jackkuo commited on
Commit
2430923
·
verified ·
1 Parent(s): 74906ef

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. -9AyT4oBgHgl3EQf3vkG/content/tmp_files/2301.00772v1.pdf.txt +3502 -0
  2. -9AyT4oBgHgl3EQf3vkG/content/tmp_files/load_file.txt +0 -0
  3. -9AzT4oBgHgl3EQfFvoN/content/tmp_files/2301.01014v1.pdf.txt +2032 -0
  4. -9AzT4oBgHgl3EQfFvoN/content/tmp_files/load_file.txt +0 -0
  5. .gitattributes +53 -0
  6. 1NE0T4oBgHgl3EQfdgAR/vector_store/index.faiss +3 -0
  7. 1NE0T4oBgHgl3EQfdgAR/vector_store/index.pkl +3 -0
  8. 1dE0T4oBgHgl3EQf_wJK/vector_store/index.pkl +3 -0
  9. 2tAyT4oBgHgl3EQf1vk5/content/2301.00739v1.pdf +3 -0
  10. 2tAyT4oBgHgl3EQf1vk5/vector_store/index.faiss +3 -0
  11. 2tAyT4oBgHgl3EQf1vk5/vector_store/index.pkl +3 -0
  12. 3tE4T4oBgHgl3EQf0g0b/content/tmp_files/2301.05282v1.pdf.txt +1481 -0
  13. 3tE4T4oBgHgl3EQf0g0b/content/tmp_files/load_file.txt +0 -0
  14. 49E2T4oBgHgl3EQfkAeM/content/tmp_files/2301.03974v1.pdf.txt +979 -0
  15. 49E2T4oBgHgl3EQfkAeM/content/tmp_files/load_file.txt +0 -0
  16. 59E4T4oBgHgl3EQfcAxC/content/tmp_files/2301.05079v1.pdf.txt +1133 -0
  17. 59E4T4oBgHgl3EQfcAxC/content/tmp_files/load_file.txt +0 -0
  18. 6NAyT4oBgHgl3EQfcffh/vector_store/index.faiss +3 -0
  19. 6dE0T4oBgHgl3EQfvwHw/content/2301.02624v1.pdf +3 -0
  20. 6dE0T4oBgHgl3EQfvwHw/vector_store/index.faiss +3 -0
  21. 6dE0T4oBgHgl3EQfvwHw/vector_store/index.pkl +3 -0
  22. 6dE2T4oBgHgl3EQfkgfV/content/tmp_files/2301.03980v1.pdf.txt +586 -0
  23. 6dE2T4oBgHgl3EQfkgfV/content/tmp_files/load_file.txt +324 -0
  24. 79AzT4oBgHgl3EQfSPsz/content/2301.01228v1.pdf +3 -0
  25. 79AzT4oBgHgl3EQfSPsz/vector_store/index.faiss +3 -0
  26. 79AzT4oBgHgl3EQfSPsz/vector_store/index.pkl +3 -0
  27. 7NE1T4oBgHgl3EQfnATd/content/tmp_files/2301.03305v1.pdf.txt +1613 -0
  28. 7NE1T4oBgHgl3EQfnATd/content/tmp_files/load_file.txt +0 -0
  29. 7dAyT4oBgHgl3EQf2vnp/content/tmp_files/2301.00758v1.pdf.txt +1713 -0
  30. 7dAyT4oBgHgl3EQf2vnp/content/tmp_files/load_file.txt +0 -0
  31. 8dE2T4oBgHgl3EQf8Ait/content/tmp_files/2301.04215v1.pdf.txt +0 -0
  32. 8dE2T4oBgHgl3EQf8Ait/content/tmp_files/load_file.txt +0 -0
  33. 9dAyT4oBgHgl3EQfdPeW/vector_store/index.pkl +3 -0
  34. AtAyT4oBgHgl3EQfq_mZ/content/2301.00553v1.pdf +3 -0
  35. AtAyT4oBgHgl3EQfq_mZ/vector_store/index.faiss +3 -0
  36. AtE1T4oBgHgl3EQfDQOp/content/2301.02875v1.pdf +3 -0
  37. AtE1T4oBgHgl3EQfDQOp/vector_store/index.faiss +3 -0
  38. AtE1T4oBgHgl3EQfDQOp/vector_store/index.pkl +3 -0
  39. BNE1T4oBgHgl3EQf9Qa6/content/2301.03555v1.pdf +3 -0
  40. BNE1T4oBgHgl3EQf9Qa6/vector_store/index.faiss +3 -0
  41. BNE1T4oBgHgl3EQf9Qa6/vector_store/index.pkl +3 -0
  42. BNFKT4oBgHgl3EQfWi5i/content/tmp_files/2301.11792v1.pdf.txt +1203 -0
  43. BNFKT4oBgHgl3EQfWi5i/content/tmp_files/load_file.txt +0 -0
  44. C9AzT4oBgHgl3EQfTvzN/content/2301.01256v1.pdf +3 -0
  45. FNE1T4oBgHgl3EQf-gbR/content/tmp_files/2301.03571v1.pdf.txt +1215 -0
  46. FNE1T4oBgHgl3EQf-gbR/content/tmp_files/load_file.txt +0 -0
  47. FdAzT4oBgHgl3EQfUfww/content/tmp_files/2301.01266v1.pdf.txt +0 -0
  48. FtE3T4oBgHgl3EQftQtg/vector_store/index.faiss +3 -0
  49. GtE1T4oBgHgl3EQf_QYB/vector_store/index.pkl +3 -0
  50. H9E1T4oBgHgl3EQfFgPu/content/tmp_files/2301.02904v1.pdf.txt +1040 -0
-9AyT4oBgHgl3EQf3vkG/content/tmp_files/2301.00772v1.pdf.txt ADDED
@@ -0,0 +1,3502 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1
2
+ A Unified Visual Information Preservation
3
+ Framework for Self-supervised Pre-training in
4
+ Medical Image Analysis
5
+ Hong-Yu Zhou, Student Member, IEEE, Chixiang Lu, Chaoqi Chen, Sibei Yang,
6
+ and Yizhou Yu, Fellow, IEEE,
7
+ Abstract—Recent advances in self-supervised learning (SSL) in computer vision are primarily comparative, whose goal is to preserve
8
+ invariant and discriminative semantics in latent representations by comparing siamese image views. However, the preserved high-level
9
+ semantics do not contain enough local information, which is vital in medical image analysis (e.g., image-based diagnosis and tumor
10
+ segmentation). To mitigate the locality problem of comparative SSL, we propose to incorporate the task of pixel restoration for explicitly
11
+ encoding more pixel-level information into high-level semantics. We also address the preservation of scale information, a powerful tool
12
+ in aiding image understanding but has not drawn much attention in SSL. The resulting framework can be formulated as a multi-task
13
+ optimization problem on the feature pyramid. Specifically, we conduct multi-scale pixel restoration and siamese feature comparison in
14
+ the pyramid. In addition, we propose non-skip U-Net to build the feature pyramid and develop sub-crop to replace multi-crop in 3D
15
+ medical imaging. The proposed unified SSL framework (PCRLv2) surpasses its self-supervised counterparts on various tasks,
16
+ including brain tumor segmentation (BraTS 2018), chest pathology identification (ChestX-ray, CheXpert), pulmonary nodule detection
17
+ (LUNA), and abdominal organ segmentation (LiTS), sometimes outperforming them by large margins with limited annotations. Codes
18
+ and models are available at https://github.com/RL4M/PCRLv2.
19
+ Index Terms—Medical image analysis, Self-supervised learning, Transfer Learning, Context restoration, Feature pyramid.
20
+ !
21
+ 1
22
+ INTRODUCTION
23
+ I
24
+ T is usual to acquire a substantial amount of manually
25
+ labeled data before training deep neural networks. This
26
+ condition is easy to meet in natural images, where labor
27
+ costs and labeling difficulties are tolerable. In medical image
28
+ analysis, however, credible annotations are mainly derived
29
+ from domain experts’ diagnoses, which are challenging to
30
+ obtain due to the rarity of the target disease, the need to safe-
31
+ guard patient privacy, and the scarcity of medical resources.
32
+ Against this background, self-supervised learning (SSL) has
33
+ been widely accepted as a viable technique to learn medical
34
+ image representations without specialistic annotations. We
35
+ usually deploy SSL in the pre-training stage to obtain well-
36
+ transferable features, which can be transferred to various
37
+ downstream tasks for performance boosting.
38
+ Recent advances in SSL are mostly based on compar-
39
+ ative learning [8], [10], [15], [17]. The rationale behind is
40
+ to learn transferable latent representations with invariant
41
+ and discriminative semantics by maximizing the mutual
42
+ information between a pair of siamese images. One potential
43
+ problem of these comparative methods is that they mainly
44
+ focus on encoding high-level global semantics in representa-
45
+
46
+ Hong-Yu Zhou, Chixiang Lu, Chaoqi Chen, and Yizhou Yu are with the
47
+ Department of Computer Science, The University of Hong Kong, Hong
48
+ Kong. Email: {whuzhouhongyu, luchixiang, cqchen1994}@gmail.com,
49
+ yizhouy@acm.org.
50
+
51
+ Sibei Yang is with ShanghaiTech University and Shanghai Engineering
52
+ Research Center of Intelligent Vision and Imaging, Shanghai, China.
53
+ Email: yangsb@shanghaitech.edu.cn.
54
+
55
+ First two authors contributed equally.
56
+
57
+ Corresponding author: Sibei Yang and Yizhou Yu.
58
+ tions but ignore the preservation of pixel-level information1.
59
+ However, in medical image analysis, the latter type of
60
+ information usually plays a vital role. For instance, in chest
61
+ pathology detection, radiologists or clinicians are required
62
+ to point out small lesions from a chest X-ray according to
63
+ their textures. Sometimes, these areas of pathologies are so
64
+ hard to identify that even medical experts have to check
65
+ pixel-level details to tell where the lesions are. Another
66
+ typical example lies in brain tumor segmentation, where the
67
+ segmentation error of one voxel may cause irreparable harm
68
+ to patients in brain surgeries, such as a permanent damage
69
+ to the cochlear nerve when trying to remove the acoustic
70
+ neuroma.
71
+ An intuitive way to preserve pixel-level information in
72
+ learned features is to restore the pixel-level content from
73
+ latent representations directly. This methodology, known
74
+ as context restoration [29], has already been adopted as a
75
+ surrogate task in pretext-based SSL for natural [23], [29], [44]
76
+ and medical images [7], [49]. Specifically, these approaches
77
+ first apply various data augmentation strategies to a given
78
+ image to generate a corrupted input, based on which deep
79
+ models are trained to restore original pixels. In this way,
80
+ we explicitly require the latent representations to preserve
81
+ information closely related to pixels. Although pure pixel-
82
+ based features are not as transferable as those from com-
83
+ parative SSL [17], [48], we hypothesize it is still beneficial
84
+ to explicitly preserve pixel-level information and global se-
85
+ 1. In 3D medical images, we often use “voxel” to denote the same
86
+ concept as the pixel does in 2D images. For simplicity, we use “pixel”
87
+ to denote the smallest addressable element in both 2D and 3D images
88
+ in the rest of this paper.
89
+ arXiv:2301.00772v1 [cs.CV] 2 Jan 2023
90
+
91
+ 2
92
+ ℱ#
93
+ ℱ$
94
+ ℱ%
95
+ ℱ"
96
+ ℱ!
97
+ Pixels
98
+ Semantics
99
+ Scales
100
+ Fig. 1. Motivation illustration. We propose a unified SSL framework
101
+ to simultaneously preserve information in visual representations from
102
+ perspectives of pixels, semantics, and scales. {F1, F2, F3, F4, F5} de-
103
+ note different levels in the feature pyramid, given an input image. Our
104
+ approach restores uncorrupted inputs from the feature maps directly
105
+ to preserve pixel-level details. In order to retain the global semantic
106
+ information, our method compares siamese one-dimensional represen-
107
+ tations. Last but not the least, the proposed methodology conducts pixel
108
+ restoration and feature comparison at different scales. The rationale
109
+ behind is to introduce multi-scale self-supervised latent representations,
110
+ making them more transferable to various downstream tasks.
111
+ mantics, especially in medical image analysis where details
112
+ matter a lot.
113
+ Besides semantics and pixels, introducing multi-scale
114
+ representations has been proven to be quite helpful in
115
+ aiding image understanding [12], [24], [26], [27], [32], [39].
116
+ The common practice of these methods is to construct a
117
+ feature pyramid during training, testing, or both stages.
118
+ Then, various tasks, such as detection, and segmentation,
119
+ can be conducted on the basis of multi-scale features. The
120
+ goal of building the feature pyramid is to endow image
121
+ representations with the ability to recognize objects at dif-
122
+ ferent scales, which is also consistent with the law of human
123
+ cognition [31]. However, the preservation of visual informa-
124
+ tion at multiple scales is rarely mentioned in SSL. Thus, it
125
+ is unclear whether introducing multi-scale self-supervised
126
+ representations provides a stronger transfer learning ability.
127
+ In Figure 1, we illustrate the motivation of the proposed
128
+ unified visual information preservation framework for SSL.
129
+ The introduced framework addresses the preservation of
130
+ information in self-supervised visual representations from
131
+ three aspects: pixels, semantics, and scales. Firstly, to re-
132
+ tain pixel-level information in latent representations, our
133
+ framework involves a reconstruction branch in the self-
134
+ supervised model to rebuild uncorrupted images from cor-
135
+ rupted inputs. Specifically, we ask the self-supervised model
136
+ to restore pixels from feature maps of randomly corrupted
137
+ inputs during training. As a result, information closely
138
+ associated with pixels can be explicitly encoded into the
139
+ latent representations. In practice, this type of information
140
+ would enhance the ability of self-supervised representations
141
+ to recognize and differentiate textures. Apart from pixel-
142
+ level information, preserving invariant and discriminative
143
+ semantics in visual representations is also necessary. To-
144
+ wards this end, we adopt the existing comparative SSL
145
+ to encode invariant semantic information by comparing
146
+ high-level representations of siamese image patches [10].
147
+ We empirically found the siamese SSL not only produces
148
+ comparably (sometimes more) transferable medical image
149
+ representations but also is much easier to implement in com-
150
+ parison to the typical contrastive manner [17]. Last but not
151
+ the least, the proposed unified framework introduces multi-
152
+ scale latent representations by conducting pixel restoration
153
+ and feature comparison in a range of scales. To achieve this
154
+ goal, we propose a non-skip U-Net (nsUNet) that constructs a
155
+ feature pyramid upon the U-shape architecture [32]. In prac-
156
+ tice, nsUNet effectively avoids the production of shortcut
157
+ solutions when performing the context restoration task. On
158
+ the basis of nsUNet, we conduct pixel-level context restora-
159
+ tion and siamese feature comparison in each level (i.e., scale)
160
+ of the feature pyramid. In this way, the proposed framework
161
+ helps improve the ability of self-supervised representations
162
+ to recognize objects (e.g., lesions and organs in medical
163
+ images) at different sizes and scales.
164
+ We summarize the contributions of this paper as follows:
165
+
166
+ We present an information preservation framework
167
+ for advancing SSL in medical image analysis. In this
168
+ framework, we unify the preservation of visual infor-
169
+ mation in latent representations from three aspects:
170
+ pixels, semantics, and scales. Towards this end, pixel
171
+ restoration and feature comparison are conducted at
172
+ different feature scales.
173
+
174
+ We introduce non-skip U-Net (nsUNet) to construct
175
+ the feature pyramid. Compared to the typical U-
176
+ shape models in medical imaging [11], [32], nsUNet
177
+ maintains more feature scales and eliminates the us-
178
+ age of the widely adopted skip connections to avoid
179
+ shortcut solutions to pixel restoration.
180
+
181
+ Inspired by multi-crop [5], we propose sub-crop to
182
+ compare global volumes against local volumes. In
183
+ order to mitigate the problem of the reduced mutual
184
+ information between global and local views in 3D
185
+ space, sub-crop restricts the cropping of local views
186
+ within the 3D minimum bounding box of global
187
+ views. Experiments on 3D medical images found that
188
+ sub-crop is more effective than multi-crop in various
189
+ downstream tasks.
190
+
191
+ We conduct extensive and comprehensive experi-
192
+ ments to validate the effectiveness of the proposed
193
+ framework. We show that the unification of pixels,
194
+ semantics, and scales can provide impressive perfor-
195
+ mance under the pre-training/fine-tuning protocol.
196
+ Specifically, the proposed framework outperforms
197
+ both self-supervised and supervised counterparts in
198
+ chest pathology classification, pulmonary nodule de-
199
+ tection, abdominal organ segmentation, and brain
200
+ tumor segmentation by substantial margins.
201
+ The conference version of this paper (PCRLv1) was pre-
202
+ sented in [47], which demonstrates the benefits of incor-
203
+ porating more pixel-level information besides the invariant
204
+ and discriminative semantics obtained by contrastive learn-
205
+ ing. In this paper, we made significant and substantial modi-
206
+ fications to PCRLv1, and we name the improved framework
207
+ as PCRLv2 (i.e., Preservational Comparative Representation
208
+ Learning). The modifications and improvements in PCRLv2
209
+ include but are not limited to (i) Besides local pixel-level
210
+ and global semantic information, scale information is also
211
+
212
+ 3
213
+ ×
214
+ ×
215
+ ×
216
+ ×
217
+ ×
218
+ ×
219
+ ×
220
+ ×
221
+ R
222
+ R
223
+ R
224
+ R
225
+ R
226
+ 𝑥
227
+ 𝑥!
228
+ 𝑥"
229
+ t!
230
+ t"
231
+ R
232
+ R
233
+ R
234
+ R
235
+ R
236
+ R Pixel restoration
237
+ Candidate scale
238
+ Chosen scale
239
+ nsUNet
240
+ Siamese nsUNet
241
+ 𝑥!
242
+ #
243
+ t!
244
+ #
245
+ 𝑥"
246
+ #
247
+ t"
248
+ #
249
+ Global aug.
250
+ Global aug. Local aug.
251
+ Local aug.
252
+ (a) Multi-scale pixel restoration
253
+ ×
254
+ ×
255
+ ×
256
+ ×
257
+ ×
258
+ ×
259
+ ×
260
+ ×
261
+ C
262
+ C
263
+ C
264
+ C
265
+ C
266
+ C Feature comparison
267
+ Candidate scale
268
+ Chosen scale
269
+ nsUNet
270
+ Siamese nsUNet
271
+ 𝑥!
272
+ 𝑥"
273
+ 𝑥!
274
+ #
275
+ 𝑥"
276
+ #
277
+ 𝑥
278
+ t!
279
+ t"
280
+ t!
281
+ #
282
+ t"
283
+ #
284
+ Global aug.
285
+ Global aug. Local aug.
286
+ Local aug.
287
+ (b) Multi-scale feature comparison
288
+ Fig. 2. The overall structure of PCRLv2. PCRLv2 performs self-supervised visual learning on siamese feature pyramids. To achieve this goal, we
289
+ propose non-skip U-Net (nsUNet). nsUNet consists of five feature scales and removes the skip connections to prevent network optimizers from
290
+ finding shortcut solutions to context restoration. On the basis of nsUNet, we propose to decouple the preservation of pixel-level, semantic, and
291
+ scale information into two tasks: (a) multi-scale pixel restoration; (b) multi-scale feature comparison. The rationale behind is to incorporate pixel
292
+ details and semantics into features at different scales. During the training stage, we randomly choose a feature scale from the feature pyramid,
293
+ on top of which we conduct pixel restoration and feature comparison. x denotes a batch of input images. t1 and t2 stand for two distinct global
294
+ augmentations, while t′
295
+ 1 and t′
296
+ 2 denote the successive local augmentations.
297
+ preserved in self-supervised visual representations. The
298
+ motivation behind is that although multiple feature scales
299
+ have been considered in various vision tasks, they have
300
+ not drawn much attention in SSL. PCRLv2 shows that
301
+ introducing multi-scale latent representations can boost the
302
+ transfer learning performance of SSL in downstream tasks.
303
+ (ii) PCRLv2 simplifies the attentional pixel restoration and
304
+ hybrid feature contrast operations of PCRLv1 into a con-
305
+ cise multi-task optimization problem. As a result, PCRLv2
306
+ is simpler and easier to implement while achieving bet-
307
+ ter performance, thus more practical. (iii) Compared to
308
+ PCRLv1 that relies on the plain U-Net architecture [32],
309
+ PCRLv2 conducts SSL on top of a new backbone, i.e., non-
310
+ skip U-Net (nsUNet). There are two inherent advantages
311
+ of nsUNet. First, the feature pyramid of nsUNet allows
312
+ performing multi-scale pixel-level context restoration and
313
+ semantic feature comparison. As a result, the unification
314
+ of pixels, semantics, and scales produces more transferable
315
+ visual representations. Second, nsUNet can effectively avoid
316
+ the production of shortcut solutions, providing obvious
317
+ performance gains over the use of the typical skip con-
318
+ nections. (iv) We integrate the idea of multi-crop [5] in
319
+ PCRLv2. Moreover, in 3D medical imaging, we propose sub-
320
+ crop to produce reliable local views with increased mutual
321
+ information by randomly cropping multiple local volumes
322
+ within the 3D minimum bounding box of global views. In
323
+ practice, we found that the proposed sub-crop has better
324
+ pre-training performance than multi-crop. (v) In 5 classifica-
325
+ tion/segmentation tasks, PCRLv2 provides more transfer-
326
+ able pre-trained visual representations, not only surpass-
327
+ ing previous self-supervised and supervised counterparts
328
+ by substantial margins but also obviously outperforming
329
+ PCRLv1 in all experiments.
330
+ 2
331
+ RELATED WORK
332
+ This section reviews related work in comparative SSL,
333
+ including contrastive and non-contrastive methods, and
334
+ lists SSL approaches that use context restoration as the
335
+ pretext task. In the third part, we collect papers that
336
+ emphasize the incorporation of multi-scale features in SSL.
337
+ Comparative SSL methodologies. One of the core ideas
338
+ behind comparative SSL is to extract and encode invari-
339
+ ant and discriminative semantics into representations via
340
+ feature-level comparison. Hjelm et al. [20] proposed Deep In-
341
+ foMax to maximize the mutual information between global
342
+ and local feature vectors of the same input image using
343
+ InfoNCE [28]. Bachman et al. [3] augmented InfoMax by
344
+ conducting a global-local comparison on feature vectors of
345
+ independently-augmented versions of each input. Tian et
346
+ al. [36] increased the number of augmented views of each
347
+ input and extended InfoNCE to multiple views. He et al. [17]
348
+ presented Momentum Contrast (MoCo), which comprises
349
+ a momentum encoder to maintain the consistency among
350
+ positive and negative feature vectors. Different from [3],
351
+ [20], MoCo performs InfoNCE on top of global feature
352
+ vectors only. Compared to MoCo, SimCLR removes the
353
+ momentum architecture and defines InfoNCE on the output
354
+ of a MLP with one hidden layer. Inspired by SimCLR,
355
+ Chen et al. [9] proposed MoCov2, which improves MoCo
356
+ with an additional MLP head and more augmentations.
357
+ SwAV [5] replaces the feature vectors in InfoNCE with
358
+ cluster assignments and introduces the multi-crop strategy
359
+ to increase the number of views of an image with affordable
360
+ computational overhead. Grill et al. [15] proposed BYOL
361
+ (bootstrap your own latent), which eliminates the use of
362
+ InfoNCE in SSL by distilling semantics from positive pairs
363
+ only. Based on BYOL, Chen et al. [10] further removed the
364
+ restriction of the momentum architecture and introduced
365
+ a simple siamese learning framework named SimSiam. In
366
+ practice, SimSiam produces comparable results to MoCov2
367
+ in various downstream tasks. Recently, Zbontar et al. [42]
368
+ simplified SimSiam by measuring the cross-correlation ma-
369
+ trix between the siamese global feature vectors and trying
370
+ to make this matrix close to the identity.
371
+ Comparative SSL, especially InfoNCE-based method-
372
+ ology, has also been widely adopted in medical image
373
+
374
+ 4
375
+ analysis. Zhou et al. [48] proposed to integrate mixup [43]
376
+ into MoCov2, increasing the diversity of both positive and
377
+ negative samples in InfoNCE. Taleb et al. [34] developed 3D
378
+ versions of existing SSL techniques and compared 2D and
379
+ 3D SSL approaches on downstream tasks. Azizi et al. [2]
380
+ incorporated multi-instance learning into SimCLR, which
381
+ helps utilize multiple views of each patient. Around the
382
+ same time, Vu et al. [37] developed a method to select posi-
383
+ tive pairs coming from views of the same patient and used
384
+ this strategy to improve MoCov2. There are also a number
385
+ of approaches [6], [40], [41] that tailored comparative SSL
386
+ for semi-supervised medical image segmentation.
387
+ However, the methodologies mentioned above fail
388
+ to
389
+ address
390
+ the
391
+ importance
392
+ of
393
+ integrating
394
+ pixel-level
395
+ information into the high-level representations with rich
396
+ semantics, which is the primary focus of the proposed
397
+ PCRL.
398
+ Context restoration for preserving pixel-level information.
399
+ Restoring original context has been treated as an important
400
+ pretext task in SSL. Pathak et al. [29] first time conducted
401
+ self-supervised feature learning by recovering masked input
402
+ images. Larsson et al. [23] and Zhang et al. [44] performed
403
+ SSL on pixels via predicting RGB color values. For medical
404
+ images, Chen et al. [7] extended the approach in [29] with
405
+ swapped image patches. Zhou et al. [49] showed that adding
406
+ more augmentations to input images brings benefits to SSL.
407
+ Tao et al. [35] presented a volume-wise context transforma-
408
+ tion for 3D medical images. Different from the approaches
409
+ mentioned above, Henaff [19] proposed to predict the next
410
+ context feature vectors following an auto-regressive manner.
411
+ We can see that context restoration is more prevalent in
412
+ medical imaging than in natural images from the above.
413
+ The underlying reason is that medical imaging tasks
414
+ require more pixel-level information to make fine-grained
415
+ yet accurate decisions. On the other hand, we observe
416
+ that comparative SSL can produce representations with
417
+ richer semantics. Thus, it can be beneficial to build a SSL
418
+ framework that simultaneously integrates pixel-level and
419
+ semantic information. As far as we are concerned, none
420
+ of these context restoration based approaches incorporate
421
+ such a combination.
422
+ Multi-scale features in SSL. Although multi-scale features
423
+ have not drawn much attention in existing SSL research,
424
+ it has already been treated as an implicit yet effective
425
+ regularization method for SSL in some methodologies. Deep
426
+ InfoMax [20] contrasts high-level feature vectors with low-
427
+ level feature maps using InfoNCE. To improve Deep Info-
428
+ Max, Bachman et al. [3] proposed to contrast global and
429
+ local feature vectors on multiple levels. In medical image
430
+ analysis, preserving scale information becomes essential,
431
+ as pathologies may show different characteristics on dif-
432
+ ferent scales. In [6], a local contrastive loss is introduced
433
+ to learn distinctive representations of local regions that are
434
+ helpful to per-pixel segmentation. At the same time, global
435
+ feature vectors are used to distill discriminative semantics
436
+ for classification tasks. A similar idea has also been used
437
+ in image registration [25] and one-shot segmentation [46],
438
+ where global and local feature vectors are employed to
439
+ provide information on semantics and position, respectively.
440
+ However, most of these methods only perform SSL on
441
+ two scales, i.e., one global and one local, which cannot fully
442
+ capture multi-scale information. Besides, although these
443
+ approaches emphasize the benefit of introducing local in-
444
+ formation to SSL, they do not exploit pixel-level information
445
+ that is helpful to encode locality. In contrast, this paper pro-
446
+ poses a unified framework that can simultaneously preserve
447
+ semantic, pixel-level, and scale information.
448
+ 3
449
+ METHODOLOGY
450
+ We provide an overview of PCRLv2 in Fig. 2. Suppose x
451
+ denotes a batch of input images. We introduce cascaded
452
+ augmentations to distort x in global and local views, respec-
453
+ tively. To be specific, the first-stage augmentations (t1 and t2
454
+ in Fig. 2) mainly consist of global transformations, such as
455
+ flip and rotation, whose goal is to distort the semantics of
456
+ input images from a global perspective. In comparison, the
457
+ second-stage augmentations (t′
458
+ 1 and t′
459
+ 2 in Fig. 2) comprise
460
+ local pixel-level transformations, such as random noise and
461
+ gaussian blur, which are leveraged to perturb the local
462
+ semantics. After two-stage augmentations, the finally aug-
463
+ mented images x′
464
+ 1 and x′
465
+ 2 are passed to siamese networks
466
+ to perform pixel restoration and feature comparison, while
467
+ the results of applying t1 and t2 to x, i.e., x1 and x2, serve
468
+ as the ground truth targets for the pixel restoration task (as
469
+ shown in Fig. 2a).
470
+ We perform SSL on the feature pyramid to encode multi-
471
+ scale visual representations. Following the standard practice
472
+ in medical image processing, we build feature pyramids
473
+ using a U-shape model named non-skip U-Net (nsUNet).
474
+ Compared to the typical U-Net architecture [11], [32],
475
+ nsUNet has more feature scales and completely removes
476
+ skip connections, both of which we empirically found help-
477
+ ful in producing better pre-trained representations. During
478
+ the training stage, one scale is first randomly chosen from all
479
+ five feature scales, after which we conduct pixel restoration
480
+ and feature comparison on the siamese feature maps at the
481
+ chosen scale. After the pre-training stage, we fine-tune the
482
+ encoder of nsUNet on various downstream tasks.
483
+ 3.1
484
+ Feature pyramid in non-skip U-Net
485
+ U-Net and its series [11], [22], [32] have been known in med-
486
+ ical imaging for their abilities to handle image segmentation
487
+ tasks. The most distinctive characteristic of these models is
488
+ the skip connection that connects equal-resolution low- and
489
+ high-level feature maps. The critical insight is to recover the
490
+ spatial information lost in down-sampling operations of the
491
+ encoder network, such as strided pooling or convolution.
492
+ U-shape models use a feature pyramid to progressively
493
+ incorporate multi-scale details brought by skip connections
494
+ into high-level semantics, making the U-shape architecture
495
+ an ideal choice for conducting context restoration.
496
+ In this paper, we explore the potential of U-shape ar-
497
+ chitecture in SSL from two perspectives: deeply fusing
498
+ semantic and pixel-level information by removing the skip
499
+ connections and introducing multi-scale latent representa-
500
+ tions by conducting SSL on the feature pyramid. For the
501
+ first perspective, we empirically found that skip connec-
502
+ tions provide shortcuts for context restoration, as the low-
503
+ level feature maps contain rich, high-resolution pixel-level
504
+
505
+ 5
506
+ ×
507
+ ×
508
+ ×
509
+ ×
510
+ 𝐻
511
+ 2 × 𝑊
512
+ 2
513
+ 𝐻×𝑊
514
+ 𝐻
515
+ 4 × 𝑊
516
+ 4
517
+ 𝐻
518
+ 8 × 𝑊
519
+ 8
520
+ 𝐻
521
+ 16 × 𝑊
522
+ 16
523
+ 𝐻
524
+ 32 × 𝑊
525
+ 32
526
+ Skip feature maps
527
+ Feature
528
+ hierarchy
529
+ Down-sampling
530
+ Up-sampling
531
+ ×
532
+ No skip connection
533
+ Conv+BN
534
+ +ReLU
535
+ ×2
536
+ Conv+BN
537
+ +ReLU
538
+ ×2
539
+ Conv+BN
540
+ +ReLU
541
+ ×2
542
+ Conv+BN
543
+ +ReLU
544
+ ×2
545
+ Conv+BN
546
+ +ReLU
547
+ ×2
548
+ Fig. 3. The architecture of non-skip U-Net (nsUNet). In comparison
549
+ to previous U-Net series, nsUNet removes skip connections, and the
550
+ associated skip feature maps to prevent shortcut solutions to the pixel
551
+ restoration and feature comparison tasks. Besides, nsUNet consists of
552
+ five levels of feature maps (denoted with different colors), where two self-
553
+ supervised tasks are further conducted. Note that this is a 2D illustration
554
+ of nsUNet.
555
+ details. This characteristic does contribute to the restoration
556
+ of context. However, it may prevent the high-level latent
557
+ representations (with rich semantics) from incorporating
558
+ more pixel-level information because the task of providing
559
+ pixel-level details is assigned to low-level feature maps. To
560
+ address this point, we remove the skip connections in U-
561
+ shape architecture and propose non-skip U-Net (nsUNet).
562
+ nsUNet relies on high-level representations without any
563
+ skip connections to restore pixel-level details. In this way,
564
+ the semantic and pixel-level information can be deeply
565
+ fused. Meanwhile, the inherent multi-scale feature maps of
566
+ nsUNet offer the opportunity to construct a feature pyra-
567
+ mid, on top of which SSL can be conducted in multiple
568
+ scales simultaneously.
569
+ Fig. 3 presents the architecture of nsUNet. The feature
570
+ pyramid in nsUNet comprises five levels, ranging from low
571
+ resolution (the down-sampling rate is 32) to full resolution
572
+ (no down-sampling). For 2D input data, we use ResNet-
573
+ 18 [18] as the encoder, while for 3D input volumes, we
574
+ build the encoder following [11]. As illustrated in Fig. 3, the
575
+ decoder of nsUNet maintains a shared architecture across
576
+ all pyramid levels, which can be summarized as:
577
+ Fi = Conv-BN-ReLU(Conv-BN-ReLU(Up(Fi−1)),
578
+ (1)
579
+ where i ∈ {1, 2, 3, 4, 5}. F0 denotes the output of the bottle-
580
+ neck block, which has the lowest spatial resolution (down-
581
+ sampling rate=32). Up represents the up-sampling opera-
582
+ tion. Conv-BN-ReLU stands for a sequence of operations,
583
+ including convolution (kernel size=3), batch normalization
584
+ (BN), and ReLU activation. As a result, the bag of feature
585
+ maps {F1, F2, F3, F4, F5} is then forwarded to following
586
+ task-dependent heads to perform pixel restoration and fea-
587
+ ture comparison, respectively and simultaneously.
588
+ 3.2
589
+ Multi-scale pixel restoration
590
+ As the name implies, multi-scale pixel restoration aims to
591
+ preserve pixel-level and scale information in latent visual
592
+ representations simultaneously. To achieve this goal, we ask
593
+ the network to recover the exact pixel-level details across
594
+ different scales, where each pair of siamese feature maps
595
+ share one pixel restoration head. In contrast, PCRLv1 only
596
+ restores pixel details at the full resolution, which inevitably
597
+ loses multi-scale properties in learned representations.
598
+ As shown in Fig. 4a, the input images x′
599
+ 1 and x′
600
+ 2 are
601
+ intentionally corrupted via various pixel-level augmenta-
602
+ tions, such as guassian blur and random noise. For each
603
+ training iteration, we first randomly choose a feature scale
604
+ Fi from {F1, F2, F3, F4, F5}. Then, we pass Fi to the pixel
605
+ restoration head f R
606
+ i (·) for the i-th scale, whose internal
607
+ processing procedure can be summarized as:
608
+ f R
609
+ i (Fi) = Conv(Conv-BN-ReLU(Fi)),
610
+ (2)
611
+ where all convolution layers use a kernel size of 3 and a
612
+ stride of 2. Similarly, we apply the shared pixel restoration
613
+ head to the paired siamese feature map Fs
614
+ i to acquire the
615
+ prediction output f R
616
+ i (Fs
617
+ i ):
618
+ f R
619
+ i (Fs
620
+ i ) = Conv(Conv-BN-ReLU(Fs
621
+ i )),
622
+ (3)
623
+ Lastly, we employ the mean square error (MSE) loss to
624
+ measure the reconstruction errors between f R
625
+ i (Fi) and x1.
626
+ For the siamese feature pyramid, we apply MSE loss to
627
+ f R
628
+ i (Fs
629
+ i ) and x2. The cost function LR of the pixel restoration
630
+ task in each training iteration (with mini-batch optimiza-
631
+ tion) is as follows:
632
+ LR =
633
+ N
634
+
635
+ j=1,
636
+ ∀i∈H
637
+ 1[i==j] [MSE(f R
638
+ i (Fi), x1) + MSE(f R
639
+ i (Fs
640
+ i ), x2)],
641
+ (4)
642
+ where N = 5 denotes the number of scales in each feature
643
+ pyramid. H = {1, 2, 3, 4, 5} stands for the scale index.
644
+ 1[i==j] is an indicator function, which is equal to 1 when
645
+ i==j is true (otherwise, 0). The explanation of LR can be
646
+ summarized as: (i) randomly choose a feature scale Fi
647
+ from all five scales; (ii) pass Fi and its siamese feature
648
+ map Fs
649
+ i to the shared task head f R
650
+ i (·); (iii) calculate the
651
+ MSE loss between the outputs of f R
652
+ i (·) and uncorrupted
653
+ images {x1, x2}. By reconstructing the same targets x1/x2
654
+ across different feature scales, LR can encode the pixel-level
655
+ information into multi-scale latent visual representations.
656
+ 3.3
657
+ Multi-scale feature comparison
658
+ PCRLv1 employs a hybrid way to conduct contrastive
659
+ learning with the help of the momentum encoder [17]
660
+ and mixup [43]. However, this contrastive deployment is
661
+ complex, making PCRLv1 heavy, thus troublesome to im-
662
+ plement and improve. To address these issues, PCRLv2
663
+ replaces the hybrid contrastive strategies in PCRLv1 with
664
+ the multi-scale comparison. Inspired by [10], multi-scale
665
+ comparison conducts SSL with siamese learning, whose
666
+ key operation is to attract the same image’s siamese views.
667
+ Different from [10] that conducts feature comparison on one
668
+ scale, we propose to preserve the discriminative semantics
669
+ across different feature scales, which forces the model to
670
+ preserve multi-scale self-supervised representations. In the
671
+ following, we provide technical details of performing the
672
+ multi-scale comparison.
673
+
674
+ 6
675
+ Conv
676
+ BN
677
+ ReLU
678
+ Conv
679
+ Conv
680
+ BN
681
+ ReLU
682
+ Conv
683
+ 𝑥!
684
+ 𝑥#
685
+ Shared
686
+ 𝑥!
687
+ "
688
+
689
+ 𝑥#
690
+ "
691
+
692
+ Siamese scale
693
+ Chosen scale
694
+ (a) Architectural details of the pixel restoration head
695
+ Siamese scale
696
+ Chosen scale
697
+ GAP
698
+ GAP
699
+ BN
700
+ FC
701
+ BN
702
+ ReLU
703
+ FC
704
+ BN
705
+ FC
706
+ BN
707
+ ReLU
708
+ FC
709
+ Predictor
710
+ Predictor
711
+ Shared
712
+ 𝑥!
713
+ "
714
+
715
+ 𝑥#
716
+ "
717
+
718
+ (b) Architectural details of the feature comparison head
719
+ Fig. 4. Architectural details of the pixel restoration and feature comparison heads. Conv, BN, GAP, and FC denote the convolution, batch
720
+ normalization, global average pooling, and fully-connected layers, respectively. The kernel size of all convolution layers is 3, and the convolution
721
+ stride is set to 1. Note that each pair of siamese feature maps share one pixel restoration head and one feature comparison head, while different
722
+ feature scales employ distinct task heads.
723
+ Given the feature maps at a randomly chosen scale Fi,
724
+ we pass them through a global average pooling layer and
725
+ a shared batch normalization layer (as shown in Fig. 4b) to
726
+ acquire 1D representations vi:
727
+ vi = BN(GAP(Fi)).
728
+ (5)
729
+ We can get vs
730
+ i by processing the siamese feature maps Fs
731
+ i in
732
+ a similar way.
733
+ Next, we forward vi to the shared predictor fP(·), whose
734
+ architecture is displayed in Fig. 4b and can be summarized
735
+ as:
736
+ fP(vi) = FC(FC-BN-ReLU(vi)).
737
+ (6)
738
+ where FC denotes the fully-connected layer. FC-BN-ReLU
739
+ stands for a sequence of layers, which are the fully-
740
+ connected layer, batch normalization layer, and ReLU ac-
741
+ tivation. Similarly, we can acquire fP(vs
742
+ i ) by passing vs
743
+ i to
744
+ the same predictor.
745
+ We measure the similarity between siamese feature vec-
746
+ tors with the cosine similarity:
747
+ cos(vi, fP(vs
748
+ i )) =
749
+ vi
750
+ ∥vi∥2
751
+ ·
752
+ fP(vs
753
+ i )
754
+ ∥fP(vs
755
+ i )∥2
756
+ ,
757
+ (7)
758
+ where || · ||2 denotes the L2 normalization. Symmetrically,
759
+ we calculate cos(fP(vi), vs
760
+ i ) as follows:
761
+ cos(fP(vi), vs
762
+ i )) =
763
+ fP(vi)
764
+ ∥fP(vi)∥2
765
+ ·
766
+ vs
767
+ i
768
+ ∥vs
769
+ i ∥2
770
+ .
771
+ (8)
772
+ Finally, the cost function LC of multi-scale feature compari-
773
+ son can be summarized as:
774
+ LC =
775
+ N
776
+
777
+ j=1,
778
+ ∀i∈H
779
+ −1
780
+ 21[i==j] [cos(sg(vi), fP(vs
781
+ i ))
782
+ + cos(fP(vi), sg(vs
783
+ i ))].
784
+ (9)
785
+ N
786
+ =
787
+ 5 denotes the number of feature scales. H
788
+ =
789
+ {1, 2, 3, 4, 5} stands for the scale index. Following [10], we
790
+ apply the stop-gradient operation (denoted as sg) in Eq. 9
791
+ to prevent the network optimizer from finding shortcut
792
+ solutions.
793
+ Minimizing LC requires the model to maximize the
794
+ similarity between siamese latent features across all feature
795
+ scales. In this way, scale invariance can be implicitly incor-
796
+ porated into the preserved latent semantics.
797
+ 1
798
+ 2
799
+ 1
800
+ 2
801
+ 3
802
+ 4
803
+ 5
804
+ 6
805
+ 7
806
+ 8
807
+ Randomly crop two
808
+ global patches with
809
+ an IoU constraint
810
+ Find the minimum 3D bounding box
811
+ Randomly crop
812
+ local patches
813
+ 1
814
+ 2
815
+ 3
816
+ 4
817
+ 5
818
+ 6
819
+ 7
820
+ 8
821
+ 3D Global views
822
+ 3D Local views
823
+ Fig. 5. Illustration of sub-crop. Given a 3D local volume, we first randomly
824
+ crop two large patches, where an intersection over union (IoU) constraint
825
+ is applied to guarantee that two patches are partly overlapped. These
826
+ two large patches are considered as x1 and x2 in Fig. 2 and will be
827
+ passed to the siamese architecture to conduct the following multi-scale
828
+ pixel restoration and feature comparison tasks. To acquire local views,
829
+ we compute the minimum 3D bounding box of two large patches, after
830
+ which random crop is applied to extract multiple local patches. Finally,
831
+ we reshape these local patches to a fixed size and forward them to the
832
+ network to extract local representations.
833
+ 3.4
834
+ From multi-crop to sub-crop
835
+ Multi-crop [5] has been known as a helpful strategy to im-
836
+ prove SSL performance in natural images, which increases
837
+ the number of input views by sampling several standard
838
+ resolution crops and more low-resolution crops from the
839
+ original input. One key insight behind multi-crop is to
840
+ capture relations between parts of a scene or an object, while
841
+ low-resolution views ensure a controllable increase in the
842
+ computational cost.
843
+ When applied to medical images, multi-crop works well
844
+ in 2D X-ray data but leads to the non-convergence of the
845
+ model in 3D volume data (such as CT and MRI). After
846
+ careful investigation, we found the root of this problem
847
+ lies in the contradiction between the limited input size and
848
+ many candidate crops in three-dimensional space. Specif-
849
+
850
+ 7
851
+ ically, on the one hand, we cannot afford large-sized 3D
852
+ inputs because processing them with 3D deep models often
853
+ costs dramatic GPU memory. On the other hand, if we
854
+ overly reduce the size of 3D inputs, the sampled views
855
+ would be too dispersed to guarantee the model capture the
856
+ local-global associations.
857
+ To mitigate the above issue, we introduce sub-crop to
858
+ replace multi-crop in 3D medical images. The core idea of
859
+ sub-crop is straightforward: reducing the sampling space. As
860
+ illustrated in Fig. 5, sub-crop mainly consists of three steps:
861
+ (i) randomly crop two extensive global views with an IoU
862
+ constraint; (ii) find the minimum 3D bounding box over the
863
+ cropped global patches; (iii) randomly crop multiple local
864
+ patches within the 3D bounding box. There are two critical
865
+ operations in sub-crop: the constraint of IoU on global views
866
+ and the sampling of local patches within the minimum
867
+ bounding box. In practice, the first operation guarantees the
868
+ global-global association by ensuring the overlap between
869
+ large patches larger than a fixed threshold. The second
870
+ operation mitigates the disperse problem of local views and
871
+ helps the model to discover local-global relations.
872
+ 3.5
873
+ Overall training objective
874
+ After applying multi-crop/sub-crop to medical images, we
875
+ can acquire two global views {g1, g2} and ˆN local views
876
+ {l1, l2, ..., l ˆ
877
+ N}. For clarification, we denote the associated in-
878
+ puts in notations of loss functions. For instance, LC(g1, g2)
879
+ means we calculate LC on top of the extracted siamese
880
+ representations of two global views, where g1 and g2 can
881
+ be regarded as a pair of siamese images. At last, the overall
882
+ training objective of PCRLv2 can be formalized as follows:
883
+ LTotal(g1, g2, l1, ..., l ˆ
884
+ N) =LR(g1, g2) + LC(g1, g2)
885
+ +
886
+
887
+ m∈{1,2}
888
+ ˆ
889
+ N
890
+
891
+ k=1
892
+ LC(lk, gm).
893
+ (10)
894
+ There are three terms in LTotal: LR(g1, g2), LC(g1, g2), and
895
+
896
+ m∈{1,2}
897
+ � ˆ
898
+ N
899
+ k=1 LC(lk, gm). The first term is designed to
900
+ preserve pixel-level details in multi-scale learned repre-
901
+ sentations. The second term addresses the importance of
902
+ encoding multi-scale semantics into latent features. The last
903
+ term aims to capture the multi-scale global-local semantic
904
+ relations.
905
+ 3.6
906
+ Short discussion: PCRLv2 vs. PCRLv1
907
+ Simpler. PCRLv1 combines the context restoration and
908
+ comparative SSL via transformation-conditioned attention
909
+ and cross-model mixup. These two components make
910
+ the framework heavy, less intuitive, and not easy to
911
+ implement. Compared to PCRLv1, PCRLv2 exploits a
912
+ simpler yet more intuitive design to incorporate pixel-level
913
+ and semantic information via multi-scale learning. As
914
+ aforementioned, PCRLv2 can be formulated as a simple
915
+ multi-task optimization problem whose objective function
916
+ maximizes the preservation of multi-level information in
917
+ latent visual representations. These characteristics make it
918
+ easier for both implementation and potential expansion.
919
+ Faster. PCRLv1 makes heavy use of mixup (to both inputs
920
+ and features) in its implementation, which is found to
921
+ deliver performance gains. In PCRLv2, we eliminate mixup
922
+ strategies and cut the training time in half. In addition,
923
+ PCRLv2 requires less running memory in GPUs during
924
+ the training stage, making it more practical in real-world
925
+ scenarios.
926
+ 4
927
+ EXPERIMENTS
928
+ In this section, we first conduct thorough ablation studies to
929
+ investigate the influence of different modules in PCRLv2.
930
+ Then, we evaluate the effectiveness of PCRLv2 on both
931
+ 2D and 3D medical imaging tasks, including chest pathol-
932
+ ogy classification, pulmonary nodule detection, abdominal
933
+ organ segmentation, and brain tumor segmentation. For
934
+ model evaluation, we follow the pre-training (on source
935
+ data)→fine-tuning (on target data) protocol and employ two
936
+ settings, which are semi-supervised learning and transfer
937
+ learning. In the first setting, the source and target data come
938
+ from the same dataset. Specifically, we first pre-train the
939
+ model using all training data without labels, and then fine-
940
+ tune the pre-trained model with limited annotations. As for
941
+ transfer learning (the second setting), we pre-train and fine-
942
+ tune the model on different datasets. Different from semi-
943
+ supervised learning, we fine-tune the pre-trained model
944
+ with both limited and full annotations in transfer learning.
945
+ 4.1
946
+ Datasets
947
+ NIH ChestX-ray (2D) [38] is made up of 112,120 X-
948
+ ray
949
+ scans
950
+ from
951
+ 30,805
952
+ patients.
953
+ There
954
+ are
955
+ fourteen
956
+ different chest pathologies in NIH ChestX-ray, including
957
+ atelectasis, cardiomegaly, consolidation, edema, effusion,
958
+ emphysema, fibrosis, hernia, infiltration, mass, nodule,
959
+ pleural thickening, pneumonia, and pneumothorax. The
960
+ labels of radiographs were automatically extracted from
961
+ associated
962
+ radiology
963
+ reports
964
+ using
965
+ natural
966
+ language
967
+ process (NLP) techniques. We use NIH ChestX-ray in
968
+ semi-supervised learning in our experiments and treat it as
969
+ the target dataset in transfer learning.
970
+ CheXpert (2D) [21] involves 224,316 chest radiographs
971
+ from 65,240 patients for the presence of 14 common
972
+ chest
973
+ radiographic
974
+ observations:
975
+ no
976
+ finding,
977
+ enlarged
978
+ cardio, cardiomegaly, lung opacity, lung Lesion, edema,
979
+ consolidation,
980
+ pneumonia,
981
+ atelectasis,
982
+ pneumothorax,
983
+ pleural
984
+ effusion,
985
+ pleural
986
+ other,
987
+ fracture,
988
+ and
989
+ support
990
+ devices. Similar to NIH ChestX-ray, an NLP labeler was
991
+ developed to detect the presence of 14 observations in
992
+ radiology
993
+ reports
994
+ automatically.
995
+ In
996
+ practice, CheXpert
997
+ serves as the source data in transfer learning.
998
+ LUNA (3D) [33] was collected for the automatic detection
999
+ of pulmonary nodules, which involves 888 annotated
1000
+ thoracic computed tomography (CT) scans. LUNA is a
1001
+ cherry-picked subset of LIDC-IDRI [1], which excludes
1002
+ scans with a slice thickness greater than 3mm, inconsistent
1003
+ slice spacing, or missing slices. In the 888 scans, a total of
1004
+ 5,855 annotations were made by the radiologists, where
1005
+ only nodules ≥ 3mm are categorized as relevant lesions,
1006
+
1007
+ 8
1008
+ and at least one radiologist checks each nodule. On LUNA,
1009
+ we perform semi-supervised learning and transfer learning
1010
+ experiments. For transfer learning, LUNA is mainly used
1011
+ for self-supervised pre-training.
1012
+ LiTS (3D) [4] releases 131 abdominal CT Volumes and
1013
+ associated annotations for training and validation. There
1014
+ are two types of labels in LiTS: the liver and tumor. In this
1015
+ paper, we only utilize the ground truth masks of the liver
1016
+ to evaluate the effectiveness of various SSL algorithms. The
1017
+ task on LiTS is abdominal organ segmentation, where LiTS
1018
+ is used for fine-tuning in transfer learning.
1019
+ BraTS (3D) has been known as a series of challenges in brain
1020
+ tumor segmentation. In this paper, we perform experiments
1021
+ on the released 351 magnetic resonance imaging (MRI) scans
1022
+ of BraTS 2018. There are three classes in BraTS: whole tumor
1023
+ (WT), tumor core (TC), and enhancing tumor (ET). Similar
1024
+ to the role of LiTS, BraTS serves as the target data in transfer
1025
+ learning.
1026
+ 4.2
1027
+ Baselines
1028
+ A variety of SSL baselines are included in our extensive
1029
+ experiments, which can be roughly divided into three
1030
+ categories: 2D specific methods, 3D specific approaches,
1031
+ and generic (2D & 3D) methodologies. Details of baselines
1032
+ in each category are listed below.
1033
+ 2D specific SSL methodologies consist of ImageNet-based
1034
+ pre-training (IN) [14], Comparing to Learn (C2L) [48],
1035
+ and Simple Siamese Learning (SimSiam) [10]. IN is the
1036
+ most widely adopted pre-training methodology, which
1037
+ conducts supervised pre-training on one of the biggest
1038
+ natural image datasets, i.e., ImageNet [14]. C2L is a recently
1039
+ proposed SSL approach based on momentum contrast (i.e.,
1040
+ MoCov1 [17] and MoCov2 [9]). SimSiam is a simple siamese
1041
+ SSL framework that eliminates the barrier of negative
1042
+ samples in contrastive learning and the use of a momentum
1043
+ encoder in BYOL [15]. Besides, we compare PCLRv2 against
1044
+ SimSiam to highlight the significance of the preserved
1045
+ pixel-level information and multi-scale features.
1046
+ 3D specific SSL methodologies include Rubik’s cube++ [35]
1047
+ and 3D-CPC [34]. Rubik’s cube++ is the most recent SSL
1048
+ approach built on top of context restoration for 3D
1049
+ medical images. It adopts a volume-wise transformation
1050
+ for context permutation. In comparison, 3D-CPC is based
1051
+ on contrastive predictive encoding [19], a variation of
1052
+ contrastive learning, and demonstrates the most superior
1053
+ performance among different SSL approaches investigated
1054
+ in [34].
1055
+ Generic SSL methodologies involve train from scratch (TS),
1056
+ Model Genesis (MG) [49], TransVW [16], and PCRLv1 [47]
1057
+ (the conference version of our approach). MG resorts to ag-
1058
+ gressive augmentations to generate corrupted input images,
1059
+ based on which the model is asked to restore the original in-
1060
+ puts. TransVW improves MG by appending an intermediate
1061
+ classification head to encode anatomical patterns explicitly.
1062
+ PCRLv1 first proposes simultaneously preserving semantic
1063
+ and pixel-level information in SSL.
1064
+ 4.3
1065
+ Implementation details
1066
+ Dataset pre-processing for pre-training. On NIH ChestX-
1067
+ ray and CheXpert, each input image is resized to 224×224
1068
+ after random crop. On LUNA, we randomly crop a
1069
+ volume from the whole CT scan with a random size
1070
+ from {64×64×32, 96×96×64, 96×96×96, 112×112×64}.
1071
+ Each cropped volume is then resized to 64×64×32. Each
1072
+ voxel’s Hounsfield Unit (HU) in the crop is truncated to
1073
+ [-1000,1000]. If a voxel’s HU is lower than -150, we regard
1074
+ it as a background voxel. In practice, if over 85% voxels
1075
+ within a crop belong to the background, we would not use
1076
+ this crop in pre-training.
1077
+ Dataset pre-processing for fine-tuning. For NIH ChestX-
1078
+ ray and CheXpert, we follow the same pre-processing
1079
+ procedures as in the pre-training stage. On LUNA, we
1080
+ randomly crop a volume for each training iteration, and the
1081
+ size of each crop is 48×48×48. On LiTS, we first localize
1082
+ the liver and expand the target volume by 30 slices on
1083
+ each axis. After random crop, the size of each crop is
1084
+ 256×256×64. Unlike LUNA, we truncate the HU of each
1085
+ voxel to [-200, 200]. For BraTS, the size of each random crop
1086
+ is 112×112×112×4.
1087
+ Data augmentation and multi-crop/sub-crop. As shown in
1088
+ Fig. 2, there are two types of augmentations, i.e., global and
1089
+ local augmentations. Specifically, for 2D tasks, the global
1090
+ augmentation includes random crop, random horizontal
1091
+ flip, and random rotation. The local augmentation involves
1092
+ random grayscale, gaussian blur, and cutout. In comparison,
1093
+ for 3D tasks, the global augmentation consists of random
1094
+ flip and random affine. Local augmentation strategies are
1095
+ applied, including Gaussian blur, random noise, random
1096
+ gamma, and random swap. Note that all 3D augmentations
1097
+ are implemented following [30]. As for multi-crop in 2D
1098
+ tasks, we resort to the scale factor of random crop2 to
1099
+ generate global and local views. Specifically, we set the
1100
+ range of scale to [0.3, 1] to generate two global views. For six
1101
+ local views, the scale range is set to [0.05, 0.3]. Both global
1102
+ and local views are resized to 224×224. As for sub-crop
1103
+ in 3D tasks, we randomly sample two global views with
1104
+ a random size from {64×64×32, 96×96×64, 96×96×96,
1105
+ 112×112×64}. The IoU constraint (i.e., threshold) between
1106
+ two global views is 0.3. Then, we find the minimum
1107
+ bounding box of global views, from which six local views
1108
+ are randomly cropped, each with a random size from
1109
+ {8×8×8, 16×16×16, 32×32×16, 32×32×32}. After random
1110
+ crop, all 3D global views are resized to 64×64×32, while all
1111
+ local views are resized to 16×16×16.
1112
+ Training and evaluation details. We use stochastic gradient
1113
+ descent (SGD) with momentum as the default optimizer,
1114
+ where the momentum is set to 0.9. The initial learning rate
1115
+ is 1e-2, and we employ the cosine annealing strategy for
1116
+ learning rate decay. We set the weight decay to 1e-5. The
1117
+ number of training epochs is 240. The batch sizes of 2D pre-
1118
+ training and fine-tuning (on NIH ChestX-ray or CheXpert)
1119
+ are 256 and 512, respectively. As for 3D pre-training, the
1120
+ 2. https://pytorch.org/vision/main/generated/torchvision.
1121
+ transforms.RandomResizedCrop.html
1122
+
1123
+ 9
1124
+ 0
1125
+ 15
1126
+ Epochs
1127
+ -0.5
1128
+ -1.5
1129
+ Training MSE loss (log10)
1130
+ w/ skip
1131
+ w/o skip
1132
+ Fig. 6. Influence of skip connections in pixel restoration. We display the
1133
+ loss curve of mean square error (MSE) in the first 15 epoches.
1134
+ TABLE 1
1135
+ Impact of skip connections on chest pathology identification (NIH
1136
+ ChestX-ray), brain tumor segmentation (BraTS), and abdominal organ
1137
+ segmentation (LiTS). On NIH, We use 95% unlabeled training data for
1138
+ pre-training, while the rest 5% data with labels are used for fine-tuning.
1139
+ On BraTS and LiTS, we use 10% labeled data for fine-tuning.
1140
+ Datasets
1141
+ w/o skip
1142
+ w/ skip
1143
+ Gain
1144
+ NIH
1145
+ 76.6
1146
+ 75.4
1147
+ 1.2
1148
+ BraTS
1149
+ 73.0
1150
+ 71.5
1151
+ 1.5
1152
+ LiTS
1153
+ 79.0
1154
+ 77.6
1155
+ 1.4
1156
+ batch size (on LUNA) is 32. For 3D fine-tuning tasks, the
1157
+ batch sizes on LUNA, LiTS, and BraTS are 32, 4, and
1158
+ 4, respectively. The evaluation metric on NIH ChestX-ray,
1159
+ CheXpert, and LUNA is AUROC (Area Under the Receiver
1160
+ Operating Characteristics). For segmentation tasks on LiTS
1161
+ and BraTS, we use Dice similarity as the evaluation metric.
1162
+ We use 70%, 10%, and 20% of the whole dataset to build the
1163
+ training, validation, and test sets. In particular, for semi-
1164
+ supervised learning, we construct the pre-training set by
1165
+ removing a specific amount of data from the entire training
1166
+ set. At the same time, the remainder is used as the training
1167
+ set for fine-tuning. Binary cross-entropy loss is used for the
1168
+ fine-tuning of NIH ChestX-ray, CheXpert, and LUNA, while
1169
+ Dice loss is used for the fine-tuning of LiTS and BraTS.
1170
+ 4.4
1171
+ Ablation studies
1172
+ Impact of skip connections on pixel restoration. In Fig. 6,
1173
+ we present the mean square error (MSE) loss (cf. Eq. 4)
1174
+ curves during the training stage. We see that the MSE
1175
+ loss, with skip connections, decreases rapidly in the first
1176
+ 15 training epochs. In comparison, the proposed nsUNet
1177
+ (w/o skip) slows down the decreasing rate of MSE loss.
1178
+ These phenomena are consistent with the role of skip
1179
+ connections, which bridges the gap between low-level
1180
+ pixel details and high-level latent semantics. The existence
1181
+ of skip connections makes it easier to restore pixels
1182
+ by incorporating pixel-level details from low-level but
1183
+ high-resolution feature maps. However, nsUNet removes
1184
+ skip connections, avoiding shortcut solutions to context
1185
+ restoration. Although this design makes it harder to restore
1186
+ pixels (higher loss values in Fig. 6), it helps encode pixel-
1187
+ level information into high-level semantic representations.
1188
+ ℱ#
1189
+ ℱ$
1190
+ ℱ%
1191
+ ℱ"
1192
+ ℱ!
1193
+ ℱ#
1194
+ &
1195
+ ℱ$
1196
+ &
1197
+ ℱ%
1198
+ &
1199
+ ℱ"
1200
+ &
1201
+ ℱ"
1202
+ &
1203
+ (a) Pairwise
1204
+ ℱ#
1205
+ ℱ$
1206
+ ℱ%
1207
+ ℱ"
1208
+ ℱ!
1209
+ ℱ#
1210
+ &
1211
+ ℱ$
1212
+ &
1213
+ ℱ%
1214
+ &
1215
+ ℱ"
1216
+ &
1217
+ ℱ"
1218
+ &
1219
+ (b) Cross-scale
1220
+ Fig. 7. Two choices of how to conduct siamese feature comparison for
1221
+ multiple feature scales. Here, we primarily consider pairwise feature
1222
+ comparison and cross-scale feature comparison.
