only one real solution (4.37). Substituting this x
1707
+ into (4.38) will show that t = tu.
1708
+
1709
+ When − a
1710
+ 4 is not a critical point, this (xtu, tu) occurs only at two FP1
1711
+ integral curves of (4.6) whose initial point is a local minimum and a local
1712
+ maximum, respectively. Among them, one curve corresponds to the case
1713
+ ˙x(tu) = +∞ and another to ˙x(tu) = −∞. Most importantly, the integral
1714
+ curve starts with globally minimum will not pass this (xtu, tu), which is the
1715
+ main discovery of this paper, and will be explained in details in Section 4.4.
1716
+ Remark 9. If there exists a critical point x′ at t = 0 such that x′ = − a
1717
+ 4, then
1718
+ (4.6) implies its fingerprint curve x(t) ≡ − a
1719
+ 4. This happens when x′ is the
1720
+ local maximizer, and other two critical points x1 and x3 satisfy x1+x3 = 2x′.
1721
+ If case is this, all three fingerprint curves meet up at x′ = − a
1722
+ 4 when t = tu,
1723
+ which will be explained in details in the following sections.
1724
+ Example 2. Consider the polynomial p(x) = x4 + 0.2x3 − 0.5x2 + 0.01x,
1725
+ the illustration is in Fig.6.
1726
+
1727
+ 26
1728
+ QIAO WANG
1729
+ -1
1730
+ -0.8
1731
+ -0.6
1732
+ -0.4
1733
+ -0.2
1734
+ 0
1735
+ 0.2
1736
+ 0.4
1737
+ 0.6
1738
+ 0.8
1739
+ 1
1740
+ x
1741
+ 0
1742
+ 0.02
1743
+ 0.04
1744
+ 0.06
1745
+ 0.08
1746
+ 0.1
1747
+ 0.12
1748
+ 0.14
1749
+ 0.16
1750
+ 0.18
1751
+ 0.2
1752
+ t
1753
+ tu
1754
+ t*
1755
+ (a) FPi, (i = 1, 2, 3), tu and t∗. Notice
1756
+ that FP1 corresponds to c = 0.01.
1757
+ -1
1758
+ -0.8
1759
+ -0.6
1760
+ -0.4
1761
+ -0.2
1762
+ 0
1763
+ 0.2
1764
+ 0.4
1765
+ 0.6
1766
+ 0.8
1767
+ 1
1768
+ x
1769
+ 0
1770
+ 0.02
1771
+ 0.04
1772
+ 0.06
1773
+ 0.08
1774
+ 0.1
1775
+ 0.12
1776
+ 0.14
1777
+ 0.16
1778
+ 0.18
1779
+ 0.2
1780
+ t
1781
+ (b) FP2, FP3 and trajectories of c =
1782
+ −0.05.
1783
+ -1
1784
+ -0.8
1785
+ -0.6
1786
+ -0.4
1787
+ -0.2
1788
+ 0
1789
+ 0.2
1790
+ 0.4
1791
+ 0.6
1792
+ 0.8
1793
+ 1
1794
+ x
1795
+ 0
1796
+ 0.02
1797
+ 0.04
1798
+ 0.06
1799
+ 0.08
1800
+ 0.1
1801
+ 0.12
1802
+ 0.14
1803
+ 0.16
1804
+ 0.18
1805
+ 0.2
1806
+ t
1807
+ (c) critical trajectories when c = −0.051,
1808
+ which are symmetric about x = − a
1809
+ 4 .
1810
+ -1
1811
+ -0.8
1812
+ -0.6
1813
+ -0.4
1814
+ -0.2
1815
+ 0
1816
+ 0.2
1817
+ 0.4
1818
+ 0.6
1819
+ 0.8
1820
+ 1
1821
+ x
1822
+ 0
1823
+ 0.02
1824
+ 0.04
1825
+ 0.06
1826
+ 0.08
1827
+ 0.1
1828
+ 0.12
1829
+ 0.14
1830
+ 0.16
1831
+ 0.18
1832
+ 0.2
1833
+ t
1834
+ (d) FP2, FP3 and trajectories of c =
1835
+ −0.2.
1836
+ -1
1837
+ -0.8
1838
+ -0.6
1839
+ -0.4
1840
+ -0.2
1841
+ 0
1842
+ 0.2
1843
+ 0.4
1844
+ 0.6
1845
+ 0.8
1846
+ 1
1847
+ x
1848
+ 0
1849
+ 0.02
1850
+ 0.04
1851
+ 0.06
1852
+ 0.08
1853
+ 0.1
1854
+ 0.12
1855
+ 0.14
1856
+ 0.16
1857
+ 0.18
1858
+ 0.2
1859
+ t
1860
+ (e) trajectories by varying c.
1861
+ (f) more trajectories by varying c.
1862
+ Figure 6. Illustration of Fingerprints and Trajectories in
1863
+ Example 2. To observe the change of trajectories with coef-
1864
+ ficient c in the polynomial, we vary it in c ∈ [−2, 2].
1865
+
1866
+ 0.2
1867
+ 2 < c < 2 tr
1868
+ 0.18
1869
+ FP2
1870
+ FP.jectories0.16
1871
+ 0.14
1872
+ 0.12
1873
+ 0.1
1874
+ 0.08
1875
+ 0.06
1876
+ 0.04
1877
+ T
1878
+ 0.02
1879
+ 0
1880
+ -1
1881
+ -0.5
1882
+ 0
1883
+ 0.5
1884
+ X1HEAT EVOLUTION
1885
+ 27
1886
+ 4.4. Heat evolution of critical points of quartic polynomials. Now
1887
+ we investigate the evolution of the critical points, and in particular, the quart
1888
+ polynomial case. Essentially, we concentrate on the case tu > 0 in which
1889
+ there exist three distinct critical points x1 < x2 < x3, and they correspond-
1890
+ ingly evolve to the critical points xt
1891
+ 1, xt
1892
+ 2, xt
1893
+ 3 when 0 < t < tu. Generally, both
1894
+ x1 and x3 are local minimizers and x2 is local maximizer. Our main con-
1895
+ cern is the behavior associate with heat evolution, characterized by equation
1896
+ (4.6).
1897
+ Our next concern is the comparison principle between two local minimums
1898
+ during heat evolution. Surprisingly, we have the very expected result for heat
1899
+ evolution of quartic polynomials:
1900
+ Theorem 14. For monic quartic polynomial p(x), assume that tu > 0,
1901
+ and denote its three critical points x1 < x2 < x3 (or x1 > x2 > x3). If
1902
+ p(x1) < p(x3), then p(xt
1903
+ 1, t) < p(xt
1904
+ 2, t).
1905
+ Before showing this result, we need several lemmas.
1906
+ Lemma 9. Let x1, x2, x3 be the critical points of quartic polynomial p(x),
1907
+ then we have
1908
+ p(x3) − p(x1) = −(x3 − x1)3 · (x1 + x3 − 2x2)/3.
1909
+ (4.40)
1910
+ Proof. Represent a, b, c in terms of x1, x2, x3, by (4.34). Thus we obtain
1911
+ p(x1) − p(x3). Factorizing it will lead to required result.
1912
+
1913
+ This Lemma 9 implies that
1914
+ Lemma 10. For any t ∈ [0, tu), p(xt
1915
+ 3, t) = p(xt
1916
+ 1, t) if and only if xt
1917
+ 1 + xt
1918
+ 3 =
1919
+ 2xt
1920
+ 2.
1921
+ Consequently, we see that
1922
+ Lemma 11. For any t ∈ [0, tu), xt
1923
+ 1 + xt
1924
+ 3 = 2xt
1925
+ 2 if and only if xt
1926
+ 2 = − a
1927
+ 4.
1928
+ Proof. Notice that the coefficient of x3 in p(x, t) is invariant with t, and the
1929
+ coefficient of x2 of ∂p
1930
+ ∂x is also invariant with t. According to Appendix A, we
1931
+ have
1932
+ xt
1933
+ 1 + xt
1934
+ 2 + xt
1935
+ 3 = −3a
1936
+ 4 .
1937
+ (4.41)
1938
+ Applying Lemma 10 will yield the result.
1939
+
1940
+ Lemma 12. Assume that x1 < x2 < x3 are three critical points of monic
1941
+ quartic polynomial p(x), if p(x1) = p(x3), then for all t ∈ (−∞, tu), we have
1942
+ p(xt
1943
+ 1, t) = p(xt
1944
+ 3, t).
1945
+ (4.42)
1946
+ Proof. Recall (4.5), and apply (4.34), we can actually represent ∂p(x,t)
1947
+ ∂x
1948
+ = 0
1949
+ in terms of x1, x2, x3 as
1950
+ x3 − (x1 + x2 + x3)x2+(x1x2 + x2x3 + x3x1 + 3t)x
1951
+ −[x1x2x3 + (x1 + x2 + x3)t] = 0.
1952
+ (4.43)
1953
+
1954
+ 28
1955
+ QIAO WANG
1956
+ Now if p(x1) = p(x3), Lemma 10 tells us x3 = 2x2−x1, thus we may simplify
1957
+ the above equation as
1958
+ x3 − 3x2x2 + (2x1x2 + 2x2
1959
+ 2 − x2
1960
+ 1 + 3t)x − (2x1x2
1961
+ 2 − x2
1962
+ 1x2 + 3x2t) = 0, (4.44)
1963
+ whose solution is
1964
+
1965
+
1966
+
1967
+
1968
+
1969
+
1970
+
1971
+
1972
+
1973
+
1974
+
1975
+ xt
1976
+ 2 = x2,
1977
+ xt
1978
+ 1 = x2 −
1979
+
1980
+ (x1 − x2)2 − 3t,
1981
+ xt
1982
+ 3 = x2 +
1983
+
1984
+ (x1 − x2)2 − 3t,
1985
+ −∞ < t < min
1986
+ �(x1 − x2)2
1987
+ 3
1988
+ , tu
1989
+
1990
+ (4.45)
1991
+ which shows that xt
1992
+ 1 + xt
1993
+ 3 = 2xt
1994
+ 2. According to Lemma 10, this leads to
1995
+ p(xt
1996
+ 1, t) = p(xt
1997
+ 3, t).
1998
+
1999
+ At present stage, we summarize all lemmas as below,
2000
+ Theorem 15. Under the same assumptions as Theorem 14, and denote
2001
+ xt
2002
+ 1 < xt
2003
+ 2 < xt
2004
+ 3 the critical points of p(x, t). Then the following statements
2005
+ are equivalent:
2006
+ (1) p(x1) = p(x3),
2007
+ (2) p(xt
2008
+ 1, t) = p(xt
2009
+ 3, t), ∀t ∈ [0, tu);
2010
+ (3) x1 + x3 = 2x2;
2011
+ (4) xt
2012
+ 1 + xt
2013
+ 3 = 2xt
2014
+ 2, ∀t ∈ [0, tu);
2015
+ (5) x2 = − a
2016
+ 4;
2017
+ (6) xt
2018
+ 2 = − a
2019
+ 4, ∀t ∈ [0, tu).
2020
+ (7) t∗ = tu.
2021
+ Proof. We will prove that (1) =⇒ (3) =⇒ (5) =⇒ (6) =⇒ (4) =⇒ (2) =⇒
2022
+ (1). In addition, (4) ⇐⇒ (7). Actually, this routine is partially repeated
2023
+ with previous proofs.
2024
+ (1) =⇒ (3) (also (4) =⇒ (2)) comes from Lemma 10, (3) =⇒ (5) (also
2025
+ (6) =⇒ (4)) from Lemma 11, (5) =⇒ (6) from differential equation (4.6).
2026
+ Finally, (4) ⇐⇒ (7) comes from (4.36) in Lemma 8 as well as the condition
2027
+ x1 < x2 < x3.
2028
+
2029
+ Proof of Theorem 10. The dynamical equation (4.6) states that the evo-
2030
+ lution of three critical points are continuous when t ∈ [0, tu).
2031
+ Thus if
2032
+ p(x1) < p(x3), we must have p(xt
2033
+ 1, t) < p(xt
2034
+ 3, t) for t ∈ [0, tu), otherwise,
2035
+ there must have
2036
+ p(xt′
2037
+ 1 , t′) = p(xt′
2038
+ 3 , t′)
2039
+ for some t′ ∈ (0, tu). But, if case is this, Theorem 15 or Lemma 12 tells us
2040
+ that p(x1) = p(x3) since t is reversible. This leads to conflict with assump-
2041
+ tion.
2042
+
2043
+ To intuitively explain this result, we suggest a triangle representation at
2044
+ Fig.20 for each t, where the cortes of triangle consists of (xt
2045
+ i, p(xt
2046
+ i, t)), (i =
2047
+ 1, 2, 3) when t < tu. Notice that the sequence of triangles when 0 ≤ t ≤ tu
2048
+
2049
+ HEAT EVOLUTION
2050
+ 29
2051
+ and continued curve ˙x(t) actually connected to global minimum of p(x, t) at
2052
+ each t ≥ 0.
2053
+ Finally, we discuss an interesting problem: if x1 < x2 < x3 are three
2054
+ critical points of quartic polynomial p(x), can we judge which one of them
2055
+ is global minimizer without valuating all these p(xi)? The answer is YES.
2056
+ Theorem 16. Let x1 < x2 < x3 be three distinct critical points of monic
2057
+ quartic polynomial p(x), then the following statements are equivalent:
2058
+ (1) x3 (resp. x1) is global minimizer;
2059
+ (2) x1 + x3 > 2x2 (resp. x1 + x3 < 2x2);
2060
+ (3) x2 < −a/4 (resp. x2 > −a/4).
2061
+ Proof. Apply Lemma 9.
2062
+
2063
+ This Theorem inspired the following very simple Euler’s algorithm with-
2064
+ out Heat Convolution for quartic polynomials.
2065
+ Theorem 17. For any monic quartic polynomial p(x) with a the coefficient
2066
+ of x3, the Euler’s algorithm with FIXED initial position x(0) = − a
2067
+ 4,
2068
+ x(k+1) = x(k) − ∆x · p′(x(k)),
2069
+ (4.46)
2070
+ MUST converge to the global minimizer of p(x).
2071
+ 4.5. Algorithm. Actually, the Euler’s algorithm may work from t > tu.
2072
+ Fortunately, we know at t ≥ tu, the p(x, t) has only single minimum about
2073
+ x. Recall the formula (A.2), in which we see that the sum of all real roots of
2074
+ fingerprint cubic equation (4.5) should be invariant under 0 ≤ t ≥ tu, thus
2075
+ we know the remaining critical point xinit at t = tu can be solved since we
2076
+ already know the information of (x(tu), tu) from Theorem 13. This means
2077
+ we may adopt
2078
+ xinit = − 3a
2079
+ 4 − 2x(tu)
2080
+ = − a
2081
+ 4 − 2
2082
+ �a3 − 4ab + 8c
2083
+ 64
2084
+ �1/3
2085
+ ,
2086
+ (4.47)
2087
+ at t = tu as initial point, then perform Euler’s algorithm for equation (4.6),
2088
+ and finally attain the global minimum of p(x). This implies the following
2089
+ result.
2090
+ Theorem 18. For the quartic polynomial (1.13), if tu ≤ 0, this polynomial
2091
+ has only one critical point. If tu > 0, the polynomial contains three distinct
2092
+ critical points, then if all the critical points satisfy x ̸= − a
2093
+ 4, we backward
2094
+ perform the differential equation
2095
+ dx(t)
2096
+ dt
2097
+ = −
2098
+ 12x + 3a
2099
+ 12x2 + 6ax + 2b + 12t,
2100
+ (4.48)
2101
+ with initial condition
2102
+ xtu = x(tu) = −a
2103
+ 4 − (a3 − 4ab + 8c)1/3
2104
+ 2
2105
+ (4.49)
2106
+
2107
+ 30
2108
+ QIAO WANG
2109
+ from t = tu > 0 to t = 0, must attain the global minimizer of (1.13) at t = 0.
2110
+ Finally, if one critical point equals − a
2111
+ 4, then p(x) has two global minimizers,
2112
+ and they are the roots of quadratic polynomial
2113
+ p(x)
2114
+ x + a
2115
+ 4
2116
+ .
2117
+ (4.50)
2118
+ So far, we may start with verifying that whether − a
2119
+ 4 is a root of cubic
2120
+ polynomial ∂p(x)
2121
+ ∂x . The global minimizer can be obtained immediately if − a
2122
+ 4
2123
+ is a root. Otherwise, set x(0) = xinit, and t(0) = tu. Then motivated by
2124
+ (4.6), the iteration process is as below
2125
+ x(i+1) =x(i) − ∆t ·
2126
+ 12x(i) + 3a
2127
+ 12x(i)2 + 6ax(i) + 2b + 12t(i) ,
2128
+ t(i+1) =t(i) − ∆t.
2129
+ (4.51)
2130
+ Here the prescribed step ∆t > 0 is small enough, and we may stop the
2131
+ iteration while t(n) ≈ 0. Finally, this algorithm provides
2132
+ lim
2133
+ i→∞ x(i) = xmin.
2134
+ (4.52)
2135
+ Instead beginning with xinit, we can also start by sufficient evolution
2136
+ p(x, t). This will cost more steps of iterations.
2137
+ 4.6. Numerical experiments.
2138
+ Example 3. This counter-example [7] is proposed against ’backward differ-
2139
+ ential flow’ method of [6] , in which p(x) = x4−8x3−18x2+56x. In our heat
2140
+ conduct framework, we have p(x, t) = x4−8x3−(18−6t)x2+32x−(18t−3t2).
2141
+ Notice that Fig. 20 demonstrates the triangle series of critical points.
2142
+ Example 4. Set p(x) = x4 + 0.2114x3 − 2.6841x2 − 0.1110x + 1.2406, then
2143
+ in Fig. 7 the most left curve x1(t) is the global minimizer of corresponding
2144
+ p(x, t) at each t ≥ 0, and Fig.8 illustrates the fingerprint FP1. The the-
2145
+ oretical minimizer is x1 = −1.2307, and our iteration algorithm provides
2146
+ x1 = −1.2308.
2147
+ Example 5. Consider p(x) = x4 − 4x3 − 2x2 + 12x, then we actually have
2148
+ three critical points x1 = −1, x2 = 1, x3 = 3. Notice that x1+x3 = 2x2 thus
2149
+ p(x1, t) = p(x3, t) and x2(t) = x2 = 1 for all x ∈ [0, tu]. One can further
2150
+ verify that in this symmetric case, we must have t∗ = tu.
2151
+ The detailed
2152
+ explain can be referred as in Theorem 15.
2153
+ 4.7. Summary of quartic polynomial case. For the global minimizer of
2154
+ quartic polynomial p(x), while generating its multi-scale version p(x, t) =
2155
+ p(x) ∗ gt(x) on account of Gaussian filter gt(x) with variance from t = 0 to
2156
+ +∞, we will see that:
2157
+
2158
+ HEAT EVOLUTION
2159
+ 31
2160
+ Require: a, b, c, d of p(x) = x4 + ax3 + bx2 + cx + d, and ∆t, pre.
2161
+ Ensure: global minimizer xmin
2162
+ 1: function Iteration(a, b, c, d)
2163
+ 2:
2164
+ [tu, xinit] ← Initialize(a, b, c)
2165
+ 3:
2166
+ 4:
2167
+ if tu < 0 or −a/4 is critical point then computing x
2168
+ 5:
2169
+ 6:
2170
+ else
2171
+ 7:
2172
+ t ← tu
2173
+ 8:
2174
+ x ← xint
2175
+ 9:
2176
+ while t > pre do
2177
+ 10:
2178
+ r ← (12x + 3a)/(12x2 + 6ax + 2b + 12t)
2179
+ 11:
2180
+ x ← x − ∆t · r
2181
+ 12:
2182
+ t ← t − ∆t
2183
+ 13:
2184
+ end while
2185
+ 14:
2186
+ 15:
2187
+ end if
2188
+ 16:
2189
+ return x
2190
+ 17: end function
2191
+ 18: function Initialize(a, b, c)
2192
+ 19:
2193
+ h ← a3 − 4ab + 8c
2194
+ 20:
2195
+ t∗ ← a2/16 − b/6
2196
+ 21:
2197
+ tu ← t∗ − h
2198
+ 2
2199
+ 3 /16
2200
+ 22:
2201
+ xinit ← −a/4 − h1/3/2
2202
+ 23:
2203
+ return tu, xinit
2204
+ 24: end function
2205
+ • If tu < 0, p(x) itself is not necessary convex, but it has unique critical
2206
+ point. Consequently, each p(x, t) has only one critical point at any
2207
+ t ≥ 0;
2208
+ • If further tu ≤ t∗ < 0, p(x) must be convex. Then each p(x, t) is
2209
+ convex about x at any t ≥ 0;
2210
+ • If tu > 0, the polynomial p(x) has three distinct critical points x1 <
2211
+ x2 < x3 when 0 ≤ t < tu;
2212
+ • When t = tu, the critical point corresponding to the global minimizer
2213
+ will evolve continuously from tu to t∗, and the local minimizer will
2214
+ meet up with local maximum xt
2215
+ 2.
2216
+ Even more, these two critical
2217
+ points will stop evolution at t = tu;
2218
+ • When tu < t < t∗, the polynomial p(x, t) has unique minimizer at
2219
+ each t.
2220
+ • When t ≥ t∗, the polynomial p(x, t) will become convex about x,
2221
+ and possesses unique minimizer.
2222
+
2223
+ 32
2224
+ QIAO WANG
2225
+ -1.5
2226
+ -1
2227
+ -0.5
2228
+ 0
2229
+ 0.5
2230
+ 1
2231
+ 1.5
2232
+ x
2233
+ -1
2234
+ -0.5
2235
+ 0
2236
+ 0.5
2237
+ 1
2238
+ 1.5
2239
+ 2
2240
+ y
2241
+ x1(0)
2242
+ x2(0)
2243
+ x3(0)
2244
+ x2(tu)=x3(tu)
2245
+ x1(tu)
2246
+ x3(t)
2247
+ x2(t)
2248
+ x1(t)
2249
+ Figure 7. An example of triangle series in (x, y) system, of
2250
+ p(x) = x4 + 0.2114x3 − 2.6841x2 − 0.1110x + 1.2406. See
2251
+ Example 4 for more details.
2252
+ -2
2253
+ -1.5
2254
+ -1
2255
+ -0.5
2256
+ 0
2257
+ 0.5
2258
+ 1
2259
+ x
2260
+ 0
2261
+ 0.5
2262
+ 1
2263
+ 1.5
2264
+ 2
2265
+ 2.5
2266
+ 3
2267
+ t
2268
+ Figure 8. Fingerprint FP1 in (x, t) system, of p(x) = x4 +
2269
+ 0.2114x3 − 2.6841x2 − 0.1110x + 1.2406. See Example 4 for
2270
+ more details.
2271
+ 5. Case study of sixth degree polynomials
2272
+ 5.1. Evolution and fingerprints. Now we consider 6 degree monic poly-
2273
+ nomial
2274
+ p(x) = x6 + bx4 + cx3 + dx2 + ex + f,
2275
+ (5.1)
2276
+ For the sake of simplicity, here we already regularize the coefficient of x5
2277
+ by setting it as zero, which is a standard technique in treating the algebraic
2278
+
2279
+ HEAT EVOLUTION
2280
+ 33
2281
+ -1.5
2282
+ -1
2283
+ -0.5
2284
+ 0
2285
+ 0.5
2286
+ 1
2287
+ 1.5
2288
+ 2
2289
+ 2.5
2290
+ 3
2291
+ x
2292
+ -10
2293
+ -8
2294
+ -6
2295
+ -4
2296
+ -2
2297
+ 0
2298
+ 2
2299
+ 4
2300
+ 6
2301
+ 8
2302
+ y
2303
+ x1(0)
2304
+ x3(0)
2305
+ x1(tu)=x2(tu)=x3(tu)
2306
+ x2(0)
2307
+ x(t), t> tu
2308
+ Figure 9. Triangle series of p(x) = x4 − 4x3 − 2x2 + 12x,
2309
+ in which there exist two global minimizers. Notice that at
2310
+ x1(tu) = x2(tu) = x3(tu), the fingerprint of x2(t) is a line
2311
+ segmentation, which is partial repeated by global minimizer
2312
+ curve x(t) after t ≥ tu. See Example 5 for more details.
2313
+ equations. This implies that the heat evolution is
2314
+ p(x, t) =p(x) + t · ∂p
2315
+ ∂t + t2
2316
+ 2
2317
+ ∂2p
2318
+ ∂t2 + t3
2319
+ 6
2320
+ ∂3p
2321
+ ∂t3
2322
+ =p(x) + t
2323
+ 2 · ∂2p
2324
+ ∂x2 + t2
2325
+ 8
2326
+ ∂4p
2327
+ ∂x4 + t3
2328
+ 48
2329
+ ∂6p
2330
+ ∂x6
2331
+ =x6 + b(t)x4 + c(t)x3 + d(t)x2 + e(t)x + f(t),
2332
+ (5.2)
2333
+ in which
2334
+
2335
+
2336
+
2337
+
2338
+
2339
+
2340
+
2341
+
2342
+
2343
+
2344
+
2345
+
2346
+
2347
+
2348
+
2349
+ b(t) = b + 15t,
2350
+ c(t) = c,
2351
+ d(t) = d + 6bt + 45t2,
2352
+ e(t) = e + 3ct,
2353
+ f(t) = f + dt + 3bt2 + 15t3.
2354
+ (5.3)
2355
+ The critical points of (5.1) satisfy the 5 degree equation
2356
+ 0 = 1
2357
+ 6
2358
+ ∂p(x, t)
2359
+ ∂x
2360
+ = x5 + B(t)x3 + C(t)x2 + D(t)x + E(t),
2361
+ (5.4)
2362
+ in which
2363
+
2364
+
2365
+
2366
+
2367
+
2368
+
2369
+
2370
+
2371
+
2372
+
2373
+
2374
+
2375
+
2376
+
2377
+
2378
+
2379
+
2380
+
2381
+
2382
+
2383
+
2384
+ B(t) = 2b
2385
+ 3 + 10t,
2386
+ C(t) = c
2387
+ 2,
2388
+ D(t) = d
2389
+ 3 + 2bt + 15t2,
2390
+ E(t) = e
2391
+ 6 + ct
2392
+ 2 .
2393
+ (5.5)
2394
+
2395
+ 34
2396
+ QIAO WANG
2397
+ -0.5
2398
+ 0
2399
+ 0.5
2400
+ 0
2401
+ 10
2402
+ 20
2403
+ 10-3
2404
+ t = 0
2405
+ -0.5
2406
+ 0
2407
+ 0.5
2408
+ -0.1
2409
+ 0
2410
+ 0.1
2411
+ 0.2
2412
+ -0.5
2413
+ 0
2414
+ 0.5
2415
+ 0
2416
+ 1
2417
+ 2
2418
+ -0.5
2419
+ 0
2420
+ 0.5
2421
+ -20
2422
+ 0
2423
+ 20
2424
+ -0.5
2425
+ 0
2426
+ 0.5
2427
+ 0
2428
+ 10
2429
+ 20
2430
+ 10-3t = 0.02
2431
+ -0.5
2432
+ 0
2433
+ 0.5
2434
+ -0.5
2435
+ 0
2436
+ 0.5
2437
+ -0.5
2438
+ 0
2439
+ 0.5
2440
+ 0
2441
+ 1
2442
+ 2
2443
+ 3
2444
+ -0.5
2445
+ 0
2446
+ 0.5
2447
+ -20
2448
+ 0
2449
+ 20
2450
+ -0.5
2451
+ 0
2452
+ 0.5
2453
+ 0
2454
+ 10
2455
+ 20
2456
+ 10-3t = 0.05
2457
+ -0.5
2458
+ 0
2459
+ 0.5
2460
+ -1
2461
+ 0
2462
+ 1
2463
+ -0.5
2464
+ 0
2465
+ 0.5
2466
+ 0
2467
+ 5
2468
+ -0.5
2469
+ 0
2470
+ 0.5
2471
+ -20
2472
+ 0
2473
+ 20
2474
+ Figure 10. The evolution of six degree polynomial p(x, t)
2475
+ and its derivatives ∂p(x,t)
2476
+ ∂x
2477
+ , ∂2p(x,t)
2478
+ ∂x2
2479
+ , ∂3p(x,t)
2480
+ ∂x3
2481
+ .
2482
+ The solution of (5.4) for t ≥ 0 consists the fingerprint FP1. Unfortunately,
2483
+ the roots of this fifth degree equation is algebraically intractable [18].
2484
+ Similarly, we may write the equation of FP2 as below,
2485
+ 1
2486
+ 30
2487
+ ∂2p
2488
+ ∂x2 = x4 +
2489
+ �2b
2490
+ 5 + 6t
2491
+
2492
+ x2 + c
2493
+ 5x + d
2494
+ 15 + 2b
2495
+ 5 t + 3t2 = 0.
2496
+ (5.6)
2497
+ Then the equation of FP3 is
2498
+ 1
2499
+ 120
2500
+ ∂3p
2501
+ ∂x3 = x3 +
2502
+ �b
2503
+ 5 + 3t
2504
+
2505
+ x + c
2506
+ 20 = 0.
2507
+ (5.7)
2508
+ Our interest is the set FP2
2509
+ � FP3. Geometrically, there exist two pair of
2510
+ real double roots of quartic equation (5.21), based on following Lemma.
2511
+ Lemma 13. If x0 be a root of both polynomial p(x) and its derivative p′(x),
2512
+ then it must be at least a double root of p(x).
2513
+ Different from quartic polynomials case, we have not an explicit represen-
2514
+ tation for FP2
2515
+ � FP3, and numerical approach is required here for exper-
2516
+ iments. Clearly, the real double root of quartic equation (5.6) must be the
2517
+ common roots of both (5.6) and (5.7). In general, there exist two 0 ≤ t1 < t2
2518
+
2519
+ HEAT EVOLUTION
2520
+ 35
2521
+ 0
2522
+ 0.005
2523
+ 0.01
2524
+ 0.015
2525
+ 0.02
2526
+ 0.025
2527
+ 0.03
2528
+ 0.035
2529
+ 0.04
2530
+ t
2531
+ -1.5
2532
+ -1
2533
+ -0.5
2534
+ 0
2535
+ 0.5
2536
+ 1
2537
+ 1.5
2538
+ 2
2539
+ (t)
2540
+ 10-6
2541
+ (t)
2542
+ 0
2543
+ Figure 11. The discriminant ∆(t) of quartic equation is
2544
+ generated from ∂2p(x,t)
2545
+ ∂x2
2546
+ = 0 and defined in (5.8). Here the
2547
+ data comes from Example 1. The two real roots of ∆(t) are
2548
+ t1 = 0.002341, at which the quartic equation possesses a real
2549
+ double root and two distinct real roots, and t2 = 0.034887,
2550
+ at which the quartic equation possesses a real double root
2551
+ and a pair of conjugate complex roots.
2552
+ such that the corresponding x1 and x2 are those two real double roots. The
2553
+ following Theorem 19 explains the process of numerical approach.
2554
+ Theorem 19. For any six degree monic polynomial p(x), the set FP2
2555
+ � FP3
2556
+ contains a pair of elements (xi, ti), i = 1, 2, or one element (x1, t1), or
2557
+ empty. Specifically,
2558
+ (1) Any ti must be the zero of discriminant
2559
+ ∆(t) = 27648t′6 + 1728c2
2560
+ 25
2561
+ t′3 − 256
2562
+ 625h2t′2 − 288
2563
+ 625c2ht′ − 256h3
2564
+ 753
2565
+ − 27c4
2566
+ 625
2567
+ (5.8)
2568
+ where
2569
+ t′ = t + b
2570
+ 15,
2571
+ h = b2 − 5d.
2572
+ (5.9)
2573
+
2574
+ 36
2575
+ QIAO WANG
2576
+ 0
2577
+ 0.005
2578
+ 0.01
2579
+ 0.015
2580
+ 0.02
2581
+ 0.025
2582
+ 0.03
2583
+ 0.035
2584
+ 0.04
2585
+ t
2586
+ -1
2587
+ -0.8
2588
+ -0.6
2589
+ -0.4
2590
+ -0.2
2591
+ 0
2592
+ 0.2
2593
+ 0.4
2594
+ 0.6
2595
+ 0.8
2596
+ 1
2597
+ x(t)
2598
+ 0.23516
2599
+ -0.078914
2600
+ Figure 12. The function x(t) is defined in (5.10), where the
2601
+ data comes from the Example 1. Here we obtain two solution
2602
+ (t1, x1) = (0.0023, 0.23516), (t2, x2) = (0.03489, −0.078914),
2603
+ which consist the set FP2
2604
+ � FP3.
2605
+ (2) Any xi is dependent of ti by the function
2606
+ x = −c ·
2607
+ 1800
2608
+
2609
+ t + b
2610
+ 15
2611
+ �2 − 4(b2 − 5d)
2612
+ 36000
2613
+
2614
+ t + b
2615
+ 15
2616
+ �3 + 80(b2 − 5d)
2617
+
2618
+ t + b
2619
+ 15
2620
+
2621
+ + 45c2 .
2622
+ (5.10)
2623
+ Proof. We will give a detailed analysis in Apendix B, based on which, we
2624
+ know that in general settings there exist at most two merge time t1 and t2
2625
+ from the discriminant equation ∆(t) = 0 of quartic equation (5.6) and (B.8),
2626
+ This function (5.8) can be verified by (B.7) and (B.8) immediately, but
2627
+ we omit the detailed computation here. We may find out its two real zeros
2628
+ t1 and t2 through numerical computation, then the real double roots x1 and
2629
+ x2 of (5.6) at t1 and t2 respectively, could be obtained according to following
2630
+ (5.10). Thus in general settings, i.e., when p′(x) has five distinct real roots,
2631
+ we will have
2632
+ FP2
2633
+
2634
+ FP3 = {(x1, t1), (x2, t2)}.
2635
+ (5.11)
2636
+ In degenerated cases, this intersection set might possess one point or even
2637
+ null. When it is null, the polynomial p(x) is globally convex.
2638
+
2639
+ HEAT EVOLUTION
2640
+ 37
2641
+ Here we may take Euclidean algorithm to reduce the degree of the poly-
2642
+ nomials about x for (5.6) and (5.7). At first, multiplying (5.7) with x, and
2643
+ subtracted both sides of (5.6) resp., we get a second degree polynomial
2644
+
2645
+ t + b
2646
+ 15
2647
+
2648
+ x2 + c
2649
+ 20x +
2650
+
2651
+ t + b
2652
+ 15
2653
+ �2
2654
+ − b2 − 5d
2655
+ 225
2656
+ = 0.
2657
+ (5.12)
2658
+ Again, multiplying with x for both sides of (5.12), and subtracted from both
2659
+ sides of (5.7) multiplied with t + b
2660
+ 15, then we obtain
2661
+ − c
2662
+ 20x2 +
2663
+
2664
+ 2
2665
+
2666
+ t + b
2667
+ 15
2668
+ �2
2669
+ + b2 − 5d
2670
+ 225
2671
+
2672
+ x + c
2673
+ 20
2674
+
2675
+ t + b
2676
+ 15
2677
+
2678
+ = 0.
2679
+ (5.13)
2680
+ Finally, eliminating the second degree term by combining (5.13) and
2681
+ (5.12) will lead to (5.10).
2682
+
2683
+ As explained in Appendix B, we can obtain the suitable t from the dis-
2684
+ criminant ∆(t) = 0 at first, then substitute it into the above (5.10), then
2685
+ choose x from (5.10).
2686
+ 5.2. The boundary of confinement zone. To characterize the boundary
2687
+ of confinement zone, we must study the singular trajectory
2688
+ dx
2689
+ dt = −
2690
+ ∂3p
2691
+ ∂x3
2692
+ 2 ∂2p
2693
+ ∂x2
2694
+ ,
2695
+ Initial Condition : (t′, x(t′)) ∈ FP2
2696
+
2697
+ FP3.
2698
+ (5.14)
2699
+ Unfortunately, this equation is singular due to its initial data, which re-
2700
+ sults in
2701
+ dx
2702
+ dt = 0
2703
+ 0.
2704
+ (5.15)
2705
+ However, at critical time t′, the multiplicity of common root x′ of both FP2
2706
+ and FP3 is different. Clearly, x′ is double root of FP2 and single root of
2707
+ FP3, thus the reciprocal equation of (5.14)
2708
+ dt
2709
+ dx = −2 ∂2p
2710
+ ∂x2
2711
+ ∂3p
2712
+ ∂x3
2713
+ ,
2714
+ Initial Condition : (t(x′), x′) ∈ FP2
2715
+
2716
+ FP3
2717
+ (5.16)
2718
+ contains only removable singularity near FP2
2719
+ � FP3.
2720
+ Fig.13 demonstrates the limitation curve satisfying (3.8) and inversely
2721
+ evolutes as t → +0 starting from the top points near FP2
2722
+ � FP3.
2723
+
2724
+ 38
2725
+ QIAO WANG
2726
+ -0.6
2727
+ -0.5
2728
+ -0.4
2729
+ -0.3
2730
+ -0.2
2731
+ -0.1
2732
+ 0
2733
+ 0.1
2734
+ 0.2
2735
+ 0.3
2736
+ 0.4
2737
+ x
2738
+ 0
2739
+ 0.005
2740
+ 0.01
2741
+ 0.015
2742
+ 0.02
2743
+ 0.025
2744
+ 0.03
2745
+ 0.035
2746
+ 0.04
2747
+ 0.045
2748
+ 0.05
2749
+ t
2750
+ X1
2751
+ LL
2752
+ X1
2753
+ LR
2754
+ X1
2755
+ RL
2756
+ X1
2757
+ RR
2758
+ X2
2759
+ LL
2760
+ X2
2761
+ RR
2762
+ X2
2763
+ LR,X2
2764
+ RL
2765
+ Figure 13.
2766
+ Numerically, the partition of Confinement
2767
+ Zone and Escape Zone associated with p(x) defined in Ex-
2768
+ ample 1 is obtained through Matlab ODE packet of @ode25,
2769
+ and
2770
+ the
2771
+ Confinement
2772
+ Zone
2773
+ is
2774
+ [−0.5082, 0.0858] �[0.1603, 0.3267].
2775
+ 5.3. Why convecification approach can not guarantee attaining the
2776
+ global minimizer? Recall the regularized polynomial (5.1), which contains
2777
+ the parameters {b, c, d, e, f}, and the FP1 equation (5.4) contains {b, c, d, e}.
2778
+ However, the FP2 equation (5.6) and the FP3 equation (5.7), contains only
2779
+ {b, c, d} and {b, c} respectively, which is independent of e. Thus FP2
2780
+ � FP3
2781
+ doesn’t contain the information of e, which actually affect the location of
2782
+ global minimizer of polynomial p(x) in (5.1)6.
2783
+ On the other hand, to determine the scope of the Confinement Zone, we
2784
+ require the equation (3.8) which depends only on {b, c, d}. Thus we must
2785
+ investigate the affection of e in the (5.1) to the global minimizer.
2786
+ Thus
2787
+ we may consider that for fixed b, c, d and vary e, the variation of global
2788
+ minimizer of corresponding p(x), and when it is included in the Escape
2789
+ 6The parameter f in (5.1) doesn’t affect the location of global minimizer of six degree
2790
+ polynomial.
2791
+
2792
+ HEAT EVOLUTION
2793
+ 39
2794
+ Zone. To this end, we may define the mapping
2795
+ R(e|b, c, d) =
2796
+
2797
+ 1,
2798
+ x∗ ∈ Escape Zone,
2799
+ −1,
2800
+ x∗ ∈ Confinement Zone.
2801
+ (5.17)
2802
+ where x∗ represents the global of polynomial p(x) in (5.1). The detailed
2803
+ analysis reveals that when
2804
+ e ∈ (−∞, 0.676739]
2805
+
2806
+ [−0.617543, −0.58523]
2807
+
2808
+ [−0.67115, +∞),
2809
+ (5.18)
2810
+ the corresponding global minimizer x∗ falls into the scape of Escape Zone.
2811
+ Fig17 illustrates the curve of x∗(e|b, c, d), the global minimizer of polynomial
2812
+ p(x) in (5.1) in which b, c, d remains invariant, while varying the parameter e.
2813
+ Those x∗ fails to be obtained by convexification approach is demonstrated.
2814
+ 5.4. Comparison principle and criterion function for evolution poly-
2815
+ nomials. Now we describe the comparison criterion for p(xi, t) > p(xj, t),
2816
+ where xi and xj are critical points of p(x, t) at time t. Apparently, it follows
2817
+ immediately from (5.1) that
2818
+ p(xi, t) − p(xj, t) = (xi − xj) · Q5(xi, xj, t)
2819
+ (5.19)
2820
+ in which Q5 is a fifth degree polynomial. However, it is too complicated for
2821
+ analysis. Instead, we can give a more concise representation for factoring
2822
+ the p(xi, t) − p(xj, t).
2823
+ Theorem 20. Let ξ = ξ(t) and η = η(t) be critical points of p(x, t), we
2824
+ have
2825
+ p(ξ, t) − p(η, t) = −(ξ − η)3
2826
+ 10
2827
+ · K(ξ, η, t),
2828
+ (5.20)
2829
+ where the criterion function
2830
+ K(ξ, η, t) = 20(ξ3 + η3) + 30(ξ2η + ξη2)
2831
+ +15a(ξ2 + η2) + 20aξη
2832
+ + (10b + 150t)(ξ + η) + (5c + 50at)
2833
+ =K(ξ, η, 0) + 150t(ξ + η) + 50at.
2834
+ (5.21)
2835
+ If both ξ and η (where we suppose that ξ ̸= η) are real critical points of p(x),
2836
+ then for sixth degree monic polynomial p(x, t),
2837
+ p(ξ, t) > p(η, t) ⇐⇒ (ξ − η)K(ξ, η, t) < 0.
2838
+ (5.22)
2839
+ Remark 10. If we transform the critical points by translation xi → xi − a
2840
+ 5,
2841
+ we may set a = 0 in (5.21), then the criterion function (5.21) can be reduced
2842
+ to
2843
+ K(ξ, η, t) = 20(ξ3 + η3) + 30(ξ2η + ξη2) + (10b + 150t)(ξ + η) + 5c. (5.23)
2844
+
2845
+ 40
2846
+ QIAO WANG
2847
+ Proof of Theorem 20. Let xi (i = 1, 2, 3, 4, 5) be the critical points of monic
2848
+ sixth degree polynomial p(x, t) at t, i.e., the roots of equation 1
2849
+ 6
2850
+ ∂p(x,t)
2851
+ ∂x
2852
+ = 0.
2853
+ Thus we may write
2854
+
2855
+
2856
+
2857
+
2858
+
2859
+
2860
+
2861
+
2862
+
2863
+
2864
+
2865
+
2866
+
2867
+
2868
+
2869
+
2870
+
2871
+
2872
+
2873
+
2874
+
2875
+
2876
+
2877
+
2878
+
2879
+
2880
+
2881
+
2882
+
2883
+
2884
+
2885
+
2886
+
2887
+ A(t) = −
2888
+
2889
+ i
2890
+ xi,
2891
+ B(t) =
2892
+
2893
+ i<j
2894
+ xixj,
2895
+ C(t) = −
2896
+
2897
+ i<j<k
2898
+ xixjxk,
2899
+ D(t) =
2900
+
2901
+ i<j<k<l
2902
+ xixjxkxl,
2903
+ E(t) = −x1x2x3x4x5.
2904
+ (5.24)
2905
+ Without loss of generality, we may set ξ = x1, η = x2. Then factoring the
2906
+ polynomial p(ξ, t) − p(η, t), we will obtain
2907
+ p(ξ, t) − p(η, t) = −(ξ − η)3
2908
+ 10
2909
+ · K(ξ, η, x3, x4, x5, t),
2910
+ But this K(ξ, η, x3, x4, x5, t) can be represented as function of compositions
2911
+ of I1 = x3 + x4 + x5, I2 = x3x4 + x4x5 + x5x3 and I3 = x3x4x5, as well as
2912
+ ξ,η and t. Note that
2913
+ I1 = x3 + x4 + x5 = −A(t) − ξ − η,
2914
+ I2 = x3x4 + x4x5 + x5x3 = B(t) − (ξ + η) · I1,
2915
+ I3 = x3x4x5 = −C(t) − (ξ + η) · I2 − ξη · I1,
2916
+ (5.25)
2917
+ then we obtain the representation of p(x, t) on account of A(t), B(t), C(t) and
2918
+ ξ, η. Finally, representing K(ξ, η, x3, x4, x5, t) as functions of ξ, η, a, b, c, d, e, f, t
2919
+ will yield the required results.
2920
+
2921
+ Remark 11. Notice that ξ and η are symmetric in function K(ξ, η, t),
2922
+ thus we may observe the information in semi-plane ξ < η, then p(ξ, t) <
2923
+ p(η, t) ⇐⇒ K(ξ, η, t) < 0, and p(ξ, t) > p(η, t) ⇐⇒ K(ξ, η, t) > 0.
2924
+ In general settings, a monic sixth degree polynomial p(x) has five real
2925
+ critical points x1 < x2 < x3 < x4 < x5, and
2926
+ p(x1) < p(x2) > p(x3) < p(x4) > p(x5).
2927
+ Accordingly, we may see that
2928
+ K(x1, x2), K(x3, x4) < 0,
2929
+ and
2930
+ K(x2, x3), K(x4, x5) > 0.
2931
+ Our main task is to determine the
2932
+ arg
2933
+ min
2934
+ i∈{1,3,5} p(xi).
2935
+ (5.26)
2936
+
2937
+ HEAT EVOLUTION
2938
+ 41
2939
+ Thus we have a graphical criterion by partition the (ξ, η) plane into P/N
2940
+ parts according to the sign of K(ξ, η, t = 0):
2941
+ Theorem 21 (criterion of global minimizer). Let x1 < x2 < x3 < x4 < x5
2942
+ be five critical points of monic sixth degree polynomial p(x), we have the
2943
+ following criterion
2944
+ (1) x1 = arg minx p(x) if and only if K(x1, x3) ≤ 0 and K(x1, x5) ≤ 0;
2945
+ (2) x3 = arg minx p(x) if and only if K(x1, x3) ≥ 0 and K(x3, x5) ≤ 0;
2946
+ (3) x5 = arg minx p(x) if and only if K(x1, x5) ≥ 0 and K(x3, x5) ≥ 0.
2947
+ In a summary, the sign of K(x1, x3), K(x1, x5) and K(x3, x5) determines
2948
+ the global minimizer.
2949
+ 5.5. The level set of criterion surface. Our main challenge is to explain
2950
+ the merge of two critical points at time evolution. That is, in general set-
2951
+ tings, we intend to find out those two merge time t1 and t2 such that at
2952
+ each merge time ti, there is a pair of critical points satisfy ξ(t1) = η(t1), and
2953
+ another pair of points satisfy ξ(t2) = η(t2). Notice that these don’t mean
2954
+ that they satisfy K(ξ, η, ti) = 0.
2955
+ In general setting, we assume that all five critical points are real and
2956
+ separate. Define the sets
2957
+ Zt(K) ={(ξ, η); K(ξ, η, t) = 0},
2958
+ Z+
2959
+ t (K) ={(ξ, η); K(ξ, η, t) > 0},
2960
+ Z−
2961
+ t (K) ={(ξ, η); K(ξ, η, t) < 0}.
2962
+ (5.27)
2963
+ Proposition 1. If the initial sixth degree polynomial contains five real crit-
2964
+ ical points xi, then the regularization a = 0 will lead to
2965
+ b ≤ 0,
2966
+ (5.28)
2967
+ And each xi satisfies
2968
+ |xi| ≤
2969
+
2970
+ −b.
2971
+ (5.29)
2972
+ Proof. We have
2973
+ 0 = a2 =
2974
+ ��
2975
+ i
2976
+ xi
2977
+ �2
2978
+ =
2979
+
2980
+ i
2981
+ x2
2982
+ i + 2
2983
+
2984
+ i<j
2985
+ xixj =
2986
+
2987
+ i
2988
+ x2
2989
+ i + b,
2990
+ (5.30)
2991
+ which results in b ≤ 0.
2992
+
2993
+ Proposition 2. Under the reduced form a = 0, the straight line ξ + η = 0
2994
+ is contained in Zt(K) (Z+
2995
+ t (K), or Z−
2996
+ t (K), resp. ) for all t if and only if
2997
+ c = 0 (c > 0, or c < 0, resp.).
2998
+ Proof. This can be immediately obtained from (5.23).
2999
+
3000
+ To understand the evolution of criterion surface, it is more natural to
3001
+ change the coordinates while setting a = 0. Define
3002
+ u = ξ + η,
3003
+ v = η − ξ,
3004
+ (5.31)
3005
+
3006
+ 42
3007
+ QIAO WANG
3008
+ we may rewrite the criterion surface K(ξ, η, t) as
3009
+ ˜K(u, v, t) = 25
3010
+ 2 u3 +
3011
+ �15
3012
+ 2 v2 + 10b + 150t
3013
+
3014
+ u + 5c.
3015
+ (5.32)
3016
+ Then the level set Zt(K) = 0 can be characterized by
3017
+ u3 +
3018
+ �3
3019
+ 5v2 + 4b
3020
+ 5 + 12t
3021
+
3022
+ u + 2c
3023
+ 5 = 0.