1223
+ Such advantage can be verified by the performance gains in
1224
+ Table 1, where removing skip connections brings over 1%
1225
+ improvement to chest pathology identification, brain tumor
1226
+ segmentation, and abdominal organ segmentation.
1227
+ How to conduct siamese feature comparison for multiple
1228
+ feature scales? We illustrate two intuitive choices in Fig. 7.
1229
+ Besides the adopted pairwise comparison manner (Fig. 7a),
1230
+ another obvious choice is to compare siamese features
1231
+ following a crossed way (a similar strategy was used
1232
+ in [3]). As shown in Fig. 7b, the cross-scale comparison
1233
+ aggressively compares siamese features across all feature
1234
+ scales. The motivation behind is to introduce multi-scale
1235
+ latent representations by coupling features across different
1236
+ scales. Table 2 reports the experimental results of pairwise
1237
+ and cross-scale siamese feature comparison. We find that
1238
+ cross-scale feature comparison slightly deteriorates the
1239
+ performance of semi-supervised pathology identification
1240
+ by 0.6 percents. The underlying reason might be that the
1241
+ features in each scale maintains distinct characteristics,
1242
+ and neglecting these discrepancies can lead to degenerate
1243
+ feature representations.
1244
+ Investigation of different modules in PCRLv2. In Table 3,
1245
+ we study and report the impact of different modules on
1246
+ the whole tumor (WT) and enhancing tumor (ET) classes of
1247
+ BraTS. Note that in practice, most instances of WT are much
1248
+ larger than instances from ET, making ET instances harder
1249
+ to segment. Besides, we also present the transfer learning
1250
+ results on NIH ChestX-ray.
1251
+ TABLE 2
1252
+ Results of pairwise and crossed siamese feature comparison
1253
+ (semi-supervised learning on NIH ChestX-ray). The ratio of unlabeled
1254
+ to labeled data is 9.5:0.5.
1255
+ Pairwise
1256
+ Crossed [3]
1257
+ Gain
1258
+ Mean AUROC
1259
+ 76.6
1260
+ 76.0
1261
+ 0.6
1262
+ First of all, we investigate the influence of pixel restora-
1263
+ tion (row 0) and feature comparison (row 1), respectively.
1264
+ We directly reconstruct the full resolution uncorrupted im-
1265
+ ages for the pixel restoration task while siamese feature
1266
+ comparison is conducted on the last-layer output of the
1267
+ encoder. Comparing row 0 with row 1, we see that the
1268
+ context restoration task is more advantageous in segmenta-
1269
+ tion of small tumor regions (i.e., ET) while the comparative
1270
+ SSL is more capable of dealing with large tumor regions
1271
+ (i.e., WT) and chest pathologies. Such comparison shows
1272
+
1273
+ 10
1274
+ TABLE 3
1275
+ Impact of different modules in PCRLv2. Res. and Comp. denote the tasks of pixel restoration and feature comparison, respectively. S (N) means
1276
+ there are N scales included. MC and SC stand for the multi-crop and proposed sub-crop strategies, respectively. WT and ET denote classes of the
1277
+ whole tumor and enhancing tumor in BraTS, respectively. In most cases, instances from WT are much larger (in size) than those of ET. We
1278
+ performed these experiments by first using LUNA for self-supervised pre-training, and then we fine-tune the pre-trained model on BraTS using
1279
+ 10% labeled data. NIH denotes the transfer learning on chest pathology identification, where we use CheXpert for pre-training and fine-tune the
1280
+ pre-trained model with 50% labeled data from NIH ChestX-ray.
1281
+ #
1282
+ Res.
1283
+ Comp.
1284
+ S (3)
1285
+ S (5)
1286
+ MC
1287
+ SC
1288
+ WT (BraTS)
1289
+ ET (BraTS)
1290
+ NIH
1291
+ 0
1292
+
1293
+ 74.2
1294
+ 64.9
1295
+ 78.2
1296
+ 1
1297
+
1298
+ 76.4
1299
+ 63.8
1300
+ 78.5
1301
+ 2
1302
+
1303
+
1304
+ 76.2
1305
+ 64.6
1306
+ 80.9
1307
+ 3
1308
+
1309
+
1310
+
1311
+ 76.9
1312
+ 66.1
1313
+ 81.5
1314
+ 4
1315
+
1316
+
1317
+
1318
+ 77.2
1319
+ 66.8
1320
+ 82.0
1321
+ 5
1322
+
1323
+
1324
+
1325
+
1326
+ fail
1327
+ fail
1328
+ 82.5
1329
+ 6
1330
+
1331
+
1332
+
1333
+
1334
+ 77.7
1335
+ 67.2
1336
+ 82.7
1337
+ that semantic information preservation may be more helpful
1338
+ to the detection of large objects, while segmenting small
1339
+ objects requires the incorporation of pixel-level information.
1340
+ In row 2, we can already acquire noticeable performance
1341
+ gains by directly combining pixel restoration and feature
1342
+ comparison.
1343
+ Next, we show that multi-scale representations benefit
1344
+ both pixel restoration and feature comparison tasks. By
1345
+ conducting both tasks on 3 scales, we observe a 0.7-percent
1346
+ improvement on WT, a 1.5-percent gain on ET, and a
1347
+ 0.6-percent improvement on chest pathology classification.
1348
+ These results show that introducing multiple scales is more
1349
+ helpful to the segmentation of small regions. Moreover,
1350
+ by increasing the number of scales from 3 to 5, we can
1351
+ improve the accuracy of all three tasks consistently. Not
1352
+ surprisingly, ET benefits the most from the introduction of
1353
+ multiple scales, indicating the necessity of utilizing multi-
1354
+ scale representations in medical image segmentation.
1355
+ Last but not the least, we investigate the significance of
1356
+ multi-crop (row 4) and sub-crop (row 5). We empirically
1357
+ found that directly applying multi-crop to 3D medical vol-
1358
+ umes leads to the failure of model training. The underlying
1359
+ reason might be that it is difficult for cropped global and
1360
+ local views to maintain clear spatial relations in the 3D
1361
+ space as in the 2D space. In contrast, sub-crop can provide
1362
+ consistent performance gains on both types of tumor regions
1363
+ by successfully preserving the spatial relations in latent
1364
+ representations. When applying sub-crop to 2D X-rays,
1365
+ we observe a marginal improvement over multi-crop. The
1366
+ underlying reason is that sub-crop is proposed to handle
1367
+ dispersed sampled views in a 3D space to guarantee the
1368
+ model captures local-global relations. However, in a 2D
1369
+ space, the sampled views usually (partly) overlap.
1370
+ 4.5
1371
+ Semi-supervised chest pathology identification
1372
+ Table 4 presents the experimental results of applying semi-
1373
+ supervised learning on NIH ChestX-ray. Specifically, we use
1374
+ a specific amount of the training set (denoted as the labeling
1375
+ ratio in Table 4) as labeled data while the remaining training
1376
+ data is used for self-supervised pre-training.
1377
+ From Table 4, we see that self-supervised pre-training
1378
+ can dramatically boost the performance compared to train
1379
+ from scratch (TS), which verify the necessity of conduct-
1380
+ ing pre-training in medical imaging. Comparing MG with
1381
+ TransVW, they show similar performance in different label-
1382
+ ing ratios. Such comparison is easy to explain as TransVW
1383
+ is built upon MG, and both are based on context restora-
1384
+ tion. TransVW performs slightly better than MG, as it
1385
+ incorporates an additional classification head to encode
1386
+ more semantics. Compared to context restoration based
1387
+ methods, comparative methodologies (C2L and SimSiam)
1388
+ display better overall and class-specific results, especially
1389
+ in small labeling ratios. The underlying reason might be
1390
+ that semantic information is more critical than pixel-level
1391
+ information in chest pathology detection. As for C2L and
1392
+ SimSiam, C2L performs better when the amount of labeled
1393
+ data is quite limited. However, SimSiam gradually produces
1394
+ better diagnosis results as the labeling ratio increases.
1395
+ After incorporating the semantic, pixel-level, and scale
1396
+ information into a unified framework, PCRLv2 outperforms
1397
+ various SSL baselines in different labeling ratios signifi-
1398
+ cantly. It surpasses the previous conference version by clear
1399
+ margins, i.e., PCRLv1. Particularly, PCRLv2 seems to have
1400
+ more advantages in small labeling ratios. For instance, when
1401
+ the labeling ratio is 5%, PCRLv2 outperforms PCRLv1 by
1402
+ 2.5 percents on average, which verifies the significance of
1403
+ multi-scale latent representations.
1404
+ 4.6
1405
+ Semi-supervised pulmonary nodule detection
1406
+ In Table 5, we report the experimental results of semi-
1407
+ supervised pulmonary nodule detection. Interestingly, we
1408
+ observe narrowed performance gaps between TS and SSL
1409
+ baselines than those reported in Table 4. One possible ex-
1410
+ planation is that the task of detecting pulmonary nodules
1411
+ is less sensitive to the amount of labeled data. Among
1412
+ all SSL baselines, Cube++ gives better performance when
1413
+ utilizing small amounts of labeled data, while 3D-CPC is
1414
+ more advantageous in large labeling ratios. In addition, we
1415
+ see TransVW quickly catching up with MG and Cube++ as
1416
+ the labeling ratio increases.
1417
+ PCRLv1 outperforms previous SSL approaches in dif-
1418
+ ferent labeling ratios by large margins. After incorporat-
1419
+ ing multi-scale latent representations, PCRLv2 consistently
1420
+ surpasses PCRLv1 in a range of labeling ratios. When the
1421
+ baseline SSL methods show similar performance as the
1422
+ labeling ratio increases, PCRLv2 can still provide impressive
1423
+ improvements over PCRLv1 and previous SSL approaches.
1424
+
1425
+ 11
1426
+ TABLE 4
1427
+ Semi-supervised chest pathology identification (on NIH ChestX-ray). The labeling ratio denotes the amount of data with labels in the training set
1428
+ that is used for fine-tuning while the remaining data in the training set is used for self-supervised pre-training. The best results are bolded.
1429
+ Labeling ratio
1430
+ Methodology
1431
+ Mean
1432
+ Atelectasis
1433
+ Cardiomegaly
1434
+ Effusion
1435
+ Infiltration
1436
+ Mass
1437
+ Nodule
1438
+ Pneumonia
1439
+ Pneumothorax
1440
+ Consolidation
1441
+ Edema
1442
+ Emphysema
1443
+ Fibrosis
1444
+ Pleural Thick.
1445
+ Hernia
1446
+ 5%
1447
+ TS
1448
+ 61.8
1449
+ 58.8
1450
+ 72.0
1451
+ 68.8
1452
+ 51.5
1453
+ 63.8
1454
+ 49.2
1455
+ 57.4
1456
+ 67.4
1457
+ 61.5
1458
+ 71.0
1459
+ 62.7
1460
+ 58.1
1461
+ 60.0
1462
+ 63.1
1463
+ MG [49]
1464
+ 66.4
1465
+ 63.4
1466
+ 74.1
1467
+ 72.9
1468
+ 53.5
1469
+ 67.2
1470
+ 54.3
1471
+ 59.9
1472
+ 71.3
1473
+ 66.5
1474
+ 77.0
1475
+ 65.8
1476
+ 64.5
1477
+ 62.8
1478
+ 76.2
1479
+ TransVW [16]
1480
+ 66.5
1481
+ 64.2
1482
+ 72.9
1483
+ 72.2
1484
+ 54.8
1485
+ 69.4
1486
+ 55.7
1487
+ 59.6
1488
+ 71.0
1489
+ 64.8
1490
+ 77.4
1491
+ 66.6
1492
+ 63.6
1493
+ 62.8
1494
+ 75.6
1495
+ C2L [48]
1496
+ 71.7
1497
+ 69.9
1498
+ 77.9
1499
+ 76.2
1500
+ 59.1
1501
+ 73.4
1502
+ 60.0
1503
+ 64.5
1504
+ 76.2
1505
+ 71.4
1506
+ 80.3
1507
+ 76.1
1508
+ 69.9
1509
+ 68.4
1510
+ 80.4
1511
+ SimSiam [10]
1512
+ 71.7
1513
+ 68.9
1514
+ 79.3
1515
+ 77.8
1516
+ 58.7
1517
+ 73.0
1518
+ 61.0
1519
+ 65.4
1520
+ 76.2
1521
+ 72.1
1522
+ 81.7
1523
+ 75.1
1524
+ 69.6
1525
+ 68.1
1526
+ 76.8
1527
+ PCRLv1 [47]
1528
+ 74.1
1529
+ 70.1
1530
+ 80.3
1531
+ 79.3
1532
+ 61.8
1533
+ 76.8
1534
+ 64.6
1535
+ 68.6
1536
+ 77.2
1537
+ 72.8
1538
+ 83.7
1539
+ 77.4
1540
+ 71.3
1541
+ 72.7
1542
+ 80.8
1543
+ PCRLv2
1544
+ 76.6
1545
+ 75.7
1546
+ 81.0
1547
+ 80.3
1548
+ 64.0
1549
+ 76.8
1550
+ 68.7
1551
+ 70.7
1552
+ 83.2
1553
+ 77.5
1554
+ 87.8
1555
+ 79.2
1556
+ 72.5
1557
+ 73.2
1558
+ 81.8
1559
+ 10%
1560
+ TS
1561
+ 68.1
1562
+ 65.8
1563
+ 77.6
1564
+ 74.4
1565
+ 57.1
1566
+ 69.4
1567
+ 54.8
1568
+ 63.0
1569
+ 72.9
1570
+ 68.3
1571
+ 78.8
1572
+ 68.2
1573
+ 64.3
1574
+ 66.4
1575
+ 72.5
1576
+ MG [49]
1577
+ 70.0
1578
+ 67.1
1579
+ 78.1
1580
+ 76.1
1581
+ 57.2
1582
+ 72.8
1583
+ 57.5
1584
+ 63.3
1585
+ 75.5
1586
+ 70.9
1587
+ 79.5
1588
+ 68.8
1589
+ 67.4
1590
+ 68.0
1591
+ 77.6
1592
+ TransVW [16]
1593
+ 70.2
1594
+ 66.6
1595
+ 78.9
1596
+ 74.9
1597
+ 58.4
1598
+ 71.2
1599
+ 59.5
1600
+ 64.8
1601
+ 72.6
1602
+ 70.4
1603
+ 79.4
1604
+ 70.7
1605
+ 67.2
1606
+ 68.3
1607
+ 79.5
1608
+ C2L [48]
1609
+ 74.1
1610
+ 72.3
1611
+ 81.7
1612
+ 79.9
1613
+ 60.2
1614
+ 74.6
1615
+ 62.7
1616
+ 67.6
1617
+ 78.7
1618
+ 73.9
1619
+ 83.5
1620
+ 78.2
1621
+ 72.8
1622
+ 69.8
1623
+ 81.4
1624
+ SimSiam [10]
1625
+ 74.0
1626
+ 71.2
1627
+ 81.4
1628
+ 78.9
1629
+ 60.2
1630
+ 75.5
1631
+ 63.2
1632
+ 67.3
1633
+ 78.7
1634
+ 73.2
1635
+ 83.5
1636
+ 77.7
1637
+ 72.5
1638
+ 71.8
1639
+ 80.8
1640
+ PCRLv1 [47]
1641
+ 76.2
1642
+ 73.6
1643
+ 82.9
1644
+ 81.2
1645
+ 64.7
1646
+ 77.1
1647
+ 66.7
1648
+ 69.7
1649
+ 79.8
1650
+ 74.5
1651
+ 86.9
1652
+ 78.8
1653
+ 75.6
1654
+ 74.2
1655
+ 81.1
1656
+ PCRLv2
1657
+ 78.2
1658
+ 77.2
1659
+ 84.3
1660
+ 84.4
1661
+ 67.4
1662
+ 77.5
1663
+ 68.9
1664
+ 71.6
1665
+ 84.4
1666
+ 77.8
1667
+ 89.0
1668
+ 79.3
1669
+ 76.1
1670
+ 74.0
1671
+ 82.4
1672
+ 20%
1673
+ TS
1674
+ 71.5
1675
+ 68.9
1676
+ 80.7
1677
+ 77.5
1678
+ 60.2
1679
+ 73.6
1680
+ 58.7
1681
+ 66.2
1682
+ 76.1
1683
+ 71.7
1684
+ 82.9
1685
+ 72.2
1686
+ 69.0
1687
+ 68.7
1688
+ 74.7
1689
+ MG [49]
1690
+ 73.9
1691
+ 71.9
1692
+ 83.0
1693
+ 80.0
1694
+ 62.3
1695
+ 75.2
1696
+ 62.2
1697
+ 67.5
1698
+ 79.0
1699
+ 73.3
1700
+ 83.6
1701
+ 73.4
1702
+ 71.0
1703
+ 70.6
1704
+ 81.4
1705
+ TransVW [16]
1706
+ 74.3
1707
+ 71.6
1708
+ 82.5
1709
+ 80.1
1710
+ 62.3
1711
+ 76.7
1712
+ 62.8
1713
+ 69.2
1714
+ 78.2
1715
+ 73.5
1716
+ 83.8
1717
+ 75.4
1718
+ 72.2
1719
+ 71.2
1720
+ 80.3
1721
+ C2L [48]
1722
+ 76.4
1723
+ 74.2
1724
+ 83.9
1725
+ 81.7
1726
+ 63.8
1727
+ 77.3
1728
+ 64.7
1729
+ 70.3
1730
+ 81.5
1731
+ 75.5
1732
+ 86.0
1733
+ 80.2
1734
+ 75.2
1735
+ 73.4
1736
+ 81.8
1737
+ SimSiam [10]
1738
+ 76.5
1739
+ 73.8
1740
+ 84.0
1741
+ 81.4
1742
+ 63.2
1743
+ 78.2
1744
+ 64.7
1745
+ 69.6
1746
+ 82.1
1747
+ 76.2
1748
+ 86.4
1749
+ 80.7
1750
+ 75.0
1751
+ 73.9
1752
+ 81.7
1753
+ PCRLv1 [47]
1754
+ 78.8
1755
+ 75.4
1756
+ 86.2
1757
+ 83.6
1758
+ 65.1
1759
+ 79.9
1760
+ 69.6
1761
+ 72.0
1762
+ 82.3
1763
+ 79.9
1764
+ 88.3
1765
+ 82.6
1766
+ 76.5
1767
+ 75.9
1768
+ 81.9
1769
+ PCRLv2
1770
+ 79.9
1771
+ 78.1
1772
+ 87.2
1773
+ 85.9
1774
+ 68.2
1775
+ 80.5
1776
+ 69.9
1777
+ 72.5
1778
+ 85.3
1779
+ 80.4
1780
+ 89.2
1781
+ 83.1
1782
+ 77.5
1783
+ 77.0
1784
+ 83.5
1785
+ 30%
1786
+ TS
1787
+ 73.4
1788
+ 70.6
1789
+ 81.9
1790
+ 79.1
1791
+ 61.6
1792
+ 75.5
1793
+ 60.7
1794
+ 68.8
1795
+ 78.3
1796
+ 72.7
1797
+ 84.3
1798
+ 74.1
1799
+ 70.3
1800
+ 70.9
1801
+ 78.9
1802
+ MG [49]
1803
+ 76.1
1804
+ 74.3
1805
+ 84.4
1806
+ 82.1
1807
+ 63.6
1808
+ 78.3
1809
+ 64.4
1810
+ 69.6
1811
+ 81.2
1812
+ 75.8
1813
+ 85.6
1814
+ 75.9
1815
+ 73.6
1816
+ 73.6
1817
+ 82.8
1818
+ TransVW [16]
1819
+ 76.7
1820
+ 74.9
1821
+ 84.1
1822
+ 81.9
1823
+ 64.9
1824
+ 79.0
1825
+ 65.3
1826
+ 70.9
1827
+ 80.3
1828
+ 76.2
1829
+ 86.5
1830
+ 78.6
1831
+ 74.5
1832
+ 74.2
1833
+ 82.1
1834
+ C2L [48]
1835
+ 77.5
1836
+ 74.3
1837
+ 84.8
1838
+ 82.6
1839
+ 64.6
1840
+ 78.3
1841
+ 66.3
1842
+ 71.5
1843
+ 83.0
1844
+ 76.8
1845
+ 87.6
1846
+ 81.3
1847
+ 76.5
1848
+ 74.4
1849
+ 82.9
1850
+ SimSiam [10]
1851
+ 78.0
1852
+ 75.4
1853
+ 85.1
1854
+ 82.9
1855
+ 65.0
1856
+ 79.4
1857
+ 67.0
1858
+ 71.4
1859
+ 83.4
1860
+ 77.4
1861
+ 87.8
1862
+ 82.8
1863
+ 76.1
1864
+ 75.5
1865
+ 82.7
1866
+ PCRLv1 [47]
1867
+ 79.0
1868
+ 75.5
1869
+ 86.6
1870
+ 83.8
1871
+ 65.9
1872
+ 80.7
1873
+ 70.2
1874
+ 72.8
1875
+ 82.9
1876
+ 80.4
1877
+ 88.9
1878
+ 83.3
1879
+ 76.6
1880
+ 76.5
1881
+ 81.9
1882
+ PCRLv2
1883
+ 81.1
1884
+ 78.4
1885
+ 87.6
1886
+ 86.6
1887
+ 69.6
1888
+ 82.8
1889
+ 72.0
1890
+ 74.0
1891
+ 86.2
1892
+ 81.0
1893
+ 89.9
1894
+ 84.4
1895
+ 79.5
1896
+ 79.0
1897
+ 84.6
1898
+ 40%
1899
+ TS
1900
+ 75.4
1901
+ 72.6
1902
+ 83.6
1903
+ 81.5
1904
+ 62.9
1905
+ 77.3
1906
+ 63.3
1907
+ 70.1
1908
+ 80.3
1909
+ 74.9
1910
+ 85.5
1911
+ 76.4
1912
+ 72.5
1913
+ 73.0
1914
+ 81.8
1915
+ MG [49]
1916
+ 77.3
1917
+ 75.4
1918
+ 86.0
1919
+ 83.3
1920
+ 65.1
1921
+ 79.0
1922
+ 65.1
1923
+ 70.8
1924
+ 82.1
1925
+ 77.0
1926
+ 87.3
1927
+ 76.7
1928
+ 74.8
1929
+ 74.9
1930
+ 83.5
1931
+ TransVW [16]
1932
+ 77.6
1933
+ 75.0
1934
+ 85.1
1935
+ 82.7
1936
+ 65.2
1937
+ 79.7
1938
+ 66.5
1939
+ 72.0
1940
+ 81.0
1941
+ 76.7
1942
+ 87.2
1943
+ 79.2
1944
+ 75.5
1945
+ 76.5
1946
+ 83.7
1947
+ C2L [48]
1948
+ 79.0
1949
+ 76.0
1950
+ 86.1
1951
+ 84.3
1952
+ 66.0
1953
+ 80.0
1954
+ 67.9
1955
+ 72.5
1956
+ 84.1
1957
+ 78.5
1958
+ 88.5
1959
+ 83.7
1960
+ 77.9
1961
+ 76.6
1962
+ 83.8
1963
+ SimSiam [10]
1964
+ 79.4
1965
+ 76.7
1966
+ 86.7
1967
+ 84.7
1968
+ 67.0
1969
+ 80.9
1970
+ 69.0
1971
+ 73.1
1972
+ 84.4
1973
+ 78.9
1974
+ 88.9
1975
+ 83.5
1976
+ 77.7
1977
+ 76.6
1978
+ 83.4
1979
+ PCRLv1 [47]
1980
+ 79.9
1981
+ 76.7
1982
+ 87.1
1983
+ 84.9
1984
+ 67.1
1985
+ 82.7
1986
+ 72.2
1987
+ 73.3
1988
+ 83.6
1989
+ 80.6
1990
+ 89.2
1991
+ 83.8
1992
+ 77.3
1993
+ 76.9
1994
+ 83.2
1995
+ PCRLv2
1996
+ 81.5
1997
+ 78.7
1998
+ 87.8
1999
+ 87.0
2000
+ 69.8
2001
+ 83.2
2002
+ 72.5
2003
+ 74.7
2004
+ 86.3
2005
+ 81.2
2006
+ 90.2
2007
+ 84.9
2008
+ 80.0
2009
+ 79.4
2010
+ 85.0
2011
+ TABLE 5
2012
+ Semi-supervised pulmonary nodule detection (on LUNA). The labeling
2013
+ ratio indicates how much data from the training set with labels is utilized
2014
+ for fine-tuning while the rest of the data is used for pre-training. Best
2015
+ results are bolded.
2016
+ Methodology
2017
+ Labeling ratio
2018
+ 10%
2019
+ 20%
2020
+ 30%
2021
+ 40%
2022
+ TS
2023
+ 78.4
2024
+ 83.0
2025
+ 85.7
2026
+ 87.5
2027
+ MG [49]
2028
+ 80.2
2029
+ 85.0
2030
+ 87.5
2031
+ 90.3
2032
+ TransVW [16]
2033
+ 79.3
2034
+ 84.5
2035
+ 87.9
2036
+ 90.5
2037
+ Cube++ [35]
2038
+ 81.4
2039
+ 85.2
2040
+ 87.9
2041
+ 90.0
2042
+ 3D-CPC [34]
2043
+ 80.2
2044
+ 85.2
2045
+ 88.3
2046
+ 90.6
2047
+ PCRLv1 [47]
2048
+ 84.4
2049
+ 87.5
2050
+ 89.8
2051
+ 92.2
2052
+ PCRLv2
2053
+ 85.5
2054
+ 88.3
2055
+ 90.3
2056
+ 93.1
2057
+ 4.7
2058
+ Transfer learning on chest pathology identification
2059
+ In Table 6, we validate the transferable ability of visual
2060
+ representations provided by different pre-training method-
2061
+ ologies. Specifically, we compare PCRLv2 against train from
2062
+ scratch, ImageNet-based pre-training (IN), different SSL
2063
+ baselines, and PCRLv1.
2064
+ Comparing MG/TransVW with IN, we see context
2065
+ restoration based SSL maintains the limited transferable
2066
+ ability. This phenomenon becomes more apparent when the
2067
+ target domain has quite limited annotations. The underlying
2068
+ reason is that semantic information plays a crucial role in
2069
+ transfer learning. In contrast, the significant performance
2070
+ gains brought by C2L and SimSiam again verify the effec-
2071
+ tiveness of comparative SSL. C2L and SimSiam still cannot
2072
+ outperform IN by significant margins, especially when con-
2073
+ sidering that IN is more advantageous when the labeling
2074
+ ratio is 10%.
2075
+ After integrating the benefits of context restoration
2076
+ based and comparative SSL, PCRLv1 is already capable of
2077
+ outperforming previous SSL methodologies by observable
2078
+ margins. Furthermore, by exploiting multi-scale semantic
2079
+ and pixel-level information, PCRLv2 achieves consistent
2080
+ improvements over PCRLv1 in overall and class-specific
2081
+ results in different labeling ratios.
2082
+ 4.8
2083
+ Transfer learning on brain tumor segmentation
2084
+ We report the experimental results of applying transfer
2085
+ learning to brain tumor segmentation in Table 7, where
2086
+ we use LUNA dataset for self-supervised pre-training and
2087
+ fine-tune the pre-trained model with different amounts of
2088
+ labeled data.
2089
+
2090
+ 12
2091
+ TABLE 6
2092
+ Transfer learning on chest pathology identification. We pre-train the model using data from CheXpert (without labels). Then, we fine-tune the
2093
+ pre-trained model on NIH ChestX-ray with different amounts of labeled data (denotes as different labeling ratios). The best results are bolded.
2094
+ Labeling ratio
2095
+ Methodology
2096
+ Mean
2097
+ Atelectasis
2098
+ Cardiomegaly
2099
+ Effusion
2100
+ Infiltration
2101
+ Mass
2102
+ Nodule
2103
+ Pneumonia
2104
+ Pneumothorax
2105
+ Consolidation
2106
+ Edema
2107
+ Emphysema
2108
+ Fibrosis
2109
+ Pleural Thick.
2110
+ Hernia
2111
+ 10%
2112
+ TS
2113
+ 68.1
2114
+ 67.6
2115
+ 63.3
2116
+ 76.8
2117
+ 57.5
2118
+ 71.5
2119
+ 61.8
2120
+ 64.2
2121
+ 76.2
2122
+ 69.8
2123
+ 80.2
2124
+ 72.4
2125
+ 62.8
2126
+ 68.0
2127
+ 61.1
2128
+ IN [28]
2129
+ 73.5
2130
+ 73.3
2131
+ 68.7
2132
+ 81.6
2133
+ 63.0
2134
+ 76.6
2135
+ 67.3
2136
+ 70.0
2137
+ 81.3
2138
+ 75.6
2139
+ 85.9
2140
+ 78.5
2141
+ 68.6
2142
+ 72.5
2143
+ 65.9
2144
+ MG [49]
2145
+ 70.1
2146
+ 69.9
2147
+ 65.6
2148
+ 79.2
2149
+ 59.4
2150
+ 72.9
2151
+ 64.3
2152
+ 67.0
2153
+ 77.9
2154
+ 72.0
2155
+ 82.3
2156
+ 75.8
2157
+ 65.9
2158
+ 69.6
2159
+ 59.4
2160
+ TransVW [16]
2161
+ 69.7
2162
+ 69.4
2163
+ 64.3
2164
+ 78.2
2165
+ 59.5
2166
+ 72.6
2167
+ 63.1
2168
+ 67.2
2169
+ 77.2
2170
+ 70.9
2171
+ 83.0
2172
+ 75.3
2173
+ 65.8
2174
+ 68.9
2175
+ 60.2
2176
+ C2L [48]
2177
+ 73.1
2178
+ 72.5
2179
+ 68.0
2180
+ 81.3
2181
+ 62.4
2182
+ 75.8
2183
+ 67.2
2184
+ 70.2
2185
+ 80.6
2186
+ 74.8
2187
+ 85.4
2188
+ 78.4
2189
+ 68.3
2190
+ 72.2
2191
+ 66.1
2192
+ SimSiam [10]
2193
+ 72.5
2194
+ 71.9
2195
+ 67.5
2196
+ 81.2
2197
+ 61.7
2198
+ 75.9
2199
+ 66.6
2200
+ 69.6
2201
+ 79.8
2202
+ 74.2
2203
+ 84.8
2204
+ 77.6
2205
+ 67.7
2206
+ 71.8
2207
+ 64.5
2208
+ PCRLv1 [47]
2209
+ 75.8
2210
+ 75.4
2211
+ 70.6
2212
+ 84.2
2213
+ 65.5
2214
+ 78.9
2215
+ 69.6
2216
+ 72.7
2217
+ 83.5
2218
+ 77.6
2219
+ 88.5
2220
+ 80.8
2221
+ 71.3
2222
+ 74.8
2223
+ 67.6
2224
+ PCRLv2
2225
+ 77.2
2226
+ 76.8
2227
+ 72.0
2228
+ 85.6
2229
+ 66.8
2230
+ 80.2
2231
+ 71.0
2232
+ 74.0
2233
+ 84.8
2234
+ 78.9
2235
+ 89.8
2236
+ 82.2
2237
+ 72.6
2238
+ 76.2
2239
+ 69.7
2240
+ 20%
2241
+ TS
2242
+ 71.4
2243
+ 71.8
2244
+ 73.1
2245
+ 78.4
2246
+ 59.6
2247
+ 72.5
2248
+ 64.5
2249
+ 66.6
2250
+ 77.7
2251
+ 71.7
2252
+ 82.0
2253
+ 75.5
2254
+ 69.8
2255
+ 68.9
2256
+ 68.2
2257
+ IN [14]
2258
+ 76.2
2259
+ 75.9
2260
+ 78.3
2261
+ 82.9
2262
+ 64.2
2263
+ 77.8
2264
+ 68.8
2265
+ 70.7
2266
+ 83.0
2267
+ 76.4
2268
+ 87.2
2269
+ 80.0
2270
+ 75.3
2271
+ 73.9
2272
+ 73.1
2273
+ MG [49]
2274
+ 73.8
2275
+ 73.9
2276
+ 75.4
2277
+ 80.2
2278
+ 61.9
2279
+ 74.9
2280
+ 66.5
2281
+ 68.3
2282
+ 80.0
2283
+ 74.0
2284
+ 85.1
2285
+ 78.1
2286
+ 72.8
2287
+ 71.5
2288
+ 71.3
2289
+ TransVW [16]
2290
+ 73.8
2291
+ 73.0
2292
+ 75.5
2293
+ 80.1
2294
+ 62.3
2295
+ 75.6
2296
+ 66.7
2297
+ 68.6
2298
+ 80.2
2299
+ 74.0
2300
+ 85.2
2301
+ 77.5
2302
+ 72.9
2303
+ 71.5
2304
+ 69.4
2305
+ C2L [48]
2306
+ 77.0
2307
+ 76.5
2308
+ 78.9
2309
+ 83.4
2310
+ 65.0
2311
+ 78.6
2312
+ 69.8
2313
+ 71.8
2314
+ 83.5
2315
+ 77.2
2316
+ 88.1
2317
+ 80.8
2318
+ 76.0
2319
+ 74.2
2320
+ 73.5
2321
+ SimSiam [10]
2322
+ 76.6
2323
+ 76.6
2324
+ 78.7
2325
+ 83.3
2326
+ 64.6
2327
+ 77.9
2328
+ 69.2
2329
+ 71.6
2330
+ 83.1
2331
+ 76.9
2332
+ 87.8
2333
+ 80.5
2334
+ 75.5
2335
+ 73.8
2336
+ 73.6
2337
+ PCRLv1 [47]
2338
+ 77.5
2339
+ 77.3
2340
+ 79.7
2341
+ 84.3
2342
+ 65.7
2343
+ 78.9
2344
+ 70.3
2345
+ 72.8
2346
+ 83.8
2347
+ 77.6
2348
+ 88.6
2349
+ 81.1
2350
+ 76.5
2351
+ 74.8
2352
+ 74.3
2353
+ PCRLv2
2354
+ 79.4
2355
+ 79.0
2356
+ 81.3
2357
+ 85.9
2358
+ 67.3
2359
+ 80.8
2360
+ 72.1
2361
+ 74.0
2362
+ 86.0
2363
+ 79.4
2364
+ 90.3
2365
+ 83.1
2366
+ 78.4
2367
+ 76.7
2368
+ 76.6
2369
+ 30%
2370
+ TS
2371
+ 73.5
2372
+ 71.7
2373
+ 79.7
2374
+ 79.9
2375
+ 60.5
2376
+ 76.5
2377
+ 68.4
2378
+ 66.8
2379
+ 79.2
2380
+ 72.8
2381
+ 83.4
2382
+ 76.9
2383
+ 71.4
2384
+ 70.5
2385
+ 71.3
2386
+ IN [14]
2387
+ 78.5
2388
+ 77.2
2389
+ 84.6
2390
+ 84.3
2391
+ 66.2
2392
+ 80.8
2393
+ 73.0
2394
+ 72.3
2395
+ 84.0
2396
+ 78.0
2397
+ 88.5
2398
+ 82.0
2399
+ 76.8
2400
+ 75.3
2401
+ 76.0
2402
+ MG [49]
2403
+ 75.6
2404
+ 74.1
2405
+ 81.8
2406
+ 81.0
2407
+ 63.3
2408
+ 77.9
2409
+ 70.1
2410
+ 69.0
2411
+ 80.9
2412
+ 74.8
2413
+ 85.4
2414
+ 79.7
2415
+ 73.6
2416
+ 72.6
2417
+ 74.2
2418
+ TransVW [16]
2419
+ 75.7
2420
+ 74.8
2421
+ 81.4
2422
+ 81.0
2423
+ 63.6
2424
+ 77.7
2425
+ 69.9
2426
+ 69.8
2427
+ 80.9
2428
+ 75.4
2429
+ 86.0
2430
+ 79.3
2431
+ 73.9
2432
+ 72.3
2433
+ 73.8
2434
+ C2L [48]
2435
+ 78.6
2436
+ 77.1
2437
+ 84.5
2438
+ 84.5
2439
+ 66.1
2440
+ 81.1
2441
+ 73.0
2442
+ 72.5
2443
+ 84.0
2444
+ 78.1
2445
+ 88.3
2446
+ 82.1
2447
+ 76.8
2448
+ 75.5
2449
+ 76.8
2450
+ SimSiam [10]
2451
+ 78.3
2452
+ 77.0
2453
+ 84.4
2454
+ 84.1
2455
+ 65.7
2456
+ 80.7
2457
+ 72.7
2458
+ 72.2
2459
+ 83.9
2460
+ 77.9
2461
+ 88.1
2462
+ 82.1
2463
+ 76.6
2464
+ 75.2
2465
+ 75.6
2466
+ PCRLv1 [47]
2467
+ 79.9
2468
+ 78.5
2469
+ 85.8
2470
+ 85.6
2471
+ 67.4
2472
+ 82.3
2473
+ 74.2
2474
+ 73.8
2475
+ 85.5
2476
+ 79.4
2477
+ 89.7
2478
+ 83.5
2479
+ 78.1
2480
+ 76.7
2481
+ 78.1
2482
+ PCRLv2
2483
+ 80.5
2484
+ 79.1
2485
+ 86.4
2486
+ 86.2
2487
+ 68.0
2488
+ 82.8
2489
+ 74.8
2490
+ 74.3
2491
+ 86.0
2492
+ 80.0
2493
+ 90.3
2494
+ 84.1
2495
+ 78.6
2496
+ 77.2
2497
+ 79.2
2498
+ 40%
2499
+ TS
2500
+ 75.4
2501
+ 72.6
2502
+ 80.0
2503
+ 81.0
2504
+ 62.5
2505
+ 76.9
2506
+ 69.2
2507
+ 68.0
2508
+ 80.7
2509
+ 74.7
2510
+ 85.1
2511
+ 79.5
2512
+ 74.0
2513
+ 71.0
2514
+ 79.8
2515
+ IN [14]
2516
+ 79.0
2517
+ 76.7
2518
+ 84.2
2519
+ 84.3
2520
+ 66.3
2521
+ 80.7
2522
+ 73.6
2523
+ 72.3
2524
+ 84.7
2525
+ 78.5
2526
+ 88.6
2527
+ 83.4
2528
+ 77.4
2529
+ 75.0
2530
+ 79.7
2531
+ MG [49]
2532
+ 76.5
2533
+ 74.1
2534
+ 81.3
2535
+ 81.7
2536
+ 63.9
2537
+ 77.9
2538
+ 71.1
2539
+ 70.1
2540
+ 82.5
2541
+ 76.1
2542
+ 85.6
2543
+ 80.6
2544
+ 74.5
2545
+ 73.1
2546
+ 77.9
2547
+ TransVW [16]
2548
+ 77.3
2549
+ 75.2
2550
+ 82.4
2551
+ 82.4
2552
+ 64.4
2553
+ 79.0
2554
+ 71.4
2555
+ 70.5
2556
+ 83.2
2557
+ 76.7
2558
+ 86.6
2559
+ 82.0
2560
+ 75.8
2561
+ 73.6
2562
+ 78.4
2563
+ C2L [48]
2564
+ 79.1
2565
+ 76.9
2566
+ 84.3
2567
+ 84.5
2568
+ 66.4
2569
+ 80.8
2570
+ 73.4
2571
+ 72.2
2572
+ 84.8
2573
+ 78.3
2574
+ 88.6
2575
+ 83.4
2576
+ 77.2
2577
+ 75.4
2578
+ 80.6
2579
+ SimSiam [10]
2580
+ 78.9
2581
+ 76.7
2582
+ 83.9
2583
+ 84.1
2584
+ 66.6
2585
+ 80.4
2586
+ 73.1
2587
+ 72.1
2588
+ 84.7
2589
+ 78.1
2590
+ 88.4
2591
+ 83.4
2592
+ 77.2
2593
+ 74.8
2594
+ 80.5
2595
+ PCRLv1 [47]
2596
+ 80.8
2597
+ 78.5
2598
+ 86.0
2599
+ 86.2
2600
+ 68.2
2601
+ 82.4
2602
+ 75.2
2603
+ 74.0
2604
+ 86.6
2605
+ 80.2
2606
+ 90.2
2607
+ 85.1
2608
+ 79.0
2609
+ 76.9
2610
+ 82.1
2611
+ PCRLv2
2612
+ 81.5
2613
+ 79.2
2614
+ 86.6
2615
+ 86.9
2616
+ 68.9
2617
+ 83.0
2618
+ 75.8
2619
+ 74.6
2620
+ 87.2
2621
+ 80.8
2622
+ 90.9
2623
+ 85.8
2624
+ 79.7
2625
+ 77.6
2626
+ 83.4
2627
+ 50%
2628
+ TS
2629
+ 77.5
2630
+ 75.2
2631
+ 82.0
2632
+ 82.0
2633
+ 64.5
2634
+ 79.6
2635
+ 71.8
2636
+ 71.3
2637
+ 82.9
2638
+ 75.8
2639
+ 86.6
2640
+ 80.9
2641
+ 76.1
2642
+ 75.5
2643
+ 80.3
2644
+ IN
2645
+ 79.5
2646
+ 77.2
2647
+ 84.5
2648
+ 84.4
2649
+ 66.6
2650
+ 81.4
2651
+ 73.6
2652
+ 73.0
2653
+ 84.6
2654
+ 78.2
2655
+ 89.1
2656
+ 82.7
2657
+ 77.9
2658
+ 77.3
2659
+ 82.0
2660
+ MG [49]
2661
+ 77.6
2662
+ 75.0
2663
+ 82.8
2664
+ 82.8
2665
+ 64.8
2666
+ 79.5
2667
+ 71.8
2668
+ 71.6
2669
+ 82.3
2670
+ 75.7
2671
+ 86.7
2672
+ 81.5
2673
+ 76.2
2674
+ 75.7
2675
+ 79.5
2676
+ TransVW [16]
2677
+ 77.3
2678
+ 74.5
2679
+ 81.9
2680
+ 82.4
2681
+ 64.8
2682
+ 78.8
2683
+ 71.5
2684
+ 71.3
2685
+ 82.4
2686
+ 75.7
2687
+ 86.8
2688
+ 80.4
2689
+ 75.7
2690
+ 74.9
2691
+ 80.6
2692
+ C2L [48]
2693
+ 79.8
2694
+ 77.6
2695
+ 84.7
2696
+ 84.5
2697
+ 67.0
2698
+ 81.6
2699
+ 73.6
2700
+ 73.4
2701
+ 84.7
2702
+ 78.5
2703
+ 89.0
2704
+ 83.1
2705
+ 78.4
2706
+ 78.0
2707
+ 82.6
2708
+ SimSiam [10]
2709
+ 80.0
2710
+ 77.7
2711
+ 84.9
2712
+ 84.8
2713
+ 67.1
2714
+ 81.7
2715
+ 74.0
2716
+ 73.5
2717
+ 84.7
2718
+ 78.3
2719
+ 89.5
2720
+ 83.6
2721
+ 78.8
2722
+ 77.7
2723
+ 83.2
2724
+ PCRLv1 [47]
2725
+ 81.2
2726
+ 78.7
2727
+ 86.1
2728
+ 86.3
2729
+ 68.3
2730
+ 82.8
2731
+ 75.4
2732
+ 74.5
2733
+ 86.8
2734
+ 80.4
2735
+ 90.5
2736
+ 85.3
2737
+ 79.5
2738
+ 78.2
2739
+ 83.5
2740
+ PCRLv2
2741
+ 82.5
2742
+ 80.0
2743
+ 87.4
2744
+ 87.3
2745
+ 69.6
2746
+ 84.1
2747
+ 76.4
2748
+ 76.1
2749
+ 87.4
2750
+ 81.0
2751
+ 91.8
2752
+ 85.9
2753
+ 81.0
2754
+ 80.4
2755
+ 86.1
2756
+ 100%
2757
+ TS
2758
+ 80.9
2759
+ 77.7
2760
+ 86.1
2761
+ 85.1
2762
+ 67.7
2763
+ 84.2
2764
+ 73.3
2765
+ 73.9
2766
+ 84.9
2767
+ 78.7
2768
+ 89.4
2769
+ 85.4
2770
+ 79.4
2771
+ 78.5
2772
+ 87.6
2773
+ IN
2774
+ 80.8
2775
+ 77.8
2776
+ 86.3
2777
+ 84.7
2778
+ 67.3
2779
+ 83.6
2780
+ 73.0
2781
+ 74.1
2782
+ 84.9
2783
+ 78.8
2784
+ 89.5
2785
+ 85.7
2786
+ 79.6
2787
+ 78.2
2788
+ 87.0
2789
+ MG [49]
2790
+ 80.8
2791
+ 77.8
2792
+ 86.3
2793
+ 84.7
2794
+ 67.3
2795
+ 83.6
2796
+ 73.0
2797
+ 74.1
2798
+ 84.9
2799
+ 78.8
2800
+ 89.5
2801
+ 85.7
2802
+ 79.6
2803
+ 78.2
2804
+ 87.0
2805
+ TransVW [16]
2806
+ 81.2
2807
+ 77.9
2808
+ 86.4
2809
+ 85.3
2810
+ 67.6
2811
+ 84.3
2812
+ 73.8
2813
+ 74.4
2814
+ 85.1
2815
+ 79.3
2816
+ 89.8
2817
+ 86.2
2818
+ 80.0
2819
+ 78.6
2820
+ 88.8
2821
+ C2L [48]
2822
+ 81.4
2823
+ 78.2
2824
+ 87.0
2825
+ 85.3
2826
+ 68.3
2827
+ 84.8
2828
+ 73.7
2829
+ 74.8
2830
+ 85.5
2831
+ 79.6
2832
+ 90.1
2833
+ 86.3
2834
+ 80.0
2835
+ 78.6
2836
+ 88.1
2837
+ SimSiam [10]
2838
+ 81.6
2839
+ 78.3
2840
+ 87.2
2841
+ 85.5
2842
+ 68.3
2843
+ 84.9
2844
+ 74.2
2845
+ 74.7
2846
+ 85.7
2847
+ 79.6
2848
+ 90.1
2849
+ 86.2
2850
+ 80.2
2851
+ 79.1
2852
+ 89.1
2853
+ PCRLv1 [47]
2854
+ 83.0
2855
+ 79.8
2856
+ 88.5
2857
+ 87.1
2858
+ 69.7
2859
+ 86.1
2860
+ 75.6
2861
+ 76.1
2862
+ 87.0
2863
+ 81.2
2864
+ 91.6
2865
+ 87.7
2866
+ 81.7
2867
+ 80.4
2868
+ 90.2
2869
+ PCRLv2
2870
+ 84.0
2871
+ 80.7
2872
+ 89.3
2873
+ 87.9
2874
+ 70.5
2875
+ 87.0
2876
+ 76.4
2877
+ 77.0
2878
+ 87.9
2879
+ 82.0
2880
+ 92.5
2881
+ 88.6
2882
+ 82.6
2883
+ 81.3
2884
+ 91.6
2885
+ TABLE 7
2886
+ Transfer learning on brain tumor segmentation (on BraTS). WT, TC, and ET stand for the whole tumor, tumor core, and enhancing tumor. For all
2887
+ SSL approaches, we use LUNA for pre-training, and then fine-tune the pre-trained model on BraTS with varying amounts of labeled data. Best
2888
+ results are bolded.
2889
+ Methodology
2890
+ 10%
2891
+ 20%
2892
+ 30%
2893
+ 40%
2894
+ 100%
2895
+ Mean
2896
+ WT
2897
+ TC
2898
+ ET
2899
+ Mean
2900
+ WT
2901
+ TC
2902
+ ET
2903
+ Mean
2904
+ WT
2905
+ TC
2906
+ ET
2907
+ Mean
2908
+ WT
2909
+ TC
2910
+ ET
2911
+ Mean
2912
+ WT
2913
+ TC
2914
+ ET
2915
+ TS
2916
+ 66.6
2917
+ 71.2
2918
+ 66.7
2919
+ 62.1
2920
+ 72.7
2921
+ 78.5
2922
+ 74.3
2923
+ 65.5
2924
+ 76.7
2925
+ 81.8
2926
+ 77.9
2927
+ 70.6
2928
+ 77.1
2929
+ 82.3
2930
+ 78.3
2931
+ 70.9
2932
+ 81.5
2933
+ 86.8
2934
+ 82.8
2935
+ 75.1
2936
+ MG [49]
2937
+ 69.6
2938
+ 72.4
2939
+ 71.4
2940
+ 65.1
2941
+ 75.5
2942
+ 80.4
2943
+ 77.3
2944
+ 68.9
2945
+ 79.6
2946
+ 84.2
2947
+ 80.6
2948
+ 74.1
2949
+ 80.4
2950
+ 85.3
2951
+ 82.0
2952
+ 74.0
2953
+ 82.4
2954
+ 87.1
2955
+ 83.6
2956
+ 76.6
2957
+ TransVW [16]
2958
+ 70.3
2959
+ 74.6
2960
+ 71.7
2961
+ 64.6
2962
+ 75.6
2963
+ 79.9
2964
+ 75.4
2965
+ 71.5
2966
+ 79.1
2967
+ 83.8
2968
+ 79.9
2969
+ 73.6
2970
+ 80.8
2971
+ 85.8
2972
+ 82.1
2973
+ 74.5
2974
+ 82.3
2975
+ 87.1
2976
+ 83.3
2977
+ 76.5
2978
+ Cube++ [35]
2979
+ 69.0
2980
+ 74.5
2981
+ 70.6
2982
+ 61.9
2983
+ 74.9
2984
+ 80.7
2985
+ 75.9
2986
+ 68.1
2987
+ 79.3
2988
+ 84.0
2989
+ 79.4
2990
+ 74.5
2991
+ 79.7
2992
+ 84.5
2993
+ 80.0
2994
+ 74.6
2995
+ 82.2
2996
+ 87.2
2997
+ 82.4
2998
+ 77.0
2999
+ 3D-CPC [34]
3000
+ 70.1
3001
+ 76.7
3002
+ 70.5
3003
+ 63.1
3004
+ 75.9
3005
+ 81.6
3006
+ 75.6
3007
+ 70.5
3008
+ 79.4
3009
+ 84.6
3010
+ 79.9
3011
+ 73.7
3012
+ 81.2
3013
+ 86.5
3014
+ 81.8
3015
+ 75.3
3016
+ 82.9
3017
+ 88.0
3018
+ 83.3
3019
+ 77.4
3020
+ PCRLv1 [47]
3021
+ 71.6
3022
+ 76.9
3023
+ 73.1
3024
+ 65.2
3025
+ 77.6
3026
+ 81.4
3027
+ 79.1
3028
+ 72.7
3029
+ 81.1
3030
+ 84.9
3031
+ 82.2
3032
+ 76.6
3033
+ 83.3
3034
+ 87.5
3035
+ 84.6
3036
+ 78.2
3037
+ 85.0
3038
+ 89.0
3039
+ 86.2
3040
+ 80.2
3041
+ PCRLv2
3042
+ 73.0
3043
+ 77.7
3044
+ 74.3
3045
+ 67.2
3046
+ 78.8
3047
+ 83.2
3048
+ 79.4
3049
+ 74.0
3050
+ 82.1
3051
+ 85.1
3052
+ 82.7
3053
+ 78.7
3054
+ 84.1
3055
+ 87.9
3056
+ 84.5
3057
+ 80.1
3058
+ 85.6
3059
+ 89.4
3060
+ 85.9
3061
+ 81.7
3062
+
3063
+ 13
3064
+ TransVW
3065
+ PCRLv1
3066
+ PCRLv2
3067
+ GT
3068
+ TransVW
3069
+ PCRLv1
3070
+ PCRLv2
3071
+ GT
3072
+ 10%
3073
+ 10%
3074
+ 10%
3075
+ 20%
3076
+ 20%
3077
+ 20%
3078
+ TransVW
3079
+ PCRLv1
3080
+ PCRLv2
3081
+ GT
3082
+ 10%
3083
+ 20%
3084
+ 30%
3085
+ TransVW
3086
+ PCRLv1
3087
+ PCRLv2
3088
+ GT
3089
+ 10%
3090
+ 20%
3091
+ 30%
3092
+ b
3093
+ c
3094
+ Atelectasis
3095
+ TransVW
3096
+ PCRLv1
3097
+ PCRLv2
3098
+ Effusion
3099
+ Infiltration
3100
+ Mass
3101
+ Nodule
3102
+ Pneumonia
3103
+ TransVW
3104
+ PCRLv1
3105
+ PCRLv2
3106
+ a
3107
+ Fig. 8. Visual interpretation of the transfer learning on chest pathology identification (a), and segmentation results of brain tumor (b) and liver (c). We
3108
+ mainly compare PCRLv2 against PCRLv1 and TransVW. Red boxes in the top figure a denote the ground-truth (GT) annotations from radiologists.
3109
+ In figure b, we present the segmentation results of the enhancing tumor (ET) from BraTS when the labeling ratios are 10% and 20%. Similarly in
3110
+ the bottom figure, we display the liver segmentation results in three different labeling ratios (10%, 20%, and 30%).
3111
+
3112
+ PORTABLEPORTABLEPORTABLE14
3113
+ TABLE 8
3114
+ Transfer learning on abdominal organ segmentation (on LiTS). We use
3115
+ LUNA for pre-training, and fine-tune the pre-trained model on LiTS with
3116
+ different amounts of labeled data. Best results are bolded.
3117
+ Methodology
3118
+ Labeling ratio
3119
+ 10%
3120
+ 20%
3121
+ 30%
3122
+ 40%
3123
+ 100%
3124
+ TS
3125
+ 71.1
3126
+ 77.2
3127
+ 84.1
3128
+ 87.3
3129
+ 90.7
3130
+ MG [49]
3131
+ 73.3
3132
+ 79.5
3133
+ 84.3
3134
+ 87.9
3135
+ 91.3
3136
+ TransVW [16]
3137
+ 73.8
3138
+ 79.3
3139
+ 85.5
3140
+ 88.2
3141
+ 91.4
3142
+ Cube++ [35]
3143
+ 74.2
3144
+ 79.3
3145
+ 84.5
3146
+ 88.2
3147
+ 91.8
3148
+ 3D-CPC [34]
3149
+ 74.8
3150
+ 80.2
3151
+ 85.6
3152
+ 88.9
3153
+ 91.9
3154
+ PCRLv1 [47]
3155
+ 77.3
3156
+ 83.5
3157
+ 87.8
3158
+ 90.1
3159
+ 93.7
3160
+ PCRLv2
3161
+ 79.0
3162
+ 86.5
3163
+ 89.3
3164
+ 90.9
3165
+ 94.5
3166
+ Nodule
3167
+ Infiltrate
3168
+ Atelectasis
3169
+ Fig. 9. Failure case analysis on chest pathology identification. Red boxes
3170
+ stand for the lesion areas delineated by radiologists. Images are from
3171
+ NIH ChestX-ray.
3172
+ Somewhat surprisingly, we find 3D-CPC does not out-
3173
+ perform context restoration based SSL (MG, TransVW, and
3174
+ Cube++) as obviously as those in Tables 4, 5, and 7.
3175
+ This comparison is consistent with our intuition: pixel-
3176
+ level information matters a lot in medical image segmen-
3177
+ tation. Again, PCRLv1 and PCRLv2 outperform previous
3178
+ SSL methodologies in all three classes by large margins.
3179
+ Compared to PCRLv1, PCRLv2 is more advantageous in
3180
+ segmenting the enhancing tumor (ET) regions, which are
3181
+ often smaller than WT and TC, and thus harder to segment.
3182
+ The performance gains on ET again verify the effectiveness
3183
+ of multi-scale latent representations, which advances the
3184
+ segmentation of small objects.
3185
+ 4.9
3186
+ Transfer learning on liver segmentation
3187
+ In Table 8, we present the results of liver segmentation.
3188
+ There exist three observable phenomena. First, we see that
3189
+ all SSL approaches provide substantial performance gains
3190
+ over train from scratch. Second, we find the comparative
3191
+ methodology, i.e., 3D-CPC, achieves comparable segmen-
3192
+ tation performance to traditional context restoration based
3193
+ SSL. This phenomenon verifies the necessity of utilizing
3194
+ pixel-level information in medical image segmentation (sim-
3195
+ ilar results also appear in Table 7). Last but not the least,
3196
+ PCRLv2 consistently outperforms PCRLv1 in all labeling
3197
+ ratios, which again validates the effectiveness of introducing
3198
+ multiple scales into SSL.
3199
+ 4.10
3200
+ Visual analysis
3201
+ In Fig. 8, we visually analyze the experimental results of
3202
+ transfer learning with limited annotations on chest pathol-
3203
+ ogy identification (10%), brain tumor segmentation (10%
3204
+ and 20%), and liver segmentation (10%, 20%, and 30%).
3205
+ Here, we compare PCRLv2 against generic SSL methodolo-
3206
+ gies. Considering TransVW was developed on top of MG,
3207
+ we exclude MG and compare PCRLv2 against PCRLv1 and
3208
+ TransVW.
3209
+ Fig. 8a presents the visual interpretation of chest pathol-
3210
+ ogy diagnoses using CAM [45] on six different pathologies.
3211
+ We find that TransVW fails to capture the correct location of
3212
+ lesions on atelectasis, infiltration, nodule, and pneumonia.
3213
+ In comparison, PCRLv1 can generate more interpretable
3214
+ diagnosis results but still yields inconsistent predictions on
3215
+ infiltration and nodule. By integrating multi-scale latent rep-
3216
+ resentations, PCRLv2 can capture the small lesion areas on
3217
+ infiltration and nodule, resulting in centralized yet accurate
3218
+ diagnosis results.
3219
+ In Fig. 8b and Fig. 8c, we visualize the segmentation
3220
+ results of the enhancing tumor (ET) on BraTS and liver
3221
+ on LiTS. Compared to TransVW and PCRLv1, PCRLv2
3222
+ reduces the false positive predictions and contains richer
3223
+ fine-grained details. We believe such superiority of PCRLv2
3224
+ can be attributed to the integration of multi-scale pixel-level
3225
+ and semantic information.
3226
+ We also provide some failure examples in Fig. 9. One
3227
+ common characteristic of these detection results is that they
3228
+ include high-confidence predictions outside the lung area.
3229
+ However, in daily clinical practice, such anomalies should
3230
+ not be located outside the lung area. Similar phenomena
3231
+ have been reported in [13], where the authors summarized
3232
+ them as “shortcuts” that are common in learning systems
3233
+ based on neural networks. To mitigate this problem in self-
3234
+ supervised learning, we can add commonsense knowledge
3235
+ to pre-trained models. Besides, it is also necessary to de-
3236
+ velop more powerful machine learning tools for model
3237
+ interpretation in various downstream tasks.
3238
+ 5
3239
+ CONCLUSION
3240
+ We present a unified visual information preservation frame-
3241
+ work for self-supervised learning in medical imaging. This
3242
+ framework aims to encode the pixel-level, semantic, and
3243
+ scale information into latent representations simultaneously.
3244
+ To achieve this goal, we conduct multi-scale pixel restora-
3245
+ tion and feature comparison on the feature pyramid, which
3246
+ non-skip U-Net supports. The proposed PCRLv2 outper-
3247
+ forms previous self-supervised pre-training approaches by
3248
+ large margins and yields consistent improvements over
3249
+ its conference version (PCRLv1) on four well-established
3250
+ datasets in both quantitative and qualitative validation. We
3251
+ will continue to explore how to optimally integrate different
3252
+ types of information into SSL in the future.
3253
+ REFERENCES
3254
+ [1]
3255
+ Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F
3256
+ McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao,
3257
+ Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al.
3258
+ The lung image database consortium (lidc) and image database
3259
+ resource initiative (idri): a completed reference database of lung
3260
+ nodules on ct scans. Medical Physics, 38(2):915–931, 2011.
3261
+ [2]
3262
+ Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan
3263
+ Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam,
3264
+ Simon Kornblith, Ting Chen, et al. Big self-supervised models ad-
3265
+ vance medical image classification. In Proceedings of the IEEE/CVF
3266
+ International Conference on Computer Vision, pages 3478–3488, 2021.
3267
+
3268
+ 74CRRIGHT15
3269
+ [3]
3270
+ Philip Bachman, R Devon Hjelm, and William Buchwalter. Learn-
3271
+ ing representations by maximizing mutual information across
3272
+ views. Advances in neural information processing systems, 32, 2019.
3273
+ [4]
3274
+ Patrick Bilic, Patrick Ferdinand Christ, Eugene Vorontsov, Grze-
3275
+ gorz Chlebus, Hao Chen, Qi Dou, Chi-Wing Fu, Xiao Han, Pheng-
3276
+ Ann Heng, J¨urgen Hesser, et al. The liver tumor segmentation
3277
+ benchmark (lits). arXiv preprint arXiv:1901.04056, 2019.
3278
+ [5]
3279
+ Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr
3280
+ Bojanowski, and Armand Joulin. Unsupervised learning of visual
3281
+ features by contrasting cluster assignments.
3282
+ Advances in Neural
3283
+ Information Processing Systems, 33:9912–9924, 2020.
3284
+ [6]
3285
+ Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender
3286
+ Konukoglu. Contrastive learning of global and local features for
3287
+ medical image segmentation with limited annotations. Advances
3288
+ in Neural Information Processing Systems, 33:12546–12558, 2020.
3289
+ [7]
3290
+ Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Mi-
3291
+ chitaka Fujiwara, and Daniel Rueckert. Self-supervised learning
3292
+ for medical image analysis using image context restoration. Medi-
3293
+ cal Image Analysis, 58:101539, 2019.
3294
+ [8]
3295
+ Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey
3296
+ Hinton. A simple framework for contrastive learning of visual
3297
+ representations.
3298
+ In International Conference on Machine Learning,
3299
+ pages 1597–1607. PMLR, 2020.
3300
+ [9]
3301
+ Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He.
3302
+ Im-
3303
+ proved baselines with momentum contrastive learning.
3304
+ arXiv
3305
+ preprint arXiv:2003.04297, 2020.
3306
+ [10] Xinlei Chen and Kaiming He. Exploring simple siamese repre-
3307
+ sentation learning. In Proceedings of the IEEE/CVF Conference on
3308
+ Computer Vision and Pattern Recognition, pages 15750–15758, 2021.
3309
+ [11]
3310
+ ¨Ozg¨un C¸ ic¸ek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas
3311
+ Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric
3312
+ segmentation from sparse annotation. In International Conference on
3313
+ Medical Image Computing and Computer-assisted Intervention, pages
3314
+ 424–432. Springer, 2016.
3315
+ [12] Navneet Dalal and Bill Triggs. Histograms of oriented gradients
3316
+ for human detection.
3317
+ In Proceedings of the IEEE Conference on
3318
+ Computer Vision and Pattern Recognition, volume 1, pages 886–893,
3319
+ 2005.
3320
+ [13] Alex J DeGrave, Joseph D Janizek, and Su-In Lee. Ai for radio-
3321
+ graphic covid-19 detection selects shortcuts over signal. Nature
3322
+ Machine Intelligence, 3(7):610–619, 2021.
3323
+ [14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-
3324
+ Fei. Imagenet: A large-scale hierarchical image database. In IEEE
3325
+ Conference on Computer Vision and Pattern Recognition, pages 248–
3326
+ 255, 2009.
3327
+ [15] Jean-Bastien Grill, Florian Strub, Florent Altch´e, Corentin Tallec,
3328
+ Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo
3329
+ Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al.
3330
+ Bootstrap your own latent-a new approach to self-supervised
3331
+ learning.
3332
+ Advances in Neural Information Processing Systems,
3333
+ 33:21271–21284, 2020.
3334
+ [16] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Zong-
3335
+ wei Zhou, Michael B Gotway, and Jianming Liang. Transferable
3336
+ visual words: Exploiting the semantics of anatomical patterns for
3337
+ self-supervised learning.
3338
+ IEEE Transactions on Medical Imaging,
3339
+ 40(10):2857–2868, 2021.
3340
+ [17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Gir-
3341
+ shick. Momentum contrast for unsupervised visual representation
3342
+ learning. In Proceedings of the IEEE Conference on Computer Vision
3343
+ and Pattern Recognition, pages 9729–9738, 2020.
3344
+ [18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep
3345
+ residual learning for image recognition. In Proceedings of the IEEE
3346
+ Conference on Computer Vision and Pattern Recognition, pages 770–
3347
+ 778, 2016.
3348
+ [19] Olivier Henaff. Data-efficient image recognition with contrastive
3349
+ predictive coding. In International Conference on Machine Learning,
3350
+ pages 4182–4192. PMLR, 2020.
3351
+ [20] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan
3352
+ Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio.
3353
+ Learning deep representations by mutual information estimation
3354
+ and maximization. In International Conference on Learning Represen-
3355
+ tations, 2018.
3356
+ [21] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana
3357
+ Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo,
3358
+ Robyn Ball, Katie Shpanskaya, et al. CheXpert: A large chest radio-
3359
+ graph dataset with uncertainty labels and expert comparison. In
3360
+ Proceedings of the AAAI conference on artificial intelligence, volume 33,
3361
+ pages 590–597, 2019.
3362
+ [22] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and
3363
+ Klaus H Maier-Hein. nnu-net: a self-configuring method for deep
3364
+ learning-based biomedical image segmentation. Nature Methods,
3365
+ 18(2):203–211, 2021.
3366
+ [23] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich.
3367
+ Learning representations for automatic colorization. In European
3368
+ Conference on Computer Vision, pages 577–593. Springer, 2016.
3369
+ [24] Tsung-Yi Lin, Piotr Doll´ar, Ross Girshick, Kaiming He, Bharath
3370
+ Hariharan, and Serge Belongie.
3371
+ Feature pyramid networks for
3372
+ object detection. In Proceedings of the IEEE Conference on Computer
3373
+ Vision and Pattern Recognition, pages 2117–2125, 2017.
3374
+ [25] Fengze Liu, Ke Yan, Adam P Harrison, Dazhou Guo, Le Lu, Alan L
3375
+ Yuille, Lingyun Huang, Guotong Xie, Jing Xiao, Xianghua Ye, et al.
3376
+ SAME: Deformable image registration based on self-supervised
3377
+ anatomical embeddings.
3378
+ In International Conference on Medical
3379
+ Image Computing and Computer-Assisted Intervention, pages 87–97.
3380
+ Springer, 2021.
3381
+ [26] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convo-
3382
+ lutional networks for semantic segmentation. In Proceedings of the
3383
+ IEEE Conference on Computer Vision and Pattern Recognition, pages
3384
+ 3431–3440, 2015.
3385
+ [27] David G Lowe. Distinctive image features from scale-invariant
3386
+ keypoints.
3387
+ International Journal of Computer Vision, 60(2):91–110,
3388
+ 2004.
3389
+ [28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representa-
3390
+ tion learning with contrastive predictive coding.
3391
+ arXiv preprint
3392
+ arXiv:1807.03748, 2018.
3393
+ [29] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Dar-
3394
+ rell, and Alexei A Efros. Context encoders: Feature learning by
3395
+ inpainting. In Proceedings of the IEEE Conference on Computer Vision
3396
+ and Pattern Recognition, pages 2536–2544, 2016.
3397
+ [30] Fernando P´erez-Garc´ıa, Rachel Sparks, and Sebastien Ourselin.
3398
+ Torchio: a python library for efficient loading, preprocessing,
3399
+ augmentation and patch-based sampling of medical images in
3400
+ deep learning.
3401
+ Computer Methods and Programs in Biomedicine,
3402
+ 208:106236, 2021.
3403
+ [31] Bart M Haar Romeny. Front-end vision and multi-scale image analysis:
3404
+ multi-scale computer vision theory and applications, written in mathe-
3405
+ matica, volume 27. Springer Science & Business Media, 2008.
3406
+ [32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox.
3407
+ U-net:
3408
+ Convolutional networks for biomedical image segmentation. In
3409
+ International Conference on Medical image computing and computer-
3410
+ assisted intervention, pages 234–241. Springer, 2015.
3411
+ [33] Arnaud Arindra Adiyoso Setio, Alberto Traverso, Thomas De Bel,
3412
+ Moira SN Berens, Cas Van Den Bogaard, Piergiorgio Cerello, Hao
3413
+ Chen, Qi Dou, Maria Evelina Fantacci, Bram Geurts, et al. Vali-
3414
+ dation, comparison, and combination of algorithms for automatic
3415
+ detection of pulmonary nodules in computed tomography images:
3416
+ the luna16 challenge. Medical Image Analysis, 42:1–13, 2017.
3417
+ [34] Aiham Taleb, Winfried Loetzsch, Noel Danz, Julius Severin,
3418
+ Thomas Gaertner, Benjamin Bergner, and Christoph Lippert. 3D
3419
+ self-supervised methods for medical imaging. Advances in Neural
3420
+ Information Processing Systems, 33:18158–18172, 2020.
3421
+ [35] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng.
3422
+ Revisiting rubik’s cube: self-supervised learning with volume-
3423
+ wise transformation for 3d medical image segmentation.
3424
+ In
3425
+ International Conference on Medical Image Computing and Computer-
3426
+ Assisted Intervention, pages 238–248. Springer, 2020.
3427
+ [36] Yonglong Tian, Dilip Krishnan, and Phillip Isola.
3428
+ Contrastive
3429
+ multiview coding.
3430
+ In European Conference on Computer Vision,
3431
+ pages 776–794. Springer, 2020.
3432
+ [37] Yen Nhi Truong Vu, Richard Wang, Niranjan Balachandar, Can
3433
+ Liu, Andrew Y Ng, and Pranav Rajpurkar. Medaug: Contrastive
3434
+ learning leveraging patient metadata improves representations
3435
+ for chest x-ray interpretation. In Machine Learning for Healthcare
3436
+ Conference, pages 755–769. PMLR, 2021.
3437
+ [38] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammad-
3438
+ hadi Bagheri, and Ronald M Summers.
3439
+ Chestx-ray8: Hospital-
3440
+ scale chest x-ray database and benchmarks on weakly-supervised
3441
+ classification and localization of common thorax diseases.
3442
+ In
3443
+ Proceedings of the IEEE Conference on Computer Vision and Pattern
3444
+ Recognition, pages 2097–2106, 2017.
3445
+ [39] Songfan Yang and Deva Ramanan. Multi-scale recognition with
3446
+ dag-cnns. In Proceedings of the IEEE Conference on Computer Vision
3447
+ and Pattern Recognition, pages 1215–1223, 2015.
3448
+ [40] Chenyu You, Ruihan Zhao, Lawrence Staib, and James S Duncan.
3449
+ Momentum contrastive voxel-wise representation learning for
3450
+ semi-supervised volumetric medical image segmentation. arXiv
3451
+ preprint arXiv:2105.07059, 2021.
3452
+ [41] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence Staib, and
3453
+ James S Duncan.
3454
+ Simcvd: Simple contrastive voxel-wise repre-
3455
+ sentation distillation for semi-supervised medical image segmen-
3456
+
3457
+ 16
3458
+ tation. IEEE Transactions on Medical Imaging, 2022.
3459
+ [42] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and St´ephane
3460
+ Deny.
3461
+ Barlow twins: Self-supervised learning via redundancy
3462
+ reduction. In International Conference on Machine Learning, pages
3463
+ 12310–12320. PMLR, 2021.
3464
+ [43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David
3465
+ Lopez-Paz. Mixup: Beyond empirical risk minimization. Interna-
3466
+ tional Conference on Learning Representations, 2017.
3467
+ [44] Richard Zhang, Phillip Isola, and Alexei A Efros.
3468
+ Split-brain
3469
+ autoencoders: Unsupervised learning by cross-channel prediction.
3470
+ In Proceedings of the IEEE Conference on Computer Vision and Pattern
3471
+ Recognition, pages 1058–1067, 2017.
3472
+ [45] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and
3473
+ Antonio Torralba.
3474
+ Learning deep features for discriminative
3475
+ localization.
3476
+ In IEEE Conference on Computer Vision and Pattern
3477
+ Recognition, pages 2921–2929, 2016.
3478
+ [46] Hong-Yu Zhou, Hualuo Liu, Shilei Cao, Dong Wei, Chixiang
3479
+ Lu, Yizhou Yu, Kai Ma, and Yefeng Zheng.
3480
+ Generalized organ
3481
+ segmentation by imitating one-shot reasoning using anatomical
3482
+ correlation. In International Conference on Information Processing in
3483
+ Medical Imaging, pages 452–464. Springer, 2021.
3484
+ [47] Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and
3485
+ Yizhou Yu.
3486
+ Preservational learning improves self-supervised
3487
+ medical image models by reconstructing diverse contexts.
3488
+ In
3489
+ Proceedings of the IEEE/CVF International Conference on Computer
3490
+ Vision, pages 3499–3509, 2021.
3491
+ [48] Hong-Yu Zhou, Shuang Yu, Cheng Bian, Yifan Hu, Kai Ma, and
3492
+ Yefeng Zheng.
3493
+ Comparing to learn: Surpassing imagenet pre-
3494
+ training on radiographs by comparing image representations. In
3495
+ International Conference on Medical Image Computing and Computer-
3496
+ Assisted Intervention, pages 398–407. Springer, 2020.
3497
+ [49] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway,
3498
+ and Jianming Liang.
3499
+ Models genesis.
3500
+ Medical Image Analysis,
3501
+ 67:101840, 2021.
3502
+
-9AyT4oBgHgl3EQf3vkG/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
-9AzT4oBgHgl3EQfFvoN/content/tmp_files/2301.01014v1.pdf.txt ADDED
@@ -0,0 +1,2032 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.01014v1 [math.DG] 3 Jan 2023
2
+ TRICHOTOMY THEOREM FOR PRESCRIBED SCALAR AND MEAN
3
+ CURVATURES ON COMPACT MANIFOLDS WITH BOUNDARIES
4
+ JIE XU
5
+ Abstract. In this article, we give results of prescribing scalar and mean curvature functions
6
+ for metrics either pointwise conformal or conformally equivalent to a Riemannian metric that is
7
+ equipped on a compact manifold with boundary, with dimensions at least 3. The results are clas-
8
+ sified by the sign of the first eigenvalue of the conformal Laplacian. This leads to a “Trichotomy
9
+ Theorem” in terms of both scalar and mean curvature functions, which is a full extension of the
10
+ “Trichotomy Theorem” given by Kazdan and Warner. We also discuss prescribing Gauss and geo-
11
+ desic curvature problems on compact Riemann surfaces with boundary for metrics either pointwise
12
+ conformal or conformally equivalent to the original metric, provided that the Euler characteristic is
13
+ negative. The key step is a general version of monotone iteration scheme which handle the zeroth
14
+ order nonlinear term on the boundary conditions.
15
+ 1. Introduction
16
+ In this article, we give a “Trichotomy Theorem” on compact manifolds ( ¯
17
+ M, g) with non-empty
18
+ smooth boundaries ∂M, n := dim M ⩾ 3, involving both the scalar and mean curvatures. This
19
+ is a full generalization of the “Trichotomy Theorem” on closed manifolds, given by Kazdan and
20
+ Warner [9]. Precisely speaking, this “Trichotomy Theorem” concerns whether the given functions
21
+ S, H can be realized as scalar and mean curvatures, respectively, of a metric ˜g either within a
22
+ conformal class [g] or conformally equivalent to the metric g. Throughout this article, we assume
23
+ that ¯
24
+ M is connected since otherwise we can easily apply arguments below equally to each connected
25
+ component. It is well-known that this problem is reduced to the existence of the positive solutions
26
+ of the nonlinear second order elliptic PDE
27
+ − a∆gu + Rgu = (S ◦ φ) up−1 in M, ∂u
28
+ ∂ν +
29
+ 2
30
+ p − 2hgu =
31
+ 2
32
+ p − 2 (H ◦ φ) u
33
+ p
34
+ 2 on ∂M.
35
+ (1)
36
+ Here Rg is the scalar curvature of the metric g, hg is the mean curvature. φ : ¯
37
+ M → ¯
38
+ M is some
39
+ diffeomorphism on
40
+ ¯
41
+ M.
42
+ When φ = Id, the PDE (1) is for prescribing functions S, H within a
43
+ conformal class [g]. The constants a, p are defined as
44
+ a = 4(n − 1)
45
+ n − 2 , p =
46
+ 2n
47
+ n − 2.
48
+ ∆g is the Laplace-Beltrami operator and ν is the unique outward unit normal vector field along
49
+ ∂M. The functions S : C∞( ¯
50
+ M) → R, and H : C∞(∂M) → R are given. We denote η1 to be the
51
+ first eigenvalue of the conformal Laplacian □g := −a∆g + Rg with associated eigenfunction ϕ, i.e.
52
+ ϕ is a positive, smooth function that solves the following PDE:
53
+ −a∆gϕ + Rgϕ = η1ϕ in M, ∂ϕ
54
+ ∂ν +
55
+ 2
56
+ p − 2hgϕ = 0 on ∂M.
57
+ When the dimension of the manifold n = 2, we also discuss the pair of functions K, σ that can be
58
+ realized as Gaussian and geodesic curvatures, respectively, either for a pointwise conformal metric
59
+ 1
60
+
61
+ 2
62
+ J. XU
63
+ or a conformally equivalent metric. The two dimensional case is reduced to the existence of the
64
+ solutions of the following elliptic PDE
65
+ − a∆gu + Kg = (K ◦ φ) e2u in M, ∂u
66
+ ∂ν + σg = (σ ◦ φ) eu on ∂M.
67
+ (2)
68
+ Here Kg and σg are Gaussian and geodesic curvatures of g, respectively.
69
+ The functions K :
70
+ C∞( ¯
71
+ M) → R and σ : C∞(∂M) → R are given. Again when the diffeomorphism φ :
72
+ ¯
73
+ M →
74
+ ¯
75
+ M
76
+ is the identity map, K, σ are prescribing Gauss and geodesic curvatures for some metric within the
77
+ conformal class [g].
78
+ The main results of this article are given as follows:
79
+ Theorem 1.1. Let ( ¯
80
+ M, g) be a connected, compact manifold with non-empty smooth boundary ∂M,
81
+ n = dim ¯
82
+ M ⩾ 3. Let S, H ∈ C∞( ¯
83
+ M) be given functions.
84
+ (i). If η1 < 0, then any function S < 0 somewhere in M can be realized as a scalar curvature
85
+ function of some metric conformally equivalent to g, with mean curvature cH for some small
86
+ enough constant c > 0 and any function H;
87
+ (ii). If η1 < 0, then any function S < 0 that changes sign in M can be realized as a scalar curvature
88
+ function of some metric conformally equivalent to g, with mean curvature cH for some small
89
+ enough constant c > 0 and any function H;
90
+ (iii). If η1 < 0, then any function S > 0 somewhere in M can be realized as a scalar curvature
91
+ function of some metric pointwise conformal to g, with mean curvature cH for some small
92
+ enough constant c > 0 and any function H.
93
+ Case (i) is given in §3 and §4; when S < 0 everywhere on ¯
94
+ M, we can improve the result within
95
+ a pointwise conformal class [g] in Theorem 3.1 and Theorem 3.2; Case (ii) is given in §6; when S
96
+ satisfies
97
+ ´
98
+ M SdVolg < 0 in addition, we can improve the result within a pointwise conformal class
99
+ [g], see [18, Thm .1.2]; and Case (iii) is given in §7. The significance of this is that we can choose
100
+ arbitrary function with small enough sup-norm as our mean curvature function, provided that the
101
+ scalar curvature function is nontrivial.
102
+ Based on our best understanding, known results in this topic are mainly for the non-positive first
103
+ eigenvalue cases or non-positive Euler characteristic cases. In [5], Cruz-Bl´azquez, Malchiodi and
104
+ Ruiz discussed prescribing negative scalar functions and mean curvature functions with arbitrary
105
+ signs by variational method, for compact manifolds with dimensions at least 2. Some of our results
106
+ overlap their results, but with a different method and different hypotheses on prescribed functions.
107
+ However, our results are classified by the sign of the first eigenvalue of the conformal Laplacian. For
108
+ zero first eigenvalue case or zero Euler characteristic case, we follow the results of [18]. We point
109
+ out that Brezis and Merle discussed the PDE −∆eu = V eu on Ω ⊂ R2 with Dirichlet boundaray
110
+ condition in [3]. Other results for the local Yamabe equation with Dirichlet condition in higher
111
+ dimensions could be found in [12]. For more discussions with respect to (2) in 2-dimensional case,
112
+ we refer to [15, Ch. 13, Ch. 14]. When the first eigenvalue of conformal Laplacian is positive, a lot
113
+ of non-existence results are given, e.g. [2] and [11], etc.
114
+ We also give results on compact Riemann surfaces with boundary, provided that χ( ¯
115
+ M) < 0.
116
+ Theorem 1.2. Let ( ¯
117
+ M, g) be a compact Riemann surface with non-empty smooth boundary ∂M.
118
+ Let K, σ ∈ C∞( ¯
119
+ M) be given functions.
120
+ (i). If K < 0 everywhere on ¯
121
+ M, then there exists a metric pointwise conformal to g with Gauss
122
+ curvature K and geodesic curvature cσ for some small enough constant c > 0 and arbitrary
123
+ function σ;
124
+ (ii). If K < 0 somewhere on ¯
125
+ M, then there exists a metric conformally equivalent to g with Gauss
126
+ curvature K and geodesic curvature cσ for some small enough constant c > 0 and arbitrary
127
+ function σ.
128
+
129
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
130
+ 3
131
+ Both results above are given in §5. Other than the related works on compact Riemann surfaces
132
+ with boundary we introduced above, many work has been done on closed Riemann surface, a
133
+ comprehensive study was given by Kazdan and Warner [8], including results for all signs of χ( ¯
134
+ M).
135
+ For Nirenberg problem, we refer to Chang and Yang [4] and Struwe [13], etc..
136
+ The most common method in analyzing this type of Kazdan-Warner problem is by calculus
137
+ of variations since we can consider the PDE as Euler-Lagrange equation with respect to some
138
+ functional; recently Morse theory is also involved. However, a new method, inspired by Kazdan
139
+ and Warner [10], has been developed recently. This new method applies monotone iteration scheme,
140
+ a local version of calculus of variation to classify the existence results by sign of the first eigenvalue
141
+ η1 of conformal Laplacian. This method has been applied to completely solve the Escobar problem
142
+ [16], the Han-Li conjecture [17], the prescribed scalar curvature problem on compact manifolds [19],
143
+ a trichotomy theorem in terms of prescribed scalar curvature with Dirichlet condition at boundary
144
+ [20], and a comprehensive study of zero first eigenvalue case on compact manifolds, possibly with
145
+ boundary, with dimensions at least 3 [18]. In this article, we apply a variation of the combination
146
+ of monotone iteration scheme and local analysis to show the results of prescribed scalar and mean
147
+ curvatures for the cases η1 > 0 and η1 < 0. We also develop a general monotone iteration scheme,
148
+ which can handle nonlinear terms both in the PDE and on the boundary condition; this new
149
+ monotone iteration scheme, see Theorem 2.3, allows us to work on 2-dimensional case without
150
+ using the calculus of variation.
151
+ This systematic procedure is powerful, but unfortunately this
152
+ direct method cannot be used to the classical manifold, the unit ball with spherical boundary. We
153
+ will explain why this direct method does not work in this case. Note that Escobar [7] has proved a
154
+ nontrivial Kazdan-Warner type obstruction of prescribed mean curvature functions for this case.
155
+ This paper is organized as follows:
156
+ In §2, we introduce the essential definitions and results that will be used throughout this article.
157
+ We assume the backgrounds of standard elliptic theory. We also introduced two versions of mono-
158
+ tone iteration schemes. Theorem 2.2 is for the PDE (1); Theorem 2.3 is more general, works for all
159
+ second order semi-linear elliptic PDE with Robin boundary conditions, possibly with zeroth order
160
+ nonlinear term on the boundary condition. Theorem 2.3 works well for the PDE like (2).
161
+ In §3, we give results for prescribing scalar curvature function S and mean curvature function
162
+ H within a conformal class [g] on ( ¯
163
+ M, g), n = dim ¯
164
+ M ⩾ 3, provided that η1 < 0. When S < 0
165
+ everywhere and arbitrary H, the results are given in Theorem 3.1 and Theorem 3.2. When S < 0
166
+ somewhere and arbitrary H, the results are given in Theorem 3.3 and Corollary 3.1 with some
167
+ restriction on S. The monotone iteration scheme plays a central role.
168
+ In §4, we give results for prescribing scalar curvature function S and mean curvature function
169
+ H for some metric conformally equivalent to g on ( ¯
170
+ M, g), n = dim ¯
171
+ M ⩾ 3, provided that η1 < 0.
172
+ It follows from Corollary 3.1. We conclude in Theorem 4.1 that any S that is negative somewhere
173
+ can be realized as a scalar curvature function of some metric conformally equivalent to g, with the
174
+ mean curvature cH for small enough constant c > 0 and arbitrary H.
175
+ In §5, we discuss prescribing Gauss and geodesic curvature functions K, σ on compact Riemann
176
+ surfaces with boundary for metrics conformally equivalent to the original metric g. We show in
177
+ Theorem 5.1 that any function K that is negative somewhere and satisfies some analytic condition
178
+ can be realized as Gaussian curvature function for some metric conformally equivalent to g, the
179
+ metric also has geodesic curvature cσ for some small enough constant c > 0 and arbitrary σ. The
180
+ result in Corollary 5.1 says that when K < 0 everywhere on ¯
181
+ M, the metric can be chosen within a
182
+ conformal class [g].
183
+ In §6, we give results for prescribing scalar function S and mean curvature function H for some
184
+ metric conformally equivalent to g on ( ¯
185
+ M, g), n = dim ¯
186
+ M ⩾ 3, provided that η1 = 0. We show
187
+ that any function S that changes sign can be realized as a scalar curvature function some metric
188
+
189
+ 4
190
+ J. XU
191
+ conformally equivalent to g, with the mean curvature cH for small enough constant c > 0 and
192
+ arbitrary H in Corollary 6.2. Obviously there is a trivial case S ≡ H ≡ 0.
193
+ In §7, we consider the prescribing scalar and mean curvature problem for η1 > 0. The results in
194
+ Theorem 7.1 and Theorem 7.2 are for the case S > 0 somewhere and arbitrary H. We also explain
195
+ why our method cannot work on closed Euclidean ball with some nontrivial mean curvature on the
196
+ boundary Sn.
197
+ 2. The Preliminaries and The Monotone Iteration Scheme
198
+ In this section, we first introduce the necessary definitions and essential results we need for the
199
+ later sections, then introduce a general version of the monotone iteration scheme given in [18], other
200
+ than many variations we have used in [16, 17, 19, 20, 21], with respect to the following Yamabe
201
+ equation with Robin boundary condition
202
+ − a∆gu + Rgu = Sup−1 in M, ∂u
203
+ ∂ν +
204
+ 2
205
+ p − 2hgu =
206
+ 2
207
+ p − 2Hu
208
+ p
209
+ 2 on ∂M.
210
+ (3)
211
+ for given functions S, H ∈ C∞( ¯
212
+ M), and n = dim ¯
213
+ M ⩾ 3. Lastly we introduce a W s,q-type regularity
214
+ for elliptic PDE with Robin boundary conditions.
215
+ First of all, we give definitions of Sobolev spaces, a local version and a global version.
216
+ Let
217
+ Ω be a connected, bounded, open subset of Rn with smooth boundary ∂Ω equipped with some
218
+ Riemannian metric g that can be extended smoothly to ¯Ω. We call (Ω, g) a Riemannian domain.
219
+ Throughout this article, we denote the space of smooth functions with compact support by C∞
220
+ c ,
221
+ smooth functions by C∞, and continuous functions by C0.
222
+ Definition 2.1. Let (Ω, g) be a Riemannian domain.
223
+ Let (M, g) be a closed Riemannian n-
224
+ manifold, and ( ¯
225
+ M, g) be a compact Riemannian n-manifold with non-empty smooth boundary, with
226
+ volume density dVolg. Let u be a real valued function. Let ⟨v, w⟩g and |v|g = ⟨v, v⟩1/2
227
+ g
228
+ denote the
229
+ inner product and norm with respect to g.
230
+ (i) For 1 ⩽ p < ∞, we define the Lebesgue spaces on Ω and ¯
231
+ M to be
232
+ Lp(Ω) is the completion of
233
+
234
+ u ∈ C∞
235
+ c (Ω) : ∥u∥p
236
+ p :=
237
+ ˆ
238
+
239
+ |u|pdx < ∞
240
+
241
+ ,
242
+ Lp(Ω, g) is the completion of
243
+
244
+ u ∈ C∞
245
+ c (Ω) : ∥u∥p
246
+ p,g :=
247
+ ˆ
248
+
249
+ |u|p dVolg < ∞
250
+
251
+ ,
252
+ Lp(M, g) is the completion of
253
+
254
+ u ∈ C∞(M) : ∥u∥p
255
+ p,g :=
256
+ ˆ
257
+ M
258
+ |u|p dVolg < ∞
259
+
260
+ .
261
+ (ii) For ∇u the Levi-Civita connection of g, and for u ∈ C∞(Ω) or u ∈ C∞( ¯
262
+ M),
263
+ |∇ku|2
264
+ g := (∇α1 . . . ∇αku)(∇α1 . . . ∇αku).
265
+ (4)
266
+ In particular, |∇0u|2
267
+ g = |u|2 and |∇1u|2
268
+ g = |∇u|2
269
+ g.
270
+
271
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
272
+ 5
273
+ (iii) For s ∈ N, 1 ⩽ p < ∞, we define the (s, p)-type Sobolev spaces on Ω and ¯
274
+ M to be
275
+ W s,p(Ω) =
276
+
277
+
278
+ u ∈ Lp(Ω) : ∥u∥p
279
+ W s,p(Ω) :=
280
+ ˆ
281
+
282
+ s
283
+
284
+ j=0
285
+ ��Dju
286
+ ��p dx < ∞
287
+
288
+
289
+  ,
290
+ (5)
291
+ W s,p(Ω, g) =
292
+
293
+
294
+ u ∈ Lp(Ω, g) : ∥u∥p
295
+ W s,p(Ω,g) =
296
+ s
297
+
298
+ j=0
299
+ ˆ
300
+
301
+ ��∇ju
302
+ ��p
303
+ g dVolg < ∞
304
+
305
+
306
+  ,
307
+ W s,p(M, g) =
308
+
309
+
310
+ u ∈ Lp(M, g) : ∥u∥p
311
+ W s,p(M,g) =
312
+ s
313
+
314
+ j=0
315
+ ˆ
316
+ M
317
+ ��∇ju
318
+ ��p
319
+ g dVolg < ∞
320
+
321
+
322
+  .
323
+ Here |Dju|p := �
324
+ |α|=j|∂αu|p in the weak sense. Similarly, W s,p
325
+ 0 (Ω) is the completion of C∞
326
+ c (Ω)
327
+ with respect to the W s,p-norm.
328
+ In particular, Hs(Ω) := W s,2(Ω) and Hs(Ω, g) := W s,2(Ω, g),
329
+ Hs(M, g) := W s,2(M, g) are the usual Sobolev spaces. We similarly define Hs
330
+ 0(Ω), Hs
331
+ 0(Ω, g) and
332
+ Hs
333
+ 0(M, g).
334
+ (iv) On closed manifolds (M, g), we say that a function u ∈ Hs(M, g) if u ∈ L2(M, g) , and for
335
+ any coordinate chart U ⊂ M, any ψ ∈ C∞
336
+ c (U), the function ψu ∈ Hs(U, g).
337
+ We assume the background of the standard elliptic theory, including the solvability of standard
338
+ linear elliptic PDEs, elliptic regularity of Hs-type, trace theorem, Sobolev embedding, Schauder
339
+ estimates, etc. We introduce a W s,q-type elliptic regularity for later use.
340
+ Theorem 2.1. [17, Thm. 2.2] Let ( ¯
341
+ M, g) be a compact manifold with smooth boundary ∂M. Let
342
+ ν be the unit outward normal vector along ∂M and q > n = dim ¯
343
+ M. Let L : C∞( ¯
344
+ M) → C∞( ¯
345
+ M)
346
+ be a uniform second order elliptic operator on M with smooth coefficients up to ∂M and can be
347
+ extended to L : W 2,q(M, g) → Lq(M, g). Let f ∈ Lq(M, g), ˜f ∈ W 1,q(M, g). Let u ∈ H1(M, g) be a
348
+ weak solution of the following boundary value problem
349
+ Lu = f in M, Bu = ∂u
350
+ ∂ν + c(x)u = ˜f on ∂M.
351
+ (6)
352
+ Here c ∈ C∞(M). Assume also that Ker(L) = {0} associated with the homogeneous Robin boundary
353
+ condition. If, in addition, u ∈ Lq(M, g), then u ∈ W 2,q(M, g) with the following estimates
354
+ ∥u∥W 2,q(M,g) ⩽ γ′ �
355
+ ∥Lu∥Lq(M,g) + ∥Bu∥W 1,q(M,g)
356
+
357
+ (7)
358
+ Here γ′ depends on L, q, c and the manifold ( ¯
359
+ M, g) and is independent of u.
360
+ We then introduce the first eigenvalue of conformal Laplacian. Note that a = 4(n−1)
361
+ n−2
362
+ and p =
363
+ 2n
364
+ n−2,
365
+ hence it only makes sense when n ⩾ 3.
366
+ Definition 2.2. Let ( ¯
367
+ M, g) be a compact manifold with non-empty smooth boundary ∂M. We
368
+ denote η1 be the first eigenvalue of conformal Laplacian with its corresponding eigenfunction ϕ > 0
369
+ if and only if the following PDE holds.
370
+ − a∆gϕ + Rgϕ = η1ϕ in M, ∂ϕ
371
+ ∂ν +
372
+ 2
373
+ p − 2hgϕ = 0 on ∂M.
374
+ (8)
375
+ We now introduce a variation of the monotone iteration scheme we used in [16], [17] and [19].
376
+ In particular, we do require hg = h > 0 to be some positive constant on ∂M, this can be done due
377
+ to the proof of the Han-Li conjecture in [17]. We will also use other versions of monotone iteration
378
+ schemes introduced in eariler work [16, 17, 19, 20, 21].
379
+ Theorem 2.2. [18, Thm. 2.4] Let ( ¯
380
+ M, g) be a compact manifold with smooth boundary ∂M. Let
381
+ ν be the unit outward normal vector along ∂M and q > dim ¯
382
+ M. Let S ∈ C∞( ¯
383
+ M) and H ∈ C∞( ¯
384
+ M)
385
+
386
+ 6
387
+ J. XU
388
+ be given functions. Let the mean curvature hg = h > 0 be some positive constant. In addition,
389
+ we assume that sup ¯
390
+ M|H| is small enough. Suppose that there exist u− ∈ C0( ¯
391
+ M) ∩ H1(M, g) and
392
+ u+ ∈ W 2,q(M, g) ∩ C0( ¯
393
+ M), 0 ⩽ u− ⩽ u+, u− ̸≡ 0 on ¯
394
+ M, some constants θ1 ⩽ 0, θ2 ⩾ 0 such that
395
+ −a∆gu− + Rgu− − Sup−1
396
+
397
+ ⩽ 0 in M, ∂u−
398
+ ∂ν +
399
+ 2
400
+ p − 2hgu− ⩽ θ1u− ⩽
401
+ 2
402
+ p − 2Hu
403
+ p
404
+ 2
405
+ − on ∂M
406
+ −a∆gu+ + Rgu+ − Sup−1
407
+ +
408
+ ⩾ 0 in M, ∂u+
409
+ ∂ν +
410
+ 2
411
+ p − 2hgu+ ⩾ θ2u+ ⩾
412
+ 2
413
+ p − 2Hu
414
+ p
415
+ 2
416
+ + on ∂M
417
+ (9)
418
+ holds weakly. In particular, θ1 can be zero if H ⩾ 0 on ∂M, and θ1 must be negative if H < 0
419
+ somewhere on ∂M; similarly, θ2 can be zero if H ⩽ 0 on ∂M, and θ2 must be positive if H > 0
420
+ somewhere on ∂M. Then there exists a real, positive solution u ∈ C∞(M) ∩ C1,α( ¯
421
+ M) of
422
+ □gu = −a∆gu + Rgu = Sup−1 in M, Bgu = ∂u
423
+ ∂ν +
424
+ 2
425
+ p − 2hgu =
426
+ 2
427
+ p − 2Hu
428
+ p
429
+ 2 on ∂M.
430
+ (10)
431
+ The following two results are necessary, which shows the existence of the solution of some local
432
+ Yamabe-type problem. When the manifold is not locally conformally flat, we need
433
+ Proposition 2.1. [18, Prop. 3.2] Let (Ω, g) be a Riemannian domain in Rn, n ⩾ 3, not locally
434
+ conformally flat, with C∞ boundary, with Volg(Ω) and the Euclidean diameter of Ω sufficiently
435
+ small. Let f ∈ Ω′ ⊃ Ω be a positive, smooth function in some open region Ω′. In addition, we
436
+ assume that the first eigenvalue of Laplace-Beltrami operator −∆g on Ω with Dirichlet condition
437
+ satisfies λ1 → ∞ as Ω shrinks. Assume Rg < 0 within the small enough closed domain ¯Ω. Then
438
+ the Dirichlet problem
439
+ − a∆gu + Rgu = fup−1 in Ω, u ≡ 0 on ∂Ω
440
+ (11)
441
+ has a real, positive, smooth solution u ∈ C∞(Ω) ∩ H1
442
+ 0(Ω, g) ∩ C0(¯Ω). The size of Ω is depending on
443
+ the function f.
444
+ When the manifold is locally conformally flat, we give the local solution of (11) provided that Ω
445
+ is not topologically trivial.
446
+ Proposition 2.2. [19, Prop. 2.5] Let (Ω, g) be a Riemannian domain in Rn, n ⩾ 3, with C∞
447
+ boundary. Let the metric g be locally conformally flat on some open subset Ω′ ⊃ ¯Ω. For any point
448
+ ρ ∈ Ω and any positive constant ǫ, denote the region Ωǫ to be
449
+ Ωǫ = {x ∈ Ω||x − ρ| > ǫ}.
450
+ Assume that Q ∈ C2(¯Ω), minx∈¯Ω Q(x) > 0 and ∇Q(ρ) ̸= 0. Then there exists some ǫ0 such that for
451
+ every ǫ ∈ (0, ǫ0) the Dirichlet problem
452
+ − a∆gu + Rgu = Qup−1 in Ωǫ, u = 0 on ∂Ωǫ
453
+ (12)
454
+ has a real, positive, smooth solution u ∈ C∞(Ωǫ) ∩ H1
455
+ 0(Ωǫ, g) ∩ C0( ¯Ωǫ).
456
+ Remark 2.1. It is straightforward to see that under conformal change ˜g = φp−2g, we have
457
+ ˜g = φp−2g ⇒ −a∆˜g + R˜g = φ− n+2
458
+ n−2 (−a∆g + Rg) φ ⇔ □˜g = φ1−p□gφ.
459
+ (13)
460
+ We call (13) the conformal invariance of the conformal Laplacian. It follows from Proposition 2.2
461
+ and (13) that if the manifold ( ¯
462
+ M, g) is locally conformally flat in the interior, the equation (12) is
463
+ equivalent to
464
+ − a∆geu = Qup−1 in Ωǫ, u = 0 on ∂Ωǫ
465
+ (14)
466
+ which admits a positive solution u ∈ C∞(Ωǫ) ∩ H1
467
+ 0(Ωǫ, g) ∩ C0( ¯Ωǫ).
468
+ As a prerequisite, we also need a result in terms of the perturbation of negative first eigenvalue
469
+ of conformal Laplacian.
470
+
471
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
472
+ 7
473
+ Proposition 2.3. Let ( ¯
474
+ M, g) be a compact Riemannian manifold with non-empty smooth boundary
475
+ ∂M, n = dim ¯
476
+ M ⩾ 3. Let β > 0 be a small enough constant. If η′
477
+ 1 < 0, then the quantity
478
+ η′
479
+ 1,β = inf
480
+ u̸=0
481
+ a
482
+ ´
483
+ M|∇gu|2dVolg +
484
+ ´
485
+ M Rgu2dVolg +
486
+ 2a
487
+ p−2
488
+ ´
489
+ ∂M(hg + β)u2dS
490
+ ´
491
+ M u2dVolg
492
+ < 0.
493
+ In particular, η′
494
+ 1,β satisfies
495
+ − a∆gϕ + Rgϕ = η′
496
+ 1,βϕ in M, ∂ϕ
497
+ ∂ν +
498
+ 2
499
+ p − 2(hg + β)ϕ = 0 on ∂M
500
+ (15)
501
+ with some positive function ϕ ∈ C∞( ¯
502
+ M).
503
+ Proof. Since η′
504
+ 1 < 0, the normalized first eigenfunction ϕ1, i.e.
505
+ ´
506
+ M ϕ2
507
+ 1dVolg = 1, satisfies
508
+ η′
509
+ 1 = a
510
+ ˆ
511
+ M
512
+ |∇gϕ1|2dVolg +
513
+ ˆ
514
+ M
515
+ Rgϕ2
516
+ 1dVolg +
517
+ 2a
518
+ p − 2
519
+ ˆ
520
+ ∂M
521
+ hgϕ2
522
+ 1dS
523
+ By characterization of η′
524
+ 1,β, we have
525
+ η′
526
+ 1,β ⩽ a
527
+ ˆ
528
+ M
529
+ |∇gϕ1|2dVolg +
530
+ ˆ
531
+ M
532
+ Rgϕ2
533
+ 1dVolg +
534
+ 2a
535
+ p − 2
536
+ ˆ
537
+ ∂M
538
+ (hg + β)ϕ2
539
+ 1dS = η′
540
+ 1 + β
541
+ ˆ
542
+ ∂M
543
+ ϕ2
544
+ 1dS.
545
+ Since ϕ1 is fixed, it follows that η′
546
+ 1,β < 0 if β > 0 is small enough.
547
+
548
+ When n = 2, i.e. M or ¯
549
+ M is a compact Riemann surface (possibly with boundary), all tools
550
+ above are not available. We thus need a new version of the monotone iteration scheme for compact
551
+ Riemann surfaces with non-empty smooth boundary. We point out that the monotone iteration
552
+ scheme below works for all compact manifolds with non-empty boundary, with dimensions at least
553
+ 2.
554
+ Theorem 2.3. Let ( ¯
555
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
556
+ dim M ⩾ 2. Let q > n be a positive integer. Let F(·, ·), G(·, ·) : ¯
557
+ M × R → R be smooth functions.
558
+ Let ν be the unit outward normal vector along ∂M. Let σ be some nonnegative, small enough
559
+ constant. If
560
+ (i) there exists two functions u+ ∈ C∞( ¯
561
+ M) and u− ∈ C0( ¯
562
+ M) ∩ H1(M, g) such that
563
+ −∆gu+ ⩾ F(·, u) in M, ∂u
564
+ ∂ν + σu ⩾ G(·, u+) on ∂M;
565
+ −∆gu− ⩽ F(·, u) in M, ∂u
566
+ ∂ν + σu ⩽ G(·, u−) on ∂M,
567
+ (16)
568
+ where the sub-solution may hold in the weak sense; and
569
+ (ii) in addition, sup ¯
570
+ M|G(·, u+)|, sup ¯
571
+ M|∇G(·, u+)| are small enough;
572
+ (iii) furthermore, u+ ⩾ u− pointwise on ¯
573
+ M;
574
+ then there exists a smooth function u ∈ C∞( ¯
575
+ M) with u− ⩽ u ⩽ u+ such that
576
+ − ∆gu = F(·, u) in M, ∂u
577
+ ∂ν + σu = G(·, u) on ∂M.
578
+ (17)
579
+ Remark 2.2. The proof of Theorem 2.3 is essentially the same as the proof of [18, Thm. 2.4],
580
+ except some minor change, e.g. here we use general smooth functions F and G but not specific
581
+ Yamabe equations. We therefore will give a relatively concise proof for Theorem 2.3.
582
+
583
+ 8
584
+ J. XU
585
+ Proof.
586
+ ¯
587
+ M is compact, so extremal values of continuous functions u+, u− can be achieved. Choose
588
+ positive constant A and nonnegative constant B such that
589
+ A ⩾ −∂F
590
+ ∂u (x, u(x)), ∀x ∈ ¯
591
+ M, u(x) ∈ [min
592
+ ¯
593
+ M u−, max
594
+ ¯
595
+ M u+];
596
+ B ⩾ σ − ∂G
597
+ ∂u (x, u(x)), ∀x ∈ ¯
598
+ M, u(x) ∈ [min
599
+ ¯
600
+ M u−, max
601
+ ¯
602
+ M u+].
603
+ (18)
604
+ Denote u0 = u+ ∈ C∞( ¯
605
+ M), and consider the iteration scheme
606
+ −∆guk + Auk = Auk−1 + F(·, uk−1) in M, k ∈ N,
607
+ ∂uk
608
+ ∂ν + Buk = Buk−1 − σuk−1 + G(·, uk−1) on ∂M, k ∈ N.
609
+ (19)
610
+ Since A > 0, B ⩾ 0, the operator
611
+
612
+ −∆g + A, ∂
613
+ ∂ν + B
614
+
615
+ is invertible due to the standard argument. Clearly when k = 1, the first iteration step in (19)
616
+ gives a unique smooth solution u1 ∈ C∞( ¯
617
+ M). The regularity argument is also standard.
618
+ We show that u− ⩽ u ⩽ u+. For u− ⩽ u, we have to use the sub-solution in the weak sense,
619
+ since u0 = u+ ⩾ u−, we pair (19) for k = 1 with arbitrary non-negative function v ∈ C∞( ¯
620
+ M), and
621
+ subtract this with the sub-solution (adding Au− and Bu− on both sides of the PDE and boundary
622
+ conditions respectively) in the weak sense, we have
623
+ ˆ
624
+ M
625
+ (A (u0 − u−) + F (x, u0) − F (x, u−)) vdVolg ⩽
626
+ ˆ
627
+ M
628
+ (−∆g (u1 − u−) + A (u1 − u−)) vdVolg
629
+
630
+ ˆ
631
+ ∂M
632
+ B (u1 − u−) vdS −
633
+ ˆ
634
+ ∂M
635
+ (B (u0 − u−) − σ (u0 − u−) + G (x, u0) − G (x, u−)) vdS
636
+ +
637
+ ˆ
638
+ M
639
+ A (u1 − u−) vdVolg +
640
+ ˆ
641
+ M
642
+ ∇g (u1 − u−) · ∇gvdVolg.
643
+ Taking v = w := max (u− − u1, 0), and applying the mean value theorem for F, G, due to the
644
+ definitions of A, B in (18), we observe that
645
+ ˆ
646
+ M
647
+ |∇gw|2 +
648
+ ˆ
649
+ ∂M
650
+ Bw2 +
651
+ ˆ
652
+ M
653
+ Aw2 ⩽ 0.
654
+ It follows that w = 0, therefore u− ⩽ u1. By a very similar argument in terms of the subtraction
655
+ between (19) and the super-solution, we conclude that u+ ⩾ u1.
656
+ Inductively, we may assume the existence of the solutions u1, . . . , uk with
657
+ u− ⩽ uk ⩽ uk−1 ⩽ . . . ⩽ u1 ⩽ u0.
658
+ By the same argument in the first iteration step, we conclude the existence of uk+1 ∈ C∞( ¯
659
+ M); in
660
+ addition, uk+1 satisfies
661
+ u− ⩽ uk+1 ⩽ uk ⩽ uk−1 ⩽ . . . ⩽ u1 ⩽ u0.
662
+ Therefore we show the existence of the sequence of solutions of (19) with the monotonicity
663
+ u− ⩽ . . . ⩽ uk+1 ⩽ uk ⩽ uk−1 ⩽ . . . ⩽ u0, k ∈ N.
664
+ (20)
665
+ We now show the uniform boundedness of ∥uk∥C1,α( ¯
666
+ M).
667
+ Since q > n, showing the uniform
668
+ boundedness of ∥uk∥C1,α( ¯
669
+ M) is equivalent to show the uniform boundedness of ∥uk∥W 2,q(M,g). We
670
+ have mentioned that the operator is invertible and thus the W s,q-type estimates (7) applies. We L
671
+ and the boundary condition c to be the operators here with associated constant γ′. Mimicking the
672
+
673
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
674
+ 9
675
+ boundedness proof in [18, Thm. 2.4], we should require σ and sup ¯
676
+ M|G(·, u)|, and sup ¯
677
+ M|∇G(·, u)|
678
+ to be small enough. Denote
679
+ C =
680
+ sup
681
+ x∈ ¯
682
+ M,u(x)∈[min ¯
683
+ M u−,max ¯
684
+ M u+]
685
+ |F(x, u(x))|;
686
+ D1 =
687
+ sup
688
+ x∈ ¯
689
+ M,u(x)∈[min ¯
690
+ M u−,max ¯
691
+ M u+]
692
+ |G(x, u(x))| ;
693
+ D2 =
694
+ sup
695
+ x∈ ¯
696
+ M,u(x)∈[min ¯
697
+ M u−,max ¯
698
+ M u+]
699
+ |∇G(x, u(x))| ;
700
+ (21)
701
+ We require that G(·, u+), D1, D2 satisfies
702
+ ∥(B − σ) u+ + G (·, u+)∥W 1,q(M,g) ⩽ 1;
703
+ (B − σ) · γ′
704
+ ��
705
+ A max
706
+ ¯
707
+ M (|u+|, |u−|) + C
708
+
709
+ · Volg(M)
710
+ 1
711
+ q + 1
712
+
713
+ + D1 · Volg(M)
714
+ 1
715
+ q
716
+ + D2 · γ′
717
+ ��
718
+ A max
719
+ ¯
720
+ M (|u+|, |u−|) + C
721
+
722
+ · Volg(M)
723
+ 1
724
+ q + 1
725
+
726
+ ⩽ 1.
727
+ (22)
728
+ By (7) and the first inequality in (22), we observe from the PDE (19) with k = 1 that
729
+ ∥u1∥W 2,q(M,g) ⩽ γ′ �
730
+ ∥Au+ + F(·, u+)∥Lq(M,g) + ∥(B − σ) u+ + G (·, u+)∥W 1,q(M,g)
731
+
732
+ ⩽ γ′
733
+ ��
734
+ A max
735
+ ¯
736
+ M |u+| + C
737
+
738
+ · Volg(M)
739
+ 1
740
+ q + 1
741
+
742
+ ⩽ γ′
743
+ ��
744
+ A max
745
+ ¯
746
+ M (|u+|, |u−|) + C
747
+
748
+ · Volg(M)
749
+ 1
750
+ q + 1
751
+
752
+ .
753
+ Inductively, assume that
754
+ ∥uk∥W 2,q(M,g) ⩽ γ′
755
+ ��
756
+ A max
757
+ ¯
758
+ M (|u+|, |u−|) + C
759
+
760
+ · Volg(M)
761
+ 1
762
+ q + 1
763
+
764
+ .
765
+ (23)
766
+ To check ∥uk+1∥W 2,q(M,g), we apply the W s,q-type elliptic estimate with the solution of (19) again,
767
+ ∥uk+1∥W 2,q(M,g) ⩽ γ′ �
768
+ ∥Auk + F(·, uk)∥Lq(M,g) + ∥(B − σ) uk + G (·, uk)∥W 1,q(M,g)
769
+
770
+ ⩽ γ′
771
+ ��
772
+ A max
773
+ ¯
774
+ M (|u+|, |u−|) + C
775
+
776
+ · Volg(M)
777
+ 1
778
+ q
779
+
780
+ + γ′ �
781
+ (B − σ)∥uk∥W 1,q(M,g) + ∥G(·, uk)∥Lq(M,g) + ∥∇G(·, uk)∥Lq(M,g)
782
+
783
+ ⩽ γ′
784
+ ��
785
+ A max
786
+ ¯
787
+ M (|u+|, |u−|) + C
788
+
789
+ · Volg(M)
790
+ 1
791
+ q
792
+
793
+ +
794
+
795
+ γ′�2 (B − σ)
796
+
797
+ A max
798
+ ¯
799
+ M ((|u+|, |u−|) + C) · Volg(M)
800
+ 1
801
+ q + 1
802
+
803
+ + γ′D1 · Volg(M)
804
+ 1
805
+ q +
806
+
807
+ γ′�2 D2
808
+
809
+ A max
810
+ ¯
811
+ M ((|u+|, |u−|) + C) · Volg(M)
812
+ 1
813
+ q + 1
814
+
815
+ ⩽ γ′
816
+ ��
817
+ A max
818
+ ¯
819
+ M (|u+|, |u−|) + C
820
+
821
+ · Volg(M)
822
+ 1
823
+ q + 1
824
+
825
+ .
826
+ It turns that ∥uk∥W 2,q(M,g) is uniformly bounded. The rest of the argument, in applying Arzela-
827
+ Ascoli, the monotonicity of the sequence, and the elliptic regularity, is essentially the same as in
828
+ [18, Thm. 2.4]. We omit the details here.
829
+ In conclusion, the sequence uk converges classically to a smooth function u which solves (17). In
830
+ addition, u− ⩽ u ⩽ u+ pointwise on ¯
831
+ M.
832
+
833
+
834
+ 10
835
+ J. XU
836
+ Remark 2.3. Theorem 2.2 is a special case of Theorem 2.3 by taking F(·, u) = −Rgu+Sup−1 and
837
+ G(·, u) = −
838
+ 2
839
+ p−2hgu +
840
+ 2
841
+ p−2Hu
842
+ p
843
+ 2 .
844
+ 3. Prescribed Scalar and Mean Curvature Functions under Pointwise Conformal
845
+ Deformation When η1 < 0
846
+ Recall the Yamabe equation with Robin condition
847
+ − a∆gu + Rgu = Sup−1 in M, ∂u
848
+ ∂ν +
849
+ 2
850
+ p − 2hgu =
851
+ 2
852
+ p − 2Hu
853
+ p
854
+ 2 on ∂M.
855
+ (24)
856
+ In this section, we consider the existence of the solution of (24) for given functions S, H ∈ C∞( ¯
857
+ M),
858
+ provided that η1 < 0. In particular, we will discuss the following cases:
859
+ (i). S < 0 in M, and H ⩽ 0 everywhere on ∂M, H ̸≡ 0, with η1 < 0;
860
+ (ii). S < 0 in M, and H > 0 somewhere on ∂M, with η1 < 0;
861
+ (iii). S changes sign in M, and H is arbitrary on ∂M, with η1 < 0.
862
+ Note that the Case (ii) above covers the possibilities when H > 0 everywhere on ∂M, or
863
+ ´
864
+ ∂M HdS >
865
+ 0. Note also that the case S < 0 everywhere in M and H = 0 on ∂M has been discussed in [19].
866
+ For Case (iii), obviously we have to impose some restrictions on S and H, as we shall see later;
867
+ there is no free choice of S especially, due to Kazdan and Warner [10]. The first result concerns
868
+ the Case (i).
869
+ Theorem 3.1. Let ( ¯
870
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
871
+ dim ¯
872
+ M ⩾ 3. Let S1 < 0 be any smooth function on
873
+ ¯
874
+ M. Let H1 ∈ C∞( ¯
875
+ M) such that H1 < 0
876
+ everywhere on ∂M. If η1 < 0, then there exists a small enough constant c > 0 such that (24)
877
+ admits a positive solution u ∈ C∞( ¯
878
+ M) with S = S1 and H = cH1. Equivalently, there exists a
879
+ Yamabe metric ˜g = up−2g such that R˜g = S1 and h˜g = cH1
880
+ ����
881
+ ∂M
882
+ .
883
+ Proof. Due to the proof of Han-Li conjecture [17, Theorem], we may assume that hg = h > 0 and
884
+ Rg < 0. Since η1 < 0, it follows that η1,β < 0 with small enough positive constant β > 0, due to
885
+ Proposition 2.3. Any constant multiple of ϕ solves (15). Denote φ = δϕ, we choose the constant
886
+ δ > 0 small enough so that
887
+ η1,β inf
888
+ ¯
889
+ M ϕ ⩾ δp−2 · inf
890
+ ¯
891
+ M S1 · sup
892
+ ¯
893
+ M
894
+ ϕp−1.
895
+ This can be done since both η1,β and S1 are negative functions. It follows that
896
+ −a∆gφ + Rgφ = η1,βφ ⩽ S1φp−1 in M.
897
+ Fix this δ. We check the boundary condition
898
+ −∂φ
899
+ ∂ν +
900
+ 2
901
+ p − 2hgφ = −β ·
902
+ 2
903
+ p − 2φ ⩽
904
+ 2
905
+ p − 2 · (cH1) φ
906
+ p
907
+ 2
908
+ for small enough positive constant c > 0. Again it works since both −β and H1 are negative. We
909
+ set
910
+ u− := φ.
911
+ (25)
912
+ The argument above shows that u− is a sub-solution of (24) with S = S1 and H = cH1 for small
913
+ enough c. For super-solution, we set
914
+ u+ := C ≫ 1.
915
+ (26)
916
+ When C large enough, we have
917
+ −a∆gu+ + Rgu+ = RgC ⩾ S1Cp−1 in M.
918
+
919
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
920
+ 11
921
+ Since H1 < 0, it is straightforward to check that for any c > 0, we have
922
+ −∂u+
923
+ ∂ν +
924
+ 2
925
+ p − 2hgu+ ⩾ 0 >
926
+ 2
927
+ p − 2 (cH1) u
928
+ p
929
+ 2
930
+ +.
931
+ We can enlarge C so that C ⩾ sup ¯
932
+ M u−. Lastly we shrink c if necessary since we require the
933
+ smallness of the sup-norm of the prescribing mean curvature function in the proof of Theorem 2.2.
934
+ Since 0 < u− ⩽ u+ and both u+ and u− are smooth functions, we conclude by Theorem 2.2 that
935
+ (24) has a positive solution u ∈ C∞( ¯
936
+ M) with S = S1 and H = cH1 for small enough c > 0.
937
+
938
+ We now consider the Case (ii) at the beginning of this section. Actually the proof is very similar
939
+ to Theorem 3.1 above.
940
+ Theorem 3.2. Let ( ¯
941
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
942
+ dim ¯
943
+ M ⩾ 3. Let S2 < 0 be any smooth function on
944
+ ¯
945
+ M. Let H2 ∈ C∞( ¯
946
+ M) such that H2 > 0
947
+ somewhere on ∂M. If η1 < 0, then there exists a small enough constant c > 0 such that (24)
948
+ admits a positive solution u ∈ C∞( ¯
949
+ M) with S = S2 and H = cH2. Equivalently, there exists a
950
+ Yamabe metric ˜g = up−2g such that R˜g = S2 and h˜g = cH2
951
+ ����
952
+ ∂M
953
+ .
954
+ Proof. The choice of the sub-solution is exactly the same as in Theorem 3.1. When we fix the sub-
955
+ solution u−, we choose u+ = C ≫ 1 with C ⩾ u−, also large enough so that the same argument in
956
+ Theorem 3.1 holds. Fix this C from now on. The only difference is that since H2 > 0 somewhere,
957
+ we may need to shrink c, if necessary, so that
958
+ ∂C
959
+ ∂ν +
960
+ 2
961
+ p − 2hgC ⩾
962
+ 2
963
+ p − 2 · sup
964
+ ∂M
965
+ (cH2)C
966
+ p
967
+ 2
968
+ The rest of the argument is exactly the same as in Theorem 3.1.
969
+
970
+ Remark 3.1. The method of monotone iteration scheme has its limits, as we cannot obtain the
971
+ prescribed mean curvature to be H, due to the technical issue, see [18, Thm. 2.4].
972
+ We now discuss the Case (iii). The following argument is inspired by Kazdan and Warner [10].
973
+ When η1 < 0, Kazdan and Warner showed that the key is to get the super-solution of (24), if we
974
+ are not using the variational method but instead the monotone iteration scheme. Next result shows
975
+ that a super-solution of (24) can be converted to another relation. We point out that the following
976
+ result is not specific for η1 < 0 case only.
977
+ Lemma 3.1. Let ( ¯
978
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
979
+ dim ¯
980
+ M ⩾ 3.
981
+ Let S, H ∈ C∞( ¯
982
+ M) be given functions.
983
+ Then there exists some positive function
984
+ u ∈ C∞( ¯
985
+ M) satisfying
986
+ − a∆gu + Rgu ⩾ Sup−1 in M, ∂u
987
+ ∂ν +
988
+ 2
989
+ p − 2hgu ⩾
990
+ 2
991
+ p − 2Hu
992
+ p
993
+ 2 on ∂M
994
+ (27)
995
+ if and only if there exists some positive function w ∈ C∞( ¯
996
+ M) satisfying
997
+ − a∆gw + (2 − p)Rgw + (p − 1)a
998
+ p − 2
999
+ · |∇gw|2
1000
+ w
1001
+ ⩽ (2 − p)S in M, ∂w
1002
+ ∂ν − 2hgw ⩽ −2Hw
1003
+ 1
1004
+ 2.
1005
+ (28)
1006
+ Moreover, the equality in (27) holds if and only if the equality in (28) holds.
1007
+ Proof. Assume that there is a positive function u ∈ C∞(M) that satisfies (27). Define
1008
+ w = u2−p.
1009
+ Note that 2 − p = −
1010
+ 4
1011
+ n−2 < 0 since n ⩾ 3 by hypothesis. We compute that
1012
+ ∇w = (2 − p)u1−p∇u ⇔ ∇u = up−1(2 − p)−1∇w,
1013
+
1014
+ 12
1015
+ J. XU
1016
+ and
1017
+ ∆gw = (2 − p)u1−p∆gu + (2 − p)(1 − p)u−p|∇gu|2.
1018
+ By the inequality (27), we have
1019
+ a∆gw = (2 − p)u1−p (a∆gu) + a(2 − p)(1 − p)u−p|∇gu|2
1020
+ ⩾ (p − 2)u1−p �
1021
+ −Rgu + Sup−1�
1022
+ + a(2 − p)(1 − p)(2 − p)−2u2p−2u−p|∇gv|2
1023
+ = (p − 2)S + (2 − p)Rgu1−p + a(p − 1)
1024
+ p − 2 up−2|∇gv|2
1025
+ = (p − 2)S + (2 − p)Rgw + a(p − 1)
1026
+ p − 2
1027
+ |∇gw|2
1028
+ w
1029
+ .
1030
+ Shifting (p − 2)S to the left side and a∆gw to the right side, we get the first part of the inequality
1031
+ (28). For the boundary condition, recall that u = w
1032
+ 1
1033
+ 2−p and p =
1034
+ 2n
1035
+ n−2, it follows that
1036
+ ∂u
1037
+ ∂ν +
1038
+ 2
1039
+ p − 2Hu
1040
+ p
1041
+ 2 ⩾
1042
+ 2
1043
+ p − 2hgu ⇔
1044
+ 1
1045
+ 2 − pw
1046
+ 1
1047
+ 2−p −1 ∂w
1048
+ ∂ν +
1049
+ 2
1050
+ p − 2hgw
1051
+ 1
1052
+ 2−p ⩾
1053
+ 2
1054
+ p − 2Hw
1055
+ p
1056
+ 2(2−p)
1057
+ ⇔ − n − 2
1058
+ 4
1059
+ w− n
1060
+ 4 − 1
1061
+ 2 ∂w
1062
+ ∂ν + n − 2
1063
+ 2
1064
+ hgw− n
1065
+ 4 + 1
1066
+ 2 ⩾ n − 2
1067
+ 2
1068
+ Hw− n
1069
+ 4
1070
+ ⇔∂w
1071
+ ∂ν − 2hgw ⩽ −2Hw
1072
+ 1
1073
+ 2 .