3024
+ (5.33)
3025
+ Notice that ξ < η is equivalent to v > 0, and ξ = η means v = 0. Thus we
3026
+ may represent the discriminant of the cubic function of u (5.32) by
3027
+ ∆(v, t) =
3028
+ �v2
3029
+ 5 + 4b
3030
+ 15 + 4t
3031
+ �3
3032
+ + c2
3033
+ 25.
3034
+ (5.34)
3035
+ The solution about u in ˜K(u, v, t) = 0 contains three distinct, (or one and
3036
+ a repeated pair, and single, resp.), real roots, when ∆(v, t) < 0, (or =, > 0
3037
+ resp.) Notice that the equation (5.32) is symmetric about v, we may consider
3038
+ only about v ≥ 0. Clearly, the sign of ∆(v, t) is the same as that of
3039
+ v2
3040
+ 5 + 4b
3041
+ 15 + 4t +
3042
+ �c
3043
+ 5
3044
+ � 2
3045
+ 3 ,
3046
+ (5.35)
3047
+ which is monotonically increasing with t. Therefore we have
3048
+ Theorem 22 (Level Set). For ˜K(u, v, t) = 0, set
3049
+ t∗
3050
+ v = −v2
3051
+ 20 − b
3052
+ 15 −
3053
+ � c
3054
+ 40
3055
+ � 2
3056
+ 3 .
3057
+ (5.36)
3058
+ Then for 0 ≤ t < t∗
3059
+ v, the equation has three distinct real roots,
3060
+
3061
+
3062
+
3063
+
3064
+
3065
+
3066
+
3067
+
3068
+
3069
+
3070
+
3071
+
3072
+
3073
+ u1(v, t) = P(v, t) + Q(v, t),
3074
+ u2(v, t) = −1
3075
+ 2 [P(v, t) + Q(v, t)] +
3076
+
3077
+ 3
3078
+ 2 i [P(v, t) − Q(v, t)] ,
3079
+ u3(v, t) = −1
3080
+ 2 [P(v, t) + Q(v, t)] −
3081
+
3082
+ 3
3083
+ 2 i [P(v, t) − Q(v, t)] ,
3084
+ (5.37)
3085
+ in which
3086
+
3087
+
3088
+
3089
+
3090
+
3091
+
3092
+
3093
+
3094
+
3095
+ P(v, t) =
3096
+ 3
3097
+
3098
+ − c
3099
+ 40 +
3100
+
3101
+ ∆(v, t),
3102
+ Q(v, t) =
3103
+ 3
3104
+
3105
+ − c
3106
+ 40 −
3107
+
3108
+ ∆(v, t).
3109
+ (5.38)
3110
+ If t > t∗
3111
+ v ≥ 0, i.e., ∆(v, t) > 0, it has only one real solution
3112
+ u(t) = P(v, t) + Q(v, t).
3113
+ (5.39)
3114
+ Finally, when t = t∗
3115
+ v, i.e., ∆(v, t) = 0, it has one real root and in addition,
3116
+ two repeated real roots,
3117
+ u1(v, t∗
3118
+ v) = −2 3
3119
+ � c
3120
+ 40,
3121
+ u2(v, t∗
3122
+ v) = u3(v, t∗
3123
+ v) =
3124
+ 3
3125
+ � c
3126
+ 40.
3127
+ (5.40)
3128
+
3129
+ HEAT EVOLUTION
3130
+ 43
3131
+ This Theorem indicates the structure of the level set Zt(K).
3132
+ Theorem 23. Under the reduced form (a = 0), define
3133
+ t∗
3134
+ max = − b
3135
+ 15 −
3136
+ � c
3137
+ 40
3138
+ � 2
3139
+ 3 .
3140
+ (5.41)
3141
+ (1) When t∗
3142
+ max < 0, the level set contains a unique continuous curve C,
3143
+ which is asymptotically by a line parallel to ξ + η = 0;
3144
+ (2) When t∗
3145
+ max > 0, for any fixed t ∈ [0, t∗
3146
+ max], there are two types of
3147
+ the configuration of the level set. If c > 0, the oval-like part of Zt is
3148
+ included in the north-east part of the plane w.r.t. the long curve C.
3149
+ If c < 0, it is contained in the south-east part.
3150
+ (3) When t∗
3151
+ max > 0, then for any fixed t ∈ [0, t∗
3152
+ max], then the oval-like
3153
+ closed curve spanned in the scope of v2 ≤ 20(t∗
3154
+ max − t). The level set
3155
+ inside this scope is characterized by (5.37). In particular, one pair
3156
+ of the vortex of this oval like curve is
3157
+ u = − 3
3158
+
3159
+ − c
3160
+ 40,
3161
+ v(t) = ±
3162
+
3163
+ 20(t∗max − t).
3164
+ (5.42)
3165
+ and another pair of vertex at v = 0 is u1,2(t), which are the solution
3166
+ of (5.37) except for the minimal one if c > 0 (or maximal one if
3167
+ c < 0).
3168
+ (4) When t∗
3169
+ max = 0, then at t = 0, the level set contains a curve C and
3170
+ a point u = − 3�− c
3171
+ 40, v = 0.
3172
+ 5.6. Symmetry and trend of the criterion function. We consider the
3173
+ case u = 0, i.e.,
3174
+ ξ + η = 0.
3175
+ (5.43)
3176
+ Clearly, we see that
3177
+ K(x, y, t) = 5c,
3178
+ if ξ = x + y = 0.
3179
+ (5.44)
3180
+ We further consider v = 0, i.e., ξ = η at zero level set, that is,
3181
+ K(ξ, ξ, t) = ˜K(u = 2ξ, v = 0, t) = 0.
3182
+ (5.45)
3183
+ According to (5.23) we can obtain a concise cubic equation
3184
+ ξ3 +
3185
+ �b
3186
+ 5 + 3t
3187
+
3188
+ ξ + c
3189
+ 20 = 0.
3190
+ (5.46)
3191
+ Remark 12. This equation formally agrees with the equation
3192
+ ∂3p(x, t)
3193
+ ∂x3
3194
+ = 0,
3195
+ (a = 0),
3196
+ (5.47)
3197
+ since that
3198
+ 1
3199
+ 120
3200
+ ∂3p(x, t)
3201
+ ∂x3
3202
+ = x3 +
3203
+ �b
3204
+ 5 + 3t
3205
+
3206
+ x + c
3207
+ 20.
3208
+ (5.48)
3209
+
3210
+ 44
3211
+ QIAO WANG
3212
+ (a) a = 0, b = −0.3726, c = 0.0574
3213
+ (b) a = 0, b = −0.2938, c = −0.0797
3214
+ Figure 14. Criterion functions and P/N partition with pa-
3215
+ rameters a = 0, b = −0.3726, c = 0.0574 (a), and a = 0, b =
3216
+ −0.2938, c = −0.0797 (b). Here the green domain is positive
3217
+ domain.
3218
+
3219
+ 150
3220
+ 100
3221
+ 50
3222
+ 0
3223
+ -50
3224
+ -100
3225
+ -150
3226
+ -1
3227
+ -0.5
3228
+ 0
3229
+ 0.5
3230
+ 1
3231
+ -1
3232
+ n150
3233
+ 100
3234
+ 50
3235
+ 0:
3236
+ -50
3237
+ -100
3238
+ -150
3239
+ -1
3240
+ -0.5
3241
+ 0
3242
+ 0.50
3243
+ nHEAT EVOLUTION
3244
+ 45
3245
+ -0.6
3246
+ -0.4
3247
+ -0.2
3248
+ 0
3249
+ 0.2
3250
+ 0.4
3251
+ -0.6
3252
+ -0.5
3253
+ -0.4
3254
+ -0.3
3255
+ -0.2
3256
+ -0.1
3257
+ 0
3258
+ 0.1
3259
+ 0.2
3260
+ 0.3
3261
+ 0.4
3262
+ (a) t = 0
3263
+ -1
3264
+ -0.5
3265
+ 0
3266
+ 0.5
3267
+ 1
3268
+ -1
3269
+ -0.5
3270
+ 0
3271
+ 0.5
3272
+ 1
3273
+ (b) t = 0.002
3274
+ Figure 15. The evolution of position and sign (”+” or
3275
+ ”−”) indicates the critical points of p(x) = x6 − 0.3726x4 +
3276
+ 0.0574x3 + 0.0306x2 − 0.0084x by level set of criterion func-
3277
+ tion K(ξ, η, t). The ”o” at ξ = η stands for the critical points
3278
+ (xi, xi) (i = 1, 2, 3, 4, 5) and their evolution or merge. Note
3279
+ that all K(x1, xi, 0) < 0, then x1 is the global minimizer of
3280
+ p(x) according to Theorem 21.
3281
+
3282
+ 46
3283
+ QIAO WANG
3284
+ Notice that the structure of the solution of equation (5.46) and (5.48) is
3285
+ only a special case of (5.33) as v = 0.
3286
+ In a summary, the intersection of level set and the line ξ = η is just the
3287
+ roots of ∂3p
3288
+ ∂x3 = 0.
3289
+ 5.7. The merge time.
3290
+ 5.7.1. Merge time of critical points. Now we discuss the merge phenomenon
3291
+ of critical points. Two critical points meet up when ξ = η at time t, that
3292
+ is, v = 0 in above equations (5.31)-(5.36). Notice that if case is this, they
3293
+ must satisfy both the fingerprint equation ∂p
3294
+ ∂x = 0 and ∂2p
3295
+ ∂x2 = 0.
3296
+ Now we apply the reduced form for both equations and by assuming that
3297
+ a = 0, and obtain
3298
+
3299
+
3300
+
3301
+
3302
+
3303
+
3304
+
3305
+ x5 +
3306
+ �2b
3307
+ 3 + 10t
3308
+
3309
+ x3 + c
3310
+ 2x2 +
3311
+
3312
+ 15t2 + 2bt + d
3313
+ 3
3314
+
3315
+ x + 3ct + e
3316
+ 6
3317
+ = 0,
3318
+ x4 +
3319
+ �2b
3320
+ 5 + 6t
3321
+
3322
+ x2 + c
3323
+ 5x +
3324
+
3325
+ 3t2 + 2bt
3326
+ 5 + d
3327
+ 15
3328
+
3329
+ = 0,
3330
+ (5.49)
3331
+ Subtracted the second equation multiplied by x from the first one, we have
3332
+ �4b
3333
+ 15 + 4t
3334
+
3335
+ x3 + 3c
3336
+ 10x +
3337
+
3338
+ 12t2 + 8bt
3339
+ 5 + 4d
3340
+ 15
3341
+
3342
+ x + 3ct + e
3343
+ 6
3344
+ = 0,
3345
+ (5.50)
3346
+ Then, multiplied by x −
3347
+ 3c
3348
+ 40(t+b/15) and subtracted from the second equation
3349
+ multiplied by 4b
3350
+ 15 + 4t, we obtain a quadratic equation like
3351
+ F(t)x2 + G(t)x + H(t) = 0
3352
+ (5.51)
3353
+ We may first consider the merge time of fingerprints FP1 and FP2. But
3354
+ the systems of fifth degree polynomial and a quadratic polynomial will not
3355
+ be explicitly expressed. We apply Euclidean method to decrease the order
3356
+ of x. We begin with
3357
+ R1(x, t) = 1
3358
+ 6
3359
+ ∂p(x, t)
3360
+ ∂x
3361
+ ,
3362
+ R2(x) = 1
3363
+ 30
3364
+ ∂2p(x, t)
3365
+ ∂x2
3366
+ .
3367
+ (5.52)
3368
+ Then we apply the Euclidean algorithm,
3369
+ Ri(x, t) = (αi(t)x + βi(t))Ri+1(x, t) + Ri+2(x, t),
3370
+ i = 1, 2, 3, 4,
3371
+ (5.53)
3372
+ where all the terms are polynomials. Finally we obtain a rational polynomial
3373
+ representation
3374
+ x(t) = −M2(t)
3375
+ M1(t),
3376
+ (5.54)
3377
+ in which
3378
+ M1(t) =
3379
+ 5
3380
+
3381
+ i=0
3382
+ Niti,
3383
+ M2(t) =
3384
+ 6
3385
+
3386
+ i=0
3387
+ diti,
3388
+ (5.55)
3389
+
3390
+ HEAT EVOLUTION
3391
+ 47
3392
+ where
3393
+ N5 = − 34992000c,
3394
+ N4 = − 6480000e − 10368000bc,
3395
+ N3 = − 1287360b2c − 1728000be + 388800cd,
3396
+ N2 = − 64512b3c − 273600b2e − 23040bcd − 94770c3 + 504000de,
3397
+ N1 = − 1536b4c − 21120b3e + 4416b2cd − 6156bc3 + 67200bde − 32400c2e − 31680cd2,
3398
+ N0 = − 768b4e + 128b3cd + 4160b2de
3399
+ − 3510bc2e − 1152bcd2 + 729c3d + 3375ce2 − 4800d2e,
3400
+ (5.56)
3401
+ and
3402
+ D6 =466560000,
3403
+ D5 =186624000b,
3404
+ D4 =34214400b2 − 15552000d,
3405
+ D3 =3594240b3 − 4147200db + 2041200c2,
3406
+ D2 =211968b4 − 322560b2d + 537840bc2 − 648000ce − 230400d2,
3407
+ D1 =6144b5 − 6144b3d + 27504b2c2 − 104400bce
3408
+ − 30720bd2 + 93960c2d + 45000e2,
3409
+ D0 =1024b4d − 384b3c2 − 1920b2ce − 5632b2d2 + 8424bc2d
3410
+ + 3000be2 − 2187c4 − 10800cde + 7680d3.
3411
+ (5.57)
3412
+ The motivation of this rational representation comes from the fact that
3413
+ at the merge time t we have dx(t)
3414
+ dt
3415
+ = ∞, which means the suitable t must be
3416
+ the singularity. Thus we intend to obtain the singularity, i.e., the zeros of
3417
+ M1(t).
3418
+ 5.7.2. merge time of inflection points. Then consider the merge time of fin-
3419
+ gerprints FP2 and FP3.
3420
+ Combining the equation (5.49) with the criterion equation (5.23) (setting
3421
+ ξ = x) will yield a quadratic equation
3422
+ �b
3423
+ 5 + 3t
3424
+
3425
+ x2 + 3c
3426
+ 20x +
3427
+
3428
+ 3t2 + 2bt
3429
+ 5 + d
3430
+ 15
3431
+
3432
+ = 0,
3433
+ (5.58)
3434
+ with quadratic discriminant
3435
+ ∆2(t) = −36t3 − 36b
3436
+ 5 t2 −
3437
+ �8b2
3438
+ 25 + 4d
3439
+ 5
3440
+
3441
+ t + 9c2
3442
+ 400 − 4bd
3443
+ 75 .
3444
+ (5.59)
3445
+ When ∆2(t) ≥ 0, the equation possesses a pair of roots
3446
+ x1,2(t) = − 3c
3447
+ 20 ±
3448
+
3449
+ ∆2(t)
3450
+ 2
3451
+ � b
3452
+ 5 + 3t
3453
+
3454
+ .
3455
+ (5.60)
3456
+
3457
+ 48
3458
+ QIAO WANG
3459
+ 0
3460
+ 0.01
3461
+ 0.02
3462
+ 0.03
3463
+ 0.04
3464
+ 0.05
3465
+ 0.06
3466
+ t
3467
+ -10
3468
+ -8
3469
+ -6
3470
+ -4
3471
+ -2
3472
+ 0
3473
+ 2
3474
+ 4
3475
+ 6
3476
+ 8
3477
+ 10
3478
+ x
3479
+ (a) The singularity of − M2(t)
3480
+ M1(t) occurs.
3481
+ 0
3482
+ 0.01
3483
+ 0.02
3484
+ 0.03
3485
+ 0.04
3486
+ 0.05
3487
+ 0.06
3488
+ t
3489
+ -0.2
3490
+ 0
3491
+ 0.2
3492
+ 0.4
3493
+ 0.6
3494
+ 0.8
3495
+ 1
3496
+ 1.2
3497
+ 1.4
3498
+ x
3499
+ (b) The zeros of M1(t)
3500
+ Figure 16. The singularity occurs
3501
+
3502
+ HEAT EVOLUTION
3503
+ 49
3504
+ Therefore, combining it with (5.37) will produce the repeated root of (5.46).
3505
+ We can actually continue using Euclidean’s algorithm to reduce the order
3506
+ of polynomials w.r.t x, and obtain
3507
+ M1(t)x + M2(t) = 0,
3508
+ (5.61)
3509
+ thus
3510
+ x = −M1(t)
3511
+ M2(t),
3512
+ (5.62)
3513
+ in which
3514
+ M1(t) =18t3 + 18b
3515
+ 5 t2 +
3516
+ �7b2
3517
+ 25 − d
3518
+ 5 − 9ac
3519
+ 40
3520
+
3521
+ t
3522
+ − b
3523
+ 5
3524
+ � d
3525
+ 15 − b2
3526
+ 25
3527
+
3528
+ + 3c
3529
+ 20
3530
+ �3c
3531
+ 20 − ab
3532
+ 10
3533
+
3534
+ ,
3535
+ M2(t) =9c − 3ab
3536
+ 10
3537
+ t2 + 6bc − ab2 − 5ad
3538
+ 50
3539
+ t + 5cd + b2c
3540
+ 500
3541
+ − abd
3542
+ 150
3543
+ (5.63)
3544
+ 5.8. Global minimizers with varying S(x) = sx and differential
3545
+ equation. Now we explain the evolution of global minimizer of
3546
+ p(x) = x6 − 0.3726x4 + 0.0574x3 + 0.0376x2 + sx,
3547
+ (5.64)
3548
+ which is a modified version of Example 1. Fig.17 shows the evolution of
3549
+ global minimizers of this seesaw polynomial w.r.t the parameter s, the coef-
3550
+ ficient of x. Actually, at two points s1 and s2, there exists two distinct global
3551
+ minimizers pair x1(s1) and x2(s1), and x1(s2) and x2(s2), which occurs at
3552
+ p(x1(s1)) = p(x2(s1)) and p(x1(s2)) = p(x2(s2)) respectively.
3553
+ We give a detailed analysis on locating the s1 and s2.
3554
+ Actually, the
3555
+ polynomial p(x|s) has at most five real critical points for each s. Among
3556
+ them, three are local minimizers.
3557
+ Assume that x1, x2, x3 are three local
3558
+ minimizers of p(x), we may apply the seesaw equation to characterize their
3559
+ evolution as s changes. At the first phase, we start with s = −2, and apply
3560
+ dx
3561
+ ds = −
3562
+ 1
3563
+ p′′(x),
3564
+ xi = xi(s = −2),
3565
+ i = 1, 2, 3.
3566
+ (5.65)
3567
+ We take s goes to +∞, in order to obtain these three continuous curves
3568
+ begin with s = −2.
3569
+ 5.9. Comparison principle and criterion function for seesaw poly-
3570
+ nomials. Similar to the comparison criterion of evolution polynomials, we
3571
+ may compare p(xi|s) > p(xj|s), where xi and xj are critical points of p(x|s)
3572
+ at seesaw parameter s.
3573
+ Theorem 24. Let ξ = ξ(s) and η = η(s) be critical points of p(x|s), we
3574
+ have
3575
+ p(ξ|s) − p(η|s) = −(ξ − η)3
3576
+ 10
3577
+ · H(ξ, η),
3578
+ (5.66)
3579
+
3580
+ 50
3581
+ QIAO WANG
3582
+ Figure 17. The global minimizers x∗(s) of six degree poly-
3583
+ nomials p(x) = Q(p) + S(p), where Q(p) = x6 − 0.3726x4 +
3584
+ 0.0574x3 + 0.0376x2, and S(p) = sx with varying parame-
3585
+ ter s. This is the modified version of Example 1. There are
3586
+ two concave domains, each one starts with a vertical dashed
3587
+ line and ends with a vertical straight line.
3588
+ Both the blue
3589
+ points and red points represent the global minimizer x∗(s),
3590
+ and they always appears at convex domain. Specifically, the
3591
+ blue point occurs at Escape Zone, means that at correspond-
3592
+ ing s the global minimizer x∗(s) can be obtained by inversely
3593
+ heat conduct algorithm. However, the red point appears at
3594
+ the Confinement Zone which means that this global mini-
3595
+ mizers x∗(s) can not be obtained immediately through the
3596
+ inverse heat conduct algorithm. However, this red part in
3597
+ confinement zone can still be accessed by solving seesaw dif-
3598
+ ferential equation (3.25) with initial global minimums from
3599
+ attainable zone in connected blue part.
3600
+ where the criterion function
3601
+ H(ξ, η) = 20(ξ3 + η3) + 30(ξ2η + ξη2)
3602
+ +15a(ξ2 + η2) + 20aξη
3603
+ + 10b(ξ + η) + 5c
3604
+ (5.67)
3605
+
3606
+ 2
3607
+ EscapeZone
3608
+ 1.5Concave
3609
+ 0.5
3610
+ S
3611
+ 0
3612
+ -0.5
3613
+ -1
3614
+ -1.5
3615
+ confinementZone
3616
+ -2
3617
+ -1
3618
+ -0.5
3619
+ 0
3620
+ 0.5
3621
+ *
3622
+ X1HEAT EVOLUTION
3623
+ 51
3624
+ -1.2
3625
+ -1
3626
+ -0.8
3627
+ -0.6
3628
+ -0.4
3629
+ -0.2
3630
+ 0
3631
+ 0.2
3632
+ 0.4
3633
+ 0.6
3634
+ 0.8
3635
+ x
3636
+ 0
3637
+ 0.01
3638
+ 0.02
3639
+ 0.03
3640
+ 0.04
3641
+ 0.05
3642
+ 0.06
3643
+ 0.07
3644
+ 0.08
3645
+ 0.09
3646
+ 0.1
3647
+ t
3648
+ * stands for global minimum
3649
+ Figure 18. An example of Fingerprint of p(x) = x6 +
3650
+ 0.6987x5 − 1.0908x4 − 0.4216x3 + 0.2177x2 + 0.1071x. Here
3651
+ ∗ stands for the global minimizer, and the Euler’s method
3652
+ along the critical point Fingerprint from large t will back-
3653
+ ward to the true global minimizer.
3654
+ If both ξ and η (where we suppose that ξ ̸= η) are real critical points of
3655
+ p(x|s),
3656
+ p(ξ|s) > p(η|s) ⇐⇒ (ξ − η)H(ξ, η) < 0.
3657
+ (5.68)
3658
+ Our interest is to find out the seesaw parameter s such that ξ(s) ̸= η(s)
3659
+ and
3660
+ min
3661
+ x p(x|s) = p(ξ(s)|s) = p(η(s)|s).
3662
+ (5.69)
3663
+ That is, the seesaw polynomial p(x|s) attains a state that occurs the jump
3664
+ phenomena: it possesses (at least) two global minimizers ξ(s) and η(s).
3665
+ 5.10. Numerical examples for 6-degree polynomials. It is extremely
3666
+ expected to generalize the heat evolution algorithm to find out global min-
3667
+ imizer of 6 or higher even degree polynomials. However, the Theorem 14
3668
+ can not be generalized to higher degree polynomials. Here we illustrate the
3669
+ positive and negative examples.
3670
+ Example 6. The fingerprint FP1 of p(x) = x6 + 0.6987x5 − 1.0908x4 −
3671
+ 0.4216x3 + 0.2177x2 + 0.1071x illustrated in Fig.18 shows that the global
3672
+ minimizer is included in the integral curve to convex p(x, t).
3673
+ Example 7. The fingerprint FP1 of p(x) = x6 − 0.8529x5 − 0.4243x4 −
3674
+ 0.2248x3 + 0.0916x2 − 0.0074x illustrated in Fig.19 shows that the global
3675
+ minimizer is NOT included in the integral curve to convex p(x, t).
3676
+
3677
+ 52
3678
+ QIAO WANG
3679
+ -1
3680
+ -0.8
3681
+ -0.6
3682
+ -0.4
3683
+ -0.2
3684
+ 0
3685
+ 0.2
3686
+ 0.4
3687
+ x
3688
+ 0
3689
+ 0.01
3690
+ 0.02
3691
+ 0.03
3692
+ 0.04
3693
+ 0.05
3694
+ 0.06
3695
+ 0.07
3696
+ 0.08
3697
+ 0.09
3698
+ 0.1
3699
+ t
3700
+ * stands for global minimum
3701
+ Figure 19. A counter-example of fingerprint of p(x) = x6 −
3702
+ 0.8529x5−0.4243x4−0.2248x3+0.0916x2−0.0074x. Here the
3703
+ ∗ stands for the global minimizer, but the most right curve
3704
+ started from large t > 0, connected only to the local mini-
3705
+ mizer.
3706
+ 6. Conclusion
3707
+ In this paper, we investigate the possibility of finiding the global mini-
3708
+ mizer of a polynomial p(x) by inversely evolution from the global minimizer
3709
+ of its conxification version p(x, t) = p(x) ∗ gt(x). We propose the concepts
3710
+ of confinement zone and escape zone, as well as attainable zone, of the poly-
3711
+ nomial p(x).
3712
+ We apply Yuille-Poggio’s fingerprint theory including the Yuille-Poggio
3713
+ equation in computer vision to characterize the critical points of p(x, t),
3714
+ and propose a seesaw decomposition which produces the seesaw polynomial
3715
+ p(x|s). We further propose a seesaw differential equation to characterize
3716
+ the change of minimizers of p(x|s). Here, the fingerprint FP2 and FP3 are
3717
+ independent of seesaw parameter s, but the information of critical points of
3718
+ p(x|s) are contained in Yuille-Poggio’s flow.
3719
+ We showed in this paper that the global minimizer x∗ of a polynomial
3720
+ p(x) can be evolved inversely from the global minimizer of its conxification
3721
+ version p(x, t) = p(x) ∗ gt(x), if and only if this x∗ is in the escape zone
3722
+ of polynomial p(x). When x∗ is not in the escape zone, we may apply the
3723
+ seesaw equation by varying s through the global minimizer x∗(s) of p(x|s)
3724
+ to obtain x∗.
3725
+ However, the characterization of escape zone and attainable zone of a
3726
+ polynomial p(x) is in general algebraically not tractable, according to the
3727
+ Galois theory. Thus efficient numerical methods, as well as various criterions
3728
+ of judge the zones are extremely expected.
3729
+
3730
+ HEAT EVOLUTION
3731
+ 53
3732
+ Some results concerning the multivariate cases will be giving in our forth-
3733
+ coming works.
3734
+
3735
+ 54
3736
+ QIAO WANG
3737
+ -4
3738
+ -2
3739
+ 0
3740
+ 2
3741
+ 4
3742
+ 6
3743
+ 8
3744
+ -900
3745
+ -800
3746
+ -700
3747
+ -600
3748
+ -500
3749
+ -400
3750
+ -300
3751
+ -200
3752
+ -100
3753
+ 0
3754
+ 100
3755
+ x1(t)
3756
+ x3(t)
3757
+ x2(t)
3758
+ x1(t) and x2(t) meets up at tu
3759
+ x3(t) is the unique critical point when t>tu
3760
+ at tu
3761
+ Figure 20. The triangle series of critical points of evolution. Here the polynomial is p(x) = x4 − 8x3 −
3762
+ 18x2 + 56x, which is explained in Example 3.
3763
+
3764
+ HEAT EVOLUTION
3765
+ 55
3766
+ -4
3767
+ -2
3768
+ 0
3769
+ 2
3770
+ 4
3771
+ 6
3772
+ 8
3773
+ x
3774
+ 0
3775
+ 5
3776
+ 10
3777
+ 15
3778
+ scale t
3779
+ convexity fingerprint
3780
+ -4
3781
+ -2
3782
+ 0
3783
+ 2
3784
+ 4
3785
+ 6
3786
+ 8
3787
+ x
3788
+ 0
3789
+ 5
3790
+ 10
3791
+ 15
3792
+ scale t
3793
+ extreme fingerprint
3794
+ -4
3795
+ -2
3796
+ 0
3797
+ 2
3798
+ 4
3799
+ 6
3800
+ 8
3801
+ x
3802
+ 0
3803
+ 5
3804
+ 10
3805
+ 15
3806
+ scale t
3807
+ mixtured fingerprints
3808
+ Figure 21. Both the fingerprints FP1 and FP2 characterize the distribution of critical points and convexity
3809
+ of heat evolved version of a quartic polynomial.
3810
+
3811
+ 56
3812
+ QIAO WANG
3813
+ Appendix A. Real solutions of cubic equation
3814
+ A.1. Representation by roots. Recall the classical theory of cubic alge-
3815
+ braic equation (cf. [14])
3816
+ x3 + αx2 + βx + γ = 0,
3817
+ (A.1)
3818
+ According to Newton’s method, we have
3819
+ Lemma 14. Let xi (i = 1, 2, 3) be the roots (real or complex) of polynomial
3820
+ equation (A.1). We have the following propositions,
3821
+ x1 + x2 + x3 = −α,
3822
+ x1x2 + x2x3 + x3x1 = β,
3823
+ x1x2x3 = −γ,
3824
+ x2
3825
+ 1 + x2
3826
+ 2 + x2
3827
+ 3 = α2 − 2β,
3828
+ x3
3829
+ 1 + x3
3830
+ 2 + x3
3831
+ 3 = −α3 + 3αβ − 3γ,
3832
+ x4
3833
+ 1 + x4
3834
+ 2 + x4
3835
+ 3 = α4 − 4α2β + 4α2 + 2β2.
3836
+ (A.2)
3837
+ When the coefficients of the equation are real, the discriminant of the
3838
+ equation is
3839
+ ∆ = (x1 − x2)2(x2 − x3)2(x3 − x1)2,
3840
+ (A.3)
3841
+ which is equivalent to
3842
+ ∆ = g2
3843
+ 4 + f3
3844
+ 27,
3845
+ (A.4)
3846
+ in which
3847
+ f = β − α2
3848
+ 3
3849
+ and g = 2α3
3850
+ 27 − αβ
3851
+ 3 + γ.
3852
+ (A.5)
3853
+ A.2. The real roots described by discriminant. Then the solution of
3854
+ this cubic equation is as follows:
3855
+ If ∆ < 0, the equation (A.1) contains three distinct real roots.
3856
+ x1 =
3857
+ 2
3858
+
3859
+ 3
3860
+
3861
+ −f sin(θ) − α
3862
+ 3 ,
3863
+ x2 = − 2
3864
+
3865
+ 3
3866
+
3867
+ −f sin
3868
+
3869
+ θ + π
3870
+ 3
3871
+
3872
+ − α
3873
+ 3 ,
3874
+ x3 =
3875
+ 2
3876
+
3877
+ 3
3878
+
3879
+ −f cos
3880
+
3881
+ θ + π
3882
+ 6
3883
+
3884
+ − α
3885
+ 3 ,
3886
+ (A.6)
3887
+ where
3888
+ θ = 1
3889
+ 3 arcsin
3890
+
3891
+ 3
3892
+
3893
+ 3g
3894
+ 2(√−f)3
3895
+
3896
+ .
3897
+ (A.7)
3898
+ If ∆ = 0, the solutions contains a single root and two repeated roots,
3899
+ x1 = −2
3900
+ �g
3901
+ 2
3902
+ � 1
3903
+ 3 − α
3904
+ 3 ,
3905
+ x2 = x3 =
3906
+ �g
3907
+ 2
3908
+ � 1
3909
+ 3 − α
3910
+ 3 .
3911
+ (A.8)
3912
+
3913
+ HEAT EVOLUTION
3914
+ 57
3915
+ Finally, when ∆ > 0, the equation has only single real root
3916
+ x =
3917
+
3918
+ −g
3919
+ 2 +
3920
+
3921
+
3922
+ � 1
3923
+ 3 +
3924
+
3925
+ −g
3926
+ 2 −
3927
+
3928
+
3929
+ � 1
3930
+ 3 .
3931
+ (A.9)
3932
+ Appendix B. Real double roots of quartic polynomials and the
3933
+ structure of FP2
3934
+ � FP3
3935
+ We consider the regularized form of real quartic equation
3936
+ x4 + βx2 + γx + δ = 0.
3937
+ (B.1)
3938
+ The structure of its roots is described in [19]. It possesses repeated roots if
3939
+ and only if its discriminant ∆ = 0, where
3940
+ ∆ = 256γ3 − 128β2δ2 + 144βγ2δ − 27γ4 + 16β4δ − 4β3γ2.
3941
+ (B.2)
3942
+ In addition, we require another four polynomials,
3943
+ P =8β,
3944
+ R =8γ,
3945
+ ∆0 =β2 + 12δ,
3946
+ D =64δ − 16β2.
3947
+ (B.3)
3948
+ The equation (B.1) has one double real roots and two other distinct real
3949
+ roots, if and only if
3950
+ ∆ = 0 and P < 0 and D < 0 and ∆0 ̸= 0.
3951
+ (B.4)
3952
+ The equation (B.1) has one double real roots and a pair of complex roots,
3953
+ if and only if
3954
+ (∆ = 0 and D > 0) or (∆ = 0 and P > 0 and (D ̸= 0 or R ̸= 0)).
3955
+ (B.5)
3956
+ To apply the above results to the quartic equation (5.23), we may repre-
3957
+ sent the variables according to
3958
+ β =2b
3959
+ 5 + 6t,
3960
+ γ =c
3961
+ 5,
3962
+ δ = d
3963
+ 15 + 2b
3964
+ 5 t + 3t2.
3965
+ (B.6)
3966
+ Finally, we see that
3967
+ ∆(t) =
3968
+ 6
3969
+
3970
+ k=0
3971
+ c6−ktk,
3972
+ (B.7)
3973
+
3974
+ 58
3975
+ QIAO WANG
3976
+ in which
3977
+ c0 =27648,
3978
+ c1 =55296b
3979
+ 5
3980
+ ,
3981
+ c2 =9216b2
3982
+ 5
3983
+ ,
3984
+ c3 =4096b3 + 1728c2
3985
+ 25
3986
+ ,
3987
+ c4 =4864b4
3988
+ 625
3989
+ + 512db2 + 1728bc2
3990
+ 125
3991
+ − 256d2
3992
+ 25
3993
+ ,
3994
+ c5 =32(b2 + 5d)(48b3 − 80db + 135c2)
3995
+ 9375
3996
+ ,
3997
+ c6 =256b4d
3998
+ 9375 − 32b3c2
3999
+ 3125 − 512b2d2
4000
+ 5625
4001
+ + 96bc2d
4002
+ 625
4003
+ − 27c4
4004
+ 625 + 256d3
4005
+ 3375
4006
+ (B.8)
4007
+ Here we should explicitly represent the conditions in (B.4) (B.5). Actually,
4008
+ based on (B.6), we have
4009
+ P < 0 ⇐⇒ t < − b
4010
+ 15,
4011
+ R ̸= 0 ⇐⇒ c ̸= 0,
4012
+ ∆0 ̸= 0 ⇐⇒ d > b2
4013
+ 5 ,
4014
+ D > 0 (= 0, < 0, resp.) ⇐⇒
4015
+
4016
+ t + b
4017
+ 15
4018
+ �2
4019
+ − 1
4020
+ 90
4021
+
4022
+ d − b2
4023
+ 5
4024
+
4025
+ > 0 (= 0, < 0, resp.).
4026
+ (B.9)
4027
+ These implies that
4028
+ (B.4) ⇐⇒ − b
4029
+ 15 −
4030
+ 1
4031
+
4032
+ 90
4033
+
4034
+ d − b2
4035
+ 5 < t < − b
4036
+ 15, d − b2
4037
+ 5 > 0, and ∆(t) = 0.
4038
+ If we denote
4039
+ �t =
4040
+ 1
4041
+
4042
+ 90
4043
+
4044
+ d − b2
4045
+ 5 − b
4046
+ 15,
4047
+ then we see that
4048
+ • If d − b2
4049
+ 5 < 0, then (B.5) ⇐⇒ ∆(t) = 0.
4050
+ • If d − b2
4051
+ 5 = 0, then D > 0 is equivalent to t ̸= − b
4052
+ 15. Thus
4053
+ (B.5) ⇐⇒ t ̸= − b
4054
+ 15 and ∆(t) = 0.
4055
+ • If d − b2
4056
+ 5 > 0, we may classify it into two cases. At first case, if
4057
+ d ≥ 3b2
4058
+ 5 , then D > 0 is equivalent to
4059
+ 0 ≤ t < − b
4060
+ 15 −
4061
+ 1
4062
+
4063
+ 90
4064
+
4065
+ d − b2
4066
+ 5 ,
4067
+ or t > − b
4068
+ 15 +
4069
+ 1
4070
+
4071
+ 90
4072
+
4073
+ d − b2
4074
+ 5 .
4075
+ (B.10)
4076
+
4077
+ HEAT EVOLUTION
4078
+ 59
4079
+ Notice that
4080
+ D > 0 ∨ (P > 0 ∧ (D ̸= 0 ∨ R ̸= 0))
4081
+ =(D > 0 ∨ P > 0) ∧ (D > 0 ∨ D ̸= 0 ∨ R ̸= 0)
4082
+ =(D > 0 ∨ P > 0) ∧ (D ̸= 0 ∨ R ̸= 0)
4083
+ (B.11)
4084
+ which indicates that
4085
+ (B.5) ⇐⇒ ∆(t) = 0, and t ∈
4086
+
4087
+ 0, − b
4088
+ 15 −
4089
+ 1
4090
+
4091
+ 90
4092
+
4093
+ d − b2
4094
+ 5
4095
+ � � �
4096
+ − b
4097
+ 15, +∞
4098
+
4099
+ ,
4100
+ excluding that both D = 0 and c = 0.
4101
+ But if 3b2
4102
+ 5 > d > b2
4103
+ 5 , then D > 0 is equivalent to
4104
+ t > − b
4105
+ 15 +
4106
+ 1
4107
+
4108
+ 90
4109
+
4110
+ d − b2
4111
+ 5 .
4112
+ (B.12)
4113
+ then (B.5) ⇐⇒ ∆(t) = 0 and t > − b
4114
+ 15 excluding that both D = 0
4115
+ and c = 0.
4116
+ In a summary, we can obtain at 0 ≤ t1 ≤ t2, respectively corresponds to
4117
+ a dual real roots x1 and x2 of the equation (B.1), thus
4118
+ {(x1, t1), (x2, t2)} = FP2
4119
+
4120
+ FP3,
4121
+ in which t2 ≥ t1 and 0 ≤ t1 < − b
4122
+ 15.
4123
+ References
4124
+ [1] J. B. Lasserre, Global optimization with polynomials and the problem of moments,
4125
+ SIAM J. Optim. Vol. 11(3), pp:796-817, 2001.
4126
+ [2] N.Z. Shor, Quadratic optimization problems, Soviet J. Comput. Systems Sci.,
4127
+ 25(1987), pp:1-11.
4128
+ [3] N.Z. Shor, Nondifferentiable optimization and polynomial problems, Kluwer Aca-
4129
+ demic Publishers, 1998
4130
+ [4] V.N. Nefedov, Polynomial optimization problem, U.S.S.R. Comput. Maths. Math.
4131
+ Phys., Vol.27, No.3, pp.l3-21, 1987
4132
+ [5] J. Zhu and X. Zhang, On global optimizations with polynomials, Optimization Let-
4133
+ ters, (2008)2: 239-249.
4134
+ [6] J. Zhu, S. Zhao and G. Liu, Solution to global minimization of polynomials by back-
4135
+ ward differential flow, J. Optim Theory Appl (2014)161: 828-836.
4136
+ [7] O. Arikan, R.S. Burachik and C.Y. Kaya, ”Backward differential flow” may not con-
4137
+ verge to a global minimizer of polynomials, J. Optim Theory Appl (2015): 167:
4138
+ 401-408.
4139
+ [8] O. Arikan, R.S. Burachik and C.Y. Kaya, Steklov regularization and trajectory meth-
4140
+ ods for univariate global optimization, J. Global Optimization, 2019.
4141
+ [9] Burachik, R.S., Kaya, C.Y. Steklov convexification and a trajectory method for
4142
+ global optimization of multivariate quartic polynomials. Math. Program. (2020).
4143
+ https://doi.org/10.1007/s10107-020-01536-8
4144
+ [10] T. Iijima, basic theory of pattern observation, Papers of Tech. Group on Automata
4145
+ and Automatic Control, IEICE, Japan, 1959 (in Japanese).
4146
+ [11] T. Iijima, basic theory on normalization of pattern (in case of typical one-dimensional
4147
+ pattern), Bulletin of the Electrotechnical Lab., Vol.26: 368-388, 1962.
4148
+ [12] A. L. Yuille and T. Poggio, Fingerprints theorems for zero crossings, J. Opt. Soc.
4149
+ Am. A, 2(5): 683-692, 1985. doi = 10.1364/JOSAA.2.000683
4150
+
4151
+ 60
4152
+ QIAO WANG
4153
+ [13] A. L. Yuille and T. A. Poggio, Scaling theorems for zero crossings, IEEE Trans.
4154
+ on Pattern Analysis and Machine Intelligence, vol.8(1), pp. 15-25, Jan. 1986, doi:
4155
+ 10.1109/TPAMI.1986.4767748.
4156
+ [14] E. Zeidler, Oxford users’ guide to mathematics, Oxford Univ. Press, 2013.
4157
+ [15] Irving Kaplansky, an introduction to differential algebra, Hermann, Paris, 1957.
4158
+ [16] Fritz John, Partial differential equations, Springer-Verlag, 4.ed., 1982.
4159
+ [17] W.A. Strauss, Partial differential equations. John Wiley and Sons Inc. 1992.
4160
+ [18] I.N. Stewart, Galois theory, Chapman and Hall/CRC, 4.ed., 2015.
4161
+ [19] E. L. Rees, Graphical discussion of the roots of a quartic equation, The American
4162
+ Mathematical Monthly, 29:2, 51-55, 1922.
4163
+ [20] Birkhoff, G. and Mac Lane, S. A survey of modern algebra, 3rd ed. New York:
4164
+ Macmillan, 1965.
4165
+ [21] Nickalls, R.W.D, The quartic equation: invariants and Euler’s solution revealed, The
4166
+ Mathematical Gazette, vol.93(526), 66-75, 2009.
4167
+ [22] Louis Nirenberg, A strong maximum principle for parabolic equations, Comm. Pure.
4168
+ Appl. Math., Vol.6, 167-177, 1953.
4169
+ School of Information Science and Engineering, Southeast University, Nan-
4170
+ jing, 210096, China
4171
+ Email address: qiaowang@seu.edu.cn
4172
+
ENAyT4oBgHgl3EQfevhN/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
EtAyT4oBgHgl3EQfevi1/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9554443d87ef9fe00060322ab2145a672ca09ab57c9f4c5c697702c4e141892b
3
+ size 73445
EtFLT4oBgHgl3EQfFS_K/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3fdbd829faf5cbe3e32713581efb3964d7724b56d29e36d7c7b944a636461970
3
+ size 4128813
FNAzT4oBgHgl3EQfG_ve/content/tmp_files/2301.01039v1.pdf.txt ADDED
@@ -0,0 +1,1143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.01039v1 [math.CA] 3 Jan 2023
2
+ BRASS-STANCU-KANTOROVICH OPERATORS ON A
3
+ HYPERCUBE∗
4
+ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA
5
+ This study is dedicated to Professor Ioan Ra¸sa on the occasion of his 70th birthday
6
+ Abstract. We deal with multivariate Brass-Stancu-Kantorovich oper-
7
+ ators depending on a non-negative integer parameter and defined on
8
+ the space of all Lebesgue integrable functions on a unit hypercube. We
9
+ prove Lp-approximation and provide estimates for the Lp-norm of the
10
+ error of approximation in terms of a multivariate averaged modulus of
11
+ continuity and of the corresponding Lp-modulus.
12
+ 1. Introduction and Historical Notes
13
+ The fundamental functions of the well-known Bernstein operators are
14
+ defined by
15
+ pn,k(x) =
16
+ � �n
17
+ k
18
+
19
+ xk(1 − x)n−k;
20
+ 0 ≤ k ≤ n
21
+ 0;
22
+ k < 0 or k > n
23
+ , x ∈ [0, 1].
24
+ (1.1)
25
+ In [23], using a probabilistic method, Stancu generalized Bernstein’s funda-
26
+ mental functions as
27
+ wn,k,r(x) :=
28
+
29
+
30
+
31
+ (1 − x) pn−r,k (x) ;
32
+ 0 ≤ k < r
33
+ (1 − x) pn−r,k (x) + xpn−r,k−r (x) ;
34
+ r ≤ k ≤ n − r
35
+ xpn−r,k−r (x) ;
36
+ n − r < k ≤ n
37
+ , x ∈ [0, 1],
38
+ (1.2)
39
+ where r is a non-negative integer parameter, n is any natural number such
40
+ that n > 2r, for which each pn−r,k is given by (1.1), and therefore, con-
41
+ structed and studied Bernstein-type positive linear operators as
42
+ Ln,r (f; x) :=
43
+ n
44
+
45
+ k=0
46
+ wn,k,r(x)f
47
+ �k
48
+ n
49
+
50
+ ,
51
+ x ∈ [0, 1],
52
+ (1.3)
53
+ for f ∈ C[0, 1]. In doing so Stancu was guided by an article of Brass [8].
54
+ This is further discussed by Gonska [11]. Among others, estimates in terms
55
+ of the second order modulus of smoothness are given there for continuous
56
+ functions.
57
+ It is clear that for x ∈ [0, 1] Stancu’s fundamental functions in (1.2) satisfy
58
+ wn,k,r(x) ≥ 0 and
59
+ n
60
+
61
+ k=0
62
+ wn,k,r(x) = 1,
63
+ Key words and phrases. Multivariate Kantorovich operator; Multivariate averaged
64
+ modulus of smoothness; Multivariate K-functional
65
+ 2010 MSC: 41A36, 41A25, 26A45
66
+ ∗This paper is an extension of a talk given in ICATA 2022.
67
+ 1
68
+
69
+ 2
70
+ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA
71
+ hence the operators Ln,r can be expressed as
72
+ Ln,r (f; x) :=
73
+ n−r
74
+
75
+ k=0
76
+ pn−r,k (x)
77
+
78
+ (1 − x) f
79
+ �k
80
+ n
81
+
82
+ + xf
83
+ �k + r
84
+ n
85
+ ��
86
+ ,
87
+ (1.4)
88
+ are defined for n ≥ r and satisfy the end point interpolation Ln,r (f; 0) =
89
+ f (0) , Ln,r (f; 1) = f (1).
90
+ It thus seems to be justified to call the Ln,r
91
+ Brass-Stancu-Bernstein (BSB) operators.
92
+ In [24] Stancu gave uniform convergence limn→∞ Ln,r (f) = f on [0, 1] for
93
+ f ∈ C[0, 1] and presented an expression for the remainder Rn,r(f; x) of the
94
+ approximation formula f(x) = Ln,r(f; x) + Rn,r(f; x) by means of second
95
+ order divided differences and also obtained an integral representation for
96
+ the remainder. Moreover, the author estimated the order of approximation
97
+ by the operators Ln,r (f) via the classical modulus of continuity. He also
98
+ studied the spectral properties of Ln,r.
99
+ In the cases r = 0 and r = 1, the operators Ln,r reduce to the classical
100
+ Bernstein operators Bn, i.e.,
101
+ Bn (f; x) =
102
+ n
103
+
104
+ k=0
105
+ pn,k(x)f
106
+ �k
107
+ n
108
+
109
+ .
110
+ What also has to be mentioned: Stancu himself in his 1983 paper observed
111
+ that ”we can optimize the error bound of the approximation of the function
112
+ f by means of Ln,rf if we take r = 0 or r = 1, when the operator Ln,r
113
+ reduces to Bernstein’s.” So there is a shortcoming.
114
+ Since Bernstein polynomials are not appropriate for approximation of
115
+ discontinuous functions (see [14, Section 1.9]), by replacing the point evalu-
116
+ ations f
117
+ � k
118
+ n
119
+
120
+ with the integral means over small intervals around the knots
121
+ k
122
+ n, Kantorovich [12] generalized the Bernstein operators as
123
+ Kn (f; x) =
124
+ n
125
+
126
+ k=0
127
+ pn,k (x) (n + 1)
128
+ k+1
129
+ n+1
130
+
131
+ k
132
+ n+1
133
+ f (t) dt,
134
+ x ∈ [0, 1], n ∈ N,
135
+ (1.5)
136
+ for Lebesgue integrable functions f on [0, 1].
137
+ On p. 239 of his mathematical memoirs [13] Kantorovich writes: ”While
138
+ I was waiting for a student who was late, I was looking over vol. XIII of
139
+ Fundamenta Math. and saw in it a note from the Moscow Mathematician
140
+ Khlodovskii related to Bernstein polynomials. In it I first caught sight of
141
+ Bernstein polynomials, which he proposed in 1912 for an elementary proof
142
+ of the well known Weierstrass theorem ... I at once wondered if it is not
143
+ possible in these polynomials to change the values of the function at certain
144
+ points into the more stable average of the function in the corresponding
145
+ interval. It turned out that this was possible, and the polynomials could be
146
+ written in such a form not only for a continuous function but also for any
147
+ Lebesgue-summable function.”
148
+ Lorentz [14] proved that lim
149
+ n→∞ ∥Kn(f) − f∥p = 0, f ∈ Lp[0, 1], 1 ≤ p < ∞.
150
+ There are a lot of articles dealing with classical Kantorovich operators,
151
+ and, in particular, their degree of approximation and the importance of
152
+
153
+ BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE
154
+ 3
155
+ second order moduli of different types. See, e.g., the work of Berens and
156
+ DeVore [5], [6], Swetits and Wood [25] and Gonska and Zhou [10].
157
+ It is
158
+ beyond the scope of this note to further discuss this matter.
159
+ As further
160
+ work on the classical case here we only mention the 1976 work of M¨uller
161
+ [16], Maier [15], and Altomare et al. [1], see also the references therein.
162
+ Similarly to Kantorovich operators Bodur et al. [7] constructed a Kan-
163
+ torovich type modification of BSB operators as
164
+ Kn,r (f; x) :=
165
+ n
166
+
167
+ k=0
168
+ wn,k,r(x)
169
+
170
+
171
+
172
+ (n + 1)
173
+ k+1
174
+ n+1
175
+
176
+ k
177
+ n+1
178
+ f (t) dt
179
+
180
+
181
+
182
+  ,
183
+ x ∈ [0, 1],
184
+ (1.6)
185
+ for f ∈ L1 [0, 1], where r is a non-negative integer parameter, n is a natural
186
+ number such that n > 2r and wn,k,r(x) are given by (1.2). And, it was
187
+ shown that If f ∈ Lp[0, 1], 1 ≤ p < ∞, then
188
+ lim
189
+ n→∞ ∥Kn,r(f) − f∥p = 0.
190
+ In addition, it was obtained that each Kn,r is variation detracting as well
191
+ [7]. Throughout the paper, we shall call the operators Kn,r given by (1.6)
192
+ ”Brass-Stancu-Kantorovich”, BSK operators.
193
+ Notice that from the definition of wn,k,r, Kn,r (f; x) can be expressed as
194
+ Kn,r (f; x)
195
+ (1.7)
196
+ =
197
+ n−r
198
+
199
+ k=0
200
+ pn−r,k (x) (n + 1)
201
+
202
+ (1 − x)
203
+ k+1
204
+ n+1
205
+
206
+ k
207
+ n+1
208
+ f (t) dt + x
209
+ k+r+1
210
+ n+1
211
+
212
+ k+r
213
+ n+1
214
+ f (t) dt
215
+
216
+ 
217
+ and in the cases r = 0 and r = 1 they reduce to the Kantorovich operators;
218
+ Kn,0 = Kn,1 = Kn given by (1.5). Again they are defined for all n ≥ r.
219
+ MULTIVARIATE SITUATION
220
+ Some work has been done in the multivariate setting for BSB and BSK
221
+ operators. For the standard simplex this was done, e.g., by Yang, Xiong
222
+ and Cao [27] and Cao [9], For example, Cao proved that multivariate Stancu
223
+ operators preserve the properties of multivariate moduli of continuity and
224
+ obtained the rate of convergence with the help of Ditzian-Totik’s modulus
225
+ of continuity.
226
+ In this work, motivated by the work Altomare et al. [3], we deal with
227
+ a multivariate extension of the BSK operators on a d-dimensional unit hy-
228
+ percube and we study Lp -approximation by these operators. For the rate
229
+ of convergence we provide an estimate in terms of the so called first order
230
+ multivariate τ-modulus, a quantity coming from the Bulgarian school of
231
+ Approximation Theory. Also, inspired by M¨uller’s approach in [17], we give
232
+ estimates for differentiable functions and such in terms of the Lp-modulus
233
+ of smoothness, using properties of the τ-modulus. Here the work of Quak
234
+ [20], [21] was helpful.
235
+
236
+ 4
237
+ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA
238
+ 2. Preliminaries
239
+ Consider the space Rd, d ∈ N.
240
+ Let ∥x∥∞ denote the max-norm of a
241
+ point x = (x1, . . . , xd) ∈ Rd;
242
+ ∥x∥∞ := ∥x∥max =
243
+ max
244
+ i∈{1,...,d} |xi|
245
+ and let 1 denote the constant function 1 : Rd → R such that 1 (x) = 1 for
246
+ x ∈ Rd. And, for each j = 1, . . . , d, let
247
+ prj : Rd → R
248
+ stand for the jth coordinate function defined for x ∈ Rd by
249
+ prj (x) = xj.
250
+ Definition 2.1. A multi-index is a d-tuple α = (α1, . . . , αd) of non-negative
251
+ integers. Its norm (length) is the quantity
252
+ |α| =
253
+ d
254
+
255
+ i=1
256
+ αi.
257
+ The differential operator Dα is defined by
258
+ Dαf = Dα1
259
+ 1 · · · Dαd
260
+ d f,
261
+ where Di, i = 1, . . . , d, is the corresponding partial derivative operator (see
262
+ [4, p. 335]).
263
+ Throughout the paper Qd := [0, 1]d, d ∈ N, will denote the d-dimensional
264
+ unit hypercube and we consider the space
265
+ Lp (Qd) = {f : Qd → R | f p-integrable on Qd} , 1 ≤ p < ∞,
266
+ with the standard norm ∥.∥p. Recall the following definition of the usual
267
+ Lp-modulus of smoothness of first order:
268
+ Definition 2.2. Let f ∈ Lp (Qd) , 1 ≤ p < ∞, h ∈ Rd and δ > 0. The
269
+ modulus of smoothness of the first order for the function f and step δ in
270
+ Lp-norm is given by
271
+ ω1 (f; δ)p =
272
+ sup
273
+ 0<∥h∥∞≤δ
274
+
275
+
276
+
277
+
278
+ Qd
279
+ |f (x + h) − f (x)|p dx
280
+
281
+
282
+
283
+ 1/p
284
+ if x, x + h ∈ Qd [21].
285
+ Let M (Qd) := {f | f bounded and measurable on Qd}. Below, we present
286
+ the concept of the first order averaged modulus of smoothness.
287
+ Definition 2.3. Let f ∈ M (Qd) , h ∈ Rd and δ > 0. The multivariate
288
+ averaged modulus of smoothness, or τ-modulus, of the first order for function
289
+ f and step δ in Lp-norm is given by
290
+ τ 1 (f, δ)p := ∥ω1 (f, .; δ)∥p , 1 ≤ p < ∞,
291
+ where
292
+ ω1 (f, x; δ) =
293
+ sup
294
+
295
+ |f (t + h) − f (t)| : t, t + h ∈ Qd, ∥t − x∥∞ ≤ δ
296
+ 2, ∥t + h − x∥∞ ≤ δ
297
+ 2
298
+
299
+
300
+ BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE
301
+ 5
302
+ is the multivariate local modulus of smoothness of first order for the function
303
+ f at the point x ∈ Qd and for step δ. [21].
304
+ For our future purposes, we need the following properties of first order
305
+ multivariate averaged modulus of smoothness:
306
+ For f ∈ M (Qd) , 1 ≤ p < ∞ and δ, λ, γ ∈ R+, there hold
307
+ τ 1) τ 1 (f, δ)p ≤ τ 1 (f, λ)p for 0 < δ ≤ λ,
308
+ τ 2) τ 1 (f, λδ)p ≤ (2 ⌊λ⌋ + 2)d+1 τ 1 (f, δ)p, where ⌊λ⌋ is the greatest inte-
309
+ ger that does not exceed λ,
310
+ τ 3) τ 1 (f, δ)p ≤ 2 �
311
+ |α|≥1
312
+ δ|α| ∥Dαf∥p , αi = 0 or 1, if Dαf ∈ Lp (Qd) for
313
+ all multi-indices α with |α| ≥ 1 and αi = 0 or 1 (see [19] or [21]).
314
+ For a detailed knowledge concerning averaged modulus of smoothness, we
315
+ refer to the book of Sendov and Popov [22].
316
+ Now, consider the Sobolev space W p
317
+ 1 (Qd) of functions f ∈ Lp (Qd) , 1 ≤
318
+ p < ∞, with (distributional) derivatives Dαf belong to Lp (Qd), where
319
+ |α| ≤ 1, with the seminorm
320
+ |f|W p
321
+ 1 =
322
+
323
+ |α|=1
324
+ ∥Dαf∥p
325
+ (see [4, p. 336]). Recall that for all f ∈ Lp (Qd) the K-functional, in Lp-
326
+ norm, is defined as
327
+ K1,p (f; t) := inf
328
+
329
+ ∥f − g∥p + t |g|W p
330
+ 1 : g ∈ W p
331
+ 1 (Qd)
332
+
333
+ (t > 0) .
334
+ (2.1)
335
+ K1,p (f; t) is equivalent with the usual first order modulus of smoothness of
336
+ f, ω1 (f; t)p; namely, there are positive constants c1 and c2 such that
337
+ c1K1,p (f; t) ≤ ω1 (f; t)p ≤ c2K1,p (f; t)
338
+ (t > 0)
339
+ (2.2)
340
+ holds for all f ∈ Lp (Qd) (see [4, Formula 4.42 in p. 341]).
341
+ The following result due to Quak [21] is an upper estimate for the Lp-norm
342
+ of the approximation error by the multivariate positive linear operators in
343
+ terms of the first order averaged modulus of smoothness. Note that this
344
+ idea was used first by Popov for the univariate case in [18].
345
+ Theorem 2.1. Let L : M (Qd) → M (Qd) be a positive linear operator that
346
+ preserves the constants. Then for every f ∈ M (Qd) and 1 ≤ p < ∞, the
347
+ following estimate holds:
348
+ ∥L(f) − f∥p ≤ Cτ1
349
+
350
+ f,
351
+ 2d√
352
+ A
353
+
354
+ p ,
355
+ where C is a positive constant and
356
+ A := sup
357
+
358
+ L
359
+
360
+ (pri ◦ ψx)2 ; x
361
+
362
+ : i = 1, . . . , d, x ∈ Qd
363
+
364
+ ,
365
+ in which ψx (y) := y − x for fixed x ∈ Qd and for every y ∈ Qd and
366
+ A ≤ 1 [21].
367
+
368
+ 6
369
+ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA
370
+ 3. Multivariate BSK-Operators
371
+ In this section, motivated by the works of Altomare et al. [1] and Al-
372
+ tomare et al. [3], we consider the multivariate extension of BSK-operators
373
+ on Lp (Qd) and study approximation properties of these operators in Lp-
374
+ norm. We investigate the rate of the convergence in terms of the first order
375
+ τ-modulus and the usual Lp-modulus of smoothness of the first order.
376
+ Let r be a given non-negative integer.
377
+ For any n ∈ N such that n >
378
+ 2r, k = (k1, . . . , kd) ∈ {0, . . . , n}d and x = (x1, . . . , xd) ∈ Qd, we set
379
+ wn,k,r(x) :=
380
+ d
381
+
382
+ i=1
383
+ wn,ki,r(xi),
384
+ (3.1)
385
+ where, wn,ki,r(xi) is Stancu’s fundamental function given by (1.2), written
386
+ for each i = 1, . . . , d, 0 ≤ ki ≤ n and xi ∈ [0, 1]. Thus, for x ∈ Qd, we have
387
+ wn,k,r(x) ≥ 0 and
388
+
389
+ k∈{0,...,n}d
390
+ wn,k,r(x) = 1.
391
+ (3.2)
392
+ For f ∈ L1 (Qd) and x = (x1, . . . , xd) ∈ Qd we consider the following
393
+ multivariate extension of the BSK-operators Kn,r given by (1.6):
394
+ Kd
395
+ n,r (f; x) =
396
+ n
397
+
398
+ k1,...,kd=0
399
+ d
400
+
401
+ i=1
402
+ wn,ki,r(xi)
403
+
404
+ Qd
405
+ f
406
+ �k1 + u1
407
+ n + 1 , . . . , kd + ud
408
+ n + 1
409
+
410
+ du1 · · · dud.
411
+ Notice that from (3.1), and denoting, as usual, any f ∈ L1 (Qd) of x =
412
+ (x1, . . . , xd) ∈ Qd by f (x) = f (x1, . . . , xd), we can express these operators
413
+ in compact form as
414
+ Kd
415
+ n,r (f; x) =
416
+
417
+ k∈{0,...,n}d
418
+ wn,k,r(x)
419
+
420
+ Qd
421
+ f
422
+ �k + u
423
+ n + 1
424
+
425
+ du.
426
+ (3.3)
427
+ It is clear that multivariate BSK-operators are positive and linear and the
428
+ cases r = 0 and 1 give the multivariate Kantorovich operators on the hyper-
429
+ cube Qd, which can be captured from [1] as a special case.
430
+ Lemma 3.1. For x ∈ Qd, we have
431
+ Kd
432
+ n,r (1; x)
433
+ =
434
+ 1,
435
+ Kd
436
+ n,r (pri; x)
437
+ =
438
+ n
439
+ n + 1xi +
440
+ 1
441
+ 2 (n + 1),
442
+ Kd
443
+ n,r
444
+
445
+ pr2
446
+ i ; x
447
+
448
+ =
449
+ n2
450
+ (n + 1)2
451
+
452
+ x2
453
+ i +
454
+
455
+ 1 + r (r − 1)
456
+ n
457
+ � xi (1 − xi)
458
+ n
459
+
460
+ + 3nxi + 1
461
+ 3 (n + 1)2 ,
462
+ for i = 1, . . . , d.
463
+ Taking this lemma into consideration, by the well-known theorem of
464
+ Volkov [26], we immediately get that
465
+ Theorem 3.1. Let r be a non-negative fixed integer and f ∈ C (Qd). Then
466
+ lim
467
+ n→∞ Kd
468
+ n,r (f) = f uniformly on Qd.
469
+
470
+ BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE
471
+ 7
472
+ Now, we need the following evaluations for the subsequent result: For
473
+ 0 ≤ xi ≤ 1, i = 1, . . . , d, we have
474
+ 1
475
+
476
+ 0
477
+ (1 − xi) pn−r,ki (xi) dxi
478
+ =
479
+ �n − r
480
+ ki
481
+
482
+ 1
483
+
484
+ 0
485
+ xki
486
+ i (1 − xi)n−r−ki+1 dxi
487
+ =
488
+ n − r − ki + 1
489
+ (n − r + 2) (n − r + 1)
490
+ when 0 ≤ ki < r and
491
+ 1
492
+
493
+ 0
494
+ xipn−r,ki−r (xi) dxi
495
+ =
496
+ �n − r
497
+ ki − r
498
+
499
+ 1
500
+
501
+ 0
502
+ xki−r+1
503
+ i
504
+ (1 − xi)n−ki dxi
505
+ =
506
+ ki − r + 1
507
+ (n − r + 2) (n − r + 1)
508
+ when n − r < ki ≤ n. Thus, from (1.1) and (1.2), it follows that
509
+ 1
510
+
511
+ 0
512
+ wn,ki,r(xi)dxi =
513
+
514
+
515
+
516
+
517
+
518
+ n−r−ki+1
519
+ (n−r+2)(n−r+1);
520
+ 0 ≤ ki < r
521
+ n−2r+2
522
+ (n−r+2)(n−r+1);
523
+ r ≤ ki ≤ n − r
524
+ ki−r+1
525
+ (n−r+2)(n−r+1);
526
+ n − r < ki ≤ n
527
+ .
528
+ (3.4)
529
+ Note that we can write the following estimates
530
+ n − r − ki + 1
531
+
532
+ n − r + 1 when 0 ≤ ki < r,
533
+ n − 2r + 2
534
+
535
+ n − r + 1 when r ≤ ki ≤ n − r,
536
+ ki − r + 1
537
+
538
+ n − r + 1 when n − r < ki ≤ n
539
+ (3.5)
540
+ for each i = 1, . . . , d, where in the middle term, we have used the hypothesis
541
+ n > 2r. Making use of (3.5), (3.4) and (3.1), we obtain
542
+
543
+ Qd
544
+ wn,k,r(x)dx =
545
+ d
546
+
547
+ i=1
548
+ 1
549
+
550
+ 0
551
+ wn,ki,r(xi)dxi ≤
552
+ 1
553
+ (n − r + 2)d .
554
+ (3.6)
555
+ Lp-approximation by the sequence of the multivariate Stancu-Kantorovich
556
+ operators is presented in the following theorem.
557
+ Theorem 3.2. Let r be a non-negative fixed integer and f ∈ Lp (Qd) , 1 ≤
558
+ p < ∞. Then lim
559
+ n→∞
560
+ ��Kd
561
+ n,r(f) − f
562
+ ��
563
+ p = 0.
564
+ Proof. Since the cases r = 0 and 1 correspond to the multivariate Kan-
565
+ torovich operators (see [1] or [3]), we consider only the cases r > 1, which is
566
+ taken as fixed. From Theorem 3.1, we obtain that lim
567
+ n→∞
568
+ ��Kd
569
+ n,r(f) − f
570
+ ��
571
+ p =
572
+ 0 for any f ∈ C (Qd). Since C (Qd) is dense in Lp (Qd), denoting the norm
573
+ of the operator Kd
574
+ n,r acting on Lp (Qd) onto itself by
575
+ ��Kd
576
+ n,r
577
+ ��, it remains to
578
+ show that there exists an Mr, where Mr is a positive constant that maybe
579
+ depends on r, such that
580
+ ��Kd
581
+ n,r
582
+ �� ≤ Mr for all n > 2r. Now, as in [3, p.604],
583
+ we adopt the notation
584
+ Qn,k :=
585
+ d
586
+
587
+ i=1
588
+
589
+ ki
590
+ n + 1, ki + 1
591
+ n + 1
592
+
593
+ ⊂ Qd;
594
+
595
+ k∈{0,...,n}d
596
+ Qn,k = Qd.
597
+
598
+ 8
599
+ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA
600
+ Making use of the convexity of the function ϕ (t) := |t|p , t ∈ R, 1 ≤ p <
601
+ ∞ (see, e.g., [2]), and (3.2), for every f ∈ Lp (Qd) , n > 2r, and x ∈ Qd, we
602
+ obtain
603
+ ���Kd
604
+ n,r (f; x)
605
+ ���
606
+ p
607
+
608
+
609
+ k∈{0,...,n}d
610
+ wn,k,r(x)
611
+
612
+ Qd
613
+ ����f
614
+ �k + u
615
+ n + 1
616
+ �����
617
+ p
618
+ du
619
+ =
620
+
621
+ k∈{0,...,n}d
622
+ wn,k,r(x) (n + 1)d
623
+
624
+ Qn,k
625
+ |f (v)|p dv.
626
+ Taking (3.6) into consideration, we reach to
627
+
628
+ Qd
629
+ ���Kd
630
+ n,r (f; x)
631
+ ���
632
+ p
633
+ dx ≤
634
+
635
+ k∈{0,...,n}d
636
+
637
+ n + 1
638
+ n − r + 2
639
+ �d �
640
+ Qn,k
641
+ |f (v)|p dv.
642
+ Since sup
643
+ n>2r
644
+
645
+ n+1
646
+ n−r+2
647
+ �d
648
+ =
649
+
650
+ 2r+2
651
+ r+3
652
+ �d
653
+ := Mr for r > 1, where 1 < 2r+2
654
+ r+3 < 2, we
655
+ get
656
+
657
+ Qd
658
+ ���Kd
659
+ n,r (f; x)
660
+ ���
661
+ p
662
+ dx ≤ Mr
663
+
664
+ Qd
665
+ |f (v)|p dv,
666
+ which implies that
667
+ ��Kd
668
+ n,r (f)
669
+ ��
670
+ p ≤ M1/p
671
+ r
672
+ ∥f∥p. Note that for the cases r = 0
673
+ and 1; we have Mr = 1 (see [3]). Therefore, the proof is completed.
674
+
675
+ 4. Estimates for the rate of convergence
676
+ In [17], M¨uller studied Lp-approximation by the sequence of the Cheney-
677
+ Sharma-Kantorovich operators (CSK). The author gave an estimate for this
678
+ approximation in terms of the univariate τ-modulus and moreover, using
679
+ some properties of the τ-modulus, he also obtained upper estimates for the
680
+ Lp-norm of the error of approximation for first order differentiable functions
681
+ as well as for continuous ones. In this part, we show that similar estimates
682
+ can also be obtained for
683
+ ��Kd
684
+ n,r (f) − f
685
+ ��
686
+ p in the multivariate setting. Our
687
+ first result is an application of Quak’s method in Theorem 2.1
688
+ Theorem 4.1. Let r be a non-negative fixed integer, f ∈ M (Qd) and 1 ≤
689
+ p < ∞. Then
690
+ ���Kd
691
+ n,r (f) − f
692
+ ���
693
+ p ≤ Cτ 1
694
+
695
+ f, 2d
696
+
697
+ 3n + 1 + 3r (r − 1)
698
+ 12 (n + 1)2
699
+
700
+ p
701
+ (4.1)
702
+ for all n ∈ N such that n > 2r, where the positive constant C does not
703
+ depend on f.
704
+ Proof. According to Theorem 2.1; by taking ψx (y) = y − x for fixed x ∈
705
+ Qd and for every y ∈ Qd, and defining
706
+ An,r := sup
707
+
708
+ Kd
709
+ n,r
710
+
711
+ (pri ◦ ψx)2 ; x
712
+
713
+ : i = 1, . . . , d, x ∈ Qd
714
+
715
+ ,
716
+ where (pri ◦ ψx)2 = pr2
717
+ i − 2xipri + x2
718
+ i 1, i = 1, . . . , d, we get the following
719
+ estimate
720
+ ���Kd
721
+ n,r (f) − f
722
+ ���
723
+ p ≤ Cτ1
724
+
725
+ f; 2d�
726
+ An,r
727
+
728
+
729
+ BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE
730
+ 9
731
+ for any f ∈ M (Qd), under the condition that An,r ≤ 1. Now, applying the
732
+ operators Kd
733
+ n,r and making use of Lemma 3.1, for every i = 1, . . . , d and
734
+ x ∈ Qd, we obtain
735
+ Kd
736
+ n,r
737
+
738
+ (pri ◦ ψx)2 ; x
739
+
740
+ =
741
+ n − 1 + r (r − 1)
742
+ (n + 1)2
743
+ xi (1 − xi) +
744
+ 1
745
+ 3 (n + 1)2
746
+
747
+ n − 1 + r (r − 1)
748
+ 4 (n + 1)2
749
+ +
750
+ 1
751
+ 3 (n + 1)2
752
+ =
753
+ 3n + 1 + 3r (r − 1)
754
+ 12 (n + 1)2
755
+ for all n ∈ N such that n > 2r, where r ∈ N ∪ {0}. Therefore, since we have
756
+ n ≥ 2r + 1, we take r ≤ n−1
757
+ 2
758
+ and obtain that An,r ≤ 3n+1+3r(r−1)
759
+ 12(n+1)2
760
+ ≤ 1 is
761
+ satisfied, which completes the proof.
762
+
763
+ Now, making use of the properties τ1)-τ3) of the multivariate first order
764
+ τ-modulus, we obtain
765
+ Theorem 4.2. Let r be a non-negative fixed integer, f ∈ Lp (Qd) , 1 ≤ p <
766
+ ∞, and Dαf ∈ Lp (Qd) for all multi-indices α with |α| ≥ 1, αi = 0 or 1.
767
+ Then
768
+ ���Kd
769
+ n,r (f) − f
770
+ ���
771
+ p ≤ 2Cr
772
+
773
+ |α|≥1
774
+
775
+ 1
776
+ 2d√n + 1
777
+ �|α|
778
+ ∥Dαf∥p ,
779
+ for all n ∈ N such that n > 2r, where Cr is a positive constant depending
780
+ on r.
781
+ Proof. Since n > 2r, we immediately have n + 1 ≥ 2 (r + 1). Thus, the
782
+ term appearing inside the 2dth root in the formula (4.1) can be estimated,
783
+ respectively, for r > 1, and r = 0, 1, as
784
+ 3n + 1 + 3r (r − 1)
785
+ 12 (n + 1)2
786
+ =
787
+ 3n + 3 + 3r (r − 1) − 2
788
+ 12(n + 1)2
789
+ =
790
+ 1
791
+ n + 1
792
+ �1
793
+ 4 + 3r (r − 1) − 2
794
+ 12(n + 1)
795
+
796
+
797
+ 1
798
+ n + 1
799
+ �1
800
+ 4 + 3r (r − 1) − 2
801
+ 24(r + 1)
802
+
803
+ =
804
+ 1
805
+ n + 1
806
+ �3r2 + 3r + 4
807
+ 24(r + 1)
808
+
809
+ and
810
+ 3n + 1
811
+ 12 (n + 1)2 =
812
+ 1
813
+ n + 1
814
+ 3n + 1
815
+ 4 (3n + 3) <
816
+ 1
817
+ 4 (n + 1).
818
+ Now, defining
819
+ Br :=
820
+
821
+ 3r2+3r+4
822
+ 24(r+1) ;
823
+ r > 1,
824
+ 1
825
+ 4;
826
+ r = 0, 1,
827
+
828
+ 10
829
+ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA
830
+ and making use of the properties τ 1)-τ 3) of τ-modulus, from (4.1), we arrive
831
+ at
832
+ ���Kd
833
+ n,r (f) − f
834
+ ���
835
+ p
836
+
837
+ Cτ1
838
+
839
+ f, 2d
840
+
841
+ 3n + 1 + 3r (r − 1)
842
+ 12 (n + 1)2
843
+
844
+ p
845
+
846
+ Cτ1
847
+ ���
848
+ f,
849
+ 2d�
850
+ Br
851
+ 1
852
+ 2d√n + 1
853
+
854
+ p
855
+
856
+ C
857
+
858
+ 2
859
+
860
+ 2d�
861
+ Br
862
+
863
+ + 2
864
+ �d+1
865
+ τ 1
866
+
867
+ f,
868
+ 1
869
+ 2d√n + 1
870
+
871
+ p
872
+
873
+ 2Cr
874
+
875
+ |α|≥1
876
+
877
+ 1
878
+ 2d√n + 1
879
+ �|α|
880
+ ∥Dαf∥p ,
881
+ where the positive constant Cr is defined as Cr := C
882
+
883
+ 2
884
+ � 2d√Br
885
+
886
+ + 2
887
+ �d+1 .
888
+
889
+ For non-differentiable functions we have the following estimate in terms
890
+ of the first order modulus of smoothness, in Lp-norm.
891
+ Theorem 4.3. Let r be a non-negative fixed integer and f ∈ Lp (Qd) , 1 ≤
892
+ p < ∞.
893
+ Then
894
+ ���Kd
895
+ n,r (f) − f
896
+ ���
897
+ p ≤ c2Cr,pω1
898
+
899
+ f;
900
+ 1
901
+ 2d√n + 1
902
+
903
+ p
904
+ ,
905
+ where ω1 is the first order multivariate modulus of smoothness of f and Cr,p
906
+ is a constant depending on r and p.
907
+ Proof. By Theorem 3.2, since Kd
908
+ n,r is bounded, with
909
+ ��Kd
910
+ n,r
911
+ ��
912
+ p ≤ M1/p
913
+ r
914
+ , for
915
+ all n ∈ N such that n > 2r, we have
916
+ ��Kd
917
+ n,r (g) − g
918
+ ��
919
+ p ≤
920
+
921
+ M1/p
922
+ r
923
+ + 1
924
+
925
+ ∥g∥p for
926
+ g ∈ Lp (Qd). Moreover, from Theorem 4.2, we can write
927
+ ���Kd
928
+ n,r (g) − g
929
+ ���
930
+ p ≤ 2Cr
931
+
932
+ |α|≥1
933
+
934
+ 1
935
+ 2d√n + 1
936
+ �|α|
937
+ ∥Dαg∥p
938
+ for those g such that Dαg ∈ Lp (Qd), for all multi-indices α with |α| ≥ 1
939
+ and αi = 0 or 1. Hence, for f ∈ Lp (Qd), it readily follows that
940
+ ���Kd
941
+ n,r (f) − f
942
+ ���
943
+ p
944
+
945
+ ���Kd
946
+ n,r (f − g) − (f − g)
947
+ ���
948
+ p +
949
+ ���Kd
950
+ n,r (g) − g
951
+ ���
952
+ p
953
+
954
+
955
+ M1/p
956
+ r
957
+ + 1
958
+
959
+
960
+
961
+ ∥f − g∥p + 2Cr
962
+
963
+ |α|≥1
964
+
965
+ 1
966
+ 2d√n + 1
967
+ �|α|
968
+ ∥Dαg∥p
969
+
970
+
971
+  .
972
+
973
+ BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE
974
+ 11
975
+ Passing to the infimum for all g ∈ W p
976
+ 1 (Qd) in the last formula, since the
977
+ infimum of a superset does not exceed that of subset, we obtain
978
+ ���Kd
979
+ n,r (f) − f
980
+ ���
981
+ p
982
+
983
+
984
+ M1/p
985
+ r
986
+ + 1
987
+
988
+ inf
989
+
990
+
991
+ ∥f − g∥p +
992
+ 2Cr
993
+ 2d√n + 1
994
+
995
+ |α|=1
996
+ ∥Dαg∥p : g ∈ W p
997
+ 1 (Qd)
998
+
999
+
1000
+
1001
+ =
1002
+
1003
+ M1/p
1004
+ r
1005
+ + 1
1006
+
1007
+ inf
1008
+
1009
+ ∥f − g∥p +
1010
+ 2Cr
1011
+ 2d√n + 1 |g|W p
1012
+ 1 : g ∈ W p
1013
+ 1 (Qd)
1014
+
1015
+ =
1016
+
1017
+ M1/p
1018
+ r
1019
+ + 1
1020
+
1021
+ K1,p
1022
+
1023
+ f;
1024
+ 2Cr
1025
+ 2d√n + 1
1026
+
1027
+ ,
1028
+ (4.2)
1029
+ where K1,p is the K-functional given by (2.1). The proof follows from the
1030
+ equivalence (2.2) of the K-functional and the first order modulus of smooth-
1031
+ ness in Lp-norm and the non-decreasingness property of the modulus. In-
1032
+ deed, we get
1033
+ K1,p
1034
+
1035
+ f;
1036
+ 2Cr
1037
+ 2d√n + 1
1038
+
1039
+
1040
+ c2ω1
1041
+
1042
+ f;
1043
+ 2Cr
1044
+ 2d√n + 1
1045
+
1046
+ p
1047
+
1048
+ c2 (2Cr + 1) ω1
1049
+
1050
+ f;
1051
+ 1
1052
+ 2d√n + 1
1053
+
1054
+ p
1055
+ .
1056
+ (4.3)
1057
+ Combining (4.3) with (4.2) and defining Cr,p :=
1058
+
1059
+ M1/p
1060
+ r
1061
+ + 1
1062
+
1063
+ (2Cr + 1),
1064
+ where M1/p
1065
+ r
1066
+ and Cr are the same as in Theorems 3.2 and 4.2, respectively,
1067
+ we obtain the desired result.
1068
+
1069
+ References
1070
+ [1] F. Altomare, M. Cappelletti Montano, V. Leonessa, On a generalization of Kan-
1071
+ torovich operators on simplices and hypercubes, Adv. Pure Appl. Math. 1 (2010), no.
1072
+ 3, 359-385.
1073
+ [2] F. Altomare, Korovkin-type Theorems and Approximation by Positive Linear Oper-
1074
+ ators, Surv. Approx. Theory 5 (2010), 92-164.
1075
+ [3] F. Altomare, M. Cappelletti Montano, V. Leonessa, I. Ra¸sa, A generalization of
1076
+ Kantorovich operators for convex compact subsets, Banach J. Math. Anal. 11 (2017),
1077
+ no. 3, 591–614.
1078
+ [4] C. Bennett, R. Sharpley, Interpolation of Operators, Academic Press Inc., 1988.
1079
+ [5] H. Berens, R. DeVore, Quantitative Korovkin theorems for Lp-spaces. In: Approx.
1080
+ Theory II, Proc. int. Symp., Austin 1976, 289–298 (1976).
1081
+ [6] H. Berens, R. DeVore, Quantitative Korovkin theorems for positive linear operators
1082
+ on Lp-spaces, Transactions AMS 245 (1978), 349–361.
1083
+ [7] M. Bodur, T. Bostancı, G. Ba¸scanbaz-Tunca, Generalized Kantorovich operators
1084
+ depending on a non-negative integer, Submitted to a Journal.
1085
+ [8] H. Brass, Eine Verallgemeinerung dwe Bernsteinschen Operatoren, Abh. Math.
1086
+ Sem.Univ. Hamburg 38 (1971), 111–122.
1087
+ [9] Fei-long Cao, Multivariate Stancu polynomials and moduli of continuity, (Chinese)
1088
+ Acta Math. Sinica (Chinese Ser.) 48 (2005), no. 1, 51–62.
1089
+ [10] H. Gonska, Xin-long Zhou, The strong converse inequality for Bernstein-Kantorovich
1090
+ polynomials. Comput. Math. Appl. 30 (1995), 103–128.
1091
+ [11] H. H. Gonska, On the composition and decomposition of positive linear operators,
1092
+ in: Kovtunets, V. V. (ed.) et al., Approximation Theory and its Applications. Proc.
1093
+
1094
+ 12
1095
+ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA
1096
+ int. conf. ded. to the memory of Vl. K. Dzyadyk, Kiev, Ukraine, May 27–31, 1999.
1097
+ Pr. Inst. Mat. Nats. Akad. Nauk Ukr., Mat. Zastos. 31, 161–180 (2000).
1098
+ [12] L. V. Kantorovich, Sur certains d´eveloppements suivant les polynˆomes de la forme
1099
+ de S. Bernstein, I, II, C. R. Acad. Sci. URSS, (1930), 563–568, 595–600.
1100
+ [13] L. V. Kantorovich, My journey in science (proposed report to the Moscow Mathe-
1101
+ matical Society), Russ. Math. Surv, 42(2) (1987), 233–270. Russian original: Uspekhi
1102
+ Mat. Nauk 42(2) (1987), 183–213.
1103
+ [14] G. G. Lorentz, Bernstein Polynomials, University of Toronto Press, Toronto, 1953
1104
+ (2.ed., Chelsea Publishing Co., New York, 1986).
1105
+ [15] V. Maier, Lp-approximation by Kantoroviˇc operators, Anal. Math., 4 (1978), 289–
1106
+ 295.
1107
+ [16] M. W. M¨uller, Die G¨ute der Lp- Approximation durch Kantoroviˇc-polynome, Math.
1108
+ Z. 151 (1976), 243–247.
1109
+ [17] M. W. M¨uller, Approximation by Cheney-Sharma-Kantoroviˇc polynomials in the Lp-
1110
+ metric, Rocky Mountain J. Math., 19(1) (1989), 281–291.
1111
+ [18] V. A. Popov, On the quantitative Korovkin theorems in Lp, Compt. Rend. Acad.
1112
+ Bulg. Sci., 35 (1982), 897–900.
1113
+ [19] V. A. Popov, V. Kh. Khristov, “Averaged moduli of smoothness for functions of
1114
+ several variables and function spaces generated by them”, Orthogonal series and
1115
+ approximations of functions, Collection of articles. Dedicated to Academician N. N.
1116
+ Luzin on the occasion of the 100th anniversary of his birth, Trudy Mat. Inst. Steklov.,
1117
+ 164, 1983, 136–141; Proc. Steklov Inst. Math., 164 (1985), 155–160
1118
+ [20] E. Quak, Uni- und multivariate Lp-Abschatzungen des Approximationsfehlers posi-
1119
+ tiver linearer Operatoren mit Hilfe des -Moduls, Ph.D. Thesis, University of Dort-
1120
+ mund 1985.
1121
+ [21] E. Quak, Multivariate Lp-error estimates for positive linear operators via the first-
1122
+ order τ-modulus, J. Approx. Theory 56 (1989), 277–286.
1123
+ [22] B. Sendov, V. A. Popov, The Averaged Moduli of Smoothness. Applications in Nu-
1124
+ merical Methods and Approximation, Chichester (UK) etc.: Wiley 1988.
1125
+ [23] D. D. Stancu, Quadrature formulas constructed by using certain linear positive op-
1126
+ erators, Numerical Integration (Proc. Conf., Oberwolfach, 1981), ISNM 57 (1982),
1127
+ 241–251, Birkh¨auser Verlag, Basel.
1128
+ [24] D. D. Stancu, Approximation of functions by means of a new generalized Bernstein
1129
+ operator, Calcolo, 20 (1983), no. 2, 211–229.
1130
+ [25] J. J. Swetits, B. Wood, Quantitative estimates for Lp approximation with positive
1131
+ linear operators, J. Approx. Theory 38 (1983), 81–89.
1132
+ [26] V. I. Volkov, On the convergence of sequences of linear positive operators in the space
1133
+ of continuous functions of two variables (Russian), Dokl. Akad. Nauk. SSSR (N.S.),
1134
+ 115 (1957), 17–19.
1135
+ [27] R. Yang, J. Xiong, F. Cao, Multivariate Stancu operators defined on a simplex, Appl.
1136
+ Math. Comput., 138 (2003), 189–198.
1137
+ Ankara University, Faculty of Science, Department of Mathematics, Str.