1074
+ Hence the second part of (28) holds. It is clear that the equality holds if an only if all inequalities
1075
+ above are equalities.
1076
+ If we assume (28) for some w, we just define u = w
1077
+ 1
1078
+ 2−p . The argument is very similar and we
1079
+ omit the details.
1080
+
1081
+ We now introduce the result of prescribing scalar and mean curvature functions for Case (iii),
1082
+ with a technical restriction very similar to the condition given by Kazdan and Warner.
1083
+ This
1084
+ technical condition, in principle, is to show the positivity of the function that satisfies (28). Due
1085
+ to the Han-Li conjecture [17, Theorem], we may assume that the initial metric g has Rg = λ < 0
1086
+ and hg = ζ > 0, since η1 < 0. Before we start with the special case, recall that if there exists a
1087
+ constant q > n, and some function u ∈ C∞( ¯
1088
+ M) satisfies
1089
+ ∥u∥W 2,q(M,g) ⩽ γ′ �
1090
+ ∥F1∥Lq(M,g) + ∥F2∥W 1,q(M,g)
1091
+
1092
+ for some functions F1 ∈ Lq(M, g) and F2 ∈ W 1,q(M, g), the H¨older estimates implies that
1093
+ ∥u∥L∞( ¯
1094
+ M) + ∥∇u∥L∞( ¯
1095
+ M) ⩽ γ
1096
+
1097
+ ∥F1∥Lq(M,g) + ∥F2∥W 1,q(M,g)
1098
+
1099
+ .
1100
+ (29)
1101
+ This inequality is due to the Sobolev embedding in [1, §2].
1102
+ Theorem 3.3. Let ( ¯
1103
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
1104
+ dim ¯
1105
+ M ⩾ 3. Assume that η1 < 0, Rg = λ < 0 and hg = ζ > 0 for some constants λ and ζ. Let
1106
+ S3, H3 ∈ C∞( ¯
1107
+ M) and q > n be a positive integer. Let γ be the constant in the estimate (29). Set
1108
+ D = (p−1)a
1109
+ p−2 . If there exists a function F ∈ C∞( ¯
1110
+ M) and a positive constant A > 0, such that
1111
+ (2 − p)S3 ⩾ F on ∂M, ∥F − A∥Lq(M,g) ⩽
1112
+ A
1113
+ 2γ (1 + (D + 1) (2 − p)λ),
1114
+ (30)
1115
+ then there exists a small enough constant c > 0 such that (24) admits a positive solution u ∈ C∞( ¯
1116
+ M)
1117
+ with S = S3 and H = cH3. Equivalently, there exists a Yamabe metric ˜g = up−2g such that R˜g = S3
1118
+ and h˜g = cH3
1119
+ ����
1120
+ ∂M
1121
+ .
1122
+
1123
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
1124
+ 13
1125
+ Proof. In this proof, we always denote Rg = λ and hg = ζ. We construct the super-solution of
1126
+ (24) first.
1127
+ Due to Lemma 3.1, it is equivalent to show the existence of some positive function
1128
+ w ∈ C∞( ¯
1129
+ M) such that (28) holds for S = S3 and H = cH3 for some constant c. Take
1130
+ δ :=
1131
+ A
1132
+ (1 + (D + 1)(2 − p)λ) > 0.
1133
+ We also choose some negative constant
1134
+ δ′ = −
1135
+ δ
1136
+ 2γVolg(M)
1137
+ 1
1138
+ q
1139
+ < 0.
1140
+ By standard elliptic theory, there exists a unique solution w of the following PDE
1141
+ −a∆gw + (2 − p)λw = F − δ in M, ∂w
1142
+ ∂ν = δ′ on ∂M.
1143
+ The uniqueness comes from the fact that (2−p)λ > 0, which implies the invertibility of the operator
1144
+
1145
+ −a∆g + (2 − p)λ, ∂
1146
+ ∂ν
1147
+
1148
+ . Clearly the constant (D + 1)δ solves the PDE
1149
+ −a∆g((D + 1)δ) + (2 − p)λ · ((D + 1)δ) = (2 − p)λ · ((D + 1)δ) in ∂M, ∂((D + 1)δ)
1150
+ ∂ν
1151
+ = 0 on ∂M.
1152
+ Denote
1153
+ w0 := w − (D + 1)δ.
1154
+ The function w0 satisfies
1155
+ −a∆gw0 + (2 − p)λw0 = F − δ − (D + 1)(2 − p)λδ = F − A in M,
1156
+ ∂w
1157
+ ∂ν = δ′ on ∂M.
1158
+ (31)
1159
+ The first line in (31) is due to the definition of δ. Since the differential operator with the boundary
1160
+ operator is invertible, we apply W s,q-type elliptic estimates (7) as well as the estimates of (29),
1161
+ ∥w0∥L∞( ¯
1162
+ M) + ∥∇w0∥L∞( ¯
1163
+ M) ⩽ γ
1164
+
1165
+ ∥F − A∥Lq(M,g) + ∥δ′∥W 1,q(M,g)
1166
+
1167
+ ⩽ γ
1168
+
1169
+ A
1170
+ 2γ (1 + (D + 1)(2 − p)λ) + |δ′| · Volg(M)
1171
+ 1
1172
+ q
1173
+
1174
+ ⩽ δ.
1175
+ (32)
1176
+ The last inequality is due to the definitions of δ and δ′. By definition of w0, the inequality (32)
1177
+ implies
1178
+ ∥w − (D + 1)δ∥L∞( ¯
1179
+ M) ⩽ δ, ∥∇w∥L∞(¯Ω) ⩽ δ.
1180
+ It follows that
1181
+ 0 < Dδ ⩽ w ⩽ (D + 2)w on ¯
1182
+ M, sup
1183
+ ¯
1184
+ M
1185
+ |∇w| ⩽ δ ⇒ (p − 1)a
1186
+ p − 2
1187
+ · |∇w|2
1188
+ w
1189
+ ⩽ δ on ¯
1190
+ M.
1191
+ (33)
1192
+ With (36), (33), we have
1193
+ − a∆gw + (2 − p)λw + (p − 1)a
1194
+ p − 2
1195
+ · |∇w|2
1196
+ w
1197
+ = F − δ + (p − 1)a
1198
+ p − 2
1199
+ · |∇w|2
1200
+ w
1201
+ ⩽(2 − p)S3;
1202
+ ∂w
1203
+ ∂ν − 2hgw = δ′ − 2hgw ⩽ −2 · (cH3) w
1204
+ 1
1205
+ 2 .
1206
+ The last inequality holds for small enough constant c > 0, regardless of the sign of H3 since
1207
+ δ′ − 2hgw < 0 by set-up. By (33) again, we conclude that w > 0 on
1208
+ ¯
1209
+ M. By Lemma 3.1, the
1210
+ positive, smooth function
1211
+ u = w
1212
+ 1
1213
+ 2−p
1214
+
1215
+ 14
1216
+ J. XU
1217
+ is a super-solution of (24) with S = S3 and H = cH3. Note that u is still a super-solution if we
1218
+ make c smaller.
1219
+ For sub-solution, we apply the perturbed eigenvalue problem in Proposition 2.3 again. There
1220
+ exists a small enough constant β > 0 such that
1221
+ − a∆gϕ + λϕ = η1,βϕ in M, ∂ϕ
1222
+ ∂ν +
1223
+ 2
1224
+ p − 2 (ζ + β) ϕ = 0 on ∂M.
1225
+ (34)
1226
+ Any scaling of ϕ solves (35). Set the positive constant ξ ≪ 1 such that
1227
+ φ := ξϕ ⩽ u on ¯
1228
+ M
1229
+ for the fixed super-solution u defined just above. We shrink ξ further, if necessary, such that
1230
+ η1,β (ξϕ) ⩽ S3 (ξϕ)p−1 in M, −
1231
+ 2
1232
+ p − 2 · β (ξϕ) ⩽
1233
+ 2
1234
+ p − 2 · (cH3) · (ξϕ)
1235
+ p
1236
+ 2 on ∂M.
1237
+ (35)
1238
+ Note that the boundary condition holds for every c, as long as we take ξ small enough. We point
1239
+ out that the choice of the constant c depends on the construction of the super-solution as well as
1240
+ the technical condition of the monotone iteration scheme, which only depends on the super-solution
1241
+ but not the sub-solution, see Equation (19) in [18]. Thus we can choose c first, then determine ξ.
1242
+ It follows that φ is a sub-solution of (24) with S = S3 and H = cH3. Furthermore, 0 < φ ⩽ u
1243
+ on ¯
1244
+ M. Applying Theorem 2.2, we conclude that there exists a positive function u ∈ C∞( ¯
1245
+ M) as
1246
+ desired.
1247
+
1248
+ The general case when η1 < 0 is a straightforward consequence of the result above.
1249
+ Corollary 3.1. Let ( ¯
1250
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
1251
+ dim ¯
1252
+ M ⩾ 3. Let S4, H4 ∈ C∞( ¯
1253
+ M) and q > n be a positive integer. Let γ be the constant in the
1254
+ estimate (29) and λ be some negative constant. Set D = (p−1)a
1255
+ p−2 . Assume that η1 < 0. If there
1256
+ exists a function F ∈ C∞( ¯
1257
+ M) and a positive constant A > 0, such that
1258
+ (2 − p)S4 ⩾ F on ∂M, ∥F − A∥Lq(M,g) ⩽
1259
+ A
1260
+ 2γ (1 + (D + 1) (2 − p)λ),
1261
+ (36)
1262
+ then there exists a small enough constant c > 0 such that (24) admits a positive solution u ∈ C∞( ¯
1263
+ M)
1264
+ with S = S4 and H = cH4. Equivalently, there exists a Yamabe metric ˜g = up−2g such that R˜g = S4
1265
+ and h˜g = cH4
1266
+ ����
1267
+ ∂M
1268
+ .
1269
+ Proof. By the result of the Han-Li conjecture [17, Theorem], there exists a conformal metric g1 =
1270
+ vp−2g such that Rg1 = λ and hg1 = ζ. We then apply Theorem 3.3 for the metric g1, i.e. there
1271
+ exists ˜g = up−2g1 with R˜g = S4 and h˜g = cH4 with small enough c > 0. The conformal change
1272
+ ˜g = (uv)p−2 g
1273
+ is the desired metric.
1274
+
1275
+ 4. Prescribed Scalar and Mean Curvature Functions for Conformal Equivalent
1276
+ Metrics When η1 < 0
1277
+ Inspired by the “Trichotomy Theorem” on closed manifolds, we would like to discuss the pre-
1278
+ scribing scalar and mean curvature problem on ( ¯
1279
+ M, g), n = dim ¯
1280
+ M ⩾ 3, but not restricted in a
1281
+ conformal class [g] only. Instead, we are interested in the conformally equivalent metrics.
1282
+
1283
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
1284
+ 15
1285
+ Definition 4.1. Let ( ¯
1286
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, we say
1287
+ that a metric ˜g is conformally equivalent to the metric g if there exists a positive, smooth function
1288
+ u ∈ C∞( ¯
1289
+ M) and a diffeomorphism φ : ¯
1290
+ M → ¯
1291
+ M such that
1292
+ φ∗˜g = up−2g.
1293
+ Within in a conformal class, the prescribing scalar and mean curvature problem for given func-
1294
+ tions S, H ∈ C∞( ¯
1295
+ M) is reduced to the PDE (1). For conformlaly equivalent metrics, the prescribing
1296
+ scalar and mean curvature problem is reduced to the existence of a positive, smooth solution of the
1297
+ following PDE
1298
+ − a∆gu + Rgu = (S ◦ φ) up−1 in M, ∂u
1299
+ ∂ν +
1300
+ 2
1301
+ p − 2hgu =
1302
+ 2
1303
+ p − 2 · (H ◦ φ) · u
1304
+ p
1305
+ 2 on ∂M.
1306
+ (37)
1307
+ Our next result extends the result of prescribing scalar curvature problem on closed manifolds with
1308
+ dimensions at least 3 [10, Thm. 3.3] to compact manifolds with non-empty smooth boundaries,
1309
+ provided that the first eigenvalue η1 of the conformal Laplacian with Robin boundary condition is
1310
+ negative. The method is essentially due to Kazdan and Warner [8, 10].
1311
+ Theorem 4.1. Let ( ¯
1312
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
1313
+ dim ¯
1314
+ M ⩾ 3. Let S5 be any smooth function on
1315
+ ¯
1316
+ M that is negative somewhere in M. Let H5 ∈
1317
+ C∞( ¯
1318
+ M) and q > n be a positive integer. If η1 < 0, then there exists a small enough constant c > 0
1319
+ and a diffeomorphism φ : ¯
1320
+ M → ¯
1321
+ M such that (37) admits a positive solution u ∈ C∞( ¯
1322
+ M) with S = S5
1323
+ and H = cH5. Equivalently, there exists a conformally equivalent metric ˜g =
1324
+
1325
+ φ−1�∗ �
1326
+ up−2g
1327
+
1328
+ such
1329
+ that R˜g = S5 and h˜g = cH5
1330
+ ����
1331
+ ∂M
1332
+ .
1333
+ Proof. By Han-Li conjecture [17, Theorem], we may assume that Rg = λ < 0 and hg = ζ > 0
1334
+ for some constants λ, ζ.
1335
+ Fix some constant q > n.
1336
+ Due to Theorem 3.3, it suffices to find a
1337
+ diffeomorphism φ : ¯
1338
+ M → ¯
1339
+ M, a smooth function F ∈ C∞( ¯
1340
+ M), a positive constant A > 0 and a small
1341
+ enough positive constant c such that
1342
+ (2 − p)S5 ◦ φ ⩾ F on ¯
1343
+ M, ∥F − A∥Lq(M,g) ⩽
1344
+ A
1345
+ 2γ (1 + (D + 1)(2 − p)λ);
1346
+ (38)
1347
+ in addition, sup ¯
1348
+ M c (H ◦ φ) is small enough. Here γ is the constant in the estimate (29), the constant
1349
+ D is defined to be
1350
+ D = (p − 1)a
1351
+ p − 2 .
1352
+ We determine φ, F and A first. If S5 < 0 everywhere on ¯
1353
+ M, we just choose φ to be the identity
1354
+ map and set
1355
+ F = A = (2 − p) max
1356
+ ¯
1357
+ M S5.
1358
+ It is straightforward to check that (38) holds.
1359
+ If S5 ⩾ 0 somewhere and changes sign, we choose A first to be any positive constant such that
1360
+ 0 < A < (2 − p) min
1361
+ ¯
1362
+ M S5.
1363
+ (39)
1364
+ Just note that (2 − p) < 0. We pick interior open submanifolds U, V ⊂ M such that
1365
+ V ⊂ ¯V ⊂ U ⊂ M ⊂ ¯
1366
+ M.
1367
+ In particular, we require that
1368
+ Volg(U − V ) ⩽
1369
+
1370
+
1371
+ A
1372
+ 2γ (1 + (D + 1)(2 − p)λ) ·
1373
+
1374
+ (2 − p)∥S3∥L∞( ¯
1375
+ M) − A
1376
+
1377
+
1378
+
1379
+ q
1380
+ .
1381
+ (40)
1382
+
1383
+ 16
1384
+ J. XU
1385
+ We select the diffeomorphism φ such that
1386
+ (2 − p)S3 ◦ φ > A in U.
1387
+ (41)
1388
+ We then take the function F to be
1389
+ F = A in V ;
1390
+ (2 − p) max
1391
+ ¯
1392
+ M S3 ◦ φ ⩽ F ⩽ A in U − V ;
1393
+ F = (2 − p) max
1394
+ ¯
1395
+ M S3 ◦ φ in ¯
1396
+ M − U.
1397
+ (42)
1398
+ Clearly F ⩽ (2 − p)S3 ◦ φ on ¯
1399
+ M by (42). The function F only differs with A in U − V , by (40), it
1400
+ is immediate to check that the second inequality in (38) holds.
1401
+ Lastly we choose c so that the condition in Theorem 2.2 holds for the function S3 ◦ φ, i.e.
1402
+ c sup ¯
1403
+ M|H5| is small enough.
1404
+ The same c applies for the smallness of c sup ¯
1405
+ M|H5 ◦ φ| since the
1406
+ diffeomorphism does not change the extremal values of a function. Therefore the function S3 ◦ φ
1407
+ and cH5 ◦ φ can be realized as prescribed scalar and mean curvature functions, respectively, for
1408
+ some metric φ∗˜g = up−2g where u is positive and smooth on ¯
1409
+ M. Equivalently, S5 and cH5 can
1410
+ be realized as prescribed scalar and mean curvature functions, respectively, for some metric ˜g =
1411
+
1412
+ φ−1�∗ up−2g.
1413
+
1414
+ Remark 4.1. The result of Theorem 4.1 indicates that on ( ¯
1415
+ M, g) with n = dim ¯
1416
+ M ⩾ 3, any
1417
+ function that is negative somewhere can be realized as a scalar curvature function of some metric
1418
+ g, meanwhile the mean curvature function of g can be some small enough scaling of any smooth
1419
+ function, provided that the manifold admits a metric with negative first eigenvalue of the conformal
1420
+ Laplacian, or equivalently, negative Yamabe invariant [6, §1].
1421
+ 5. Prescribed Gauss and Geodesic Curvature Functions When χ( ¯
1422
+ M) < 0
1423
+ In this section, we discuss the prescribing Gauss and geodesic curvatures problem within a
1424
+ conformal class [g] of compact manifolds ( ¯
1425
+ M, g) with non-empty smooth boundary ∂M, provided
1426
+ that χ( ¯
1427
+ M) < 0 and n = dim ¯
1428
+ M = 2. This is a 2-dimensional analogy of prescribing scalar and
1429
+ mean curvatures problem with η1 < 0, provided that the dimension is at least 3.
1430
+ Let K, σ ∈ C∞( ¯
1431
+ M) be given functions. This type of Kazdan-Warner problem is reduced to the
1432
+ existence of a smooth solution u of the following PDE
1433
+ − a∆gu + Kg = Ke2u in M, ∂u
1434
+ ∂ν + σg = σeu on ∂M.
1435
+ (43)
1436
+ Here Kg and σg are Gaussian and geodesic curvatures of g, respectively. The solvability of this
1437
+ PDE implies that the metric ˜g = e2ug has Gauss curvature K˜g = K and geodesic curvature σ˜g = σ.
1438
+ We mainly discuss to cases:
1439
+ (i). K ⩽ 0 everywhere in ¯
1440
+ M, and arbitrary σ, with χ( ¯
1441
+ M) < 0;
1442
+ (ii). K > 0 somewhere in ¯
1443
+ M and changes sign, σ is an arbitrary function, with χ( ¯
1444
+ M) < 0.
1445
+ We would like to apply the monotone iteration scheme to solve (43), it is equivalent to construct the
1446
+ sub- and super-solutions of (43). The key is to construct the super-solution. As in §3, we convert
1447
+ the super-solution of (43) into another inequality involving derivatives.
1448
+ Lemma 5.1. Let ( ¯
1449
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
1450
+ dim ¯
1451
+ M = 2. Let K, σ ∈ C∞( ¯
1452
+ M) be given functions. Then there exists some function u ∈ C∞( ¯
1453
+ M)
1454
+ satisfying
1455
+ − ∆gu + Kg ⩾ Ke2u in M, ∂u
1456
+ ∂ν + σg ⩾ σeu on ∂M
1457
+ (44)
1458
+
1459
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
1460
+ 17
1461
+ if and only if there exists some positive function w ∈ C∞( ¯
1462
+ M) satisfying
1463
+ − ∆gw − 2wKg + |∇gw|2
1464
+ w
1465
+ ⩽ −2K in M, ∂w
1466
+ ∂ν − 2wσg ⩽ −2σw
1467
+ 1
1468
+ 2 on ∂M.
1469
+ (45)
1470
+ Moreover, the equality in (44) holds if and only if the equality in (45) holds; and the inequality in
1471
+ (44) is in the reverse direction if and only if the inequality in (45) is in the reverse direction.
1472
+ Proof. Assume (44) for some function u first. Define
1473
+ w := e−2u
1474
+ We observe that
1475
+ ∇gw = −2e−2u∇gu ⇒ ∇gu = −1
1476
+ 2e2u∇gw,
1477
+ ∆gw = −2e−2u∆gu + 4e−2u|∇gu|2 = −2e−2u∆gu + e2u|∇gw|2.
1478
+ Thus we have
1479
+ −∆gw = 2e−2u∆gu − e2u|∇gw|2 ⩽ 2e−2u �
1480
+ Kg − Ke2u�
1481
+ − |∇gw|2
1482
+ w
1483
+ = 2wKg − 2K − |∇gw|2
1484
+ w
1485
+ ⇒ − ∆gw − 2wKg + |∇gw|2
1486
+ w
1487
+ ⩽ −2K in M.
1488
+ For the boundary condition, we have
1489
+ ∂w
1490
+ ∂ν = ∂e−2u
1491
+ ∂ν
1492
+ = −2e−2u ∂u
1493
+ ∂ν ⩽ −2e−2u (−σg + σeu)
1494
+ = 2wσg − 2σw
1495
+ 1
1496
+ 2
1497
+ ⇒∂w
1498
+ ∂ν − 2wσg ⩽ −2σw
1499
+ 1
1500
+ 2 on ∂M.
1501
+ Therefore (45) holds for w = e−2u > 0 on ¯
1502
+ M. It is clear that equality holds when all inequalities
1503
+ above are equalities. It is also straightforward to see that the inequalities are in the reverse directions
1504
+ if and only if the inequalities are in the reverse directions in each step above.
1505
+ For the opposite direction, we assume (45) holds for some positive, smooth function w. Define
1506
+ u = −1
1507
+ 2 log w.
1508
+ We can show that u satisfies (44).
1509
+ The argument is quite similar to above and we omit the
1510
+ details.
1511
+
1512
+ Due to the uniformization theorem, we may assume Kg = −1 and σg = 0 in (43) from now on, as
1513
+ our model case up to some pointwise conformal change, provided that χ( ¯
1514
+ M) < 0. In 2-dimensional
1515
+ case, we also have the W s,q-type estimates from Theorem 2.1. We choose q = 3, the estimate in
1516
+ (7) plus the Sobolev embedding into H¨older space, the inequality in (29) becomes
1517
+ ∥u∥L∞( ¯
1518
+ M) + ∥∇u∥L∞( ¯
1519
+ M) ⩽ γ
1520
+
1521
+ ∥F1∥L3(M,g) + ∥F2∥W 1,3(M,g)
1522
+
1523
+ .
1524
+ (46)
1525
+ Here F1, F2 and u comes from the PDE (6) with the operators L = −∆g + 2 and B =
1526
+
1527
+ ∂ν , so is the
1528
+ constant γ. Our main result of this section is the following, which covers both Case (i) and Case
1529
+ (ii) at the beginning of this section.
1530
+
1531
+ 18
1532
+ J. XU
1533
+ Theorem 5.1. Let ( ¯
1534
+ M, g) be a compact Riemann surface with non-empty smooth boundary ∂M.
1535
+ Let K1, σ1 ∈ C∞( ¯
1536
+ M) be given functions. Let γ be the constant in the estimate (46). Assume that
1537
+ χ( ¯
1538
+ M) < 0. If there exists a function F ∈ C∞( ¯
1539
+ M) and a positive constant A > 0, such that
1540
+ − 2K1 ⩾ F on ∂M, ∥F − A∥L3(M,g) ⩽ A
1541
+ 6γ ,
1542
+ (47)
1543
+ then there exists a small enough constant c > 0 such that (43) admits a positive solution u ∈ C∞( ¯
1544
+ M)
1545
+ with K = K1 and σ = cσ1. Equivalently, there exists a Yamabe metric ˜g = e2ug such that K˜g = K1
1546
+ and σ˜g = cσ1
1547
+ ����
1548
+ ∂M
1549
+ .
1550
+ Proof. The proof is essentially the same as in Theorem 4.1. By Lemma 43, the construction of the
1551
+ super-solution is equivalent to the construction of a function w that satisfies (45) for K1, σ1 and
1552
+ some small enough positive constant c. We set
1553
+ δ = A
1554
+ 3 , δ′ = −
1555
+ δ
1556
+ 2γVolg(M)
1557
+ 1
1558
+ 3
1559
+ .
1560
+ (48)
1561
+ There is a unique solution for the PDE
1562
+ −∆gw + 2w = F − δ in M, ∂w
1563
+ ∂ν = δ′ on ∂M.
1564
+ Define
1565
+ w0 = w − 2δ,
1566
+ it follows that w0 satisfies the PDE
1567
+ − ∆gw0 + 2w0 = F − 3δ = F − A in M, ∂w0
1568
+ ∂ν = δ′ on ∂M.
1569
+ (49)
1570
+ Apply the estimate (46) for w0 in (49), it follows that
1571
+ ∥w0∥L∞( ¯
1572
+ M) + ∥∇w0∥L∞( ¯
1573
+ M) ⩽ γ
1574
+
1575
+ ∥F − A∥L3(M,g) + ∥δ′∥W 1,3(M,g)
1576
+
1577
+ ⩽ δ.
1578
+ It follows from the definition of w0 that
1579
+ 0 < δ ⩽ w ⩽ 3δ on ¯
1580
+ M, ∥∇w∥L∞( ¯
1581
+ M) ⩽ δ.
1582
+ Therefore we conclude that
1583
+ −∆gw + 2w + |∇w|2
1584
+ w
1585
+ = F − δ + |∇w|2
1586
+ w
1587
+ ⩽ F ⩽ −2K1 in M.
1588
+ In addition, we take c small enough so that
1589
+ ∂w
1590
+ ∂ν = δ′ ⩽ −2cσ1w
1591
+ 1
1592
+ 2 on ∂M.
1593
+ This can be done since δ′ < 0. It follows that the function
1594
+ u+ := −1
1595
+ 2 log w
1596
+ is a super-solution of (43) with K = K1 and σ = cσ1. Clearly u+ ∈ C∞( ¯
1597
+ M).
1598
+ We construct a sub-solution now. Consider the PDE
1599
+ −∆gu0 = 1
1600
+ 2 in M, ∂u0
1601
+ ∂ν = C on ∂M.
1602
+ By standard elliptic PDE theory, see e.g. [14, Prop. 7.7, Ch. 4], the above PDE is solvable by some
1603
+ smooth function u0 ∈ C∞( ¯
1604
+ M) if −
1605
+ ´
1606
+ M
1607
+ 1
1608
+ 2dVolg =
1609
+ ´
1610
+ ∂M CdSg. We choose the constant C < 0 so that
1611
+ the compatibility condition just mentioned holds. Clearly
1612
+ u− := u0 + C1
1613
+
1614
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
1615
+ 19
1616
+ solves the PDE above also for any constant C1. We just choose C1 to be very negative such that
1617
+ u− ⩽ u+ on ∂M;
1618
+ In addition,
1619
+ −∆gu− − 1 = −1
1620
+ 2 ⩽ K1e2u− = K1e2u0 · e2C1 in M,
1621
+ ∂u−
1622
+ ∂ν = C ⩽ cσ1eu− = cσ1eu0 · eC1 on ∂M.
1623
+ These can be done since the constants on the left sides of the inequalities are both negative. Note
1624
+ that both the super-solution and sub-solution holds for smaller constant c by adjusting the constant
1625
+ C1 only.
1626
+ Note that when F(·, u) = K1e2u + 1, G(·, u) = cσ1eu, the condition (22) is independent of the
1627
+ sub-solution u− as we can see the very similar case in Theorem 2.2 for the Yamabe equation. Thus
1628
+ we take c small enough so that the hypotheses in Theorem 2.3 holds. It follows that there exists
1629
+ some smooth function u that solves (43) with K = K1 and σ = cσ1.
1630
+
1631
+ We can partially answer the two cases we are interested in. For Case (ii), not every function that
1632
+ changes sign can be a prescribed scalar curvature function unless it is not too positive too often. We
1633
+ show that every function that is negative everywhere can be realized as a scalar curvature function,
1634
+ meanwhile, a small enough scaling of any function can be realized as prescribed mean curvature
1635
+ function, under pointwise conformal deformation. This is Case (i).
1636
+ Corollary 5.1. Let ( ¯
1637
+ M, g) be a compact Riemann surface with non-empty smooth boundary ∂M.
1638
+ Let K2, σ2 ∈ C∞( ¯
1639
+ M) be given functions. Assume that K2 < 0 everywhere on
1640
+ ¯
1641
+ M. If χ( ¯
1642
+ M) < 0,
1643
+ then there exists a small enough constant c, a smooth function u ∈ C∞( ¯
1644
+ M) such that u solves (43)
1645
+ with K = K2 and σ = cσ2. It is equivalent to say that the metric ˜g = e2ug has Gauss curvature
1646
+ K˜g = K2 and geodesic curvature σ˜g = cσ2.
1647
+ Proof. We show that the condition (47) holds. Since K2 < 0 everywhere, we just choose
1648
+ F = A = −2 max
1649
+ ¯
1650
+ M K2 ⇒ −2K2 ⩾ F, ∥F − A∥L3(M,g) = 0.
1651
+ We just need to choose a small enough c such that the hypotheses in Theorem 5.1 and Theorem
1652
+ 2.3 hold.
1653
+
1654
+ For Case (ii), we can get a more comprehensive answer by considering the class of conformally
1655
+ equivalent metrics.
1656
+ Analogous to §4, we are looking for a metric ˜g =
1657
+
1658
+ φ−1�∗ e2ug with some
1659
+ diffeomorphism φ :
1660
+ ¯
1661
+ M →
1662
+ ¯
1663
+ M and smooth function u ∈ C∞( ¯
1664
+ M) such that the scalar and mean
1665
+ curvatures of ˜g are given functions K, σ ∈ C∞( ¯
1666
+ M), respectively. This problem is reduced to the
1667
+ PDE
1668
+ − ∆gu + Kg = (K ◦ φ) e2u in M, ∂u
1669
+ ∂ν + σgu = (σ ◦ φ) eu on ∂M.
1670
+ (50)
1671
+ Similar to Theorem 4.1 for dimensions at least 3, we introduce the following result for compact
1672
+ Riemann surfaces.
1673
+ Corollary 5.2. Let ( ¯
1674
+ M, g) be a compact Riemann surface with non-empty smooth boundary ∂M.
1675
+ Let σ3 ∈ C∞( ¯
1676
+ M) be any function and K3 ∈ C∞( ¯
1677
+ M) be a function that is negative somewhere in M.
1678
+ If χ( ¯
1679
+ M) < 0, then there exists a small enough constant c, a smooth function u ∈ C∞( ¯
1680
+ M) and a
1681
+ diffeomorphism φ : ¯
1682
+ M → ¯
1683
+ M such that u solves (50) with K = K3 and σ = cσ3. It is equivalent to
1684
+ say that the metric ˜g =
1685
+
1686
+ φ−1�∗ e2ug has Gauss curvature K˜g = K3 and geodesic curvature σ˜g = cσ3.
1687
+
1688
+ 20
1689
+ J. XU
1690
+ Proof. The proof is essentially the same as in Theorem 4.1. We determine φ, F, A first so that (47)
1691
+ holds; then determine the constant c. We may assume that K3 is negative somewhere but not
1692
+ everywhere since otherwise it is reduced to the result of Corollary 5.1.
1693
+ We choose A first to be any positive constant such that
1694
+ 0 < A < −2 min
1695
+ ¯
1696
+ M K3.
1697
+ (51)
1698
+ We pick interior open submanifolds U, V ⊂ M such that
1699
+ V ⊂ ¯V ⊂ U ⊂ M ⊂ ¯
1700
+ M.
1701
+ In particular, we require that
1702
+ Volg(U − V ) ⩽
1703
+
1704
+
1705
+ A
1706
+ 6γ ·
1707
+
1708
+ 2∥K3∥L∞( ¯
1709
+ M) − A
1710
+
1711
+
1712
+
1713
+ 3
1714
+ .
1715
+ (52)
1716
+ We select the diffeomorphism φ such that
1717
+ − 2K3 ◦ φ > A in U.
1718
+ (53)
1719
+ We then take the function F to be
1720
+ F = A in V ;
1721
+ − 2 max
1722
+ ¯
1723
+ M K3 ◦ φ ⩽ F ⩽ A in U − V ;
1724
+ F = −2 max
1725
+ ¯
1726
+ M K3 ◦ φ in ¯
1727
+ M − U.
1728
+ (54)
1729
+ Clearly F ⩽ −2K3 ◦ φ on ¯
1730
+ M by (54). The function F only differs with A in U − V , by (52), it is
1731
+ immediate to check that the second inequality in (47) holds.
1732
+ Lastly we choose c so that the condition in Theorem 2.3 holds for the function K3 ◦ φ, i.e.
1733
+ c sup ¯
1734
+ M|σ3| is small enough.
1735
+ The same c applies for the smallness of c sup ¯
1736
+ M|σ3 ◦ φ| since the
1737
+ diffeomorphism does not change the extremal values of a function. Therefore the function K3 ◦ φ
1738
+ and cσ3 ◦ φ can be realized as prescribed scalar and mean curvature functions, respectively, for
1739
+ some metric φ∗˜g = up−2g where u is positive and smooth on ¯
1740
+ M. Equivalently, K3 and cσ3 can
1741
+ be realized as prescribed scalar and mean curvature functions, respectively, for some metric ˜g =
1742
+
1743
+ φ−1�∗ up−2g.
1744
+
1745
+ Remark 5.1. The result of Corollary 5.2, combining Theorem 4.1 indicate that on ( ¯
1746
+ M, g) with
1747
+ n = dim ¯
1748
+ M ⩾ 2, any function that is negative somewhere can be realized as a scalar/Gauss
1749
+ curvature function of some metric g, meanwhile the mean/geodesic curvature function of g can be
1750
+ some small enough scaling of any smooth function, provided that the manifold admits a metric with
1751
+ negative first eigenvalue of the conformal Laplacian, or negative Euler characteristics, respectively,
1752
+ depending on the dimension of the manifold. This improve the result mentioned in Remark 4.1.
1753
+ 6. Prescribed Scalar and Mean Curvature Functions for Conformally Equivalent
1754
+ Metrics When η1 = 0
1755
+ In this section, we discuss the prescribing scalar and mean curvatures problem for metrics con-
1756
+ formally equivalent to the metric g on compact manifolds ( ¯
1757
+ M, g) with non-empty smooth boundary
1758
+ ∂M, provided that η1 = 0 and n = dim ¯
1759
+ M ⩾ 3. We gave a comprehensive study for manifolds
1760
+ with dimensions at least 3 in [18] for pointwise conformal change. Here we consider whether there
1761
+ exists some smooth function u ∈ C∞( ¯
1762
+ M) and some diffeomorphism φ :
1763
+ ¯
1764
+ M →
1765
+ ¯
1766
+ M such that the
1767
+ metric ˜g =
1768
+
1769
+ φ−1�∗ up−2g has scalar curvature S and mean curvature H for some given functions
1770
+
1771
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
1772
+ 21
1773
+ S, H ∈ C∞( ¯
1774
+ M). Since the model case for zero first eigenvalue case is Rg = hg = 0, the problem
1775
+ above is reduced to the existence of the solution of the following PDE
1776
+ − a∆gu = (S ◦ φ) · up−1 in M, ∂u
1777
+ ∂ν =
1778
+ 2
1779
+ p − 2 · (H ◦ φ) · u
1780
+ p
1781
+ 2 on ∂M.
1782
+ (55)
1783
+ Recall the result of prescribing scalar and mean curvature problems for conformal metrics on ( ¯
1784
+ M, g).
1785
+ Theorem 6.1. [18, Thm. 1.4] Let ( ¯
1786
+ M, g) be a compact manifold with non-empty smooth boundary
1787
+ ∂M, n = dim ¯
1788
+ M ⩾ 3. Let S, H ∈ C∞( ¯
1789
+ M) be given nonzero functions. Assume that η1 = 0. If the
1790
+ function S satisfies
1791
+ S changes sign and
1792
+ ˆ
1793
+ M
1794
+ SdVolg < 0,
1795
+ then there exists a pointwise conformal metric ˜g ∈ [g] that has scalar curvature R˜g = S and h˜g = cH
1796
+ for some small enough positive constant c.
1797
+ The conformally equivalent case follows from the result of Theorem 6.1, we show it below. Note
1798
+ that the case S = H = 0 is the trivial case.
1799
+ Theorem 6.2. Let ( ¯
1800
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
1801
+ dim ¯
1802
+ M ⩾ 3. Let S6, H6 ∈ C∞( ¯
1803
+ M) be given nonzero functions. Assume that η1 = 0. If the function
1804
+ S satisfies
1805
+ S6 changes sign,
1806
+ then there exists a diffeomorphism φ : ¯
1807
+ M → ¯
1808
+ M and a small enough constant c > 0 such that (55)
1809
+ has a smooth solution u ∈ C∞( ¯
1810
+ M) for φ, S = S6 and H = cH6. It is equivalent to say that the
1811
+ conformally equivalent metric ˜g =
1812
+
1813
+ φ−1�∗ up−2g has scalar curvature R˜g = S6 and mean curvature
1814
+ h˜g = cH6.
1815
+ Proof. Due to Theorem 6.1, it suffices to show that there exist a diffeomorphism φ : ¯
1816
+ M → ¯
1817
+ M such
1818
+ that
1819
+ ˆ
1820
+ M
1821
+ (S6 ◦ φ) dVolg < 0.
1822
+ Due to the same reason in [9, 8], it is straightforward that such a diffeomorphism does exist since
1823
+ S6 changes sign. The smallness of c is then determined by S6 ◦ φ, sup ¯
1824
+ M|H6| as well as the choice
1825
+ of sub- and super-solutions in the proofs of [18, Thm. 5.1, Cor. 5.1, Cor. 5.2].
1826
+ Note that any
1827
+ diffeomorphism φ will not change the supremum of |H6| on ¯
1828
+ M.
1829
+
1830
+ Remark 6.1. The result of Theorem 6.2 indicates that on ( ¯
1831
+ M, g) with n = dim ¯
1832
+ M ⩾ 3, any
1833
+ function that changes sign or identically zero can be realized as a scalar curvature function of some
1834
+ metric g, meanwhile the mean curvature function of g can be some small enough scaling of any
1835
+ smooth function or zero function, respectively, provided that the manifold admits a metric with
1836
+ zero first eigenvalue of the conformal Laplacian, or equivalently, zero Yamabe invariant [6, §1].
1837
+ 7. Prescribed Scalar and Mean Curvature Functions When η1 > 0
1838
+ In this section, we seek for a positive, smooth solution of the following PDE
1839
+ − a∆gu + Rgu = Sup−1 in M, ∂u
1840
+ ∂ν +
1841
+ 2
1842
+ p − 2hgu =
1843
+ 2
1844
+ p − 2Hu
1845
+ p
1846
+ 2 on ∂M.
1847
+ (56)
1848
+ on compact manifolds ( ¯
1849
+ M, g) with non-empty smooth boundary ∂M, n = dim ¯
1850
+ M ⩾ 3, for given
1851
+ functions S, H ∈ C∞( ¯
1852
+ M), provided that η1 > 0.
1853
+ As we have shown in [16], [17] and [19], we
1854
+ need to use local analysis, gluing a super-solution, and then apply monotone iteration scheme here.
1855
+ According to the “Trichotomy Theorem” in [20], we expect few restrictions on prescribed scalar
1856
+ and mean curvature functions. We will discuss the following case:
1857
+
1858
+ 22
1859
+ J. XU
1860
+ (i). S > 0 somewhere in M, and H > 0 somewhere on ∂M, with η1 > 0;
1861
+ (ii). S > 0 somewhere in M, and H ⩽ 0 everywhere on ∂M but H ̸≡ 0, with η1 > 0.
1862
+ Note that we have discussed the case S > 0 somewhere and H ≡ 0 in [19]. Currently we do not see
1863
+ how to apply our method to the case mentioned in [7],
1864
+ − ∆eu = 0 in Bn, ∂u
1865
+ ∂ν +
1866
+ 2
1867
+ p − 2hgu =
1868
+ 2
1869
+ p − 2Hu
1870
+ p
1871
+ 2 on ∂Bn, u > 0
1872
+ (57)
1873
+ for some given function H. Escobar showed that there is an obstruction for the choice of H
1874
+ ˆ
1875
+ ∂Bn X · ∇gHdS = 0.
1876
+ Here X is some conformal Killing field on ∂Bn. With standard Euclidean metric in Bn and the
1877
+ induced metric on ∂Bn, the first eigenvalue of conformal Laplacian with Robin condition is positive.
1878
+ However, since the right side is zero, we are not able to get a nontrivial local solution of the Dirichlet
1879
+ problem
1880
+ −∆eu = 0 in Ω, u = 0 on ∂Ω.
1881
+ Therefore we may need some alternative method to resolve this issue.
1882
+ However, we can get some interesting results provided that S ̸≡ 0. According to the detailed
1883
+ analysis in [19, §5], we know that there will be obstructions for the choices of prescribed scalar
1884
+ curvature functions on Sn/Γ for some Kleinian group Γ. The map Sn → Sn/Γ must be a covering
1885
+ map since otherwise Sn/Γ cannnot be a manifold. It follows that Sn/Γ has empty boundary, which
1886
+ follows that there will be no obstruction for the choice of prescribed scalar curvature functions on
1887
+ ( ¯
1888
+ M, g).
1889
+ The first result concerns the Case (i) above:
1890
+ Theorem 7.1. Let ( ¯
1891
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
1892
+ dim ¯
1893
+ M ⩾ 3. Let S7 > 0 somewhere be any smooth function on
1894
+ ¯
1895
+ M. Let H7 ∈ C∞( ¯
1896
+ M) such that
1897
+ H7 > 0 somewhere on ∂M. If η1 > 0, then there exists a small enough constant c > 0 such that
1898
+ (56) admits a positive solution u ∈ C∞( ¯
1899
+ M) with S = S7 and H = cH7. Equivalently, there exists a
1900
+ Yamabe metric ˜g = up−2g such that R˜g = S7 and h˜g = cH7
1901
+ ����
1902
+ ∂M
1903
+ .
1904
+ Proof. Without loss of generality, we may assume that Sg > 0 and hg = h > 0 with positive
1905
+ constant h, by Theorem 2.1. According to Proposition 2.3, we fix some β < 0 small enough so that
1906
+ η1,β > 0 and satisfies
1907
+ − a∆gϕ + Rgϕ = η1,βϕ in M, ∂ϕ
1908
+ ∂ν +
1909
+ 2
1910
+ p − 2hgϕ = 0 on ∂M.
1911
+ (58)
1912
+ Here ϕ > 0 on ¯
1913
+ M. Any scaling of ϕ solves (58). Denote φ = δϕ for some δ > 0. We choose δ > 0
1914
+ small enough so that
1915
+ η1,β inf
1916
+ ¯
1917
+ M ϕ ⩾ δp−2 sup
1918
+ ¯
1919
+ M
1920
+ S7 sup
1921
+ ¯
1922
+ M
1923
+ ϕp−1.
1924
+ It follows that
1925
+ −a∆gφ + Rgφ ⩾ S7φp−1 in M.
1926
+ Fix this δ. We then choose c > 0 small enough so that
1927
+ βφ ⩾ (cH7) φ
1928
+ p
1929
+ 2 on ∂M.
1930
+ It follows that
1931
+ ∂φ
1932
+ ∂ν +
1933
+ 2
1934
+ p − 2hgφ ⩾
1935
+ 2
1936
+ p − 2 (cH7) φ
1937
+ p
1938
+ 2 on ∂M.
1939
+ (59)
1940
+ Note that (59) still holds for any smaller c. For the sub-solution, we apply Proposition 2.1 or
1941
+ Proposition 2.2, depending on the vanishing of the Weyl tensor in the interior M, to construct local
1942
+
1943
+ TRICHOTOMY THEOREM: PRESCRIBED SCALAR AND MEAN CURVATURES
1944
+ 23
1945
+ solution u0 of the Yamabe equation with Dirichlet boundary condition on some domain Ω. Apply
1946
+ Lemma 3.2 in [19], we can construct a local super-solution f of the Yamabe equation in Ω such
1947
+ that f = φ near ∂Ω. We then define
1948
+ u− =
1949
+
1950
+ u0 in Ω
1951
+ 0 in M\Ω
1952
+ u+ :=
1953
+
1954
+ f in Ω
1955
+ φ in M\Ω .
1956
+ Since u− ≡ 0 on ∂M, it follows from the same argument in Lemma 3.1 in [19] that u− is a sub-
1957
+ solution of the (56) with S = S7 and H = cH7 for any constant c. According to the construction in
1958
+ Lemma 3.2 of [19], we conclude that 0 ⩽ u− ⩽ u+, u− ̸≡ 0. In addition, u− ∈ H1(M, g) ∩ C0( ¯
1959
+ M),
1960
+ and u+ ∈ C∞( ¯
1961
+ M). According to (59), we have seen that u+ is a super-solution of the (56) with
1962
+ S = S7 and H = cH7 for small enough c. Shrinking c, if necessary, so that the hypotheses of
1963
+ smallness of c sup ¯
1964
+ M|H7| holds. A direct application of Theorem 2.2 indicates the existence of a
1965
+ positive solution u ∈ C∞( ¯
1966
+ M) with S = S7 and H = cH7.
1967
+
1968
+ The proof of the Case (ii) is very similar as in Theorem 7.1.
1969
+ Theorem 7.2. Let ( ¯
1970
+ M, g) be a compact manifold with non-empty smooth boundary ∂M, n =
1971
+ dim ¯
1972
+ M ⩾ 3. Let S8 > 0 somewhere be any smooth function on
1973
+ ¯
1974
+ M. Let H8 ∈ C∞( ¯
1975
+ M) such that
1976
+ H8 ⩽ 0 everywhere on ∂M. If η1 > 0, then there exists a small enough constant c > 0 such that
1977
+ (56) admits a positive solution u ∈ C∞( ¯
1978
+ M) with S = S8 and H = cH8. Equivalently, there exists a
1979
+ Yamabe metric ˜g = up−2g such that R˜g = S8 and h˜g = cH8
1980
+ ����
1981
+ ∂M
1982
+ .
1983
+ Proof. Everything is exactly the same as in Theorem 7.1, except at (59), there is no restriction for
1984
+ the choice of the constant c. However, c should be small enough so that the hypotheses in Theorem
1985
+ 2.2 holds.
1986
+
1987
+ Remark 7.1. The result of Theorem 7.1 and Theorem 7.2 indicate that on ( ¯
1988
+ M, g) with n =
1989
+ dim ¯
1990
+ M ⩾ 3, any function that is positive somewhere can be realized as a scalar curvature function
1991
+ of some metric g, meanwhile the mean curvature function of g can be some small enough scaling
1992
+ of any smooth function, provided that the manifold admits a metric with positive first eigenvalue
1993
+ of the conformal Laplacian, or equivalently, positive Yamabe invariant [6, §1].
1994
+ References
1995
+ [1] T. Aubin. Nonlinear Analysis on Manifolds. Monge-Amp´ere Equations. Grundlehren der mathematischen Wis-
1996
+ senschaften. Springer, Berlin, Heidelberg, New York, 1982.
1997
+ [2] S. Brendle and F. Marques. Recent progress on the Yamabe problem. arXiv:1040.4960.
1998
+ [3] H. Brezis and F. Merle. Uniform esitmates and blow-up behavior for solutions of −δu = v(x)eu in two dimensions.
1999
+ Commun. Partial. Differ., 16(8-9):1223–1253, 1991.
2000
+ [4] A. Chang and P. Yang. Prescribing Gaussian curvature on S2. Acta Math., 159:215–259, 1987.
2001
+ [5] S. Cruz-Bl´azquez, A. Malchiodi, and D. Ruiz. Conformal metrics with prscribed scalar and mean curvature.
2002
+ arXiv:2105.04185.
2003
+ [6] J. Escobar. The Yamabe problem on manifolds with boundary. J. Differential Geom., 35:21–84, 1992.
2004
+ [7] J. Escobar. Conformal metrices with prescribed mean curvature on the boundary. Calc. Var. Partial Differential
2005
+ Equations, 4:559–592, 1996.
2006
+ [8] J. Kazdan and F. Warner. Curvature functions for compact 2−manifolds. Ann. of Math., 99:14–47, 1974.
2007
+ [9] J. Kazdan and F. Warner. Existence and conformal deformations of metrices with prescribed Gaussian and scalar
2008
+ curvatures. Ann. of Math. (2), 101(2):317–331, 1975.
2009
+ [10] J. Kazdan and F. Warner. Scalar curvature and conformal deformation of Riemannian structure. J. Differential
2010
+ Geom., 10:113–134, 1975.
2011
+
2012
+ 24
2013
+ J. XU
2014
+ [11] A. Malchiodi and M. Mayer. Prescribing Morse scalar curvatures: Pinching and Morse theory. Commun. Pure
2015
+ Appl. Math, 2021.
2016
+ [12] S. Rosenberg and J. Xu. Solving the Yamabe problem by an iterative method on a small Riemannian domain.
2017
+ arXiv:2110.14543.
2018
+ [13] M. Struwe. A flow approach to Nirenberg’s problem. Duke Math. J., 128(19-64), 2005.
2019
+ [14] M. Taylor. Partial Differential Equations I. Springer-Verlag, New York, New York, 2011.
2020
+ [15] M. Taylor. Partial Differential Equations III. Springer-Verlag, New York, New York, 2011.
2021
+ [16] J. Xu. The boundary Yamabe problem, I: Minimal boundaray case. arXiv:2111:03219.
2022
+ [17] J. Xu. The boundary Yamabe problem, II: General constant mean curvature case. arXiv:2112.05674.
2023
+ [18] J. Xu. The conformal Laplacian and the Kazdan-Warner problem: Zero first eigenvalue case. arXiv:2211.15024.
2024
+ [19] J. Xu. Prescribed scalar curvature on compact manifolds under conformal deformation. arXiv:2205.15453.
2025
+ [20] J. Xu. Prescribed scalar curvature problem under conformal deformation of a Riemannian metric with Dirichlet
2026
+ boundary condition. arXiv:2208.11318.
2027
+ [21] J. Xu. Solving the Yamabe-type equations on closed manifolds by iteration schemes. arXiv: 2110.15436.
2028
+ Department of Mathematics and Statistics, Boston University, Boston, MA, U.S.A.
2029
+ Email address: xujie@bu.edu
2030
+ Institute for Theoretical Sciences, Westlake University, Hangzhou, Zhejiang Province, China
2031
+ Email address: xujie67@westlake.edu.cn
2032
+
-9AzT4oBgHgl3EQfFvoN/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
.gitattributes CHANGED
@@ -2475,3 +2475,56 @@ KdE3T4oBgHgl3EQfAQmY/content/2301.04256v1.pdf filter=lfs diff=lfs merge=lfs -tex
2475
  MtE3T4oBgHgl3EQfYwqL/content/2301.04491v1.pdf filter=lfs diff=lfs merge=lfs -text
2476
  NtFLT4oBgHgl3EQfOS8u/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2477
  8tE4T4oBgHgl3EQf3A1I/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2475
  MtE3T4oBgHgl3EQfYwqL/content/2301.04491v1.pdf filter=lfs diff=lfs merge=lfs -text
2476
  NtFLT4oBgHgl3EQfOS8u/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2477
  8tE4T4oBgHgl3EQf3A1I/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2478
+ SNAzT4oBgHgl3EQf0f5U/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2479
+ UtAyT4oBgHgl3EQfhfhI/content/2301.00377v1.pdf filter=lfs diff=lfs merge=lfs -text
2480
+ AtAyT4oBgHgl3EQfq_mZ/content/2301.00553v1.pdf filter=lfs diff=lfs merge=lfs -text
2481
+ wtE4T4oBgHgl3EQfXgy3/content/2301.05042v1.pdf filter=lfs diff=lfs merge=lfs -text
2482
+ odAzT4oBgHgl3EQf5P7V/content/2301.01858v1.pdf filter=lfs diff=lfs merge=lfs -text
2483
+ e9E5T4oBgHgl3EQfhA8F/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2484
+ BNE1T4oBgHgl3EQf9Qa6/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2485
+ odAzT4oBgHgl3EQf5P7V/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2486
+ HtE2T4oBgHgl3EQfTwfs/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2487
+ ptAzT4oBgHgl3EQfq_1g/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2488
+ L9E4T4oBgHgl3EQfiw1J/content/2301.05136v1.pdf filter=lfs diff=lfs merge=lfs -text
2489
+ 6NAyT4oBgHgl3EQfcffh/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2490
+ BNE1T4oBgHgl3EQf9Qa6/content/2301.03555v1.pdf filter=lfs diff=lfs merge=lfs -text
2491
+ y9E3T4oBgHgl3EQfmQpK/content/2301.04614v1.pdf filter=lfs diff=lfs merge=lfs -text
2492
+ jNE1T4oBgHgl3EQf0AVB/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2493
+ LNAyT4oBgHgl3EQfTve8/content/2301.00113v1.pdf filter=lfs diff=lfs merge=lfs -text
2494
+ wtE4T4oBgHgl3EQfXgy3/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2495
+ LNAyT4oBgHgl3EQfTve8/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2496
+ L9E4T4oBgHgl3EQfiw1J/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2497
+ hdE3T4oBgHgl3EQfIQm9/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2498
+ MtE3T4oBgHgl3EQfYwqL/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2499
+ d9FKT4oBgHgl3EQfrC7t/content/2301.11877v1.pdf filter=lfs diff=lfs merge=lfs -text
2500
+ UtAyT4oBgHgl3EQfhfhI/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2501
+ jNE1T4oBgHgl3EQf0AVB/content/2301.03450v1.pdf filter=lfs diff=lfs merge=lfs -text
2502
+ iNE1T4oBgHgl3EQffgTu/content/2301.03220v1.pdf filter=lfs diff=lfs merge=lfs -text
2503
+ pNFAT4oBgHgl3EQfeR0E/content/2301.08574v1.pdf filter=lfs diff=lfs merge=lfs -text
2504
+ VdFJT4oBgHgl3EQfOSwV/content/2301.11481v1.pdf filter=lfs diff=lfs merge=lfs -text
2505
+ KdE3T4oBgHgl3EQfAQmY/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2506
+ HdAyT4oBgHgl3EQf5frD/content/2301.00807v1.pdf filter=lfs diff=lfs merge=lfs -text
2507
+ 2tAyT4oBgHgl3EQf1vk5/content/2301.00739v1.pdf filter=lfs diff=lfs merge=lfs -text
2508
+ d9FKT4oBgHgl3EQfrC7t/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2509
+ AtE1T4oBgHgl3EQfDQOp/content/2301.02875v1.pdf filter=lfs diff=lfs merge=lfs -text
2510
+ X9E4T4oBgHgl3EQfNwwL/content/2301.04958v1.pdf filter=lfs diff=lfs merge=lfs -text
2511
+ 6dE0T4oBgHgl3EQfvwHw/content/2301.02624v1.pdf filter=lfs diff=lfs merge=lfs -text
2512
+ ttFAT4oBgHgl3EQfhR0m/content/2301.08592v1.pdf filter=lfs diff=lfs merge=lfs -text
2513
+ AtE1T4oBgHgl3EQfDQOp/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2514
+ AtAyT4oBgHgl3EQfq_mZ/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2515
+ C9AzT4oBgHgl3EQfTvzN/content/2301.01256v1.pdf filter=lfs diff=lfs merge=lfs -text
2516
+ gNFLT4oBgHgl3EQfYy_o/content/2301.12067v1.pdf filter=lfs diff=lfs merge=lfs -text
2517
+ HdAyT4oBgHgl3EQf5frD/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2518
+ 2tAyT4oBgHgl3EQf1vk5/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2519
+ VdFJT4oBgHgl3EQfOSwV/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2520
+ ttFAT4oBgHgl3EQfhR0m/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2521
+ RdE2T4oBgHgl3EQfBwYp/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2522
+ 1NE0T4oBgHgl3EQfdgAR/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2523
+ X9E4T4oBgHgl3EQfNwwL/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2524
+ 79AzT4oBgHgl3EQfSPsz/content/2301.01228v1.pdf filter=lfs diff=lfs merge=lfs -text
2525
+ y9E3T4oBgHgl3EQfmQpK/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2526
+ 6dE0T4oBgHgl3EQfvwHw/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2527
+ 79AzT4oBgHgl3EQfSPsz/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2528
+ atFQT4oBgHgl3EQfgDa2/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2529
+ FtE3T4oBgHgl3EQftQtg/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
2530
+ Z9FJT4oBgHgl3EQf7y3s/content/2301.11680v1.pdf filter=lfs diff=lfs merge=lfs -text
1NE0T4oBgHgl3EQfdgAR/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a00858eade89b4a35048fd7b2641f0064da2fba9f21ecd588200eb175f778743
3
+ size 6291501
1NE0T4oBgHgl3EQfdgAR/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:30474d8c7312d16b19e52e5049d065718d968a8e1dc7954ff6291870bc8c8469
3
+ size 206302
1dE0T4oBgHgl3EQf_wJK/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ed7a4ac9cc0581af336bdd48149c42fbd341e58fe6106bb6ae9563cafb87b2ca
3
+ size 373299
2tAyT4oBgHgl3EQf1vk5/content/2301.00739v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a6beb570641b29d9e19348385ccb90e6b9ca10873e7df556c0fe32f5657cf4ff
3
+ size 234688
2tAyT4oBgHgl3EQf1vk5/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:405b578ad54607bd1846680b2e551a3afce3d9140d60fd32b13e6753408551f2
3
+ size 983085
2tAyT4oBgHgl3EQf1vk5/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1084dc181450401a60e63f63d0862cee85a54b7973e67f4477505d38c3be184e
3
+ size 39772
3tE4T4oBgHgl3EQf0g0b/content/tmp_files/2301.05282v1.pdf.txt ADDED
@@ -0,0 +1,1481 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1
2
+
3
+ Non-centrosymmetric Sr2IrO4 obtained under High Pressure
4
+
5
+ Haozhe Wang1‡, Madalynn Marshall2‡, Zhen Wang3, Kemp W. Plumb4, Martha Greenblatt2,
6
+ Yimei Zhu3, David Walker5, Weiwei Xie1*
7
+
8
+
9
+ 1. Department of Chemistry, Michigan State University, East Lansing, Michigan 48824, USA
10
+ 2. Department of Chemistry and Chemical Biology, Rutgers University, Piscataway, New Jersey
11
+ 08854, USA
12
+ 3. Condensed Matter Physics and Materials Science Department, Brookhaven National
13
+ Laboratory, Upton, New York 11973, USA
14
+ 4. Department of Physics, Brown University, Providence, Rhode Island 02912, USA
15
+ 5. Lamont Doherty Earth Observatory, Columbia University, Palisades, New York 10964, USA
16
+
17
+
18
+ ‡ H.W. and M.M. contributed equally. * Email: xieweiwe@msu.edu
19
+
20
+
21
+
22
+ Abstract
23
+
24
+ Sr2IrO4 with strong spin-orbit coupling (SOC) and Hubbard repulsion (U) hosts Mott
25
+ insulating states. The similar crystal structure, magnetic and electronic properties, particularly the
26
+ d-wave gap observed in Sr2IrO4 enhanced the analogies to cuprate high-Tc superconductor,
27
+ La2CuO4. The incomplete analogy was due to the lack of broken inversion symmetry phases
28
+ observed in Sr2IrO4. Here, under high pressure and high temperature conditions, we report a non-
29
+ centrosymmetric Sr2IrO4. The crystal structure and its noncentrosymmetric character were
30
+ determined by single crystal X-ray diffraction and high-resolution scanning transmission electron
31
+ microscopy (HR-STEM). The magnetic characterization confirms the Ir4+ with S = 1/2 at low
32
+ temperature in Sr2IrO4 with magnetic ordering occurred at around 86 K, where a larger moment is
33
+ observed than the ambient pressure Sr2IrO4. Moreover, the resistivity measurement shows three-
34
+ dimensional Mott variable-range hopping existed in the system. This non-centrosymmetric Sr2IrO4
35
+ phase appears to be a unique material to offer further understanding of high-Tc superconductivity.