1138
+ D¨ogol 06100, Bes¸evler, Ankara, Turkey
1139
+ Email address: tunca@science.ankara.edu.tr
1140
+ University of Duisburg-Essen, Faculty of Mathematics, Forsthausweg 2,
1141
+ D-47057 Duisburg, Germany
1142
+ Email address: heiner.gonska@uni-due.de and gonska.sibiu@gmail.com
1143
+
FNAzT4oBgHgl3EQfG_ve/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,532 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf,len=531
2
+ page_content='arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
3
+ page_content='01039v1 [math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
4
+ page_content='CA] 3 Jan 2023 BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE∗ G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA This study is dedicated to Professor Ioan Ra¸sa on the occasion of his 70th birthday Abstract.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
5
+ page_content=' We deal with multivariate Brass-Stancu-Kantorovich oper- ators depending on a non-negative integer parameter and defined on the space of all Lebesgue integrable functions on a unit hypercube.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
6
+ page_content=' We prove Lp-approximation and provide estimates for the Lp-norm of the error of approximation in terms of a multivariate averaged modulus of continuity and of the corresponding Lp-modulus.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
7
+ page_content=' 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
8
+ page_content=' Introduction and Historical Notes The fundamental functions of the well-known Bernstein operators are defined by pn,k(x) = � �n k � xk(1 − x)n−k;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
9
+ page_content=' 0 ≤ k ≤ n 0;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
10
+ page_content=' k < 0 or k > n , x ∈ [0, 1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
11
+ page_content=' (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
12
+ page_content='1) In [23], using a probabilistic method, Stancu generalized Bernstein’s funda- mental functions as wn,k,r(x) := \uf8f1 \uf8f2 \uf8f3 (1 − x) pn−r,k (x) ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
13
+ page_content=' 0 ≤ k < r (1 − x) pn−r,k (x) + xpn−r,k−r (x) ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
14
+ page_content=' r ≤ k ≤ n − r xpn−r,k−r (x) ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
15
+ page_content=' n − r < k ≤ n , x ∈ [0, 1], (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
16
+ page_content='2) where r is a non-negative integer parameter, n is any natural number such that n > 2r, for which each pn−r,k is given by (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
17
+ page_content='1), and therefore, con- structed and studied Bernstein-type positive linear operators as Ln,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
18
+ page_content=' x) := n � k=0 wn,k,r(x)f �k n � , x ∈ [0, 1], (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
19
+ page_content='3) for f ∈ C[0, 1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
20
+ page_content=' In doing so Stancu was guided by an article of Brass [8].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
21
+ page_content=' This is further discussed by Gonska [11].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
22
+ page_content=' Among others, estimates in terms of the second order modulus of smoothness are given there for continuous functions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
23
+ page_content=' It is clear that for x ∈ [0, 1] Stancu’s fundamental functions in (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
24
+ page_content='2) satisfy wn,k,r(x) ≥ 0 and n � k=0 wn,k,r(x) = 1, Key words and phrases.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
25
+ page_content=' Multivariate Kantorovich operator;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
26
+ page_content=' Multivariate averaged modulus of smoothness;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
27
+ page_content=' Multivariate K-functional 2010 MSC: 41A36, 41A25, 26A45 ∗This paper is an extension of a talk given in ICATA 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
28
+ page_content=' 1 2 G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA hence the operators Ln,r can be expressed as Ln,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
29
+ page_content=' x) := n−r � k=0 pn−r,k (x) � (1 − x) f �k n � + xf �k + r n �� , (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
30
+ page_content='4) are defined for n ≥ r and satisfy the end point interpolation Ln,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
31
+ page_content=' 0) = f (0) , Ln,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
32
+ page_content=' 1) = f (1).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
33
+ page_content=' It thus seems to be justified to call the Ln,r Brass-Stancu-Bernstein (BSB) operators.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
34
+ page_content=' In [24] Stancu gave uniform convergence limn→∞ Ln,r (f) = f on [0, 1] for f ∈ C[0, 1] and presented an expression for the remainder Rn,r(f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
35
+ page_content=' x) of the approximation formula f(x) = Ln,r(f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
36
+ page_content=' x) + Rn,r(f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
37
+ page_content=' x) by means of second order divided differences and also obtained an integral representation for the remainder.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
38
+ page_content=' Moreover, the author estimated the order of approximation by the operators Ln,r (f) via the classical modulus of continuity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
39
+ page_content=' He also studied the spectral properties of Ln,r.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
40
+ page_content=' In the cases r = 0 and r = 1, the operators Ln,r reduce to the classical Bernstein operators Bn, i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
41
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
42
+ page_content=', Bn (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
43
+ page_content=' x) = n � k=0 pn,k(x)f �k n � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
44
+ page_content=' What also has to be mentioned: Stancu himself in his 1983 paper observed that ”we can optimize the error bound of the approximation of the function f by means of Ln,rf if we take r = 0 or r = 1, when the operator Ln,r reduces to Bernstein’s.” So there is a shortcoming.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
45
+ page_content=' Since Bernstein polynomials are not appropriate for approximation of discontinuous functions (see [14, Section 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
46
+ page_content='9]), by replacing the point evalu- ations f � k n � with the integral means over small intervals around the knots k n, Kantorovich [12] generalized the Bernstein operators as Kn (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
47
+ page_content=' x) = n � k=0 pn,k (x) (n + 1) k+1 n+1 � k n+1 f (t) dt, x ∈ [0, 1], n ∈ N, (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
48
+ page_content='5) for Lebesgue integrable functions f on [0, 1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
49
+ page_content=' On p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
50
+ page_content=' 239 of his mathematical memoirs [13] Kantorovich writes: ”While I was waiting for a student who was late, I was looking over vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
51
+ page_content=' XIII of Fundamenta Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
52
+ page_content=' and saw in it a note from the Moscow Mathematician Khlodovskii related to Bernstein polynomials.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
53
+ page_content=' In it I first caught sight of Bernstein polynomials, which he proposed in 1912 for an elementary proof of the well known Weierstrass theorem .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
54
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
55
+ page_content=' I at once wondered if it is not possible in these polynomials to change the values of the function at certain points into the more stable average of the function in the corresponding interval.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
56
+ page_content=' It turned out that this was possible, and the polynomials could be written in such a form not only for a continuous function but also for any Lebesgue-summable function.” Lorentz [14] proved that lim n→∞ ∥Kn(f) − f∥p = 0, f ∈ Lp[0, 1], 1 ≤ p < ∞.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
57
+ page_content=' There are a lot of articles dealing with classical Kantorovich operators, and, in particular, their degree of approximation and the importance of BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE 3 second order moduli of different types.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
58
+ page_content=' See, e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
59
+ page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
60
+ page_content=', the work of Berens and DeVore [5], [6], Swetits and Wood [25] and Gonska and Zhou [10].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
61
+ page_content=' It is beyond the scope of this note to further discuss this matter.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
62
+ page_content=' As further work on the classical case here we only mention the 1976 work of M¨uller [16], Maier [15], and Altomare et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
63
+ page_content=' [1], see also the references therein.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
64
+ page_content=' Similarly to Kantorovich operators Bodur et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
65
+ page_content=' [7] constructed a Kan- torovich type modification of BSB operators as Kn,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
66
+ page_content=' x) := n � k=0 wn,k,r(x) \uf8eb \uf8ec \uf8ec \uf8ed(n + 1) k+1 n+1 � k n+1 f (t) dt \uf8f6 \uf8f7 \uf8f7 \uf8f8 , x ∈ [0, 1], (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
67
+ page_content='6) for f ∈ L1 [0, 1], where r is a non-negative integer parameter, n is a natural number such that n > 2r and wn,k,r(x) are given by (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
68
+ page_content='2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
69
+ page_content=' And, it was shown that If f ∈ Lp[0, 1], 1 ≤ p < ∞, then lim n→∞ ∥Kn,r(f) − f∥p = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
70
+ page_content=' In addition, it was obtained that each Kn,r is variation detracting as well [7].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
71
+ page_content=' Throughout the paper, we shall call the operators Kn,r given by (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
72
+ page_content='6) ”Brass-Stancu-Kantorovich”, BSK operators.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
73
+ page_content=' Notice that from the definition of wn,k,r, Kn,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
74
+ page_content=' x) can be expressed as Kn,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
75
+ page_content=' x) (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
76
+ page_content='7) = n−r � k=0 pn−r,k (x) (n + 1) \uf8ee \uf8ef\uf8ef\uf8f0(1 − x) k+1 n+1 � k n+1 f (t) dt + x k+r+1 n+1 � k+r n+1 f (t) dt \uf8f9 \uf8fa\uf8fa\uf8fb and in the cases r = 0 and r = 1 they reduce to the Kantorovich operators;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
77
+ page_content=' Kn,0 = Kn,1 = Kn given by (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
78
+ page_content='5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
79
+ page_content=' Again they are defined for all n ≥ r.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
80
+ page_content=' MULTIVARIATE SITUATION Some work has been done in the multivariate setting for BSB and BSK operators.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
81
+ page_content=' For the standard simplex this was done, e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
82
+ page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
83
+ page_content=', by Yang, Xiong and Cao [27] and Cao [9], For example, Cao proved that multivariate Stancu operators preserve the properties of multivariate moduli of continuity and obtained the rate of convergence with the help of Ditzian-Totik’s modulus of continuity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
84
+ page_content=' In this work, motivated by the work Altomare et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
85
+ page_content=' [3], we deal with a multivariate extension of the BSK operators on a d-dimensional unit hy- percube and we study Lp -approximation by these operators.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
86
+ page_content=' For the rate of convergence we provide an estimate in terms of the so called first order multivariate τ-modulus, a quantity coming from the Bulgarian school of Approximation Theory.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
87
+ page_content=' Also, inspired by M¨uller’s approach in [17], we give estimates for differentiable functions and such in terms of the Lp-modulus of smoothness, using properties of the τ-modulus.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
88
+ page_content=' Here the work of Quak [20], [21] was helpful.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
89
+ page_content=' 4 G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
90
+ page_content=' Preliminaries Consider the space Rd, d ∈ N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
91
+ page_content=' Let ∥x∥∞ denote the max-norm of a point x = (x1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
92
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
93
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
94
+ page_content=' , xd) ∈ Rd;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
95
+ page_content=' ∥x∥∞ := ∥x∥max = max i∈{1,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
96
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
97
+ page_content=',d} |xi| and let 1 denote the constant function 1 : Rd → R such that 1 (x) = 1 for x ∈ Rd.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
98
+ page_content=' And, for each j = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
99
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
100
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
101
+ page_content=' , d, let prj : Rd → R stand for the jth coordinate function defined for x ∈ Rd by prj (x) = xj.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
102
+ page_content=' Definition 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
103
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
104
+ page_content=' A multi-index is a d-tuple α = (α1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
105
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
106
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
107
+ page_content=' , αd) of non-negative integers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
108
+ page_content=' Its norm (length) is the quantity |α| = d � i=1 αi.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
109
+ page_content=' The differential operator Dα is defined by Dαf = Dα1 1 · · · Dαd d f, where Di, i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
110
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
111
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
112
+ page_content=' , d, is the corresponding partial derivative operator (see [4, p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
113
+ page_content=' 335]).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
114
+ page_content=' Throughout the paper Qd := [0, 1]d, d ∈ N, will denote the d-dimensional unit hypercube and we consider the space Lp (Qd) = {f : Qd → R | f p-integrable on Qd} , 1 ≤ p < ∞, with the standard norm ∥.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
115
+ page_content='∥p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
116
+ page_content=' Recall the following definition of the usual Lp-modulus of smoothness of first order: Definition 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
117
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
118
+ page_content=' Let f ∈ Lp (Qd) , 1 ≤ p < ∞, h ∈ Rd and δ > 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
119
+ page_content=' The modulus of smoothness of the first order for the function f and step δ in Lp-norm is given by ω1 (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
120
+ page_content=' δ)p = sup 0<∥h∥∞≤δ \uf8eb \uf8ec \uf8ed � Qd |f (x + h) − f (x)|p dx \uf8f6 \uf8f7 \uf8f8 1/p if x, x + h ∈ Qd [21].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
121
+ page_content=' Let M (Qd) := {f | f bounded and measurable on Qd}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
122
+ page_content=' Below, we present the concept of the first order averaged modulus of smoothness.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
123
+ page_content=' Definition 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
124
+ page_content='3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
125
+ page_content=' Let f ∈ M (Qd) , h ∈ Rd and δ > 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
126
+ page_content=' The multivariate averaged modulus of smoothness, or τ-modulus, of the first order for function f and step δ in Lp-norm is given by τ 1 (f, δ)p := ∥ω1 (f, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
127
+ page_content=' ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
128
+ page_content=' δ)∥p , 1 ≤ p < ∞, where ω1 (f, x;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
129
+ page_content=' δ) = sup � |f (t + h) − f (t)| : t, t + h ∈ Qd, ∥t − x∥∞ ≤ δ 2, ∥t + h − x∥∞ ≤ δ 2 � BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE 5 is the multivariate local modulus of smoothness of first order for the function f at the point x ∈ Qd and for step δ.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
130
+ page_content=' [21].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
131
+ page_content=' For our future purposes, we need the following properties of first order multivariate averaged modulus of smoothness: For f ∈ M (Qd) , 1 ≤ p < ∞ and δ, λ, γ ∈ R+, there hold τ 1) τ 1 (f, δ)p ≤ τ 1 (f, λ)p for 0 < δ ≤ λ, τ 2) τ 1 (f, λδ)p ≤ (2 ⌊λ⌋ + 2)d+1 τ 1 (f, δ)p, where ⌊λ⌋ is the greatest inte- ger that does not exceed λ, τ 3) τ 1 (f, δ)p ≤ 2 � |α|≥1 δ|α| ∥Dαf∥p , αi = 0 or 1, if Dαf ∈ Lp (Qd) for all multi-indices α with |α| ≥ 1 and αi = 0 or 1 (see [19] or [21]).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
132
+ page_content=' For a detailed knowledge concerning averaged modulus of smoothness, we refer to the book of Sendov and Popov [22].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
133
+ page_content=' Now, consider the Sobolev space W p 1 (Qd) of functions f ∈ Lp (Qd) , 1 ≤ p < ∞, with (distributional) derivatives Dαf belong to Lp (Qd), where |α| ≤ 1, with the seminorm |f|W p 1 = � |α|=1 ∥Dαf∥p (see [4, p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
134
+ page_content=' 336]).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
135
+ page_content=' Recall that for all f ∈ Lp (Qd) the K-functional, in Lp- norm, is defined as K1,p (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
136
+ page_content=' t) := inf � ∥f − g∥p + t |g|W p 1 : g ∈ W p 1 (Qd) � (t > 0) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
137
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
138
+ page_content='1) K1,p (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
139
+ page_content=' t) is equivalent with the usual first order modulus of smoothness of f, ω1 (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
140
+ page_content=' t)p;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
141
+ page_content=' namely, there are positive constants c1 and c2 such that c1K1,p (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
142
+ page_content=' t) ≤ ω1 (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
143
+ page_content=' t)p ≤ c2K1,p (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
144
+ page_content=' t) (t > 0) (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
145
+ page_content='2) holds for all f ∈ Lp (Qd) (see [4, Formula 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
146
+ page_content='42 in p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
147
+ page_content=' 341]).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
148
+ page_content=' The following result due to Quak [21] is an upper estimate for the Lp-norm of the approximation error by the multivariate positive linear operators in terms of the first order averaged modulus of smoothness.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
149
+ page_content=' Note that this idea was used first by Popov for the univariate case in [18].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
150
+ page_content=' Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
151
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
152
+ page_content=' Let L : M (Qd) → M (Qd) be a positive linear operator that preserves the constants.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
153
+ page_content=' Then for every f ∈ M (Qd) and 1 ≤ p < ∞, the following estimate holds: ∥L(f) − f∥p ≤ Cτ1 � f, 2d√ A � p , where C is a positive constant and A := sup � L � (pri ◦ ψx)2 ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
154
+ page_content=' x � : i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
155
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
156
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
157
+ page_content=' , d, x ∈ Qd � , in which ψx (y) := y − x for fixed x ∈ Qd and for every y ∈ Qd and A ≤ 1 [21].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
158
+ page_content=' 6 G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
159
+ page_content=' Multivariate BSK-Operators In this section, motivated by the works of Altomare et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
160
+ page_content=' [1] and Al- tomare et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
161
+ page_content=' [3], we consider the multivariate extension of BSK-operators on Lp (Qd) and study approximation properties of these operators in Lp- norm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
162
+ page_content=' We investigate the rate of the convergence in terms of the first order τ-modulus and the usual Lp-modulus of smoothness of the first order.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
163
+ page_content=' Let r be a given non-negative integer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
164
+ page_content=' For any n ∈ N such that n > 2r, k = (k1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
165
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
166
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
167
+ page_content=' , kd) ∈ {0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
168
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
169
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
170
+ page_content=' , n}d and x = (x1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
171
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
172
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
173
+ page_content=' , xd) ∈ Qd, we set wn,k,r(x) := d � i=1 wn,ki,r(xi), (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
174
+ page_content='1) where, wn,ki,r(xi) is Stancu’s fundamental function given by (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
175
+ page_content='2), written for each i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
176
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
177
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
178
+ page_content=' , d, 0 ≤ ki ≤ n and xi ∈ [0, 1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
179
+ page_content=' Thus, for x ∈ Qd, we have wn,k,r(x) ≥ 0 and � k∈{0,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
180
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
181
+ page_content=',n}d wn,k,r(x) = 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
182
+ page_content=' (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
183
+ page_content='2) For f ∈ L1 (Qd) and x = (x1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
184
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
185
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
186
+ page_content=' , xd) ∈ Qd we consider the following multivariate extension of the BSK-operators Kn,r given by (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
187
+ page_content='6): Kd n,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
188
+ page_content=' x) = n � k1,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
189
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
190
+ page_content=',kd=0 d � i=1 wn,ki,r(xi) � Qd f �k1 + u1 n + 1 , .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
191
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
192
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
193
+ page_content=' , kd + ud n + 1 � du1 · · · dud.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
194
+ page_content=' Notice that from (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
195
+ page_content='1), and denoting, as usual, any f ∈ L1 (Qd) of x = (x1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
196
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
197
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
198
+ page_content=' , xd) ∈ Qd by f (x) = f (x1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
199
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
200
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
201
+ page_content=' , xd), we can express these operators in compact form as Kd n,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
202
+ page_content=' x) = � k∈{0,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
203
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
204
+ page_content=',n}d wn,k,r(x) � Qd f �k + u n + 1 � du.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
205
+ page_content=' (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
206
+ page_content='3) It is clear that multivariate BSK-operators are positive and linear and the cases r = 0 and 1 give the multivariate Kantorovich operators on the hyper- cube Qd, which can be captured from [1] as a special case.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
207
+ page_content=' Lemma 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
208
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
209
+ page_content=' For x ∈ Qd, we have Kd n,r (1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
210
+ page_content=' x) = 1, Kd n,r (pri;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
211
+ page_content=' x) = n n + 1xi + 1 2 (n + 1), Kd n,r � pr2 i ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
212
+ page_content=' x � = n2 (n + 1)2 � x2 i + � 1 + r (r − 1) n � xi (1 − xi) n � + 3nxi + 1 3 (n + 1)2 , for i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
213
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
214
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
215
+ page_content=' , d.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
216
+ page_content=' Taking this lemma into consideration, by the well-known theorem of Volkov [26], we immediately get that Theorem 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
217
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
218
+ page_content=' Let r be a non-negative fixed integer and f ∈ C (Qd).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
219
+ page_content=' Then lim n→∞ Kd n,r (f) = f uniformly on Qd.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
220
+ page_content=' BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE 7 Now, we need the following evaluations for the subsequent result: For 0 ≤ xi ≤ 1, i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
221
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
222
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
223
+ page_content=' , d, we have 1 � 0 (1 − xi) pn−r,ki (xi) dxi = �n − r ki � 1 � 0 xki i (1 − xi)n−r−ki+1 dxi = n − r − ki + 1 (n − r + 2) (n − r + 1) when 0 ≤ ki < r and 1 � 0 xipn−r,ki−r (xi) dxi = �n − r ki − r � 1 � 0 xki−r+1 i (1 − xi)n−ki dxi = ki − r + 1 (n − r + 2) (n − r + 1) when n − r < ki ≤ n.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
224
+ page_content=' Thus, from (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
225
+ page_content='1) and (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
226
+ page_content='2), it follows that 1 � 0 wn,ki,r(xi)dxi = \uf8f1 \uf8f4 \uf8f2 \uf8f4 \uf8f3 n−r−ki+1 (n−r+2)(n−r+1);' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
227
+ page_content=' 0 ≤ ki < r n−2r+2 (n−r+2)(n−r+1);' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
228
+ page_content=' r ≤ ki ≤ n − r ki−r+1 (n−r+2)(n−r+1);' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
229
+ page_content=' n − r < ki ≤ n .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
230
+ page_content=' (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
231
+ page_content='4) Note that we can write the following estimates n − r − ki + 1 ≤ n − r + 1 when 0 ≤ ki < r, n − 2r + 2 ≤ n − r + 1 when r ≤ ki ≤ n − r, ki − r + 1 ≤ n − r + 1 when n − r < ki ≤ n (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
232
+ page_content='5) for each i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
233
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
234
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
235
+ page_content=' , d, where in the middle term, we have used the hypothesis n > 2r.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
236
+ page_content=' Making use of (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
237
+ page_content='5), (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
238
+ page_content='4) and (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
239
+ page_content='1), we obtain � Qd wn,k,r(x)dx = d � i=1 1 � 0 wn,ki,r(xi)dxi ≤ 1 (n − r + 2)d .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
240
+ page_content=' (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
241
+ page_content='6) Lp-approximation by the sequence of the multivariate Stancu-Kantorovich operators is presented in the following theorem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
242
+ page_content=' Theorem 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
243
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
244
+ page_content=' Let r be a non-negative fixed integer and f ∈ Lp (Qd) , 1 ≤ p < ∞.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
245
+ page_content=' Then lim n→∞ ��Kd n,r(f) − f �� p = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
246
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
247
+ page_content=' Since the cases r = 0 and 1 correspond to the multivariate Kan- torovich operators (see [1] or [3]), we consider only the cases r > 1, which is taken as fixed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
248
+ page_content=' From Theorem 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
249
+ page_content='1, we obtain that lim n→∞ ��Kd n,r(f) − f �� p = 0 for any f ∈ C (Qd).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
250
+ page_content=' Since C (Qd) is dense in Lp (Qd), denoting the norm of the operator Kd n,r acting on Lp (Qd) onto itself by ��Kd n,r ��, it remains to show that there exists an Mr, where Mr is a positive constant that maybe depends on r, such that ��Kd n,r �� ≤ Mr for all n > 2r.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
251
+ page_content=' Now, as in [3, p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
252
+ page_content='604], we adopt the notation Qn,k := d � i=1 � ki n + 1, ki + 1 n + 1 � ⊂ Qd;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
253
+ page_content=' � k∈{0,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
254
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
255
+ page_content=',n}d Qn,k = Qd.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
256
+ page_content=' 8 G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA Making use of the convexity of the function ϕ (t) := |t|p , t ∈ R, 1 ≤ p < ∞ (see, e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
257
+ page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
258
+ page_content=', [2]), and (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
259
+ page_content='2), for every f ∈ Lp (Qd) , n > 2r, and x ∈ Qd, we obtain ���Kd n,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
260
+ page_content=' x) ��� p ≤ � k∈{0,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
261
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
262
+ page_content=',n}d wn,k,r(x) � Qd ����f �k + u n + 1 ����� p du = � k∈{0,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
263
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
264
+ page_content=',n}d wn,k,r(x) (n + 1)d � Qn,k |f (v)|p dv.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
265
+ page_content=' Taking (3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
266
+ page_content='6) into consideration, we reach to � Qd ���Kd n,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
267
+ page_content=' x) ��� p dx ≤ � k∈{0,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
268
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
269
+ page_content=',n}d � n + 1 n − r + 2 �d � Qn,k |f (v)|p dv.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
270
+ page_content=' Since sup n>2r � n+1 n−r+2 �d = � 2r+2 r+3 �d := Mr for r > 1, where 1 < 2r+2 r+3 < 2, we get � Qd ���Kd n,r (f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
271
+ page_content=' x) ��� p dx ≤ Mr � Qd |f (v)|p dv, which implies that ��Kd n,r (f) �� p ≤ M1/p r ∥f∥p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
272
+ page_content=' Note that for the cases r = 0 and 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
273
+ page_content=' we have Mr = 1 (see [3]).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
274
+ page_content=' Therefore, the proof is completed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
275
+ page_content=' □ 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
276
+ page_content=' Estimates for the rate of convergence In [17], M¨uller studied Lp-approximation by the sequence of the Cheney- Sharma-Kantorovich operators (CSK).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
277
+ page_content=' The author gave an estimate for this approximation in terms of the univariate τ-modulus and moreover, using some properties of the τ-modulus, he also obtained upper estimates for the Lp-norm of the error of approximation for first order differentiable functions as well as for continuous ones.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
278
+ page_content=' In this part, we show that similar estimates can also be obtained for ��Kd n,r (f) − f �� p in the multivariate setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
279
+ page_content=' Our first result is an application of Quak’s method in Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
280
+ page_content='1 Theorem 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
281
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
282
+ page_content=' Let r be a non-negative fixed integer, f ∈ M (Qd) and 1 ≤ p < ∞.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
283
+ page_content=' Then ���Kd n,r (f) − f ��� p ≤ Cτ 1 � f, 2d � 3n + 1 + 3r (r − 1) 12 (n + 1)2 � p (4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
284
+ page_content='1) for all n ∈ N such that n > 2r, where the positive constant C does not depend on f.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
285
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
286
+ page_content=' According to Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
287
+ page_content='1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
288
+ page_content=' by taking ψx (y) = y − x for fixed x ∈ Qd and for every y ∈ Qd, and defining An,r := sup � Kd n,r � (pri ◦ ψx)2 ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
289
+ page_content=' x � : i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
290
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
291
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
292
+ page_content=' , d, x ∈ Qd � , where (pri ◦ ψx)2 = pr2 i − 2xipri + x2 i 1, i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
293
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
294
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
295
+ page_content=' , d, we get the following estimate ���Kd n,r (f) − f ��� p ≤ Cτ1 � f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
296
+ page_content=' 2d� An,r � BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE 9 for any f ∈ M (Qd), under the condition that An,r ≤ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
297
+ page_content=' Now, applying the operators Kd n,r and making use of Lemma 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
298
+ page_content='1, for every i = 1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
299
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
300
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
301
+ page_content=' , d and x ∈ Qd, we obtain Kd n,r � (pri ◦ ψx)2 ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
302
+ page_content=' x � = n − 1 + r (r − 1) (n + 1)2 xi (1 − xi) + 1 3 (n + 1)2 ≤ n − 1 + r (r − 1) 4 (n + 1)2 + 1 3 (n + 1)2 = 3n + 1 + 3r (r − 1) 12 (n + 1)2 for all n ∈ N such that n > 2r, where r ∈ N ∪ {0}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
303
+ page_content=' Therefore, since we have n ≥ 2r + 1, we take r ≤ n−1 2 and obtain that An,r ≤ 3n+1+3r(r−1) 12(n+1)2 ≤ 1 is satisfied, which completes the proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
304
+ page_content=' □ Now, making use of the properties τ1)-τ3) of the multivariate first order τ-modulus, we obtain Theorem 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
305
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
306
+ page_content=' Let r be a non-negative fixed integer, f ∈ Lp (Qd) , 1 ≤ p < ∞, and Dαf ∈ Lp (Qd) for all multi-indices α with |α| ≥ 1, αi = 0 or 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
307
+ page_content=' Then ���Kd n,r (f) − f ��� p ≤ 2Cr � |α|≥1 � 1 2d√n + 1 �|α| ∥Dαf∥p , for all n ∈ N such that n > 2r, where Cr is a positive constant depending on r.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
308
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
309
+ page_content=' Since n > 2r, we immediately have n + 1 ≥ 2 (r + 1).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
310
+ page_content=' Thus, the term appearing inside the 2dth root in the formula (4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
311
+ page_content='1) can be estimated, respectively, for r > 1, and r = 0, 1, as 3n + 1 + 3r (r − 1) 12 (n + 1)2 = 3n + 3 + 3r (r − 1) − 2 12(n + 1)2 = 1 n + 1 �1 4 + 3r (r − 1) − 2 12(n + 1) � ≤ 1 n + 1 �1 4 + 3r (r − 1) − 2 24(r + 1) � = 1 n + 1 �3r2 + 3r + 4 24(r + 1) � and 3n + 1 12 (n + 1)2 = 1 n + 1 3n + 1 4 (3n + 3) < 1 4 (n + 1).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
312
+ page_content=' Now, defining Br := � 3r2+3r+4 24(r+1) ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
313
+ page_content=' r > 1, 1 4;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
314
+ page_content=' r = 0, 1, 10 G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA and making use of the properties τ 1)-τ 3) of τ-modulus, from (4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
315
+ page_content='1), we arrive at ���Kd n,r (f) − f ��� p ≤ Cτ1 � f, 2d � 3n + 1 + 3r (r − 1) 12 (n + 1)2 � p ≤ Cτ1 � f, 2d� Br 1 2d√n + 1 � p ≤ C � 2 � 2d� Br � + 2 �d+1 τ 1 � f, 1 2d√n + 1 � p ≤ 2Cr � |α|≥1 � 1 2d√n + 1 �|α| ∥Dαf∥p , where the positive constant Cr is defined as Cr := C � 2 � 2d√Br � + 2 �d+1 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
316
+ page_content=' □ For non-differentiable functions we have the following estimate in terms of the first order modulus of smoothness, in Lp-norm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
317
+ page_content=' Theorem 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
318
+ page_content='3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
319
+ page_content=' Let r be a non-negative fixed integer and f ∈ Lp (Qd) , 1 ≤ p < ∞.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
320
+ page_content=' Then ���Kd n,r (f) − f ��� p ≤ c2Cr,pω1 � f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
321
+ page_content=' 1 2d√n + 1 � p , where ω1 is the first order multivariate modulus of smoothness of f and Cr,p is a constant depending on r and p.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
322
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
323
+ page_content=' By Theorem 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
324
+ page_content='2, since Kd n,r is bounded, with ��Kd n,r �� p ≤ M1/p r , for all n ∈ N such that n > 2r, we have ��Kd n,r (g) − g �� p ≤ � M1/p r + 1 � ∥g∥p for g ∈ Lp (Qd).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
325
+ page_content=' Moreover, from Theorem 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
326
+ page_content='2, we can write ���Kd n,r (g) − g ��� p ≤ 2Cr � |α|≥1 � 1 2d√n + 1 �|α| ∥Dαg∥p for those g such that Dαg ∈ Lp (Qd), for all multi-indices α with |α| ≥ 1 and αi = 0 or 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
327
+ page_content=' Hence, for f ∈ Lp (Qd), it readily follows that ���Kd n,r (f) − f ��� p ≤ ���Kd n,r (f − g) − (f − g) ��� p + ���Kd n,r (g) − g ��� p ≤ � M1/p r + 1 � \uf8f1 \uf8f2 \uf8f3∥f − g∥p + 2Cr � |α|≥1 � 1 2d√n + 1 �|α| ∥Dαg∥p \uf8fc \uf8fd \uf8fe .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
328
+ page_content=' BRASS-STANCU-KANTOROVICH OPERATORS ON A HYPERCUBE 11 Passing to the infimum for all g ∈ W p 1 (Qd) in the last formula, since the infimum of a superset does not exceed that of subset, we obtain ���Kd n,r (f) − f ��� p ≤ � M1/p r + 1 � inf \uf8f1 \uf8f2 \uf8f3∥f − g∥p + 2Cr 2d√n + 1 � |α|=1 ∥Dαg∥p : g ∈ W p 1 (Qd) \uf8fc \uf8fd \uf8fe = � M1/p r + 1 � inf � ∥f − g∥p + 2Cr 2d√n + 1 |g|W p 1 : g ∈ W p 1 (Qd) � = � M1/p r + 1 � K1,p � f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
329
+ page_content=' 2Cr 2d√n + 1 � , (4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
330
+ page_content='2) where K1,p is the K-functional given by (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
331
+ page_content='1).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
332
+ page_content=' The proof follows from the equivalence (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
333
+ page_content='2) of the K-functional and the first order modulus of smooth- ness in Lp-norm and the non-decreasingness property of the modulus.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
334
+ page_content=' In- deed, we get K1,p � f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
335
+ page_content=' 2Cr 2d√n + 1 � ≤ c2ω1 � f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
336
+ page_content=' 2Cr 2d√n + 1 � p ≤ c2 (2Cr + 1) ω1 � f;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
337
+ page_content=' 1 2d√n + 1 � p .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
338
+ page_content=' (4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
339
+ page_content='3) Combining (4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
340
+ page_content='3) with (4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
341
+ page_content='2) and defining Cr,p := � M1/p r + 1 � (2Cr + 1), where M1/p r and Cr are the same as in Theorems 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
342
+ page_content='2 and 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
343
+ page_content='2, respectively, we obtain the desired result.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
344
+ page_content=' □ References [1] F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
345
+ page_content=' Altomare, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
346
+ page_content=' Cappelletti Montano, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
347
+ page_content=' Leonessa, On a generalization of Kan- torovich operators on simplices and hypercubes, Adv.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
348
+ page_content=' Pure Appl.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
349
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
350
+ page_content=' 1 (2010), no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
351
+ page_content=' 3, 359-385.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
352
+ page_content=' [2] F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
353
+ page_content=' Altomare, Korovkin-type Theorems and Approximation by Positive Linear Oper- ators, Surv.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
354
+ page_content=' Approx.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
355
+ page_content=' Theory 5 (2010), 92-164.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
356
+ page_content=' [3] F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
357
+ page_content=' Altomare, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
358
+ page_content=' Cappelletti Montano, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
359
+ page_content=' Leonessa, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
360
+ page_content=' Ra¸sa, A generalization of Kantorovich operators for convex compact subsets, Banach J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
361
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
362
+ page_content=' Anal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
363
+ page_content=' 11 (2017), no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
364
+ page_content=' 3, 591–614.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
365
+ page_content=' [4] C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
366
+ page_content=' Bennett, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
367
+ page_content=' Sharpley, Interpolation of Operators, Academic Press Inc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
368
+ page_content=', 1988.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
369
+ page_content=' [5] H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
370
+ page_content=' Berens, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
371
+ page_content=' DeVore, Quantitative Korovkin theorems for Lp-spaces.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
372
+ page_content=' In: Approx.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
373
+ page_content=' Theory II, Proc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
374
+ page_content=' int.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
375
+ page_content=' Symp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
376
+ page_content=', Austin 1976, 289–298 (1976).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
377
+ page_content=' [6] H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
378
+ page_content=' Berens, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
379
+ page_content=' DeVore, Quantitative Korovkin theorems for positive linear operators on Lp-spaces, Transactions AMS 245 (1978), 349–361.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
380
+ page_content=' [7] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
381
+ page_content=' Bodur, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
382
+ page_content=' Bostancı, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
383
+ page_content=' Ba¸scanbaz-Tunca, Generalized Kantorovich operators depending on a non-negative integer, Submitted to a Journal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
384
+ page_content=' [8] H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
385
+ page_content=' Brass, Eine Verallgemeinerung dwe Bernsteinschen Operatoren, Abh.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
386
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
387
+ page_content=' Sem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
388
+ page_content='Univ.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
389
+ page_content=' Hamburg 38 (1971), 111–122.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
390
+ page_content=' [9] Fei-long Cao, Multivariate Stancu polynomials and moduli of continuity, (Chinese) Acta Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
391
+ page_content=' Sinica (Chinese Ser.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
392
+ page_content=') 48 (2005), no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
393
+ page_content=' 1, 51–62.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
394
+ page_content=' [10] H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
395
+ page_content=' Gonska, Xin-long Zhou, The strong converse inequality for Bernstein-Kantorovich polynomials.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
396
+ page_content=' Comput.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
397
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
398
+ page_content=' Appl.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
399
+ page_content=' 30 (1995), 103–128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
400
+ page_content=' [11] H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
401
+ page_content=' H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
402
+ page_content=' Gonska, On the composition and decomposition of positive linear operators, in: Kovtunets, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
403
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
404
+ page_content=' (ed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
405
+ page_content=') et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
406
+ page_content=', Approximation Theory and its Applications.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
407
+ page_content=' Proc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
408
+ page_content=' 12 G¨ULEN BAS¸CANBAZ-TUNCA AND HEINER GONSKA int.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
409
+ page_content=' conf.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
410
+ page_content=' ded.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
411
+ page_content=' to the memory of Vl.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
412
+ page_content=' K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
413
+ page_content=' Dzyadyk, Kiev, Ukraine, May 27–31, 1999.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
414
+ page_content=' Pr.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
415
+ page_content=' Inst.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
416
+ page_content=' Mat.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
417
+ page_content=' Nats.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
418
+ page_content=' Akad.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
419
+ page_content=' Nauk Ukr.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
420
+ page_content=', Mat.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
421
+ page_content=' Zastos.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
422
+ page_content=' 31, 161–180 (2000).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
423
+ page_content=' [12] L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
424
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
425
+ page_content=' Kantorovich, Sur certains d´eveloppements suivant les polynˆomes de la forme de S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
426
+ page_content=' Bernstein, I, II, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
427
+ page_content=' R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
428
+ page_content=' Acad.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
429
+ page_content=' Sci.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
430
+ page_content=' URSS, (1930), 563–568, 595–600.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
431
+ page_content=' [13] L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
432
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
433
+ page_content=' Kantorovich, My journey in science (proposed report to the Moscow Mathe- matical Society), Russ.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
434
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
435
+ page_content=' Surv, 42(2) (1987), 233–270.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
436
+ page_content=' Russian original: Uspekhi Mat.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
437
+ page_content=' Nauk 42(2) (1987), 183–213.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
438
+ page_content=' [14] G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
439
+ page_content=' G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
440
+ page_content=' Lorentz, Bernstein Polynomials, University of Toronto Press, Toronto, 1953 (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
441
+ page_content='ed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
442
+ page_content=', Chelsea Publishing Co.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
443
+ page_content=', New York, 1986).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
444
+ page_content=' [15] V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
445
+ page_content=' Maier, Lp-approximation by Kantoroviˇc operators, Anal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
446
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
447
+ page_content=', 4 (1978), 289– 295.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
448
+ page_content=' [16] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
449
+ page_content=' W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
450
+ page_content=' M¨uller, Die G¨ute der Lp- Approximation durch Kantoroviˇc-polynome, Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
451
+ page_content=' Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
452
+ page_content=' 151 (1976), 243–247.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
453
+ page_content=' [17] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
454
+ page_content=' W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
455
+ page_content=' M¨uller, Approximation by Cheney-Sharma-Kantoroviˇc polynomials in the Lp- metric, Rocky Mountain J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
456
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
457
+ page_content=', 19(1) (1989), 281–291.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
458
+ page_content=' [18] V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
459
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
460
+ page_content=' Popov, On the quantitative Korovkin theorems in Lp, Compt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
461
+ page_content=' Rend.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
462
+ page_content=' Acad.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
463
+ page_content=' Bulg.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
464
+ page_content=' Sci.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
465
+ page_content=', 35 (1982), 897–900.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
466
+ page_content=' [19] V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
467
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
468
+ page_content=' Popov, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
469
+ page_content=' Kh.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
470
+ page_content=' Khristov, “Averaged moduli of smoothness for functions of several variables and function spaces generated by them”, Orthogonal series and approximations of functions, Collection of articles.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
471
+ page_content=' Dedicated to Academician N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
472
+ page_content=' N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
473
+ page_content=' Luzin on the occasion of the 100th anniversary of his birth, Trudy Mat.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
474
+ page_content=' Inst.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
475
+ page_content=' Steklov.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
476
+ page_content=', 164, 1983, 136–141;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
477
+ page_content=' Proc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
478
+ page_content=' Steklov Inst.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
479
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
480
+ page_content=', 164 (1985), 155–160 [20] E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
481
+ page_content=' Quak, Uni- und multivariate Lp-Abschatzungen des Approximationsfehlers posi- tiver linearer Operatoren mit Hilfe des -Moduls, Ph.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
482
+ page_content='D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
483
+ page_content=' Thesis, University of Dort- mund 1985.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
484
+ page_content=' [21] E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
485
+ page_content=' Quak, Multivariate Lp-error estimates for positive linear operators via the first- order τ-modulus, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
486
+ page_content=' Approx.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
487
+ page_content=' Theory 56 (1989), 277–286.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
488
+ page_content=' [22] B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
489
+ page_content=' Sendov, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
490
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
491
+ page_content=' Popov, The Averaged Moduli of Smoothness.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
492
+ page_content=' Applications in Nu- merical Methods and Approximation, Chichester (UK) etc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
493
+ page_content=' : Wiley 1988.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
494
+ page_content=' [23] D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
495
+ page_content=' D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
496
+ page_content=' Stancu, Quadrature formulas constructed by using certain linear positive op- erators, Numerical Integration (Proc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
497
+ page_content=' Conf.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
498
+ page_content=', Oberwolfach, 1981), ISNM 57 (1982), 241–251, Birkh¨auser Verlag, Basel.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
499
+ page_content=' [24] D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
500
+ page_content=' D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
501
+ page_content=' Stancu, Approximation of functions by means of a new generalized Bernstein operator, Calcolo, 20 (1983), no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
502
+ page_content=' 2, 211–229.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
503
+ page_content=' [25] J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
504
+ page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
505
+ page_content=' Swetits, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
506
+ page_content=' Wood, Quantitative estimates for Lp approximation with positive linear operators, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
507
+ page_content=' Approx.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
508
+ page_content=' Theory 38 (1983), 81–89.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
509
+ page_content=' [26] V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
510
+ page_content=' I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
511
+ page_content=' Volkov, On the convergence of sequences of linear positive operators in the space of continuous functions of two variables (Russian), Dokl.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
512
+ page_content=' Akad.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
513
+ page_content=' Nauk.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
514
+ page_content=' SSSR (N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
515
+ page_content='S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
516
+ page_content=' ), 115 (1957), 17–19.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
517
+ page_content=' [27] R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
518
+ page_content=' Yang, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
519
+ page_content=' Xiong, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
520
+ page_content=' Cao, Multivariate Stancu operators defined on a simplex, Appl.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
521
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
522
+ page_content=' Comput.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
523
+ page_content=', 138 (2003), 189–198.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
524
+ page_content=' Ankara University, Faculty of Science, Department of Mathematics, Str.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
525
+ page_content=' D¨ogol 06100, Bes¸evler, Ankara, Turkey Email address: tunca@science.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
526
+ page_content='ankara.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
527
+ page_content='edu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
528
+ page_content='tr University of Duisburg-Essen, Faculty of Mathematics, Forsthausweg 2, D-47057 Duisburg, Germany Email address: heiner.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
529
+ page_content='gonska@uni-due.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
530
+ page_content='de and gonska.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
531
+ page_content='sibiu@gmail.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
532
+ page_content='com' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/FNAzT4oBgHgl3EQfG_ve/content/2301.01039v1.pdf'}
FtE1T4oBgHgl3EQfXARM/content/tmp_files/2301.03121v1.pdf.txt ADDED
@@ -0,0 +1,1863 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ AdS/BCFT correspondence and Horndeski gravity in the presence of gauge
2
+ fields: from holographic paramagnetism/ferromagnetism phase transition
3
+ Fabiano F. Santosa,∗ Mois´es Bravo-Gaeteb,† Oleksii Sokoliuk c,d,‡ and Alexander Baransky c§
4
+ aInstituto de F´ısica, Universidade Federal do Rio de Janeiro,
5
+ Caixa Postal 68528, Rio de Janeiro-RJ, 21941-972 – Brazil.
6
+ bFacultad de Ciencias B´asicas, Universidad Cat´olica del Maule, Casilla 617, Talca, Chile.
7
+ c Astronomical Observatory, Taras Shevchenko National University of Kyiv,
8
+ 3 Observatorna St., 04053 Kyiv, Ukraine, and
9
+ dMain Astronomical Observatory of the NAS of Ukraine (MAO NASU), Kyiv, 03143, Ukraine.
10
+ This paper presents a dual gravity model for a (2+1)-dimensional system with a limit
11
+ on finite charge density and temperature, which will be used to study the properties of the
12
+ holographic phase transition to paramagnetism-ferromagnetism in the presence of Horndeski
13
+ gravity terms. In our model, the non-zero charge density is supported by a magnetic field.
14
+ As a result, the radius ρ/B indicates a localized condensate, as we increase the Horndeski
15
+ gravity parameter, that is represented by γ. Furthermore, such condensate shows quantum
16
+ Hall-type behavior. This radius is also inversely related to the total action coefficients of
17
+ our model. It was observed that increasing the Horndeski parameter decreases the critical
18
+ temperature of the holographic model and leads to the harder formation of the magnetic
19
+ moment at the bottom of the black hole.
20
+ However, when removing the magnetic field,
21
+ the ferromagnetic material presents a disorder of its magnetic moments, which is observed
22
+ through the entropy of the system. We also found that at low temperatures, spontaneous
23
+ magnetization and ferromagnetic phase transition.
24
+ I.
25
+ INTRODUCTION
26
+ For almost thirty years, the Anti-de Sitter/Conformal Field Theory (AdS/CFT) correspondence
27
+ has been a bridge that allows us to relate gravity and strongly coupled conformal field theories
28
+ [1, 2]. Following this spirit, a new holographic dual of a CFT arises, which is defined on a manifold
29
+ M with a boundary ∂M, denoted as Boundary Conformal Field Theory (BCFT), proposed by
30
+ Takayanagi [3] and Takayanagi et al. [4], extending the AdS/CFT duality. This new holographic
31
+ ∗Electronic address: fabiano.ffs23-at-gmail.com
32
+ †Electronic address: mbravo-at-ucm.cl
33
+ ‡Electronic address: oleksii.sokoliuk-at-mao.kiev.ua
34
+ §Electronic address: abaransky-at-ukr.net
35
+ arXiv:2301.03121v1 [hep-th] 8 Jan 2023
36
+
37
+ 2
38
+ dual denoted as AdS/BCFT correspondence, is defined on a manifold boundary in a D-dimensional
39
+ manifold M to a (D+1)-dimensional asymptotically AdS space N in order to ∂N = M∪Q. Here,
40
+ Q corresponds to a D-dimensional manifold that satisfies ∂Q = ∂M (see Figure 1).
41
+ FIG. 1: Schematic representation of the AdS/BCFT correspondence. Here, M represents the manifold with
42
+ boundary ∂M where the CFT is present. On the other hand, the gravity side is represented by N, which is
43
+ asymptotically AdS is M. Together with the above, ∂M is extended into the bulk AdS, which constitutes
44
+ the boundary of the D−dimensional manifold Q.
45
+ At the moment to explore the AdS/CFT correspondence, we impose the Dirichlet boundary
46
+ condition at the boundary of AdS, and therefore we require the Dirichlet boundary condition on M.
47
+ Nevertheless, according to [3, 4], for AdS/BCFT duality a Neumann boundary condition (NBC)
48
+ on Q is required, given that this boundary should be dynamical, from the viewpoint of holography,
49
+ and there is no natural definite metric on Q specified from the CFT side [5].
50
+ On the other hand, the AdS/BCFT conjecture appears in many scenarios of the transport
51
+ coefficients, where black holes take a providential role, such for example Hawking-Page phase
52
+ transition, the Hall conductivity and the fluid/gravity correspondence [4, 6–11]. Together with the
53
+ above, this duality finds its natural roots in the holographic derivation of entanglement entropy
54
+ [12] as well as in the Randall-Sundrum model [13]. In fact, this extension of the CFT’s boundary
55
+ inside the bulk of the AdS-space is considered a modification of a thin Randall-Sundrum brane,
56
+ which intersects the AdS boundary. For this brane to be a dynamical object, we need to impose, as
57
+ was shown before, NBC where the discontinuity in the bulk extrinsic curvature across the defect,
58
+ is compensated by the tension from the brane. Furthermore, these boundaries are known as the
59
+ Randall-Sundrum (RS) branes in the literature.
60
+ Following the above, Fujita et al. [14] propose a model with gauge fields in the AdS4 background
61
+ with boundary RS branes. In this setup, the authors show that the additional boundary conditions
62
+ impose relevant constraints on the gauge field parameters, deriving the Hall conductivity behavior
63
+
64
+ M
65
+ Q
66
+ N3
67
+ in the dual field theory. Nevertheless, this approach does not consider the back reaction of the gauge
68
+ fields on the geometry, constraining the geometry of the empty AdS space. A natural extension
69
+ and generalization from the above work was constructed in [6].
70
+ In the present paper, we are interested in constructing configurations describing a physical sys-
71
+ tem at finite temperature and charge density. For this, we consider the most common playground,
72
+ provided by the charged AdS4 black holes. This background has already been shown to encode
73
+ many interesting condensed-matter-like phenomena such as superconductivity/superfluidity [15, 16]
74
+ and strange metallic behaviors [17], via an action characterized by the well-known Einstein-Hilbert
75
+ structure together with a cosmological constant and Abelian gauge fields. It is interesting to note
76
+ that the above toy model can be extended in the presence of boundaries within a special case of
77
+ the Horndeski gravity [18], (see for example [19–26]). Here, the gravity theory is given through the
78
+ Lagrangian
79
+ LH = κ
80
+
81
+ (R − 2Λ) − 1
82
+ 2(αgµν − γ Gµν)∇µφ∇νφ
83
+
84
+ ,
85
+ (1)
86
+ where R, Gµν and Λ are the scalar curvature, the Einstein tensor, and the cosmological constant
87
+ respectively, φ = φ(r) is a scalar field, α and γ are coupling constants, while that κ = 1/(16πGN),
88
+ where GN is the Newton Gravitational constant. The Lagrangian (1) has been exhaustively ex-
89
+ plored from the perspective of hairy black hole configurations [27–31], boson and neutron stars
90
+ [33–35], Hairy Taub-NUT/Bolt-AdS solutions [36], as well as holographic applications such that
91
+ quantum complexity and shear viscosity [37–41].
92
+ On the other hand, through this work the physical system analyzed is based on the model
93
+ proposed by [6, 14]. Here, as we will see in the following lines, we start from the same Lagrangian
94
+ for a Horndeski-Maxwell system, this is (1), together with the Maxwell Lagrangian
95
+ LM = − κ
96
+ 4e2 F µνFµν,
97
+ (2)
98
+ where e is a coupling constant and Fµν = ∂µAν − ∂νAµ is the Maxwell stress tensor, describing
99
+ the gravity dual of a field theory on a half-plane. In the simple plane-symmetric black hole ansatz,
100
+ we have that only tensionless RS branes are allowed, and that the background solution must be
101
+ not allowed to model the situation with external electric fields, as in [14]. Even more, as a result
102
+ of the NBC for the gauge fields, and showing in [6], the charge density ρ in the dual field theory
103
+ must be supported by an external magnetic field B, where the ratio ρ/B, which is equal to the
104
+ Hall conductivity, is a constant inversely proportional to the coefficients. In our prescription, this
105
+ represents the topological terms present in the gravity action: namely, a m2 in the bulk action, that
106
+
107
+ 4
108
+ is, an antisymmetric tensor field Mµν which is the effective polarization tensor of the term in the
109
+ boundary action on the RS branes [42–44]. Such behaviors are expected for a quantum Hall system
110
+ tuned to a quantized value of the conductivity. Furthermore, we provided similar results in the
111
+ AdS/BCFT holographic model, where, for example, we will see how accurately it can account for
112
+ the physical behaviors expected in a quantum Hall system where, as was showed before, through
113
+ AdS/BCFT construction the Hall conductivity is inversely proportional to the coefficients of the
114
+ terms that appear in the gravity Lagrangian. Additionally, the ratio ρ/B will indicate a localized
115
+ condensate [45, 46].
116
+ Just for completeness, as discussed in [6], for the classical Hall effect, the charge density and the
117
+ external magnetic field are independent quantities, that is, the ρ/B ratio depends on the density
118
+ of conductance electrons. On the other hand, in the quantum Hall Effect (QHE) the transverse
119
+ conductivity given by σH, has plateaus that are independent of either ρ or B. These plateaus are
120
+ generally attributed to disorder [47–49], being responsible for the existence of localized electron
121
+ states [6]. Here, the localized states fill the gaps between the Landau levels. Nevertheless, there is
122
+ no active participation in the Hall conductivity.
123
+ Finally, we study the properties of holographic paramagnetism-ferromagnetism phase transition
124
+ in the presence of Horndeski gravity (1). Here, from the matter field part, we consider the effects
125
+ of the Maxwell field (2) on the phase transition of this system, following [50, 51], introducing a
126
+ massive 2-form coupled field, and neglect the effects of this 2-form field and gauge fields on the
127
+ background geometry. In our analysis, we observe that increasing the strength of parameter γ,
128
+ given in (1), decreases the temperature of the holographic model and leads to a harder formation
129
+ of the magnetic moment in the black hole background. On the other hand, at low temperatures,
130
+ spontaneous magnetization, and ferromagnetic phase transition happen, but when removes the
131
+ external magnetic field, this magnetization disappears. As we know, ferromagnetic materials have
132
+ coercivity, which is the ability to keep their elementary magnets stuck in a certain position. This
133
+ position can be modified by placing the magnetized material in the presence of an external magnetic
134
+ field. In this way, a material with high coercivity its elementary magnets resists the change of
135
+ position. In the material science, experimental framework [52], there is a close relationship between
136
+ the magnetic related to viscosity and coercivity, this relationship was predicted theoretically and
137
+ observed experimentally. Thus, we have a fundamental role in both cases, that is, between viscosity
138
+ and coercivity, where they play the so-called activation volume, which is the relevant volume where
139
+ thermally activated and field-induced magnetization processes occur, respectively. In our work, we
140
+ will study this way for the paramagnetic material to resist the external magnetic field, through the
141
+
142
+ 5
143
+ viscosity/entropy ratio. In our model, this relationship depends on the external magnetic field, the
144
+ Horndeski parameters, and the boundary size ∆ yQ of the RS brane in a non-trivial way.
145
+ This work is organized as follows: In Section II we consider the gravitational setup, which con-
146
+ tains all the information with respect to the AdS4/BCFT3 duality, showing the solution. Together
147
+ with the above, in Section III the charge density is obtained for then, in Section IV to present
148
+ the boundary Q profile. In Section V, we perform a holographic renormalization, computing the
149
+ Euclidean on-shell action, which is related to the free energy of the corresponding thermodynamic
150
+ system, where in particular we will focus on the black hole entropy, present in Section VI, and
151
+ the holographic paramagnetism/ferromagnetism phase transition, given in Section VII. Finally,
152
+ Section VIII is devoted to our conclusions and discussions.
153
+ II.
154
+ BLACK HOLE AS A PROBE OF ADS/BCFT
155
+ As was shown in the introduction, we will present our setup starting with the total action,
156
+ which contains all information related to AdS4/BCFT3 correspondence with probe approximation,
157
+ so that:
158
+ S = SN
159
+ H + SN
160
+ M + SN
161
+ 2−FF + SN
162
+ mat + SQ
163
+ bdry + SQ
164
+ mat + SQ
165
+ ct,
166
+ (3)
167
+ where
168
+ SN
169
+ H =
170
+
171
+ N
172
+ d4x√−g LH,
173
+ SN
174
+ M =
175
+
176
+ N
177
+ d4x√−g LM,
178
+ (4)
179
+ with LH and LM given previously in (1)-(2) respectively, while that SN
180
+ mat is the action associated
181
+ to matter sources and:
182
+ SQ
183
+ bdry = 2κ
184
+
185
+ Q
186
+ d3x
187
+
188
+ −hLbdry
189
+ SQ
190
+ mat = 2
191
+
192
+ Q
193
+ d3x
194
+
195
+ −hLmat,
196
+ SQ
197
+ ct = 2κ
198
+
199
+ ct
200
+ d3x
201
+
202
+ −hLct ,
203
+ (5)
204
+ with
205
+ Lbdry = (K − Σ) − γ
206
+ 4(∇µφ∇νφnµnν − (∇φ)2)K − γ
207
+ 4∇µφ∇νφKµν ,
208
+ (6)
209
+ Lct = c0 + c1R + c2RijRij + c3R2 + b1(∂iφ∂iφ)2 + · · · ,
210
+ (7)
211
+ where in our notations (∇φ)2 = ∇µφ∇µφ. In Eq.(6), Lbdry corresponds to the Gibbons-Hawking γ-
212
+ dependent terms associated with the Horndeski gravity (1), where Kµν = h β
213
+ µ ∇βnν is the extrinsic
214
+
215
+ 6
216
+ curvature, K = hµνKµν is the trace of the extrinsic curvature, hµν is the induced metric, nµ is an
217
+ outward pointing unit normal vector to the boundary of the hypersurface Q, Σ is the boundary
218
+ tension on Q.
219
+ Lmat is the matter Lagrangian on Q, while that in Eq. (7) Lct represents the
220
+ boundary counterterms, which do not affect the bulk dynamics and will be neglected.
221
+ Following the procedures presented by [3, 4, 6, 10, 11] we have imposed the NBC:
222
+ Kαβ − hαβ(K − Σ) − γ
223
+ 4Hαβ = κSQ
224
+ αβ ,
225
+ (8)
226
+ where
227
+ Hαβ ≡ (∇σφ∇ρφ nσnρ − (∇φ)2)(Kαβ − hαβK) − (∇αφ∇βφ)K ,
228
+ (9)
229
+ SQ
230
+ αβ = −
231
+ 2
232
+
233
+ −h
234
+ δSQ
235
+ mat
236
+ δhαβ .
237
+ (10)
238
+ Considering the matter stress-energy tensor on Q as a constant (this is SQ
239
+ αβ = 0), we can write
240
+ Kαβ − hαβ(K − Σ) − γ
241
+ 4Hαβ = 0 .
242
+ (11)
243
+ On the other hand, from the gravitational part, given by the Einstein-Horndeski theory and as-
244
+ suming that SN
245
+ mat is constant, varying SN
246
+ H and SQ
247
+ bdry with respect to the dynamical fields, we have:
248
+ Eαβ = −
249
+ 2
250
+ √−g
251
+ δSN
252
+ δgαβ ,
253
+ Eφ = −
254
+ 2
255
+ √−g
256
+ δSN
257
+ δφ ,
258
+ Fφ = −
259
+ 2
260
+
261
+ −h
262
+ δSQ
263
+ bdry
264
+ δφ
265
+ ,
266
+ (12)
267
+ where
268
+ Eµν = Gµν + Λgµν − α
269
+ 2
270
+
271
+ ∇µφ∇νφ − 1
272
+ 2gµν∇λφ∇λφ
273
+
274
+ + γ
275
+ 2
276
+ �1
277
+ 2∇µφ∇νφR − 2∇λφ∇(µφRλ
278
+ ν) − ∇λφ∇ρφRµλνρ
279
+
280
+ + γ
281
+ 2
282
+
283
+ −(∇µ∇λφ)(∇ν∇λφ) + (∇µ∇νφ)2φ + 1
284
+ 2Gµν(∇φ)2
285
+
286
+ − γgµν
287
+ 2
288
+
289
+ −1
290
+ 2(∇λ∇ρφ)(∇λ∇ρφ) + 1
291
+ 2(2φ)2 − (∇λφ∇ρφ)Rλρ
292
+
293
+ ,
294
+ (13)
295
+ Eφ = ∇µ [(αgµν − γGµν) ∇νφ] ,
296
+ (14)
297
+ Fφ = −γ
298
+ 4(∇µ∇νφnµnν − (∇2φ))K − γ
299
+ 4(∇µ∇νφ)Kµν ,
300
+ (15)
301
+ and note that, Eφ = Fφ, from the Euler-Lagrange equation.
302
+ Together with the above, and according to [27–31] , we have a condition that deals to static
303
+ black hole configurations, avoiding no-hair theorems [32]. Here, we need to require that the square
304
+
305
+ 7
306
+ of the radial component of the conserved current must vanish identically without restricting the
307
+ radial dependence of the scalar field. Such discussion implies that in Eq. (14):
308
+ αgrr − γGrr = 0 ,
309
+ (16)
310
+ and defining φ′(r) ≡ ψ(r), where (′) denotes the derivative with respect to r, we can show that the
311
+ equations Eφ = 0 = Err are satisfied. In our setup, the four dimensional metric is defined via the
312
+ following line element
313
+ ds2 = L2
314
+ r2
315
+
316
+ −f(r) dt2 + dx2 + dy2 + dr2
317
+ f(r)
318
+
319
+ ,
320
+ (17)
321
+ where x1 ≤ x ≤ x2 and y1 ≤ y ≤ y2, while that from Refs.[10, 20, 30], f(r) is the metric function
322
+ which takes the form
323
+ f(r) = αL2
324
+
325
+
326
+ 1 −
327
+ � r
328
+ rh
329
+ �3�
330
+ ,
331
+ (18)
332
+ while that ψ(r) reads
333
+ ψ2(r) = (φ′(r))2 = −2L2(α + γΛ)
334
+ αγr2f(r)
335
+ ,
336
+ (19)
337
+ where
338
+ φ(r) = ±2
339
+
340
+ −6(α + Λγ)
341
+
342
+ tanh−1
343
+ ��
344
+ 1 − r3
345
+ r3
346
+ h
347
+
348
+ + φ0.
349
+ (20)
350
+ Here, φ0 and rh are integration constants, where the last one is related to the location of the event
351
+ horizon. Following the steps of [10, 20], performing the transformations
352
+ f(r) → αL2
353
+ 3γ f(r),
354
+ t → 3γ
355
+ αL2 t,
356
+ x →
357
+
358
+
359
+ αL2 x,
360
+ y →
361
+
362
+
363
+ αL2 y,
364
+ L →
365
+ � α
366
+ 3γ L2,
367
+ (21)
368
+ we have that the line element (17) is invariant, but now the metric function f(r) takes the form
369
+ f(r) = 1 −
370
+ � r
371
+ rh
372
+ �3
373
+ (22)
374
+ while the square of the derivative of the scalar field ψ2(r) takes the form given previously in (19).
375
+ Here is important to note that from Eqs. (19)-(20) we can see that to have a real scalar field,
376
+ α + Λγ ≤ 0,
377
+ where it vanishes when α = −Λγ.
378
+ It is important to note that, from the action (3), we can see that there is another contribution,
379
+ denoted as SN
380
+ 2−FF, which is responsible to construct the ferromagnetic/paramagnetic model. The
381
+ above will be explained in the following section.
382
+
383
+ 8
384
+ III.
385
+ THE FINITE CHARGE DENSITY
386
+ As was shown in the previous section, in the action (3) appears the additional contribution
387
+ SN
388
+ 2−FF = λ2
389
+
390
+ N
391
+ d4x√−g L2−FF,
392
+ where
393
+ L2−FF = − 1
394
+ 12(dM)2 − m2
395
+ 4 MµνMµν − 1
396
+ 2MµνFµν − J
397
+ 8 V (M).
398
+ (23)
399
+ Here, the above action defined from the seminal works [42, 43], is coupled through the constant λ
400
+ and constructed via the 2-form Mµν, dM is the exterior differential of the 2-form field Mµν, this is
401
+ (dM)τµν = 3∇[τMµν] and (dM)2 = 9∇[τMµν]∇[τMµν], m is a constant related to the mass, while
402
+ that V (M) describes the self-interaction of polarization tensor, with J a constant, which reads
403
+ V (M) = (∗MµνMµν)2 = [∗(M ∧ M)]2,
404
+ (24)
405
+ where (∗) is the Hodge star operator, this is ∗Mµν =
406
+ 1
407
+ 2!εαβ
408
+ µνMαβ and εαβ
409
+ µν is the Levi-Civita
410
+ Tensor. In the following lines, will restrict our analysis to the probe approximation, that is, from
411
+ the action Eq. (3), one can derive the corresponding equations of motions for matter fields in the
412
+ probe approximation, that is, e2 → ∞ and λ → 0, so that:
413
+ ∇µ
414
+
415
+ Fµν + λ2
416
+ 4 Mµν
417
+
418
+ = 0,
419
+ (25)
420
+ ∇τ(dM)τµν − m2Mµν − J(∗MτσMτσ)(∗Mµν) − Fµν = 0 .
421
+ (26)
422
+ Given that we are focusing on the probe limit approximation, we are going to disregard any
423
+ back reaction coming from the two-form field Mµν. In order to analyze the holographic paramag-
424
+ netism/ferromagnetism and paraelectric/ferroelectric phase transition, we consider the gauge fields
425
+ Mµν and Aµ we consider the following ansatz:
426
+ Mµν = −p(r) dt ∧ dr + ρ(r) dx ∧ dy,
427
+ (27)
428
+ Aµ = At(r) dt + Bx dy,
429
+ F = dA,
430
+ (28)
431
+ where B is the external magnetic field. Using (17), (27)-(28) in the background (22), the field
432
+ equations (25) and (26) are given by
433
+ A′
434
+ t +
435
+
436
+ m2 − 4 J r4 ρ2
437
+ L4
438
+
439
+ p = 0,
440
+ (29)
441
+ ρ′′
442
+ L2 +
443
+ �f′
444
+ f + 2
445
+ r
446
+ � ρ′
447
+ L2 −
448
+ �4 J r2 p2
449
+ fL4
450
+ + m2
451
+ r2 f
452
+
453
+ ρ − B
454
+ r2 f = 0,
455
+ (30)
456
+ A′′
457
+ t + λ2
458
+ 4 p′ = 0 ,
459
+ (31)
460
+
461
+ 9
462
+ As we work with probe approximation, the back reaction can be neglected. Together with the
463
+ above, given that the behaviors are asymptotically AdS4, we can solve the field equations (29)-(31)
464
+ near to the boundary (this is r → 0). Here, asymptotic solutions are given by
465
+ At(r) ∼ µ − σr,
466
+ (32)
467
+ p(r) ∼ σ
468
+ m2 ,
469
+ (33)
470
+ ρ(r) ∼ ρ+r∆+ + ρ−r∆− − B
471
+ m2 ,
472
+ (34)
473
+ ∆± = −1 ±
474
+
475
+ 1 + 4L2m2
476
+ 2
477
+ .
478
+ (35)
479
+ Here, ρ+ and ρ− correspond to the source and vacuum expectation value of the dual operator in
480
+ the boundary field theory (up to a normalization factor), respectively. It is worth pointing out
481
+ that one should take ρ+ = 0, in order to obtain condensation spontaneously [43]. From Eq. (34),
482
+ we can define ρ+ and ρ− as
483
+ ρ+ ≡ r−∆+
484
+ h
485
+ ,
486
+ ρ− ≡ r−∆−
487
+ h
488
+ ,
489
+ (36)
490
+ yielding to the asymptotic solution ρ(r) the following structure
491
+ ρ(r) ∼
492
+ � r
493
+ rh
494
+ �∆+
495
+ +
496
+ � r
497
+ rh
498
+ �∆−
499
+ − B
500
+ m2 .
501
+ (37)
502
+ Additionally, and according to [8], we can to analyze the electromagnetic field, extracted from the
503
+ four dimensional electromagnetic duality, in a sense that the theory is invariant under
504
+ Fµν →∗ Fµν = 1
505
+ 2εµναβF αβ,
506
+ (38)
507
+ where, as before, εαβµν is the Levi-Civita Tensor, transforming the electric field into a magnetic field
508
+ and vice versa. Such duality gives that, from the action (2), FµνF µν = (∗Fµν)(∗F µν), showing that
509
+ is invariant under (38). Besides, the transformation (38) shows that Frt → (∗Frt) = Fxy = σ = B,
510
+ where σ (B) is the constant related to the electric (magnetic) field.
511
+ IV.
512
+ Q-BOUNDARY PROFILE
513
+ In this section, we present the boundary Q profile, we assume that Q is parameterized by the
514
+ equation y = yQ(r), analyzing the influence of the Horndeski action (1), (4). For this, to find the
515
+ extrinsic curvature, one has to consider the induced metric on this surface, which reads
516
+ ds2
517
+ ind = L2
518
+ r2
519
+
520
+ −f(r)dt2 + dx2 + g2(r)dr2
521
+ f(r)
522
+
523
+ ,
524
+ (39)
525
+
526
+ 10
527
+ where g2(r) = 1 + y′2(r)f(r) and (′) denotes the derivative with respect to the coordinate r. Here,
528
+ the normal vectors on Q are represented by
529
+ nµ =
530
+ r
531
+ Lg(r)
532
+
533
+ 0, 0, 1, −f(r)y′(r)
534
+
535
+ .
536
+ (40)
537
+ Considering the field equation Fφ = 0 (15), one can solve the Eq. (11), yielding
538
+ y′(r) =
539
+ (ΣL)
540
+
541
+
542
+
543
+ �4 + γψ2(r)
544
+ 4
545
+ − (ΣL)2
546
+
547
+ 1 −
548
+ � r
549
+ rh
550
+ �3� ,
551
+ (41)
552
+ and, with ψ2(r) given previously in Eq. (19), we have
553
+ y′(r) =
554
+ (ΣL)
555
+
556
+
557
+
558
+
559
+
560
+
561
+ 4 −
562
+ ξL2
563
+ 2r2
564
+
565
+ 1 −
566
+ � r
567
+ rh
568
+ �3� − (ΣL)2
569
+
570
+ 1 −
571
+ � r
572
+ rh
573
+ �3� ,
574
+ (42)
575
+ where we define
576
+ ξ = α + γΛ
577
+ α
578
+ .
579
+ (43)
580
+ With all this information, we can plot the yQ profile from Eq. (42), representing the holographic
581
+ description of BCFT considering the theory (1).
582
+ FIG. 2: The figure shows the numerical solution for Q boundary profile from Eq. (42) for the black hole
583
+ within Horndeski gravity, considering the values for θ′ = 2π/3, θ = π − θ′, Λ = −1, α = 8/3 with γ = 0
584
+ (pink curve), γ = 0.1 (blue dashed curve ), γ = 0.2 (red dot dashed curve), and γ = 0.3 (green thick curve).
585
+ The dashed parallel vertical lines represent the UV solution, Eq. (46), that is, Randall-Sundrum branes.
586
+ The region between the curves Q represents the bulk N.
587
+
588
+ Yo
589
+ -yo
590
+ 0
591
+ -
592
+ Q
593
+ N
594
+ N
595
+ rh
596
+ r11
597
+ On the other hand, following the steps of [6, 7], we have that the NBC on the gauge field
598
+ is nµFµν|Q = 0, and B = σ. The holographic model (AdS4/BCFT3) predicts that a constant
599
+ boundary current in the bulk induces a constant current on the boundary Q.
600
+ Such boundary
601
+ current can be measured in materials graphene-like. Furthermore, nµMµν|Q = 0 provide
602
+ ρ(r)
603
+ B
604
+ = f(r)y′(r)
605
+ m2
606
+ .
607
+ (44)
608
+ Here, the density ρ and the magnetic field B are no longer two independent parameters. As the
609
+ ratio is the Hall conductivity, this is very reminiscent of the quantum Hall effect (QHE), where this
610
+ ratio is independent of both ρ and B and is inversely proportional to the topological coefficients,
611
+ which in our case are the coupling constant γ presents in the Horndeski gravity, together with the
612
+ parameter from the antisymmetric tensor field Mµν, this is m2. In our case, the equation of y′ from
613
+ (42) and then the ρ/B ratio (44) can be analyzed by numerical calculations, being represented in
614
+ Fig. 3. Here, we show the ratio ρ/B with respect to external magnetic field B for different values
615
+ of the Horndeski gravity parameter γ, where we introduced ΣL = cos(θ′), where θ′ represents
616
+ the angle between the positive direction of the y axis and Q. At the boundary Q, the curves of
617
+ solutions in the (ρ, B) plane will be a localized condensate [45, 46].
618
+ 0.05
619
+ 0.06
620
+ 0.07
621
+ 0.08
622
+ 0.09
623
+ 0.10
624
+ 0.11
625
+ 0.000
626
+ 0.005
627
+ 0.010
628
+ 0.015
629
+ 0.020
630
+ 0.025
631
+ 0.030
632
+ r
633
+ Ρ
634
+ B
635
+ FIG. 3: Graphic of the ratio ρ/B with respect to external magnetic field B versus r, for different values of
636
+ the Horndeski parameter γ. Here, we consider the values rh = 0.1, L = 1, θ′ = 2π/3, Λ = −1, α = 0.5,
637
+ m = 1, and γ = 1 (represented through the blue curve), γ = 4 (represented through the red curve), and
638
+ γ = 8 (represented through the green curve).
639
+
640
+ 12
641
+ Together with the above, in addition to the above numerical solution, we can analyze some
642
+ particular cases regarding the study of the UV and IR regimes. Thus, for the first case, performing
643
+ an expansion at r → 0 with, as before, ΣL = cos(θ′), the equation (42) becomes
644
+ yUV (r) = y0 +
645
+
646
+ 2
647
+ −ξL2 r cos(θ′),
648
+ (45)
649
+ where y0 is an integration constant. In the above equation, considering ξ → −∞, we have
650
+ yUV (r) = y0 = constant.
651
+ (46)
652
+ This is equivalent to keeping ξ finite and a zero tension limit Σ → 0, considering the cases θ′ = π/2
653
+ and θ′ = 3π/2. Now, for this regime, we have that the ρ/B ratio takes the form
654
+ ρ
655
+ B =
656
+
657
+ 2
658
+ −ξL2
659
+ cos(θ′)
660
+ m2
661
+ .
662
+ (47)
663
+ Here, it turns out a straightforward generalization of a known AdS4/CFT3 solution, given by
664
+ the plane-symmetric charged four-dimensional AdS black hole, where only allows for tensionless
665
+ RS branes in the AdS4/BCFT3 construction [6]. In this case, requires that the static uniform
666
+ charge density is supported by a magnetic field. Specifically, we found that ρ/B is a constant
667
+ proportional to a ratio of the coefficients appearing in the Horndeski gravity.
668
+ These analyses
669
+ indicate a generalization of the AdS4 black hole can describe a quantum Hall system at a plateau
670
+ of the transverse conductivity. Additionally, the AdS/BCFT setup yields that the Hall conductivity
671
+ is inversely proportional to a sum of the coefficients of the topological terms appearing in the gravity
672
+ Lagrangian. That is, we obtain that σH = ρ/B, which from the equation (47)
673
+ σH =
674
+
675
+ 2
676
+ −ξL2
677
+ cos(θ′)
678
+ m2
679
+ ,
680
+ (48)
681
+ where, as was shown in the introduction, in QHE the conductivity is related to the number of filled
682
+ Landau levels (filling fraction), namely, by
683
+ h
684
+ e2 σH =
685
+
686
+ 2
687
+ −ξL2
688
+ cos(θ′)
689
+ m2
690
+ ,
691
+ (49)
692
+ where e2/h is the magnetic flux quantum.
693
+ In this way, the holographic description seems to
694
+ provide results similar to the description of the QHE obtained in [48, 49]. In our case, we have an
695
+ extension of the covariant form of the Hall relation ρ = σHB.
696
+ For the IR case, we take r → +∞ so that from Eq. (42) implies that limr→+∞(φ′(r))2 = 0, and
697
+ then φ = constant, which ensures a genuine vacuum solution. Plugging this result in Eq. (42), in
698
+ the limit r → ∞, we have
699
+ y′
700
+ IR(r) ∼
701
+ �rh
702
+ r
703
+ �3/2
704
+ + O
705
+ � 1
706
+ r2
707
+
708
+ ,
709
+ (50)
710
+
711
+ 13
712
+ and y
713
+
714
+ IR(r) → 0 when r → +∞, which implies from (47) that ρ/B → 0. Such value becomes the
715
+ on-shell action finite.
716
+ For the sake of completeness, an approximate analytical solution for y(r) can be obtained by
717
+ performing an expansion for ξ very small from Eq. (42), this is
718
+ y′
719
+ Q =
720
+ cos(θ′)
721
+
722
+ 4 − cos2(θ′)f(r)
723
+ +
724
+ L2 cos(θ)ξ
725
+ 4r2f(r)(4 − cos2(θ′)f(r))3/2 + O(ξ2),
726
+ with f given previously in (22), and considering this expansion up to the first order, we obtain
727
+ yQ(r) = y0 +
728
+ r cos(θ′)
729
+
730
+ (r3 − r3
731
+ h) cos2(θ′) + 4r3
732
+ h
733
+
734
+ 4r3
735
+ h − (r3 − r3
736
+ h) cos(2θ′)
737
+ 4 − cos2(θ′)
738
+ ×2F1
739
+ �1
740
+ 3, 1
741
+ 2; 4
742
+ 3; −
743
+ r3 cos2(θ′)
744
+ r3
745
+ h(4 − cos2(θ′))
746
+
747
+ + ξ
748
+
749
+ L2 cos(θ)
750
+ 4r2f(r)(4 − cos2(θ′)f(r))3/2 dr + O(ξ2),
751
+ (51)
752
+ where 2F1(a, b; c; x) is the hypergeometric function.
753
+ V.
754
+ HOLOGRAPHIC RENORMALIZATION
755
+ In our setup, we will compute the Euclidean on-shell action, which is related to the free energy
756
+ of the corresponding thermodynamic system. Thus, our holographic scheme takes into account
757
+ the contributions of AdS4/BCFT3 correspondence within Horndeski gravity. Let us start with the
758
+ Euclidean action given by IE = Ibulk + 2Ibdry, i.e.,
759
+ Ibulk = −
760
+ 1
761
+ 16πGN
762
+
763
+ N
764
+ √gd4x
765
+
766
+ R − 2Λ + γ
767
+ 2Gµν∇µφ∇νφ
768
+
769
+
770
+ 1
771
+ 8πGN
772
+
773
+ M
774
+ d3x√¯γLM,
775
+ (52)
776
+ LM = K(¯γ) − Σ(¯γ) − γ
777
+ 4(∇µφ∇νφnµnν − (∇φ)2)K(¯γ) − γ
778
+ 4∇µφ∇νφK(¯γ)
779
+ µν .
780
+ (53)
781
+ Together with the above, g is the determinant of the metric gµν on the bulk N, while that ¯γ is the
782
+ induced metric, the surface tension on M is represented with Σ(¯γ), and K(¯γ) corresponds to the
783
+ extrinsic curvature on M. On the other hand, for the boundary, we have the following expressions
784
+ Ibdry = −
785
+ 1
786
+ 16πGN
787
+
788
+ N
789
+ √gd4x
790
+
791
+ R − 2Λ + γ
792
+ 2Gµν∇µφ∇νφ
793
+
794
+
795
+ 1
796
+ 8πGN
797
+
798
+ Q
799
+ d3x
800
+
801
+ hLbdry,
802
+ (54)
803
+ Lbdry = (K − Σ) − γ
804
+ 4(∇µφ∇νφnµnν − (∇φ)2)K − γ
805
+ 4∇µφ∇νφKµν.
806
+ (55)
807
+ Thus, in order to compute the bulk action Ibulk, we consider the induced metric on the bulk, which
808
+ is obtained from (17) after the transformation τ = it, given by
809
+ ds2
810
+ ind = ¯γµνdxµdxν = L2
811
+ r2
812
+
813
+ f(r)dτ 2 + dx2 + dy2 + dr2
814
+ f(r)
815
+
816
+ .
817
+ (56)
818
+
819
+ 14
820
+ Here, we have that 0 ≤ τ ≤ β, where from Eq. (22)
821
+ β = 1
822
+ T =
823
+ �|f′(rh)|
824
+
825
+ �−1
826
+ = 4πrh
827
+ 3
828
+ ,
829
+ (57)
830
+ where T is the Hawking Temperature, the above allows us to obtain the following quantities:
831
+ R = − 12
832
+ L2 ,
833
+ Λ = − 3
834
+ L2 ,
835
+ K(¯γ) = 3
836
+ L,
837
+ Σ(¯γ) = 2
838
+ L.
839
+ Thus, we have all elements needed to construct the bulk action Ibulk. In the process of holographic
840
+ renormalization, we need to introduce a cutoff ϵ to remove the IR divergence on the bulk side and
841
+ we can provide that:
842
+ Ibulk =
843
+ 1
844
+ 16πGN
845
+
846
+ d2x
847
+
848
+ 4πrh
849
+ 3
850
+ 0
851
+
852
+ � rh
853
+ ϵ
854
+ dr√g
855
+
856
+ R − 2Λ + γ
857
+ 2Grrψ2(r)
858
+
859
+ +
860
+ 1
861
+ 16πGN
862
+
863
+ d2x
864
+
865
+ 4πrh
866
+ 3
867
+ 0
868
+ dτ L2�
869
+ f(ϵ)
870
+ ϵ3
871
+ ,
872
+ (58)
873
+ Ibulk = − L2V
874
+ 8r2
875
+ hG
876
+
877
+ 1 − ξ
878
+ 4
879
+
880
+ ,
881
+ (59)
882
+ with ξ given previously in (61) and, in our notations, V =
883
+
884
+ d2x = ∆x∆y = (x2 − x1)(y2 − y1).
885
+ Now, computing the Ibdry, we introduce a cutoff ϵ to remove the UV divergence on the boundary
886
+ side, and with this information, we have:
887
+ Ibdry = rhL2∆yQ
888
+ 2GN
889
+
890
+ 1 − ξ
891
+ 4
892
+ � � rh
893
+ ϵ
894
+ ∆yQ(r)
895
+ r4
896
+ dr − rhL2 sec(θ′)∆yQ
897
+ 2GN
898
+ � rh
899
+ ϵ
900
+ ∆yQ(r)
901
+ r3
902
+ dr
903
+ (60)
904
+ Here, ∆yQ is a constant and ∆yQ(r) := yQ(r) − y0 is obtained from the equation (51). As we
905
+ know, from the point of view of AdS/CFT correspondence, IR divergences in AdS correspond to
906
+ UV divergences in CFT. This relationship is known as the IR-UV connection. Thus, based on
907
+ this duality, we can reduce the above equation (60) after some eliminations of terms that produce
908
+ divergences to the following form:
909
+ Ibdry = −L2∆ yQ
910
+ 2GN
911
+
912
+ 1 − ξ
913
+ 4
914
+ � �
915
+ ξ L2b(θ′)
916
+ 5r4
917
+ h
918
+ + q(θ
919
+ ′)
920
+ 4r2
921
+ h
922
+
923
+ +L2 sec(θ′)∆ yQ
924
+ 2GN
925
+
926
+ ξ L2b(θ′)
927
+ 4r3
928
+ h
929
+ + q(θ
930
+ ′)
931
+ 2rh
932
+
933
+ ,
934
+ (61)
935
+ where
936
+ b(θ′) =
937
+ cos(θ′)
938
+ 4(4 − cos2(θ′))3/2 ,
939
+ q(θ′) =
940
+ cos(θ′)
941
+
942
+ 4 − cos2(θ′)
943
+ .
944
+ (62)
945
+ With all the above information, from Eqs. (59) and (61)-(62), we can compute IE = Ibulk + 2Ibdry
946
+ as:
947
+
948
+ 15
949
+ IE = − L2V
950
+ 8r2
951
+ hGN
952
+
953
+ 1 − ξ
954
+ 4
955
+
956
+ − L2∆ yQ
957
+ GN
958
+
959
+ 1 − ξ
960
+ 4
961
+ � �
962
+ ξ L2b(θ′)
963
+ 5r4
964
+ h
965
+ + q(θ
966
+ ′)
967
+ 4r2
968
+ h
969
+
970
+ +L2 sec(θ′)∆ yQ
971
+ GN
972
+
973
+ ξ L2b(θ′)
974
+ 4r3
975
+ h
976
+ + q(θ
977
+ ′)
978
+ 2rh
979
+
980
+ .
981
+ (63)
982
+ Here, IE is the approximated analytical expression for the Euclidean action.
983
+ This equation is
984
+ essential to construct the free energy and extract all thermodynamic quantities in our setup, as we
985
+ show in the next section.
986
+ VI.
987
+ BLACK HOLE ENTROPY
988
+ Now, we will compute the entropy related to the black hole considering the contributions of the
989
+ AdS/BCFT correspondence in the Horndeski gravity. Free energy is defined as
990
+ Ω = TIE ,
991
+ (64)
992
+ one can obtain the corresponding entropy as:
993
+ S = −∂ Ω
994
+ ∂T
995
+ (65)
996
+ where T is the Hawking Temperature. By plugging the Euclidean on-shell action IE from Eq.(63),
997
+ and replacing T obtained previously in (57), we have
998
+ Stotal = Sbulk + Sbdry,
999
+ (66)
1000
+ where
1001
+ Sbulk =
1002
+ L2V
1003
+ 4r2
1004
+ hGN
1005
+
1006
+ 1 − ξ
1007
+ 4
1008
+
1009
+ ,
1010
+ (67)
1011
+ Sbdry = L2∆ yQ
1012
+ GN
1013
+
1014
+ 1 − ξ
1015
+ 4
1016
+ � �
1017
+ ξ L2b(θ′)
1018
+ 5r4
1019
+ h
1020
+ + q(θ
1021
+ ′)
1022
+ 4r2
1023
+ h
1024
+
1025
+ − L2 sec(θ′)∆ yQ
1026
+ GN
1027
+
1028
+ ξ L2b(θ′)
1029
+ 4r3
1030
+ h
1031
+ + q(θ
1032
+ ′)
1033
+ 2rh
1034
+
1035
+ .
1036
+ (68)
1037
+ The interpretation for this total entropy can be identified with the Bekenstein-Hawking formula
1038
+ for the black hole:
1039
+ SBH =
1040
+ A
1041
+ 4GN
1042
+ ,
1043
+ (69)
1044
+
1045
+ 16
1046
+ where, in this case
1047
+ A = L2V
1048
+ 2r2
1049
+ h
1050
+
1051
+ 1 − ξ
1052
+ 4
1053
+
1054
+ + 4L2∆ yQ
1055
+
1056
+ 1 − ξ
1057
+ 4
1058
+ � �
1059
+ ξ L2b(θ′)
1060
+ 5r4
1061
+ h
1062
+ + q(θ
1063
+ ′)
1064
+ 4r2
1065
+ h
1066
+
1067
+ −4L2 sec(θ′)∆ yQ
1068
+
1069
+ ξ L2b(θ′)
1070
+ 4r3
1071
+ h
1072
+ + q(θ
1073
+ ′)
1074
+ 2rh
1075
+
1076
+ .
1077
+ (70)
1078
+ Here, A is the total area of the AdS black hole in the Horndeski contribution terms for the bulk
1079
+ and the boundary Q. We can see that the information is bounded by the black hole area. Then,
1080
+ the equation (70) suggests that the information storage increases with increasing |ξ|, as long as
1081
+ ξ < 0.
1082
+ Together with the above, with respect to the boundary contribution of (68), we have that this
1083
+ expression is the entropy of the BCFT corrected by the Horndeski terms parametrized by ξ, given
1084
+ previously in (43). In this case, the results presented in Refs. [6, 11] are recovered in the limit
1085
+ ξ → 0. Besides, still analyzing Eq. (68), due to the effects of the Horndeski gravity, there is a
1086
+ non-zero boundary entropy even if we consider the zero temperature scenario, similar to an extreme
1087
+ black hole. This can be seen if one takes the limit T → 0 (or rh → ∞) in Eq.(68), then we do not
1088
+ get the denominated residual boundary entropy, as discussed in [10].
1089
+ On the other hand, through Eq. (47) we have
1090
+ Smagnetic
1091
+ bdry
1092
+ = L2∆ yQ
1093
+ GN
1094
+
1095
+ 1 − ξ
1096
+ 4
1097
+ � �
1098
+ −2B2 cos2(θ
1099
+ ′)
1100
+ m2ρ2
1101
+ b(θ′)
1102
+ 5r4
1103
+ h
1104
+ + q(θ
1105
+ ′)
1106
+ 4r2
1107
+ h
1108
+
1109
+ − L2 sec(θ′)∆ yQ
1110
+ GN
1111
+
1112
+ −2B2 cos2(θ
1113
+ ′)
1114
+ m2ρ2
1115
+ b(θ′)
1116
+ 4r3
1117
+ h
1118
+ + q(θ
1119
+ ′)
1120
+ 2rh
1121
+
1122
+ ,
1123
+ (71)
1124
+ where m2 > −1/(4L2). For the entropy bound, the restriction on m2 comes from Eq. (35). A
1125
+ well-defined probe limit demands that the charge density contributed by the polarization should
1126
+ be finite. At low temperatures, below the critical, in the ferromagnetic region, we can observe that
1127
+ our entropy is Sbdry
1128
+ magnetic ∝ B2, that is, has a square dependence on the external magnetic field
1129
+ and this is a characteristic of ferromagnetic systems. Furthermore, we can observe that Smagnetic
1130
+ bdry
1131
+ is the magnetic entropy of the boundary Q, and we can observe that for ferromagnetic materials,
1132
+ the magnetic entropy is associated with the disorder of the magnetic moments. In addition, these
1133
+ materials have spontaneous magnetization. So when we remove the applied magnetic field, they
1134
+ still show magnetization.
1135
+
1136
+ 17
1137
+ VII.
1138
+ HOLOGRAPHIC PARAMAGNETISM/FERROMAGNETISM PHASE
1139
+ TRANSITION
1140
+ In this section, we present the holographic paramagnetism/ferromagnetism phase transition
1141
+ through the boundary contribution from the entropy (71). For this, we start considering the free
1142
+ energy Ω from (63) -(64), where the first law of black holes thermodynamics, considering the
1143
+ canonical ensemble, takes the form
1144
+ dΩ = −PdV − SdT,
1145
+ (72)
1146
+ where, in addition to the entropy S as well as the Hawking temperature T, the pressure P and the
1147
+ volume V appear, yielding
1148
+ Ω = ϵ − TS,
1149
+ where ϵ takes the role of the energy density.
1150
+ As a first thermodynamic quantity to study, we will consider the entropy S, from Eq. (66),
1151
+ calculated in the previous section, and represented graphically in Fig.
1152
+ 4, with respect to the
1153
+ Hawking temperature T (57).
1154
+ Here, in the right panel (left panel) there is (not) an external
1155
+ magnetic field B. Concretely, we see that the right panel exhibit similar behavior as analyzed in
1156
+ [53], as for example ferromagnetic materials with nearly zero coercivity and hysteresis. On the
1157
+ other hand, in the left panel, when the external magnetic field is removed (this is B = 0), we still
1158
+ have a disorder of magnetic moments, this is a characteristic of paramagnetism.
1159
+ The second parameter that we analyze is the heat capacity CV , which allows us to analyze local
1160
+ thermodynamic stability, defined in the following form
1161
+ CV = T
1162
+ �∂S
1163
+ ∂T
1164
+
1165
+ V
1166
+ = −T
1167
+ �∂2Ω
1168
+ ∂T 2
1169
+
1170
+ V
1171
+ ,
1172
+ (73)
1173
+ where the sub-index V from Eq. (73) represents at volume constant. From Fig. 5, we can see
1174
+ that in the right panel, the black hole can switch between stable (CV > 0) and unstable (CV < 0)
1175
+ phases, depending on the sign of heat capacity CV .
1176
+ This phase transition occurs, due to the
1177
+ spontaneous electric polarization, which was realized in our model from the application of the
1178
+ magnetic external field. Moreover, in the region CV > 0, we have structures built like magnetic
1179
+ domes on the boundary Q. Additionally, in Fig. 5, one can see the influence of Horndeski gravity
1180
+ (represented via the constant γ) with respect to the temperature T, where the phase transition
1181
+ occurs for some ranges of values for T when the external magnetic field is null, that is, B = 0.
1182
+
1183
+ 18
1184
+ 0.0
1185
+ 0.2
1186
+ 0.4
1187
+ 0.6
1188
+ 0.8
1189
+ 1.0
1190
+ 0.00
1191
+ 0.05
1192
+ 0.10
1193
+ 0.15
1194
+ 0.20
1195
+ T
1196
+ S
1197
+ 0
1198
+ 2
1199
+ 4
1200
+ 6
1201
+ 8
1202
+ 10
1203
+ 0
1204
+ 20
1205
+ 40
1206
+ 60
1207
+ 80
1208
+ 100
1209
+ 120
1210
+ T
1211
+ S�B�0�
1212
+ FIG. 4: Right panel: The behavior of the entropy S with the temperature T with different values for
1213
+ α = 8/3, m = 1/8, B = (4/5)T, ρ = 1/4, Λ = −1, V = 1, GN = 1, θ′ = 2π/3 with γ = 1 (pink curve),
1214
+ γ = 4 (red dot dashed curve), γ = 8 (green thick curve). Left panel: The behavior of the entropy S with
1215
+ respect the temperature T, with different values for B = 0.
1216
+ 0.0
1217
+ 0.2
1218
+ 0.4
1219
+ 0.6
1220
+ 0.8
1221
+ 1.0
1222
+ �6
1223
+ �4
1224
+ �2
1225
+ 0
1226
+ T
1227
+ CV
1228
+ 0.0
1229
+ 0.2
1230
+ 0.4
1231
+ 0.6
1232
+ 0.8
1233
+ 1.0
1234
+ 0.0
1235
+ 0.5
1236
+ 1.0
1237
+ 1.5
1238
+ T
1239
+ CV �B�0�
1240
+ FIG. 5: Right panel: The behavior of the heat capacity CV with the temperature T with different values
1241
+ for α = 8/3, m = 1/8, B = (4/5)T, ρ = 1/4, Λ = −1, θ′ = 2π/3 with γ = 1 (pink curve), γ = 4 (red dot
1242
+ dashed curve), γ = 8 (green thick curve). Left panel: The behavior of the heat capacity CV with respect
1243
+ the temperature T, with different values for B = 0.
1244
+ Additionally, we can obtain the heat capacity at constant pressure CP , which reads
1245
+ CP = T
1246
+ �∂S
1247
+ ∂T
1248
+
1249
+ P
1250
+ ,
1251
+ (74)
1252
+ and, from Fig. 6, we can see that in the right panel, the black hole can switch between stable
1253
+ (CP > 0), describing a ferromagnetic material, and unstable (CP < 0), describing a paramagnetic
1254
+ material, depending on the sign of heat capacity. This phase transition occurs, as in the previous
1255
+ case, due to spontaneous electric polarization. Moreover, in the region CP > 0, we have structures
1256
+
1257
+ 19
1258
+ built like magnetic domes on the boundary Q, wherein the experimental specific frame, these heat
1259
+ curves without magnetic field can represent a material like DyAl2 [53]. On the other hand, the
1260
+ left panel represents the heat capacity CP where B = 0, where we can see, that is locally unstable
1261
+ (CP < 0).
1262
+ 0.0
1263
+ 0.2
1264
+ 0.4
1265
+ 0.6
1266
+ 0.8
1267
+ 1.0
1268
+ �4
1269
+ �2
1270
+ 0
1271
+ 2
1272
+ 4
1273
+ T
1274
+ CP
1275
+ 0.0
1276
+ 0.2
1277
+ 0.4
1278
+ 0.6
1279
+ 0.8
1280
+ 1.0
1281
+ �4
1282
+ �2
1283
+ 0
1284
+ 2
1285
+ 4
1286
+ T
1287
+ CP �B�0�
1288
+ FIG. 6: Right panel: The behavior of the CP with respect to the temperature T with different values for
1289
+ α = 8/3, m = 1/8, B = (4/5)T, ρ = 1/4, Λ = −1, θ′ = 2π/3 with γ = 1 (pink curve), γ = 4 (red dot dashed
1290
+ curve), γ = 8 (green thick curve). Left panel: The behavior of CP with respect T, with different values for
1291
+ B = 0.
1292
+ Additionally, we can derive other quantities, as for example the magnetization density m, and
1293
+ magnetic susceptibility χ, following the steps of [46], given by
1294
+ m = −
1295
+ �∂ Ω
1296
+ ∂B
1297
+
1298
+ = L2∆ yQT
1299
+ GN
1300
+
1301
+ 1 − ξ
1302
+ 4
1303
+ � �
1304
+ 4 cos2(θ
1305
+ ′)
1306
+ m2ρ2
1307
+ b(θ′)
1308
+ 5r4
1309
+ h
1310
+
1311
+ − L2 sec(θ′)∆ yQT
1312
+ GN
1313
+
1314
+ cos(θ
1315
+ ′)
1316
+ m2ρ2
1317
+ b(θ′)
1318
+ 4r3
1319
+ h
1320
+
1321
+ ,
1322
+ (75)
1323
+ χ =
1324
+ � ∂2Ω
1325
+ ∂B2
1326
+
1327
+ = −L2∆ yQT
1328
+ GN
1329
+
1330
+ 1 − ξ
1331
+ 4
1332
+ � �
1333
+ 4B cos2(θ
1334
+ ′)
1335
+ m2ρ2
1336
+ b(θ′)
1337
+ 5r4
1338
+ h
1339
+
1340
+ + L2 sec(θ′)∆ yQT
1341
+ GN
1342
+
1343
+ B cos(θ
1344
+ ′)
1345
+ m2ρ2
1346
+ b(θ′)
1347
+ 4r3
1348
+ h
1349
+
1350
+ . (76)
1351
+ As we can see from equations (75) and (76), the RS brane behaves like a paramagnetism material,
1352
+ that is, when we remove the external magnetic field, the equation (76) disappears and the entropy
1353
+ linked disorder increases, as shown in Fig. 4. On the other hand, from the equation (75), the
1354
+ magnetization density is not null for zero magnetic fields (this is B = 0). Thus, we can conclude
1355
+ that paramagnetic materials have a low coercivity, that is, their ability to remain magnetized is
1356
+ very low. Thus, one way to analyze coercivity is through viscosity η in our model [52].
1357
+
1358
+ 20
1359
+ In order to be as clear as possible, the details about the computation of the shear viscosity and
1360
+ entropy density ratio are present in Appendix A. In particular, we will focus on the η/S ratio, where
1361
+ from Eq. A11 and Fig. 7, we can analyze the dependence of the viscosity on the magnetic field,
1362
+ characterizing a magnetic side effect, and describing the slow relaxation of the magnetization of
1363
+ paramagnetic materials when they acquire magnetization in the presence of an external magnetic
1364
+ field B (left panel of Fig. 7). In the right panel, we can observe that under an interval of the
1365
+ temperature T, the η/S ratio is an increasing function when B = 0.
1366
+ 0.0
1367
+ 0.5
1368
+ 1.0
1369
+ 1.5
1370
+ 2.0
1371
+ 0.0
1372
+ 0.2
1373
+ 0.4
1374
+ 0.6
1375
+ 0.8
1376
+ 1.0
1377
+ 1.2
1378
+ T
1379
+ Η
1380
+ S
1381
+ 0.0
1382
+ 0.2
1383
+ 0.4
1384
+ 0.6
1385
+ 0.8
1386
+ 1.0
1387
+ 0
1388
+ 20
1389
+ 40
1390
+ 60
1391
+ 80
1392
+ 100
1393
+ T
1394
+ Η
1395
+ S
1396
+ �B�0�
1397
+ FIG. 7: Right panel: The behavior of the η/S ratio as a function of the temperature T for different values
1398
+ for α = 8/3, B = (4/5)T, ρ = 1/4, Λ = −1, γ = 1 (pink curve), γ = 2 (red dot dashed curve), γ = 2.5
1399
+ (green thick curve). Left panel: The behavior of η/s for B = 0.
1400
+ 0.0
1401
+ 0.5
1402
+ 1.0
1403
+ 1.5
1404
+ 2.0
1405
+ 0
1406
+ 1
1407
+ 2
1408
+ 3
1409
+ 4
1410
+ B
1411
+ Η
1412
+ S
1413
+ FIG. 8: The behavior of η/S with respect to the magnetic field B, for different values for α = 8/3, T = 4/5,
1414
+ ρ = 1/4, Λ = −1, γ = 1 (pink curve), γ = 2 (red dot dashed curve), γ = 2.5 (green thick curve).
1415
+ On the other hand, and as we can see from Fig. 8 at a temperature T fixed when we observe
1416
+ as the paramagnetic material, represented by the RS brane, we can obtain a relation between η/S
1417
+
1418
+ 21
1419
+ with respect to the magnetic field B, which is a decreasing function. Here, when B becomes large,
1420
+ we have that η/S → 0.
1421
+ We finalize this section showing the magnetic moment N at a low temperature T, corresponding
1422
+ to order parameter ρ in the absence of an external magnetic field, setting B = 0, and then compute
1423
+ the value of N, defined as
1424
+ N = λ2rh
1425
+ 2L
1426
+ � 1
1427
+ 0
1428
+ ρ(r)dr = −λ2rh
1429
+ 2L
1430
+
1431
+ − B
1432
+ m2 +
1433
+ 1
1434
+ (∆+ + 1)r∆+
1435
+ h
1436
+ +
1437
+ 1
1438
+ (∆− + 1)r∆−
1439
+ h
1440
+
1441
+ .
1442
+ (77)
1443
+ In Fig. 9, it can be found that as the temperature decreases, the magnetization increases and
1444
+ the system is in the perfect order with the maximum magnetization at zero temperature. Thus,
1445
+ increasing the Horndeski parameters lowers the magnetization value and the critical temperature.
1446
+ Indeed, we have that the effect of a larger value of the parameters γ and m2 makes the magnetization
1447
+ harder and the ferromagnetic phase transition happen, which is in good agreement with previous
1448
+ works [50, 51].
1449
+ 0.0
1450
+ 0.2
1451
+ 0.4
1452
+ 0.6
1453
+ 0.8
1454
+ 1.0
1455
+ 1.2
1456
+ 0
1457
+ 2
1458
+ 4
1459
+ 6
1460
+ 8
1461
+ 10
1462
+ T
1463
+ N
1464
+ Λ2
1465
+ FIG. 9: The behavior of magnetic moment N with different values for B = 0, α = 8/3 with γ = 1; m2 = 2
1466
+ (blue curve), γ = 4; m2 = 4 (red curve), γ = 8; m2 = 6 (green curve). We consider in the Eq. 77 the
1467
+ transformations Eq.∼(21).
1468
+ Finally, we present the susceptibility density χ of the materials as a response to the magnetic
1469
+ moment. Thus, this behavior is an essential property of ferromagnetic materials. In order to study
1470
+ χ of the ferromagnetic materials in the Horndeski gravity and to consider the transformations Eq.
1471
+ (21), we follow the definition
1472
+ χ
1473
+ λ2 = lim
1474
+ B→0
1475
+ ∂N
1476
+ ∂B =
1477
+
1478
+ 3
1479
+ 8πm2L2
1480
+ � 1
1481
+ T .
1482
+ (78)
1483
+
1484
+ 22
1485
+ 0.0
1486
+ 0.2
1487
+ 0.4
1488
+ 0.6
1489
+ 0.8
1490
+ 1.0
1491
+ 1.2
1492
+ 0.0
1493
+ 0.5
1494
+ 1.0
1495
+ 1.5
1496
+ 2.0
1497
+ T
1498
+ Λ2
1499
+ Χ
1500
+ 0.0
1501
+ 0.2
1502
+ 0.4
1503
+ 0.6
1504
+ 0.8
1505
+ 1.0
1506
+ 1.2
1507
+ 0
1508
+ 2
1509
+ 4
1510
+ 6
1511
+ 8
1512
+ 10
1513
+ T
1514
+ Χ
1515
+ Λ2
1516
+ FIG. 10: The behavior of 1/χ in the function of the temperature T with different values for α = 8/3 with
1517
+ γ = 1; m2 = 2 (blue curve), γ = 4; m2 = 4 (red curve), γ = 8; m2 = 6 (green curve). We consider in the Eq.
1518
+ (78) the transformations given in Eq.(21).
1519
+ In Fig.10, we have the behavior of 1/χ and χ as a function of the temperature T for different
1520
+ choices of m2 and γ. In our case, in the right panel, we have that increasing each one of these pa-
1521
+ rameters makes the susceptibility value decrease when the temperature increases. This fact agrees
1522
+ with our expectation of paramagnetic materials because when we remove the external magnetic
1523
+ field, the paramagnetic substance loses its magnetism. Its magnetic susceptibility is very small,
1524
+ but positive, and decreases with increasing temperature. In fact, this magnetic susceptibility is
1525
+ only part of the background black hole and the other part of the polarization field. For pure dionic
1526
+ Reissner-Nordstr¨om-AdS black hole, we have a diamagnetic material. In this sense, in the chemical
1527
+ reference, we have that a particle (atom, ion, or molecule) is paramagnetic or diamagnetic when
1528
+ the electrons in the particle are paired due to the external magnetic field [50, 51].
1529
+ VIII.
1530
+ CONCLUSIONS AND DISCUSSIONS
1531
+ In four dimensions, we analyzed an AdS/BCFT model of a condensed matter system at finite
1532
+ temperature and charge density living on a 2+1-dimensional space with a boundary, showing an
1533
+ extension of the previous work presented in [10], where in addition to the contributions of the
1534
+ theory together with the boundary terms, we include the components Aµ and Mµν, responsible to
1535
+ construct the ferromagnetic/paramagnetic model.
1536
+ Via the resolution of the field equations, and using the no-hair theorem, we extend to the
1537
+ four-dimensional configuration obtained in [10, 30]. From the above solution, we present the Q
1538
+ profile, found a numerical solution, and present it in Fig. 9, where the Horndeski parameter γ
1539
+
1540
+ 23
1541
+ takes an important role. Together with the above, we show that components of Mµν can be viewed
1542
+ as dual fields of the order parameter in the paraelectric/ferroelectric phase transition in dielectric
1543
+ materials. Through the NBC over nµM|Q, we found the ratio ρ/B, where for some particular
1544
+ cases is a constant proportional to a ratio of the coefficients appearing in the gravity action. These
1545
+ properties resemble a quantum Hall system, which suggests at the boundary Q in the (ρ, B) plane
1546
+ will be a localized condensate.
1547
+ Additionally, via the solution we performed a holographic renormalization, calculating the Eu-
1548
+ clidean on-shell action, which is related to the free energy Ω, and allowing us to obtain the entropy
1549
+ S and the heat capacities CV , CP , thanks to the contribution to the bulk as well as the boundary.
1550
+ With respect to the entropy S, we show that when the magnetic field is present we see it exhibits
1551
+ similar behavior as for example ferromagnetic materials with nearly zero coercivity and hysteresis.
1552
+ Nevertheless, when B = 0 the disorder entropy of the magnetic moments increases, being a char-
1553
+ acteristic of paramagnetism. Together with the above, with respect to CV and CP , we obtained
1554
+ for both cases stable and unstable phases, due to the spontaneous electric polarization, which was
1555
+ realized in our model from the application of the magnetic external field B, being influence via
1556
+ the Horndeski gravity, represented through γ. We also show that the specific heat CP behaves
1557
+ like a material of the type DyAl2, having a growth behavior similar to that expected from the
1558
+ experimental point of view, as presented by [53].
1559
+ Currently, we can observe that the microscopic differences between real experimental systems,
1560
+ in relation to theories with gravitational dual suggest that, in the near future, we will have measure-
1561
+ ments of these values for experimental quantities obtained holographically. So many measurements
1562
+ can realistically aspire to more than useful benchmarks. Furthermore, it is important to highlight
1563
+ in this regard the need to take the big limit N in holographic calculations [1]. We now have a
1564
+ clarity of the value of the ratio between shear viscosity and entropy density, η/S = 1/4π, which is
1565
+ universal in classical gravity to usual classical gravity [54]. Furthermore, in the Horndeski gravity,
1566
+ these relations are modified by the parameter γ. However, there are controlled corrections 1/N
1567
+ for this result, which can be both positive and negative and which for realistic values of N show
1568
+ significant changes in the numerical value of the ratio. As we show in our model, the violation of
1569
+ this universal bound in the Horndeski gravity with gauge fields changes the η/S ratio (see Fig.7
1570
+ and Fig.8), where this behavior is similar to the results of [55]. Furthermore, as γ increases, we
1571
+ can observe a translational symmetry breaking that survives the lower energy scales. According to
1572
+ Fig. 8, we have η/S → 0 at low temperatures.
1573
+ One of the strongest motivations for working with AdS/BCFT for condensed matter physics
1574
+
1575
+ 24
1576
+ rests on two pillars. The first is that, although theories with holographic duals may exhibit spe-
1577
+ cific exotic features, they also have features that are expected to be generic to tightly coupled
1578
+ theories, for example, the quantum critiques. In this sense, theories with gravitational duals are
1579
+ computationally tractable examples of generic tightly coupled field theories, and we can use them
1580
+ both to test our generic expectations and to guide us in refining those expectations. Thus, the
1581
+ examples discussed here are special cases of the fact that real-time finite temperature transport is
1582
+ much easier to calculate via AdS/BCFT than almost any other microscopic theory.
1583
+ Acknowledgments
1584
+ F.S. would like to thank the group of Instituto de F´ısica da UFRJ for fruitful discussions about
1585
+ the paramagnetic systems. In special to the E. Capossoli, Diego M. Rodrigues and Henrique Boschi-
1586
+ Filho. S.O. performed the work in the frame of the ”Mathematical modeling in interdisciplinary
1587
+ research of processes and systems based on intelligent supercomputer, grid and cloud technologies”
1588
+ program of the NAS of Ukraine. M.B. is supported by PROYECTO INTERNO UCM-IN-22204,
1589
+ L´ıNEA REGULAR.
1590
+ Appendix A: Shear viscosity and entropy density ratio with magnetic field
1591
+ We will present the calculation of the ratio η/S following the procedures [20, 38, 39, 54, 55].
1592
+ For this, we consider a perturbation along the xy direction in the metric Eq.17 [20, 38], in this
1593
+ sense, we have
1594
+ ds2 = L2
1595
+ r2
1596
+
1597
+ −f(r)dt2 + dx2 + dy2 + 2Ψ(r, t)dxdy + dr2
1598
+ f(r)
1599
+
1600
+ .
1601
+ (A1)
1602
+ From the overview point of the holographic dictionary, this procedure maps the fluctuation of the
1603
+ diagonal in the bulk metric in the off-diagonal components of the dual energy-momentum tensor.