36
+
37
+
38
+
39
+ 2
40
+
41
+ Introduction
42
+
43
+ Iridates with strong spin-orbit coupling effects can generate exotic quantum phenomena,
44
+ such as quantum spin liquid phases, Kitaev magnetism, and possible superconductivity.1-4
45
+ Different from most 3d transition metal oxides in which the spin and orbit can be distinct in the
46
+ energy scale, the spin and orbit interact heavily in 5d transition metal oxides. Among the Mott
47
+ insulating 5d transition metal oxides, Sr2IrO45-9 has attracted significant of attention due to its
48
+ similarity to cuprate high-temperature superconductor, La2CuO410-12. As a single-layer
49
+ Ruddlesden–Popper compound, Sr2IrO4 crystallizes in a tetragonal lattice with an inversion center
50
+ (I41/acd, #142) at ambient pressure. Sr2IrO4 contains stacked IrO2 square lattices where the unit
51
+ cell is doubled compared to the CuO2 square lattices in high-Tc cuprates as a result of a staggered
52
+ rotation of IrO6 octahedron. Although superconductivity is not yet confirmed, many phenomena
53
+ characteristic of the superconducting cuprates have been observed in electron-and hole-doped
54
+ iridates including pseudogaps, Fermi arcs, and d-wave gaps .13-15 The Ir-d5 electrons in regular
55
+ IrO6 octahedron occupy the t2g orbitals, which can be approximated as two fully filled spin-orbital
56
+ coupled Jeff = 3/2 bands and one half-filled Jeff = 1/2 band. The Jeff band is split into an upper and
57
+ lower Hubbard band by on-site Coulomb interaction. According to a previous study, as Sr2IrO4 is
58
+ cooled below its Néel temperature (TN, ~230 K), the spin-orbit coupled Jeff = 1/2 moments order
59
+ into a basal plane commensurate Néel state. Octahedral rotations in Sr2IrO4 allow for non-zero
60
+ Dzyaloshinskii-Moriya (DM) interactions that results in a canting of the ordered moments away
61
+ from the crystallographic axis and a weak ferromagnetic moment per layer.16 Such a magnetic
62
+ transition maintains the inversion symmetry but lowers the rotational symmetry of the system from
63
+ C4 to C2. However, no additional symmetry breaking has been observed by neutron or X-ray
64
+ diffraction, which makes the comparison of the iridate to cuprate phenomenology incomplete. To
65
+ date, multiple methods have been used to tune the Mott insulating states in Sr2IrO4, for example,
66
+ isovalent Rh doping on the Ir site.5,17-23 After partially substituting Ir with Rh, an insulator-to-metal
67
+ transition can be detected. However, high pressure was also used to tune the electronic states up
68
+ to 55 GPa without observing any metallic state in Sr2IrO4.24,25
69
+
70
+ In this report, we applied the high-pressure (6 GPa) high-temperature (1400 °C) method
71
+ for synthesizing Sr2IrO4. Under such extreme conditions, the obtained Sr2IrO4 remains in a
72
+ tetragonal structure but without an inversion center. The space group was determined by single-
73
+
74
+ 3
75
+
76
+ crystal X-ray diffraction (SC-XRD) as I4mm (#107). Unlike the ambient pressure phase, the high-
77
+ pressure phase consists of the single layered IrO2 square lattice, just like CuO2 square in cuprate.
78
+ Magnetic susceptibility measurement on high pressure Sr2IrO4 indicate a magnetic ordering
79
+ temperature of approximately 86 K, which is dramatically lower than ambient pressure Sr2IrO4.
80
+ Interestingly, the resistivity data shows three-dimensional Mott variable-range hopping of charge
81
+ carriers between states localized by disorder with negligible long-range Coulomb interactions.
82
+ Discovering the non-centrosymmetric phase in Sr2IrO4 may accelerate the realization of
83
+ superconductivity and unravel the puzzle in cuprate high-Tc superconductors.
84
+
85
+
86
+
87
+
88
+ 4
89
+
90
+ Experimental Section
91
+
92
+ High-Pressure Synthesis. The ambient pressure Sr2IrO4 phase was prepared accordingly by
93
+ thoroughly mixing and pelletizing the materials SrCO3 and IrO2 and subsequently heating them to
94
+ 900 °C then regrinding and reannealing at 1000 °C and subsequently reannealing at 1100 °C.26
95
+ The ambient pressure Sr2IrO4 was pressurized to 6 GPa in 24 hours. After that, the sample was
96
+ heated up to 1400 °C and stayed at 1400 °C for 4 hours. Another sample was heated to 1400 °C
97
+ and stayed up to 28 hours to explore the optimal condition. The sample was cooled down to room
98
+ temperature before depressurizing to the ambient pressure. The high-pressure synthesis was
99
+ performed by statically compressing the sample using the Walker type multi-anvil press27 where
100
+ the original Sr2IrO4 was placed in a Pt capsule inside an Al2O3 crucible that was inserted into a
101
+ Cermacast 646 octahedra pressure medium lined on the inside with a LaCrO3 heater.
102
+
103
+ Phase Analysis and Chemical Composition Determinations. The phase identity and purity were
104
+ examined using a Bruker D2 Phaser powder X-ray diffractometer with Cu K������������ radiation (������������ =
105
+ 1.5406 Å). Room temperature measurements were performed with a step size of 0.004° at a scan
106
+ speed of 0.55°/min over a Bragg angle (2������������) range of 5–90°. FullProf Suite software28,29 was
107
+ utilized to analyze the phase information and lattice parameters from a Rietveld refinement.
108
+
109
+ Structure Determination. The room temperature and low temperature (100 K) crystal structure
110
+ was determined using a Bruker D8 Quest Eco single crystal X-ray diffractometer, equipped with
111
+ Mo radiation (������������������������������������ = 0.71073 Å) with an ������������ of 2.0° per scan and an exposure time of 10 s per frame.
112
+ A SHELXTL package with the direct methods and full-matrix least-squares on the F2 model was
113
+ used to determine the crystal structure of Sr2IrO4.30,31 To confirm the crystal structure, high-
114
+ resolution scanning transmission electron microscopy (HR-STEM) images were collected and
115
+ electron diffraction was conducted using a 200 kV JEOL ARM electron microscope equipped with
116
+ double aberration correctors. Samples for TEM analysis were crushed in an agate mortar and
117
+ deposited directly onto a holey carbon copper grid.
118
+
119
+ Physical Properties Measurement. Temperature and field-dependent magnetization, resistivity,
120
+ and heat capacity measurements were performed with a Quantum Design physical property
121
+ measurement system (PPMS) under a temperature range of 1.85–300 K and applied fields up to 9
122
+
123
+ 5
124
+
125
+ T. Electrical resistivity measurements were accomplished with a four-probe method using
126
+ platinum wires on a pelletized sample of Sr2IrO4. The polycrystalline Sr2IrO4 was pressed up to 6
127
+ GPa and heated at a lower temperature (100 °C) to eliminate the contribution of grain boundary
128
+ effect but also keep the phase stable.
129
+
130
+
131
+ 6
132
+
133
+ Results and Discussions
134
+
135
+ Exploring New Phase. The new Sr2IrO4 phase (I4mm, #107) was formed at 6 GPa from the
136
+ starting material, ambient pressure Sr2IrO4 (I41/acd, #142). The synthesis temperatures were set
137
+ up at 1200 °C and 1400 °C. The high pressure Sr2IrO4 phase was only produced at 1400 °C. To
138
+ increase the yield and grow larger crystals, the longer heating duration of 28 hours was tested.
139
+ However, the secondary tetragonal phase Sr3Ir2O7 simultaneously forms once the heating duration
140
+ was increased. As a result, only 4 hours heating process can produce the specimen consisting
141
+ mostly of pure phase. The resulting Le Bail fitting of the PXRD patterns for the high-pressure
142
+ phase Sr2IrO4 is shown in Fig. 1. An overlay of the PXRD patterns in Fig. S1 demonstrates the
143
+ formation of the secondary Sr3Ir2O7 phase. The pure phase synthesized at 1400 °C for 4 hours was
144
+ used for the physical property measurements below.
145
+
146
+
147
+ Fig. 1 Powder X-ray diffraction pattern of the high-pressure Sr2IrO4 phase. The experimental
148
+ data (red dots) was modeled with a Rietveld refinement (black line). The blue line indicates the
149
+ corresponding residual pattern (difference between observed and calculated patterns) along with
150
+ Bragg peak positions for Sr2IrO4 (green) and Al2O3 (purple) represented by the vertical tick marks.
151
+
152
+
153
+
154
+
155
+ Calc
156
+ Diff
157
+ Obs
158
+ Intensity (a.u.)
159
+ Sr2lrO4
160
+ Al203
161
+ 10
162
+ 30
163
+ 50
164
+ 70
165
+ 90
166
+ Sr,IrO.
167
+ 20 (degree)
168
+ 14mm (#107)7
169
+
170
+ Crystal Structure and Phase Determination. After 4 hours of treatment at 6 GPa and 1400 °C,
171
+ single crystals of Sr2IrO4 were formed, subsequently selected, and measured at both 300 K and 100
172
+ K using the single crystal X-ray diffractometer. High-pressure Sr2IrO4 crystallizes with good
173
+ agreement into the tetragonal space group I4mm, as indicated by the single crystal X-ray diffraction
174
+ (SCXRD) refinement information listed in Table S1. Similar to ambient pressure Sr2IrO4, the high
175
+ pressure Sr2IrO4 phase contains the layers of IrO6 octahedra with intercalated Sr atoms. The
176
+ differences between these two are half-c lattice, the disappearance of the inversion center because
177
+ of the nonsymmetric distortion of IrO6 octahedra, and the disappearance of IrO6 octahedral
178
+ rotations in the ab-plane in high-pressure Sr2IrO4 compared to the ambient pressure phase. Shown
179
+ in Fig. 2 are crystal structures and IrO6 octahedra stacking view of ambient pressure Sr2IrO4
180
+ (I41/acd), high-pressure Sr2IrO4 (I4mm), and previously reported La2CuO4 (I4/mmm), with Ir-O
181
+ atomic distance in the IrO6 octahedra highlighted. Atomic site vacancies and site disorder were
182
+ considered and refined to reveal the O3 atomic site is slightly displaced from the closer ideal 4b
183
+ site (1/2, 0, z) to the 8d site (x, 0, z) having a statistical occupancy of 0.5. The disordered model
184
+ yielded a more reasonable refinement with an R factor of 4.35 and goodness of fit (GOF) of 1.177
185
+ while having only one O3 atomic site resulted in an R factor of 4.62 and GOF of 1.305. As such
186
+ an angle ������������ can be determined from (1/2 ± ������������, 0, z) with respect to an IrO6 octahedra where the O3
187
+ atoms occupy the 4b site. This structural disorder has been thoroughly discussed for the ambient
188
+ pressure Sr2IrO4 structure.32 Additionally, the high-pressure Sr2IrO4 phase possesses a
189
+ nonsymmetric IrO6 octahedra elongation along the c axis, ranging in Ir-O atomic distance from
190
+ 1.94(6)–2.27(6) Å, as indicated in Fig. 2b, which is in fact the cause of noncentrosymmetric
191
+ structural character. This behavior is kind of similar to the prominent feature of ambient pressure
192
+ Sr2IrO4 that has been speculated to originate from a Jahn Teller distortion.33-35 Previous studies
193
+ under high-pressure have revealed an increase in the IrO6 octahedra elongation with
194
+ pressurization.36,37 Compared to ambient pressure Sr2IrO4, one Ir-O along the c -axis is
195
+ significantly elongated, with the other almost remains the same, i.e., one oxygen atom is driven
196
+ away from the Ir atom, and thus the repulsion between Ir and the oxygen ligand is reduced. This
197
+ will lower the energy of orbitals that contains z contribution and split eg and t2g orbitals, making
198
+ the crystal field split of Ir d orbitals even more complicated. Together with spin-orbit coupling,
199
+ this may further remove orbital degeneracies. Moreover, as pressure applied for Sr2IrO4, the Ir-O-
200
+
201
+ 8
202
+
203
+ Ir angle was pushed close to 180°, which is the angle in Cu-O-Cu in La2CuO4. The structural
204
+ disorder was further confirmed at 100 K and the SCXRD refinement details can be found in SI.
205
+
206
+ Fig. 2 Crystal structure illustration. Crystal structures, octahedra stacking view along a axis,
207
+ and along c axis of (a) ambient pressure Sr2IrO4, (b) as-synthesized high pressure Sr2IrO4, and (c)
208
+ previously reported La2CuO4, with Ir(Cu)O6 octahedra and Ir(Cu)-O atomic distances presented.
209
+ Green, blue, dark green, dark blue, and red atoms represent Sr, Ir, La, Cu, and O atoms,
210
+ respectively. Single-layer square net is also highlighted.
211
+
212
+
213
+
214
+ (a)
215
+ (b)
216
+ 1.98(1) A
217
+ 2.27(6) A
218
+ 1.95(1) A
219
+ 1.93(1) A
220
+ 1.94(1) A
221
+ 1.94(6) A
222
+ 2.11(1) A
223
+ Sr,IrO4
224
+ Sr,lrO4
225
+ La,CuO
226
+ 14,/acd (#142)
227
+ 14mm (#107)
228
+ 14/mmm(#139)9
229
+
230
+ Transmission Electron Microscopy. The non-centrosymmetric space group and loss of IrO6
231
+ octahedral rotation, as well as the oxygen distortion and defects in Sr2IrO4, can at first, be
232
+ surprisingly interesting, thus high-pressure Sr2IrO4 was investigated by transmission electron
233
+ microscopy (TEM) to characterize its crystallographic nature. The High-angle annular dark-field
234
+ scanning transmission electron microscopy (HAADF-STEM) image was obtained along the a axis
235
+ shown in Fig. 3a. The TEM diffraction patterns projected down the crystalline [100] axis (Fig. 3b)
236
+ allowed for the determination of the orientation of the images through the d002 spacing. The c-axis
237
+ parameter is ~12.8 Å, agreeing with the single crystal XRD results. The electron diffraction and
238
+ imaging study confirmed the high quality of the nanoscale ordering in the specimen. However, the
239
+ fractional spots 1/2 (110)/(1-10) were observed by TEM electron diffraction in Fig. 3d. As is
240
+ known that IrO6 tilt/rotation along the c-axis would not introduce these fractional spots. Such
241
+ fractional reflection spots are related to the ordering of oxygen vacancy, which is consistent with
242
+ single crystal X-ray diffraction results in Fig. 3e.
243
+
244
+
245
+ Fig. 3 Transmission electron microscopy study of high pressure Sr2IrO4 phase. (a) HAADF-
246
+ STEM image taken along a axis from a large area showing the high quality of the crystal Sr2IrO4.
247
+ (b) The zoom-in HAADF image shows the projected structure in the [100] direction, with a crystal
248
+ model superimposed, where Sr (green), Ir (blue), and O (orange). (c) The diffraction pattern took
249
+ along the [100] direction which is consistent with the simulated pattern (Fig. 3e) based on the
250
+ crystal model determined by single crystal X-ray diffraction. (d) SAED pattern along the [001]
251
+ direction showing fractional spots of 1/2 (110)/(1-10). (e) Simulated diffraction pattern and (f)
252
+ projected crystal structure along the [001] direction based on the crystal structure determined by
253
+
254
+ 220
255
+ 020
256
+ X
257
+ 200
258
+ 000
259
+ 003
260
+ 220
261
+ 020
262
+ 20
263
+ [100]
264
+ 110
265
+ 200
266
+ 10
267
+ [001]10
268
+
269
+ SCXRD. The fractional spots observed in TEM were marked in red. The single crystal structure
270
+ of Sr2IrO4 with oxygen distortion was confirmed by both single crystal X-ray diffraction and TEM.
271
+
272
+
273
+ Weak Ferromagnetic Ordering. To study the magnetic properties of the high pressure Sr2IrO4
274
+ phase, the temperature-dependent susceptibility was measured under field cooled warming (FCW)
275
+ and field cooled cooling (FCC) mode at 0.1 T shown in Fig. 4a. No significant differences between
276
+ FCW and FCC were observed. At about 150 K, the susceptibility goes below 0, indicating a
277
+ diamagnetic contribution in the system, which suggests the possible breakdown of Curie-Weiss
278
+ behavior at high temperatures in the system. The data between 80–140 K was modeled with the
279
+ modified Curie-Weiss law (Eqn. 1), shown in Fig. 4b and Fig. S2b,
280
+ ������������ = ������������0 +
281
+ ������������
282
+ ������������ − ������������cw
283
+ (1)
284
+ where ������������������������������������ is the paramagnetic Curie temperature, ������������0 is the temperature independent susceptibility
285
+ and ������������ is the Curie constant. From the fitting, the Curie temperature, ������������������������������������, of 86(7) K was found to
286
+ be comparable to the magnetic ordering temperature ������������������������ ~84 K, as determined from the minimum
287
+ in the temperature derivative of ������������ (See Fig. S2a for details). The magnetic ordering temperature,
288
+ consequently, decreases when compared to ambient pressure Sr2IrO4, which has a ������������������������ ~240 K.39,40
289
+ On the other hand, it can be assumed that the Tc significantly decreases as the angle of Ir-O-Ir is
290
+ more close to 180 °, which is the one observed in Cu-O-Cu in high Tc superconductor La2CuO4.
291
+ The fitting also gave a negative ������������0 of -2.9(9)×10-3 emu mol-1 Oe-1, which provided a potential
292
+ opportunity to extrapolate our Curie-Weiss fit to higher temperature. Finally, up to 160 K was
293
+ included (Fig. S2c) and the fit yielded the effective moment ������������eff = 1.2(2) µB/Ir, which is more
294
+ agreeable with the Hund’s-rule value of 1.73 µB/Ir for S = 1/2 than the reported ������������eff = 0.33 µB/Ir
295
+ for ambient pressure Sr2IrO4.
296
+
297
+ Furthermore, the magnetization of high pressure Sr2IrO4 was measured as shown in Fig.
298
+ 4c up to 9 T at different temperatures. It appears to saturate at ~3 T at which the magnetic saturation
299
+ moment (������������������������������������������������) was determined to be ~0.046 µB/Ir. This value is significantly lower than the
300
+ theoretical value of 1/3 µB f.u−1, however, similar to the previously reported moment for the
301
+ ambient pressure Sr2IrO4 phase, which originates from spin canted antiferromagnetic (AFM)
302
+ order.39 This could also explain why the weak ferromagnetic behavior observed in the temperature
303
+
304
+ 11
305
+
306
+ dependence of magnetic susceptibility gives such a low value of moment. However, unlike the
307
+ ambient pressure Sr2IrO4 phase, the magnetization reaches a maximum at around 3 T at which
308
+ point the magnetization decreases. It turned out that diamagnetic transition was observed under
309
+ higher fields at the respective temperatures (e.g., see the 50 K and 100 K data). At 300 K, a
310
+ complete diamagnetic behavior was shown, consistent with ������������ < 0 shown in Fig. 4c. Subtracting
311
+ this by linearly fitting data from 7–9 T, the ������������������������������������������������ was modified to be 0.067 µB/Ir at 2 K and 0.014
312
+ µB/Ir at 100 K, as presented in Fig. 4d and 4e. Magnetic hysteresis was observed in the system
313
+ under 2 K from -0.6 T to 0.6 T, presented in Fig. S3, which could be interpreted as small canting
314
+ of the moments existed in the system.
315
+
316
+ 12
317
+
318
+
319
+ Fig. 4 Magnetization in the dependence of temperature and field. (a) Temperature dependence
320
+ of magnetic susceptibility ������������ at 1000 Oe under FCW and FCC mode ranging from 2–300 K. No
321
+ significant difference was observed. (b) The modified inverse magnetic susceptibility data (FCW,
322
+ 80–140 K, blue hollow circle) fitted with the modified Curie-Weiss model (orange line). (c) Field
323
+ dependence of magnetization up to 9 T at different temperatures. (d) Derivation of ������������sat at 2 K by
324
+ linearly fitting the magnetization data from 7–9 T. (e) Derivation of ������������sat at 100 K.
325
+
326
+ No Magnetically Induced Anomalies Observed in Specific Heat Measurement. To confirm the
327
+ magnetic transition, the specific heat over the temperature range of 2–200 K was measured under
328
+ 0 T with a polycrystalline pelletized sample of Sr2IrO4, as presented in Fig. 5a. Measurements
329
+
330
+ (a)
331
+ (b)
332
+ 1e-2
333
+ 1e3
334
+ 2
335
+ FCW
336
+ Cw fit
337
+ Oe)
338
+ FCC
339
+ 0
340
+ FCW
341
+ mol
342
+ 0
343
+ 8
344
+ oo
345
+ 0
346
+ 0
347
+ 0
348
+ 1
349
+ 090
350
+ 0
351
+ 4
352
+ 8
353
+ 08
354
+ X
355
+ 8
356
+ 0
357
+ 0
358
+ 0
359
+ 100
360
+ 200
361
+ 300
362
+ 80
363
+ 130
364
+ 180
365
+ 230
366
+ 280
367
+ T (K)
368
+ T (K)
369
+ (c)
370
+ (d)
371
+ (e)
372
+ 1e-2
373
+ 8
374
+ 0
375
+ 10
376
+ 20 K
377
+ (μB per Ir ion)
378
+ 4
379
+ 0
380
+ .4
381
+ M
382
+ 50 K
383
+ -8
384
+ 100 K
385
+ 0 2K
386
+ 100K
387
+ 300 K
388
+ "2 K"
389
+ "100 K"
390
+ -9
391
+ -6-3
392
+ 3
393
+ 6
394
+ 9
395
+ 0
396
+ 36
397
+ 9
398
+ 0
399
+ 36
400
+ 9
401
+ μoH (T)
402
+ μoH (T)
403
+ μoH (T)13
404
+
405
+ under applied fields of 0.05 T and 1 T in Fig. S3 were additionally tested to conclude no significant
406
+ deviation from the 0 T specific heat. No ������������ shape anomalies were observed at the whole temperature
407
+ regime studied, which may result from higher temperature regions being heavily dominated by the
408
+ phonon contribution. The specific heat data were fitted by the Debye model (Eqn. 2), and Einstein
409
+ model (Eqn. 3), shown in Fig. S4a and b. The Debye and Einstein temperatures could then be
410
+ determined as 417(2) K and 306(2) K, respectively. However, neither of these two described the
411
+ experimental data well.
412
+ ������������D = 9������������������������ � ������������
413
+ ������������D
414
+
415
+ 3
416
+
417
+ ������������4������������������������
418
+ (������������������������ − 1)2 ������������������������
419
+ ������������D ������������
420
+
421
+ 0
422
+ (2)
423
+ where ������������ is the number of atoms per formula unit, ������������ is the gas constant, and ������������������������ is the Debye
424
+ temperature.
425
+ ������������E = 3������������������������ �������������E
426
+ ������������ �
427
+ 2
428
+ ������������
429
+ ������������E
430
+ ������������ �������������
431
+ ������������E
432
+ ������������ − 1�
433
+ −2
434
+ (3)
435
+ where ������������ is the number of atoms per formula unit, ������������ is the gas constant, and ������������������������ is the Einstein
436
+ temperature.
437
+
438
+ The specific heat data was further fitted with two Debye model (Eqn. 4) and weighted
439
+ Debye model (Eqn. 5), with and without the electronic contribution included, shown in Fig. 5a
440
+ and Fig. S4c, d, and e. The data was found to be described well with two Debye model (Fig. 5a),
441
+ and the Debye temperatures, ������������������������1 of 235(1) K, ������������������������2 of 708(5) K was obtained. At low temperatures,
442
+ the first Debye mode has a larger contribution to the specific heat. Within the temperature regime
443
+ studied, the expected Dulong-Petit value of 3������������������������ is not recovered, and this can be explained by the
444
+ high value of ������������������������2, which means that the specific heat will plateau at ������������ ≫ ������������������������2. The fitting also yields
445
+ ������������������������1 of 3.20(3) and ������������������������2 of 4.51(2). The sum of these two seems a little larger than the expected
446
+ value of 7 for Sr2IrO4, which may be attributed to the impurity of Srn+1IrnO3n+1, lack of electron
447
+ contribution, or overestimation of photon contribution in the model. Once the electron contribution
448
+ term was included, ������������������������1 was slightly shifted to 238(2) K and the sum of ������������������������1 and ������������������������2 went down to
449
+ 7.52(11).
450
+ ������������ = 9������������D1������������ � ������������
451
+ ������������D1
452
+
453
+ 3
454
+
455
+ ������������4������������������������
456
+ (������������������������ − 1)2 ������������������������
457
+ ������������D1 ������������
458
+
459
+ 0
460
+ + 9������������D2������������ � ������������
461
+ ������������D2
462
+
463
+ 3
464
+
465
+ ������������4������������������������
466
+ (������������������������ − 1)2 ������������������������
467
+ ������������D2 ������������
468
+
469
+ 0
470
+ (+������������������������)
471
+ (4)
472
+
473
+ 14
474
+
475
+ where ������������������������1 and ������������������������2 are Debye temperatures, ������������������������1 and ������������������������2 are the oscillator strengths, and ������������������������ is the
476
+ electron contribution.
477
+ ������������ = 9������������D������������ � ������������
478
+ ������������D
479
+
480
+ 3
481
+
482
+ ������������4������������������������
483
+ (������������������������ − 1)2 ������������������������
484
+ ������������D ������������
485
+
486
+ 0
487
+ + 3������������E������������ �������������E
488
+ ������������ �
489
+ 2
490
+ ������������
491
+ ������������E
492
+ ������������ �������������
493
+ ������������E
494
+ ������������ − 1�
495
+ −2
496
+ (+������������������������)
497
+ (5)
498
+ where ������������������������ and ������������������������ are the Debye and Einstein temperatures, ������������������������ and ������������������������ are the oscillator strengths.
499
+
500
+ It should be noted that the magnetic contribution cannot be quantitatively extracted from
501
+ the specific heat data as the phonon contribution cannot be distinguished from the magnetic
502
+ contribution due to the lack of a nonmagnetic analog.
503
+
504
+ At a low-temperature regime, of 2–20 K, the specific heat was measured, as shown in Fig.
505
+ S5. The data ranging from 2–3.2 K was fitted with Eqn. 6, shown in Fig. 5b.
506
+ ������������p
507
+ ������������ = ������������ + ������������������������2
508
+ (6)
509
+ From this fitting, a ������������ and ������������ value of 0.0153(2) J mol-1 K-2 and 7.1(2) × 10-4 J mol-1 K-3
510
+ corresponding to the electronic and phonon contributions to the specific heat, respectively, could
511
+ be obtained. The ������������ value recovered the Debye temperature (Eqn. 7) to be 268(2) K, which is much
512
+ closer to ������������������������1 rather than ������������������������2. It falls out of the temperature interval, 300–350 K, where iridates
513
+ most commonly exhibit Debye temperatures.41
514
+ ������������D = �12������������4
515
+ 5������������ �������������������������
516
+ 1
517
+ 3
518
+ (7)
519
+
520
+
521
+ 15
522
+
523
+
524
+ Fig. 5 Specific heat data fitting of high pressure Sr2IrO4. (a) Temperature dependence of
525
+ specific heat over temperature (��������������p ������������
526
+ ⁄ ) for high-pressure Sr2IrO4 fitted by two Debye model in
527
+ orange. Green and red dotted lines refer to the 1st and 2nd Debye model. (b) ������������p ������������
528
+ ⁄ vs ������������2 between
529
+ 2–3.2 K fitted with Eqn. 6 (orange dotted line).
530
+
531
+ Mott Variable-range Hopping (VRH). It is critical to investigate the electrical conductivity in
532
+ the high pressure Sr2IrO4 phase to compare to the Mott insulator ambient pressure Sr2IrO4.
533
+ Temperature-dependent resistivity measurements were performed from 2–300 K with an applied
534
+ field up to 9 T on a pelletized polycrystalline sample of the high pressure Sr2IrO4 phase, shown in
535
+ Fig. 6a. No significant field dependence was observed, which indicates the insignificance of
536
+ magnetoresistance for the high pressure Sr2IrO4 phase. This may be not unexpected considering
537
+ the small saturation moment under fields (see the discussion above). At room temperature and 0
538
+ T, the resistivity is relatively low, only around 4 Ω cm. However, the resistivity is increases by 6
539
+ orders of magnitude upon cooling, indicating the semiconducting character of the high-pressure
540
+ Sr2IrO4 phase.
541
+
542
+ To further analyze its behavior, we first tried to model the temperature dependence of ������������ with the
543
+ Arrhenius law (Eqn. 8),
544
+ ������������ = ������������0������������������������������������ ������������������������
545
+
546
+ (8)
547
+
548
+ (a)
549
+ (b)
550
+ 1e-1
551
+ 1e-2
552
+ two Debye
553
+ OT
554
+ ... y+βT2
555
+ OOT
556
+ 8
557
+ Cp/T ( mol-1 K-2)
558
+ 8
559
+ 2.4
560
+ 6
561
+ 4
562
+ 2.0
563
+ 6
564
+ 2
565
+ Debyel
566
+ Debye2
567
+ 0
568
+ 0
569
+ 50
570
+ 100
571
+ 150
572
+ 200
573
+ 4
574
+ 8
575
+ 12
576
+ 16
577
+ T (K)
578
+ T2 (K2)16
579
+
580
+ where ������������0 is the residual resistivity, ������������������������ is the activation energy, and ������������ is the Boltzmann constant.
581
+ However, ������������ could not be fitted well to a ������������������������, shown in Fig. S7a, i.e., the Arrhenius law is not well
582
+ obeyed. Then its temperature dependence was fitted by law in the form (Eqn. 9) with ������������ of 1/2 and
583
+ 1/4,
584
+ ������������ = ������������0������������(������������0 ������������
585
+ ⁄ )������������
586
+ (9)
587
+ where ������������0 is the residual resistivity, and ������������0 is the characteristic temperature. The fitting results were
588
+ presented in Fig. 6b, and Fig. S8, with parameters summarized in Table S4. The value ������������ of 1/4 is
589
+ favored over 1/2. While both of them indicate three-dimensional Mott variable-range hopping of
590
+ charge carriers between localized states, the weaker temperature dependence with ������������ of 1/4 implies
591
+ negligible long-range Coulomb interactions between localized electrons in the temperature regime
592
+ studied. This behavior is also reported in the ambient pressure Sr2IrO4.42 To explore the harboring
593
+ quantum states in the high-pressure Sr2IrO4 phase, further examination of its transport properties
594
+ is warranted.
595
+
596
+
597
+
598
+ Fig. 6 Details of field and temperature dependent resistivity. (a) Temperature dependence of
599
+ resistivity data for high-pressure phase Sr2IrO4 under fields up to 9 T. No significant derivation
600
+ was observed. (b) The resistivity ������������ (blue hollow circle) ranging from 80–300 K was fitted by Eqn.
601
+ 9 with ������������ of 1/4 (orange line). A linear relationship was obtained.
602
+
603
+
604
+
605
+
606
+ (a)
607
+ (b)
608
+ 107
609
+ OT
610
+ fit
611
+ 1 T
612
+ o data
613
+ 3 T
614
+ 8
615
+ 105
616
+ 5 T
617
+ In(p/(2 cm))
618
+ p (Q cm)
619
+ 7 T
620
+ 9T
621
+ 103
622
+ 4
623
+ 101
624
+ 0
625
+ 0
626
+ 100
627
+ 200
628
+ 300
629
+ 0.2
630
+ 0.3
631
+ 0.4
632
+ 0.5
633
+ T (K)
634
+ T-1/4 (K-1/4)17
635
+
636
+ Conclusion
637
+
638
+ In summary, we reported the non-centrosymmetric Sr2IrO4 phase obtained under high
639
+ pressure and high temperature conditions. The ferromagnetic ordering temperature decreases
640
+ significantly to ������������c ~86 K from ~240 K in the ambient pressure Sr2IrO4, while there may be a
641
+ possible breakdown of the Curie-Weiss law under higher temperatures. Diamagnetism was
642
+ observed under room temperature and higher fields. No anomalies indicating magnetic ordering
643
+ were observed in the specific heat measurements, where a greater photon contribution was
644
+ obtained from the low-temperature regime. Temperature-dependent resistivity revealed three-
645
+ dimensional Mott variable-range hopping of charge carriers between states localized by disorder
646
+ with negligible long-range Coulomb repulsions. Further transport measurements, together with
647
+ first-principal calculation, are expected to explore the electronic properties of the high-pressure
648
+ Sr2IrO4 phase. Such a system may offer a promising platform to unravel the mystery of high-Tc
649
+ superconductivity in cuprates.
650
+
651
+
652
+
653
+ Acknowledgments
654
+
655
+ The work at Rutgers was supported by U.S. DOE-BES under Contract DE-SC0022156.
656
+ The electron microscopy work at BNL was supported by U.S. DOE-BES, Materials Sciences and
657
+ Engineering Division under Contract No. DESC0012704.
658
+
659
+ Supporting Information
660
+
661
+ Single crystal X-ray diffraction data at room temperature and 100 K; Anisotropic
662
+ displacement parameters; Atomic coordinates and equivalent isotropic displacement parameters;
663
+ PXRD overlay of Sr2IrO4; Magnetic susceptibility and Curie-Weiss fitting; Magnetic hysteresis;
664
+ Field dependence of specific heat; Specific heat data fitted by Debye and Einstein model; Low
665
+ temperature specific heat data (2–20 K); Temperature dependence of resistivity; Resistivity data
666
+ fitted by Eqn. 9 with ������������ of 1/2 and 1/4; Summary of fitting parameters for resistivity data.
667
+
668
+
669
+
670
+
671
+ 18
672
+
673
+ References
674
+ 1. Takagi, H.; Takayama, T.; Jackeli, G.; Khaliullin, G.; Nagler, S. E. Concept and realization of
675
+ Kitaev quantum spin liquids. Nat. Rev. Phys. 2019, 1, 264-280.
676
+ 2. Revelli, A.; Moretti Sala, M.; Monaco, G.; Hickey, C.; Becker, P.; Freund, F.; Jesche, A.;
677
+ Gegenwart, P.; Eschmann, T.; Buessen, F. L.; Trebst, S.; van Loosdrecht, P. H. M.; van den Brink,
678
+ J.; Grüninger, M. Fingerprints of Kitaev physics in the magnetic excitations of honeycomb iridates.
679
+ Phys. Rev. Res. 2020, 2, 043094.
680
+ 3. Gao, Y.; Zhou, T.; Huang, H.; Wang, Q.-H. Possible superconductivity in Sr2IrO4 probed by
681
+ quasiparticle interference. Sci. Rep. 2015, 5, 9251.
682
+ 4. Mitchell, J. F. Sr2IrO4: Gateway to cuprate superconductivity? APL Mater. 2015, 3, 062404.
683
+ 5. Cao, Y.; Wang, Q.; Waugh, J. A.; Reber, T. J.; Li, H.; Zhou, X.; Parham, S.; Park, S. R.; Plumb,
684
+ N. C.; Rotenberg, E.; Bostwick, A.; Denlinger, J. D.; Qi, T.; Hermele, M. A.; Cao, G.; Dessau, D.
685
+ S. Hallmarks of the Mott-metal crossover in the hole-doped pseudospin-1/2 Mott insulator Sr2IrO4.
686
+ Nat. Commun. 2016, 7, 11367.
687
+ 6. Nichols, J.; Bray-Ali, N.; Ansary, A.; Cao, G.; Ng, K.-W. Tunneling into the Mott insulator
688
+ Sr2IrO4. Phys. Rev. B 2014, 89, 085125.
689
+ 7. Kim, B. J.; Jin, H.; Moon, S. J.; Kim, J. Y.; Park, B. G.; Leem, C. S.; Yu, J.; Noh, T. W.; Kim,
690
+ C.; Oh, S. J.; Park, J. H.; Durairaj, V.; Cao, G.; Rotenberg, E. Novel Jeff=1/2 Mott State Induced
691
+ by Relativistic Spin-Orbit Coupling in Sr2IrO4. Phys. Rev. Lett. 2008, 101, 076402.
692
+ 8. Kim, B. J.; Ohsumi, H.; Komesu, T.; Sakai, S.; Morita, T.; Takagi, H.; Arima, T. Phase-Sensitive
693
+ Observation of a Spin-Orbital Mott State in Sr2IrO4. Science 2009, 323, 1329-1332.
694
+ 9. Ye, F.; Chi, S.; Cao, H.; Chakoumakos, B. C.; Fernandez-Baca, J. A.; Custelcean, R.; Qi, T. F.;
695
+ Korneta, O. B.; Cao, G. Direct evidence of a zigzag spin-chain structure in the honeycomb lattice:
696
+ A neutron and x-ray diffraction investigation of single-crystal Na2IrO3. Phys. Rev. B 2012, 85,
697
+ 180403.
698
+ 10. Grant, P. M.; Parkin, S. S. P.; Lee, V. Y.; Engler, E. M.; Ramirez, M. L.; Vazquez, J. E.; Lim,
699
+ G.; Jacowitz, R. D.; Greene, R. L. Evidence for superconductivity in La2CuO4. Phys. Rev. Lett.
700
+ 1987, 58, 2482-2485.
701
+ 11. Dean, M. P. M.; Springell, R. S.; Monney, C.; Zhou, K. J.; Pereiro, J.; Božović, I.; Dalla Piazza,
702
+ B.; Rønnow, H. M.; Morenzoni, E.; van den Brink, J.; Schmitt, T.; Hill, J. P. Spin excitations in a
703
+ single La2CuO4 layer. Nat. Mater. 2012, 11, 850-854.
704
+ 12. Attfield, J. P.; Kharlanov, A. L.; McAllister, J. A. Cation effects in doped La2CuO4
705
+ superconductors. Nature 1998, 394, 157-159.
706
+ 13. Battisti, I.; Bastiaans, K. M.; Fedoseev, V.; de la Torre, A.; Iliopoulos, N.; Tamai, A.; Hunter,
707
+ E. C.; Perry, R. S.; Zaanen, J.; Baumberger, F.; Allan, M. P. Universality of pseudogap and
708
+ emergent order in lightly doped Mott insulators. Nat. Phys. 2017, 13, 21-25.
709
+ 14. Kim, Y. K.; Sung, N. H.; Denlinger, J. D.; Kim, B. J. Observation of a d-wave gap in electron-
710
+ doped Sr2IrO4. Nat. Phys. 2016, 12, 37-41.
711
+
712
+ 19
713
+
714
+ 15. He, J.; Hafiz, H.; Mion, T. R.; Hogan, T.; Dhital, C.; Chen, X.; Lin, Q.; Hashimoto, M.; Lu, D.
715
+ H.; Zhang, Y.; Markiewicz, R. S.; Bansil, A.; Wilson, S. D.; He, R.-H. Fermi Arcs vs. Fermi
716
+ Pockets in Electron-doped Perovskite Iridates. Sci. Rep. 2015, 5, 8533.
717
+ 16. Zhao, L.; Torchinsky, D. H.; Chu, H.; Ivanov, V.; Lifshitz, R.; Flint, R.; Qi, T.; Cao, G.; Hsieh,
718
+ D. Evidence of an odd-parity hidden order in a spin–orbit coupled correlated iridate. Nat. Phys.
719
+ 2016, 12, 32-36.
720
+ 17. Chikara, S.; Fabbris, G.; Terzic, J.; Cao, G.; Khomskii, D.; Haskel, D. Charge partitioning and
721
+ anomalous hole doping in Rh-doped Sr2IrO4. Phys. Rev. B 2017, 95, 060407.
722
+ 18. Sohn, C. H.; Cho, D.-Y.; Kuo, C. T.; Sandilands, L. J.; Qi, T. F.; Cao, G.; Noh, T. W. X-ray
723
+ Absorption Spectroscopy Study of the Effect of Rh doping in Sr2IrO4. Sci. Rep. 2016, 6, 23856.
724
+ 19. Qi, T. F.; Korneta, O. B.; Li, L.; Butrouna, K.; Cao, V. S.; Wan, X.; Schlottmann, P.; Kaul, R.
725
+ K.; Cao, G. Spin-orbit tuned metal-insulator transitions in single-crystal Sr2Ir1-xRhxO4. Phys. Rev.
726
+ B 2012, 86, 125105.
727
+ 20. Clancy, J. P.; Lupascu, A.; Gretarsson, H.; Islam, Z.; Hu, Y. F.; Casa, D.; Nelson, C. S.;
728
+ LaMarra, S. C.; Cao, G.; Kim, Y.-J. Dilute magnetism and spin-orbital percolation effects in Sr2Ir1-
729
+ xRhxO4. Phys. Rev. B 2014, 89, 054409.
730
+ 21. Ye, F.; Wang, X.; Hoffmann, C.; Wang, J.; Chi, S.; Matsuda, M.; Chakoumakos, B. C.;
731
+ Fernandez-Baca, J. A.; Cao, G. Structure symmetry determination and magnetic evolution in
732
+ Sr2Ir1-xRhxO4. Phys. Rev. B 2015, 92, 201112.
733
+ 22. Brouet, V.; Mansart, J.; Perfetti, L.; Piovera, C.; Vobornik, I.; Le Fèvre, P.; Bertran, F.; Riggs,
734
+ S. C.; Shapiro, M. C.; Giraldo-Gallo, P.; Fisher, I. R. Transfer of spectral weight across the gap of
735
+ Sr2IrO4 induced by La doping. Phys. Rev. B 2015, 92, 081117.
736
+ 23. Chikara, S.; Haskel, D.; Sim, J.-H.; Kim, H.-S.; Chen, C.-C.; Fabbris, G.; Veiga, L. S. I.; Souza-
737
+ Neto, N. M.; Terzic, J.; Butrouna, K.; Cao, G.; Han, M. J.; van Veenendaal, M. Sr2Ir1−xRhxO4 (x <
738
+ 0.5): An inhomogeneous jeff = 1/2 Hubbard system. Phys. Rev. B 2015, 92, 081114.
739
+ 24. Haskel, D.; Fabbris, G.; Zhernenkov, M.; Kong, P. P.; Jin, C. Q.; Cao, G.; van Veenendaal, M.
740
+ Pressure Tuning of the Spin-Orbit Coupled Ground State in Sr2IrO4. Phys. Rev. Lett. 2012, 109,
741
+ 027204.
742
+ 25. Chen, C.; Zhou, Y.; Chen, X.; Han, T.; An, C.; Zhou, Y.; Yuan, Y.; Zhang, B.; Wang, S.;
743
+ Zhang, R.; Zhang, L.; Zhang, C.; Yang, Z.; DeLong, L. E.; Cao, G. Persistent insulating state at
744
+ megabar pressures in strongly spin-orbit coupled Sr2IrO4. Phys. Rev. B 2020, 101, 144102.
745
+ 26. Bhatti, I. N.; Rawat, R.; Banerjee, A.; Pramanik, A. K. Temperature evolution of magnetic and
746
+ transport behavior in 5d Mott insulator Sr2IrO4: significance of magneto-structural coupling. J.
747
+ Phys.: Condens. Matter 2014, 27, 016005.
748
+ 27. Walker, D.; Carpenter, M. A.; Hitch, C. M. Some simplifications to multianvil devices for high
749
+ pressure experiments. Am. Mineral. 1990, 75, 1020-1028.
750
+ 28. Rodríguez-Carvajal, J. Recent advances in magnetic structure determination by neutron
751
+ powder diffraction. Physica B: Condensed Matter 1993, 192, 55-69.
752
+ 29. Dinnebier, R. E.; Billinge, S. J. L., Chapter 1 Principles of Powder Diffraction. In Powder
753
+ Diffraction: Theory and Practice, The Royal Society of Chemistry: 2008; pp 1-19.
754
+
755
+ 20
756
+
757
+ 30. Sheldrick, G. Crystal structure refinement with SHELXL. Acta Crystallogr., Sect. C 2015, 71,
758
+ 3-8.
759
+ 31. Sheldrick, G. SHELXT - Integrated space-group and crystal-structure determination. Acta
760
+ Crystallogr., Sect. A 2015, 71, 3-8.
761
+ 32. Huang, Q.; Soubeyroux, J. L.; Chmaissem, O.; Sora, I. N.; Santoro, A.; Cava, R. J.; Krajewski,
762
+ J. J.; Peck, W. F. Neutron Powder Diffraction Study of the Crystal Structures of Sr2RuO4 and
763
+ Sr2IrO4 at Room Temperature and at 10 K. J. Solid State Chem. 1994, 112, 355-361.
764
+ 33. Plotnikova, E. M.; Daghofer, M.; van den Brink, J.; Wohlfeld, K. Jahn-Teller Effect in Systems
765
+ with Strong On-Site Spin-Orbit Coupling. Phys. Rev. Lett. 2016, 116, 106401.
766
+ 34. Dikushina, E. A.; Avvakumov, I. L. Study of the influence of a spin-orbit exciton on the
767
+ magnetic ordering in Sr2IrO4. J. Phys. Conf. Ser. 2016, 741, 012016.
768
+ 35. Crawford, M. K.; Subramanian, M. A.; Harlow, R. L.; Fernandez-Baca, J. A.; Wang, Z. R.;
769
+ Johnston, D. C. Structural and magnetic studies of Sr2IrO4. Phys. Rev. B 1994, 49, 9198-9201.
770
+ 36. Samanta, K.; Tartaglia, R.; Kaneko, U. F.; Souza-Neto, N. M.; Granado, E. Anisotropic lattice
771
+ compression and pressure-induced electronic phase transitions in Sr2IrO4. Phys. Rev. B 2020, 101,
772
+ 075121.
773
+ 37. Samanta, K.; Ardito, F. M.; Souza-Neto, N. M.; Granado, E. First-order structural transition
774
+ and pressure-induced lattice/phonon anomalies in Sr2IrO4. Phys. Rev. B 2018, 98, 094101.
775
+ 38. Longo, J. M.; Raccah, P. M. The structure of La2CuO4 and LaSrVO4. J. Solid State Chem.
776
+ 1973, 6, 526-531.
777
+ 39. Ye, F.; Chi, S.; Chakoumakos, B. C.; Fernandez-Baca, J. A.; Qi, T.; Cao, G. Magnetic and
778
+ crystal structures of Sr2IrO4: A neutron diffraction study. Phys. Rev. B 2013, 87, 140406.
779
+ 40. Kini, N. S.; Strydom, A. M.; Jeevan, H. S.; Geibel, C.; Ramakrishnan, S. Transport and thermal
780
+ properties of weakly ferromagnetic Sr2IrO4. J. Phys.: Condens. Matter 2006, 18, 8205-8216.
781
+ 41. Pallecchi, I.; Buscaglia, M. T.; Buscaglia, V.; Gilioli, E.; Lamura, G.; Telesio, F.; Cimberle,
782
+ M. R.; Marré, D. Thermoelectric behavior of Ruddlesden–Popper series iridates. J. Phys.: Condens.
783
+ Matter 2016, 28, 065601.
784
+ 42. Cao, G.; Bolivar, J.; McCall, S.; Crow, J. E.; Guertin, R. P. Weak ferromagnetism, metal-to-
785
+ nonmetal transition, and negative differential resistivity in single-crystal Sr2IrO4. Phys. Rev. B
786
+ 1998, 57, R11039-R11042.
787
+
788
+
789
+
790
+
791
+ 21
792
+
793
+ Non-centrosymmetric Sr2IrO4 obtained under high pressure
794
+
795
+
796
+ Haozhe Wang1‡, Madalynn Marshall2‡, Zhen Wang3, Kemp W. Plumb4, Martha Greenblatt2,
797
+ Yimei Zhu3, David Walker5, Weiwei Xie1*
798
+
799
+ 1. Department of Chemistry, Michigan State University, East Lansing, Michigan 48824, USA
800
+ 2. Department of Chemistry and Chemical Biology, Rutgers University, Piscataway, New Jersey
801
+ 08854, USA
802
+ 3. Condensed Matter Physics and Materials Science Department, Brookhaven National
803
+ Laboratory, Upton, New York 11973, USA
804
+ 4. Department of Physics, Brown University, Providence, Rhode Island 02912, USA
805
+ 5. Lamont Doherty Earth Observatory, Columbia University, Palisades, New York 10964, USA
806
+
807
+
808
+ ‡ H.W. and M.M. contributed equally. * Email: xieweiwe@msu.edu
809
+
810
+
811
+
812
+ Supporting Information
813
+
814
+ Table S1 Single crystal X-ray diffraction data at room temperature and 100 K .......................... S2
815
+ Table S2 Anisotropic displacement parameters ........................................................................... S3
816
+ Table S3 Atomic coordinates and equivalent isotropic displacement parameters ....................... S4
817
+ Figure S1 PXRD overlay of Sr2IrO4 ............................................................................................ S5
818
+ Figure S2 Magnetic susceptibility and Curie-Weiss fitting ......................................................... S6
819
+ Figure S3 Magnetic hysteresis ..................................................................................................... S7
820
+ Figure S4 Field dependence of specific heat ............................................................................... S8
821
+ Figure S5 Specific heat data fitted by Debye and Einstein model ............................................... S9
822
+ Figure S6 Low temperature specific heat data (2–20 K) ........................................................... S10
823
+ Figure S7 Temperature dependence of resistivity ...................................................................... S11
824
+ Figure S8 Resistivity data fitted by Equation 9 with ������������ of 1/2 and 1/4 ..................................... S12
825
+ Table S4 Summary of fitting parameters for resistivity data ..................................................... S13
826
+
827
+
828
+ 22
829
+
830
+ Table S1 Single crystal X-ray diffraction data at room temperature and 100 K.
831
+ Temperature
832
+ Room Temperature
833
+ 100 K
834
+ Refined formula
835
+ Sr2IrO4
836
+ Sr2IrO4
837
+ FW (g/mol)
838
+ 431.44
839
+ 431.44
840
+ Space group
841
+ I4mm
842
+ I4mm
843
+ a (Å)
844
+ 3.8860(5)
845
+ 3.8777(5)
846
+ c (Å)
847
+ 12.826(2)
848
+ 12.825(2)
849
+ V (Å3)
850
+ 193.69(6)
851
+ 192.85(6)
852
+ Extinction Coefficient
853
+ N/A
854
+ N/A
855
+ ������������ range (°)
856
+ 3.177–33.030
857
+ 3.177–33.075
858
+ # of reflections; Rint
859
+ 1088; 0.0627
860
+ 1286; 0.0591
861
+ # of independent reflections
862
+ 267
863
+ 264
864
+ # of parameters
865
+ 23
866
+ 23
867
+ R1; ωR2 (������������ > ������������������������(������������))
868
+ 0.0409; 0.0651
869
+ 0.0312; 0.0443
870
+ Goodness of fit (GOF)
871
+ 1.177
872
+ 1.125
873
+ Diffraction peak and hole (e-/ Å3)
874
+ 3.658, -3.492
875
+ 2.359, -1.96
876
+
877
+
878
+
879
+
880
+ 23
881
+
882
+ Table S2 Anisotropic displacement parameters for Sr2IrO4 at room temperature and 100 K.
883
+ Sr2IrO4 at Room Temperature
884
+ Atom
885
+ U11
886
+ U22
887
+ U33
888
+ U23
889
+ U13
890
+ U12
891
+ Ir1
892
+ -0.0018(6)
893
+ -0.0018(6)
894
+ -0.0021(6)
895
+ 0
896
+ 0
897
+ 0
898
+ Sr1
899
+ 0.026(7)
900
+ 0.026(7)
901
+ 0.007(7)
902
+ 0
903
+ 0
904
+ 0
905
+ Sr2
906
+ -0.001(4)
907
+ -0.001(4)
908
+ 0.005(6)
909
+ 0
910
+ 0
911
+ 0
912
+ O1
913
+ 0.03(2)
914
+ 0.03(2)
915
+ -0.02(2)
916
+ 0
917
+ 0
918
+ 0
919
+ O2
920
+ -0.006(10)
921
+ -0.006(10)
922
+ -0.023(18)
923
+ 0
924
+ 0
925
+ 0
926
+ O3
927
+ 0.04(3)
928
+ 0.003(11)
929
+ -0.01(3)
930
+ 0
931
+ 0.02(3)
932
+ 0
933
+
934
+ Sr2IrO4 at 100 K
935
+ Atom
936
+ U11
937
+ U22
938
+ U33
939
+ U23
940
+ U13
941
+ U12
942
+ Ir1
943
+ -0.0004(3)
944
+ -0.0004(3)
945
+ 0.0012(7)
946
+ 0
947
+ 0
948
+ 0
949
+ Sr1
950
+ 0.005(8)
951
+ 0.005(8)
952
+ 0.005(4)
953
+ 0
954
+ 0
955
+ 0
956
+ Sr2
957
+ 0.002(8)
958
+ 0.002(8)
959
+ 0.000(4)
960
+ 0
961
+ 0
962
+ 0
963
+ O1
964
+ 0.005(8)
965
+ 0.005(8)
966
+ -0.033(15)
967
+ 0
968
+ 0
969
+ 0
970
+ O2
971
+ 0.012(10)
972
+ 0.012(10)
973
+ -0.033(14)
974
+ 0
975
+ 0
976
+ 0
977
+ O3
978
+ 0.009(10)
979
+ 0.005(7)
980
+ 0.006(8)
981
+ 0
982
+ 0.01(3)
983
+ 0
984
+
985
+
986
+
987
+
988
+ 24
989
+
990
+ Table S3 Atomic coordinates and equivalent isotropic displacement parameters for Sr2IrO4 at room
991
+ temperature and 100 K. (Ueq is defined as one-third of the trace of the orthogonalized Uij tensor (Å2)).
992
+ Sr2IrO4 at Room Temperature
993
+ Atom
994
+ Wyck.
995
+ x
996
+ y
997
+ z
998
+ Occ.
999
+ Ueq
1000
+ Ir1
1001
+ 2a
1002
+ 0
1003
+ 0
1004
+ 0.1513(13)
1005
+ 1
1006
+ -0.0019(4)
1007
+ Sr2
1008
+ 2a
1009
+ 0
1010
+ 0
1011
+ 0.5044(4)
1012
+ 1
1013
+ 0.020(4)
1014
+ Sr1
1015
+ 2a
1016
+ 0
1017
+ 0
1018
+ 0.79985(2)
1019
+ 1
1020
+ 0.001(3)
1021
+ O1
1022
+ 2a
1023
+ 0
1024
+ 0
1025
+ 0.328(4)
1026
+ 1
1027
+ 0.013(18)
1028
+ O2
1029
+ 2a
1030
+ 0
1031
+ 0
1032
+ 0.000(4)
1033
+ 1
1034
+ -0.011(7)
1035
+ O3
1036
+ 8d
1037
+ 0.419(9)
1038
+ 0
1039
+ 0.661(7)
1040
+ 0.5
1041
+ 0.010(15)
1042
+
1043
+ Sr2IrO4 at 100 K
1044
+ Atom
1045
+ Wyck.
1046
+ x
1047
+ y
1048
+ z
1049
+ Occ.
1050
+ Ueq
1051
+ Ir1
1052
+ 2a
1053
+ 0
1054
+ 0
1055
+ 0.1489(7)
1056
+ 1
1057
+ 0.0001(3)
1058
+ Sr2
1059
+ 2a
1060
+ 0
1061
+ 0
1062
+ 0.5019(4)
1063
+ 1
1064
+ 0.005(5)
1065
+ Sr1
1066
+ 2a
1067
+ 0
1068
+ 0
1069
+ 0.7978(2)
1070
+ 1
1071
+ 0.002(5)
1072
+ O1
1073
+ 2a
1074
+ 0
1075
+ 0
1076
+ 0.321(3)
1077
+ 1
1078
+ -0.008(6)
1079
+ O2
1080
+ 2a
1081
+ 0
1082
+ 0
1083
+ 0.000(3)
1084
+ 1
1085
+ -0.003(8)
1086
+ O3
1087
+ 8d
1088
+ 0.412(4)
1089
+ 0
1090
+ 0.649(6)
1091
+ 0.5
1092
+ 0.007(4)
1093
+
1094
+
1095
+
1096
+
1097
+ 25
1098
+
1099
+ Figure S1 Powder X-ray diffraction pattern overlay. The experimental data of high pressure Sr2IrO4
1100
+ phase synthesized at 1400 °C for ~4 hrs (black line) and ~28 hrs (red line) were presented. Bragg peak
1101
+ positions are indicated as Sr2IrO4 and Sr3Ir2O7 with green and purple vertical tick marks, respectively.
1102
+
1103
+
1104
+
1105
+
1106
+
1107
+
1108
+ TT-1412 ~4 hrs
1109
+ GG-1418 -28 hrs
1110
+ Intensity (a.u.)
1111
+ Sr2lrO4
1112
+ Sr3lr2Q7
1113
+ 10
1114
+ 30
1115
+ 50
1116
+ 70
1117
+ 90
1118
+ 2e (degree)26
1119
+
1120
+ Figure S2 Magnetic susceptibility and Curie-Weiss fitting. (a) Temperature derivative of magnetic
1121
+ susceptibility ������������ at 1000 Oe under FCW mode. The minimum at around 84 K was highlighted by red circle
1122
+ and an arrow. (b) The inverse magnetic susceptibility data (FCW, 80–140 K, blue hollow circle) fitted with
1123
+ the modified Curie-Weiss model (orange line). (c) The Curie-Weiss fit was further extrapolated to 160 K.
1124
+
1125
+
1126
+
1127
+
1128
+
1129
+
1130
+ (a)
1131
+ (b)
1132
+ 1e-3
1133
+ 1e3
1134
+ K-1)
1135
+ Cw fit
1136
+ 1.2 -
1137
+ FCW
1138
+ -1
1139
+ 00
1140
+ 0.6 -
1141
+ 8
1142
+ -2
1143
+ 1/x
1144
+ 000
1145
+ 0.0 -
1146
+ 000
1147
+ 50
1148
+ 60
1149
+ 70
1150
+ 80 °
1151
+ 90
1152
+ 100
1153
+ 80
1154
+ 100
1155
+ 120
1156
+ 140
1157
+ T (K)
1158
+ T (K)
1159
+ (c)
1160
+ 1e2
1161
+ mol Oe)
1162
+ 4
1163
+ CW fit
1164
+ FCW
1165
+ N
1166
+ 0%
1167
+ 00
1168
+ 0
1169
+ 80
1170
+ 100
1171
+ 120
1172
+ 140
1173
+ 160
1174
+ T (K)27
1175
+
1176
+ Figure S3 Magnetic hysteresis. Magnetic hysteresis observed in the high pressure Sr2IrO4 phase at 2 K
1177
+ ranging from -0.6 T to 0.6 T.
1178
+
1179
+
1180
+
1181
+
1182
+
1183
+
1184
+ 1e-2
1185
+ 0 2K
1186
+ 1
1187
+ M (μB per Ir ion)
1188
+ -1
1189
+ -0.6
1190
+ -0.3
1191
+ 0.0
1192
+ 0.3
1193
+ 0.6
1194
+ μoH (T)28
1195
+
1196
+ Figure S4 Field dependence of specific heat. Temperature dependence of specific heat data over
1197
+ temperature (������������p/������������) for high pressure Sr2IrO4 phase, under 0 T (blue), 0.05 T (orange), and 1 T (green). No
1198
+ significant differences were observed. No ������������ shape anomalies emerged in the whole temperature regime
1199
+ studied under either case.
1200
+
1201
+
1202
+
1203
+
1204
+
1205
+
1206
+ le-1
1207
+ 8
1208
+ 6
1209
+ 4
1210
+ O T
1211
+ 2
1212
+ 0.05 T
1213
+ 0
1214
+ 1 T
1215
+ 0
1216
+ 0
1217
+ 50
1218
+ 100
1219
+ 150
1220
+ 200
1221
+ T (K)29
1222
+
1223
+ Figure S5 Specific heat data fitted by Debye and Einstein model. Temperature dependence of
1224
+ specific heat data over temperature under 0 T (������������p/������������, blue hollow circle) for high pressure Sr2IrO4 phase,
1225
+ fitted by (a) Debye model, (b) Einstein model, (c) two Debye model with the electronic contribution
1226
+ included, and weighted Debye model (d) without and (e) with the electronic contribution included.
1227
+
1228
+
1229
+
1230
+
1231
+
1232
+ (a)
1233
+ (b)
1234
+ le-1
1235
+ le-1
1236
+ 8
1237
+ K-2)
1238
+ 6
1239
+ 6
1240
+ 4
1241
+ 4
1242
+ Debye
1243
+ 2
1244
+ Einstein
1245
+ 0
1246
+ 10
1247
+ 10
1248
+ 0
1249
+ 0
1250
+ 0
1251
+ 50
1252
+ 100
1253
+ 150
1254
+ 200
1255
+ 0
1256
+ 50
1257
+ 100
1258
+ 150
1259
+ 200
1260
+ T (K)
1261
+ T (K)
1262
+ (c)
1263
+ le-1
1264
+ (d)
1265
+ le-1
1266
+ two Debye + yT
1267
+ OOT
1268
+ weightedDebye
1269
+ 10
1270
+ VT
1271
+ 8
1272
+ 8 -
1273
+ Cp/T (I mol-1 K-2)
1274
+ K-2)
1275
+ 6
1276
+ 6
1277
+ 4
1278
+ 4
1279
+ 2
1280
+ 2
1281
+ Debyel
1282
+ Debye
1283
+ Debye2
1284
+ Einstein
1285
+ 0
1286
+ 0
1287
+ 0
1288
+ 50
1289
+ 100
1290
+ 150
1291
+ 200
1292
+ 0
1293
+ 50
1294
+ 100
1295
+ 150
1296
+ 200
1297
+ T (K)
1298
+ T (K)
1299
+ (e)
1300
+ le-1
1301
+ weighted Debye + yT o O T
1302
+ Cp/T (I mol-1 K-2)
1303
+ 8
1304
+ 6
1305
+ Debye
1306
+ 2
1307
+ Einstein
1308
+ yT
1309
+ 0
1310
+ 0
1311
+ 50
1312
+ 100
1313
+ 150
1314
+ 200
1315
+ T (K)30
1316
+
1317
+ Figure S6 Low temperature specific heat data (2–20 K). Specific heat data over temperature (������������p/������������)
1318
+ plotted versus ������������2 under low temperature regime, 2–20 K, providing the possibility to derivate the
1319
+ Sommerfeld parameter, ������������.
1320
+
1321
+
1322
+
1323
+
1324
+
1325
+
1326
+ le-1
1327
+ 1oo
1328
+ 2
1329
+ 0
1330
+ 0
1331
+ 0
1332
+ 0
1333
+ 1 :
1334
+ 0
1335
+ 0
1336
+ 0
1337
+ 888
1338
+ 0
1339
+ .
1340
+ 0
1341
+ 1
1342
+ 2
1343
+ 3
1344
+ 4
1345
+ T2 (K2)
1346
+ 1e231
1347
+
1348
+ Figure S7 Temperature dependence of resistivity. Temperature dependence of the resistivity data ������������
1349
+ plotted as ln ������������ versus (a) ������������−1, (b) ������������−1/2, and (c) ������������−1/4 under 0 T.