1604
+ In this sense, we have a linear regime where fluctuations are associated with shear waves in the
1605
+ boundary fluid. Thus, substituting this metric (A1) in the Horndeski equation (Eµν = 0) for µ = x
1606
+ and ν = y, one obtains:
1607
+ P1Ψ
1608
+ ′′(r, t) + P2Ψ
1609
+ ′(r, t) + P3 ¨Ψ(r, t) = 0 ,
1610
+ (A2)
1611
+ where we defined
1612
+ P1 = 9γ2(α − γΛ)f2(r),
1613
+ P2 = −3γ(α − γΛ)f(r)(2αL2 − 6γr3/r3
1614
+ h),
1615
+
1616
+ 25
1617
+ P3 = −9γ2r(3α + γΛ).
1618
+ (A3)
1619
+ Using the ansatz:
1620
+ Ψ(r, t) = e−iωtΦ(r),
1621
+ (A4)
1622
+ Φ(r) = exp
1623
+
1624
+ −iωK ln
1625
+ �6γ2r3f(r)
1626
+ G
1627
+ ��
1628
+ ,
1629
+ G = L2V
1630
+ GN
1631
+
1632
+ 1 − ξ
1633
+ 4
1634
+
1635
+ ,
1636
+ (A5)
1637
+ we obtain
1638
+ K =
1639
+ 1
1640
+ 4πT
1641
+
1642
+ 3α + γΛ
1643
+ α − γΛ ,
1644
+ (A6)
1645
+ with T the Hawking temperature given previously in (57). At this point, we must evaluate the
1646
+ Lagrangian (1), using the metric function from (22), and expand it up to quadratic terms in Ψ
1647
+ and its derivatives [38]. In this way, we can study the boundary field theory using the AdS/CFT
1648
+ correspondence where the quadratic terms in the Lagrangian, after removing the second derivative
1649
+ contributions using the Gibbons-Hawking term, can be written as
1650
+ Hshear = P1Ψ2(r, t) + P2 ˙Ψ(r, t) + P3Ψ
1651
+ ′2(r, t) + P4Ψ(r, t)Ψ
1652
+ ′(r, t),
1653
+ (A7)
1654
+ where
1655
+ P1 = − 48L2
1656
+ 9r7f(r),
1657
+ P2 = 4γ L2
1658
+ r7
1659
+ ,
1660
+ P3 =
1661
+ 6γ2
1662
+ r3f(r),
1663
+ P4 = (α + γΛ) 2γ2L4
1664
+ α r7f(r).
1665
+ (A8)
1666
+ Here, (˙) denotes the derivative with respect t. Finally, viscosity η is determined from the term
1667
+ P3Ψ(r, t)Ψ
1668
+ ′(r, t) which reads
1669
+ η = 1
1670
+
1671
+ G
1672
+ 4r2
1673
+ h
1674
+
1675
+ 3α + γΛ
1676
+ α − γΛ ,
1677
+ (A9)
1678
+ where the entropy, from (66)-(68), can be written as
1679
+ S = GF
1680
+ 4r2
1681
+ h
1682
+ ,
1683
+ (A10)
1684
+ with
1685
+ F = 1 +
1686
+
1687
+ B2 cos2(θ′)b(θ′)
1688
+ 5m2ρ2
1689
+ �4πT
1690
+ 3
1691
+ �4
1692
+ + q(θ
1693
+ ′)
1694
+ 4
1695
+ �4πT
1696
+ 3
1697
+ �2�
1698
+
1699
+ sec(θ′)
1700
+
1701
+ 1 − ξ
1702
+ 4
1703
+
1704
+
1705
+ −B2 cos2(θ′)b(θ′)
1706
+ 2m2ρ2
1707
+ �4πT
1708
+ 3
1709
+ �3
1710
+ + q(θ
1711
+ ′)
1712
+ 2
1713
+ �4πT
1714
+ 3
1715
+ ��
1716
+ ,
1717
+
1718
+ 26
1719
+ and T given in (57). Thus, after algebraic manipulation and imposing V = 1, we have:
1720
+ η
1721
+ S =
1722
+ 1
1723
+ 4πF
1724
+
1725
+ 3α + γΛ
1726
+ α − γΛ ,
1727
+ (A11)
1728
+ where B = 0 and θ′ = π/2, we recover the result of [38].
1729
+ [1] J. M. Maldacena, The Large N limit of superconformal field theories and supergravity, Int. J. Theor.
1730
+ Phys. 38, 1113 (1999) [Adv. Theor. Math. Phys. 2, 231 (1998)] [hep-th/9711200].
1731
+ [2] E. Witten, Adv. Theor. Math. Phys. 2 (1998), 253-291 doi:10.4310/ATMP.1998.v2.n2.a2 [arXiv:hep-
1732
+ th/9802150 [hep-th]].
1733
+ [3] T. Takayanagi, “Holographic Dual of BCFT,” Phys. Rev. Lett. 107, 101602 (2011), [arXiv:1105.5165
1734
+ [hep-th]].
1735
+ [4] M. Fujita, T. Takayanagi and E. Tonni, “Aspects of AdS/BCFT,” JHEP 1111, 043 (2011),
1736
+ [arXiv:1108.5152 [hep-th]].
1737
+ [5] M. Nozaki, T. Takayanagi and T. Ugajin, Central Charges for BCFTs and Holography, JHEP 06, 066
1738
+ (2012) doi:10.1007/JHEP06(2012)066 [arXiv:1205.1573 [hep-th]].
1739
+ [6] D. Melnikov, E. Orazi and P. Sodano, On the AdS/BCFT Approach to Quantum Hall Systems, JHEP
1740
+ 05, 116 (2013) doi:10.1007/JHEP05(2013)116 [arXiv:1211.1416 [hep-th]].
1741
+ [7] F. F. dos Santos, AdS/BCFT correspondence and BTZ black hole within electric field, JHAP 4, no.1,
1742
+ 81-92 (2022) doi:10.22128/jhap.2022.504.1018 [arXiv:2206.09502 [hep-th]].
1743
+ [8] R. X. Miao, Holographic BCFT with Dirichlet Boundary Condition. JHEP 1902, 025 (2019),
1744
+ [arXiv:1806.10777 [hep-th]].
1745
+ [9] O. Sokoliuk, F. F. Santos and A. Baransky, AdS/BCFT correspondence and Lovelock theory in the
1746
+ presence of canonical scalar field, [arXiv:2206.04054 [hep-th]].
1747
+ [10] F. F. Santos, E. F. Capossoli and H. Boschi-Filho, “AdS/BCFT correspondence and BTZ black hole
1748
+ thermodynamics within Horndeski gravity,” Phys. Rev. D 104, no.6, 066014 (2021) [arXiv:2105.03802
1749
+ [hep-th]].
1750
+ [11] J. M. Mag´an, D. Melnikov and M. R. O. Silva, Black Holes in AdS/BCFT and Fluid/Gravity Corre-
1751
+ spondence, JHEP 11 (2014), 069 doi:10.1007/JHEP11(2014)069 [arXiv:1408.2580 [hep-th]].
1752
+ [12] S. Ryu and T. Takayanagi, Holographic derivation of entanglement entropy from AdS/CFT, Phys. Rev.
1753
+ Lett. 96, 181602 (2006) doi:10.1103/PhysRevLett.96.181602 [arXiv:hep-th/0603001 [hep-th]].
1754
+ [13] L. Randall and R. Sundrum, An Alternative to compactification, Phys. Rev. Lett. 83, 4690-4693 (1999)
1755
+ doi:10.1103/PhysRevLett.83.4690 [arXiv:hep-th/9906064 [hep-th]].
1756
+ [14] M. Fujita, M. Kaminski and A. Karch, SL(2,Z) Action on AdS/BCFT and Hall Conductivities, JHEP
1757
+ 07, 150 (2012) doi:10.1007/JHEP07(2012)150 [arXiv:1204.0012 [hep-th]].
1758
+
1759
+ 27
1760
+ [15] S. S. Gubser, Breaking an Abelian gauge symmetry near a black hole horizon, Phys. Rev. D 78, 065034
1761
+ (2008) doi:10.1103/PhysRevD.78.065034 [arXiv:0801.2977 [hep-th]].
1762
+ [16] S. A. Hartnoll, C. P. Herzog and G. T. Horowitz, Building a Holographic Superconductor, Phys. Rev.
1763
+ Lett. 101, 031601 (2008) doi:10.1103/PhysRevLett.101.031601 [arXiv:0803.3295 [hep-th]].
1764
+ [17] H. Liu, J. McGreevy and D. Vegh, Non-Fermi liquids from holography, Phys. Rev. D 83, 065029 (2011)
1765
+ doi:10.1103/PhysRevD.83.065029 [arXiv:0903.2477 [hep-th]].
1766
+ [18] G. W. Horndeski, Int. J. Theor. Phys. 10 (1974), 363-384 doi:10.1007/BF01807638
1767
+ [19] F. A. Brito and F. F. Santos, Braneworlds in Horndeski gravity, Eur. Phys. J. Plus 137, no.9, 1051
1768
+ (2022) doi:10.1140/epjp/s13360-022-03270-w [arXiv:1810.08196 [hep-th]].
1769
+ [20] F. Brito and F. Santos, “Black branes in asymptotically Lifshitz spacetime and viscosity/entropy ratios
1770
+ in Horndeski gravity,” EPL 129, no.5, 50003 (2020), [arXiv:1901.06770 [hep-th]].
1771
+ [21] F. F. Santos, R. M. P. Neves and F. A. Brito, Modeling dark sector in Horndeski gravity at first-order
1772
+ formalism, Adv. High Energy Phys. 2019, 3486805 (2019) doi:10.1155/2019/3486805 [arXiv:1906.11821
1773
+ [hep-th]].
1774
+ [22] F. F. Santos, Rotating black hole with a probe string in Horndeski Gravity, Eur. Phys. J. Plus 135,
1775
+ no.10, 810 (2020) [arXiv:2005.10983 [hep-th]].
1776
+ [23] F. F. Santos and F. A. Brito, Domain walls in Horndeski gravity, [arXiv:2105.00343 [hep-th]].
1777
+ [24] F. F. Dos Santos,
1778
+ Entanglement entropy in Horndeski gravity,
1779
+ JHAP 3,
1780
+ no.1,
1781
+ 1-14 (2022)
1782
+ doi:10.22128/jhap.2022.488.1015 [arXiv:2201.02500 [hep-th]].
1783
+ [25] F. F. Santos, O. Sokoliuk and A. Baransky, Holographic complexity of braneworld in Horndeski gravity,
1784
+ [arXiv:2210.11596 [hep-th]].
1785
+ [26] F. F. Santos and F. A. Brito, Thick branes in Horndeski gravity, [arXiv:2210.15003 [hep-th]].
1786
+ [27] M. Rinaldi, Black holes with non-minimal derivative coupling, Phys. Rev. D 86 (2012), 084048
1787
+ doi:10.1103/PhysRevD.86.084048 [arXiv:1208.0103 [gr-qc]].
1788
+ [28] E. Babichev and C. Charmousis, Dressing a black hole with a time-dependent Galileon, JHEP 08 (2014),
1789
+ 106 doi:10.1007/JHEP08(2014)106 [arXiv:1312.3204 [gr-qc]].
1790
+ [29] A. Anabalon, A. Cisterna and J. Oliva, Asymptotically locally AdS and flat black holes in Horndeski
1791
+ theory, Phys. Rev. D 89 (2014), 084050 doi:10.1103/PhysRevD.89.084050 [arXiv:1312.3597 [gr-qc]].
1792
+ [30] M. Bravo-Gaete and M. Hassaine, Thermodynamics of a BTZ black hole solution with an Horndeski
1793
+ source, Phys. Rev. D 90 (2014) no.2, 024008 doi:10.1103/PhysRevD.90.024008 [arXiv:1405.4935 [hep-
1794
+ th]].
1795
+ [31] M. Bravo-Gaete and M. Hassaine, Lifshitz black holes with a time-dependent scalar field in a Horndeski
1796
+ theory, Phys. Rev. D 89 (2014), 104028 doi:10.1103/PhysRevD.89.104028 [arXiv:1312.7736 [hep-th]].
1797
+ [32] L. Hui and A. Nicolis, No-Hair Theorem for the Galileon, Phys. Rev. Lett. 110 (2013), 241104
1798
+ doi:10.1103/PhysRevLett.110.241104 [arXiv:1202.1296 [hep-th]].
1799
+ [33] Y. Brihaye, A. Cisterna and C. Erices, Boson stars in biscalar extensions of Horndeski gravity, Phys.
1800
+ Rev. D 93 (2016) no.12, 124057 doi:10.1103/PhysRevD.93.124057 [arXiv:1604.02121 [hep-th]].
1801
+
1802
+ 28
1803
+ [34] A. Cisterna, T. Delsate and M. Rinaldi, Neutron stars in general second order scalar-tensor
1804
+ theory:
1805
+ The case of nonminimal derivative coupling,
1806
+ Phys. Rev. D 92 (2015) no.4,
1807
+ 044050
1808
+ doi:10.1103/PhysRevD.92.044050 [arXiv:1504.05189 [gr-qc]].
1809
+ [35] A. Cisterna, T. Delsate, L. Ducobu and M. Rinaldi, Slowly rotating neutron stars in the non-
1810
+ minimal derivative coupling sector of Horndeski gravity, Phys. Rev. D 93 (2016) no.8, 084046
1811
+ doi:10.1103/PhysRevD.93.084046 [arXiv:1602.06939 [gr-qc]].
1812
+ [36] E. Arratia, C. Corral, J. Figueroa and L. Sanhueza, Hairy Taub-NUT/bolt-AdS solutions in Horndeski
1813
+ theory, Phys. Rev. D 103 (2021) no.6, 064068 doi:10.1103/PhysRevD.103.064068 [arXiv:2010.02460
1814
+ [hep-th]].
1815
+ [37] X. H. Feng and H. S. Liu, Holographic Complexity Growth Rate in Horndeski Theory, Eur. Phys. J. C
1816
+ 79 (2019) no.1, 40 doi:10.1140/epjc/s10052-019-6547-4 [arXiv:1811.03303 [hep-th]].
1817
+ [38] X. H. Feng, H. S. Liu, H. L¨u and C. N. Pope, Black Hole Entropy and Viscosity Bound in Horndeski
1818
+ Gravity, JHEP 11, 176 (2015) doi:10.1007/JHEP11(2015)176 [arXiv:1509.07142 [hep-th]].
1819
+ [39] M. Bravo-Gaete, F. F. Santos and H. Boschi-Filho, Shear viscosity from black holes in gen-
1820
+ eralized scalar-tensor theories in arbitrary dimensions, Phys. Rev. D 106 (2022) no.6, 066010
1821
+ doi:10.1103/PhysRevD.106.066010 [arXiv:2201.07961 [hep-th]].
1822
+ [40] M. Bravo-Gaete and M. M. Stetsko, Planar black holes configurations and shear viscosity in arbitrary
1823
+ dimensions with shift and reflection symmetric scalar-tensor theories, Phys. Rev. D 105 (2022) no.2,
1824
+ 024038 doi:10.1103/PhysRevD.105.024038 [arXiv:2111.10925 [hep-th]].
1825
+ [41] M. Bravo-Gaete and F. F. Santos, Complexity of four-dimensional hairy anti-de-Sitter black holes with
1826
+ a rotating string and shear viscosity in generalized scalar tensor theories, Eur. Phys. J. C 82 (2022)
1827
+ no.2, 101 doi:10.1140/epjc/s10052-022-10064-y [arXiv:2010.10942 [hep-th]].
1828
+ [42] R. G. Cai and R. Q. Yang, Paramagnetism-Ferromagnetism Phase Transition in a Dyonic Black Hole
1829
+ Phys. Rev. D 90, no.8, 081901 (2014) [arXiv:1404.2856 [hep-th]].
1830
+ [43] R. G. Cai and R. Q. Yang, Antisymmetric tensor field and spontaneous magnetization in holographic
1831
+ duality, Phys. Rev. D 92 (2015) no.4, 046001 [arXiv:1504.00855 [hep-th]].
1832
+ [44] B. B. Ghotbabadi, A. Sheykhi, G. H. Bordbar and A. Montakhab, Holographic paramagnetic-
1833
+ ferromagnetic phase transition of Power-Maxwell-Gauss-Bonnet black holes, [arXiv:2102.08053 [hep-
1834
+ th]].
1835
+ [45] S. A. Hartnoll, C. P. Herzog and G. T. Horowitz, Holographic Superconductors, JHEP 12, 015 (2008)
1836
+ doi:10.1088/1126-6708/2008/12/015 [arXiv:0810.1563 [hep-th]].
1837
+ [46] S. A. Hartnoll, Lectures on holographic methods for condensed matter physics, Class. Quant. Grav. 26,
1838
+ 224002 (2009) doi:10.1088/0264-9381/26/22/224002 [arXiv:0903.3246 [hep-th]].
1839
+ [47] R. B. Laughlin, Quantized Hall conductivity in two-dimensions, Phys. Rev. B 23, 5632-5733 (1981)
1840
+ doi:10.1103/PhysRevB.23.5632
1841
+ [48] G. W. Moore and N. Read, Nonabelions in the fractional quantum Hall effect, Nucl. Phys. B 360,
1842
+ 362-396 (1991) doi:10.1016/0550-3213(91)90407-O
1843
+
1844
+ 29
1845
+ [49] Avron, Joseph E. and Seiler, Ruedi, Nonabelions in the fractional quantum Hall effect, Phys. Rev. Lett.
1846
+ 54, 259-262 (1985) doi:10.1103/PhysRevLett.54.259
1847
+ [50] C. Y. Zhang, Y. B. Wu, Y. N. Zhang, H. Y. Wang and M. M. Wu, Holographic paramagnetism-
1848
+ ferromagnetism phase transition with the nonlinear electrodynamics, Nucl. Phys. B 914, 446-460 (2017)
1849
+ doi:10.1016/j.nuclphysb.2016.11.017 [arXiv:1609.09318 [hep-th]].
1850
+ [51] Y. B. Wu, C. Y. Zhang, J. W. Lu, B. Fan, S. Shu and Y. C. Liu, Holographic paramagnetism-
1851
+ ferromagnetism phase transition in the Born-Infeld electrodynamics, Phys. Lett. B 760, 469-474 (2016)
1852
+ doi:10.1016/j.physletb.2016.07.031
1853
+ [52] K. H. M¨uller Magnetic Viscosity, Reference Module in Materials Science and Materials Engineering,
1854
+ Elsevier, 2016, doi = https://doi.org/10.1016/B978-0-12-803581-8.02807-1.
1855
+ [53] Lima, A. L. and Tsokol, A. O. and Gschneidner, K. A. and Pecharsky, V. K. and Lograsso, T. A. and
1856
+ Schlagel, D. L., Magnetic properties of single-crystal DyAl2 Phys. Rev. B. 72, 024403 (2005), doi =
1857
+ 10.1103/PhysRevB.72.024403.
1858
+ [54] P. Kovtun, D. T. Son and A. O. Starinets, Viscosity in strongly interacting quantum field theories from
1859
+ black hole physics, Phys. Rev. Lett. 94, 111601 (2005) doi:10.1103/PhysRevLett.94.111601 [arXiv:hep-
1860
+ th/0405231 [hep-th]].
1861
+ [55] S. A. Hartnoll, D. M. Ramirez and J. E. Santos, Entropy production, viscosity bounds and bumpy black
1862
+ holes, JHEP 03, 170 (2016) doi:10.1007/JHEP03(2016)170 [arXiv:1601.02757 [hep-th]].
1863
+
FtE1T4oBgHgl3EQfXARM/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
FtE3T4oBgHgl3EQfVwq1/content/2301.04463v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:df735fa91fc3a518953f400dc2966df7d568acd949678c98a34293dfb0b1508b
3
+ size 1315827
FtE3T4oBgHgl3EQfVwq1/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e2cf4236d27fdefe1c253edef6921f3ed291b4851e116c82a0eca1c4083850cf
3
+ size 10944557
FtE3T4oBgHgl3EQfVwq1/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:90d2402203242e470cba0cd7ee52213ad4192cdc2b3b1e5fa5f5085856239972
3
+ size 343092
GdE2T4oBgHgl3EQfTQfv/content/tmp_files/2301.03802v1.pdf.txt ADDED
@@ -0,0 +1,1384 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Predicting Drivers’ Route Trajectories in Last-Mile Delivery Using A Pair-wise
2
+ Attention-based Pointer Neural Network
3
+ Baichuan Moa, Qing Yi Wanga,∗, Xiaotong Guoa, Matthias Winkenbachb, Jinhua Zhaoc
4
+ aDepartment of Civil and Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139
5
+ bCenter for Transportation and Logistics, Massachusetts Institute of Technology, Cambridge, MA 20139
6
+ cDepartment of Urban Studies and Planning, Massachusetts Institute of Technology, Cambridge, MA 20139
7
+ Abstract
8
+ In last-mile delivery, drivers frequently deviate from planned delivery routes because of their tacit knowledge
9
+ of the road and curbside infrastructure, customer availability, and other characteristics of the respective service
10
+ areas. Hence, the actual stop sequences chosen by an experienced human driver may be potentially preferable
11
+ to the theoretical shortest-distance routing under real-life operational conditions. Thus, being able to predict
12
+ the actual stop sequence that a human driver would follow can help to improve route planning in last-mile
13
+ delivery. This paper proposes a pair-wise attention-based pointer neural network for this prediction task using
14
+ drivers’ historical delivery trajectory data. In addition to the commonly used encoder-decoder architecture
15
+ for sequence-to-sequence prediction, we propose a new attention mechanism based on an alternative specific
16
+ neural network to capture the local pair-wise information for each pair of stops. To further capture the global
17
+ efficiency of the route, we propose a new iterative sequence generation algorithm that is used after model
18
+ training to identify the first stop of a route that yields the lowest operational cost. Results from an extensive
19
+ case study on real operational data from Amazon’s last-mile delivery operations in the US show that our
20
+ proposed method can significantly outperform traditional optimization-based approaches and other machine
21
+ learning methods (such as the Long Short-Term Memory encoder-decoder and the original pointer network)
22
+ in finding stop sequences that are closer to high-quality routes executed by experienced drivers in the field.
23
+ Compared to benchmark models, the proposed model can increase the average prediction accuracy of the
24
+ first four stops from around 0.2 to 0.312, and reduce the disparity between the predicted route and the actual
25
+ route by around 15%.
26
+ Keywords: Route planning, Trajectory prediction, Sequence-to-sequence model, Last-mile delivery,
27
+ Pointer network, Attention
28
+ 1. Introduction
29
+ The optimal planning and efficient execution of last-mile delivery routes is becoming increasingly
30
+ important for the business operations of many logistics service providers around the globe for a variety of
31
+ reasons. E-commerce volumes are growing rapidly and make up a constantly growing share of overall retail
32
+ sales. For instance, in the US, the share of e-commerce sales in total retail sales has grown from around 4% in
33
+ 2010 to around 13% in 2021. Even by the end of 2019, i.e., before the outbreak of the COVID-19 pandemic,
34
+ ∗Corresponding author
35
+ Preprint submitted to Elsevier
36
+ January 11, 2023
37
+ arXiv:2301.03802v1 [cs.LG] 10 Jan 2023
38
+
39
+ it had reached 11% (US Census Bureau, 2021). Undoubtedly, the pandemic further accelerated the growth
40
+ of e-commerce (postnord, 2021; McKinsey & Company, 2021). In the medium to long run, its growth will
41
+ continue to be fueled by an ongoing trend towards further urbanization, which is particularly pronounced in
42
+ developing and emerging economies (United Nations Department of Economic and Social Affairs, 2019).
43
+ The share of the global population living in urban areas is currently projected to rise from around 55% in
44
+ 2018 to around 68% by 2050. The associated increase in population density in most urban areas will likely
45
+ lead to growing operational uncertainties for logistics service providers, as increasing congestion levels, less
46
+ predictable travel times, and scarce curb space make efficient and reliable transport of goods into and out of
47
+ urban markets increasingly challenging (Rose et al., 2016).
48
+ As a result of the continued boom of e-commerce and constantly growing cities, global parcel delivery
49
+ volumes have been increasing rapidly in recent years and are expected to continue to do so. Across the
50
+ 13 largest global markets, including the US, Brazil, and China, the volume of parcels delivered more than
51
+ tripled from 43 billion in 2014 to 131 billion in 2020 (Pitney Bowes, 2020). At the same time, customer
52
+ expectations towards last-mile logistics services are rising. For instance, there is a growing demand for
53
+ shorter delivery lead times, including instant delivery services and same-day delivery, as well as customer-
54
+ defined delivery preferences when it comes to the time and place of delivery (Lim and Winkenbach, 2019;
55
+ Cortes and Suzuki, 2021; Snoeck and Winkenbach, 2021). The rapid growth and increasing operational
56
+ complexity of urban parcel delivery operations also amplifies their negative externalities, including their
57
+ contribution to greenhouse gas and other pollutant emissions, public health safety risks, as well as overall
58
+ urban congestion and a corresponding decline in overall mobility and accessibility of cities (Jaller et al.,
59
+ 2013; World Economic Forum, 2020).
60
+ When applied to realistically sized instances of a last-mile delivery problem, solving the underlying
61
+ traveling salesman problem (TSP) or vehicle routing problem (VRP) to (near) optimality becomes chal-
62
+ lenging, as both problem classes are known to be NP-hard. Traditional TSP and VRP formulations aim to
63
+ minimize the total distance or duration of the route(s) required to serve a given set of delivery stops. The
64
+ operations research literature has covered the TSP, VRP, and their many variants extensively, and in recent
65
+ years important advances have been made with regards to solution quality and computational cost. However,
66
+ in practice, many drivers, with their own tacit knowledge of delivery routes and service areas, divert from
67
+ seemingly optimal routes for reasons that are difficult to encode in an optimization model directly. For exam-
68
+ ple, experienced drivers may have a better understanding of which roads are hard to navigate, at which times
69
+ traffic is likely to be bad, when and where they can easily find parking, and which stops can be conveniently
70
+ served together. Therefore, compared to the theoretically optimal (i.e., distance or time minimizing) route,
71
+ the deviated actual route sequence chosen by an experienced human driver is potentially preferable under
72
+ real-life operational conditions.
73
+ An important challenge in today’s last-mile delivery route planning is therefore to leverage historical
74
+ route execution data to propose planned route sequences that are close to the actual trajectories that would
75
+ be executed by drivers, given the delivery requests and their characteristics. Note that, while distance and
76
+ time-based route efficiency is still an important factor for planning route sequences, it is not the sole objective,
77
+ as tacit driver knowledge is also incorporated in the proposed route sequences. Unlike a typical VRP in
78
+ which the number of vehicles and their respective route sequences need to be determined simultaneously, in
79
+ this study, we focus on solving a problem that is similar to a TSP at the individual vehicle level. That is,
80
+ we aim to solve a stop sequence to serve a given set of delivery requests, and expect that the proposed stop
81
+ sequence is as close to the actual trajectories that would be executed by drivers as possible.
82
+ 2
83
+
84
+ To this end, we propose a pair-wise attention-based pointer neural network to predict the actual route
85
+ sequence taken by delivery drivers using drivers’ historical delivery trajectory data. The proposed model
86
+ follows a typical encoder-decoder architecture for the sequence-to-sequence prediction. However, unlike
87
+ previous studies, we propose a new attention mechanism based on an alternative specific neural network
88
+ (ASNN) to capture the local pair-wise information for each stop pair. To further capture the global efficiency
89
+ of the route (i.e., its operational cost in terms of total distance or duration), after model training, we propose a
90
+ new sequence generation algorithm that iterates over different first stops and selects the route with the lowest
91
+ operational cost.
92
+ The main contribution of this paper is three-fold: First, we propose a new ASNN-based attention
93
+ mechanism to capture the local information between pairs of stops (e.g., travel time, geographical relation),
94
+ which can be well adapted to the original pointer network framework for sequence prediction. Second, we
95
+ propose a new sequence generation algorithm that iterates over different first stops in the predicted route
96
+ sequences and selects the lowest operational cost route. The intuition is that the stop-to-stop relationship
97
+ (referred to as the local view) is easier to learn from data than the stop sequence of the route as a whole
98
+ (referred to as the global view). Lastly, we apply our proposed method to a large set of routes executed by
99
+ Amazon delivery drivers in the US. The results show that our proposed model can outperform traditional
100
+ optimization-based approaches and other machine learning methods in finding stop sequences that are closer
101
+ to high-quality routes executed by experienced drivers in the field.
102
+ The remainder of this paper is structured as follows. In Section 2 we define the problem setting under
103
+ investigation in a more formal way. Section 3 then reviews previous studies in the literature related to this
104
+ paper. Section 4 presents our methodology and elaborates on the detailed architecture of the proposed
105
+ pair-wise attention-based pointer neural network. Section 5 presents the experimental setup and numerical
106
+ results of our case study, applying our proposed method to real-world data made available by the Amazon
107
+ Last-Mile Routing Research Challenge (Merchán et al., 2022; Winkenbach et al., 2021). Section 6 concludes
108
+ this paper and discusses future research directions.
109
+ 2. Problem Setting
110
+ In the last-mile delivery routing problem considered here, a set of stops S = {s1, ..., sn} to be served
111
+ by a given delivery vehicle is given to the route planner. The planner’s objective is to find the optimal
112
+ stop sequence that has the minimal operational cost. In this case, we consider total cost as total travel
113
+ time.
114
+ The planner is given the expected operational cost (i.e., travel times) between all pairs of stops
115
+ (si, sj). The theoretically optimal stop sequence, denoted by (sT
116
+ (1), ..., sT
117
+ (n)), can be found by solving a TSP
118
+ formulation. This stop sequence is referred to as the planned stop sequence. However, as discussed in
119
+ Section 1, minimizing the theoretical operational cost (i.e., total travel time) of the route may not capture
120
+ drivers’ tacit knowledge about the road network, infrastructure, and recipients. Therefore, the actual driver
121
+ executed stop sequence (s(1), ..., s(n)) can be different from the planned route sequence. Note that here,
122
+ s(i) ∈ S denotes the i-th stop that is actually visited by the driver.
123
+ The objective of the model presented in this study is to predict the actual driver executed sequence
124
+ (s(1), ..., s(n)) given a set of stops S and the corresponding delivery requests and characteristics XS (such
125
+ as the number of packages, estimated service time for each package, geographical information for each stop,
126
+ travel time between each stop pairs, etc.). All drivers are assumed to start their routes from a known depot
127
+ DS and return back to DS. Therefore, the complete trajectory should be a tour (DS, s(1), ..., s(n), DS). For
128
+ the convenience of model description, we ignore the depot station in the sequence.
129
+ 3
130
+
131
+ Figure 1 provides a simple example for illustration. In this example, we are given four stops S =
132
+ {s1, s2, s3, s4} and a depot DS. The planned stop sequence for the driver is (s4, s1, s2, s3), while the actual
133
+ stop sequence executed by the driver is (s4, s2, s1, s3). The proposed model aims to predict the actual
134
+ sequence (s4, s2, s1, s3) given the depot location DS, the set of stops to be visited S, and characteristics of
135
+ the stops XS. This problem setup is inspired by the Amazon Last-Mile Routing Research Challenge (cf.,
136
+ Winkenbach et al., 2021). Note that this study only focuses on the stop sequence prediction. The routing
137
+ between stops is not considered. It is assumed that the drivers always take the optimal route between stops,
138
+ which is reflected by the travel time matrix between stops in our problem setup.
139
+ Figure 1: Illustrative example of the problem setting.
140
+ 3. Literature Review
141
+ The problem setting defined in Section 2 involves both solving a cost-minimizing routing problem (i.e.,
142
+ the TSP) and capturing tacit driver knowledge to learn systematic deviation of drivers from the planned and
143
+ theoretically optimal stop sequences. Therefore, we will first review the extant literature on the TSPs and
144
+ its most relevant variants. We will then go through various machine learning approaches that have been
145
+ proposed by the extant literature to generate sequences, with a section on methods specifically for solving the
146
+ TSP. Note that although these machine learning approaches are used to solve the TSP instead of the actual
147
+ routes taken by drivers, their architectures may be helpful to learn the actual route as well.
148
+ 3.1. Travelling salesman problems
149
+ First, given the travel times between stops, a solution to the TSP, which finds the route with the minimum
150
+ cost or distance (i.e., the planned route), can be a close approximation of the actual route. Since the drivers
151
+ are paid for the number of packages delivered, all drivers’ goal is to deliver the packages in the minimum
152
+ amount of time. Most of the drivers do follow large parts of the planned routes.
153
+ The TSP is a well-known NP-hard problem that has been studied extensively over the last century, with a
154
+ lot of books and review papers published on its history, formulations, solution approaches, and applications
155
+ (Applegate et al., 2006; Matai et al., 2010; Davendra and Bialic-Davendra, 2020). An overview of the
156
+ relevant TSP variants and solution approaches are presented below.
157
+ The basic setup of TSP has one traveler and requires the traveler to return to the starting point after
158
+ visiting each node exactly once, and that the traveling cost matrix (represented by distance and/or time) is
159
+ 4
160
+
161
+ Depot Ds
162
+ Input
163
+ Output
164
+ S,Xs,Ds
165
+ Model
166
+ Actual route
167
+ S3
168
+ S1
169
+ Actual route
170
+ S2
171
+ - Planned routesymmetric (cost between i and j is the same with that between j and i). In most real-world applications,
172
+ the basic setup needs to be modified. For example, the cost matrix, if represented by travel times, is likely
173
+ asymmetric. This variant of TSP is thus named asymmetric TSP (ATSP) (Jonker and Volgenant, 1983).
174
+ In some applications, the vehicle does not need to return to the original depot (Traub et al., 2021), or
175
+ it can charge/refuel and potentially load additional delivery items at intermediate stops (Küçükoğlu et al.,
176
+ 2019). In many last-mile delivery applications, some packages are time-sensitive, and therefore time window
177
+ constraints to their delivery need to be considered in a so-called TSP with time windows (TSPTW) (da Silva
178
+ and Urrutia, 2010; Mladenović et al., 2012). In large systems, there might be more than one salesman serving
179
+ a set of stops, resulting in multiple traveling salesmen problems (MTSPs) (Cheikhrouhou and Khoufi, 2021).
180
+ Different variants of TSP further impose different constraints on the solution. While some problems
181
+ can be reduced to the basic setup in the formulation stage, others require more versatile solution algorithms.
182
+ In general, the solution approaches to the TSP can be divided into exact approaches and approximate
183
+ approaches. Exact approaches include branch-and-cut (Yuan et al., 2020) and branch-and-bound (Salman
184
+ et al., 2020). Since the TSP is a well-known NP-hard problem, exact approaches can only be applied on
185
+ problems of smaller scale, or aid in heuristics to cut the solution space. Among approximate approaches,
186
+ there are heuristics designed for the TSP specifically, as well as meta-heuristics that are generic and treat the
187
+ problem like a blackbox. The most commonly used heuristics and meta-heuristics include nearest neighbor
188
+ searches, local searches, simulated annealing, and genetic algorithms. A more comprehensive review of
189
+ existing solution approaches can be found in Halim and Ismail (2017); Purkayastha et al. (2020). Despite
190
+ the TSP being NP-hard, modern mixed-integer optimization solvers (e.g., Gurobi, CPLEX, or GLPK) can
191
+ solve it efficiently for real-world instances by combining exact approaches with heuristics.
192
+ 3.2. Sequence-to-sequence prediction using deep learning
193
+ The TSP and its variants are a viable option for sequence generation only when the objective is clearly de-
194
+ fined. They fall short when the sequence generation problem does not have a well-defined cost-minimization
195
+ objective. In a lot of applications, the rule of sequence generation cannot be simply defined and optimized.
196
+ A standard example for a sequence learning problem is machine translation, where a sequence of words
197
+ in one language needs to be translated to another language. Another type of sequence learning is time series
198
+ modeling, where a sequence of historical observations is given to predict future states of the system. In both
199
+ cases, the primary modeling task is to learn the sequence generation rules. In recent years, deep learning
200
+ has successfully achieved great performance in various settings of sequence generation. These models are
201
+ often referred to as sequence-to-sequence (seq2seq) models.
202
+ seq2seq models often consist of an encoder and a decoder, where the encoder encodes the input sequence
203
+ into a fixed-length vector representation, and the decoder generates a sequence based on the generated vector
204
+ representation. Most encoder-decoder architectures adopt recurrent neural network (RNN) layers and its
205
+ variants such as Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) and gated recurrent
206
+ layers (GRU) (Cho et al., 2014) to learn long-range dependencies. Early works using LSTM alone were able
207
+ to generate plausible texts (Graves, 2013) and translate between English and French (Sutskever et al., 2014)
208
+ with long-range dependencies. Chung et al. (2014) demonstrate the superiority of GRU compared to LSTMs
209
+ in music and speech signal modeling.
210
+ Attention-based mechanisms, first introduced by Bahdanau et al. (2015), have been shown to be a great
211
+ addition since it allows the decoder to selectively attend to parts of the input sequence and relieves the encoder
212
+ of the task of encoding all the information into a fixed-length vector representation. Most sequence generation
213
+ 5
214
+
215
+ problems benefit from keeping track of long-range dependencies and global context while decoding. To
216
+ address that, multi-level attention was proposed to capture the local and global dependency, and has shown
217
+ to be effective in speech recognition (Chorowski et al., 2015), text generation (Liu et al., 2018), and machine
218
+ translation tasks (Luong et al., 2015).
219
+ The encoder-decoder architecture combined with attention is very versatile, and it can be combined
220
+ with other deep learning architectures to perform sequence learning in addition to language tasks. The
221
+ LSTM and attention architecture is applied to semantic trajectory prediction (Karatzoglou et al., 2018),
222
+ text summarization (Liang et al., 2020), demand modelling (Ren et al., 2020), and wind power forecasting
223
+ (Zhang et al., 2020).
224
+ When the goal is set to recover the original sequence, unsupervised learning of
225
+ molecule embedding can be obtained for downstream classification tasks (Xu et al., 2017). When the spatial
226
+ dimension is added, a convolutional neural network (CNN) layer can be added, and the dimension of the
227
+ sequence generated can be expanded. For example, Wang et al. (2020a) predict a city’s crowd flow patterns,
228
+ and Wu et al. (2020) generate 3D shapes via sequentially assembling different parts of the shapes.
229
+ While RNN-based architectures are still a widely adopted choice for seq2seq modeling, attention can also
230
+ be used as a standalone mechanism for seq2seq translations independent of RNNs. The idea was proposed
231
+ by Vaswani et al. (2017) in an architecture named transformer. Without recurrence, the network allows for
232
+ significantly more parallelization, and is shown to achieve superior performance in experiments, and powered
233
+ the popularity of transformer-based architectures in various sequence generation tasks (Huang et al., 2018;
234
+ Lu et al., 2021). A separate line of work by Zhang et al. (2019) also demonstrated that a hierarchical CNN
235
+ model with attention outperforms the traditional RNN-based models.
236
+ 3.3. Using deep learning to generate TSP solutions
237
+ The above seq2seq translation mechanisms work well when the input data is naturally organized as
238
+ a sequence, and the output sequence corresponds to the input sequence, such as in music and language.
239
+ However, in our paper, the input is an unordered sequence, and the output has the same but re-ordered
240
+ elements of the same input sequence.
241
+ In this case, the concept of attention is helpful and has been
242
+ successfully used to produce solutions to the TSP. The pointer network, proposed by Vinyals et al. (2015)
243
+ and further developed in Vinyals et al. (2016), uses attention to select a member of the input sequence at
244
+ each decoder step. While it is not required that the input sequence is ordered, an informative ordering could
245
+ improve the performance (Vinyals et al., 2016).
246
+ While the original pointer network was solved as a classification problem and cross-entropy loss was
247
+ used, it is not necessarily the most efficient choice. The cross-entropy loss only distinguishes between
248
+ a correct prediction and an incorrect prediction. But in instances like routing, the distances between the
249
+ predicted position and the correct position, as well as the ordering of subsequences, could incur different
250
+ costs in practice. Further developments in solving TSP with machine learning methods involve reinforcement
251
+ learning (RL), which enables the optimization of custom evaluation metrics (Bello et al., 2019; Kool et al.,
252
+ 2019; Ma et al., 2019; Liu et al., 2020). Joshi et al. (2019) compared the performance of RL and supervised
253
+ learning (SL) on TSP solutions and found that SL and RL models achieve similar performance when the
254
+ graphs are of similar sizes in training and testing, whereas RL models have better generalizability over variable
255
+ graph sizes. However, RL models require significantly more data points and computational resources, which
256
+ is not always feasible.
257
+ Although this seq2seq and attention framework has only been used to reproduce TSP solutions, it provides
258
+ an opportunity to learn and incorporate additional information beyond the given travel times and potentially
259
+ 6
260
+
261
+ learn individual differences when more information is given to the neural network. In this paper, we combine
262
+ the ideas of seq2seq modeling and attention to predict the actual route executed by a driver.
263
+ 4. Methodology
264
+ This section details the methodology proposed to address the problem. First, the high-level seq2seq
265
+ modelling framework is introduced, followed by the explanation of the novel pair-wise attention and sequence
266
+ generation and selection mechanism used within the modelling framework.
267
+ 4.1. Sequence-to-sequence modeling framework
268
+ Let the input sequence be an arbitrarily-ordered sequence (s1, ..., sn). Denote the output sequence as
269
+ (ˆs(1), ..., ˆs(n)). Let ci indicate the “position index” of stop ˆs(i) with respect to the input sequence (where
270
+ ci ∈ {1, ..., n}). For example, for input sequence (B, A, C) and output sequence (A, B, C), we have c1 = 2,
271
+ c2 = 1, c3 = 3, which means the first output stop A is in the second position of the input sequence (B, A, C)
272
+ and so on.
273
+ The seq2seq model computes the conditional probability P(c1, ..., cn | S; θ) using a parametric neural
274
+ network (e.g., recurrent neural network) with parameter θ, i.e.,
275
+ P(c1, ..., cn | S, XS; θ) = P(c1 | S, XS; θ) ·
276
+ n
277
+
278
+ i=2
279
+ P(ci | c1, ..., ci−1, S, XS; θ)
280
+ (1)
281
+ The parameters of the model are learnt by empirical risk minimization (maximizing the conditional
282
+ probabilities on the training set), i.e.,
283
+ θ∗ = arg max
284
+ θ
285
+
286
+ S
287
+ P(c1, ..., cn | S, XS; θ)
288
+ (2)
289
+ where the summation of S is over all training routes. In the following section, we will elaborate how
290
+ P(ci | c1, ..., ci−1, S, XS; θ) is calculated using the pair-wise attention-based pointer neural network.
291
+ 4.2. Pair-wise attention-based pointer neural network
292
+ Figure 2 uses a four-stop example to illustrate the architecture of the proposed model. The whole model
293
+ is based on the LSTM encoder and decoder structure. In particular, we use one LSTM (i.e., encoder) to
294
+ read the input sequence, one time step at a time, to obtain a large fixed dimensional vector representation,
295
+ and then to use another LSTM (i.e., decoder) to extract the output sequence. However, different from the
296
+ typical seq2seq model, we borrow the idea of the pointer network (Vinyals et al., 2015) to add a pair-wise
297
+ attention mechanism to predict the output sequence based on the attention mask over the input sequence. The
298
+ pair-wise attention is calculated based on an ASNN which was previously used for travel mode prediction
299
+ (Wang et al., 2020b). Model details will be shown in the following sections.
300
+ Intuitively, the LSTM encoder and decoder aim to capture the global view of the input information
301
+ (i.e., overall sequence pattern) by embedding the input sequence to hidden vector representation. While the
302
+ ASNN-based pair-wise attention aims to capture the local view (i.e., the relationship between two stops).
303
+ Our experiments in Section 5 demonstrate the importance of both global and local views in the sequence
304
+ prediction.
305
+ 7
306
+
307
+ Figure 2: Overall architecture of the pair-wise attention-based pointer neural network (adapted from Vinyals et al. (2015))
308
+ 4.2.1. LSTM encoder.
309
+ Given an arbitrary stop sequence (s1, ..., sn) as the input, let xi ∈ RK be the features of stop si, where
310
+ xi may include the package information, the customer information, and the geographical information of the
311
+ stop si. K is the number of features. The encoder computes a sequence of encoder output vectors (e1, ..., en)
312
+ by iterating the following:
313
+ hE
314
+ i , ei = LSTM(xi, hE
315
+ i−1; θE)
316
+ ∀i = 1, ..., n
317
+ (3)
318
+ where hE
319
+ i ∈ RKE
320
+ h is the encoder hidden vector with hE
321
+ 0 := 0. ei ∈ RKe is the encoder output vector. KE
322
+ h
323
+ and Ke are corresponding vector dimensions. θE is the learnable parameters in an encoder LSTM cell.
324
+ The calculation details of an LSTM cell can be found in Appendix A. The encoding process transforms a
325
+ sequence of features (x1, ..., xn) into a sequence of embedded representation (e1, ..., en). And the hidden
326
+ vector of the last time step (hE
327
+ n) includes the global information of the whole sequence, which will be used
328
+ for the LSTM decoder.
329
+ Figure 3: Illustration of LSTM ecnoder
330
+ 4.2.2. LSTM decoder.
331
+ The role of a decoder in the traditional seq2seq model (Figure 4) is to predict a new sequence one time step
332
+ at a time. However, in the pointer network structure with attention, the role of the decoder becomes producing
333
+ 8
334
+
335
+ AsNN Attention Component
336
+ Predict next is S3
337
+ Predict next is S1
338
+ Predict next is S2
339
+ Predict next is S4
340
+ S4
341
+ S4
342
+ Encoder
343
+ Decoderen
344
+ e1
345
+ e2
346
+ he
347
+ h2
348
+ he
349
+ LSTM
350
+ LSTM
351
+ LSTM
352
+ Decoder
353
+ X2
354
+ x1
355
+ Xna vector to modulate the pair-wise attention over inputs. Denote the output sequence as (ˆs(1), ..., ˆs(n)). Let
356
+ x(i) be the feature of stop ˆs(i).
357
+ At decoder step i, we have
358
+ hD
359
+ (i+1), d(i) = LSTM
360
+ ��
361
+ x(i)
362
+ w(i)
363
+
364
+ , hD
365
+ (i); θD
366
+
367
+ ∀i = 0, 1, ..., n
368
+ (4)
369
+ where hD
370
+ (i) ∈ RKD
371
+ h is the decoder hidden vector with hD
372
+ (0) = hE
373
+ n, d(i) ∈ RKd is the decoder output vector, KD
374
+ h
375
+ and Kd are corresponding vector dimensions, and θD are learnable parameters of the decoder LSTM cell.
376
+ Note that we set x(0) = xD and d(0) = dD, representing the features and the decoder output of the depot,
377
+ respectively. w(i) is the context vector calculated from the attention component, which will be explained in
378
+ the next section.
379
+ Figure 4: Illustration of LSTM decoder
380
+ 4.2.3. ASNN-based pair-wise attention.
381
+ The pair-wise attention aims to aggregate the global and local information to predict the next stop.
382
+ Specifically, at each decoder time step i ∈ {0, ..., n}, we know that the last predicted stop is ˆs(i). To predict
383
+ ˆs(i+1), we consider all candidate stops sj ∈ S, which is the set of all stops not yet visited. We want to
384
+ evaluate how possible that sj will be the next stop of ˆs(i). The information of the stop pair ˆs(i) and sj can be
385
+ represented by the following concatenated vector:
386
+ vj
387
+ (i) = concat(zj
388
+ (i), φ(x(i), xj), d(i), ej)
389
+ (5)
390
+ where zj
391
+ (i) is a vector of features associated with the stop pair (such as travel time from ˆs(i) to sj), and
392
+ φ(x(i), xj) represents a feature processing function to extract the pair-wise information from x(i) and xj. For
393
+ example, φ(·) may return geographical relationship between stops ˆs(i) and sj, and it may also drop features
394
+ not useful for the attention calculation. Intuitively, zj
395
+ (i) and φ(x(i), xj) contains only local information of the
396
+ stop pair, while d(i) and ej contain the global information of the whole stop set and previously visited stops.
397
+ 9
398
+
399
+ Predict next is S(n)
400
+ Output W(n)
401
+ .
402
+ Predict next is S(3)
403
+ ASNN
404
+ Output W(3)
405
+ Attention
406
+ Component
407
+ Predict next is S(2)
408
+ Output W(2)
409
+ Predict next is S(1)
410
+ Output W(1)
411
+ dp
412
+ d(1)
413
+ d(2)
414
+ d(n)
415
+ d(n-1)
416
+ 4
417
+ Encoder
418
+ LSTM
419
+ LSTM
420
+ LSTM
421
+ LSTM
422
+ LSTM
423
+ []
424
+ [x(1)
425
+ x(2)
426
+ x(n-1)
427
+ x(n)
428
+ W(1)
429
+ W(2))
430
+ W(n-1))
431
+ W(n)]Figure 5: Illustration of ASNN-based pair-wise attention
432
+ Given the pair-wise information vector vj
433
+ (i), we can calculate the attention of stop ˆs(i) to stop sj as:
434
+ uj
435
+ (i) = ASNN(vj
436
+ (i); θA)
437
+ ∀i, j = 1, ..., n
438
+ (6)
439
+ aj
440
+ (i) =
441
+ exp(uj
442
+ (i))
443
+ �n
444
+ j′=1 exp(uj′
445
+ (i))
446
+ ∀i, j = 1, ..., n
447
+ (7)
448
+ where aj
449
+ (i) ∈ R is attention of stop ˆs(i) to stop sj. ASNN(·; θA)) is a multilayer perception (MLP) with
450
+ the output dimension of one (i.e., uj
451
+ (i) ∈ R). θA are the learnable parameters of the ASNN. The name
452
+ “alternative specific” is because the same parametric network will be applied on all alternative stops sj ∈ S
453
+ separately (Wang et al., 2020b). Finally, we calculate the conditional probability to make the prediction:
454
+ P(ci+1 = j | c1, ..., ci, S, XS; θ) = aj
455
+ (i)
456
+ ∀i = 0, 1, ..., n, j = 1, ..., n
457
+ (8)
458
+ ˆs(i+1) = arg max
459
+ sj∈S\SV
460
+ (i)
461
+ aj
462
+ (i)
463
+ ∀i = 0, 1, ..., n
464
+ (9)
465
+ where SV
466
+ (i) = {ˆs(1), ..., ˆs(i)} is the set of stops that have been predicted (i.e., previously visited) until decoder
467
+ step i. Eqs. 8 and 9 indicate that the predicted next stop at step i is the one with highest attention among all
468
+ stops that have not been visited.
469
+ The pair-wise attention framework also leverages the attention information as the input for the next step.
470
+ This was achieved by introducing the context vector (Bahdanau et al., 2015):
471
+ w(i) =
472
+ n
473
+
474
+ j=1
475
+ aj
476
+ (i) · ej
477
+ (10)
478
+ The context vector is a weighted sum of all the encoder output vectors with attention as the weights. As
479
+ the attention provides the emphasis for stop prediction, w(i) helps to incorporate the encoded representation
480
+ of the last predicted stop for the next stop prediction. The inputs for the next LSTM cell thus will be the
481
+ 10
482
+
483
+ Predict next is S(i+1)
484
+ Output W(i+1)
485
+ Softmax
486
+ u
487
+ ASNN
488
+ ASNN
489
+ ASNN
490
+ e1
491
+ e2
492
+ en
493
+ he
494
+ D
495
+ LSTM
496
+ LSTM
497
+ LSTM
498
+ LSTM
499
+ X1
500
+ X2
501
+ Xn
502
+ x(i)
503
+ W(i)concatenation of the stop features and w(i), i.e.,
504
+
505
+ x(i)
506
+ w(i)
507
+
508
+ .
509
+ It is worth noting that, the specific architecture of ASNN(·; θA)) can be flexible depending on the input
510
+ pair-wise information. For example, if the information includes images or networks, convolutional neural
511
+ network or graph convolutional networks can be used for better extract features. In this study, we use the
512
+ MLP for simplification as it already outperforms benchmark models. The key idea is of the ASNN is to
513
+ share the same trainable parameter θA for all stop pairs so as to better capture various pair-wise information
514
+ in the training process.
515
+ 4.3. Sequence generation and selection
516
+ During inference, given a stop set S, the trained model with learned parameters θ∗ are used to generate
517
+ the sequence. Typically, in the seq2seq modeling framework, the final output sequence is selected as the one
518
+ with the highest probability, i.e.,
519
+ (sj∗
520
+ 1, ..., sj∗n), where j∗
521
+ 1, ..., j∗
522
+ n = arg max
523
+ j1,...,jn∈CS P(c1 = j1, ..., cn = jn | S, XS; θ∗)
524
+ (11)
525
+ where CS = {All permutations of {1, ..., n}}
526
+ Finding this optimal sequence is computationally impractical because of the combinatorial number of
527
+ possible output sequences. And so it is usually done with the greedy algorithm (i.e., always select the
528
+ most possible next stop) or the beam search procedure (i.e., find the best possible sequence among a set of
529
+ generated sequences given a beam size). However, in this study, we observe that the first predicted stop ˆs(1)
530
+ is critical for the quality of the generated sequence. The reason may be that the local relationship between a
531
+ stop pair (i.e., given the last stop to predict the next one) is easier to learn than the global relationship (i.e.,
532
+ predict the whole sequence). Hence, in this study, we first generate sequences using the greedy algorithm
533
+ with different initial stops, and select the one with the lowest operational cost. The intuition behind this
534
+ process is that, once the first stop is given, the model can follow the learned pair-wise relationship to generate
535
+ the sequence with relatively high accuracy. For all the generated sequences with different first stops, the
536
+ one with the lowest operation cost captures the global view of the sequence’s quality. Therefore, the final
537
+ sequence generation and selection algorithm is as follows:
538
+ Algorithm 1 Sequence generation
539
+ Input: Trained model, S
540
+ Output: Predicted stop sequence
541
+ 1: for s in S do
542
+ 2:
543
+ Let the first predicted stop be ˆs(1) = s
544
+ 3:
545
+ Predict the following stop sequence (ˆs(2), ..., ˆs(n)) using the greedy algorithm. Denote the predicted sequence
546
+ as Ps.
547
+ 4:
548
+ Calculate the total operation cost of the whole sequence (including depot), denoted as OCs.
549
+ return Ps∗ where s∗ = arg mins∈S OCs
550
+ 5. Case Study
551
+ 5.1. Dataset
552
+ The data used in our case study was made available as part of the Amazon Last Mile Routing Research
553
+ Challenge (Merchán et al., 2022). The dataset contains a total of 6,112 actual Amazon driver trajectories
554
+ 11
555
+
556
+ for the last-mile delivery from 5 major cities in the US: Austin, Boston, Chicago, Los Angeles, and Seattle.
557
+ Each route consists of a sequence of stops. Each stop represents the actual parking location of the driver, and
558
+ the package information (package numbers, package size, and planned service time) associated with each
559
+ stop is given. The stops are characterized by their latitudes and longitudes, and expected travel time between
560
+ stops are known.
561
+ Figure 6 shows the distribution of the number of stops per route and an example route. Most routes have
562
+ around 120 to 180 stops, and the maximum observed number of stops is around 250. Figure 6b shows an
563
+ example of an actual driver trajectory in Boston. Since the depot is far from the delivery stops, we attach the
564
+ complete route (with the depot indicated by a red dot) at the bottom left of the figure, while the main plot
565
+ only shows the delivery stops.
566
+ In this data set, each stop is associated with a zone ID (indicated by different colors in Figure 6b). When
567
+ Amazon generates planned routes for drivers, they usually expect drivers to finish the delivery for one zone
568
+ first, then go to another zone. And the actual driver trajectories also follow this pattern as shown in Figure
569
+ 6b (but the actual zone sequence may be different from the planned one). Therefore, in this study, we focus
570
+ on the problem of zone sequence prediction. That is, si in the case study section now represents a specific
571
+ zone, S represents the set of zones, and XS represents zone features. This transformation does not affect the
572
+ model structure proposed in Section 4. The only difference is that the new problem has a relatively smaller
573
+ scale compared to the stop sequence prediction because the number of zones in a route is smaller than that
574
+ of stops. The zone-to-zone travel time is calculated as the average travel time of all stop pairs between the
575
+ two zones. Figure 7 presents an illustrative example of the relationship between zone and stop sequences.
576
+ As the dataset does not contain the original planned sequence, we assume the planned zone sequence is the
577
+ one with the lowest total travel time (generated by a TSP solver, (sT
578
+ 1, ..., sT
579
+ n)). After generating the zone
580
+ sequence, we can restore the whole stop sequence by assuming that drivers within a specific zone follow an
581
+ optimal TSP tour. Details of the zone sequence to stop sequence generation can be found in Appendix B.
582
+ Figure 7: Relationship between stop sequence and zone sequence.
583
+ 5.2. Experimental setup
584
+ We randomly select 4,889 routes for model training and cross-validation, and the remaining 1,223 routes
585
+ are used to evaluate/test model performance.
586
+ We consider a one-layer LSTM for both the encoder and decoder with the hidden unit sizes of 32 (i.e.,
587
+ KD
588
+ h = Ke = KE
589
+ h = Kd = 32). And the ASNN is set with 2 hidden layers with 128 hidden units in each
590
+ layer. We train the model using Adam optimizer with a default learning rate of 0.001 and 30 training epochs.
591
+ To utilize the planned route information, the input zone sequence for the LSTM encoder is set as the TSP
592
+ 12
593
+
594
+ Zone sequence
595
+ Zone 1
596
+ Zone 2
597
+ Zone 3
598
+ B
599
+ C
600
+ D
601
+ E
602
+ G
603
+ H
604
+ A
605
+ Depot
606
+ Depot
607
+ Stop sequence(a) Number of stops distribution
608
+ (b) Actual route example
609
+ Figure 6: Description of dataset
610
+ 13
611
+
612
+ 400
613
+ 350
614
+ 300
615
+ 250
616
+ Counts
617
+ 200
618
+ 150
619
+ 100
620
+ 50
621
+ 0
622
+ 50
623
+ 100
624
+ 150
625
+ 200
626
+ Number of stops per routeASTBOSTON
627
+ Air
628
+ LOPREST
629
+ ZOVESTE-ET
630
+ Complete route
631
+ CHELSEA
632
+ CHARLESTOWN
633
+ BOSTONresult (i.e., lowest travel time). That is, the input sequence (s1, ..., sn) = (sT
634
+ 1, ..., sT
635
+ n).
636
+ In the case study, xi represents zone features, including the latitude and longitude of the zone center,
637
+ number of stops in the zone, number of intersections in the zone, number of packages in the zone, total
638
+ service time in the zone, total package size in the zone, and the travel time from this zone to all other zones.
639
+ The zone pair features zj
640
+ (i) includes the travel time from ˆs(i) to sj and zone ID relationship characteristics.
641
+ For example, the zone IDs “B-6.2C” and “B-6.3A” signal that they belong to the higher-level cluster “B-6”.
642
+ As we assume all pair-wise features are captured by zj
643
+ (i), φ(x(i), xj) is not specified in this case study.
644
+ Consistent with the Amazon Last Mile Routing Research Challenge, we evaluate the quality of the
645
+ predicted stop sequences using a “disparity score” defined as follows:
646
+ R(A, B) = SD(A, B) · ERPnorm(A, B)
647
+ ERPe(A, B)
648
+ (12)
649
+ where R(A, B) is the disparity score for the actual sequence A and predicted sequence B, and SD(A, B) is
650
+ the sequence deviation defined as
651
+ SD(A, B) =
652
+ 2
653
+ n(n − 1)
654
+ n
655
+
656
+ i=2
657
+
658
+ |c[Bi] − c[Bi−1]| − 1
659
+
660
+ (13)
661
+ where n is the total number of stops, Bi is the i-th stop of sequence B, c[Bi] is the index of stop Bi in the
662
+ actual sequence A (i.e., its position in sequence A). In the case of A = B (i.e., perfectly predicted), we have
663
+ c[Bi] − c[Bi−1] = 1 for all i = 2, ..., n, and SD(A, B) = 0.
664
+ ERPnorm(A, B) is the Edited Distance with Real Penalty (ERP) defined by the following recursive
665
+ formula:
666
+ ERPnorm(A, B) = ERPnorm(A2:|A|, B2:|B|) + Timenorm(A1, B1)
667
+ (14)
668
+ where Timenorm(si, sj) =
669
+ Time(si,sj)
670
+
671
+ j′∈{1,...,n} Time(si,sj′) is the normalized travel time from stop si to stop
672
+ sj.
673
+ ERPe(A, B) is the number of edit operations (insertions, substitutions, or deletions) required to
674
+ transform sequence A to sequence B as when executing the recursive ERPnorm formulation. Hence, the
675
+ ratio ERPnorm(A,B)
676
+ ERPe(A,B) represents the average normalized travel time between the two stops involved in each ERP
677
+ edit operation. In the case of A = B, we have ERPnorm(A,B)
678
+ ERPe(A,B)
679
+ = 0.
680
+ The disparity score R(A, B) describes how well the model-estimated sequence matches the known actual
681
+ sequence. Lower score indicates better model performance. A score of zero means perfect prediction. The
682
+ final model performance is evaluated by the mean score over all routes in the test set.
683
+ In addition to the disparity score, we also evaluate the prediction accuracy of the first four zones in each
684
+ route. We choose the first four because the minimum number of zones in a route is four.
685
+ 5.3. Benchmark models
686
+ The following optimization and machine learning models are used as benchmarks to compare with the
687
+ proposed approach.
688
+ Conventional TSP. The first benchmark model is the zone sequence generated by conventional TSP,
689
+ which we treat as the planned route with the lowest travel time.
690
+ ASNN model. The ASNN component can be trained to predict the next zone given the current zone, and
691
+ the prediction sequence can be constructed in a greedy way starting from the given depot. The training zone
692
+ 14
693
+
694
+ pairs (including from depot to the first zone) are extracted from all sequences in the training routes. And the
695
+ input features are the same as the ASNN component in the proposed model except for (d(i), ej) (i.e., output
696
+ vectors from LSTM decoder and encoder, respectively). All hyper-parameters of the ASNN model are the
697
+ same as the attention component.
698
+ Inspired by the importance of the first zone, we also implement another sequence generation method
699
+ similar to Section 4.3. That is, we go through all zones in a route and assume it is the first zone, then use the
700
+ trained ASNN to predict the remaining sequence. The final sequence is selected as the one with the lowest
701
+ travel time.
702
+ LSTM-encoder-decoder. The LSTM-encoder-decoder (LSTM-E-D) architecture is a typical seq2seq
703
+ model proposed by Sutskever et al. (2014). The model structure is shown in Figure 8. In the decoder stage,
704
+ the model outputs the predicted zone based on last predicted zone’s information. The model formulation can
705
+ be written as
706
+ hE
707
+ i , ei = LSTM(xi, hE
708
+ i−1; θE)
709
+ ∀i = 1, ..., n
710
+ (15)
711
+ hD
712
+ (i+1), d(i) = LSTM(x(i), hD
713
+ (i); θD)
714
+ ∀i = 0, 1, ..., n
715
+ (16)
716
+ The decoder output vector d(i) are, then feed into a fully-connected (FC) layer to calculate probability of the
717
+ next stop:
718
+ g(i) = FC(d(i); θFC)
719
+ ∀i = 1, ..., n
720
+ (17)
721
+ P(ci+1 | c1, ..., ci, S, XS; θ) = Softmax(g(i))
722
+ ∀i = 1, ..., n
723
+ (18)
724
+ where g(i) ∈ RKz, Kz is the maximum number of zones in the dataset. And the next predicted stop is
725
+ selected by maximizing P(ci+1 = j | c1, ..., ci, S, XS; θ) for all sj ∈ S \ SV
726
+ (i) (i.e., the zones that are not in
727
+ the route and that have been visited are excluded).
728
+ Figure 8: Model architecture of the LSTM-E-D seq2seq prediction model.
729
+ Original Pointer Network. Another benchmark model is the original pointer network (Pnt Net) proposed
730
+ by (Vinyals et al., 2015). The overall architecture of the pointer network is similar to the proposed model
731
+ 15
732
+
733
+ S4
734
+ S2
735
+ S1
736
+ S3
737
+ End
738
+ FC + Softmax
739
+ S1
740
+ S2
741
+ S3
742
+ S4
743
+ Ds
744
+ S4
745
+ S2
746
+ S3
747
+ S
748
+ Encoder
749
+ Decoderexcept for the attention component. Specifically, the pointer network calculates attention as:
750
+ uj
751
+ (i) = W T
752
+ 1 tanh(W2ej + W3d(i))
753
+ ∀i, j = 1, ..., n
754
+ (19)
755
+ aj
756
+ (i) =
757
+ exp(uj
758
+ (i))
759
+ �n
760
+ j′=1 exp(uj′
761
+ (i))
762
+ ∀i, j = 1, ..., n
763
+ (20)
764
+ The original pointer network does not include the pair-wise local information (zj
765
+ (i), φ(x(i), xj)), and the
766
+ attention calculation is only quantified from three learnable parameters W1, W2, and W3, which may limit
767
+ its capacity in prediction. We observe that the original pointer network without local information performs
768
+ extremely badly. For a fair comparison, we add the local information with the similar format in Eq. 19 as:
769
+ uj
770
+ (i) = W T
771
+ 1 tanh(W2ej + W3d(i)) + W4
772
+
773
+ zj
774
+ (i)
775
+ φ(x(i), xj)
776
+
777
+ ∀i, j = 1, ..., n
778
+ (21)
779
+ After training the model, we generate the final sequence with the greedy algorithm and Algorithm 1,
780
+ respectively.
781
+ 5.4. Results
782
+ 5.4.1. Model comparison.
783
+ Table 1 presents the performance of different models. Note that for all approaches except for the TSP, we
784
+ generate sequences based on two different methods (greedy and Algorithm 1) for comparison. The standard
785
+ deviation of disparity scores is taken over all testing routes. Results show that sequence generation with
786
+ Algorithm 1 (i.e., iterating different first zones) can consistently reduce the disparity score for all machine
787
+ learning methods.
788
+ It implies that the first zone prediction and the global view (i.e., shortest path) are
789
+ important for estimating the driver’s trajectory.
790
+ The proposed method outperforms all other models, both in disparity scores and prediction accuracy.
791
+ This means the proposed pair-wise ASNN-based attention (Eq. 6) has better performance than the original
792
+ content-based attention (Eq. 21). The comparison between LSTM-E-D and Pnt Net models demonstrates
793
+ the effectiveness of the attention mechanism.
794
+ All machine learning models except for LSTM-E-D can
795
+ outperform the baseline TSP sequence with Algorithm 1 sequence generation method, suggesting that the
796
+ hidden trajectory patterns can be learned from the data.
797
+ Another observation is that, the prediction accuracy and disparity score do not always move in the same
798
+ direction. For example, the LSTM-E-D model with Algorithm 1 sequence generation, though has lower
799
+ accuracy, shows a better disparity score. This is because the accuracy metric does not differentiate “how
800
+ wrong an erroneous prediction is”. By the definition of disparity score, if a stop is si but the prediction is
801
+ sj, and sj and si are geographically close to each other, the score does not worsen too much. This suggests
802
+ a future research direction in using disparity score as the loss function (e.g., training by RL) instead of
803
+ cross-entropy loss.
804
+ Figure 9 shows the distribution of disparity scores for our proposed method with Algorithm 1 sequence
805
+ generation (i.e., the best model). We observe that the prediction performance varies a lot across different
806
+ routes. There is a huge proportion of routes with very small disparity scores (less than 0.01). The mean
807
+ score is impacted by outlier routes. The median score is 0.0340, which is smaller than the mean value.
808
+ 16
809
+
810
+ Table 1: Model performance
811
+ Sequence generation
812
+ Model
813
+ Disparity score
814
+ Prediction accuracy
815
+ Mean
816
+ Std. Dev
817
+ 1st zone
818
+ 2nd zone
819
+ 3rd zone
820
+ 4th zone
821
+ -
822
+ TSP
823
+ 0.0443
824
+ 0.0289
825
+ 0.207
826
+ 0.185
827
+ 0.163
828
+ 0.168
829
+ Greedy
830
+ ASNN
831
+ 0.0470
832
+ 0.0289
833
+ 0.150
834
+ 0.141
835
+ 0.119
836
+ 0.123
837
+ LSTM-E-D
838
+ 0.0503
839
+ 0.0313
840
+ 0.207
841
+ 0.183
842
+ 0.161
843
+ 0.166
844
+ Pnt Net
845
+ 0.0460
846
+ 0.0309
847
+ 0.224
848
+ 0.204
849
+ 0.186
850
+ 0.165
851
+ Ours
852
+ 0.0417
853
+ 0.0306
854
+ 0.241
855
+ 0.231
856
+ 0.224
857
+ 0.221
858
+ Algorithm 1
859
+ ASNN
860
+ 0.0429
861
+ 0.0299
862
+ 0.221
863
+ 0.213
864
+ 0.203
865
+ 0.195
866
+ LSTM-E-D
867
+ 0.0501
868
+ 0.0305
869
+ 0.182
870
+ 0.156
871
+ 0.142
872
+ 0.149
873
+ Pnt Net
874
+ 0.0382
875
+ 0.0301
876
+ 0.286
877
+ 0.273
878
+ 0.262
879
+ 0.274
880
+ Ours
881
+ 0.0369
882
+ 0.0301
883
+ 0.320
884
+ 0.310
885
+ 0.303
886
+ 0.314
887
+ Figure 9: Disparity score distribution of the best model
888
+ 5.4.2. Factors on trajectory predictability.
889
+ As our proposed model exhibits various levels of predictability across different routes, we aim to
890
+ investigate which attributes of a route cause high (or low) predictability. This can be done by running a
891
+ regression model with the disparity score as the dependent variable and route attributes (e.g., locations,
892
+ departure time, package numbers) as independent variables. The variables used are defined as follows:
893
+ • Total planned service time: The estimated time to deliver all packages in the route (service time only,
894
+ excluding travel time).
895
+ • Earliest time window constraint: The earliest due time to deliver packages with time window constraint
896
+ minus the vehicle departure time. The smaller the value, the tighter the time limit.
897
+ • Avg. # traffic signals: Average number of traffic signals in each zone of the route (obtained from
898
+ OpenStreetMap data).
899
+ 17
900
+
901
+ 250
902
+ Mean = 0.0369
903
+ Median = 0.034
904
+ 200
905
+ 150
906
+ Counts
907
+ 100
908
+ 50
909
+ 0.00
910
+ 0.05
911
+ 0.10
912
+ 0.15
913
+ 0.20
914
+ Disparity scores• If high-quality route: A dummy variable indicating whether the route is labeled as “high quality” by
915
+ Amazon or not (Yes = 1). High quality means the actual travel time of the route is similar to or better
916
+ than Amazon’s expectation.
917
+ • If in Location: A dummy variable indicating whether the route is in a specific city or not (Yes = 1).
918
+ • If departure Time: A dummy variable indicating the (local) departure time (e.g., before 7AM, after
919
+ 10AM).
920
+ Table 2 shows the results of the regression. Since the dependent variable is disparity scores, a negative
921
+ sign indicates a positive impact on the predictability. We observe that routes with tighter time window
922
+ constraints and more stops are easier to predict. This may be due to the fact that these routes are usually
923
+ harder to deliver. Hence, to avoid the risk of violating time constraints or delay, drivers tend to follow the
924
+ planned routes and thus the route sequences are easier predict. We also find that routes associated with larger
925
+ vans (i.e., larger vehicle capacity) are more predictable. The reason may be that larger vans are less flexible
926
+ in choosing different routes, thus drivers are more likely to follow the navigation. Another important factor
927
+ for better predictability is high-quality routes. This may be because high-quality routes are closer to the TSP
928
+ sequence which we use as inputs. Finally, routes in LA are more predictable than in other areas such as
929
+ Chicago and Boston.
930
+ Table 2: Factors on trajectory predictability
931
+ Variables
932
+ Coefficients (×10−3)
933
+ Variables
934
+ Coefficients (×10−3)
935
+ Intercept
936
+ 91.07 **
937
+ If high quality route
938
+ -1.66×10−14 **
939
+ Total # of packages
940
+ 0.059
941
+ If in LA
942
+ -4.998 *
943
+ Total planned service time
944
+ -0.476
945
+ If in Chicago
946
+ 0.783
947
+ Earliest time window constraint
948
+ -3.047 **
949
+ If in Boston
950
+ -3.354
951
+ Avg. # traffic signals
952
+ -3.255
953
+ If on weekends
954
+ 1.775
955
+ Total # of stops
956
+ -0.142 **
957
+ If departure before 7AM
958
+ 0.582
959
+ Vehicle capacity (m3)
960
+ -6.041 *
961
+ If departure after 10AM
962
+ -2.704
963
+ Number of routes: 1,002.
964
+ R2: 0.065;
965
+ ∗∗: p-value < 0.01; ∗: p-value < 0.05.
966
+ 5.4.3. Impact of input sequence.
967
+ All machine learning models in Table 1 (except for ASNN) have the LSTM encoder component, which
968
+ requires the specification of input zone sequence. As mentioned in Section 5.2, we currently use the TSP
969
+ sequence as input. It is worth exploring the model performance if we use a random zone sequence instead,
970
+ which corresponds to the scenario without planned route information. Table 3 shows the model performance
971
+ without the TSP sequence information. Since the ASNN result does not rely on TSP information, it is not
972
+ listed in the table. Results show that the LSTM-E-D model becomes much worse with a random sequence as
973
+ inputs, while the performance of Pnt Net and our method is only slightly affected. Even without the planned
974
+ route information, the proposed model can still provide a reasonable estimation of driver trajectories.
975
+ 18
976
+
977
+ Table 3: Model performance without TSP information
978
+ Sequence generation
979
+ Model
980
+ Disparity score
981
+ Prediction accuracy
982
+ Mean
983
+ Std. Dev
984
+ 1st zone
985
+ 2nd zone
986
+ 3rd zone
987
+ 4th zone
988
+ Greedy
989
+ LSTM-E-D
990
+ 0.1176
991
+ 0.0498
992
+ 0.045
993
+ 0.047
994
+ 0.041
995
+ 0.050
996
+ Pnt Net
997
+ 0.0512
998
+ 0.0323
999
+ 0.090
1000
+ 0.096
1001
+ 0.097
1002
+ 0.096
1003
+ Ours
1004
+ 0.0426
1005
+ 0.0311
1006
+ 0.204
1007
+ 0.192
1008
+ 0.195
1009
+ 0.196
1010
+ Algorithm 1
1011
+ LSTM-E-D
1012
+ 0.1054
1013
+ 0.0463
1014
+ 0.103
1015
+ 0.061
1016
+ 0.049
1017
+ 0.052
1018
+ Pnt Net
1019
+ 0.0398
1020
+ 0.0311
1021
+ 0.298
1022
+ 0.284
1023
+ 0.273
1024
+ 0.273
1025
+ Ours
1026
+ 0.0376
1027
+ 0.0307
1028
+ 0.316
1029
+ 0.298
1030
+ 0.302
1031
+ 0.298
1032
+ 5.5. Summary
1033
+ Our numerical results show that our proposed model outperforms its benchmarks in terms of disparity
1034
+ scores and prediction accuracy, meaning that it can better predict the actual route trajectories taken by drivers.
1035
+ The comparison with benchmark models shows that our proposed ASNN-based pair-wise attention mecha-
1036
+ nism and our sequence generation algorithm (Algorithm 1) are both helpful for the prediction. Moreover,
1037
+ we can observe that the predictive performance varies across different routes. Factors such as route quality,
1038
+ delivery time windows, and the total number of stops of a route affect predictability. Finally, the proposed
1039
+ model is insensitive to the input sequence. The prediction performance only slightly decreases when the
1040
+ input sequence is changed from the TSP solution to a random stop sequence. This property implies that we
1041
+ only need the set of stops to implement the model and obtain high-quality solution, while information on the
1042
+ planned route sequence is not strictly required.
1043
+ 6. Conclusion and Future Research
1044
+ In this paper, we propose a pair-wise attention-based pointer neural network that predicts actual driver
1045
+ trajectories on last-mile delivery routes for given sets of delivery stops. Compared to previously proposed
1046
+ pointer networks, this study leverages a new alternative specific neural network-based attention mechanism
1047
+ to incorporate pair-wise local information (such as relative distances and locations of stops) for the attention
1048
+ calculation. To better capture the global efficiency of a route in terms of operational cost (i.e., total travel
1049
+ time), we further propose a new sequence generation algorithm that finds the lowest-cost route sequence by
1050
+ iterating through different first stops.
1051
+ We apply our proposed method to a large set of real operational route data provided by the Amazon
1052
+ Last-Mile Routing Research Challenge in 2021. The results show that our proposed method can outperform
1053
+ a wide range of benchmark models in terms of both the disparity score and prediction accuracy, meaning
1054
+ that the predicted route sequence is closer to the actual sequence executed by drivers. Compared to the best
1055
+ benchmark model (original pointer network), our method reduces the disparity score from 0.0382 to 0.0369,
1056
+ and increases the average prediction accuracy of the first four zones from 0.229 to 0.312. Moreover, our
1057
+ proposed sequence generation method can consistently improve the prediction performance for all models.
1058
+ The disparity scores are reduced by 10-20% across different models. Lastly, we show that the proposed
1059
+ methodology is robust against changes in the input sequence pattern. Compared to an optimal TSP solution
1060
+ as the input sequence, a random input sequence only slightly increases the disparity score from 0.0369 to
1061
+ 0.0376.
1062
+ 19
1063
+
1064
+ The data-driven route planning method proposed in this paper has several highly relevant practical
1065
+ implications. First, our proposed model performs well at predicting stop sequences that would be preferable
1066
+ to delivery drivers in a real operational environment, even if it is not provided with a theoretically optimal
1067
+ (i.e., minimal route duration) planned TSP sequence as an input. Therefore, the model can be used to
1068
+ generate a predicted actual stop sequence that a driver would likely be taking for a given set of delivery
1069
+ stops. The prediction can serve as a new ‘empirical’ planned route that is informed by historical driver
1070
+ behavior and thus more consistent with the driver’s experience and preferences. Second, by comparing
1071
+ the stop sequence predicted by our model with the traditional, TSP-based planned stop sequence, a route
1072
+ planner may infer potential reasons for the drivers’ deviations and adjust the company’s planning procedures
1073
+ and/or driver incentives if necessary. Third, as stop sequence generation using machine learning models is
1074
+ computationally more efficient than traditional optimization-based approaches, a trained machine learning
1075
+ model can be applied in real-time to quickly re-optimize routes when drivers are unexpectedly forced to
1076
+ deviate from their original stop sequence (e.g., due to road closures) and need updated routing strategies.
1077
+ Based on the work presented in this paper, a number of fruitful future research avenues arise. First, instead
1078
+ of focusing on stop sequence prediction, future work may improve the interpretability of such prediction
1079
+ models and develop machine learning approaches that better explain which factors cause drivers to deviate
1080
+ from a planned stop sequence and how they affect their actual route trajectories. Second, future work should
1081
+ attempt to combine the strengths of optimization-based route planning approaches and machine learning by
1082
+ incorporating tacit driver knowledge learned via machine learning models into route optimization algorithms.
1083
+ Appendices
1084
+ Appendix A. Mathematical Formulation of a LSTM Cell
1085
+ The details of an LSTM cell, ht, et = LSTM(xt, ht−1; θ), is shown below:
1086
+ ft = σg(Wfxt + Ufht−1 + bf)
1087
+ (A.1)
1088
+ it = σg(Wixt + Uiht−1 + bi)
1089
+ (A.2)
1090
+ ot = σg(Woxt + Uoht−1 + bo)
1091
+ (A.3)
1092
+ ˜ct = σc(Wcxt + Ucht−1 + bc)
1093
+ (A.4)
1094
+ ct = ft ◦ ct−1 + it ◦ ˜ct
1095
+ (A.5)
1096
+ ht = ot ◦ σh(ct)
1097
+ (A.6)
1098
+ et = ht (if this is a single layer one-directional LSTM)
1099
+ (A.7)
1100
+ where [Wf, Wi, Wo, Wc, Uf, Ui, Uo, Uc, bf, bi, bo, bc] = θ is the vector of learnable parameters. xt is the
1101
+ input vector to the LSTM unit. ft is the forget gate’s activation vector. it is the input/update gate’s activation
1102
+ vector. ot is the output gate’s activation vector. ht is the hidden state vector. et is the output vector of
1103
+ the LSTM. Note that for a multi-layer or bidirectional LSTM, et may not equal to ht. In this study, we
1104
+ use a single layer one-directional LSTM and thus have et = ht. More details on the output vector can be
1105
+ found in Pytorch (2021). ˜ct is the cell input activation vector. ct is the cell state vector. “◦” indicates the
1106
+ component-wise multiplication.
1107
+ 20
1108
+
1109
+ Appendix B. From Zone Sequence to Stop Sequence
1110
+ The complete stop sequence is generated based on the given zone sequence. The detailed generation
1111
+ process is shown in Algorithm 2.
1112
+ Algorithm 2 Complete sequence generation. Input: zone sequence (ˆz(1), .., ˆz(n)), depot DS, set of stops in
1113
+ each zone S(i), i = 1, ..., n. PathTSP(S, sfirst, slast) and TourTSP(S) are two oracle functions for solving
1114
+ path and tour TSP problems given the set of stops S, first stop sfirst and last stop slast to be visited.
1115
+ 1: function CompleteSeqGeneration((ˆz(1), .., ˆz(n)), {S(i), i = 1, , , n})
1116
+ 2:
1117
+ sprev ← DS
1118
+ 3:
1119
+ s∗
1120
+ complete ← (sprev)
1121
+ ▷ Initialize the complete stop sequence with depot
1122
+ 4:
1123
+ for i ∈ {1, ..., n − 1} do
1124
+ 5:
1125
+ Sfirst ← Set of three stops in S(i) that are closest to sprev
1126
+ 6:
1127
+ Slast ← Set of three stops in S(i) that are closest to all stops in Si+1 on average
1128
+ 7:
1129
+ P(i) ← ∅
1130
+ ▷ Initialize the set of optimal paths in zone ˆz(i)
1131
+ 8:
1132
+ for sfirst ∈ Sfirst do
1133
+ 9:
1134
+ for slast ∈ Slast do
1135
+ 10:
1136
+ if sfirst = slast then
1137
+ 11:
1138
+ ˆptemp, ttemp = TourTSP(S(i))
1139
+ ▷ Solve the optimal tour and travel time for zone ˆz(i)
1140
+ 12:
1141
+ Delete the last edge back to sfirst in the tour ˆptemp. Let the new path and travel time be ˆp′
1142
+ temp
1143
+ and t′
1144
+ temp
1145
+ 13:
1146
+ Add ˆp′
1147
+ temp and t′
1148
+ temp to P(i)
1149
+ 14:
1150
+ else
1151
+ 15:
1152
+ ˆptemp, ttemp = PathTSP(S(i), sfirst, slast) ▷ Solve the optimal path and travel time for zone i
1153
+ 16:
1154
+ Add ˆptemp and ttemp to P(i)
1155
+ 17:
1156
+ ˆp(i) ← Path in P(i) with the minimum travel time
1157
+ 18:
1158
+ s∗
1159
+ complete ← (s∗
1160
+ complete, ˆp(i))
1161
+ ▷ Concatenate two sequence
1162
+ 19:
1163
+ sprev ← Last stop of path ˆp(i)
1164
+ 20:
1165
+ s∗
1166
+ complete ← (s∗
1167
+ complete, DS)
1168
+ ▷ Concatenate the last stop as the depot
1169
+ 21:
1170
+ return s∗
1171
+ complete
1172
+ Consider an optimal zone sequence, (ˆz(1), .., ˆz(n)), generated from the proposed machine learning
1173
+ method. We can always add the depot before the first and after last zone (i.e., (DS, ˆz(1), .., ˆz(n), DS)) and
1174
+ make the whole zone sequence a loop. For each zone ˆz(i), we aim to generate a within-zone path ˆp(i), and
1175
+ the final stop sequence will be (DS, ˆp(i), ..., ˆp(n), DS).
1176
+ When generating ˆp(i) for zone ˆz(i), we assume ˆp(i−1) is known (generated from the last step and
1177
+ ˆp(0) = (DS)). Let the set of all stops in zone ˆz(i) be S(i). We identify three potential first stops and last
1178
+ stops of path ˆp(i) based on following rules:
1179
+ • Three potential first stops of ˆp(i) are the three most closest stops (in travel time) to ˆp(i−1)’s last stop.
1180
+ • Three potential last stops of ˆp(i) are the three most closest stops (in travel time) to all stops in S(i+1)
1181
+ on average. Note that S(n+1) = {DS}
1182
+ With three potential first stops and last stops, we then solve path TSP problems between any first and last
1183
+ stop pair to generate the potential optimal inner zone path with the shortest travel time. In this step, at most
1184
+ 21
1185
+
1186
+ nine small-scale path TSP problems will be solved since there might be overlapping between the first and
1187
+ the last stops. If the first and the last stops are identical, we solve a tour TSP problem and output the path by
1188
+ deleting the last edge which traverses back to the first stop in the tour.
1189
+ After having all potential inner zone paths and total path travel time between any first and last stop pair,
1190
+ we keep the path with the minimum travel time as the inner zone sequence, ˆp(i). The key assumption we
1191
+ make here about drivers is that they will deliver packages within a zone following a path that minimizes their
1192
+ total travel time. With the optimal inner zone stop sequence of the current zone, we then move to the next
1193
+ visited zone in the optimal zone sequence and repeat the same procedure until we generate the complete stop
1194
+ sequence.
1195
+ References
1196
+ Applegate, D.L., Bixby, R.E., Chvatal, V., Cook, W.J., 2006. The Travelling Salesman Problem. Princeton
1197
+ University Press. URL: http://www.jstor.org/stable/j.ctt7s8xg.9.
1198
+ Bahdanau, D., Cho, K.H., Bengio, Y., 2015. Neural machine translation by jointly learning to align and
1199
+ translate, in: 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track
1200
+ Proceedings.
1201
+ Bello, I., Pham, H., Le, Q.V., Norouzi, M., Bengio, S., 2019. Neural combinatorial optimization with
1202
+ reinforcement learning. 5th International Conference on Learning Representations, ICLR 2017 - Workshop
1203
+ Track Proceedings , 1–15.
1204
+ Cheikhrouhou, O., Khoufi, I., 2021. A comprehensive survey on the multiple traveling salesman problem:
1205
+ Applications, approaches and taxonomy. Computer Science Review 40, 100369. URL: https://doi.
1206
+ org/10.1016/j.cosrev.2021.100369, doi:10.1016/j.cosrev.2021.100369.
1207
+ Cho, K., van Merrienboer, B., Gulcehre, C., Bougares, F., Schwenk, H., Bengio, Y., 2014. Learning phrase
1208
+ representations using rnn encoder-decoder for statistical machine translation, in: Conference on Empirical
1209
+ Methods in Natural Language Processing (EMNLP 2014).
1210
+ Chorowski, J., Bahdanau, D., Serdyuk, D., Cho, K., Bengio, Y., 2015. Attention-based models for speech
1211
+ recognition. arXiv preprint arXiv:1506.07503 .
1212
+ Chung, J., Gulcehre, C., Cho, K., Bengio, Y., 2014. Empirical evaluation of gated recurrent neural networks
1213
+ on sequence modeling. arXiv preprint arXiv:1412.3555 .
1214
+ Cortes, J.D., Suzuki, Y., 2021. Last-mile delivery efficiency: en route transloading in the parcel delivery
1215
+ industry. International Journal of Production Research 0, 1–18. URL: https://doi.org/00207543.
1216
+ 2021.1907628, doi:10.1080/00207543.2021.1907628.
1217
+ Davendra, D., Bialic-Davendra, M., 2020. Introductory chapter: Traveling salesman problem - an overview,
1218
+ in: Novel Trends in the Traveling Salesman Problem. IntechOpen. URL: https://doi.org/10.5772/
1219
+ intechopen.94435, doi:10.5772/intechopen.94435.
1220
+ Graves, A., 2013. Generating Sequences With Recurrent Neural Networks URL: http://arxiv.org/
1221
+ abs/1308.0850.
1222
+ Halim, A.H., Ismail, I., 2017. Combinatorial optimization: Comparison of heuristic algorithms in travelling
1223
+ salesman problem. Archives of Computational Methods in Engineering 26, 367–380. URL: https:
1224
+ //doi.org/10.1007/s11831-017-9247-y, doi:10.1007/s11831-017-9247-y.
1225
+ Hochreiter, S., Schmidhuber, J., 1997. Long Short-Term Memory. Neural Computation 9, 1735–1780.
1226
+ URL:
1227
+ https://doi.org/10.1162/neco.1997.9.8.1735,
1228
+ doi:10.1162/neco.1997.9.8.1735,
1229
+ arXiv:https://direct.mit.edu/neco/article-pdf/9/8/1735/813796/neco.1997.9.8.1735.pdf.
1230
+ 22
1231
+
1232
+ Huang, C.Z.A., Vaswani, A., Uszkoreit, J., Shazeer, N., Simon, I., Hawthorne, C., Dai, A.M., Hoffman,
1233
+ M.D., Dinculescu, M., Eck, D., 2018. Music transformer. arXiv preprint arXiv:1809.04281 .
1234
+ Jaller, M., Holguín-Veras, J., Hodge, S., 2013. Parking in the city. Transportation Research Record ,
1235
+ 46–56doi:10.3141/2379-06.
1236
+ Jonker, R., Volgenant, T., 1983. Transforming asymmetric into symmetric traveling salesman problems. Op-
1237
+ erations Research Letters 2, 161–163. URL: https://doi.org/10.1016/0167-6377(83)90048-2,
1238
+ doi:10.1016/0167-6377(83)90048-2.
1239
+ Joshi, C.K., Laurent, T., Bresson, X., 2019. On Learning Paradigms for the Travelling Salesman Problem.
1240
+ Advances in Neural Information Processing Systems , 1–9URL: http://arxiv.org/abs/1910.07210.
1241
+ Karatzoglou, A., Jablonski, A., Beigl, M., 2018. A seq2seq learning approach for modeling semantic
1242
+ trajectories and predicting the next location, in: Proceedings of the 26th ACM SIGSPATIAL International
1243
+ Conference on Advances in Geographic Information Systems, Association for Computing Machinery,
1244
+ New York, NY, USA. p. 528–531. URL: https://doi.org/10.1145/3274895.3274983, doi:10.
1245
+ 1145/3274895.3274983.
1246
+ Kool, W., Van Hoof, H., Welling, M., 2019. Attention, learn to solve routing problems! 7th International
1247
+ Conference on Learning Representations, ICLR 2019 , 1–25.
1248
+ Küçükoğlu, İ., Dewil, R., Cattrysse, D., 2019. Hybrid simulated annealing and tabu search method for
1249
+ the electric travelling salesman problem with time windows and mixed charging rates. Expert Systems
1250
+ with Applications 134, 279–303. URL: https://doi.org/10.1016/j.eswa.2019.05.037, doi:10.
1251
+ 1016/j.eswa.2019.05.037.
1252
+ Liang, Z., Du, J., Li, C., 2020. Abstractive social media text summarization using selective reinforced
1253
+ seq2seq attention model. Neurocomputing 410, 432–440.
1254
+ Lim, S.F.W., Winkenbach, M., 2019. Configuring the last-mile in business-to-consumer e-retailing. Califor-
1255
+ nia Management Review 61, 132–154.
1256
+ Liu, S., Jiang, H., Chen, S., Ye, J., He, R., Sun, Z., 2020. Integrating dijkstra’s algorithm into deep inverse
1257
+ reinforcement learning for food delivery route planning. Transportation Research Part E: Logistics and
1258
+ Transportation Review 142, 102070. URL: https://www.sciencedirect.com/science/article/
1259
+ pii/S1366554520307213, doi:https://doi.org/10.1016/j.tre.2020.102070.
1260
+ Liu, T., Wang, K., Sha, L., Chang, B., Sui, Z., 2018. Table-to-text generation by structure-aware seq2seq
1261
+ learning, in: Thirty-Second AAAI Conference on Artificial Intelligence.
1262
+ Lu, Y., Rai, H., Chang, J., Knyazev, B., Yu, G., Shekhar, S., Taylor, G.W., Volkovs, M., 2021. Context-
1263
+ aware scene graph generation with seq2seq transformers, in: Proceedings of the IEEE/CVF International
1264
+ Conference on Computer Vision (ICCV), pp. 15931–15941.
1265
+ Luong, M.T., Pham, H., Manning, C.D., 2015. Effective approaches to attention-based neural machine
1266
+ translation. arXiv preprint arXiv:1508.04025 .
1267
+ Ma, Q., Ge, S., He, D., Thaker, D., Drori, I., 2019. Combinatorial Optimization by Graph Pointer Networks
1268
+ and Hierarchical Reinforcement Learning URL: http://arxiv.org/abs/1911.04936.
1269
+ Matai, R., Singh, S., Lal, M., 2010. Traveling salesman problem: an overview of applications, formulations,
1270
+ and solution approaches, in: Traveling Salesman Problem, Theory and Applications. InTech. URL:
1271
+ https://doi.org/10.5772/12909, doi:10.5772/12909.
1272
+ McKinsey
1273
+ &
1274
+ Company,
1275
+ 2021.
1276
+ How
1277
+ e-commerce
1278
+ share
1279
+ of
1280
+ retail
1281
+ soared
1282
+ across
1283
+ the
1284
+ globe:
1285
+ A
1286
+ look
1287
+ at
1288
+ eight
1289
+ countries.
1290
+ URL:
1291
+ https://www.mckinsey.com/featured-insights/
1292
+ coronavirus-leading-through-the-crisis/charting-the-path-to-the-next-normal/
1293
+ 23
1294
+
1295
+ how-e-commerce-share-of-retail-soared-across-the-globe-a-look-at-eight-countries.
1296
+ Merchán, D., Arora, J., Pachon, J., Konduri, K., Winkenbach, M., Parks, S., Noszek, J., 2022. 2021 amazon
1297
+ last mile routing research challenge: Data set. Transportation Science .
1298
+ Mladenović, N., Todosijević, R., Urošević, D., 2012. An efficient GVNS for solving traveling salesman
1299
+ problem with time windows.
1300
+ Electronic Notes in Discrete Mathematics 39, 83–90.
1301
+ URL: https:
1302
+ //doi.org/10.1016/j.endm.2012.10.012, doi:10.1016/j.endm.2012.10.012.
1303
+ Pitney Bowes, 2020. Pitney Bowes Parcel Shipping Index. URL: https://www.pitneybowes.com/us/
1304
+ shipping-index.html.
1305
+ postnord, 2021.
1306
+ E-commerce in Europe 2020 - How the pandemic is changing e-commerce in
1307
+ Europe.
1308
+ Technical Report.
1309
+ URL: https://www.postnord.se/siteassets/pdf/rapporter/
1310
+ e-commerce-in-europe-2020.pdf.
1311
+ Purkayastha, R., Chakraborty, T., Saha, A., Mukhopadhyay, D., 2020.
1312
+ Study and analysis of vari-
1313
+ ous heuristic algorithms for solving travelling salesman problem—a survey, in: Advances in Intelli-
1314
+ gent Systems and Computing. Springer Singapore, pp. 61–70. URL: https://doi.org/10.1007/
1315
+ 978-981-15-2188-1_5, doi:10.1007/978-981-15-2188-1_5.
1316
+ Pytorch, 2021. Pytorch LSTM document. URL: https://pytorch.org/docs/stable/generated/
1317
+ torch.nn.LSTM.html.
1318
+ Ren, S., Choi, T.M., Lee, K.M., Lin, L., 2020. Intelligent service capacity allocation for cross-border-
1319
+ e-commerce related third-party-forwarding logistics operations:
1320
+ A deep learning approach.
1321
+ Trans-
1322
+ portation Research Part E: Logistics and Transportation Review 134, 101834.
1323
+ URL: https://
1324
+ www.sciencedirect.com/science/article/pii/S1366554519311688, doi:https://doi.org/
1325
+ 10.1016/j.tre.2019.101834.
1326
+ Rose, W.J., Mollenkopf, D.A., Autry, C., Bell, J., 2016. Exploring urban institutional pressures on logistics
1327
+ service providers. International Journal of Physical Distribution & Logistics Management 46. doi:10.
1328
+ 1108/09600035199500001.
1329
+ Salman, R., Ekstedt, F., Damaschke, P., 2020. Branch-and-bound for the precedence constrained generalized
1330
+ traveling salesman problem. Operations Research Letters 48, 163–166.
1331
+ da Silva, R.F., Urrutia, S., 2010. A general VNS heuristic for the traveling salesman problem with time
1332
+ windows. Discrete Optimization 7, 203–211. URL: https://doi.org/10.1016/j.disopt.2010.
1333
+ 04.002, doi:10.1016/j.disopt.2010.04.002.
1334
+ Snoeck, A., Winkenbach, M., 2021. A discrete simulation-based optimization algorithm for the design of
1335
+ highly responsive last-mile distribution networks. Transportation Science .
1336
+ Sutskever, I., Vinyals, O., Le, Q.V., 2014. Sequence to sequence learning with neural networks, in: Advances
1337
+ in Neural Information Processing Systems, pp. 3104–3112.
1338
+ Traub, V., Vygen, J., Zenklusen, R., 2021.
1339
+ Reducing path TSP to TSP.
1340
+ SIAM Journal on Com-
1341
+ puting , STOC20–24–STOC20–53URL: https://doi.org/10.1137/20m135594x, doi:10.1137/
1342
+ 20m135594x.
1343
+ United Nations Department of Economic and Social Affairs, 2019. World Urbanization Prospects: The
1344
+ 2018 Revision (ST/ESA/SER.A/420). Technical Report. New York: United Nations. URL: https:
1345
+ //population.un.org/wup/Publications/Files/WUP2018-Report.pdf.
1346
+ US Census Bureau, 2021. Quarterly e-commerce retail sales 2nd quarter 2021. Technical Report. U.S. Cen-
1347
+ sus Bureau of the Department of Commerce. URL: http://www2.census.gov/retail/releases/
1348
+ historical/ecomm/07q4.pdf.
1349
+ 24
1350
+
1351
+ Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.,
1352
+ 2017. Attention Is All You Need. Advances in Neural Information Processing Systems 2017-Decem,
1353
+ 5999–6009. URL: http://arxiv.org/abs/1706.03762, arXiv:1706.03762.