1350
+
1351
+
1352
+
1353
+
1354
+
1355
+
1356
+ (a)
1357
+ (b)
1358
+ 16
1359
+ 16
1360
+ OT
1361
+ OT
1362
+ 8
1363
+ 12
1364
+ 12
1365
+ In(p/(α2 cm))
1366
+ In(p/(Ω2 cm))
1367
+ 8
1368
+ 8 -
1369
+ 4
1370
+ 4 -
1371
+ 0
1372
+ 0.0
1373
+ 0.1
1374
+ 0.2
1375
+ 0.0
1376
+ 0.2
1377
+ 0.4
1378
+ T-1 (K-1)
1379
+ T-1/2 (K-1/2)
1380
+ (c)
1381
+ 16
1382
+ O T
1383
+ 8
1384
+ 12
1385
+ In(p/(2 cm))
1386
+ 8
1387
+ 4 -
1388
+ 0
1389
+ 0.2
1390
+ 0.4
1391
+ 0.6
1392
+ T-1/4 (K-1/4)32
1393
+
1394
+ Figure S8 Resistivity data fitted by Equation 9 with ������������ of 1/2 and 1/4. (a) The resistivity data ������������
1395
+ (blue hollow circle) ranging from 110–300 K fitted by Equation 9 with ������������ of 1/2 (orange line). (b) The
1396
+ resistivity data ������������ in the low temperature regime ranging from 8–20 K fitted by Equation 9 with ������������ of 1/2.
1397
+ (c) The resistivity data ������������ in the low temperature regime ranging from 10–20 K fitted by Equation 9 with ������������
1398
+ of 1/4. Fitting parameters were summarized in Table S4. The value ������������ of 1/4 is favored over 1/2.
1399
+
1400
+
1401
+
1402
+
1403
+
1404
+
1405
+ (a)
1406
+ (b)
1407
+ 16
1408
+ fit
1409
+ fit
1410
+ 00
1411
+ data
1412
+ data
1413
+ 8 -
1414
+ In(p/(α2 cm))
1415
+ In(p/(Ω cm)
1416
+ 12
1417
+ 4
1418
+ 8
1419
+ 0
1420
+ 4
1421
+ 0.1
1422
+ 0.2
1423
+ 0.1
1424
+ 0.2
1425
+ 0.3
1426
+ 0.4
1427
+ T-1/2 (K-1/2)
1428
+ T-1/2 (K-1/2)
1429
+ (c)
1430
+ 16
1431
+ fit
1432
+ 8
1433
+ data
1434
+ 12
1435
+ In(p/(2 cm))
1436
+ 8 -
1437
+ 4 -
1438
+ 0.3
1439
+ 0.5
1440
+ 0.7
1441
+ T-1/4 (K-1/4)33
1442
+
1443
+
1444
+ Table S4. Summary of fitting parameters for resistivity data. Summary of fitting parameters for
1445
+ the resistivity data ������������ by Equation 9. R2 is the coefficient of determination.
1446
+
1447
+ ������������ = 1 2
1448
+
1449
+ Temperature Range / K
1450
+ ������������0 / (٠cm)
1451
+ ������������0 / K
1452
+ R2
1453
+ 110–300
1454
+ 3.01(4) × 10-5
1455
+ 6682
1456
+ 0.9996
1457
+ 8–20
1458
+ 79.6(37)
1459
+ 583
1460
+ 0.9996
1461
+
1462
+ ������������ = 1 4
1463
+
1464
+ Temperature Range / K
1465
+ ������������0 / (٠cm)
1466
+ ������������0 / K
1467
+ R2
1468
+ 80–300
1469
+ 7.82(12) × 10-5
1470
+ 3.83 × 106
1471
+ 0.9998
1472
+ 10–20
1473
+ -2.23(4)
1474
+ 4.10 × 105
1475
+ 0.9999
1476
+
1477
+
1478
+
1479
+
1480
+
1481
+
3tE4T4oBgHgl3EQf0g0b/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
49E2T4oBgHgl3EQfkAeM/content/tmp_files/2301.03974v1.pdf.txt ADDED
@@ -0,0 +1,979 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Hydrogen storage in Li functionalized [2,2,2]paracyclophane
2
+ at cryogenic to room temperatures: A computational quest
3
+ Rakesh K. Sahoo, Sridhar Sahu
4
+
5
+ Computational Materials Research Lab, Department of Physics, Indian Institute of Technology
6
+ (Indian School of Mines) Dhanbad, India
7
+ Abstract
8
+ In this work, we have studied the hydrogen adsorption-desorption properties, and storage capacities of Li
9
+ functionalized [2,2,2]paracyclophane (PCP222) using dispersion-corrected density functional theory and
10
+ molecular dynamic simulation. The Li atom was found bonded strongly with the benzene ring of PCP222
11
+ via Dewar interaction. Subsequently, the calculation of the diffusion energy barrier revealed a significantly
12
+ high energy barrier of 1.38 eV, preventing the Li clustering on PCP222. The host material, PCP222-3Li
13
+ adsorbed up to 15H2 molecules via charge polarization mechanism with an average adsorption energy of
14
+ 0.145 eV/5H2, suggesting physisorption type of adsorption. The PCP222 functionalized with three Li atom
15
+ showed maximum hydrogen uptake capacity up to 8.32 wt% which was fairly above the US-DOE criterion.
16
+ The practical H2storage estimation revealed that the PCP222-3Li desorbed 100% of adsorbed H2 molecules
17
+ at the temperature range of 260 K-300 K and pressure range of 1-10 bar. The maximum H2 desorption
18
+ temperature estimated by the Vant-Hoff relation was found to be 219 K and 266 K at 1 bar and 5 bar,
19
+ respectively. The ADMP molecular dynamics simulations assured the reversibility of adsorbed H2 and the
20
+ structural integrity of the host material at sufficiently above the desorption temperature (300K and 500K).
21
+ Therefore, the Li-functionalized PCP222 can be considered as a thermodynamically viable and potentially
22
+ reversible H2 storage material below room temperature.
23
+ Keywords: Hydrogen storage, DFT, Van’t-Hoff equation, ADMP, [2,2,2]paracyclophane,
24
+ PCP222, ESP
25
+ 1 Introduction
26
+ The excessive consumption of traditional fossil fuels has not only led to the depletion of the energy
27
+ supplies but also has emerged as the prime cause of environmental pollution. The global
28
+ consumption of petroleum and other traditional fossil fuel is anticipated to expand up to 56% by
29
+ the year 2040 and the crude oil supply is expected to endure until 2060 if the current demand trend
30
+ continues[1]. Thus, it is essential to develop alternative energy sources that are free from the
31
+ drawbacks of traditional fossil fuels. To meet the world’s energy demand and reduce the pollution
32
+ caused by fossil fuels, hydrogen has been considered as a plausible alternative due to its natural
33
+ abundance, environmental friendliness, and regenerative properties. One of the distinctive quality
34
+
35
+ of hydrogen is that it produces a large amount of energy per unit mass (120 MJ/kg) without
36
+ releasing any pollutant by-products [2, 3]. Despite these benefits, however, the use of hydrogen in
37
+ practice is limited due to the obstacle of finding the most appropriate and affordable way to store
38
+ and deliver hydrogen under normal environmental conditions. As per the criteria proposed by the
39
+ United State department of energy (DOE-US) an effective hydrogen storage material should have
40
+ a minimum storage capacity of up to 5.5 wt% by the year 2025 at moderate thermodynamics[5, 6].
41
+ In addition, as reported by many authors, the adsorption energy of hydrogen molecules of an
42
+ effective storage materials should be in the range of 0.1 eV/H2 to 0.6 eV/H2[4].
43
+ Though, numerous varieties of materials such as; metal hydrides [7, 8], graphene [9, 10],
44
+ metal alloys [11, 12], metal-organic frameworks (MOF) [13, 14], covalent-organic frameworks
45
+ (COF) [15] and carbon nanostructures [16, 17] etc have been investigated both theoretically and
46
+ experimentally as potential hydrogen storage materials, but there are many drawbacks and
47
+ unsolved issues to handle. The metal hydrides and complex hydrides store hydrogen via
48
+ chemisorption process which is highly irreversible and prevents easy desorption of hydrogen [18].
49
+ For example, Al(BH4)3, which yields hydrogen uptake capacity of up to 16.9 wt%, has high
50
+ desorption temperature (about 1000 K) that makes the material non-effective practical reversible
51
+ hydrogen storage applications[19]. Under ambient conditions, Mg-metal hydrides have a storage
52
+ capacity up to 7.6 wt%; however, it can only be used for 2-3 cycles[20]. Tavhare et al. studied the
53
+ hetero atom substituted Ti-benzene and reported an H2 uptake capacity up to 5.85 wt%, but at
54
+ relatively high desorption temperature (1193 K)[21]. Furthermore, MOF and COF applications are
55
+ constrained in the practical H2 storage field due to the difficulties of their heavy structure and
56
+ challenging step-wise production [22].
57
+ The efficient use of carbonaceous materials as hydrogen storage media was initially
58
+ reported by Dillon et al. [23]. Carbonaceous materials are appropriate for H2 storage due to their
59
+ unique qualities such as, large surface area, high porosity, better stabilities, and low densities.
60
+ However, the early findings have shown that these pure materials are weakly interact with the
61
+ hydrogen molecules (with BE ~4-5 kJ/mol), thus impractical for realistic hydrogen storage at
62
+ ambient environment [24, 25]. Meanwhile, carbon-based pure substrates are excellent materials
63
+ for hydrogen storage at cryogenic temperatures. For instance, pure single wall carbon nanotube
64
+
65
+ (SWCNT) can store hydrogen molecules up to 8.25 wt%, with a substantially lower desorption
66
+ temperature of 80 K [26].
67
+ It has been reported that the H2 interaction strength and the desorption temperature can be
68
+ tuned by integrating pure carbon substrates with alkali metal (AM)(Li, Na, and K), alkali earth
69
+ metals (Be, Mg, Ca), and transition metals (TM)(Sc, Ti, V, Y.)[27, 28, 29]. Numerous theoretical
70
+ investigations showed that integrating AMs and TMs with the carbon/borane substrates can bind
71
+ H2 molecules via charge polarization and the Kubas mechanism [30, 31]. The metallic atom
72
+ decorated fullerenes were first explored to investigate the impact of metal integration on pure
73
+ carbon substrates. According to studies by Sun et al., Li decorated fullerene could show a storage
74
+ capacity of 9 wt%; however, the hydrogen adsorption energy was estimated to be 0.075 eV/H2,
75
+ which is much lower than the DOE criterion [32]. The Li and Na-loaded C60 revealed H2 uptake
76
+ capacities of 4.5 wt% and 4 wt%, respectively, that were significantly below the target of DoE [33].
77
+ Experimental studies of transition metals like V and Pd decorated CNT reveal 0.66 wt% and 0.69
78
+ wt% of hydrogen capacity respectively, while pure CNT has 0.53 wt% of storage capacity [34].
79
+ Sahoo et al. reported storage of H2 on Li and Sc doped C8N8 cage via Niu-Rao-Jena and Kubas
80
+ interaction and estimated a desorption temperature of 286 K and 456 K, respectively [35]. The Li
81
+ and Na decorated on C24 fullerene could adsorb H2 molecules, with average hydrogen binding
82
+ energies of 0.198 eV/H2 and 0.164 eV/H2 and led to storage capacity up to 12.7 wt% and 10 wt %,
83
+ respectively [36]. Recently, we have investigated the H2 storage on alkali metal decorated
84
+ C20 fullerene and found the molecular hydrogen are physisorbed on host material via charge
85
+ polarization mechanism with desorption temperature of 182 -191 K [37]. Each Li and Na atom on
86
+ C20 could uptake up to 5H2 molecules with a total gravimetric storage capacity of 13.08 wt % and
87
+ 10.82 wt%, respectively, and the H2 binding energies found in the range of 0.12 eV—0.13 eV/H2.
88
+ Other carbonaceous materials such as functionalized organometallic compounds,
89
+ macrocyclic compounds have also been reported recently as potential candidates for hydrogen
90
+ storage. For example, Mahamiya et al. revealed the H2 storage capacities of 11. 9 wt % in K and
91
+ Ca decorated biphenylene with an average adsorption energy of 0.24-0.33 eV [38]. Y atom doped
92
+ zeolite shows high capacity adsorption of H2 with binding energy 0.35 eV/H2 and the desorption
93
+ energy of 437K for fuel cells[39]. Lithium-doped Calixarenes show an excellent hydrogen storage
94
+ behaviour but at very low up to 100 K [40]. Calix[4]arene functionalized with Li metal reveals 10
95
+
96
+ wt% storage capacity via Kubas—Niu—Rao—Jena interaction, and all most all H2 desorbed at a
97
+ temperature of 273 K [41].
98
+ Macrocyclic compounds such as, paracyclophane (PCP), a subgroup derivative of
99
+ cyclophanes, contains aromatic benzene rings, and their nomenclature is established on the arene
100
+ substitution pattern. For a [n,n]paracyclophane, the number of -CH2- moiety connecting the
101
+ successive benzene rings is indicated by the number in the square bracket [42]. Due to the
102
+ existence of aromatic benzene rings in the geometry, PCPs are easy to synthesize experimentally
103
+ and can be functionalized with metal atoms, making them a viable choice for hydrogen storage
104
+ candidates. A report on Li and Sc functionalized [4,4]paracyclophane revealed the hydrogen
105
+ uptake capacity up to 11.8 wt% and 13.7 wt% with an average adsorption energy of 0.08 eV/H2 and
106
+ 0.3 eV/H2 respectively [43]. Sahoo et al. recently studied the H2 storage capacity of
107
+ [1,1]paracyclophane functionalized with Sc and Y metals and found an H2 gravimetric storage
108
+ capacity of 8.22 wt% and 6.33 wt%, respectively, with an average adsorption energy 0.36
109
+ eV/H2[44]. They reported the H2 desorption temperature of 439 K and 412 K for Sc and Y doped
110
+ PCP11, respectively, at 1 atm. The hydrogen molecules are physisorbed on Li, and Sc decorated
111
+ paracyclophane via Kubas-Niu-Jena interaction and show a storage capacity of 10.3 wt%, as
112
+ reported by Sathe et al. [45]. Many more alkali metal-doped macrocyclic compounds have also
113
+ been investigated for hydrogen storage candidates and found the storage capacity above the DOE
114
+ target; however very few reported the practical H2 capacity at various thermodynamic
115
+ conditions[46, 47]. Though few of PCP-based hydrogen storage systems are available in literature,
116
+ the [2,2,2]paracyclophane (PCP222) which is experimentally synthesized by Tabushi et al.[48] is
117
+ yet to be explored as hydrogen storage material. Because Li, the lightest alkali metal atom and can
118
+ hold H2 molecules via charge polarization mechanism, it can serve as better sorption center on
119
+ PCP.
120
+ Therefore, in the current work, we intend to investigate the hydrogen storage properties
121
+ and potential of Li functionalized [2,2,2]paracyclophane (PCP222). We chose the PCP222 for
122
+ hydrogen storage because it is already experimentally synthesized and can be decorated with metal
123
+ atoms to form a hydrogen storage media with a high hydrogen uptake capacity. The Li atoms are
124
+ functionalized as sorption centers; this is because the light-weight metal doping method is an
125
+ effective way to increase the capacity of H2 storage. Li being the lightest alkali metal atom,
126
+
127
+ received a lot of attention to for hydrogen sorption application. Though there are few reports
128
+ available based on hydrogen adsorption mechanism on metal doped macrocyclic organic
129
+ molecules and other Li decorated nanostructures, our work is the first to reveal the efficiency of
130
+ Li functionalized PCP222 using the atomistic MD simulation, practical storage capacity and
131
+ diffusion energy barrier estimation
132
+ 2 Theory and Computation
133
+ The theoretical computations are carried out on [2.2.2] paracyclophane (PCP222) and their
134
+ hydrogenated derivatives within the framework of density functional theory (DFT)[49]. The
135
+ modern range separated hybrid functional wB97Xd is used, and molecular orbitals (MO) are
136
+ defined as linear combination of atom centered basis functions, with all atoms using the valence
137
+ diffuse and polarization function 6-311+G(d,p) basis sets. The wB97Xd, a long range separated
138
+ form of Becke’s 97 functional, also adds Grimme’s D2 dispersion correction[50, 51]. It is worth
139
+ mentioning that the wB97Xd is a reliable approach to investigate the non-covalent interaction of
140
+ metal doped organic molecules and their thermochemistry. The harmonic frequencies of all the
141
+ studied structures are calculated to confirm that they are truly in the ground state on the potential
142
+ surface.
143
+ Some of the crucial quantitative metrics, including, binding energy of metal atom on host,
144
+ average H2 adsorption energy and successive H2 desorption energy must be determined in order to
145
+ analyze the mechanism of hydrogen storage.
146
+ The binding strength of Li atom on the PCP222 is calculated by the following expression[44];
147
+ 𝐸𝑏 =
148
+ 1
149
+ 𝑚 [𝐸𝑃𝐶𝑃222 + 𝑚𝐸𝐿𝑖 − 𝐸𝑃𝐶𝑃222+𝑚𝐿𝑖]
150
+
151
+ (1)
152
+ Where 𝐸𝑃𝐶𝑃222, 𝐸𝐿𝑖, and 𝐸𝑃𝐶𝑃222+𝑚𝐿𝑖 are symbolize for the total energy of PCP222, energy of
153
+ single isolated Li atom and energy of Li-decorated PCP222 respectively. m denotes for the number
154
+ of Li atoms used to functionalized the PCP222.
155
+ The average adsorption energy of H2 molecules with Li functionalized PCP222 is estimated
156
+ as [52];
157
+ 𝐸𝑎𝑑𝑠 =
158
+ 1
159
+ 𝑛 [𝐸𝑃𝐶𝑃222+𝑚𝐿𝑖 + 𝑛𝐸𝐻2 − 𝐸𝑃𝐶𝑃222+𝑚𝐿𝑖+𝑛𝐻2]
160
+ (2)
161
+
162
+ Where, EH2, and EPCP222+mLi+nH2 represents the energy of isolated single H2 molecule and
163
+ hydrogen adsorbed PCP222+mLi, respectively. n is the number of H2 molecules adsorbed in each
164
+ Li functionalized PCP222.
165
+ The successive desorption energy of adsorbed H2 molecules is estimated using following
166
+ equation[52].
167
+ 𝐸𝑑𝑒𝑠 =
168
+ 1
169
+ 𝑛 [𝐸𝐻2 + 𝐸𝐻𝑜𝑠𝑡+(𝑛−1)𝐻2 − 𝐸𝐻𝑜𝑠𝑡+𝑛𝐻2]
170
+
171
+ (3)
172
+ where 𝐸𝐻𝑜𝑠𝑡+(𝑛−1)𝐻2is the energy of previous H2 molecules adsorbed 𝐸𝐻𝑜𝑠𝑡+𝑛𝐻2.
173
+ The energy gap between the highest occupied molecular orbital (HOMO) and the lowest
174
+ unoccupied molecular orbital (LUMO) is calculated to ensure the kinetic stability of the Li
175
+ functionalized PCP222 and their hydrogen derivatives. The Hirshfeld charges and electrostatic
176
+ potential map (ESP) was used to study electronic charge transfer and interaction mechanism.
177
+ Further, to understand the metal and hydrogen interaction we have performed the partial density
178
+ of states (PDOS), and topological using the Bader’s quantum theory of atoms in molecules
179
+ (QTAIM). To investigate the structural integrity of the host material and H2 reversibility of the
180
+ system, atomistic molecular dynamic simulations were carried out using the expanded lagrangian
181
+ approach, atom-centered density matrix propagation (ADMP).
182
+ To determine the H2 adsorption capacity, gravimetric density (wt%) of hydrogen can be calculated
183
+ using the following expression [53]:
184
+ 𝐻2(𝑤𝑡%) =
185
+ 𝑀𝐻2
186
+ 𝑀𝐻2+𝑀𝐻𝑜𝑠𝑡 × 100
187
+
188
+
189
+
190
+ (4)
191
+ Here MH2 represent the mass of the total number of H2molecules adsorbed and MHost represent the
192
+ mass of Li functionalized PCP222.
193
+ 3 Results and Discussion
194
+ 3.1 Structural properties of PCP222
195
+ Figure 1 depicts the ground state geometrical structure of PCP222. The PCP222 comprises three
196
+ benzene rings, that are linked via two CH2 moiety as bridge between the adjacent rings. The lengths
197
+ of the nearest CH2-CH2, and the CH2 across the benzene rings are observed to be 1.5 and 5.84 Å ,
198
+
199
+ respectively, that agrees with the empirically reported value by Cohen-Addad et al [54]. To
200
+ confirm the aromaticity of the relaxed PCP222, we calculated the Nucleus Independent Chemical
201
+ Shift (NICS) from center to to 3 Å above the benzene ring by increment of 1 Å. The NICS(1) is
202
+ found to have negative maximum (-10.1 ppm), demonstrating the aromatic character of
203
+ PCP222[55, 56]. This suggest that the cyclic rings of PCP222 are -electron rich and most
204
+ probably can bind the metal atom above (outside of PC222) the benzene rings. The Li atom then
205
+ functionalized above the benzene rings and on every possible site of PCP222 and allowed to relax
206
+ as discussed below.
207
+
208
+ Figure 1: (a) Optimized structure of PCP222 with adsorption site marked with red-colored
209
+ text, (b) Li functionalized PCP222.
210
+ 3.2 Functionalization of Li atom on PCP222
211
+ To explore the hydrogen adsorption capacity in Li-functionalized PCP222, we must first carefully
212
+ examine the suitable adsorption site for Li atoms on the PCP222. In order to do this, we
213
+ investigated several PCP222 adsorption site, including the C-C bridge of benzene ring (B1),
214
+ CH2 moiety and benzene bridge (B2), CH2 - CH2 bridge (B3), and above the center of benzene
215
+ (Rc). All the possible Li adsorption sites of PCP222 are depicted in Figure 1(a). A single Li atom
216
+ is placed nearly 2 Å above the several probable adsorption sites of PCP222 and the structure is
217
+ allowed to get optimized. It is observed that functionalization of Li atom over B1 and B2 sites, it
218
+ migrate towards the Rc site following the optimization. On optimization of Li atom over B3 site,
219
+ the it moves away from the PCP222 and does not bind to the surface. We found that the Li atom
220
+ is stable on Rc site with binding energy of 0.32 eV that is 0.1 eV higher than that of Li on PCP44,
221
+ reported by Sathe et al. [43]. The Li atom is supposed to be functionalized on PCP222 via Dewar
222
+ mechanism, in which is due to the electronic charge transfer between the p-complex and s- orbitals
223
+
224
+ 5.84
225
+ 5.858
226
+ B1
227
+ 1.541
228
+ 1.543
229
+ (a)
230
+ (b)of Li atom [43, 45]. After functionalization of Li, the estimated Hirshfeld charge on benzene ring
231
+ of PCP222 is increased to -0.08 e.u from -0.03 e.u (in bare PCP222). These charges are transferred
232
+ from the metal atom, with the Hirshfeld charges on Li atom being +0.35 e.u after functionalization,
233
+ which make the Li atom ionic. The ionic Li atom is exposed to the guest H2 molecules and can
234
+ bind them via charge polarization mechanism as proposed by Niu et al. [57]. No significant change
235
+ in geometrical bond distances is observed after the functionalization of Li. The thermal stability
236
+ of the structures (host) is discussed in the molecular dynamic simulations section (section 3.5). All
237
+ the hydrogen adsorption/desorption simulations are performed by functionalizing the Li atom
238
+ above the center of benzene ring of PCP222.
239
+ 3.2.1 Diffusion energy barrier calculation
240
+
241
+ Figure 2: Diffusion energy barrier plot between energy difference and diffusion coordinates of
242
+ Li atom on PCP222
243
+
244
+ The clustering of metal atoms on the substrate can reduce the hydrogen uptake capacity of the
245
+ system as reported earlier [17]. The barrier of metal atoms diffusion energy ultimately decides
246
+ whether or not the clustering will occur. With a small rise in temperature, if the Li atom migrated
247
+ from its adsorption location, the possibility of metal-metal clustering would increase. Since, the
248
+ binding energy of Li atom on the PCP222 is less than the cohesive energy of the isolated Li atom
249
+ (1.63eV), we calculate if there is an energy barrier for diffusion of Li atom on PCP222 that can
250
+ avoid the possibility of metal clustering. To calculate the energy barrier, we shift the Li atom over
251
+ its adsorption site (on the benzene ring) by a small distance along the path shown in the
252
+ Figure 2 and carried out the single point energy calculation. Then we exhibit the energy difference
253
+ between initial and current step energy with the diffusion coordinate as illustrated in Figure 2. The
254
+
255
+ 1.4-
256
+ AE=1.38eV
257
+ 1.2-
258
+ 1.0-
259
+ 0.8-
260
+ ev
261
+ 0.6-
262
+ 4
263
+ 1-3)
264
+ 0.4
265
+ 0.2 -
266
+ 0.0 -
267
+ 1
268
+ -
269
+ 0
270
+ 1
271
+ 2
272
+ 3
273
+ 4
274
+ 5
275
+ Diffusion coordinatesfigure shows presence of an energy barrier of 1.38 eV, that is sufficient to stop the Li atom from
276
+ diffusing across the PCP222 and thus prevent the metal clustering. Therefore, our calculated
277
+ energy barrier for diffusion of Li atom is high enough to prevent metal clustering over the studided
278
+ PCP222 compound.
279
+ 3.3 Interaction of H2 with PCP222-Li
280
+ 3.3.1 Adsorption Energy
281
+
282
+ Table 1: Average bond distance between carbon bridge (C-C), center of PCP222 benzene ring (Rc)
283
+ and Lithium atom (Rc-Li), Lithium and hydrogen molecules (Li-H2), and hydrogen Hydrogen
284
+ (H-H) in Å. Average adsorption energy and successive desorption energy of PCP222-Li-
285
+ nH2 (n=1-5)
286
+
287
+ Name of complex
288
+ Bridge C-C Rc-Li
289
+ Li-H
290
+ H-H
291
+ Eads (eV) Edes (eV)
292
+ PCP222-Li
293
+ 1.542
294
+ 1.735
295
+ -
296
+ -
297
+ -
298
+ -
299
+ PCP222-Li-1H2
300
+ 1.542
301
+ 1.745
302
+ 2.124
303
+ 0.753
304
+ 0.171
305
+ 0.171
306
+ PCP222-Li-2H2
307
+ 1.541
308
+ 1.742
309
+ 2.083
310
+ 0.757
311
+ 0.159
312
+ 0.147
313
+ PCP222-Li-3H2
314
+ 1.541
315
+ 1.767
316
+ 2.159
317
+ 0.753
318
+ 0.148
319
+ 0.127
320
+ PCP222-Li-4H2
321
+ 1.541
322
+ 1.811
323
+ 2.243
324
+ 0.752
325
+ 0.134
326
+ 0.089
327
+ PCP222-Li-5H2
328
+ 1.541
329
+ 1.813
330
+ 2.478
331
+ 0.751
332
+ 0.113
333
+ 0.030
334
+ To explore the storage capacity and characteristics of Li functionalized PCP222, we introduced
335
+ the H2 molecules in a sequential manner to PCP222-Li. Firstly we introduced a single H2 molecule
336
+ at around 2Å above the Li atom on PCP222 and allowed the structure to get relaxed. It is observed
337
+ that, the H2 molecule is adsorbed at a distance of 2.124 Å from the Li atom with an adsorption
338
+ energy of 0.171 eV and the H-H bond length elongated by 0.01 Å. Sathe et al. studied the hydrogen
339
+ storage capacity of Li functionalized PCP11 (PCP22) and reported the adsorption energy of first
340
+ H2 molecule ~0.13 eV (0.11 eV) [46, 45]. Our calculated adsorption energy is slightly higher,
341
+ which is important in alkali metal doped H2 storage material and leads to the increase in the
342
+ desorption temperature. Further, we optimized the structures by adding H2 molecules sequentially
343
+ onto the PCP222-Li. On addition of second H2 molecule to the system, the average H2 adsorption
344
+ energy calculated to be 0.159 eV/H2. In this way, adsorption of 3rd, 4th and 5th H2 molecules to
345
+
346
+ PCP222-Li, the average H2 adsorption energy reduces to 0.148, 0.134 and 0.113
347
+ eV/H2respectively. When of more than five H2 molecules are added to the system, they fly away
348
+ from the Li atom and adsorption energy fall below 0.1 eV. We observed that the average adsorption
349
+ energy decreases with increase in number of H2 molecules in the system which is due the steric
350
+ hindrance between the adsorbed H2 crowed around the sorption centers and the increase in Li-
351
+ H2 distances (Table 1). The estimated data of adsorption energy and geometrical parameters of all
352
+ the bare hydrogenated systems and presented in Table 1.
353
+
354
+ Figure 3: Optimized geometry of hydrogenated Li functionalized PCP222, (a) PCP222-Li-
355
+ 1H2, (b) PCP222-Li-2H2, (c) PCP222-Li-3H2, (d) PCP222-Li-4H2, (e) PCP222-Li-5H2.
356
+ 3.3.2 Electrostatics potential and Hirshfeld charges
357
+ To get a qualitative picture of electronic charge distribution over the surface of Li functionalized
358
+ PCP222 and their hydrogen adsorbed systems during the hydrogen adsorbed, we generate and
359
+ plotted the electrostatic potential map (ESP map) on the total electron density as depicted in
360
+ Figure 4. The electronic charge distribution is used to identify the active adsorption site, where the
361
+ hydrogen molecules can be introduced. The red and blue regions in the ESP plot reflects the
362
+ aggregation and reduction of electronic charge density respectively. The variation in the charge
363
+ density is plotted with the sequence of color code as red (highest electron density)> orange >
364
+ yellow > green > blue (lowest electron density). The ESP map of PCP222-Li shows that the Li
365
+ atom has the deficiency of electronic charges as marked by the dark blue region over the Li atom,
366
+ this indicate that the Li atom is somewhat ionic and is prone to bind the guest H2 molecules. When
367
+ the first H2 molecule added to the Li atom, the colour of the region over the Li changes from dark
368
+
369
+ (a)
370
+ b
371
+ C
372
+ (d)
373
+ eblue to light blue, demonstrating the charge transfer from C atom of PCP222 and adsorbed H2 to
374
+ the Li atom. Further sequential adsorption of H2 molecules to PCP222-Li changes the colour of Li
375
+ region from blue to light blue indicating additional charge transfer. The blue region over Li almost
376
+ disappears on the adsorption of 5th H2 molecules suggesting the saturation of hydrogen uptake and
377
+ more guest H2 are unlikely to be adsorbed. The exact charge transfer is determined by calculating
378
+ the hirshfeld charges as discussed below.
379
+
380
+
381
+ Figure 4: Electrostatics potential map of (a) PCP222-Li, (b) PCP222-Li-1H2, (c) PCP222-Li-
382
+ 2H2, (d) PCP222-Li-3H2, (e) PCP222-Li-4H2, (f) PCP222-Li-5H2.
383
+ We have performed the hirshfeld charge analysis to quantify the charge transfer distributions on
384
+ the Li functionalized PCP222 and their H2 adsorbed systems. The computed average Hirshfeld
385
+ charges on C atoms of benzene ring (Li functionalized site), Li atom, and adsorbed H2 molecules
386
+ with the number of hydrogen molecules is depicted in Figure 5. The average charges on C atom
387
+ of benzene ring is noted to be -0.031 e which raises to -0.084 e with the functionalization of Li
388
+ atom. The charge on Li atom of PCP222-Li is noted to be +0.354 e, which illustrate the transfer of
389
+ charges from benzene ring to Li atom making the sorption center (Li) ionic and more suitable for
390
+ H2 adsorption. These results agree well with the aforesaid ESP analysis. On adsorption of the first
391
+ H2 molecule to PCP222-Li, the charge on C atom is reduced by 2.38% and at the same time the
392
+ charge on Li atom is increased by 16.7 %. Further addition of hydrogen molecules follows the
393
+ trend of decrease in charge on benzene ring and increase in charge on Li atom (Figure 5). These
394
+ observations suggest that, the ionic Li atom polarize the guest H2 molecules and the H2 molecules
395
+ are adsorbed to the sorption center via a charge polarization mechanism due to induced dipole
396
+ developed in H2 as suggested by the Neu-Rao-Jena [30]. It is noted that the electronic charge on
397
+
398
+ - 4.000 e-2
399
+ + 4.000 e-2
400
+ (a)
401
+ (b)
402
+ (c)
403
+ (d)
404
+ (e)
405
+ (f)
406
+ Sideview
407
+ Top viewLi atom is raised by 41.36 % after the adsorption of the 5th H2 molecule. The adsorbed
408
+ H2 molecules are found to have an average charge of 0.027e to 0.013 e.
409
+
410
+ Figure 5: Hirshfeld charges before and after hydrogen adsorption on PCP222-Li
411
+ 3.3.3 Bader’s topological analysis and PDOS
412
+ The nature of interaction between the Li functionalized PCP222 and the adsorbed hydrogen
413
+ molecules is analyzed using the topological Bader’s quantum theory of atoms in molecules
414
+ (QTAIM). The parameters of electron density distribution at the bond critical point (BCP),
415
+ including the electron density (BCP), total electron energy density (ℋBCP), and Laplacian (2BCP),
416
+ are computed and given in Table S1 (in Supporting Information). The electron density (r) on C-C,
417
+ and C-Li, of hydrogenated PCP222-Li estimated to be almost equal to that of bare host material,
418
+ suggesting the post-adsorption chemical stability of the material. Additionally, the
419
+ average BCP values on H-H in PCP222-Li-5H2 is 0.258 a.u which is same as that on isolated bare
420
+ H2 molecules (-0.263). This reveal that the adsorbed hydrogens are in molecular form during the
421
+ adsorption. According to Kumar et al., the positive value of 2BCP indicated an electron density
422
+ depletion in the region of bonding and implied a close-shell kind of interaction. We noticed there
423
+ is no BCP between the Li and H atoms which implies no chemical bond between the Li atom and
424
+ the adsorbed H2 molecules and the interaction is purely closed-shell type resulting from the charge
425
+ polarization as proposed by the Neu-Rao-Jena.
426
+ Figure 6 illustrate the density of state plot of Li and adsorbed H atoms of the hydrogenated
427
+ PCP222-Li including the first and last (5th) H2 molecules adsorbed on the system. When one
428
+ hydrogen molecule is bound to the sorption center (Li), the s-orbital of the H2 molecule appears
429
+ below the Fermi level (E = 0) and stays unaffected as in the case of bare H2in Figure S2. This
430
+
431
+ 0.6
432
+ Ring CbeforeLi decoration
433
+ 0.5
434
+ Ring C after Li decoration
435
+ -Liatom
436
+ 0.4-
437
+ -Hatom
438
+ Hirshfeld Charges (eu)
439
+ 0.3
440
+ 0.2
441
+ 0.1 -
442
+ 0.0
443
+ 0.1
444
+ -0.2
445
+ -0.3
446
+ 0.4
447
+ -
448
+ *
449
+ *
450
+ 0
451
+ 1
452
+ 2
453
+ 3
454
+ 4
455
+ 5
456
+ Number of H, molecules, nsignifies that there is no hybridization between the Li and adsorbed H2. This implies that the
457
+ adsorption of H2 molecule is owing to the induced dipole produced by charge polarization in H2.
458
+ With the adsorption of 5H2 molecules on PCP222-Li, the orbital of H atom splits into multiple
459
+ peaks ranging from -16 eV to -4 eV. This implies that the adsorption weakens as the quantity of
460
+ H2 molecules increases in the host.
461
+
462
+ Figure 6: Partial density of state on Li and H atoms of PCP222-Li-1H2 and PCP222-Li-5H2
463
+ 3.4 Thermodynamics and storage capacity
464
+ 3.4.1 Storage Capacity
465
+
466
+ Figure 7: Optimized geometry of (a) PCP222-3Li, (b) PCP222-3Li-3H2, (c) PCP222-3Li-6H2,
467
+ (d) PCP222-3Li-9H2, (e) PCP222-3Li-12H2, (f) PCP222-3Li-15H2.
468
+
469
+
470
+ 3.0
471
+ Li
472
+ PCP222-Li-1H2
473
+ 2.5
474
+ H
475
+ 2.0
476
+ 1.5
477
+ 1.0
478
+ HOMO
479
+ LUMO
480
+ -7.97eV
481
+ 0.51eV
482
+ W
483
+ 0.5
484
+ 0.0
485
+ -18
486
+ 16
487
+ 14
488
+ -12
489
+ -10
490
+ -8
491
+ -6
492
+ -4
493
+ .
494
+ -2
495
+ 0
496
+ 2
497
+ 4
498
+ Energy (ev)
499
+ (a)
500
+ 4.0
501
+ Li
502
+ PCP222-Li-5H,
503
+ 3.5
504
+ H
505
+ 3.0
506
+ 2.5
507
+ PDOS
508
+ 2.0
509
+ 1.5
510
+ HOMO
511
+ LUMO
512
+ 1.0
513
+ -7.91eV
514
+ 0.45eV
515
+ 0.5
516
+ 0.0
517
+ -18
518
+ 16
519
+ 14
520
+ 12
521
+ 10
522
+ -8
523
+ -6
524
+ -4
525
+ T
526
+ -2
527
+ 0
528
+ 2
529
+ 4
530
+ Energy (ev)
531
+ (b)+3H2
532
+ (a)
533
+ (b)
534
+ 15H2
535
+ +3H2
536
+ + 3H,To investigate the optimum hydrogen storage capacity of the studied system, we functionalized
537
+ the maximum possible number of Li atoms over each benzene ring of PCP222. The geometrical
538
+ structure of three Li functionalized PCP222 ( PCP222-3Li) is shown in Figure 7 Further, we
539
+ introduced H2 molecules to each Li atom of PCP222-3Li sequentially as discussed in previous
540
+ section (3.3.1). The computed average hydrogen adsorption energy and the geometrical parameters
541
+ of all the hydrogenated systems are provided in the Table S2 (in Supporting Information). It is
542
+ noticed that, the adsorption process of hydrogen molecules on PCP22-3Li is found similar to that
543
+ of on PCP222-Li. On saturation of H2 adsorption on PCP222-3Li, we found each Li atom can
544
+ adsorb a maximum of 5H2 molecules resulting in total gravimetric density of 8.32 wt%. The
545
+ estimated value of hydrogen storage capacity is fairly above the requirement of US-DOE for
546
+ effective hydrogen storage systems. Our results can be compared with earlier reported
547
+ H2 gravimetric density on metal decorated carbon-based materials for hydrogen storage, for
548
+ example, Li-decorated C41 allotrope (7.12 wt%) [58], Li doped MOF impregnated with Li-coated
549
+ fullerenes[59], Li-doped B4C3 monolayer (6.22 wt%) [4].
550
+ To develop a realistically usable hydrogen storage system, a significant quantity of hydrogen
551
+ molecules must be adsorbed by the host material under achievable storage conditions. Further the
552
+ adsorbed hydrogen molecules must also be efficiently desorbed at suitable temperature (T) and
553
+ pressure (P). Thus, we estimated the quantity of adsorbed hydrogen that could be used at a
554
+ accessible range of temperature (T) and pressure (P). To calculate the number of H2 molecules
555
+ remain adsorbed on PCP222-3Li (Occupation number) at different T and P, we calculated the
556
+ empirical value of hydrogen gas chemical potential (µ). Then the occupation number (N) is
557
+ estimated by the following expression and plotted with various T and P in Figure 8(b)[60].
558
+ 𝑁 =
559
+
560
+ 𝑛𝑔𝑛𝑒[𝑛(𝜇−𝐸𝑎𝑑𝑠)/𝐾𝐵𝑇]
561
+ 𝑁𝑚𝑎𝑥
562
+ 𝑛=0
563
+
564
+ 𝑔𝑛𝑒[𝑛(𝜇−𝐸𝑎𝑑𝑠)/𝐾𝐵𝑇]
565
+ 𝑛𝑚𝑎𝑥
566
+ 𝑛=0
567
+
568
+
569
+
570
+
571
+ (5)
572
+ Here Nmax is the maximum number of H2 molecules adsorbed at each Li atom on
573
+ PCP222, n and gn represents the number of H2 molecules adsorbed and configurational
574
+ degeneracy for a n respectively. kB is the Boltzmann constant and -Eads indicates the average
575
+ adsorption energy of H2 molecules to PCP222-3Li. m is the empirical value of chemical potential
576
+ of hydrogen gas at specific T and P, and is obtained by using the following expression [61].
577
+ 𝜇 = 𝐻0(𝑇) − 𝐻0(0) − 𝑇𝑆0(𝑇) + 𝐾𝐵𝑇 ln (
578
+ 𝑃
579
+ 𝑃0)
580
+
581
+ (6)
582
+
583
+ Here H0(T), S0(T) are the enthalpy and entropy of H2 at pressure P0 (1 bar).
584
+ We can see in Figure 8(b) that the PCP222-3Li can adsorbed H2 molecules giving rise to
585
+ maximum hydrogen uptake capacity of ~8.32 wt% up to the temperature of 80 K and pressure of
586
+ 30-60 bar. When the temperature rises beyond 80 K, the H2 molecules begin to desorb from the
587
+ PCP222-3Li and the gravimetric density closes to ~5.5 wt% (target of US-DOE by 2025) when
588
+ the temperature reaches 180 K under the pressure of 30-60 bar. Further rise in temperature, the
589
+ storage capacity of the PCP222-3Li fall below 4 wt% at 220 K and 40-bar. At a temperature range
590
+ of 260 K-300 K and pressure range of 1-10 bar, the studied system shows a 100% desorption of
591
+ hydrogen. Thus, we can propose the Li functionalized PCP222 as a low-temperature-adsorption
592
+ and room-temperature-desorption hydrogen storage material. Under the room temperature (300
593
+ K), the studied system shows up to 8.32 wt % of usable hydrogen storage capacity with 100%
594
+ reversibility. Thus, we believe that, our studied material Li functionalized PCP222 can be used as
595
+ an efficient hydrogen storage material satisfying the criteria of US-DOE.
596
+
597
+ Figure 8: Plot of Van’t-Hoff desorption temperature for Li functionalized PCP222 at different
598
+ temperature and pressure.
599
+ 3.4.2 Desorption temperature
600
+ For a reversible hydrogen storage media, it is crucial to estimate the desorption temperature of
601
+ hydrogen molecules. We have estimated the desorption temperature (T D) of H2 for the Li
602
+ functionalized PCP222 using the Van’t Hoff equation [17].
603
+ 𝑇𝐷 = (
604
+ 𝐸𝑎𝑑𝑠
605
+ 𝐾𝐵 ) (
606
+ ∆𝑆
607
+ 𝑅 − ln 𝑝)
608
+ −1
609
+
610
+
611
+
612
+
613
+
614
+ (7)
615
+
616
+ 280
617
+ 81
618
+ 7,488
619
+ 6.656
620
+ 7-
621
+ 260-
622
+ 5.824
623
+ 6-
624
+ 4.992
625
+ Desorptiontemperature
626
+ 240
627
+ 219K
628
+ 5-
629
+ 4.160
630
+ 220
631
+ G,wt%
632
+ 2.496
633
+ 200
634
+ 3
635
+ 1.664
636
+ 182K
637
+ 2-
638
+ 180
639
+ 0.000
640
+ 1
641
+ 160
642
+ 145K
643
+ averageT,
644
+ 140-
645
+ 一minT,
646
+ 1.5
647
+ 2.0
648
+ 2.5
649
+ 3.0
650
+ 1.0
651
+ 3.5
652
+ 4.5
653
+ 5.0
654
+ Pressure(atm)
655
+ 300
656
+ (a)
657
+ (b)Where, Eads represents the computed hydrogen adsorption energy, KB, and R denotes for the
658
+ Boltzmann constant and R the gas constant respectively. P represent s the equilibrium pressure
659
+ (we take a range of 1 to 5 atm with an increment of 0.5 atm) and △S is the entropy change of
660
+ hydrogen from its gaseous state to liquid state [62]. Using the highest and lowest adsorption energy
661
+ of system (with minimum and maximum H2 gravimetric density, respectively), the maximum and
662
+ minimum desorption temperatures (TDmax∕TDmin) are determined. While, TDmin denotes the
663
+ minimum temperature necessary to initiate the desorption of H2 molecules, the TDmax is the
664
+ temperature required for complete desorption process. The estimated desorption temperatures
665
+ along with the equilibrium pressure is depicted in Figure 8(a). The minimum and maximum
666
+ temperatures for H2 desorption are determined to be 145 K and 219 K, respectively, at 1 atm
667
+ pressure. The estimated average TD of Li functionalized PCP222 is 182 K at 1 atm. This result
668
+ reveals that, the system can adsorb its full capacity H2 at cryogenic temperature and desorb all the
669
+ H2 molecules bellow room temperature at 1 atm pressure. However, the desorption temperature
670
+ can be increases by increase in the equilibrium pressure as presented in Figure 8(a) and as
671
+ discussed above.
672
+ 3.5 Molecular dynamics simulations
673
+
674
+ Figure 9: (a) Potential energy trajectories of hydrogenated PCP222-3Li and (b) Time
675
+ evolution trajectory of average bond length between the Li atom and C atoms of PCP222 at
676
+ 300K and 500K,
677
+
678
+ 968.88
679
+ 968.90
680
+ 300K
681
+ (Hartree
682
+ 500K
683
+ 968.92
684
+ 968.94
685
+ energy
686
+ 968.96
687
+ 968.98
688
+ -969.00
689
+ Potential
690
+ -969.02
691
+ 969.04
692
+ -969.06
693
+ -969.08
694
+ 0
695
+ 100
696
+ 200
697
+ 300
698
+ 400
699
+ 500
700
+ 600
701
+ 700
702
+ 800
703
+ 900
704
+ 1000
705
+ Time (fs)
706
+ (a)
707
+ 2.6
708
+ C-Lidistance@300K
709
+ Average C-Li distance (A)
710
+ 2.5
711
+ C-Lidistance@50oK
712
+ 2.4
713
+ 2.3
714
+ 2.2
715
+ 2.1
716
+ 2.0
717
+ 1.9
718
+ 1.8
719
+ 1.7
720
+ 1.6
721
+ 300K,1000fs
722
+ 500K,1000fs
723
+ 1.5
724
+ 0
725
+ 100
726
+ 200
727
+ 300
728
+ 400
729
+ 500
730
+ 600
731
+ 700
732
+ 800
733
+ 006
734
+ 1000
735
+ Time (fs)
736
+ (b)
737
+ To validate the reversibility of hydrogen molecules on PCP222-3Li estimated by the DFT
738
+ computation, we have carried out molecular dynamics (MD) simulations using the atomistic
739
+ density matrix propagation (ADMP). ADMP is an extended Lagrangian approach to MD, that uses
740
+ the gaussian basis function and propagates the density matrix. The ADMP-MD simulations is
741
+ performed on system with highest storage capacity (PCP222-3Li-15H2), at two different
742
+ temperatures of 300K and 500 K for total time of 1 ps with the time step of 1fs. During the
743
+ simulations the temperature (kinetic energy thermostat) is maintained by the velocity scaling
744
+ approach and at every 10 fs, time step, the temperature is checked and corrected. The time
745
+ evolution potential energy trajectories and the snapshots are depicted in Figure 9(a) and Figure S3
746
+ (in supporting Information) respectively. The MD simulations at 300 K and 1ps illustrate that
747
+ almost all the H2 molecules fly away from the sorption centers, except 1H2 at each center.
748
+ Simulations at 500 K shows that all the H2 molecules are desorbed from the host material keeping
749
+ the host structure intact. This result suggests that the hydrogen storage in Li functionalized PCP222
750
+ is reversible in process.
751
+ For a viable reversible hydrogen storage material, it is important that the host material must not
752
+ distorted above the hydrogen desorption temperature. To investigate the solidity of host material
753
+ (PCP222-3Li), we performed the MD simulations on the bare host structure (PCP222-3Li) at room
754
+ temperature (300 K) and considerably above the H2 desorption temperature (500K) using ADMP.
755
+ The molecular dynamics simulations are performed for 1 ps with a time step of 1 fs. The time
756
+ evolution trajectory of average distance between Li atom and the carbon atoms of PCP222 benzene
757
+ rings is plotted in Figure 9(b). We noticed that the PCP222-3Li structure stays stable at 500 K and
758
+ almost no change in C-C and C-H bond distance is observed. The trajectory of average bond length
759
+ between the Li atom and C atoms of PCP222 benzene rings seem oscillate but the mean value
760
+ (2.25 Å) and the variation is minimal. This validates the structural integrity of the host material
761
+ above the H2 desorption temperature. Moreover no Li clustering is also noticed after desorption as
762
+ discussed earlier in Section 3.2.1. Thus, we believe that PCP222-3Li can be considered for feasible
763
+ reversible hydrogen storage material.
764
+
765
+
766
+
767
+ 4 Summery and Conclusion
768
+ In this study, we investigated the thermodynamical stability and hydrogen storage capacity of Li
769
+ functionalized [2,2,2]paracyclophane, using the density functional theory. The Li atoms are found
770
+ to bind with the PCP222 via Dewar mechanism and no clustering of Li atoms over PCP222 was
771
+ noticed. Each Li atom on PCP222 could adsorb up to 5H2 molecules via charge polarization
772
+ mechanism with an average H2 adsorption energy in the range of 0.12 - 0.17 eV/H2, indicating
773
+ physisorption type of adsorption. Moreover, the average H-H bond distance got elongated by 0.01
774
+ Å, during the adsorption process, which implied that the adsorbed H2 were in molecular form and
775
+ this fact was also confirmed by the charge distribution analysis. When three Li atoms were
776
+ functionalized on PCP222, the H2 gravimetric capacity of the system was up to 8.32 wt% which
777
+ was fairly above the US-DOE requirements for practical hydrogen applications. During saturation
778
+ of H2 adsorption, the host material displayed no significant change in geometry. The
779
+ thermodynamic usable hydrogen capacity was found up to ~8.32 wt% at the temperature of 80 K
780
+ and pressure of 30-60 bar. On further increase in temperature, up to 180 K under the pressure of
781
+ 30-60 bar, the PCP222-3Li hydrogen uptake capacity approached 5.5wt% which is the target of
782
+ DOE by 2025. At a temperature range of 260 K-300 K and pressure range of 1-10 bar, the PCP222-
783
+ 3Li system showed 100% desorption of H2. Molecular dynamic simulation confirmed that at 300
784
+ K, almost all the H2 molecules flied away except 1H2 at each center. Simulations at 500 K showed
785
+ that all the H2 molecules are desorbed from the host material keeping the structure of the host
786
+ structure intact. Since, there is no experimental works reported on Li functionalized PCP222 for
787
+ hydrogen storage, we hope our computational work will contribute significantly to the research of
788
+ hydrogen storage in macrocyclic compounds and provide supporting reference for the future
789
+ experiments.
790
+ References
791
+ [1] Sachin P. Shet, S. Shanmuga Priya, K. Sudhakar, Muhammad Tahir, A review on current
792
+ trends in potential use of metal-organic framework for hydrogen storage, International
793
+ Journal of Hydrogen Energy, 2021, 46, (21), 11782-11803.
794
+ https://doi.org/10.1016/j.ijhydene.2021.01.020
795
+ [2] Jena, P. Materials for hydrogen storage: past, present, and future. The Journal of Physical
796
+ Chemistry Letters. 2011;2(3):206-211. https://pubs.acs.org/doi/abs/10.1021/jz1015372
797
+
798
+ [3] Das GP, Bhattacharya S. Simulation, modelling and design of hydrogen storage materials.
799
+ Proc Indian Natn Sci Acad. 2015;8: 939—951.
800
+ http://scinet.science.ph/union/Downloads/Vol81_2015_4_Art18_336317.pdf
801
+ [4] Rahimi, R., & Solimannejad, M. (2022). Empowering hydrogen storage performance of
802
+ B4C3 monolayer through decoration with lithium: A DFT study. Surfaces and Interfaces, 29,
803
+ 101723.
804
+ [5] DOE technical system targets for onboard hydrogen storage for light-duty fuel cell vehicles.
805
+ https://www.energy.gov/ eere/fuelcells/doe-technical-targets-onboardhydrogenstorage-
806
+ light-duty-vehicles.
807
+ [6] Hassan, I. A., Ramadan, H. S., Saleh, M. A., Hissel, D. Hydrogen storage technologies for
808
+ stationary and mobile applications: Review, analysis and perspectives. Renewable and
809
+ Sustainable Energy Reviews. 2021; 149:111311.
810
+ https://www.sciencedirect.com/science/article/pii/S1364032121005980
811
+ [7] Von Colbe, J. B., Ares, J. R., Barale, J., Baricco, M., Buckley, C., Capurso, G., Dornheim,
812
+ M. Application of hydrides in hydrogen storage and compression: Achievements, outlook
813
+ and perspectives. international journal of hydrogen energy. 2019;44(15):7780-7808.
814
+ [8] Sakintuna, B., Lamari-Darkrim, F., Hirscher, M. Metal hydride materials for solid hydrogen
815
+ storage: a review. International journal of hydrogen energy. 2007;32(9): 1121-1140.
816
+ https://www.sciencedirect.com/science/article/pii/S0360319906005866.
817
+ [9] Shiraz, H. G., Tavakoli, O. Investigation of graphene-based systems for hydrogen storage.
818
+ Renewable and Sustainable Energy Reviews, 2017;74:104-109.
819
+ https://www.sciencedirect.com/science/article/pii/S136403211730271X
820
+ [10] Nagar, R., Vinayan, B. P., Samantaray, S. S., Ramaprabhu, S. Recent advances in
821
+ hydrogen storage using catalytically and chemically modified graphene nanocomposites.
822
+ Journal of Materials Chemistry A. 2017;5(44):22897-22912.
823
+ https://pubs.rsc.org/en/content/articlehtml/2017/ta/c7ta05068b
824
+ [11] Ma, M., Duan, R., Ouyang, L., Zhu, X., Chen, Z., Peng, C., & Zhu, M. (2017). Hydrogen
825
+ storage and hydrogen generation properties of CaMg2-based alloys. Journal of Alloys and
826
+ Compounds, 691, 929-935. doi: 10.1016/j.jallcom.2016.08.307.
827
+ [12] Edalati, K., Uehiro, R., Ikeda, Y., Li, H. W., Emami, H., Filinchuk, Y., ... & Horita, Z.
828
+ (2018). Design and synthesis of a magnesium alloy for room temperature hydrogen storage.
829
+ Acta Materialia, 149, 88-96.
830
+ [13] Murray, L. J., Dincă, M., Long, J. R. Hydrogen storage in metal—organic frameworks.
831
+ Chemical Society Reviews. 2009;38(5):1294-1314.
832
+ https://pubs.rsc.org/en/content/articlelanding/2009/CS/b802256a.
833
+ [14] Cao, Y., Dhahad, H. A., Zare, S. G., Farouk, N., Anqi, A. E., Issakhov, A., Raise, A.
834
+ Potential application of metal-organic frameworks (MOFs) for hydrogen storage:
835
+ Simulation by artificial intelligent techniques. International Journal of Hydrogen Energy,
836
+ 2021;46(73), 36336-36347. https://doi.org/10.1016/j.ijhydene.2021.08.167
837
+
838
+ [15] Li, Y., & Yang, R. T. (2008). Hydrogen storage in metal-organic and covalent-organic
839
+ frameworks by spillover. AIChE Journal, 54(1), 269-279.
840
+ [16] Gaboardi, M., Amade, N. S., Aramini, M., Milanese, C., Magnani, G., Sanna, S., Pontiroli,
841
+ D. Extending the hydrogen storage limit in fullerene. Carbon. 2017;120:77- 82.
842
+ https://www.sciencedirect.com/science/article/pii/S0008622317304712.
843
+ [17] Mahamiya, V., Shukla, A., Chakraborty, B. Scandium decorated C24 fullerene as high
844
+ capacity reversible hydrogen storage material: Insights from density functional theory
845
+ simulations. Applied Surface Science, 2022, 573, 151389.
846
+ https://doi.org/10.1016/j.apsusc.2021.151389.
847
+ [18] Mohan, M., Sharma, V. K., Kumar, E. A., & Gayathri, V. (2019). Hydrogen storage in
848
+ carbon materials˜ A review. Energy Storage, 1(2), e35.
849
+ [19] Dovgaliuk, I, Safin, D, Tumanov, N, Morelle, F, Moulai, A, Cerný, R, Lodziana, Z,
850
+ Devillers, M & Filinchuk, Y , ’Solid Aluminum Borohydrides for Prospective Hydrogen
851
+ Storage’, ChemSusChem, 2017, vol. 10, no. 23, pp. 4725-4734.
852
+ https://doi.org/10.1002/cssc.201701629
853
+ [20] Sakintuna, B., Lamari-Darkrim, F., & Hirscher, M. . Metal hydride materials for solid
854
+ hydrogen storage: a review. International journal of hydrogen energy, 2007, 32(9), 1121-
855
+ 1140.
856
+ [21] Tavhare, P., Chaudhari, A. Nitrogen substitution effect on hydrogen adsorption properties
857
+ of Tidecorated benzene. Structural Chemistry, 2019, 30(6), 2151-2158.
858
+ https://doi.org/10.1007/s11224- 019-01340-x.
859
+ [22] Suh, M. P., Park, H. J., Prasad, T. K., & Lim, D. W. . Hydrogen storage in metal—organic
860
+ frameworks. Chemical reviews, 2012,112(2), 782-835.
861
+ [23] Niemann, M. U., Srinivasan, S. S., Phani, A. R., Kumar, A., Goswami, D. Y., &
862
+ Stefanakos, E. K.. Nanomaterials for hydrogen storage applications: a review. Journal of
863
+ Nanomaterials, 2008.
864
+ [24] Vatsal J, Balasubramanian K. Functionalized graphene materials for hydrogen storage. J
865
+ Mater Sci 2020;55:1865e903. https://doi.org/10.1007/s10853-019-04150-y.
866
+ [25] Tozzini V, Pellegrini V. Prospects for hydrogen storage in graphene. Phys Chem Chem
867
+ Phys 2013;15:80e9. https:// doi.org/10.1039/c2cp42538f.
868
+ [26] Cheng, H. M., Yang, Q. H., & Liu, C. . Hydrogen storage in carbon nanotubes. Carbon,
869
+ 2001,39(10), 1447-1454. https://doi.org/10.1016/S0008-6223(00)00306-7.
870
+ [27] Jaiswal, A., Sahoo, R. K., Ray, S. S., & Sahu, S. Alkali metals decorated silicon clusters
871
+ (SinMn, n= 6, 10; M= Li, Na) as potential hydrogen storage materials: A DFT study.
872
+ International Journal of Hydrogen Energy, 2022, 47(3), 1775-1789.
873
+ https://doi.org/10.1016/j.ijhydene.2021.10.228.
874
+ [28] Ray, S. S., Sahoo, R. K., Sahu, S. Reversible hydrogen storage capacity of vanadium
875
+ decorated small boron clusters (BnV2, n= 6-10): A dispersion corrected density functional
876
+ study. Computational and Theoretical Chemistry, 2022, 1217, 113899.
877
+ https://doi.org/10.1016/j.comptc.2022.113899.
878
+
879
+ [29] Sahoo, R. K., Ray, S. S., Sahu, S. A first principle study of hydrogen storage in titanium-
880
+ doped small carbon clusters (C2nTin, n= 2—6). Structural Chemistry, 2021, 32(4), 1673-
881
+ 1683. https://doi.org/10.1007/s11224-020-01692-9.
882
+ [30] Niu, J., Rao, B. K., & Jena, P. . Binding of hydrogen molecules by a transition-metal ion.
883
+ Physical review letters, 1992, 68(15), 2277.
884
+ [31] Kubas, G. J. (. Hydrogen activation on organometallic complexes and H2 production,
885
+ utilization, and storage for future energy. Journal of Organometallic Chemistry, 2009,
886
+ 694(17), 2648-2653.
887
+ [32] Sun, Q.; Jena, P.; Wang, Q.; Marquez, M. First-Principles Study of Hydrogen Storage on
888
+ Li12C60. J. Am. Chem. Soc. 2006, 128, 9741- 9745. https://doi.org/10.1021/ja058330c.
889
+ [33] Ren, H., Cui, C., Li, X., Liu, Y. . A DFT study of the hydrogen storage potentials and
890
+ properties of Na-and Li-doped fullerenes. International Journal of Hydrogen Energy, 2017,
891
+ 42(1), 312-321.
892
+ [34] Zacharia, R., Kim, K. Y., Kibria, A. F., Nahm, K. S. . Enhancement of hydrogen storage
893
+ capacity of carbon nanotubes via spill-over from vanadium and palladium nanoparticles.
894
+ Chemical physics letters, 2005,412(4-6), 369-375.
895
+ [35] Sahoo, R. K., Sahu, S . Reversible hydrogen storage capacity of Li and Sc doped novel
896
+ C8N8 cage: Insights from density functional theory. International Journal of Energy
897
+ Research. 2022, doi.org/10.1002/er.8562
898
+ [36] Zhang, Y., & Cheng, X. Hydrogen storage property of alkali and alkaline-earth metal
899
+ atoms decorated C24 fullerene: A DFT study. Chemical Physics, 2018,505, 26-33.
900
+ [37] Sahoo, R. K., Chakraborty, B., Sahu, S. Reversible hydrogen storage on alkali metal (Li
901
+ and Na) decorated C20 fullerene: A density functional study. International Journal of
902
+ Hydrogen Energy, 2021,46(80), 40251-40261.
903
+ [38] Mahamiya, V., Shukla, A., Chakraborty, B. . Ultrahigh reversible hydrogen storage in K
904
+ and Ca decorated 4-6-8 biphenylene sheet. International Journal of Hydrogen Energy. 2022,
905
+ https://doi.org/10.1016/j.ijhydene.2022.01.216.
906
+ [39] Kundu, A., Trivedi, R., Garg, N., Chakraborty, B. (). Novel permeable material “yttrium
907
+ decorated zeolite templated carbon” for hydrogen storage: Perspectives from density
908
+ functional theory. International Journal of Hydrogen Energy. 2022,
909
+ https://doi.org/10.1016/j.ijhydene.2022.06.159.
910
+ [40] Venkataramanan, N. S., Sahara, R., Mizuseki, H., Kawazoe, Y. . Hydrogen adsorption on
911
+ lithium-functionalized calixarenes: a computational study. The Journal of Physical
912
+ Chemistry C, 2008, 112(49), 19676-19679.
913
+ [41] Kumar, S., Dhilip Kumar, T. J. (). Fundamental study of reversible hydrogen storage in
914
+ titanium-and lithium-functionalized calix [4] arene. The Journal of Physical Chemistry C,
915
+ 2017,121(16), 8703-8710.
916
+ [42] Tobe, Y., Ueda, K., Kaneda, T., Kakiuchi, K., Odaira, Y., Kai, Y., Kasai, N. Synthesis and
917
+ molecular structure of (Z)-[6] Paracycloph-3-enes. Journal of the American Chemical
918
+ Society, 1987; 109(4), 1136-1144.
919
+
920
+ [43] Sathe, R. Y., Kumar, S., Kumar, T. J. D. (2018). First-principles study of hydrogen storage
921
+ in metal functionalized [4, 4] paracyclophane. International Journal of Hydrogen Energy,
922
+ 43(11), 5680-5689.
923
+ [44] Sahoo, R. K., Kour, P., Sahu, S.. Reversible hydrogen storage capacity of Sc and Y
924
+ functionalized [1, 1] paracyclophane: Insights from density functional study. Int. J.
925
+ Hydrogen Energy, 47 (2022), 29881-29895. doi.org/10.1016/j.ijhydene.2022.06.294.
926
+ [45] Sathe, R. Y., Kumar, T. D. . Paracyclophane functionalized with Sc and Li for hydrogen
927
+ storage. Chemical Physics Letters, 2018, 692, 253-257.
928
+ [46] Sathe, R. Y., Kumar, T. D. . Reversible hydrogen adsorption in Li functionalized [1, 1]
929
+ paracyclophane. International Journal of Hydrogen Energy, 2020,45(23), 12940-12948.
930
+ [47] Kumar, S., Sathe, R. Y., Kumar, T. D. . First principle study of reversible hydrogen storage
931
+ in Sc grafted Calix [4] arene and Octamethylcalix [4] arene. International Journal of
932
+ Hydrogen Energy, 2019, 44(10), 4889-4896.
933
+ [48] Tabushi, I., Yamada, H., Yoshida, Z., Oda, R. Preparations and properties of tris [2, 2, 2]
934
+ paracyclophane derivatives. Tetrahedron, 1971, 27(19), 4845-4853.
935
+ [49] Frisch MJ, Trucks GW, Schlegel HB, Scuseria GE, Robb MA, Cheeseman JR, et al.
936
+ Gaussian 09, revision E.01. Wallingford CT: Gaussian, Inc; 2013.
937
+ [50] Chai, J. D., Head-Gordon, M. Long-range corrected hybrid density functionals with
938
+ damped atom—atom dispersion corrections. Physical Chemistry Chemical Physics,
939
+ 2008;10(44):6615-6620. https://doi.org/10.1039/B810189B.
940
+ [51] Halsey-Moore, C., Jena, P., McLeskey Jr, J. T. Tuning range-separated DFT functionals
941
+ for modeling the peak absorption of MEH-PPV polymer in various solvents. Computational
942
+ and Theoretical Chemistry, 2019;1162:112506.
943
+ https://doi.org/10.1016/j.comptc.2019.112506.
944
+ [52] Kumar, S., Samolia, M., Dhilip Kumar, T. J. Hydrogen storage in Sc and Li decorated
945
+ metal—inorganic framework. ACS Applied Energy Materials, 2018, 1(3), 1328-1336.
946
+ https://doi.org/10.1021/acsaem.8b00034.
947
+ [53] Surucu, G., Gencer, A., Candan, A., Gullu, H. H., Isik, M. CaXH3 (X= Mn, Fe, Co)
948
+ perovskite-type hydrides for hydrogen storage applications. International Journal of Energy
949
+ Research, 2020, 44(3), 2345-2354. https://doi.org/10.1002/er.5062.
950
+ [54] Cohen-Addad, C., Baret, P., Chautemps, P., & Pierre, J. L. . Structures cristallines du
951
+ [2.2.2] paracyclophane (I)(C24H24) et de son complexe avec le perchlorate d’argent
952
+ (II)(C24H24. AgClO4). Acta Crystallographica Section C: Crystal Structure
953
+ Communications, 1983, 39(10), 1346-1349.
954
+ [55] Grimme, S. On the Importance of Electron Correlation Effects for the p- p Interactions in
955
+ Cyclophanes. Chemistry—A European Journal. 2004;10(14):3423- 3429.
956
+ https://doi.org/10.1002/chem.200400091.
957
+ [56] Schleyer, P. V. R., Maerker, C., Dransfeld, A., Jiao, H., van Eikema Hommes, N. J.
958
+ Nucleus-independent chemical shifts: a simple and efficient aromaticity probe. Journal of
959
+
960
+ the American Chemical Society. 1996;118(26):6317-6318.
961
+ https://doi.org/10.1021/ja960582d.
962
+ [57] Niu, J., Rao, B. K., Jena, P., Manninen, M. . Interaction of H2 and He with metal atoms,
963
+ clusters, and ions. Physical Review B, 1995, 51(7), 4475.
964
+ [58] Yadav, S., Tam, J., Singh, C. V. A first principles study of hydrogen storage on lithium
965
+ decorated two dimensional carbon allotropes. international journal of hydrogen energy,
966
+ 2015, 40(18), 6128-6136. https://doi.org/10.1016/j.ijhydene.2015.03.038
967
+ [59] Rao, D., Lu, R., Xiao, C., Kan, E., Deng, K. Lithium-doped MOF impregnated with
968
+ lithium-coated fullerenes: A hydrogen storage route for high gravimetric and volumetric
969
+ uptakes at ambient temperatures. Chemical Communications, 2011, 47(27), 7698-7700.
970
+ [60] Lee, H., Choi, W. I., Nguyen, M. C., Cha, M. H., Moon, E., Ihm, J. Ab initio study of
971
+ dihydrogen binding in metal-decorated polyacetylene for hydrogen storage. Physical
972
+ Review B, 2007, 76(19), 195110. https://doi.org/10.1103/PhysRevB.76.195110
973
+ [61] Wassmann T., Seitsonen A. P., Saitta A. M., Lazzeri M., Mauri F. Structure, stability, edge
974
+ states, and aromaticity of graphene ribbons. Physical review letters, 2008;101(9), 096402.
975
+ https://doi.org/10.1103/PhysRevLett.101.096402.
976
+ [62] Lide, D. R. (Ed.). (2004). CRC handbook of chemistry and physics (Vol. 85). CRC press.
977
+
978
+
979
+
49E2T4oBgHgl3EQfkAeM/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
59E4T4oBgHgl3EQfcAxC/content/tmp_files/2301.05079v1.pdf.txt ADDED
@@ -0,0 +1,1133 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Deep learning enhanced noise spectroscopy of a spin qubit environment
2
+ Stefano Martina,1, 2, ∗ Santiago Hern´andez-G´omez,3, 2, † Stefano
3
+ Gherardini,4, 2, ‡ Filippo Caruso,1, 2, § and Nicole Fabbri5, 2, ¶
4
+ 1Dipartimento di Fisica e Astronomia, Universit`a di Firenze, I-50019, Sesto Fiorentino, Italy
5
+ 2European Laboratory for Non-linear Spectroscopy (LENS),
6
+ Universit`a di Firenze, I-50019 Sesto Fiorentino, Italy
7
+ 3Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, MA 02139
8
+ 4Istituto Nazionale di Ottica del Consiglio Nazionale delle Ricerche (CNR-INO),
9
+ Area Science Park, Basovizza, I-34149 Trieste, Italy
10
+ 5Istituto Nazionale di Ottica del Consiglio Nazionale delle Ricerche (CNR-INO), I-50019 Sesto Fiorentino, Italy
11
+ (Dated: January 13, 2023)
12
+ The undesired interaction of a quantum system with its environment generally leads to a coherence
13
+ decay of superposition states in time. A precise knowledge of the spectral content of the noise induced
14
+ by the environment is crucial to protect qubit coherence and optimize its employment in quantum
15
+ device applications. We experimentally show that the use of neural networks can highly increase
16
+ the accuracy of noise spectroscopy, by reconstructing the power spectral density that characterizes
17
+ an ensemble of carbon impurities around a nitrogen-vacancy (NV) center in diamond.
18
+ Neural
19
+ networks are trained over spin coherence functions of the NV center subjected to different Carr-
20
+ Purcell sequences, typically used for dynamical decoupling (DD). As a result, we determine that
21
+ deep learning models can be more accurate than standard DD noise-spectroscopy techniques, by
22
+ requiring at the same time a much smaller number of DD sequences.
23
+ I.
24
+ INTRODUCTION
25
+ Quantum sensing combines theoretical results with ex-
26
+ perimental and engineering techniques to carry out infer-
27
+ ence of signals with improved accuracy and/or less com-
28
+ putation time by making use of quantum physics [1, 2].
29
+ A quantum sensor takes advantage of the fragility of
30
+ its quantum properties, such as quantum coherence or
31
+ entanglement, to improve the detection of external per-
32
+ turbations with higher accuracy compared to any classic
33
+ sensor.
34
+ However, this same property implies that the
35
+ quantum sensor is subjected to detrimental noise stem-
36
+ ming from the coupling with its environment. For this
37
+ reason, it is desirable to fully characterize the sensor’s
38
+ environment, either to filter out its detrimental effect, or
39
+ to take it into account when detecting external signals,
40
+ for example, in algorithms using quantum optimal con-
41
+ trol [3–6].
42
+ Neural networks (NN) [7, 8], i.e., algorithmic models
43
+ provided by the interconnection of a group of nodes com-
44
+ monly called neurons, could be a powerful tool to infer
45
+ the sensor’s environment. In this context, deep learning
46
+ has been already proposed theoretically for the classifi-
47
+ cation and detection of quantum noise features [9–11],
48
+ and employed experimentally for the following tasks. (a)
49
+ Estimating the spectra of minuscule amounts of complex
50
+ molecules [12] for nano nuclear magnetic resonance; (b)
51
+ the sensing of magnetic-field strength at room temper-
52
+ ∗ stefano.martina@unifi.it; Equal contribution to this work
53
+ † shergom@mit.edu; Equal contribution to this work
54
+ ‡ stefano.gherardini@ino.cnr.it
55
+ § filippo.caruso@unifi.it
56
+ ¶ fabbri@lens.unifi.it
57
+ ature with high precision [13, 14] by using nitrogen va-
58
+ cancy (NV) centers; (c) performing error mitigation [15]
59
+ and noise learning [16–18]; (d) the tracking of quantum
60
+ trajectories [19]; (e) classification of many-body quantum
61
+ states [20] in superconducting quantum circuits. How-
62
+ ever, to our knowledge, experimental noise spectroscopy
63
+ in single color centers in diamond via deep learning is
64
+ still missing.
65
+ In this paper, we demonstrate that NN can be used
66
+ to process the data obtained by a qubit, operating as
67
+ a quantum sensor, and then reconstruct the noise spec-
68
+ trum that induces dephasing into the qubit itself. In par-
69
+ ticular, we focus on a qubit under dynamical decoupling
70
+ (DD) control sequences [21, 22] in the presence of classical
71
+ random noise with an unknown power density spectrum,
72
+ usually denoted as noise spectral density (NSD). Beyond
73
+ testing numerically our machine learning models, we use
74
+ a single NV center in diamond as a spin qubit sensor and
75
+ we perform a spectroscopic reconstruction of the mag-
76
+ netic noise of its local environment. The latter comprises
77
+ 13C nuclear spins randomly distributed in the diamond
78
+ lattice [23–25] (see Fig. 1). The dephasing affecting the
79
+ qubit sensor is analyzed by applying a set of DD con-
80
+ trol pulses that realize filter functions [21, 22, 26, 27] in
81
+ the frequency domain. The filter functions are designed
82
+ to select specific noise components, without sensing all
83
+ other system-bath interactions. A widely used DD con-
84
+ trol pulse is the Carr-Purcell (CP) sequence [1, 28] that
85
+ is given by N equidistant π pulses, performed between an
86
+ initial and a final π/2 pulse. CP sequences act in the fre-
87
+ quency domain approximately as Dirac comb filters [29];
88
+ hence, they have been used to perform spectroscopy of in-
89
+ tricate signals, e.g., for noise spectroscopy [30, 31]. With
90
+ this protocol, the requirement to achieve high values of
91
+ the noise reconstruction accuracy is to perform sequences
92
+ arXiv:2301.05079v1 [quant-ph] 12 Jan 2023
93
+
94
+ 2
95
+ with a high number of pulses meaning N ∈ [30, 120] (as in
96
+ Ref. [32]) or higher, so that the Dirac comb filter approx-
97
+ imation remains valid (in fact, N determines the filter
98
+ width). This usually leads to long experiments to recon-
99
+ struct the whole spectrum of the noise. Other techniques
100
+ using non-equidistant or even more sophisticated DD se-
101
+ quences [4, 33–36] have proved to be effective for noise
102
+ sensing, but sometimes at the price of a higher computa-
103
+ tional burden.
104
+ For our sensing task, NN are designed to solve a re-
105
+ gression problem, i.e., the reconstruction of the NSD.
106
+ Here, we assume that the NSD of the bath of spins has a
107
+ Gaussian profile [32, 37, 38]. The Gaussian NSD is thus
108
+ parametrized as a function of key parameters, i.e., the
109
+ mean value, variance, offset and noise power that we aim
110
+ to reconstruct. Note that our proposal can be adapted to
111
+ other parametrized NSD functions. The NN are trained
112
+ over a set of synthetic data generated by simulating how
113
+ the coherence of the qubit sensor decays over time under
114
+ the influence of both the CP control pulses and the NSD.
115
+ Moreover, to make the measurement statistics as close as
116
+ possible to the ones obtained from the experiments, extra
117
+ artificial errors are added.
118
+ Our approach using NN entails the following advan-
119
+ tages that we have proven experimentally. (i) NN have
120
+ the capability to predict never-before-seen experimental
121
+ data, and they can work with a better reconstruction
122
+ accuracy (even up to 7 times better, as shown in the
123
+ section Results below) than standard noise spectroscopy,
124
+ as the ´Alvarez-Suter method [31], by making use at the
125
+ same time of DD control sequence with a much smaller
126
+ number of pulses. (ii) The training dataset, which can
127
+ contain both synthetic and experimental data, is gener-
128
+ ated just once and then it can be applied several times,
129
+ as long as the new collected data reproduce the physical
130
+ context under analysis. In connection with (i), we are
131
+ going to show that the amount of data used as input to
132
+ the NN can be smaller than the one needed to resolve the
133
+ NSD by means of standard noise spectroscopy methods.
134
+ From our knowledge, this work is the first experimen-
135
+ tal proof of enhanced reconstruction performance with
136
+ NN for carrying out noise spectroscopy in single color
137
+ centers in diamond. We thus expect that the techniques
138
+ discussed here could fast become a novel standard spec-
139
+ troscopy tool both for such quantum systems and other
140
+ quantum platforms in which regression problems have to
141
+ be solved.
142
+ II.
143
+ RESULTS
144
+ A.
145
+ Generation of training dataset
146
+ The training dataset is composed of synthetic data that
147
+ are originated by simulating the coherence decay of the
148
+ qubit sensor in a noise spectroscopy experiment based on
149
+ DD, as the one depicted in Fig 1. This standard sensing
150
+ procedure, which stems from Ramsey interferometry [1],
151
+ maps information about the quantum coherence of the
152
+ sensor into the population in |0⟩ that is then effectively
153
+ recorded. After having initialized the qubit sensor in the
154
+ ground state |0⟩, a π/2 pulse is applied such that the
155
+ qubit state |ψ⟩ is the superposition (|0⟩+|1⟩)/
156
+
157
+ 2. Then,
158
+ we perform a CP control sequence consisting in a train
159
+ of π pulses that flips repeatedly the qubit, and finally, a
160
+ second π/2 pulse is applied in order to map the phase of
161
+ the qubit into its population. The probability that the
162
+ state of the quantum sensor is |0⟩, which corresponds to
163
+ the observable population, equals to [1, 32]
164
+ P = 1
165
+ 2 (1 + C(τ, N)) ,
166
+ (1)
167
+ where N is the number of π pulses and τ is the time
168
+ between them. The coherence function C(τ, N) is sim-
169
+ ulated numerically, for a set of different values of τ and
170
+ N, to generate the training dataset.
171
+ Let us now introduce the decoherence function that
172
+ quantifies how the quantum coherence C(τ, N) is modi-
173
+ fied under the action of both the external bath of spins
174
+ and a set of CP control pulses.
175
+ The control sequence
176
+ has the effect to modulate the coherence content of the
177
+ qubit sensor, while the interaction with the bath, asso-
178
+ ciated to the NSD S(ω), tends on average to destroy
179
+ such coherence. Overall, under the joint presence of con-
180
+ trol fields and a noise source, the coherence decays as
181
+ C(τ, N) ≡ e−χ(τ,N), where χ(τ, N) denotes the decoher-
182
+ ence function [27, 39–41]:
183
+ χ(τ, N) =
184
+
185
+
186
+ πω2 F(ω, τ, N)S(ω) .
187
+ (2)
188
+ In Eq. (2), the filter function F(ω, τ, N) ≡ |Y (ω, τ, N)|2
189
+ is the square modulus of the Fourier transform of the
190
+ so-called modulation function y(t, τ, N).
191
+ The latter is
192
+ constant piecewise, with values ±1, and switches sign at
193
+ the times t = τ/2, 3τ/2, . . . , (N − 1/2)τ where each π
194
+ pulse is applied [2]. Notice that we are assuming that
195
+ the π pulses are instantaneous, a reasonable assumption
196
+ for our experimental setup where a π pulse duration is
197
+ ∼ 0.1 µs and the time between pulses is τ ∈ [3.3, 6.1] µs.
198
+ Let us now recall the expression, in the frequency domain,
199
+ of the filter function for a CP sequence with even N:
200
+ F(ω, τ, N) = 8 sin2
201
+ �ωτN
202
+ 2
203
+
204
+ sec2 �ωτ
205
+ 2
206
+
207
+ sin4 �ωτ
208
+ 4
209
+
210
+ , (3)
211
+ while for odd N, sin2(ωτN/2) has to be replaced with
212
+ cos2(ωτN/2) [2, 26].
213
+ In order to generate the training dataset, the NSD
214
+ S(ω) is parameterized as
215
+ S(ω) = s0 + A exp
216
+
217
+ −(ω − ωc)2
218
+ 2σ2
219
+
220
+ .
221
+ (4)
222
+ Thus, being a Gaussian distribution, the NSD is fully de-
223
+ scribed by the offset s0, amplitude A, width σ and center
224
+ ωc.
225
+ For the training dataset in the paper, the values
226
+
227
+ 3
228
+ FIG. 1: NV center and Neural Networks for noise spectroscopy. The NV center is surrounded by an
229
+ ensemble of 13C nuclear spins (orange spheres) that collectively induce dephasing to the NV electronic spin (blue
230
+ sphere). The NV electronic spin is controlled with a DD sequence (specifically, a Carr-Purcell (CP) sequence) with
231
+ the aim to measure its dephasing, and therefore characterize the NSD of the nuclear spin bath, i.e., S(ω; s0, A, σ).
232
+ The CP sequence is formed by N equidistant π pulses in between an initial and a final π/2 pulse. The time τ
233
+ between the π pulses determines the measurement total time T = Nτ, given that the time between the first π/2 and
234
+ the train of π pulse and the time between the last π and π/2 pulses are both equal to τ/2. Then, we measure the
235
+ output of this experiment, which is the probability P = 1
236
+ 2(1 + C(t)) that the NV center remains in the initial state
237
+ |0⟩. The spin coherence function C(t) – evaluated at previously-determined times in the set T ∈ {t1, t2, . . . , tn} (the
238
+ tk’s are obtained by changing τ with N fixed) – is the input of the designed Neural Networks (NN). After being
239
+ trained, the NN return the estimation of the NSD parameters.
240
+ of these parameters are taken from the following inter-
241
+ vals: s0 ∈ [4 · 10−4, 4 · 10−3] MHz; A ∈ [0.3, 0.7] MHz;
242
+ σ ∈ [2 · 10−3, 9 · 10−3] MHz.
243
+ Instead, ωc is kept con-
244
+ stant.
245
+ This is because in our experimental setup the
246
+ NSD stems from the interaction with a large ensemble
247
+ of unresolved 13C impurities (nuclear spin bath) around
248
+ the NV electronic spin. Therefore, the center of the NSD
249
+ corresponds to the Larmor frequency ωc = γB, where
250
+ γ = 1.0705 kHz/G is the gyromagnetic ratio of the 13C
251
+ nuclear spins, and B is the amplitude of a static mag-
252
+ netic field aligned with the NV quantization axis, z. Such
253
+ static magnetic field is well known during the experimen-
254
+ tal procedure since it determines the NV electronic spin
255
+ resonances (B = 403.2 ± 2 G).
256
+ The training dataset is generated by uniformly sam-
257
+ pling 104 sets of parameters within the chosen intervals.
258
+ Hence, overall we consider 104 distinct sequences of NSD
259
+ parameters that are used to simulate different coherence
260
+ curves C(τ, N). These sequences are taken in the time
261
+ intervals τ ∈ [3.3, 3.66] µs and [5.5, 6.1] µs with sampling
262
+ time ∆τ = 1 ns (∆τ = 20 ns in the experimental case, see
263
+ below), and for N = {1, 8, 16, 24, 32, 40, 48}. These inter-
264
+ vals are significant for our study because they include the
265
+ values of τ at which the coherence decay curve exhibits
266
+ the first and second order collapses induced on the qubit
267
+ sensor by the bath of 13C impurities [42].
268
+ Finally, in
269
+ order to make the synthetic data used to train the NN
270
+ closer to the experimental setting, extra artificial errors
271
+ sampled by a normal distribution with standard devia-
272
+ tion equal to 0.05 (comparable with the expected error
273
+ in our experimental measurements) are added to every
274
+ point of the generated coherence decay curves. In this
275
+ way, one may mitigate the over-fitting of the employed
276
+ machine learning models that are thus expected to better
277
+ generalize to unseen data. In general, a model trained on
278
+ synthetic data cannot be successfully applied to real data
279
+ without fine tuning it. But in our case, it becomes possi-
280
+ ble, probably due to the fact that the simulated data of
281
+ the coherence decay are quite close to the experimentally
282
+ observed decay data induced by the environment.
283
+ As final remark, notice that, from the 104 simulated
284
+ curves C(τ, N), 6000 are used for the training of the NN
285
+ and 2000 for their validation. Instead, the test step is
286
+ performed either by using the remaining 2000 simulated
287
+ curves, or by using experimental data as described below.
288
+ B.
289
+ Neural networks working principles
290
+ Let us describe the main working features of the NN
291
+ employed in this paper to carry out noise spectroscopy.
292
+ Specifically, we are going to use the multi-layer percep-
293
+ tron (MLP) that is composed of fully-connected layers,
294
+ each of them with a variable number of artificial neurons.
295
+ A single artificial neuron returns as output the scalar
296
+ ˆy ≡ Σ(wT · x + b)
297
+ (5)
298
+ that, by definition, is provided by applying the non-linear
299
+ function Σ : R → R to the weighted sum of the input
300
+ vector x ∈ Rk to which the bias term b ∈ R is added.
301
+ w ∈ Rk denotes the vector of weights. In our analysis,
302
+ the activation function Σ is chosen equal to the rectifier
303
+
304
+ 4
305
+ Σ(x) ≡ max(0, x) [43, 44]. Thus, a MLP layer composed
306
+ of q neurons (each with k inputs) returns the vector
307
+ ˆy ≡ Σ(W T x + b),
308
+ (6)
309
+ where ˆy ∈ Rq, W ∈ Rk×q is the matrix of weights (W
310
+ collects all the weight vectors of the single neurons), and
311
+ b ∈ Rq is the vector of the biases. Hence, a MLP with L
312
+ layers is ruled by the recursion equation
313
+ h[ℓ] ≡ Σ
314
+
315
+ W[ℓ]T h[ℓ − 1] + b[ℓ]
316
+
317
+ ,
318
+ (7)
319
+ where ℓ = 1, . . . , L is the index over the number of layers
320
+ and h[0] ≡ x. In Eq. (7), W[ℓ] and b[ℓ] are, respectively,
321
+ the weights and the biases of the ℓ-th layer. The output
322
+ vector of the MLP is ˆy ≡ h[L]. It is worth noting that
323
+ the number, dimension and activation functions (they
324
+ are usually denoted as the hyperparameters ξ) of the NN
325
+ layers are chosen through a single optimization routine
326
+ (cfr. Methods).
327
+ Let us now introduce the supervised learning process.
328
+ Ideally, the purpose of the latter is to find the parameters
329
+ θ∗ = argminθRD(θ, ξ) that minimize the theoretical risk
330
+ function
331
+ RD(θ, ξ) ≡ E(x,y)∼D [L (ˆy, y)] ,
332
+ (8)
333
+ where θ ≡ {W[1], b[1], . . . , W[L], b[L]}, and ˆy are the
334
+ estimated values of y.
335
+ By definition, RD is the ex-
336
+ pected value of the loss function L for (x, y) sampled
337
+ from the distribution D that generates the dataset [45].
338
+ The loss function L is a differentiable function that mea-
339
+ sures the distance between the prediction ˆy (output of
340
+ the MLP) and the desired output y. However, since one
341
+ can only dispose of a finite set S = {(x, y)1, . . . , (x, y)m}
342
+ of samples to train, validate and test the employed ML
343
+ models, the theoretical risk function is approximated by
344
+ the empirical risk function.
345
+ Considering the partition
346
+ {Str, Sva, Ste} of S in training (Str), validation (Sva) and
347
+ test (Ste) sets, the empirical risk function is defined by:
348
+ RStr(θ, ξ) ≡
349
+ 1
350
+ |Str|
351
+
352
+ (x,y)∈Str
353
+ L (ˆy, y) ,
354
+ (9)
355
+ where |Str| is the cardinality of the training set. In fact,
356
+ RStr is the arithmetic mean of the loss function L eval-
357
+ uated on the samples of the training set Str.
358
+ In our paper, we take the loss function L equal to the
359
+ Mean Squared Error (MSE), also called L2 loss:
360
+ L(ˆy, y) = 1
361
+ q
362
+ q
363
+
364
+ i=1
365
+ (ˆyi − yi)2
366
+ (10)
367
+ for the q outputs of the last layer (in our case three,
368
+ corresponding to the noise parameters s0, A, σ). The
369
+ MLP is trained by minimizing (step-by-step over time)
370
+ the empirical risk function RStr(θ, ξ) with respect to θ
371
+ by means of the mini-batch gradient descent method, so
372
+ as to obtain the optimal value θ∗ of the NN parameters.
373
+ Each gradient descent step is defined by
374
+ θt+1 = θt − η∇θ
375
+ 1
376
+ B
377
+ B
378
+
379
+ b=1
380
+ L(ˆyb,t, yb,t),
381
+ (11)
382
+ where θ0 is a randomly chosen starting point, η is the
383
+ learning rate that defines the length of the step and
384
+ ∇θ 1
385
+ B
386
+ �B
387
+ b=1 L(ˆyt,b, yt,b) is the gradient of the loss func-
388
+ tion. The gradient is calculated for any time t on a batch
389
+ of B elements taken from the training set, and the sub-
390
+ script θ in ∇θ indicates that the variables of L during
391
+ the gradient evaluation are the weights of the NN. In
392
+ this paper, RStr is minimized by means of Adam [46]
393
+ that is a gradient-based optimization algorithm perform-
394
+ ing the adaptive estimation of lower-order moments. The
395
+ minimization is stopped when the time-derivative of the
396
+ risk function evaluated on the validation set RSva(θ∗, ξ)
397
+ becomes positive (early stopping strategy) or after a pre-
398
+ defined number of gradient steps using all the data of the
399
+ training set (called epochs). Then, we use RSva(θ∗, ξ) to
400
+ check if the MLP works also for unseen data and tune
401
+ the hyperparameters ξ (cfr. Methods). Finally, the test
402
+ set Ste is employed to calculate the metrics (discussed in
403
+ detail below) used to generate the figures with the results
404
+ that we are going to illustrate.
405
+ C.
406
+ Training and numerical test of neural networks
407
+ We now show the results obtained by using the trained
408
+ machine learning models to infer the value of the NSD
409
+ parameters {s0, A, σ}. As already mentioned, the NN are
410
+ tested with 2000 different NSD parameters. For each of
411
+ these sets of parameters, the curves C(τ, N) have been
412
+ simulated as described in the previous subsections.
413
+ In order to determine the smallest amount of data re-
414
+ quired to reconstruct the NSD, we perform the training,
415
+ validation and test of the NN with sub-sets of the simu-
416
+ lated curves. These sub-sets are defined by introducing
417
+ the variable N that denotes the upper bound for the num-
418
+ ber of pulses N ≤ N considered during the whole process.
419
+ For example, for N = 16 only the curves C(τ, N) with
420
+ N ∈ {1, 8, 16} are considered. Note that the sub-sets de-
421
+ fined for each value of N contain the curves for all the
422
+ different NSD parameters (6000 for training, 2000 for val-
423
+ idation, and 2000 for testing), and for all the times τ in
424
+ the intervals defined before.
425
+ The results of this analysis are shown in Fig. 2 (orange
426
+ data), where the MSE (the loss function) between the in-
427
+ ferred parameters ( ˆs0, ˆA, ˆσ) and the original parameters
428
+ (s0, A, σ) used to generate the dataset is plotted as a
429
+ function of N. Remarkably, the MSE seems to achieve
430
+ its minimum value after N = 16. This entails that the
431
+ NN do not significantly improve their precision on the
432
+ reconstruction of the NSD by using more data to train
433
+ the NN beyond this point.
434
+
435
+ 5
436
+ 0
437
+ 10
438
+ 20
439
+ 30
440
+ 40
441
+ 50
442
+ ¯N
443
+ 0.0
444
+ 0.1
445
+ 0.2
446
+ 0.3
447
+ MSE(s0, A, σ)
448
+ FIG. 2: Mean-square-errors (MSE) between original
449
+ and estimated NSD parameters for a set of 2000 test
450
+ cases. Orange bullets with dash-dotted line are the
451
+ mean values returned by NN. Blue squares with dotted
452
+ line are the mean values provided by the HS method.
453
+ Finally, shaded areas denote the standard deviation,
454
+ taking into account all the 2000 cases.
455
+ To establish how accurately a NN reconstructs the
456
+ NSD, we need to compare the corresponding results with
457
+ those of a different method. In particular, we concentrate
458
+ on the method used in Ref. [32], which is itself based on
459
+ Refs. [30, 31]. According to them, the decay of the coher-
460
+ ence function C(τ, N) is analyzed as a function of N, for
461
+ each fixed value of τi, i.e., for each fixed frequency com-
462
+ ponent of the filter functions. In the limit of high N, the
463
+ decay of the coherence is exponential, with a rate that is
464
+ inversely proportional to the amplitude of the NSD [30].
465
+ In other words, the amplitude of the NSD is directly es-
466
+ timated for a discrete set of frequencies (each propor-
467
+ tional to 1/τ). In contrast with the original proposals in
468
+ Refs. [30, 31], the method in Ref [32] demonstrates that
469
+ it is better to use the harmonics of the filter functions
470
+ to reconstruct the NSD, in order to avoid extra broad-
471
+ ening of the reconstructed spectrum. For this reason, we
472
+ denote this method as Harmonics Spectroscopy (HS).
473
+ We have analyzed the same 2000 different curves
474
+ C(τ, N) (used to test the machine learning models) also
475
+ with the HS method. The results are collected and shown
476
+ in Fig. 2 (blue data), where the first point is for N = 16.
477
+ This is due to the fact that, by definition, the HS method
478
+ fits the decay of the coherence as a function of N. This
479
+ is possible only for a dataset with at least three points
480
+ (in this case N = 1, 8, 16). As one can observe in Fig. 2,
481
+ the MSE values for the HS method (blue region) are al-
482
+ ways above the MSE values for the NN method (orange
483
+ region), especially for lower values of N. These results
484
+ demonstrate that the NN method can predict the pa-
485
+ rameters of the NSD with an improved accuracy (up to
486
+ 5 times larger) with respect to the HS method. The test
487
+ presented in this subsection have been performed with
488
+ simulated data. In the next subsection we are going to
489
+ repeat the same test but with experimental data.
490
+ D.
491
+ Experimental test of neural networks
492
+ By this point we know that NN can reliably predict the
493
+ NSD from noisy simulated data. In this section, we want
494
+ to use the NN (trained and validated with noisy simu-
495
+ lated data) to reconstruct the NSD using experimental
496
+ data.
497
+ As quantum sensor we use a spin qubit encoded in the
498
+ electronic spin of the ground state of a single nitrogen-
499
+ vacancy (NV) center in a bulk diamond at room temper-
500
+ ature. This system has proven as a sensitive quantum
501
+ probe of magnetic fields, with outstanding spacial reso-
502
+ lution and sensitivity [47, 48]. The diamond sample in
503
+ our experiments has a natural abundance of 13C impu-
504
+ rities (1.1%) that are randomly distributed in the dia-
505
+ mond lattice [23–25]. The 13C nuclear spins constitute
506
+ the external environment of the NV center.
507
+ They act
508
+ as a collective bath of spins that induces dephasing into
509
+ the NV electronic spin, limiting the its coherence time
510
+ T2 ≈ 100 µs. In the presence of strong bias magnetic
511
+ field (≥ 150G) [32, 49], the weak coupling of the NV spin
512
+ with these carbon impurities can be modeled as a clas-
513
+ sical stochastic field. The latter has a power spectrum
514
+ density (here called NSD) that follows a Gaussian dis-
515
+ tribution centered at the Larmor frequency of the 13C
516
+ nuclear spins. In order to measure the NV spin coher-
517
+ ence function C(τ, N), we apply a train of π pulses (in
518
+ our case a CP sequence) to the NV spin qubit following
519
+ the DD protocol described in Fig 1. For more details on
520
+ the experimental implementation and Hamiltonian of the
521
+ system see Ref. [32]. We have performed this experiment
522
+ for N = {1, 8, 16, 24, 32, 40, 48}, and for τ ∈ [3.3, 3.66] µs
523
+ and [5.5, 6.1] µs with sampling time ∆t = 20 ns. The
524
+ results are shown in Fig.3(a) (blue bullets). Then, the
525
+ collected coherence functions have been processed and
526
+ employed to reconstruct the NSD parameters by means
527
+ of both the NN (trained with the generated dataset) and
528
+ the HS method. In contrast with the test using simu-
529
+ lated data in the previous section, in the experimental
530
+ case we do not know the exact values of the NSD pa-
531
+ rameters.
532
+ Therefore, we cannot calculate the MSE to
533
+ quantify the accuracy of the reconstructed parameters.
534
+ In order to estimate such accuracy we have used the fol-
535
+ lowing procedure: from the inferred NSD, the coherence
536
+ curves C(τ, N) are simulated and then compared with
537
+ the experimental results. An example of this comparison
538
+ is shown in Fig.3(a), where C(τ, N) is simulated under
539
+ the assumption that the NSD parameters are inferred ei-
540
+ ther by the machine learning models (orange) or by the
541
+ HS method (red), both for N = 16. Qualitatively it is
542
+ clear that the orange curves are much closer to the ex-
543
+ perimental data, than the red curves.
544
+ There are several options to quantitatively compare
545
+ the experimental data and the simulation results. Here
546
+ we use both the reduced chi-squared χ2
547
+ ν [50], and the
548
+
549
+ 6
550
+ FIG. 3: (a) Coherence function C(τ, N). The experimental data (blue bullets) are shown together with the
551
+ simulated ones using the NSD predicted respectively by the HS method (red lines) and machine learning models
552
+ (orange lines), both for N = 16. (b) Reduced chi-squared χ2
553
+ ν, obtained by comparing simulation and experimental
554
+ data, as a function of N. As in panel (a), orange and red curves refer to the NN and HS method, respectively.
555
+ Instead, the dashed line denotes the value of the reduced chi-squared for the HS method when we employ additional
556
+ measurements for N = 56, 64, 72, 80 in the interval τ ∈ [5.5, 6.1] µs. Inset: Same results but quantified by the
557
+ Mean-Absolute-Error (MAE) between the experimental data and the predicted C(τ, N).
558
+ Mean-Absolute-Error (MAE) [51] between the exper-
559
+ imental data and the predicted coherence functions
560
+ C(τ, N) (see Methods for more details). The results of
561
+ this comparison are shown in Fig. 3(b), where χ2
562
+ ν and
563
+ the MAE are plotted as a function of N. Remarkably,
564
+ the NSD reconstructed by the NN for N = 16 behaves
565
+ better that any case using the HS method. It is worth
566
+ observing that the same experimental data used to infer
567
+ the NSD parameters are partially used to estimate the
568
+ χ2
569
+ ν and MAE(C(t)). For example, for N = 16, only the
570
+ data for N = 1, 8, 16 are used to reconstruct the NSD,
571
+ but we employ all the data N = 1, 8, 16, . . . , 48 to ob-
572
+ tain the χ2
573
+ ν and MAE(C(t)). Overall, we have observed
574
+ enhanced performance in reconstructing the NSD of the
575
+ collective bath of spins, with a maximum improvement
576
+ (about 7 times higher) for N = 16. In other words, for
577
+ N = 16, once we reconstruct the NSD, the quantum sen-
578
+ sor dynamics can be predicted with an average square
579
+ deviation of ≃ 1.86 experimental error-bars by using the
580
+ NN method, or with an average square deviation of ≃ 13
581
+ error-bars if we use the HS method.
582
+ III.
583
+ DISCUSSION
584
+ As shown pictorially in Fig. 1, the NN takes as input
585
+ the spin qubit coherence functions (the coherence of the
586
+ quantum sensor decays due to the presence of the exter-
587
+ nal bath) obtained by using a set of different CP control
588
+ sequences. The NN returns as output the parameters of
589
+ the unknown NSD in the frequency domain.
590
+ One can
591
+ thus note that the NN, once validated, acts as a “time-
592
+ frequency converter” (making use of a quite complicated
593
+ deconvolution) from the measured signals living in the
594
+ time domain – the spin coherence functions – to the NSD
595
+ defined in the frequency domain.
596
+ The results shown in the previous section, and sum-
597
+ marized in Figs. 2 and 3(b), demonstrate that NN can
598
+ be used to reconstruct the NSD affecting a quantum sen-
599
+ sor, achieving higher precision and with considerable less
600
+ data than the standard HS method.
601
+ Improved values
602
+ of the reconstruction accuracy have been obtained with
603
+ simulated and experimental data. Both the HS and NN
604
+ methods are comparable – in terms of NSD reconstruc-
605
+ tion accuracy – for high values of N, but not for small
606
+ ones, where NN gives significantly better results. More-
607
+ over, the main result of our study is that NN trained
608
+ with data obtained for N = 16 reconstruct the NSD
609
+ more accurately than the best estimate provided by the
610
+ HS method with N = 48. This improvement is remark-
611
+ able by itself, but it becomes more significant when we
612
+ consider that the time required to complete these exper-
613
+ iments has a growth faster than a linear function with
614
+ respect to N, following an arithmetic progression. As an
615
+ example, the total time to perform all the experiments in
616
+ the case of N = 16 and 48 is respectively ≃ 10 minutes
617
+ and ≃ 1.2 hours [52]. This is an under-estimation of the
618
+ time difference between methods, because we are only
619
+ considering the bare measurement time, without taking
620
+ into account the time delay between different experi-
621
+ ments. Furthermore, it is worth stressing that our results
622
+ also show that deep learning has a predictive power since
623
+ it can be applied to never-before-seen data. This natu-
624
+ rally provides to the employed machine learning models
625
+ a connotation of robustness that is crucial in real appli-
626
+ cations.
627
+
628
+ 7
629
+ Let us observe that regression tasks, which are suc-
630
+ cessfully solved by multi-layer perceptrons (one of the
631
+ easiest form of NN), are less common with respect to the
632
+ ones to carry out classification; a review of some exam-
633
+ ple datasets and methods for regression is in Ref. [53].
634
+ Hence, we expect that the synthetic data used in this
635
+ work could be useful as a test bed also to the audience
636
+ of machine learning researchers and developers solving
637
+ regression problems in different contexts. With this in
638
+ mind, we share the training dataset with synthetic data
639
+ and our codes for their generation, as well as the code for
640
+ machine learning experiments and NSD reconstruction
641
+ [available on the GitHub repository (see Section “Data
642
+ and code availability”)]. In this way, we promote the im-
643
+ provement of machine learning models for noise sensing
644
+ purposes and their use to solve different regression tasks
645
+ in the quantum estimation framework.
646
+ Conclusions & outlooks
647
+ In this paper, we use NN to carry out noise spec-
648
+ troscopy with a quantum sensor using dynamical decou-
649
+ pling sequences with a much smaller number of π pulses
650
+ and, at the same time, achieving a higher reconstruc-
651
+ tion accuracy than standard methods (e.g., HS proto-
652
+ col). This means that with our proposal the noise spec-
653
+ troscopy procedure will take less time and give better
654
+ results. More in detail, we experimentally demonstrate
655
+ the capability of NN to reconstruct the NSD of the collec-
656
+ tive nuclear spin bath that surrounds an electronic spin
657
+ qubit, i.e., the ground state of a single nitrogen-vacancy
658
+ center in bulk diamond at room temperature.
659
+ To conclude, we outline some possible outlooks for our
660
+ work.
661
+ First of all, one may evaluate the performance
662
+ of NN that are trained over input data obtained using
663
+ DD control sequences with more degrees of freedom than
664
+ the CP ones [54–58]. Secondly, deep learning might be
665
+ applied to noise spectroscopy techniques beyond the HS
666
+ methods, as for example optimal band-limited control
667
+ protocols [34, 35] and even non-Gaussian noise charac-
668
+ terization [59–61]. In addition, it might be worth inves-
669
+ tigating how deep learning can be integrated to quantum
670
+ sensing procedures that rely on the so-called stochastic
671
+ quantum Zeno effect [62, 63], whereby the quantum probe
672
+ is subjected to a sequence of quantum measurements that
673
+ in the ideal case are designed to confine the dynamics of
674
+ the probe around the initial (nominal) state [33, 64, 65].
675
+ We are also confident that the extent of our results can
676
+ be quite easily replicated in other experimental settings,
677
+ as e.g., superconducting flux qubits [66, 67], trapped
678
+ ions [68, 69], cold atoms [70, 71], quantum dots [72, 73],
679
+ NMR experiments in molecules [31, 74], and nanoelec-
680
+ tronic devices [75]. For such a purpose, one might slightly
681
+ adapt the deep learning techniques used here to methods
682
+ tailored for time series.
683
+ IV.
684
+ METHODS
685
+ A.
686
+ Technical details on the training of NN
687
+ The NN models are developed using the PyTorch
688
+ framework [76] on a machine with 32 CPU cores, 126Gb
689
+ of RAM and a GeForce RTX 3090 GPU. The training
690
+ time, including the optimization of the hyperparameters,
691
+ is around 12 hours for each N .
692
+ The hyperparameters optimization is implemented by
693
+ means of the Ray Tune library [77]. The Hyperopt pack-
694
+ age [78] uses the Tree-structured Parzen Estimators [79]
695
+ algorithm as a Bayesian optimization to search for the
696
+ best choice of the hyperparameters within a predefined
697
+ search space. Hyperopt suggest the likely better configu-
698
+ rations of the hyperparameters and the underlying model
699
+ is updated after each trial that is run. The ASHA sched-
700
+ uler [80] is then used to stop the run of the least promising
701
+ trials chosen by the search algorithm, thus speeding up
702
+ the hyperparameters optimization process.
703
+ The optimized hyperparameters are the following. (1)
704
+ The number of hidden layers decides the value of L −
705
+ 1 in Eq. (7). The hidden layers are between the input
706
+ layer h[0] and the output layer h[L]. (2) The dimension
707
+ of the hidden layers is the value of q in Eq. (6) that,
708
+ for the sake of simplicity, is equal for all the layers in
709
+ Eq. (7). Both the number and dimension of the hidden
710
+ layers are chosen by sampling log-uniformly an integer
711
+ value from the space [1, 32) and [1, 1024), respectively.
712
+ (3) The learning rate is responsible for the length of the
713
+ gradient descent step and it is optimized with a choice
714
+ between 10−2, 10−3 and 10−4. (4) The batch size denotes
715
+ the dimension of the batch on which the loss function is
716
+ summed for the gradient calculation in a single descent
717
+ step. The batch size is chosen between 2, 4, 8, 16, 32.
718
+ (5) The dropout is a regularization strategy that aims
719
+ to reduce the overfitting by randomly turn off the NN
720
+ neurons with a predefined probability. Such probability
721
+ is one among 0 (no dropout), 0.2 and 0.5. (6) The weight
722
+ decay is another regularization technique that adds to the
723
+ loss function the squared weights of the NN multiplied
724
+ by a decay value. The latter value is optimized choosing
725
+ between 0 (no decay), 10−6, 10−5, 10−4 and 10−3.
726
+ B.
727
+ Definition of quantifiers for reconstruction
728
+ accuracy
729
+ The accuracy NN and HS methods can be estimated by
730
+ using the reconstructed NSD to simulate the coherence
731
+ function C(τ, N), and ‘measuring’ the distance between
732
+ the simulated data and the experimental values. To do
733
+ so, we use the reduced chi-squared χ2
734
+ ν, and the Mean-
735
+ Absolute-Error (MAE(C)): We define Ce ± δCe (Cs) as
736
+ the experimental (simulated) values of C(τ, N), where
737
+ δCe is the standard deviation of the experimental data.
738
+
739
+ 8
740
+ Then we can write reduced chi-squared and the MAE as
741
+ χ2
742
+ ν ≡ 1
743
+ ν
744
+
745
+ n,N
746
+ (Ce(τn, N) − Cs(τn, N))2
747
+ δCe(τn, N)2
748
+ (12)
749
+ MAE(C) ≡ 1
750
+ ν
751
+
752
+ n,N
753
+ |Ce(τn, N) − Cs(τn, N)| ,
754
+ (13)
755
+ where N = {1, 8, 16, 24, . . . , N}, {τn} are the values of
756
+ the time between pulses within the time intervals defined
757
+ in main text, and ν is the total number of elements in the
758
+ sum. Notice that χ2
759
+ ν takes into account the experimental
760
+ precision to scale the difference between experiment and
761
+ simulation. The results showing both χ2
762
+ ν and the MAE
763
+ are in Fig. 3.
764
+ DATA AND CODE AVAILABILITY
765
+ The source codes for the generation of the train-
766
+ ing dataset and the machine learning experiments
767
+ are available on GitHub
768
+ at the following address:
769
+ https://github.com/trianam/noiseSpectroscopyNV
770
+ ACKNOWLEDGEMENTS
771
+ This work was supported by the European Com-
772
+ mission’s
773
+ Horizon
774
+ Europe
775
+ Framework
776
+ Programme
777
+ under
778
+ the
779
+ Research
780
+ and
781
+ Innovation
782
+ Action
783
+ GA
784
+ n. 101070546–MUQUABIS, and by the European De-
785
+ fence Agency under the project Q-LAMPS Contract No
786
+ B PRJ-RT-989.
787
+ S. H. G. acknowledges support from
788
+ CNR-FOE-LENS-2020. F. C. and S. M. acknowledge the
789
+ European Union’s Horizon 2020 research and innovation
790
+ programme under FET-OPEN GA n. 828946–PATHOS.
791
+ [1] C. L. Degen, F. Reinhard, and P. Cappellaro, Quantum
792
+ sensing, Rev. Mod. Phys. 89, 035002 (2017).
793
+ [2] S. Hern´andez-G´omez and N. Fabbri, Quantum con-
794
+ trol for nanoscale spectroscopy with diamond nitrogen-
795
+ vacancy centers:
796
+ A short review, Front. Phys. 8,
797
+ 10.3389/fphy.2020.610868 (2021).
798
+ [3] F. Poggiali, P. Cappellaro, and N. Fabbri, Optimal con-
799
+ trol for one-qubit quantum sensing, Phys. Rev. X 8,
800
+ 021059 (2018).
801
+ [4] M. M. M¨uller, S. Gherardini, and F. Caruso, Noise-robust
802
+ quantum sensing via optimal multi-probe spectroscopy,
803
+ Scientific Reports 8, 14278 (2018).
804
+ [5] P. Rembold, N. Oshnik, M. M. M¨uller, S. Montangero,
805
+ T. Calarco, and E. Neu, Introduction to quantum optimal
806
+ control for quantum sensing with nitrogen-vacancy cen-
807
+ ters in diamond, AVS Quantum Science 2, 024701 (2020).
808
+ [6] A.
809
+ Marshall,
810
+ T.
811
+ Reisser,
812
+ P.
813
+ Rembold,
814
+ C.
815
+ M¨uller,
816
+ J. Scheuer, M. Gierse, T. Eichhorn, J. M. Steiner,
817
+ P. Hautle, T. Calarco, F. Jelezko, M. B. Plenio, S. Mon-
818
+ tangero, I. Schwartz, M. M. M¨uller, and P. Neumann,
819
+ Macroscopic hyperpolarization enhanced with quantum
820
+ optimal control, Phys. Rev. Res. 4, 043179 (2022).
821
+ [7] C. Bishop, Pattern Recognition and Machine Learning,
822
+ Information Science and Statistics (Springer, 2006).
823
+ [8] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learn-
824
+ ing (MIT Press, 2016) http://www.deeplearningbook.
825
+ org.
826
+ [9] A. Youssry, G. A. Paz-Silva, and C. Ferrie, Characteriza-
827
+ tion and control of open quantum systems beyond quan-
828
+ tum noise spectroscopy, npj Quantum Information 6, 95
829
+ (2020).
830
+ [10] S. Martina, S. Gherardini, and F. Caruso, Machine learn-
831
+ ing approach for quantum non-Markovian noise classifi-
832
+ cation, arXiv:2101.03221 (2021).
833
+ [11] D. F. Wise, J. J. Morton, and S. Dhomkar, Using deep
834
+ learning to understand and mitigate the qubit noise en-
835
+ vironment, PRX Quantum 2, 010316 (2021).
836
+ [12] N. Aharon, A. Rotem, L. P. McGuinness, F. Jelezko,
837
+ A. Retzker, and Z. Ringel, Nv center based nano-nmr
838
+ enhanced by deep learning, Scientific Reports 9, 17802
839
+ (2019).
840
+ [13] R. Santagati, A. A. Gentile, S. Knauer, S. Schmitt,
841
+ S. Paesani, C. Granade, N. Wiebe, C. Osterkamp, L. P.
842
+ McGuinness, J. Wang, M. G. Thompson, J. G. Rarity,
843
+ F. Jelezko, and A. Laing, Magnetic-field learning using a
844
+ single electronic spin in diamond with one-photon read-
845
+ out at room temperature, Phys. Rev. X 9, 021019 (2019).
846
+ [14] K. Jung, M. H. Abobeih, J. Yun, G. Kim, H. Oh,
847
+ A. Henry, T. H. Taminiau, and D. Kim, Deep learning en-
848
+ hanced individual nuclear-spin detection, npj Quantum
849
+ Information 7, 41 (2021).
850
+ [15] A. Strikis, D. Qin, Y. Chen, S. C. Benjamin, and Y. Li,
851
+ Learning-based quantum error mitigation, PRX Quan-
852
+ tum 2, 040330 (2021).
853
+ [16] R. Harper, S. T. Flammia, and J. J. Wallman, Effi-
854
+ cient learning of quantum noise, Nature Physics 16, 1184
855
+ (2020).
856
+ [17] S. Martina, L. Buffoni, S. Gherardini, and F. Caruso,
857
+ Learning the noise fingerprint of quantum devices, Quan-
858
+ tum Machine Intelligence 4, 1 (2022).
859
+ [18] S. Martina, S. Gherardini, L. Buffoni, and F. Caruso,
860
+ Noise fingerprints in quantum computers: Machine learn-
861
+ ing software tools, Software Impacts 12, 100260 (2022).
862
+ [19] G. Koolstra, N. Stevenson, S. Barzili, L. Burns, K. Siva,
863
+ S. Greenfield, W. Livingston, A. Hashim, R. K. Naik,
864
+ J. M. Kreikebaum, K. P. O’Brien, D. I. Santiago, J. Dres-
865
+ sel, and I. Siddiqi, Monitoring fast superconducting qubit
866
+ dynamics using a neural network, Phys. Rev. X 12,
867
+ 031017 (2022).
868
+ [20] M. Gong, H.-L. Huang, S. Wang, C. Guo, S. Li, Y. Wu,
869
+ Q. Zhu, Y. Zhao, S. Guo, H. Qian, Y. Ye, C. Zha,
870
+ F. Chen, C. Ying, J. Yu, D. Fan, D. Wu, H. Su, H. Deng,
871
+ H. Rong, K. Zhang, S. Cao, J. Lin, Y. Xu, L. Sun,
872
+ C. Guo, N. Li, F. Liang, A. Sakurai, K. Nemoto, W. J.
873
+ Munro, Y.-H. Huo, C.-Y. Lu, C.-Z. Peng, X. Zhu, and J.-
874
+
875
+ 9
876
+ W. Pan, Quantum Neuronal Sensing of Quantum Many-
877
+ Body States on a 61-Qubit Programmable Superconduct-
878
+ ing Processor, arXiv:2201.05957 (2022).
879
+ [21] L. Viola and S. Lloyd, Dynamical suppression of deco-
880
+ herence in two-state quantum systems, Phys. Rev. A 58,
881
+ 2733 (1998).
882
+ [22] L. Faoro and L. Viola, Dynamical suppression of 1/f
883
+ noise processes in qubit systems, Phys. Rev. Lett. 92,
884
+ 117905 (2004).
885
+ [23] J. M. Taylor, P. Cappellaro, L. Childress, L. Jiang,
886
+ D. Budker, P. R. Hemmer, A. Yacoby, R. Walsworth,
887
+ and M. D. Lukin, High-sensitivity diamond magnetome-
888
+ ter with nanoscale resolution, 4, 810 (2008).
889
+ [24] G. Goldstein, P. Cappellaro, J. R. Maze, J. S. Hodges,
890
+ L. Jiang, A. S. Sorensen, and M. D. Lukin, Environment
891
+ assisted precision measurement, Phys. Rev. Lett. 106,
892
+ 140502 (2011).
893
+ [25] M. H. Abobeih, J. Cramer, M. A. Bakker, N. Kalb,
894
+ M. Markham, D. J. Twitchen, and T. H. Taminiau, One-
895
+ second coherence for a single electron spin coupled to a
896
+ multi-qubit nuclear-spin environment, Nature Communi-
897
+ cations 9 (2018).
898
+ [26] L. Cywi´nski, R. M. Lutchyn, C. P. Nave, and S. Das-
899
+ Sarma, How to enhance dephasing time in superconduct-
900
+ ing qubits, Phys. Rev. B 77, 174509 (2008).
901
+ [27] M. J. Biercuk, A. C. Doherty, and H. Uys, Dynamical
902
+ decoupling sequence construction as a filter-design prob-
903
+ lem, Journal of Physics B: Atomic, Molecular and Optical
904
+ Physics 44, 154002 (2011).
905
+ [28] H. Y. Carr and E. M. Purcell, Effects of diffusion on free
906
+ precession in nuclear magnetic resonance experiments,
907
+ Phys. Rev. 94, 630 (1954).
908
+ [29] J. R. Maze, P. L. Stanwix, J. S. Hodges, S. Hong, J. M.
909
+ Taylor, P. Cappellaro, L. Jiang, A. Zibrov, A. Yacoby,
910
+ R. Walsworth, and M. D. Lukin, Nanoscale magnetic
911
+ sensing with an individual electronic spin qubit in dia-
912
+ mond, Nature 455, 644 (2008).
913
+ [30] T. Yuge, S. Sasaki, and Y. Hirayama, Measurement of
914
+ the noise spectrum using a multiple-pulse sequence, Phys.
915
+ Rev. Lett. 107, 170504 (2011).
916
+ [31] G. A. ´Alvarez and D. Suter, Measuring the spectrum of
917
+ colored noise by dynamical decoupling, Phys. Rev. Lett.
918
+ 107, 230501 (2011).
919
+ [32] S. Hern´andez-G´omez, F. Poggiali, P. Cappellaro, and
920
+ N. Fabbri, Noise spectroscopy of a quantum-classical
921
+ environment with a diamond qubit, Phys. Rev. B 98,
922
+ 214307 (2018).
923
+ [33] H.-V. Do, C. Lovecchio, I. Mastroserio, N. Fabbri, F. S.
924
+ Cataliotti, S. Gherardini, M. M. M¨uller, N. D. Pozza, and
925
+ F. Caruso, Experimental proof of quantum zeno-assisted
926
+ noise sensing, New Journal of Physics 21, 113056 (2019).
927
+ [34] V. M. Frey, S. Mavadia, L. M. Norris, W. de Ferranti,
928
+ D. Lucarelli, L. Viola, and M. J. Biercuk, Application of
929
+ optimal band-limited control protocols to quantum noise
930
+ sensing, Nature Communications 8, 2189 (2017).
931
+ [35] V. Frey, L. M. Norris, L. Viola, and M. J. Biercuk, Simul-
932
+ taneous spectral estimation of dephasing and amplitude
933
+ noise on a qubit sensor via optimally band-limited con-
934
+ trol, Phys. Rev. Applied 14, 024021 (2020).
935
+ [36] G. Wang, Y. Zhu, B. Li, C. Li, L. Viola, A. Cooper, and
936
+ P. Cappellaro, Digital noise spectroscopy with a quantum
937
+ sensor, arXiv:2212.09216 (2022).
938
+ [37] P. Sza´nkowski, G. Ramon, J. Krzywda, D. Kwiatkowski,
939
+ and �L. Cywi´nski, Environmental noise spectroscopy with
940
+ qubits subjected to dynamical decoupling, Journal of
941
+ Physics: Condensed Matter 29, 333001 (2017).
942
+ [38] P. Sza´nkowski and L. Cywi´nski, Accuracy of dynamical-
943
+ decoupling-based spectroscopy of gaussian noise, Phys.
944
+ Rev. A 97, 032101 (2018).
945
+ [39] G. Gordon, N. Erez, and G. Kurizki, Universal dynam-
946
+ ical decoherence control of noisy single- and multi-qubit
947
+ systems, Journal of Physics B: Atomic, Molecular and
948
+ Optical Physics 40, S75 (2007).
949
+ [40] G. Gordon, G. Kurizki, and D. A. Lidar, Optimal dy-
950
+ namical decoherence control of a qubit, Phys. Rev. Lett.
951
+ 101, 010403 (2008).
952
+ [41] N. D. Pozza, S. Gherardini, M. M. M¨uller, and F. Caruso,
953
+ Role of the filter functions in noise spectroscopy, Inter-
954
+ national Journal of Quantum Information 17, 1941008
955
+ (2019).
956
+ [42] For the coherence curves, the first and second order of
957
+ the collapses refer to the harmonics of the filter functions
958
+ F(ω, τ, N). For more details see Ref. [32].
959
+ [43] X. Glorot, A. Bordes, and Y. Bengio, Deep sparse rectifier
960
+ neural networks, in Proceedings of the Fourteenth Inter-
961
+ national Conference on Artificial Intelligence and Statis-
962
+ tics, Proceedings of Machine Learning Research, Vol. 15,
963
+ edited by G. Gordon, D. Dunson, and M. Dud´ık (PMLR,
964
+ Fort Lauderdale, FL, USA, 2011) pp. 315–323.
965
+ [44] V. Nair and G. E. Hinton, Rectified linear units improve
966
+ restricted boltzmann machines, in Proceedings of the 27th
967
+ International Conference on International Conference on
968
+ Machine Learning, ICML’10 (Omnipress, Madison, WI,
969
+ USA, 2010) p. 807–814.
970
+ [45] S. Shalev-Shwartz and S. Ben-David, Understanding ma-
971
+ chine learning: From theory to algorithms (Cambridge
972
+ University Press, 2014).
973
+ [46] D. P. Kingma and J. Ba, Adam: A method for stochastic
974
+ optimization, arXiv:1412.6980 (2014).
975
+ [47] M. W. Doherty, N. B. Manson, P. Delaney, F. Jelezko,
976
+ J. Wrachtrup, and L. C. Hollenberg, The nitrogen-
977
+ vacancy colour centre in diamond, Physics Reports 528,
978
+ 1 (2013), the nitrogen-vacancy colour centre in diamond.
979
+ [48] L. Rondin, J.-P. Tetienne, T. Hingant, J.-F. Roch,
980
+ P. Maletinsky, and V. Jacques, Magnetometry with
981
+ nitrogen-vacancy
982
+ defects
983
+ in
984
+ diamond,
985
+ Reports
986
+ on
987
+ Progress in Physics 77, 056503 (2014).
988
+ [49] F. Reinhard, F. Shi, N. Zhao, F. Rempp, B. Naydenov,
989
+ J. Meijer, L. T. Hall, L. Hollenberg, J. Du, R.-B. Liu, and
990
+ J. Wrachtrup, Tuning a spin bath through the quantum-
991
+ classical transition, Phys. Rev. Lett. 108, 200402 (2012).
992
+ [50] I. Hughes and T. Hase, Measurements and their uncer-
993
+ tainties: a practical guide to modern error analysis (OUP
994
+ Oxford, 2010) pp. 107–107.
995
+ [51] C. Sammut and G. I. Webb, eds., Mean absolute error, in
996
+ Encyclopedia of Machine Learning (Springer US, Boston,
997
+ MA, 2010) pp. 652–652.
998
+ [52] For this estimation we consider 105 repetitions as in our
999
+ experiments. We recall that the total time for each repe-
1000
+ tition of the single experiment is T = Nτ.
1001
+ [53] M. Fern´andez-Delgado,
1002
+ M. S. Sirsat,
1003
+ E. Cernadas,
1004
+ S. Alawadi, S. Barro, and M. Febrero-Bande, An exten-
1005
+ sive experimental survey of regression methods, Neural
1006
+ Networks 111, 11 (2019).
1007
+ [54] G. S. Uhrig, Keeping a quantum bit alive by optimized
1008
+ π-pulse sequences, Phys. Rev. Lett. 98, 100504 (2007).
1009
+
1010
+ 10
1011
+ [55] N. Zhao, J.-L. Hu, S.-W. Ho, J. T. K. Wan, and R. B.
1012
+ Liu, Atomic-scale magnetometry of distant nuclear spin
1013
+ clusters via nitrogen-vacancy spin in diamond.,
1014
+ 6, 242
1015
+ (2011).
1016
+ [56] A. M. Souza, G. A. ´Alvarez, and D. Suter, Robust dy-
1017
+ namical decoupling for quantum computing and quantum
1018
+ memory, Phys. Rev. Lett. 106, 240501 (2011).
1019
+ [57] N. Zhao, J. Wrachtrup, and R.-B. Liu, Dynamical decou-
1020
+ pling design for identifying weakly coupled nuclear spins
1021
+ in a bath, Phys. Rev. A 90, 032319 (2014).
1022
+ [58] J. Casanova, Z.-Y. Wang, J. F. Haase, and M. B. Plenio,
1023
+ Robust dynamical decoupling sequences for individual-
1024
+ nuclear-spin addressing, Phys. Rev. A 92, 042304 (2015).
1025
+ [59] G. A. Paz-Silva and L. Viola, General transfer-function
1026
+ approach to noise filtering in open-loop quantum control,
1027
+ Phys. Rev. Lett. 113, 250501 (2014).
1028
+ [60] L. M. Norris, G. A. Paz-Silva, and L. Viola, Qubit noise
1029
+ spectroscopy for non-gaussian dephasing environments,
1030
+ Phys. Rev. Lett. 116, 150503 (2016).
1031
+ [61] Y. Sung, F. Beaudoin, L. M. Norris, F. Yan, D. K. Kim,
1032
+ J. Y. Qiu, U. von L¨upke, J. L. Yoder, T. P. Orlando,
1033
+ S. Gustavsson, L. Viola, and W. D. Oliver, Non-gaussian
1034
+ noise spectroscopy with a superconducting qubit sensor,
1035
+ Nature Communications 10, 3715 (2019).
1036
+ [62] A. I. Shushin, The effect of measurements, randomly
1037
+ distributed in time, on quantum systems:
1038
+ stochastic
1039
+ quantum zeno effect, J. Phys. A: Math. Theor. 44,
1040
+ 10.1088/1751-8113/44/5/055303 (2011).
1041
+ [63] S. Gherardini, S. Gupta, F. S. Cataliotti, A. Smerzi,
1042
+ F. Caruso, and S. Ruffo, Stochastic quantum zeno by
1043
+ large deviation theory, New J. Phys. 18, 10.1088/1367-
1044
+ 2630/18/1/013048 (2016).
1045
+ [64] M. M. M¨uller, N. Dalla Pozza, S. Gherardini, and
1046
+ F.
1047
+ Caruso,
1048
+ Noise
1049
+ sensing
1050
+ via
1051
+ stochastic
1052
+ quantum
1053
+ zeno, Phys. Lett. A 384, 10.1016/j.physleta.2020.126244
1054
+ (2020).
1055
+ [65] S. Virz`ı, A. Avella, F. Piacentini, M. Gramegna, T. c. v.
1056
+ Opatrn´y, A. G. Kofman, G. Kurizki, S. Gherardini,
1057
+ F. Caruso, I. P. Degiovanni, and M. Genovese, Quantum
1058
+ Zeno and Anti-Zeno Probes of Noise Correlations in Pho-
1059
+ ton Polarization, Phys. Rev. Lett. 129, 030401 (2022).
1060
+ [66] J. Bylander, S. Gustavsson, F. Yan, F. Yoshihara,
1061
+ K. Harrabi, G. Fitch, D. G. Cory, Y. Nakamura, J.-S.
1062
+ Tsai, and W. D. Oliver, Noise spectroscopy through dy-
1063
+ namical decoupling with a superconducting flux qubit,
1064
+ Nature Physics 7, 565 (2011).
1065
+ [67] F. Yoshihara, Y. Nakamura, F. Yan, S. Gustavsson,
1066
+ J. Bylander, W. D. Oliver, and J.-S. Tsai, Flux qubit
1067
+ noise spectroscopy using rabi oscillations under strong
1068
+ driving conditions, Phys. Rev. B 89, 020503 (2014).
1069
+ [68] M. J. Biercuk, H. Uys, A. P. VanDevender, N. Shiga,
1070
+ W. M. Itano, and J. J. Bollinger, Optimized dynamical
1071
+ decoupling in a model quantum memory, Nature 458,
1072
+ 996 (2009).
1073
+ [69] S. Kotler, N. Akerman, Y. Glickman, A. Keselman, and
1074
+ R. Ozeri, Single-ion quantum lock-in amplifier, Nature
1075
+ 473, 61 (2011).
1076
+ [70] Y. Sagi, I. Almog, and N. Davidson, Process tomography
1077
+ of dynamical decoupling in a dense cold atomic ensemble,
1078
+ Phys. Rev. Lett. 105, 053201 (2010).
1079
+ [71] I. Almog, Y. Sagi, G. Gordon, G. Bensky, G. Kur-
1080
+ izki, and N. Davidson, Direct measurement of the sys-
1081
+ tem–environment coupling as a tool for understand-
1082
+ ing decoherence and dynamical decoupling, Journal of
1083
+ Physics B: Atomic, Molecular and Optical Physics 44,
1084
+ 154006 (2011).
1085
+ [72] K. W. Chan, W. Huang, C. H. Yang, J. C. C. Hwang,
1086
+ B. Hensen, T. Tanttu, F. E. Hudson, K. M. Itoh,
1087
+ A. Laucht, A. Morello, and A. S. Dzurak, Assessment of
1088
+ a silicon quantum dot spin qubit environment via noise
1089
+ spectroscopy, Phys. Rev. Applied 10, 044017 (2018).
1090
+ [73] F. K. Malinowski, F. Martins, L. Cywi´nski, M. S. Rud-
1091
+ ner, P. D. Nissen, S. Fallahi, G. C. Gardner, M. J. Man-
1092
+ fra, C. M. Marcus, and F. Kuemmeth, Spectrum of the
1093
+ nuclear environment for gaas spin qubits, Phys. Rev.
1094
+ Lett. 118, 177702 (2017).
1095
+ [74] Y. Fu, Y. Wu, Y. Dai, X. Qin, X. Rong, and J. Du,
1096
+ Molecular-spin-qubit noise spectroscopy through dynam-
1097
+ ical decoupling, Phys. Rev. Applied 15, L061001 (2021).
1098
+ [75] J. T. Muhonen, J. P. Dehollain, A. Laucht, F. E. Hud-
1099
+ son, R. Kalra, T. Sekiguchi, K. M. Itoh, D. N. Jamieson,
1100
+ J. C. McCallum, A. S. Dzurak, and A. Morello, Storing
1101
+ quantum information for 30 seconds in a nanoelectronic
1102
+ device, Nature Nanotechnology 9, 986 (2014).
1103
+ [76] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury,
1104
+ G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga,
1105
+ A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Rai-
1106
+ son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang,
1107
+ J. Bai, and S. Chintala, Pytorch: An imperative style,
1108
+ high-performance deep learning library, in Advances in
1109
+ Neural Information Processing Systems, Vol. 32, edited
1110
+ by H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch´e-
1111
+ Buc, E. Fox, and R. Garnett (Curran Associates, Inc.,
1112
+ 2019).
1113
+ [77] R. Liaw, E. Liang, R. Nishihara, P. Moritz, J. E. Gon-
1114
+ zalez, and I. Stoica, Tune: A Research Platform for Dis-
1115
+ tributed Model Selection and Training, arXiv:1807.05118
1116
+ (2018).
1117
+ [78] J. Bergstra, D. Yamins, and D. Cox, Making a science of
1118
+ model search: Hyperparameter optimization in hundreds
1119
+ of dimensions for vision architectures, in Proceedings of
1120
+ the 30th International Conference on Machine Learn-
1121
+ ing, Proceedings of Machine Learning Research, Vol. 28,
1122
+ edited by S. Dasgupta and D. McAllester (PMLR, At-
1123
+ lanta, Georgia, USA, 2013) pp. 115–123.
1124
+ [79] J. Bergstra, R. Bardenet, Y. Bengio, and B. K´egl, Al-
1125
+ gorithms for hyper-parameter optimization, Advances in
1126
+ neural information processing systems 24 (2011).
1127
+ [80] L. Li, K. Jamieson, A. Rostamizadeh, E. Gonina, J. Ben-
1128
+ tzur, M. Hardt, B. Recht, and A. Talwalkar, A system
1129
+ for massively parallel hyperparameter tuning, in Proceed-
1130
+ ings of Machine Learning and Systems, Vol. 2, edited by
1131
+ I. Dhillon, D. Papailiopoulos, and V. Sze (2020) pp. 230–
1132
+ 246.
1133
+
59E4T4oBgHgl3EQfcAxC/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
6NAyT4oBgHgl3EQfcffh/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:37ea4f6850d0264f534b79f8cd45e541da5c6be778505c95b078f37a48807ef5
3
+ size 917549
6dE0T4oBgHgl3EQfvwHw/content/2301.02624v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:242b909e0e6c2f7a0fb5555cf395a6e27942b38fc6104664018ea3fe3eb7ead1
3
+ size 206271
6dE0T4oBgHgl3EQfvwHw/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d808220971d8f99cdd04cec2f8d5e364ab843cf6bfb747274ffe82dc57aea29
3
+ size 2621485
6dE0T4oBgHgl3EQfvwHw/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0487082bbb8c5c3c11f4aeb6ed78979d543df24d467931a01af6aacfc175ee7d
3
+ size 104603
6dE2T4oBgHgl3EQfkgfV/content/tmp_files/2301.03980v1.pdf.txt ADDED
@@ -0,0 +1,586 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
2
+ Language Models sounds the Death Knell of
3
+ Knowledge Graphs
4
+ Kunal Suri
5
+ Optum, India
6
+ kunal_suri@optum.com
7
+ Swapna Sourav Rout
8
+ Optum, India
9
+ rout.swapnasourav@optum.com
10
+ Atul Singh
11
+ Optum, India
12
+ atul_singh18@optum.com
13
+
14
+ Prakhar Mishra
15
+ Optum, India
16
+ prakhar_mishra29@optum.com
17
+ Rajesh Sabapathy
18
+ Optum,India
19
+ rajesh_sabapathy@uhc.com
20
+ Abstract—Healthcare
21
+ domain
22
+ generates
23
+ a
24
+ lot
25
+ of
26
+ unstructured and semi-structured text. Natural Language
27
+ processing (NLP) has been used extensively to process this data.
28
+ Deep Learning based NLP especially Large Language Models
29
+ (LLMs) such as BERT have found broad acceptance and are
30
+ used extensively for many applications. A Language Model is a
31
+ probability distribution over a word sequence. Self-supervised
32
+ Learning on a large corpus of data automatically generates deep
33
+ learning-based language models. BioBERT and Med-BERT are
34
+ language models pre-trained for the healthcare domain.
35
+ Healthcare uses typical NLP tasks such as question answering,
36
+ information extraction, named entity recognition, and search to
37
+ simplify and improve processes. However, to ensure robust
38
+ application of the results, NLP practitioners need to normalize
39
+ and standardize them. One of the main ways of achieving
40
+ normalization and standardization is the use of Knowledge
41
+ Graphs. A Knowledge Graph captures concepts and their
42
+ relationships for a specific domain, but their creation is time-
43
+ consuming and requires manual intervention from domain
44
+ experts,
45
+ which
46
+ can
47
+ prove
48
+ expensive.
49
+ SNOMED
50
+ CT
51
+ (Systematized Nomenclature of Medicine - Clinical Terms),
52
+ Unified Medical Language System (UMLS), and Gene Ontology
53
+ (GO) are popular ontologies from the healthcare domain.
54
+ SNOMED CT and UMLS capture concepts such as disease,
55
+ symptoms and diagnosis and GO is the world's largest source of
56
+ information on the functions of genes. Healthcare has been
57
+ dealing with an explosion in information about different types
58
+ of drugs, diseases, and procedures. This paper argues that using
59
+ Knowledge Graphs is not the best solution for solving problems
60
+ in this domain. We present experiments using LLMs for the
61
+ healthcare domain to demonstrate that language models
62
+ provide the same functionality as knowledge graphs, thereby
63
+ making knowledge graphs redundant.
64
+ Keywords—Medical
65
+ data,
66
+ Language
67
+ Models,
68
+ Natural
69
+ Language Processing, Knowledge Graphs, Deep Learning
70
+ I. INTRODUCTION
71
+ Knowledge graphs (KG) are knowledge bases that capture
72
+ concepts and their relationships for a specific domain using a
73
+ graph-structured data model. Systematized Nomenclature of
74
+ Medicine – Clinical Terms (SNOMED CT) (SNOMED),
75
+ Unified Medical Language Systems(UMLS) [Bodenreider O.
76
+ 2004], etc., are some of the popular KG in the healthcare
77
+ domain. Fig. 1 shows a sample from a representative medical
78
+ entity, KG. On the other hand, a language model is a
79
+ probability distribution over a word sequence and is the
80
+ backbone of modern natural language processing (NLP).
81
+ Language models try to capture any language's linguistic
82
+ intuition and writing, and large language models like BERT
83
+ [Devlin et al., 2019] and GPT-2 [Radford et al., 2019] have
84
+ shown remarkable performance. The paper presents a study
85
+ demonstrating that language models' ability to learn
86
+ relationships among different entities makes knowledge
87
+ graphs redundant for many applications.
88
+
89
+ This paper uses similar terms from SNOMED-CT KG and
90
+ passes them through a language model for the healthcare
91
+ domain BioRedditBERT to get a 768-dimensional dense
92
+ vector representation. The paper presents the results for
93
+ analyzing these embeddings. The experiments presented in
94
+ the paper validate that similar terms cluster together. The
95
+ paper uses simple heuristics to assign names to clusters. The
96
+ results show that the cluster names match the names in the
97
+ KG. Finally, the experiments demonstrate that the cosine
98
+ similarity of vector representation of similar terms is high and
99
+ vice versa.
100
+
101
+ Our contributions include: (i) We propose a study to
102
+ demonstrate the value and application of Large Language
103
+ Models (LLMs) in comparison to Knowledge Graph-based
104
+ approaches for the task of synonym extraction. (ii) We
105
+ extensively evaluate our approach on a standard, widely
106
+ accepted dataset, and the results are encouraging.
107
+
108
+
109
+ Fig 1. Medical entity Knowledge Graph Representation
110
+ The rest of the paper is organized as follows: Section II
111
+ presents the background required to understand the work
112
+ presented in this paper. Section III presents a literature survey
113
+ of related work on knowledge graphs and language models.
114
+ Section IV presents our understanding of how current days
115
+ language models are making knowledge graphs redundant.
116
+ Section V describes our proposed approach. Section VI
117
+ describes the experiments conducted and the results obtained.
118
+ Finally, section VII summarizes our work and discusses
119
+ possible directions for future study.
120
+ II. BACKGROUND
121
+ This section defines and describes Language Models and
122
+ Knowledge Graphs as used in this paper:
123
+
124
+
125
+ Medicine
126
+ Fever
127
+ Allergy
128
+ Dolo
129
+ ClaritinA. Language Models
130
+
131
+ A Language Model predicts the probability of a sequence of
132
+ words in a human language such as English. In the equation
133
+ below P(w1,…wm) is the probability of the word sequence
134
+ S, where S = (w1, w2, …, wm) and wi is the ith word in the
135
+ sequence.
136
+
137
+
138
+
139
+ Large Language Models (LLMs) are language models
140
+ trained on large general corpora that learn associations and
141
+ relationships
142
+ among
143
+ different
144
+ word
145
+ entities
146
+ in
147
+ an
148
+ unsupervised manner. Large Language Models (LLMs) are
149
+ considered universal language learners. LLMs such as BERT
150
+ and GPTare deep neural networks based on transformer
151
+ architecture. One of many reasons for the immense popularity
152
+ of LLMs is that these models are pre-trained self-supervised
153
+ models and can be adapted or fine-tuned to cater to a wide
154
+ range of NLP tasks. Few-shot learning has enabled these
155
+ LLMs to be adapted to a given NLP task using fewer training
156
+ samples.
157
+
158
+ Another reason for the immense popularity of LLMs is that
159
+ a single language model is applicable for multiple
160
+ downstream applications such as Token classification, Text
161
+ classification, and Question answering. LLMs generate
162
+ embeddings or word vectors for words, and these embeddings
163
+ capture the context of the word in the corpus. This ability of
164
+ LLMs to generate embeddings based on the corpus makes
165
+ them ubiquitous in almost NLP tasks.
166
+
167
+ In this paper, we use BioRedditBERT [Basaldella et al.,
168
+ 2020], a variant of BERT trained for the healthcare domain.
169
+ It is a domain-specific language representation model trained
170
+ on large-scale biomedical corpora from Reddit.
171
+
172
+ B. Knowledge Graphs
173
+
174
+ Knowledge Graphs (KGs) organize data and capture
175
+ relationships between different entities for a domain. Domain
176
+ experts create KGs to map domain-based relations between
177
+ various entities.
178
+
179
+ Knowledge graphs are Graph data structures with nodes
180
+ and edges. Nodes or vertices represent entities of interest, and
181
+ edges represent relations between them, as shown in Fig 1.
182
+ KGs can map and model direct and latent relationships
183
+ between entities of interest. Typically, KGs are used to model
184
+ and map information from model sources. Once KGs are
185
+ designed, typically, NLP is used to populate & create the
186
+ knowledge base from unstructured text corpora.
187
+
188
+ Knowledge graphs play a crucial role in healthcare
189
+ knowledge representation. There are many widely used
190
+ knowledge graphs like SNOMED and UMLS etc. In
191
+ healthcare, KGs are used for drug discovery drugs,
192
+ identifying tertiary symptoms for diseases and augmented
193
+ decision-making, etc.
194
+
195
+ COMETA: A Corpus for Medical Entity Linking in social
196
+ media [Basaldella et al., 2020] – a corpus containing four
197
+ years of content in 68 health-themed subreddits and
198
+ annotating the most frequent with their corresponding
199
+ SNOMED-CT entities. In this paper, we have used COMETA
200
+ to obtain synonyms from SNOMED-CT.
201
+ III. RELATED WORK
202
+ In 2019, Jawahar et al. performed experiments to understand
203
+ the underlying language structure learned by a language
204
+ model like BERT [Ganesh Jawahar et al. 2019]. The authors
205
+ show that BERT captures the semantic information from the
206
+ language hierarchically through experiments. BERT captures
207
+ surface features in the bottom layer, syntactic elements in the
208
+ middle and semantic features in the top layer. The work
209
+ presented in this paper treats the BERT model as a black box
210
+ and demonstrates that BERT can learn the information in a
211
+ knowledge graph through experiments on real-life healthcare
212
+ use cases.
213
+
214
+ There have been studies to generate a knowledge graph
215
+ directly from the output of LLMs. [Wang C et al., 2020;
216
+ Wang X et al. 2022] proposes a mechanism to create a KG
217
+ directly from LLMs. This mechanism talks about a two-step
218
+ mechanism to generate a KG from LLM. In the first step,
219
+ different candidate triplets are created from the text corpus.
220
+ Attention weights from a pre-trained LLM are used to get the
221
+ best-matched candidate triplets and then validated through a
222
+ beam search. In the second stage, the matched candidate
223
+ triplets are mapped to a pre-defined KG for validation, and
224
+ the unmatched candidates are used to create an open
225
+ knowledge graph. The work demonstrates the feasibility of
226
+ the idea presented in this paper that LLM can be used as a
227
+ substitute for knowledge graphs, especially since they
228
+ contain the information in the KG.
229
+
230
+ There is a body of research on integrating Knowledge
231
+ graphs and LLMs. Structured knowledge from Knowledge
232
+ Graphs is effectively integrated into Language models to
233
+ enhance the pre-trained language models [Lei He et al.,
234
+ 2021]. However, these approaches have found limited
235
+ success, thereby strengthening the position in this paper that
236
+ LLMs contain information from KGs.
237
+ IV. LANGUAGE MODELS FOR KNOWLEDGE GRAPHS
238
+ Language Models can find associations between different
239
+ words based on the attention weight matrix. The
240
+ methodology to use attention weights as a measure of
241
+ relationship among the entities indicates that Knowledge
242
+ graphs are getting replaced by LLMs as they learn more
243
+ generic relationships in an unsupervised way. The proposed
244
+ methodology in this paper is built on this idea to demonstrate
245
+ that Knowledge graphs are increasingly getting redundant for
246
+ many NLP tasks.
247
+ V. PROPOSED APPROACH
248
+ The paper demonstrates that language models' ability to learn
249
+ relationships among different entities makes knowledge
250
+ graphs redundant for many applications. To illustrate this, we
251
+ have used word embeddings for all the synonyms of a set of
252
+ medical terms from a large language model. This work uses
253
+
254
+ m
255
+ P(w1,..., Wm) =|[P(wi I Wi,..., Wi-1)
256
+ i=1COMETA data to obtain synonyms for a set of medical terms.
257
+ In COMETA data, the work focuses on the following
258
+ columns: a) Example column, which contains the sentences
259
+ from health-themed forums on Reddit, b) Term column
260
+ contains the medical terms present in the Example column, c)
261
+ General SNOMED Label column; contains the literal
262
+ meaning of the Term column from the SNOMED Knowledge
263
+ Graphs. To obtain synonyms, we use the different values
264
+ from the Terms column for a specific value of the General
265
+ SNOMED Label column. For example, for Abdominal Wind
266
+ Pain General SNOMED label, we have the following three
267
+ synonyms that we can obtain from the Terms column: gas
268
+ pains, painful gas, and gas pain.
269
+
270
+ To calculate the word embeddings of every synonym term,
271
+ we
272
+ use
273
+ the
274
+ word_vector
275
+ function
276
+ from
277
+ the
278
+ biobert_embeddings python module [Jitendra Jangid, 2020].
279
+ Since the original code was incompatible with the current
280
+ version of Pytorch [Paszke, A. et al., 2019] and Huggingface
281
+ [Wolf et al., 2020], we modified it just enough to satisfy the
282
+ current version requirements – the core logic remains the
283
+ same. We tokenize every Term using HuggingFace
284
+ tokenizers
285
+ and
286
+ pass
287
+ the
288
+ tokenized
289
+ Term
290
+ through
291
+ BioRedditBERT model. The previous step gives us
292
+ embedding for the Term (or sub-terms if the model didn't see
293
+ the Term before). If the model has not seen the Term before,
294
+ then we sum up the embedding of all the subterms). We then
295
+ store all the embeddings for the next steps.
296
+
297
+ We perform the following two experiments after
298
+ generating the word embeddings for the synonyms of a set of
299
+ medical terms. In the first experiment, we cluster the word
300
+ embeddings for the synonyms of a set of medical terms and
301
+ assign names to clusters. The word embeddings are passed
302
+ into UMAP to generate a 2-dimensional representation. We
303
+ plot the 2-dimensional representation to examine how the
304
+ term cluster visually. UMAP is used as the dimensionality
305
+ reduction technique over PCA because it is a non-linear
306
+ dimensionality reduction technique and does very well to
307
+ preserve the local and global structure of the data as
308
+ compared to PCA. However, unlike PCA [Karl Pearson
309
+ F.R.S. , 1901], UMAP is very sensitive to hyperparameters
310
+ that we chose, so we visualize the embeddings for several
311
+ values of number of neighbours (n_neighbors) and minimum
312
+ distance (min_dist). This step will help us visually validate
313
+ that a fine-tuned LLM indeed groups together similar terms
314
+ while ensuring different terms are further apart.
315
+
316
+ After identifying clusters from the above step, we use
317
+ Humans in the Loop approach to identify all terms that belong
318
+ together and run KMeans Clustering Algorithm [Lloyd,
319
+ Stuart P., 1982] on them. We identify the term closest to the
320
+ cluster's centroid, which becomes the Parent Node – one of
321
+ the core uses of Knowledge Graphs.
322
+
323
+ In the second experiment, we analyze the similarity
324
+ between the word embeddings of the synonyms of the set of
325
+ medical terms. In this step, we compute the cosine similarity
326
+ between all the word embeddings and then we examine the
327
+ similarity to demonstrate that the synonyms for the same term
328
+ are similar with a small cosine distance between them.
329
+ VI. EXPERIMENTS AND RESULTS
330
+ We use Term and General SNOMED Label columns from
331
+ COMETA dataset for our experiments. To calculate the
332
+ embeddings of every term, we use word_vector function from
333
+ biobert_embeddings package [Jitendra Jangid, 2020]. Since
334
+ the original code was incompatible with current version of
335
+ Pytorch [Paszke, A. et al., 2019] and Huggingface [Wolf et al.,
336
+ 2020], we modified it just enough to satisfy the current version
337
+ requirements – the core logic remains the same.
338
+ To test the rich representation of language models for our
339
+ use case, we perform 2 experiments, (1) Cluster the word
340
+ embeddings for the synonyms of a set of medical terms and
341
+ assign names to clusters (2) Analyze the similarity between
342
+ the word embeddings of the synonyms of the set of medical
343
+ terms.
344
+ For the reasons discussed in Sec. III, we use UMAP as our
345
+ choice of dimensionality reduction. For experiment (1), Fig. 2
346
+ shows that entities having similar nature are grouped together
347
+ and dissimilar entities are further apart which proves utility of
348
+ a Fine-tuned Language Models.
349
+
350
+ Fig 2. Clusters resulting from UMAP dimensionality reduction
351
+ Next we perform KMeans clustering on mentions
352
+ belonging to same group using cosine similarity. The centroid
353
+ of each clusters were then used to identify concepts by finding
354
+ terms that were closest to the centers by cosine similarity. We
355
+ found the following terms for the concepts visible in Table. 1.
356
+
357
+ Concept (General SNOMED
358
+ Label)
359
+ Term (closest to the cluster)
360
+ Oral contraception
361
+ hormonal BC pills
362
+ Crohn's disease
363
+ crohns disease
364
+ Diabetes mellitus type 2
365
+ T2 diabetes
366
+ Analgesic
367
+ Pain Medication
368
+ Diabetes mellitus type 1
369
+ T1 diabetic
370
+ Autoimmune disease
371
+ autoimmune disease
372
+ Hypoglycemia
373
+ low blood sugars
374
+ Headache
375
+ head pain
376
+ Tachycardia
377
+ heart racing
378
+
379
+ 10
380
+ general_snomed_label
381
+ Oral contraception
382
+ Crohn's disease
383
+ Diabetes mellitus type 2
384
+ Analgesic
385
+ Diabetes mellitus type 1
386
+ Autoimmunedisease
387
+ 5
388
+ Hypoglycemia
389
+ E
390
+ Headache
391
+ Tachycardia
392
+ Tired
393
+ 4
394
+ Itching
395
+ 0
396
+ 2
397
+ 4
398
+ 6
399
+ 8
400
+ 10
401
+ dim_0Tired
402
+ feel tired
403
+ Itching
404
+ itching
405
+ Table 1. Terms closest to the cluster center of each Concept
406
+ While Fig. 2 illustrates global and local structure among
407
+ different mentions of a concept, as a part of experiment (2),
408
+ we also analyze distribution of similarity scores (which are
409
+ calculated by using cosine similarity) to visualize distribution
410
+ of cosine similarity among terms belonging to same concept
411
+ (Fig. 3 and 4) and terms belonging to different concepts (Fig.
412
+ 5). We can see that distribution of mentions belonging to same
413
+ concept are closer to each other on average as compared to
414
+ mentions from different concepts. This point again validates
415
+ the utility of Language Model in finding different mentions of
416
+ a concept in multiple documents.
417
+
418
+ Fig 3. Cosine similarity between mentions from Oral Contraception
419
+
420
+ Fig 4. Cosine similarity between mentions from Cron’s disease
421
+
422
+ In addition to these plots, we also analyze similarity
423
+ between unrelated terms, and we see the following trend –
424
+
425
+ Fig 5. Cosine similarity between mentions from different concepts
426
+ VII. CONCLUSION AND FUTURE WORK
427
+ In this paper we have empirically shown how Language
428
+ Models fine-tuned on domain specific data can be used to
429
+ replace Knowledge Graphs for tasks where identifying
430
+ synonyms is involved.
431
+
432
+ Language Models do a very good job in calculating
433
+ embeddings which contains semantic information about
434
+ terms that can be used to identify if two terms are close to
435
+ each other or not. This information is used in this paper to
436
+ identify terms which are closer to each other, and which are
437
+ not. Once groups of similar terms have been identifying using
438
+ non-linear dimensionality techniques, using Humans in the
439
+ Loop approach we can annotate such groups. After
440
+ annotating the groups, we use KMeans to identify centroids
441
+ of each cluster which are then used the identify terms with
442
+ the closest cosine distance from them. These terms can then
443
+ be used as parent nodes for their respective clusters. The
444
+ primary way in which our algorithm improves over current
445
+ Knowledge Graph based approaches is that unlike KGs which
446
+ are created by subject matter experts, our algorithm doesn’t
447
+ require subject matter experts for annotation.
448
+
449
+ Our current algorithm handles synonym mapping quite
450
+ well, but it requires human intervention and for next steps, we
451
+ would be exploring ways in which we can extract Knowledge
452
+ Graphs from Language Models themselves. This would be
453
+ required to remove the human intervention in the current
454
+ process and handling cases where hypernyms are involved.
455
+ REFERENCES
456
+
457
+ [1] Bodenreider O. 2004. The Unified Medical Language System (UMLS):
458
+ integrating biomedical terminology. Nucleic Acids Res. 2004 Jan
459
+ 1;32(Database issue):D267-70.
460
+ [2] SNOMED. URL: http://www.snomed.org/
461
+ [3] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova.
462
+ 2019. BERT: Pre-training of Deep Bidirectional Transformers for
463
+ Language Understanding. In Proceedings of the 2019 Conference of
464
+ the North American Chapter of the Association for Computational
465
+ Linguistics: Human Language Technologies, Volume 1 (Long and
466
+ Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association
467
+ for Computational Linguistics.
468
+ [4] Tokenizer. URL:https://huggingface.co/docs/tokenizers
469
+ [5] Marco Basaldella, Fangyu Liu, Ehsan Shareghi, and Nigel Collier.
470
+ 2020. COMETA: A Corpus for Medical Entity Linking in the Social
471
+ Media. In Proceedings of the 2020 Conference on Empirical Methods
472
+ in Natural Language Processing (EMNLP), pp.3122–3137, Online.
473
+ Association for Computational Linguistics.
474
+
475
+ [6] McInnes, Leland and Healy, John and Saul, Nathaniel and Grossberger,
476
+ Lukas. 2018. UMAP: Uniform Manifold Approximation and
477
+ Projection. The Journal of Open Source Software. arXiv:1802.03426v3
478
+ [7] Karl Pearson F.R.S. , 1901. LIII. On lines and planes of closest fit to
479
+ systems of points in space. The London, Edinburgh, and Dublin
480
+ Philosophical Magazine and Journal of Science, 2(11), pp.559–572.
481
+ [8] Lloyd, Stuart P., 1982. Least squares quantization in PCM. Information
482
+ Theory, IEEE Transactions on 28.2, pp.129-137
483
+ [9] Jitendra
484
+ Jangid,
485
+ 2020.
486
+ https://github.com/Overfitter/biobert_embedding
487
+ [10] Wolf et al., 2020. Transformers: State-of-the-Art Natural Language
488
+ Processing. EMNLP
489
+ [11] Paszke, A. et al., 2019. PyTorch: An Imperative Style, High-
490
+ Performance Deep Learning Library. In Advances in Neural
491
+ Information Processing Systems 32. Curran Associates, Inc., pp. 8024–
492
+ 8035. Available at: http://papers.neurips.cc/paper/9015-pytorch-an-
493
+ imperative-style-high-performance-deep-learning-library.pdf.
494
+
495
+ Oralcontraception
496
+ Alesse
497
+ BCP
498
+ Cyclen
499
+ 0.95
500
+ Lolo
501
+ OC, s
502
+ 0.9
503
+ Qlaira
504
+ uirth control pills
505
+ 0.85
506
+ contraceptive pills
507
+ hormonal rirth control pills
508
+ 0.8
509
+ honone pill
510
+ 0.75
511
+ triphasic pills
512
+ contro
513
+ aceptlvepIlls
514
+ onal
515
+ mone
516
+ haslc
517
+ blrth control pWl:Crohn's disease
518
+ CD
519
+ Chrohns
520
+ Crohn
521
+ Crohn ' s
522
+ Crohn ' s flare
523
+ 0.95
524
+ Crohn disease
525
+ Crohn et *$
526
+ Crohn at' " s disease
527
+ 0.9
528
+ Crohnie
529
+ Crohnies
530
+ crohns
531
+ crohns colitis
532
+ 0.85
533
+ crohns disease
534
+ Crohn ' s disease
535
+ CrohnsDisease
536
+ 0.8
537
+ crohns flare
538
+ 3
539
+ olitlis
540
+ diseas eDissimilarityMatrix
541
+ hormonal μC pills
542
+ crohns disease
543
+ T2 diabetes
544
+ 0.95
545
+ Pain Medicatior
546
+ T1 diabetic
547
+ 0.9
548
+ autoimmune disease
549
+ low blood sugars
550
+ head pair
551
+ 0.85
552
+ heart racing
553
+ feel tired
554
+ 0.8
555
+ itching[12] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I.
556
+ (2019). Language Models are Unsupervised Multitask Learners.
557
+ [13] Ganesh Jawahar etal, What does BERT learn about the structure of
558
+ language? ;Proceedings of the 57th Annual Meeting of the Association
559
+ for Computational Linguistics, pages 3651–3657, 2019
560
+ [14] Wang, C., Liu, X., & Song, D.X. (2020). Language Models are Open
561
+ Knowledge Graphs. ArXiv, abs/2010.11967.
562
+ [15] Wang, X., He, Q., Liang, J., & Xiao, Y. (2022). Language Models as
563
+ Knowledge Embeddings. Proceedings of the Thirty-First International
564
+ Joint Conference on Artificial Intelligence (IJCAI-22)
565
+ [16] Lei He, Suncong Zheng, Tao Yang, and Feng Zhang. 2021. KLMo:
566
+ Knowledge Graph Enhanced Pretrained Language Model with Fine-
567
+ Grained
568
+ Relationships.
569
+ In Findings
570
+ of
571
+ the
572
+ Association
573
+ for
574
+ Computational Linguistics: EMNLP 2021, pages 4536–4542, Punta
575
+ Cana,
576
+ Dominican
577
+ Republic.
578
+ Association
579
+ for
580
+ Computational
581
+ Linguistics.
582
+
583
+
584
+
585
+
586
+
6dE2T4oBgHgl3EQfkgfV/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,324 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf,len=323
2
+ page_content='XXX-X-XXXX-XXXX-X/XX/$XX.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
3
+ page_content='00 ©20XX IEEE Language Models sounds the Death Knell of Knowledge Graphs Kunal Suri Optum, India kunal_suri@optum.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
4
+ page_content='com Swapna Sourav Rout Optum, India rout.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
5
+ page_content='swapnasourav@optum.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
6
+ page_content='com Atul Singh Optum, India atul_singh18@optum.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
7
+ page_content='com Prakhar Mishra Optum, India prakhar_mishra29@optum.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
8
+ page_content='com Rajesh Sabapathy Optum,India rajesh_sabapathy@uhc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
9
+ page_content='com Abstract—Healthcare domain generates a lot of unstructured and semi-structured text.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
10
+ page_content=' Natural Language processing (NLP) has been used extensively to process this data.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
11
+ page_content=' Deep Learning based NLP especially Large Language Models (LLMs) such as BERT have found broad acceptance and are used extensively for many applications.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
12
+ page_content=' A Language Model is a probability distribution over a word sequence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
13
+ page_content=' Self-supervised Learning on a large corpus of data automatically generates deep learning-based language models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
14
+ page_content=' BioBERT and Med-BERT are language models pre-trained for the healthcare domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
15
+ page_content=' Healthcare uses typical NLP tasks such as question answering, information extraction, named entity recognition, and search to simplify and improve processes.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
16
+ page_content=' However, to ensure robust application of the results, NLP practitioners need to normalize and standardize them.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
17
+ page_content=' One of the main ways of achieving normalization and standardization is the use of Knowledge Graphs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
18
+ page_content=' A Knowledge Graph captures concepts and their relationships for a specific domain, but their creation is time- consuming and requires manual intervention from domain experts, which can prove expensive.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
19
+ page_content=' SNOMED CT (Systematized Nomenclature of Medicine - Clinical Terms), Unified Medical Language System (UMLS), and Gene Ontology (GO) are popular ontologies from the healthcare domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
20
+ page_content=" SNOMED CT and UMLS capture concepts such as disease, symptoms and diagnosis and GO is the world's largest source of information on the functions of genes." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
21
+ page_content=' Healthcare has been dealing with an explosion in information about different types of drugs, diseases, and procedures.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
22
+ page_content=' This paper argues that using Knowledge Graphs is not the best solution for solving problems in this domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
23
+ page_content=' We present experiments using LLMs for the healthcare domain to demonstrate that language models provide the same functionality as knowledge graphs, thereby making knowledge graphs redundant.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
24
+ page_content=' Keywords—Medical data, Language Models, Natural Language Processing, Knowledge Graphs, Deep Learning I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
25
+ page_content=' INTRODUCTION Knowledge graphs (KG) are knowledge bases that capture concepts and their relationships for a specific domain using a graph-structured data model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
26
+ page_content=' Systematized Nomenclature of Medicine – Clinical Terms (SNOMED CT) (SNOMED), Unified Medical Language Systems(UMLS) [Bodenreider O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
27
+ page_content=' 2004], etc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
28
+ page_content=', are some of the popular KG in the healthcare domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
29
+ page_content=' Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
30
+ page_content=' 1 shows a sample from a representative medical entity, KG.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
31
+ page_content=' On the other hand, a language model is a probability distribution over a word sequence and is the backbone of modern natural language processing (NLP).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
32
+ page_content=" Language models try to capture any language's linguistic intuition and writing, and large language models like BERT [Devlin et al." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
33
+ page_content=', 2019] and GPT-2 [Radford et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
34
+ page_content=', 2019] have shown remarkable performance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
35
+ page_content=" The paper presents a study demonstrating that language models' ability to learn relationships among different entities makes knowledge graphs redundant for many applications." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
36
+ page_content=' This paper uses similar terms from SNOMED-CT KG and passes them through a language model for the healthcare domain BioRedditBERT to get a 768-dimensional dense vector representation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
37
+ page_content=' The paper presents the results for analyzing these embeddings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
38
+ page_content=' The experiments presented in the paper validate that similar terms cluster together.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
39
+ page_content=' The paper uses simple heuristics to assign names to clusters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
40
+ page_content=' The results show that the cluster names match the names in the KG.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
41
+ page_content=' Finally, the experiments demonstrate that the cosine similarity of vector representation of similar terms is high and vice versa.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
42
+ page_content=' Our contributions include: (i) We propose a study to demonstrate the value and application of Large Language Models (LLMs) in comparison to Knowledge Graph-based approaches for the task of synonym extraction.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
43
+ page_content=' (ii) We extensively evaluate our approach on a standard, widely accepted dataset, and the results are encouraging.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
44
+ page_content=' Fig 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
45
+ page_content=' Medical entity Knowledge Graph Representation The rest of the paper is organized as follows: Section II presents the background required to understand the work presented in this paper.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
46
+ page_content=' Section III presents a literature survey of related work on knowledge graphs and language models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
47
+ page_content=' Section IV presents our understanding of how current days language models are making knowledge graphs redundant.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
48
+ page_content=' Section V describes our proposed approach.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
49
+ page_content=' Section VI describes the experiments conducted and the results obtained.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
50
+ page_content=' Finally, section VII summarizes our work and discusses possible directions for future study.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
51
+ page_content=' II.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
52
+ page_content=' BACKGROUND This section defines and describes Language Models and Knowledge Graphs as used in this paper: Medicine Fever Allergy Dolo ClaritinA.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
53
+ page_content=' Language Models A Language Model predicts the probability of a sequence of words in a human language such as English.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
54
+ page_content=' In the equation below P(w1,…wm) is the probability of the word sequence S, where S = (w1, w2, …, wm) and wi is the ith word in the sequence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
55
+ page_content=' Large Language Models (LLMs) are language models trained on large general corpora that learn associations and relationships among different word entities in an unsupervised manner.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
56
+ page_content=' Large Language Models (LLMs) are considered universal language learners.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
57
+ page_content=' LLMs such as BERT and GPTare deep neural networks based on transformer architecture.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
58
+ page_content=' One of many reasons for the immense popularity of LLMs is that these models are pre-trained self-supervised models and can be adapted or fine-tuned to cater to a wide range of NLP tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
59
+ page_content=' Few-shot learning has enabled these LLMs to be adapted to a given NLP task using fewer training samples.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
60
+ page_content=' Another reason for the immense popularity of LLMs is that a single language model is applicable for multiple downstream applications such as Token classification, Text classification, and Question answering.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
61
+ page_content=' LLMs generate embeddings or word vectors for words, and these embeddings capture the context of the word in the corpus.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
62
+ page_content=' This ability of LLMs to generate embeddings based on the corpus makes them ubiquitous in almost NLP tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
63
+ page_content=' In this paper, we use BioRedditBERT [Basaldella et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
64
+ page_content=', 2020], a variant of BERT trained for the healthcare domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
65
+ page_content=' It is a domain-specific language representation model trained on large-scale biomedical corpora from Reddit.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
66
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
67
+ page_content=' Knowledge Graphs Knowledge Graphs (KGs) organize data and capture relationships between different entities for a domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
68
+ page_content=' Domain experts create KGs to map domain-based relations between various entities.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
69
+ page_content=' Knowledge graphs are Graph data structures with nodes and edges.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
70
+ page_content=' Nodes or vertices represent entities of interest, and edges represent relations between them, as shown in Fig 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
71
+ page_content=' KGs can map and model direct and latent relationships between entities of interest.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
72
+ page_content=' Typically, KGs are used to model and map information from model sources.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
73
+ page_content=' Once KGs are designed, typically, NLP is used to populate & create the knowledge base from unstructured text corpora.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
74
+ page_content=' Knowledge graphs play a crucial role in healthcare knowledge representation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
75
+ page_content=' There are many widely used knowledge graphs like SNOMED and UMLS etc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
76
+ page_content=' In healthcare, KGs are used for drug discovery drugs, identifying tertiary symptoms for diseases and augmented decision-making, etc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
77
+ page_content=' COMETA: A Corpus for Medical Entity Linking in social media [Basaldella et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
78
+ page_content=', 2020] – a corpus containing four years of content in 68 health-themed subreddits and annotating the most frequent with their corresponding SNOMED-CT entities.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
79
+ page_content=' In this paper, we have used COMETA to obtain synonyms from SNOMED-CT.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
80
+ page_content=' III.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
81
+ page_content=' RELATED WORK In 2019, Jawahar et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
82
+ page_content=' performed experiments to understand the underlying language structure learned by a language model like BERT [Ganesh Jawahar et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
83
+ page_content=' 2019].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
84
+ page_content=' The authors show that BERT captures the semantic information from the language hierarchically through experiments.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
85
+ page_content=' BERT captures surface features in the bottom layer, syntactic elements in the middle and semantic features in the top layer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
86
+ page_content=' The work presented in this paper treats the BERT model as a black box and demonstrates that BERT can learn the information in a knowledge graph through experiments on real-life healthcare use cases.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
87
+ page_content=' There have been studies to generate a knowledge graph directly from the output of LLMs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
88
+ page_content=' [Wang C et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
89
+ page_content=', 2020;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
90
+ page_content=' Wang X et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
91
+ page_content=' 2022] proposes a mechanism to create a KG directly from LLMs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
92
+ page_content=' This mechanism talks about a two-step mechanism to generate a KG from LLM.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
93
+ page_content=' In the first step, different candidate triplets are created from the text corpus.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
94
+ page_content=' Attention weights from a pre-trained LLM are used to get the best-matched candidate triplets and then validated through a beam search.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
95
+ page_content=' In the second stage, the matched candidate triplets are mapped to a pre-defined KG for validation, and the unmatched candidates are used to create an open knowledge graph.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
96
+ page_content=' The work demonstrates the feasibility of the idea presented in this paper that LLM can be used as a substitute for knowledge graphs, especially since they contain the information in the KG.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
97
+ page_content=' There is a body of research on integrating Knowledge graphs and LLMs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
98
+ page_content=' Structured knowledge from Knowledge Graphs is effectively integrated into Language models to enhance the pre-trained language models [Lei He et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
99
+ page_content=', 2021].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
100
+ page_content=' However, these approaches have found limited success, thereby strengthening the position in this paper that LLMs contain information from KGs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
101
+ page_content=' IV.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
102
+ page_content=' LANGUAGE MODELS FOR KNOWLEDGE GRAPHS Language Models can find associations between different words based on the attention weight matrix.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
103
+ page_content=' The methodology to use attention weights as a measure of relationship among the entities indicates that Knowledge graphs are getting replaced by LLMs as they learn more generic relationships in an unsupervised way.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
104
+ page_content=' The proposed methodology in this paper is built on this idea to demonstrate that Knowledge graphs are increasingly getting redundant for many NLP tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
105
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
106
+ page_content=" PROPOSED APPROACH The paper demonstrates that language models' ability to learn relationships among different entities makes knowledge graphs redundant for many applications." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
107
+ page_content=' To illustrate this, we have used word embeddings for all the synonyms of a set of medical terms from a large language model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
108
+ page_content=' This work uses m P(w1,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
109
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
110
+ page_content=', Wm) =|[P(wi I Wi,.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
111
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
112
+ page_content=', Wi-1) i=1COMETA data to obtain synonyms for a set of medical terms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
113
+ page_content=' In COMETA data, the work focuses on the following columns: a) Example column, which contains the sentences from health-themed forums on Reddit, b) Term column contains the medical terms present in the Example column, c) General SNOMED Label column;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
114
+ page_content=' contains the literal meaning of the Term column from the SNOMED Knowledge Graphs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
115
+ page_content=' To obtain synonyms, we use the different values from the Terms column for a specific value of the General SNOMED Label column.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
116
+ page_content=' For example, for Abdominal Wind Pain General SNOMED label, we have the following three synonyms that we can obtain from the Terms column: gas pains, painful gas, and gas pain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
117
+ page_content=' To calculate the word embeddings of every synonym term, we use the word_vector function from the biobert_embeddings python module [Jitendra Jangid, 2020].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
118
+ page_content=' Since the original code was incompatible with the current version of Pytorch [Paszke, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
119
+ page_content=' et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
120
+ page_content=', 2019] and Huggingface [Wolf et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
121
+ page_content=', 2020], we modified it just enough to satisfy the current version requirements – the core logic remains the same.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
122
+ page_content=' We tokenize every Term using HuggingFace tokenizers and pass the tokenized Term through BioRedditBERT model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
123
+ page_content=" The previous step gives us embedding for the Term (or sub-terms if the model didn't see the Term before)." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
124
+ page_content=' If the model has not seen the Term before, then we sum up the embedding of all the subterms).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
125
+ page_content=' We then store all the embeddings for the next steps.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
126
+ page_content=' We perform the following two experiments after generating the word embeddings for the synonyms of a set of medical terms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
127
+ page_content=' In the first experiment, we cluster the word embeddings for the synonyms of a set of medical terms and assign names to clusters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
128
+ page_content=' The word embeddings are passed into UMAP to generate a 2-dimensional representation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
129
+ page_content=' We plot the 2-dimensional representation to examine how the term cluster visually.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
130
+ page_content=' UMAP is used as the dimensionality reduction technique over PCA because it is a non-linear dimensionality reduction technique and does very well to preserve the local and global structure of the data as compared to PCA.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
131
+ page_content=' However, unlike PCA [Karl Pearson F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
132
+ page_content='R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
133
+ page_content='S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
134
+ page_content=' , 1901], UMAP is very sensitive to hyperparameters that we chose, so we visualize the embeddings for several values of number of neighbours (n_neighbors) and minimum distance (min_dist).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
135
+ page_content=' This step will help us visually validate that a fine-tuned LLM indeed groups together similar terms while ensuring different terms are further apart.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
136
+ page_content=' After identifying clusters from the above step, we use Humans in the Loop approach to identify all terms that belong together and run KMeans Clustering Algorithm [Lloyd, Stuart P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
137
+ page_content=', 1982] on them.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
138
+ page_content=" We identify the term closest to the cluster's centroid, which becomes the Parent Node – one of the core uses of Knowledge Graphs." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
139
+ page_content=' In the second experiment, we analyze the similarity between the word embeddings of the synonyms of the set of medical terms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
140
+ page_content=' In this step, we compute the cosine similarity between all the word embeddings and then we examine the similarity to demonstrate that the synonyms for the same term are similar with a small cosine distance between them.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
141
+ page_content=' VI.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
142
+ page_content=' EXPERIMENTS AND RESULTS We use Term and General SNOMED Label columns from COMETA dataset for our experiments.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
143
+ page_content=' To calculate the embeddings of every term, we use word_vector function from biobert_embeddings package [Jitendra Jangid, 2020].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
144
+ page_content=' Since the original code was incompatible with current version of Pytorch [Paszke, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
145
+ page_content=' et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
146
+ page_content=', 2019] and Huggingface [Wolf et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
147
+ page_content=', 2020], we modified it just enough to satisfy the current version requirements – the core logic remains the same.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
148
+ page_content=' To test the rich representation of language models for our use case, we perform 2 experiments, (1) Cluster the word embeddings for the synonyms of a set of medical terms and assign names to clusters (2) Analyze the similarity between the word embeddings of the synonyms of the set of medical terms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
149
+ page_content=' For the reasons discussed in Sec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
150
+ page_content=' III, we use UMAP as our choice of dimensionality reduction.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
151
+ page_content=' For experiment (1), Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
152
+ page_content=' 2 shows that entities having similar nature are grouped together and dissimilar entities are further apart which proves utility of a Fine-tuned Language Models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
153
+ page_content=' Fig 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
154
+ page_content=' Clusters resulting from UMAP dimensionality reduction Next we perform KMeans clustering on mentions belonging to same group using cosine similarity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
155
+ page_content=' The centroid of each clusters were then used to identify concepts by finding terms that were closest to the centers by cosine similarity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
156
+ page_content=' We found the following terms for the concepts visible in Table.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
157
+ page_content=' 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
158
+ page_content=' ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
159
+ page_content='Concept (General SNOMED ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
160
+ page_content='Label) ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
161
+ page_content='Term (closest to the cluster) ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
162
+ page_content='Oral contraception ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
163
+ page_content='hormonal BC pills ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
164
+ page_content="Crohn's disease " metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
165
+ page_content='crohns disease ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
166
+ page_content='Diabetes mellitus type 2 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
167
+ page_content='T2 diabetes ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
168
+ page_content='Analgesic ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
169
+ page_content='Pain Medication ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
170
+ page_content='Diabetes mellitus type 1 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
171
+ page_content='T1 diabetic ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
172
+ page_content='Autoimmune disease ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
173
+ page_content='autoimmune disease ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
174
+ page_content='Hypoglycemia ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
175
+ page_content='low blood sugars ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
176
+ page_content='Headache ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
177
+ page_content='head pain ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
178
+ page_content='Tachycardia ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
179
+ page_content='heart racing ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
180
+ page_content='10 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
181
+ page_content='general_snomed_label ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
182
+ page_content='Oral contraception ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
183
+ page_content="Crohn's disease " metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
184
+ page_content='Diabetes mellitus type 2 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
185
+ page_content='Analgesic ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
186
+ page_content='Diabetes mellitus type 1 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
187
+ page_content='Autoimmunedisease ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
188
+ page_content='5 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
189
+ page_content='Hypoglycemia ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
190
+ page_content='E ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
191
+ page_content='Headache ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
192
+ page_content='Tachycardia ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
193
+ page_content='Tired ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
194
+ page_content='4 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
195
+ page_content='Itching ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
196
+ page_content='0 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
197
+ page_content='2 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
198
+ page_content='4 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
199
+ page_content='6 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
200
+ page_content='8 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
201
+ page_content='10 ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
202
+ page_content='dim_0Tired ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
203
+ page_content='feel tired ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
204
+ page_content='Itching ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
205
+ page_content='itching ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
206
+ page_content='Table 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
207
+ page_content=' Terms closest to the cluster center of each Concept While Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
208
+ page_content=' 2 illustrates global and local structure among different mentions of a concept, as a part of experiment (2), we also analyze distribution of similarity scores (which are calculated by using cosine similarity) to visualize distribution of cosine similarity among terms belonging to same concept (Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
209
+ page_content=' 3 and 4) and terms belonging to different concepts (Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
210
+ page_content=' 5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
211
+ page_content=' We can see that distribution of mentions belonging to same concept are closer to each other on average as compared to mentions from different concepts.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
212
+ page_content=' This point again validates the utility of Language Model in finding different mentions of a concept in multiple documents.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
213
+ page_content=' Fig 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
214
+ page_content=' Cosine similarity between mentions from Oral Contraception Fig 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
215
+ page_content=' Cosine similarity between mentions from Cron’s disease In addition to these plots, we also analyze similarity between unrelated terms, and we see the following trend – Fig 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
216
+ page_content=' Cosine similarity between mentions from different concepts VII.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
217
+ page_content=' CONCLUSION AND FUTURE WORK In this paper we have empirically shown how Language Models fine-tuned on domain specific data can be used to replace Knowledge Graphs for tasks where identifying synonyms is involved.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
218
+ page_content=' Language Models do a very good job in calculating embeddings which contains semantic information about terms that can be used to identify if two terms are close to each other or not.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
219
+ page_content=' This information is used in this paper to identify terms which are closer to each other, and which are not.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
220
+ page_content=' Once groups of similar terms have been identifying using non-linear dimensionality techniques, using Humans in the Loop approach we can annotate such groups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
221
+ page_content=' After annotating the groups, we use KMeans to identify centroids of each cluster which are then used the identify terms with the closest cosine distance from them.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
222
+ page_content=' These terms can then be used as parent nodes for their respective clusters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
223
+ page_content=' The primary way in which our algorithm improves over current Knowledge Graph based approaches is that unlike KGs which are created by subject matter experts, our algorithm doesn’t require subject matter experts for annotation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
224
+ page_content=' Our current algorithm handles synonym mapping quite well, but it requires human intervention and for next steps, we would be exploring ways in which we can extract Knowledge Graphs from Language Models themselves.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
225
+ page_content=' This would be required to remove the human intervention in the current process and handling cases where hypernyms are involved.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
226
+ page_content=' REFERENCES [1] Bodenreider O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
227
+ page_content=' 2004.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
228
+ page_content=' The Unified Medical Language System (UMLS): integrating biomedical terminology.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
229
+ page_content=' Nucleic Acids Res.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
230
+ page_content=' 2004 Jan 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
231
+ page_content='32(Database issue):D267-70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
232
+ page_content=' [2] SNOMED.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
233
+ page_content=' URL: http://www.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
234
+ page_content='snomed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
235
+ page_content='org/ [3] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
236
+ page_content=' 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
237
+ page_content=' BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
238
+ page_content=' In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
239
+ page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
240
+ page_content=' [4] Tokenizer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
241
+ page_content=' URL:https://huggingface.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
242
+ page_content='co/docs/tokenizers [5] Marco Basaldella, Fangyu Liu, Ehsan Shareghi, and Nigel Collier.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
243
+ page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
244
+ page_content=' COMETA: A Corpus for Medical Entity Linking in the Social Media.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
245
+ page_content=' In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
246
+ page_content='3122–3137, Online.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
247
+ page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
248
+ page_content=' [6] McInnes, Leland and Healy, John and Saul, Nathaniel and Grossberger, Lukas.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
249
+ page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
250
+ page_content=' UMAP: Uniform Manifold Approximation and Projection.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
251
+ page_content=' The Journal of Open Source Software.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
252
+ page_content=' arXiv:1802.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
253
+ page_content='03426v3 [7] Karl Pearson F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
254
+ page_content='R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
255
+ page_content='S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
256
+ page_content=' , 1901.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
257
+ page_content=' LIII.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
258
+ page_content=' On lines and planes of closest fit to systems of points in space.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
259
+ page_content=' The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11), pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
260
+ page_content='559–572.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
261
+ page_content=' [8] Lloyd, Stuart P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
262
+ page_content=', 1982.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
263
+ page_content=' Least squares quantization in PCM.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
264
+ page_content=' Information Theory, IEEE Transactions on 28.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
265
+ page_content='2, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
266
+ page_content='129-137 [9] Jitendra Jangid, 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
267
+ page_content=' https://github.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
268
+ page_content='com/Overfitter/biobert_embedding [10] Wolf et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
269
+ page_content=', 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
270
+ page_content=' Transformers: State-of-the-Art Natural Language Processing.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
271
+ page_content=' EMNLP [11] Paszke, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
272
+ page_content=' et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
273
+ page_content=', 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
274
+ page_content=' PyTorch: An Imperative Style, High- Performance Deep Learning Library.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
275
+ page_content=' In Advances in Neural Information Processing Systems 32.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
276
+ page_content=' Curran Associates, Inc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
277
+ page_content=', pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
278
+ page_content=' 8024– 8035.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
279
+ page_content=' Available at: http://papers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
280
+ page_content='neurips.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
281
+ page_content='cc/paper/9015-pytorch-an- imperative-style-high-performance-deep-learning-library.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
282
+ page_content='pdf.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
283
+ page_content=' Oralcontraception Alesse BCP Cyclen 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
284
+ page_content='95 Lolo OC, s 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
285
+ page_content='9 Qlaira uirth control pills 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
286
+ page_content='85 contraceptive pills hormonal rirth control pills 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
287
+ page_content='8 honone pill 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
288
+ page_content="75 triphasic pills contro aceptlvepIlls onal mone haslc blrth control pWl:Crohn's disease CD Chrohns Crohn Crohn ' s Crohn ' s flare 0." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
289
+ page_content='95 Crohn disease Crohn et *$ Crohn at\' " s disease 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
290
+ page_content='9 Crohnie Crohnies crohns crohns colitis 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
291
+ page_content="85 crohns disease Crohn ' s disease CrohnsDisease 0." metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
292
+ page_content='8 crohns flare 3 olitlis diseas eDissimilarityMatrix hormonal μC pills crohns disease T2 diabetes 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
293
+ page_content='95 Pain Medicatior T1 diabetic 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
294
+ page_content='9 autoimmune disease low blood sugars head pair 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
295
+ page_content='85 heart racing feel tired 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
296
+ page_content='8 itching[12] Radford, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
297
+ page_content=', Wu, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
298
+ page_content=', Child, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
299
+ page_content=', Luan, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
300
+ page_content=', Amodei, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
301
+ page_content=', & Sutskever, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
302
+ page_content=' (2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
303
+ page_content=' Language Models are Unsupervised Multitask Learners.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
304
+ page_content=' [13] Ganesh Jawahar etal, What does BERT learn about the structure of language?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
305
+ page_content=' ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
306
+ page_content='Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651–3657, 2019 [14] Wang, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
307
+ page_content=', Liu, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
308
+ page_content=', & Song, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
309
+ page_content='X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
310
+ page_content=' (2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
311
+ page_content=' Language Models are Open Knowledge Graphs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
312
+ page_content=' ArXiv, abs/2010.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
313
+ page_content='11967.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
314
+ page_content=' [15] Wang, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
315
+ page_content=', He, Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
316
+ page_content=', Liang, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
317
+ page_content=', & Xiao, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
318
+ page_content=' (2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
319
+ page_content=' Language Models as Knowledge Embeddings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
320
+ page_content=' Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI-22) [16] Lei He, Suncong Zheng, Tao Yang, and Feng Zhang.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
321
+ page_content=' 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
322
+ page_content=' KLMo: Knowledge Graph Enhanced Pretrained Language Model with Fine- Grained Relationships.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
323
+ page_content=' In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4536–4542, Punta Cana, Dominican Republic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
324
+ page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/6dE2T4oBgHgl3EQfkgfV/content/2301.03980v1.pdf'}
79AzT4oBgHgl3EQfSPsz/content/2301.01228v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1cfff4374d9cdaf117b05aad9fb5bcf8b085a14eec1bcb89685a78849129bec8
3
+ size 150273
79AzT4oBgHgl3EQfSPsz/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3b51221e0a419e4ba3811e2ecd039d0d35a2796526a866dae15412668a6574e
3
+ size 852013
79AzT4oBgHgl3EQfSPsz/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:de5c2b7cbf9d176267e428642f0750264d420c949ffa728990966e0ff8c2f4ac
3
+ size 34668
7NE1T4oBgHgl3EQfnATd/content/tmp_files/2301.03305v1.pdf.txt ADDED
@@ -0,0 +1,1613 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Topological charge quantization on localized imperfections in crystalline insulators
2
+ and