1354
+ Vinyals, O., Bengio, S., Kudlur, M., 2016. Order matters: Sequence to sequence for sets. 4th International
1355
+ Conference on Learning Representations, ICLR 2016 - Conference Track Proceedings , 1–11.
1356
+ Vinyals, O., Meire, F., Navdeep, J., 2015. Pointer Networks. Advances in Neural Information Processing
1357
+ Systems , 1–9.
1358
+ Wang, S., Cao, J., Chen, H., Peng, H., Huang, Z., 2020a. Seqst-gan: Seq2seq generative adversarial nets for
1359
+ multi-step urban crowd flow prediction 6. URL: https://doi.org/10.1145/3378889, doi:10.1145/
1360
+ 3378889.
1361
+ Wang, S., Mo, B., Zhao, J., 2020b. Deep neural networks for choice analysis: Architecture design with
1362
+ alternative-specific utility functions. Transportation Research Part C: Emerging Technologies 112, 234–
1363
+ 251. doi:10.1016/J.TRC.2020.01.012.
1364
+ Winkenbach, M., Parks, S., Noszek, J., 2021. Technical Proceedings of the Amazon Last Mile Routing
1365
+ Research Challenge URL: https://dspace.mit.edu/handle/1721.1/131235.
1366
+ World Economic Forum, 2020. The Future of the Last-Mile Ecosystem. Technical Report January. URL:
1367
+ https://www.weforum.org/reports/the-future-of-the-last-mile-ecosystem.
1368
+ Wu, R., Zhuang, Y., Xu, K., Zhang, H., Chen, B., 2020. Pq-net: A generative part seq2seq network for
1369
+ 3d shapes, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition
1370
+ (CVPR).
1371
+ Xu, Z., Wang, S., Zhu, F., Huang, J., 2017. Seq2seq fingerprint: An unsupervised deep molecular embedding
1372
+ for drug discovery, in: Proceedings of the 8th ACM International Conference on Bioinformatics, Compu-
1373
+ tational Biology,and Health Informatics, Association for Computing Machinery, New York, NY, USA. p.
1374
+ 285–294. URL: https://doi.org/10.1145/3107411.3107424, doi:10.1145/3107411.3107424.
1375
+ Yuan, Y., Cattaruzza, D., Ogier, M., Semet, F., 2020. A branch-and-cut algorithm for the generalized
1376
+ traveling salesman problem with time windows. European Journal of Operational Research 286, 849–866.
1377
+ Zhang, Y., Li, D., Wang, Y., Fang, Y., Xiao, W., 2019. Abstract text summarization with a convolutional
1378
+ seq2seq model. Applied Sciences 9. URL: https://www.mdpi.com/2076-3417/9/8/1665, doi:10.
1379
+ 3390/app9081665.
1380
+ Zhang, Y., Li, Y., Zhang, G., 2020. Short-term wind power forecasting approach based on seq2seq model
1381
+ using nwp data. Energy 213, 118371. URL: https://www.sciencedirect.com/science/article/
1382
+ pii/S036054422031478X, doi:https://doi.org/10.1016/j.energy.2020.118371.
1383
+ 25
1384
+
GdE2T4oBgHgl3EQfTQfv/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
JNFRT4oBgHgl3EQfzTgG/content/2301.13649v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4dc806adf6a3eadd0491a8526f72810ba95a7f7e2ac60441004fdbb939ed1099
3
+ size 668532
JNFRT4oBgHgl3EQfzTgG/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a150eb6ad68aca1de5c2d0c1db183900ac641c5d39dde0f4717c4123f5fdbb5f
3
+ size 1572909
JNFRT4oBgHgl3EQfzTgG/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5ceb262ac1cbe515803c7aad03a717f8c7c43cc573cdf8a0902e562e7a2a422a
3
+ size 57433
JtAzT4oBgHgl3EQfIPso/content/tmp_files/2301.01057v1.pdf.txt ADDED
@@ -0,0 +1,895 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ BS3D: Building-scale 3D Reconstruction from
2
+ RGB-D Images
3
+ Janne Mustaniemi1, Juho Kannala2, Esa Rahtu3,
4
+ Li Liu1, and Janne Heikkilä1
5
+ 1 Center for Machine Vision and Signal Analysis, University of Oulu, Finland
6
+ 2 Department of Computer Science, Aalto University, Finland
7
+ 3 Tampere University, Finland
8
+ janne.mustaniemi@oulu.fi
9
+ Abstract. Various datasets have been proposed for simultaneous local-
10
+ ization and mapping (SLAM) and related problems. Existing datasets
11
+ often include small environments, have incomplete ground truth, or lack
12
+ important sensor data, such as depth and infrared images. We propose
13
+ an easy-to-use framework for acquiring building-scale 3D reconstruction
14
+ using a consumer depth camera. Unlike complex and expensive acquisi-
15
+ tion setups, our system enables crowd-sourcing, which can greatly bene-
16
+ fit data-hungry algorithms. Compared to similar systems, we utilize raw
17
+ depth maps for odometry computation and loop closure refinement which
18
+ results in better reconstructions. We acquire a building-scale 3D dataset
19
+ (BS3D) and demonstrate its value by training an improved monocular
20
+ depth estimation model. As a unique experiment, we benchmark visual-
21
+ inertial odometry methods using both color and active infrared images.
22
+ Keywords: Depth camera · SLAM · Large-scale.
23
+ 1
24
+ Introduction
25
+ Simultaneous localization and mapping (SLAM) is an essential component in
26
+ robot navigation, virtual reality (VR), and augmented reality (AR) systems. Var-
27
+ ious datasets and benchmarks have been proposed for SLAM [11,35,39] and re-
28
+ lated problems, including visual-intertial odometry [30,6], camera re-localization
29
+ [29,32,15], and depth estimation [21,33]. Currently, there exists only a few building-
30
+ scale SLAM datasets [28] that include ground truth camera poses and dense 3D
31
+ geometry. Such datasets enable, for example, evaluation of algorithms needed in
32
+ large-scale AR applications.
33
+ The lack of building-scale SLAM datasets is explained by the difficulty of
34
+ acquiring ground truth data. Some have utilized a high-end LiDAR for obtaining
35
+ 3D geometry of the environment [26,2,4,28]. Ground truth camera poses may
36
+ be acquired using a motion capture (MoCap) system when the environment is
37
+ small enough [35,40]. The high cost of equipment, complex sensor setup, and
38
+ slow capturing process make these approaches less attractive and inconvenient
39
+ for crowd-sourced data collection.
40
+ arXiv:2301.01057v1 [cs.CV] 3 Jan 2023
41
+
42
+ 2
43
+ Mustaniemi et al.
44
+ An alternative is to perform 3D reconstruction using a monocular, stereo, or
45
+ depth camera. Consumer RGB-D cameras, in particular, are interesting because
46
+ of their relatively good accuracy, fast acquisition speed, low-cost, and effective-
47
+ ness in textureless environments. RGB-D cameras have been used to collect
48
+ datasets for depth estimation [21,33], scene understanding [8], and camera re-
49
+ localization [32,38], among other tasks. The problem is that existing RGB-D
50
+ reconstruction systems (e.g. [22,9,5]) are limited to room-scale and apartment-
51
+ scale environments.
52
+ Synthetic SLAM datasets have also been proposed [20,39,27] that include per-
53
+ fect ground truth. The challenge is that data such as time-of-flight (ToF) depth
54
+ maps and infrared images are difficult to synthesize realistically. Consequently,
55
+ training and evaluation done using synthetic data may not reflect algorithm’s
56
+ real-world performance. To address the domain gap problem, algorithms are
57
+ often fine-tuned using real data.
58
+ We propose a framework to create building-scale 3D reconstructions using a
59
+ consumer depth camera (Azure Kinect). Unlike existing approaches, we register
60
+ color images and depth maps using color-to-depth (C2D) strategy. This allows us
61
+ to directly utilize the raw depth maps captured by the wide field-of-view (FoV)
62
+ infrared camera. Coupled with an open-source SLAM library [19], we acquire a
63
+ building-scale 3D vision dataset (BS3D) that is considerably larger than similar
64
+ datasets as shown in Fig. 1. The BS3D dataset includes 392k synchronized color
65
+ images, depth maps and infrared images, inertial measurements, camera poses,
66
+ enhanced depth maps, surface reconstructions, and laser scans. Our framework
67
+ will be released for the public to enable fast, easy and affordable indoor 3D
68
+ reconstruction.
69
+ 240 m
70
+ 80 m
71
+ 8 m
72
+ Zoomed
73
+ Fig. 1. Building-scale 3D reconstruction (4300 m2) obtained using an RGB-D camera
74
+ and the proposed framework. The magnified area (90 m2) is larger than any recon-
75
+ struction in the ScanNet dataset [8].
76
+ 2
77
+ Related work
78
+ This section introduces commonly used RGB-D SLAM datasets and correspond-
79
+ ing data acquisition processes. A summary of the datasets is provided in Table
80
+
81
+ BS3D: Building-scale 3D Reconstruction from RGB-D Images
82
+ 3
83
+ 1. As there exist countless SLAM datasets, the scope is restricted to real-world
84
+ indoor scenarios. We leave out datasets focusing on aerial scenarios (e.g. Eu-
85
+ RoC MAV [2]) and autonomous driving (e.g. KITTI [11]). We also omit RGB-D
86
+ datasets captured with a stationary scanner (e.g. Matterport3D [4]) as they can-
87
+ not be used for SLAM evaluation. Synthetic datasets, such as SceneNet RGB-D
88
+ [20], TartanAir [39], and ICL [27] are also omitted.
89
+ ADVIO [6] dataset is a realistic visual-inertial odometry benchmark that in-
90
+ cludes building-scale environments. Ground truth trajectory is computed using
91
+ an inertial navigation system (INS) together with manual location fixes. The
92
+ main limitation of the dataset is that it does not come with ground truth 3D
93
+ geometry. LaMAR [28] is a large-scale SLAM benchmark that utilizes high-end
94
+ mapping platforms (NavVis M6 trolley and VLX backpack) for ground truth
95
+ generation. Although the capturing setup includes a variety of devices (e.g.
96
+ HoloLens2 and iPad Pro), it does not include a dedicated RGB-D camera.
97
+ OpenLORIS-Scene [31] focuses on the lifelong SLAM scenario where environ-
98
+ ments are dynamic and changing, similar to LaMAR [28]. The data is collected
99
+ over an extended period of time using wheeled robots equipped with various
100
+ sensors, including RGB-D, stereo, IMU, wheel odometry, and LiDAR. Ground
101
+ truth poses are acquired using an external motion capture (MoCap) system, or
102
+ with a 2D laser SLAM method. The dataset is not ideal for handheld SLAM
103
+ evaluation because of the limited motion patterns of a ground robot.
104
+ TUM RGB-D SLAM [35] is one of the most popular SLAM datasets. The
105
+ RGB-D images are acquired using a consumer depth camera Microsoft Kinect
106
+ v1. Ground truth trajectory is incomplete because the MoCap system can only
107
+ cover a small part of the environment. CoRBS [40] consists of four room-scale
108
+ environments. It also utilizes MoCap for acquiring ground truth trajectories for
109
+ Microsoft Kinect v2. Unlike [35], CoRBS provides ground truth 3D geometry
110
+ acquired using a laser scanner. The data also includes infrared images, but not
111
+ inertial measurements, unlike our dataset.
112
+ 7-Scenes [32] and 12-Scenes [38] are commonly used for evaluating camera lo-
113
+ calization. 7-Scenes includes seven scenes captured using Kinect v1. KinectFusion
114
+ [22] is used to obtain ground truth poses and dense 3D models from the RGB-
115
+ D images. 12-Scenes consists of multiple rooms captured using the Structure.io
116
+ depth sensor and iPad color camera. The reconstructions are larger compared
117
+ to 7-Scenes, about 37 m3 on average. They are acquired using the VoxelHashing
118
+ framework [23], an alternative to KinectFusion with better scalability.
119
+ ScanNet [8] is an RGB-D dataset containing 2.5M views acquired in 707
120
+ distinct spaces. It includes estimated calibration parameters, camera poses, 3D
121
+ surface reconstructions, textured meshes, and object-level semantic segmenta-
122
+ tions. The hardware consists of a Structure.io depth sensor attached to a tablet
123
+ computer. Pose estimation is done using BundleFusion [9], after which volumet-
124
+ ric integration is performed through VoxelHashing [23].
125
+ Sun3D [43] is a large RGB-D database with camera poses, point clouds,
126
+ object labels, and refined depth maps. The reconstruction process is based on
127
+ structure from motion (SfM) where manual object annotations are utilized to
128
+
129
+ 4
130
+ Mustaniemi et al.
131
+ reduce drift and loop-closure failures. Refined depth maps are obtained via vol-
132
+ umetric fusion similar to KinectFusion [22]. We emphasize that ScanNet [8] and
133
+ Sun3D [43] reconstructions are considerably smaller and have lower quality than
134
+ those provided in our dataset. Unlike [28,31,35], our system also does not require
135
+ a complex and expensive capturing setup, or manual annotation [6,43].
136
+ Table 1. List of indoor RGB-D SLAM datasets. The BS3D acquisition setup does
137
+ not require high-end LiDARs [40,31,28], MoCap systems [40,31,36], or manual annota-
138
+ tion [43,6]. BS3D is building-scale, unlike [32,36,8,40,38,43]. Note that ADVIO [6] and
139
+ LaMAR [28] do not have a dedicated depth camera.
140
+ Dataset
141
+ Scale
142
+ Depth
143
+ IMU
144
+ IR
145
+ Ground truth
146
+ 7-Scenes [32]
147
+ room
148
+ Kinect v1
149
+ -
150
+ -
151
+ RGBD-recons.
152
+ TUM RGBD [36]
153
+ room
154
+ Kinect v1
155
+
156
+ -
157
+ MoCap
158
+ ScanNet [8]
159
+ room
160
+ Structure.io
161
+
162
+ -
163
+ RGBD-recons.
164
+ CoRBS [40]
165
+ room
166
+ Kinect v2
167
+ -
168
+
169
+ MoCap+LiDAR
170
+ 12-Scenes [38]
171
+ apartment
172
+ Structure.io
173
+ -
174
+ -
175
+ RGBD-recons.
176
+ Sun3D [43]
177
+ apartment
178
+ Xtion Pro Live
179
+ -
180
+ -
181
+ RGBD+manual
182
+ OpenLORIS [31]
183
+ building
184
+ RS-D435i
185
+
186
+ -
187
+ MoCap+LiDAR
188
+ ADVIO [6]
189
+ building
190
+ Tango
191
+
192
+ -
193
+ INS+manual
194
+ LaMAR [28]
195
+ building
196
+ HoloLens2
197
+
198
+
199
+ LiDAR+VIO+SfM
200
+ BS3D (ours)
201
+ building
202
+ Azure Kinect
203
+
204
+
205
+ RGBD-recons.
206
+ 3
207
+ Reconstruction framework
208
+ In this section, we introduce the RGB-D reconstruction framework shown in
209
+ Fig. 2. The framework produces accurate 3D reconstructions of building-scale
210
+ environments using low-cost hardware. The system is fully automatic and robust
211
+ against poor lighting conditions and fast motions. Color images are only used for
212
+ loop closure detection as they are susceptible to motion blur and rolling shutter
213
+ distortion. Raw depth maps enable accurate odometry and the refinement of
214
+ loop closure transformations.
215
+ 3.1
216
+ Hardware
217
+ Data is captured using the Azure Kinect depth camera, which is well-suited for
218
+ crowd-sourcing due to its popularity and affordability. We capture synchronized
219
+ depth, color, and infrared images at 30 Hz using the official recorder application
220
+ running on a laptop computer. We use the wide FoV mode of the infrared camera
221
+ with 2x2 binning to extend the Z-range. The resolution of raw depth maps and
222
+ IR images is 512 x 512 pixels. Auto-exposure is enabled when capturing color
223
+ images at the resolution of 720 x 1280 pixels. We also record accelerometer and
224
+ gyroscope readings at 1.6 kHz.
225
+
226
+ BS3D: Building-scale 3D Reconstruction from RGB-D Images
227
+ 5
228
+ Fig. 2. Overview of the RGB-D reconstruction system.
229
+ .
230
+ 3.2
231
+ Color-to-depth alignment
232
+ Most RGB-D reconstruction systems expect that color images and depth maps
233
+ have been spatially and temporally aligned. Modern depth cameras typically
234
+ produce temporally synchronized images so the main concern is the spatial align-
235
+ ment. Conventionally, raw depth maps are transformed to the coordinate system
236
+ of the color camera, which we refer to as the depth-to-color (D2C) alignment.
237
+ In the case of Azure Kinect, the color camera’s FoV is much narrower (90
238
+ x 59 degrees) compared to the infrared camera (120 x 120 degrees). Thus, the
239
+ D2C alignment would not take advantage of the infrared camera’s wide FoV
240
+ because depth maps would be heavily cropped. Moreover, the D2C alignment
241
+ might introduce artefacts to the raw depth maps.
242
+ We propose an alternative called color-to-depth (C2D) alignment where color
243
+ images are transformed instead. In the experiments, we show that this drastically
244
+ improves the quality of the reconstructions. The main challenge of C2D is that
245
+ it requires a fully dense depth map. Fortunately, a reasonably good alignment
246
+ can be achieved even with a low quality depth map. This is because the baseline
247
+ between the cameras is narrow and missing depths often appear in areas that
248
+ are far away from the camera.
249
+ For the C2D alignment, we first perform depth inpainting using linear in-
250
+ terpolation. Then, the color image is transformed to the raw depth frame. To
251
+ keep as much of the color information as possible, the output resolution will be
252
+ higher (1024 x 1024 pixels) compared to the raw depth maps . After that, holes
253
+ in the color image due to occlusions are inpainted using the OpenCV library’s
254
+ implementation of [37]. We note that minor artefacts in the aligned color images
255
+ will have little impact on the SIFT-based loop closure detection.
256
+ 3.3
257
+ RGB-D Mapping
258
+ We process the RGB-D sequences using an open-source SLAM library called
259
+ RTAB-Map [19]. Odometry is computed from the raw depth maps using the
260
+ point-to-plane variant of the iterative closest point (ICP) algorithm [25]. We use
261
+ the scan-to-map odometry strategy [19] where incoming frames are registered
262
+ against a point cloud map created from past keyframes. The wide FoV ensures
263
+ that ICP-odometry rarely fails, but in case it does, a new map is initialized.
264
+
265
+ RGBD
266
+ RGB
267
+ Depth
268
+ RGBD
269
+ Color-to-depth
270
+ Loop closures
271
+ Volumetric
272
+ (C2D)
273
+ (PnP + ICP)
274
+ fusion
275
+ Poses
276
+ Normals
277
+ Depth (raw)
278
+ Mesh
279
+ (optimized)
280
+ Poses
281
+ Odometry
282
+ (odometry)
283
+ Graph
284
+ Render
285
+ (ICP)
286
+ optimization6
287
+ Mustaniemi et al.
288
+ Loop closure detection is needed for drift correction and merging of individual
289
+ maps. For this purpose, SIFT features are extracted from the valid area of the
290
+ aligned color images. Loop closures are detected using the bag-of-words approach
291
+ [18], and transformations are estimated using the Perspective-n-Point RANSAC
292
+ algorithm and refined using ICP [25]. Graph optimization is done using the
293
+ GTSAM library [10] and Gauss-Newton algorithm.
294
+ RTAB-Map supports multi-session mapping which is a necessary feature
295
+ when reconstructing building-scale environments. It is not practical to collect
296
+ possibly hours of data at once. Furthermore, having the ability to later update
297
+ and expand the map is a useful feature. In practise, individual sequences are
298
+ first processed separately, followed by multi-session mapping. The sessions are
299
+ merged by finding loop closures and by performing graph optimization. The in-
300
+ put is a sequence of keyframes along with odometry poses and SIFT features
301
+ computed during single-session mapping. The sessions are processed in such or-
302
+ der that there is at least some overlap between the current session and the global
303
+ map build so far.
304
+ 3.4
305
+ Surface reconstruction
306
+ It is often useful to have a 3D surface representation of the environment. There
307
+ exists many classical [14,22] and learning-based [41,1] surface reconstruction ap-
308
+ proaches. Methods that utilize deep neural networks, such as NeuralFusion [41],
309
+ have produced impressive results on the task of depth map fusion. Neural ra-
310
+ diance fields (NeRFs) have also been adapted to RGB-D imagery [1] showing
311
+ good performance. We did not use learning-based approaches in this work be-
312
+ cause they are limited to small scenes, at least for the time being. Moreover,
313
+ scene-specific learning [1] takes several hours even with powerful hardware.
314
+ Surface reconstruction is done in segments due to the large scale of the en-
315
+ vironment and the vast number of frames. To that end, we first create a point
316
+ cloud from downsampled raw depth maps. Every point includes a view index
317
+ along with 3D coordinates. The point cloud is partitioned into manageable seg-
318
+ ments using the K-means algorithm. A mesh is created for each segment using
319
+ the scalable TSDF fusion implementation [46] that is based on [7,22]. It uses a
320
+ hierarchical hashing structure to support large scenes.
321
+ 4
322
+ BS3D dataset
323
+ The BS3D dataset was collected at the university campus using Azure Kinect
324
+ (Section 3.1). Figure 3 shows example frames from the dataset. The collection
325
+ was done in multiple sessions due to large scale of the environment. The record-
326
+ ings were processed using the framework described in Section 3.
327
+ 4.1
328
+ Dataset features
329
+ The reconstruction shown in Fig. 1 consists of 47 overlapping recording sessions.
330
+ Additional 14 sessions, including 3D laser scans, were recorded for evaluation
331
+
332
+ BS3D: Building-scale 3D Reconstruction from RGB-D Images
333
+ 7
334
+ Cafeteria
335
+ Stairs
336
+ Study
337
+ Corridor
338
+ Lobby
339
+ Fig. 3. Example frames from the dataset. Environments are diverse and challenging,
340
+ including cafeterias, stairs, study areas, corridors, and lobbies.
341
+ purposes. Most sessions begin and end at the same location to encourage loop
342
+ closure detection. The total duration of the sessions is 3 hours and 38 minutes
343
+ and the combined trajectory length is 6.4 kilometers. The reconstructed floor
344
+ area is approximately 4300 m2.
345
+ The dataset consists of 392k frames, including color images, raw depth maps,
346
+ and infrared images. Color images and depth maps are provided in both coordi-
347
+ nate frames (color and infrared camera). The images have been undistorted for
348
+ convenience, but the original recordings are also included. We provide camera
349
+ poses in a global reference frame for every image. Data also includes inertial mea-
350
+ surements, enhanced depth maps and surface normals that have been rendered
351
+ from the mesh as visualized in Fig. 4.
352
+ Color
353
+ Infrared
354
+ Normals (render)
355
+ Mesh
356
+ Depth
357
+ Depth (raw)
358
+ Depth (render)
359
+ Fig. 4. The BS3D dataset includes color and infrared images, depth maps, IMU data,
360
+ camera parameters, and surface reconstructions. Enhanced depth maps and surface
361
+ normals are rendered from the mesh.
362
+ 4.2
363
+ Laser scan
364
+ We utilize the FARO 3D X 130 laser scanner for acquiring ground truth 3D
365
+ geometry. The scanned area includes a lobby and corridors of different sizes (800
366
+ m2). The 28 individual scans were registered using the SCENE software that
367
+ comes with the laser scanner. Noticeable artefacts, e.g. those caused by mirrors,
368
+
369
+ 8
370
+ Mustaniemi et al.
371
+ were manually removed. The laser scan is used to evaluate the reconstruction
372
+ framework in Section 5. However, this data also enables, for example, training
373
+ and evaluation of RGB-D surface reconstruction algorithms.
374
+ 5
375
+ Experiments
376
+ We compare our framework with the state-of-the-art RGB-D reconstruction
377
+ methods [5,9,3]. The value of the BS3D dataset is demonstrated by training
378
+ a recent monocular depth estimation model [44]. We also benchmark visual-
379
+ inertial odometry approaches [12,34,3] using either color or infrared images to
380
+ further highlight the unique aspects of the BS3D dataset.
381
+ 5.1
382
+ Reconstruction framework
383
+ In this experiment, we compare the framework against Redwood [5], Bundle-
384
+ Fusion [9], and ORB-SLAM3 [3]. RGBD images are provided for [5,9,3] in the
385
+ coordinate frame of the color camera. Given the estimated camera poses, we cre-
386
+ ate a point cloud and compare it to the laser scan (Section 4.2). The evaluation
387
+ metrics include overlap of the point clouds and RMSE of inlier correspondences.
388
+ Before comparison, we create uniformly sampled point clouds using voxel down-
389
+ sampling (1 cm3 voxel) that computes the centroid of the points in each voxel.
390
+ The overlap is defined as the ratio of inlier correspondences and the number of
391
+ ground truth points. A 3D point is considered to be an inlier if the distance to
392
+ the closest ground truth point is below threshold γ.
393
+ Table 2 shows the results for environments of different sizes. All methods
394
+ are able to reconstruct the small environment (35 m2) consisting of 2.8k frames.
395
+ The differences between the methods become more evident when reconstructing
396
+ the medium-size environment (160 m2) consisting of 7.3k frames. BundleFusion
397
+ [9] only produces a partial reconstruction because of odometry failures. The
398
+ proposed approach gives the most accurate reconstructions as visualized in Fig.
399
+ 5. Note that it is not possible to achieve 100 % overlap because the depth camera
400
+ does not observe all parts of the ground truth.
401
+ The largest environment (350 m2) consists of 24k frames acquired in four
402
+ sessions. Redwood [5] does not scale to input sequences of this long. ORB-SLAM3
403
+ [3] frequently loses the odometry in open spaces which leads to incomplete and
404
+ less accurate reconstructions. Our method suffers the same problem when C2D
405
+ is disabled. Unreliable odometry is likely due to the color camera’s limited FoV,
406
+ rolling shutter distortion, and motion blur. The C2D alignment improves the
407
+ accuracy and robustness of ICP-based odometry and loop closures. Without
408
+ C2D, the frequent odometry failures result in disconnected maps and noticeable
409
+ drift. We note that the reconstruction in Fig. 1 was computed from ∼300k frames
410
+ which is far more than [5,9,3] can handle.
411
+
412
+ BS3D: Building-scale 3D Reconstruction from RGB-D Images
413
+ 9
414
+ Table 2. Comparison of RGB-D reconstruction methods in small, medium and large-
415
+ scale environments (from top to bottom). Overlap of the point clouds and inlier RMSE
416
+ computed for distance thresholds γ (mm). Some methods only work in small and/or
417
+ medium scale environments.
418
+ γ = 10 (mm)
419
+ γ = 20 (mm)
420
+ γ = 50 (mm)
421
+ Method
422
+ Overlap ↑
423
+ RMSE ↓
424
+ Overlap ↑
425
+ RMSE ↓
426
+ Overlap ↑
427
+ RMSE ↓
428
+ Redwood [5]
429
+ 66.5
430
+ 5.6
431
+ 77.9
432
+ 7.6
433
+ 87.1
434
+ 12.6
435
+ BundleFusion [9]
436
+ 72.1
437
+ 5.5
438
+ 80.8
439
+ 6.9
440
+ 88.3
441
+ 11.7
442
+ ORB-SLAM3 [3]
443
+ 78.2
444
+ 5.3
445
+ 85.2
446
+ 6.5
447
+ 91.3
448
+ 10.6
449
+ Prop. (w/o C2D)
450
+ 66.8
451
+ 5.7
452
+ 77.8
453
+ 7.5
454
+ 87.0
455
+ 12.7
456
+ Proposed
457
+ 78.4
458
+ 5.2
459
+ 85.7
460
+ 6.5
461
+ 91.6
462
+ 10.6
463
+ Redwood [5]
464
+ 30.4
465
+ 6.2
466
+ 44.5
467
+ 9.8
468
+ 63.9
469
+ 19.9
470
+ BundleFusion [9]
471
+ 8.1
472
+ 6.2
473
+ 11.1
474
+ 9.2
475
+ 14.8
476
+ 18.8
477
+ ORB-SLAM3 [3]
478
+ 44.3
479
+ 6.0
480
+ 57.7
481
+ 8.7
482
+ 71.0
483
+ 16.2
484
+ Prop. (w/o C2D)
485
+ 36.5
486
+ 6.1
487
+ 49.2
488
+ 9.0
489
+ 64.3
490
+ 18.3
491
+ Proposed
492
+ 54.1
493
+ 5.7
494
+ 64.8
495
+ 7.7
496
+ 73.2
497
+ 13.4
498
+ ORB-SLAM3 [3]
499
+ 9.5
500
+ 6.3
501
+ 14.4
502
+ 9.9
503
+ 20.8
504
+ 20.7
505
+ Prop. (w/o C2D)
506
+ 23.1
507
+ 6.7
508
+ 40.6
509
+ 10.9
510
+ 64.7
511
+ 22.4
512
+ Proposed
513
+ 34.7
514
+ 6.4
515
+ 52.7
516
+ 10.0
517
+ 75.0
518
+ 19.8
519
+ ORB-SLAM3 [3]
520
+ Proposed
521
+ Redwood [5]
522
+ Proposed (w/o C2D)
523
+ ϵ < 20 mm
524
+ 20 ≤ ϵ < 50
525
+ 50 ≤ ϵ < 100
526
+ 100 ≤ ϵ < 200
527
+ ϵ ≥ 200 mm
528
+ Fig. 5. Reconstructions obtained using Redwood [5], ORB-SLAM3 [3], and the pro-
529
+ posed method. Colors depict errors (distance to the closest ground truth point).
530
+
531
+ 10
532
+ Mustaniemi et al.
533
+ 5.2
534
+ Depth estimation
535
+ We investigate whether the BS3D dataset can be used to train better models
536
+ for monocular depth estimation. For this experiment, we use the state-of-the-
537
+ art LeReS model [44] based on ResNet50. The original model has been trained
538
+ using 354k samples taken from various datasets [45,24,16,13,42]. We finetune
539
+ the model using 16.5k samples from BS3D. We set the learning rate to 2e-5 and
540
+ train only 4 epochs to avoid overfitting. Other training details, including loss
541
+ functions are the same as in [44].
542
+ For testing, we use NYUD-v2 [21] and iBims-1 [17] that are not seen during
543
+ training. We also evaluate using BS3D by sampling 535 images from an unseen
544
+ part of the building. Table 3 shows that finetuning improves the performance on
545
+ iBims-1 and BS3D. The finetuned model performs marginally worse on NYUD-
546
+ v2 which is not surprising considering that NYUD-v2 mainly contains room-scale
547
+ scenes that are not present in BS3D. The qualitative comparison in Fig. 6 also
548
+ shows a clear improvement over the pretrained model on iBims-1 that contains
549
+ both small and large scenes. The model trained only using BS3D cannot compete
550
+ with other models, except on BS3D on which the performance is surprisingly
551
+ good. The poor performance on other datasets is not surprising because of the
552
+ small training set.
553
+ Table 3. Monocular depth estimation using LeReS [44] trained from scratch using
554
+ BS3D, pretrained model, and finetuned model. NUYD-v2 [21], iBims-1 [17], and BS3D
555
+ are used for testing.
556
+ NYUD-v2 [21]
557
+ iBims-1 [17]
558
+ BS3D
559
+ Training data
560
+ AbsRel ↓
561
+ δ1 ↑
562
+ AbsRel ↓
563
+ δ1 ↑
564
+ AbsRel ↓
565
+ δ1 ↑
566
+ BS3D
567
+ 0.181
568
+ 0.764
569
+ 0.188
570
+ 0.763
571
+ 0.144
572
+ 0.828
573
+ Pretrained
574
+ 0.096
575
+ 0.913
576
+ 0.115
577
+ 0.890
578
+ 0.157
579
+ 0.785
580
+ Pre. + BS3D
581
+ 0.100
582
+ 0.907
583
+ 0.098
584
+ 0.901
585
+ 0.115
586
+ 0.881
587
+ Color
588
+ Pretrained
589
+ Finetuned
590
+ Ground truth
591
+ Fig. 6. Comparison of pretrained and finetuned (BS3D) monocular depth estimation
592
+ model LeReS [44] on an independent iBims-1 [17] dataset unseen during training.
593
+
594
+ BS3D: Building-scale 3D Reconstruction from RGB-D Images
595
+ 11
596
+ 5.3
597
+ Visual-inertial odometry
598
+ The BS3D dataset includes active infrared images along with color and IMU
599
+ data. This opens interesting possibilities, for example, the comparison of color
600
+ and infrared as inputs for visual-inertial odometry. Infrared-inertial odometry
601
+ is an attractive approach in the sense that it does not require external light,
602
+ meaning it would work in completely dark environments.
603
+ We evaluate OpenVINS [12], ORB-SLAM3 [3], and DM-VIO [34] using color-
604
+ inertial and infrared-inertial inputs. Note that ORB-SLAM3 has an unfair ad-
605
+ vantage because it has a loop closure detector that cannot be disabled. In the
606
+ case of infrared images, we apply a power law transformation (I = 0.04 · I0.6)
607
+ to increase brightness. As supported by [34], we provide a mask of valid pix-
608
+ els to ignore black areas near the edges of the infrared images. We adjust the
609
+ parameters related to feature detection when using infrared images with [12,3].
610
+ We use the standard error metrics, namely absolute trajectory error (ATE) and
611
+ relative pose error (RPE) which measures the drift per second. The methods are
612
+ evaluated 5 times on each of the 10 sequences (Table 4).
613
+ From the results in Table 5, we can see that ORB-SLAM3 has the lowest
614
+ ATE when evaluating color-inertial odometry, mainly because of loop closure
615
+ detection. In most cases, ORB-SLAM3 and OpenVINS fail to initialize when
616
+ using infrared images. We conclude that off-the-shelve feature detectors (FAST
617
+ and ORB) are quite poor at detecting good features from infrared images. Inter-
618
+ estingly, DM-VIO performs better when using infrared images instead of color
619
+ which is likely due to the infrared camera’s global shutter and wider FoV. This
620
+ result reveals the great potential of using active infrared images for visual-inertial
621
+ odometry and the need for new research.
622
+ Table 4. Evaluation sequences used in the visual-inertial odometry experiment. Last
623
+ column shows if the camera returns to the starting point (chance for a loop closure).
624
+ Sequence
625
+ Duration (s)
626
+ Length (m)
627
+ Dimensions (m)
628
+ Loop
629
+ cafeteria
630
+ 200
631
+ 90.0
632
+ 12.4 x 15.7 x 0.8
633
+
634
+ central
635
+ 242
636
+ 155.0
637
+ 25.5 x 42.1 x 5.3
638
+
639
+ dining
640
+ 192
641
+ 109.2
642
+ 33.8 x 25.0 x 5.5
643
+
644
+ corridor
645
+ 174
646
+ 77.6
647
+ 31.1 x 4.7 x 2.4
648
+
649
+ foobar
650
+ 75
651
+ 37.1
652
+ 5.4 x 14.4 x 0.6
653
+
654
+ hub
655
+ 124
656
+ 52.3
657
+ 11.4 x 5.9 x 0.7
658
+ -
659
+ juice
660
+ 103
661
+ 42.7
662
+ 6.3 x 8.6 x 0.5
663
+ -
664
+ lounge
665
+ 222
666
+ 94.2
667
+ 14.4 x 10.3 x 1.1
668
+
669
+ study
670
+ 87
671
+ 40.0
672
+ 5.6 x 9.8 x 0.6
673
+ -
674
+ waiting
675
+ 139
676
+ 60.1
677
+ 9.8 x 6.7 x 0.9
678
+
679
+
680
+ 12
681
+ Mustaniemi et al.
682
+ Table 5. Comparison of visual-inertial odometry methods using color-inertial and
683
+ infrared-inertial inputs. Average absolute trajectory error (ATE) and relative pose error
684
+ (RPE). Last column shows the percentage of successful runs.
685
+ Color-inertial odometry
686
+ Infrared-inertial odometry
687
+ Method
688
+ ATE ↓
689
+ (m)
690
+ RPE ↓
691
+ (deg/s)
692
+ RPE ↓
693
+ (m/s)
694
+ Succ. ↑
695
+ (%)
696
+ ATE ↓
697
+ (m)
698
+ RPE ↓
699
+ (deg/s)
700
+ RPE ↓
701
+ (m/s)
702
+ Succ. ↑
703
+ (%)
704
+ OpenVINS [12]
705
+ 0.347
706
+ 0.37
707
+ 0.031
708
+ 76.0
709
+ 0.597
710
+ 0.42
711
+ 0.057
712
+ 36.0
713
+ ORB-SLAM3 [3]
714
+ 0.298
715
+ 0.29
716
+ 0.026
717
+ 100.0
718
+ 0.193
719
+ 0.29
720
+ 0.025
721
+ 24.0
722
+ DM-VIO [34]
723
+ 0.491
724
+ 0.29
725
+ 0.033
726
+ 100.0
727
+ 0.433
728
+ 0.29
729
+ 0.025
730
+ 100.0
731
+ 6
732
+ Conclusion
733
+ We presented a framework for acquiring high-quality 3D reconstructions using
734
+ a consumer depth camera. The ability to produce building-scale reconstructions
735
+ is a significant improvement over existing methods that are limited to smaller
736
+ environments such as rooms or apartments. The proposed C2D alignment en-
737
+ ables the use of raw depth maps, resulting in more accurate 3D reconstructions.
738
+ Our approach is fast, easy to use, and requires no expensive hardware, making
739
+ it ideal for crowd-sourced data collection. We acquire building-scale 3D dataset
740
+ (BS3D) and demonstrate its value for monocular depth estimation. BS3D is
741
+ unique also because it includes active infrared images, which are often miss-
742
+ ing in other datasets. We employ infrared images for visual-inertial odometry,
743
+ discovering a promising new research direction.
744
+ References
745
+ 1. Azinović, D., Martin-Brualla, R., Goldman, D.B., Nießner, M., Thies, J.: Neural
746
+ RGB-D surface reconstruction. In: Conference on Computer Vision and Pattern
747
+ Recognition (CVPR). pp. 6290–6301 (2022)
748
+ 2. Burri, M., Nikolic, J., Gohl, P., Schneider, T., Rehder, J., Omari, S., Achtelik,
749
+ M.W., Siegwart, R.: The EuRoC micro aerial vehicle datasets. The International
750
+ Journal of Robotics Research 35(10), 1157–1163 (2016)
751
+ 3. Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J.M., Tardós, J.D.: ORB-
752
+ SLAM3: An accurate open-source library for visual, visual-inertial, and multimap
753
+ SLAM. IEEE Transactions on Robotics 37(6), 1874–1890 (2021)
754
+ 4. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song,
755
+ S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor
756
+ environments. arXiv preprint arXiv:1709.06158 (2017)
757
+ 5. Choi, S., Zhou, Q.Y., Koltun, V.: Robust reconstruction of indoor scenes. In: IEEE
758
+ Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5556–5565
759
+ (2015)
760
+ 6. Cortés, S., Solin, A., Rahtu, E., Kannala, J.: ADVIO: An authentic dataset for
761
+ visual-inertial odometry. In: European Conference on Computer Vision (ECCV).
762
+ pp. 419–434 (2018)
763
+
764
+ BS3D: Building-scale 3D Reconstruction from RGB-D Images
765
+ 13
766
+ 7. Curless, B., Levoy, M.: A volumetric method for building complex models from
767
+ range images. In: Conference on Computer Graphics and Interactive Techniques.
768
+ pp. 303–312 (1996)
769
+ 8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scan-
770
+ Net: Richly-annotated 3D reconstructions of indoor scenes. In: IEEE Conference
771
+ on Computer Vision and Pattern Recognition (CVPR). pp. 5828–5839 (2017)
772
+ 9. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: BundleFusion: Real-
773
+ time globally consistent 3D reconstruction using on-the-fly surface reintegration.
774
+ ACM Transactions on Graphics (ToG) 36(4), 1 (2017)
775
+ 10. Dellaert, F.: Factor graphs and GTSAM: A hands-on introduction. Tech. rep.,
776
+ Georgia Institute of Technology (2012)
777
+ 11. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI
778
+ dataset. The International Journal of Robotics Research 32(11), 1231–1237 (2013)
779
+ 12. Geneva, P., Eckenhoff, K., Lee, W., Yang, Y., Huang, G.: OpenVINS: A research
780
+ platform for visual-inertial estimation. In: International Conference on Robotics
781
+ and Automation (ICRA). pp. 4666–4672. IEEE (2020)
782
+ 13. Hua, Y., Kohli, P., Uplavikar, P., Ravi, A., Gunaseelan, S., Orozco, J., Li,
783
+ E.: Holopix50k: A large-scale in-the-wild stereo image dataset. arXiv preprint
784
+ arXiv:2003.11172 (2020)
785
+ 14. Kazhdan, M., Bolitho, M., Hoppe, H.: Poisson surface reconstruction. In: Pro-
786
+ ceedings of the Fourth Eurographics Symposium on Geometry Processing. vol. 7
787
+ (2006)
788
+ 15. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A convolutional network for real-
789
+ time 6-dof camera relocalization. In: IEEE International Conference on Computer
790
+ Vision (ICCV). pp. 2938–2946 (2015)
791
+ 16. Kim, Y., Jung, H., Min, D., Sohn, K.: Deep monocular depth estimation via in-
792
+ tegration of global and local predictions. IEEE Transactions on Image Processing
793
+ 27(8), 4131–4144 (2018)
794
+ 17. Koch, T., Liebel, L., Fraundorfer, F., Korner, M.: Evaluation of CNN-based single-
795
+ image depth estimation methods. In: European Conference on Computer Vision
796
+ (ECCV) Workshops. pp. 0–0 (2018)
797
+ 18. Labbe, M., Michaud, F.: Appearance-based loop closure detection for online large-
798
+ scale and long-term operation. IEEE Transactions on Robotics 29(3), 734–745
799
+ (2013)
800
+ 19. Labbé, M., Michaud, F.: RTAB-Map as an open-source lidar and visual simultane-
801
+ ous localization and mapping library for large-scale and long-term online operation.
802
+ Journal of Field Robotics 36(2), 416–446 (2019)
803
+ 20. McCormac, J., Handa, A., Leutenegger, S., Davison, A.J.: SceneNet RGB-D: Can
804
+ 5m synthetic images beat generic ImageNet pre-training on indoor segmentation?
805
+ In: IEEE International Conference on Computer Vision (ICCV). pp. 2678–2687
806
+ (2017)
807
+ 21. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and sup-
808
+ port inference from RGBD images. In: European Conference on Computer Vision
809
+ (ECCV) (2012)
810
+ 22. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J.,
811
+ Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense
812
+ surface mapping and tracking. In: IEEE International Symposium on Mixed and
813
+ Augmented Reality. pp. 127–136. IEEE (2011)
814
+ 23. Nießner, M., Zollhöfer, M., Izadi, S., Stamminger, M.: Real-time 3D reconstruction
815
+ at scale using voxel hashing. ACM Transactions on Graphics (ToG) 32(6), 1–11
816
+ (2013)
817
+
818
+ 14
819
+ Mustaniemi et al.
820
+ 24. Niklaus, S., Mai, L., Yang, J., Liu, F.: 3D Ken Burns effect from a single image.
821
+ ACM Transactions on Graphics (ToG) 38(6), 1–15 (2019)
822
+ 25. Pomerleau, F., Colas, F., Siegwart, R., Magnenat, S.: Comparing ICP variants on
823
+ real-world data sets. Autonomous Robots 34(3), 133–148 (Feb 2013)
824
+ 26. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A.,
825
+ Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.:
826
+ Habitat-Matterport 3D Dataset (HM3D): 1000 large-scale 3D environments for
827
+ embodied AI. arXiv preprint arXiv:2109.08238 (2021)
828
+ 27. Saeedi, S., Carvalho, E.D., Li, W., Tzoumanikas, D., Leutenegger, S., Kelly, P.H.,
829
+ Davison, A.J.: Characterizing visual localization and mapping datasets. In: Inter-
830
+ national Conference on Robotics and Automation (ICRA). pp. 6699–6705. IEEE
831
+ (2019)
832
+ 28. Sarlin, P.E., Dusmanu, M., Schönberger, J.L., Speciale, P., Gruber, L., Larsson,
833
+ V., Miksik, O., Pollefeys, M.: LaMAR: Benchmarking localization and mapping
834
+ for augmented reality. In: European Conference on Computer Vision (ECCV). pp.
835
+ 686–704. Springer (2022)
836
+ 29. Sattler, T., Maddern, W., Toft, C., Torii, A., Hammarstrand, L., Stenborg, E.,
837
+ Safari, D., Okutomi, M., Pollefeys, M., Sivic, J., et al.: Benchmarking 6dof out-
838
+ door visual localization in changing conditions. In: IEEE Conference on Computer
839
+ Vision and Pattern Recognition (CVPR). pp. 8601–8610 (2018)
840
+ 30. Schubert, D., Goll, T., Demmel, N., Usenko, V., Stückler, J., Cremers, D.: The
841
+ TUM VI benchmark for evaluating visual-inertial odometry. In: IEEE/RSJ Inter-
842
+ national Conference on Intelligent Robots and Systems (IROS). pp. 1680–1687.
843
+ IEEE (2018)
844
+ 31. Shi, X., Li, D., Zhao, P., Tian, Q., Tian, Y., Long, Q., Zhu, C., Song, J., Qiao, F.,
845
+ Song, L., et al.: Are we ready for service robots? the OpenLORIS-scene datasets for
846
+ lifelong SLAM. In: IEEE International Conference on Robotics and Automation
847
+ (ICRA). pp. 3139–3145. IEEE (2020)
848
+ 32. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene
849
+ coordinate regression forests for camera relocalization in RGB-D images. In: IEEE
850
+ Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2930–2937
851
+ (2013)
852
+ 33. Song, S., Lichtenberg, S.P., Xiao, J.: Sun RGB-D: A RGB-D scene understanding
853
+ benchmark suite. In: IEEE Conference on Computer Vision and Pattern Recogni-
854
+ tion (CVPR). pp. 567–576 (2015)
855
+ 34. von Stumberg, L., Cremers, D.: DM-VIO: Delayed marginalization visual-inertial
856
+ odometry. IEEE Robotics and Automation Letters 7(2), 1408–1415 (2022)
857
+ 35. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for
858
+ the evaluation of RGB-D SLAM systems. In: International Conference on Intelli-
859
+ gent Robot Systems (IROS) (Oct 2012)
860
+ 36. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for
861
+ the evaluation of RGB-D SLAM systems. In: IEEE/RSJ International Conference
862
+ on Intelligent Robots and Systems. pp. 573–580. IEEE (2012)
863
+ 37. Telea, A.: An image inpainting technique based on the fast marching method.
864
+ Journal of graphics tools 9(1), 23–34 (2004)
865
+ 38. Valentin, J., Dai, A., Nießner, M., Kohli, P., Torr, P., Izadi, S., Keskin, C.: Learning
866
+ to navigate the energy landscape. In: Fourth International Conference on 3D Vision
867
+ (3DV). pp. 323–332. IEEE (2016)
868
+ 39. Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang, C., Hu, Y., Kapoor, A.,
869
+ Scherer, S.: TartanAir: A dataset to push the limits of visual SLAM. In: IEEE/RSJ
870
+
871
+ BS3D: Building-scale 3D Reconstruction from RGB-D Images
872
+ 15
873
+ International Conference on Intelligent Robots and Systems (IROS). pp. 4909–
874
+ 4916. IEEE (2020)
875
+ 40. Wasenmüller, O., Meyer, M., Stricker, D.: CoRBS: Comprehensive RGB-D bench-
876
+ mark for SLAM using Kinect v2. In: IEEE Winter Conference on Applications of
877
+ Computer Vision (WACV). pp. 1–7. IEEE (2016)
878
+ 41. Weder, S., Schonberger, J.L., Pollefeys, M., Oswald, M.R.: NeuralFusion: Online
879
+ depth fusion in latent space. In: Conference on Computer Vision and Pattern
880
+ Recognition (CVPR). pp. 3162–3172 (2021)
881
+ 42. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking
882
+ loss for single image depth prediction. In: IEEE/CVF Conference on Computer
883
+ Vision and Pattern Recognition (CVPR). pp. 611–620 (2020)
884
+ 43. Xiao, J., Owens, A., Torralba, A.: Sun3D: A database of big spaces reconstructed
885
+ using SfM and object labels. In: Proceedings of the IEEE international conference
886
+ on computer vision. pp. 1625–1632 (2013)
887
+ 44. Yin, W., Zhang, J., Wang, O., Niklaus, S., Mai, L., Chen, S., Shen, C.: Learning
888
+ to recover 3d scene shape from a single image. In: Conference on Computer Vision
889
+ and Pattern Recognition (CVPR) (2021)
890
+ 45. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy:
891
+ Disentangling task transfer learning. In: IEEE Conference on Computer Vision and
892
+ Pattern Recognition (CVPR). pp. 3712–3722 (2018)
893
+ 46. Zhou, Q.Y., Park, J., Koltun, V.: Open3D: A modern library for 3D data process-
894
+ ing. arXiv preprint arXiv:1801.09847 (2018)
895
+
JtAzT4oBgHgl3EQfIPso/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff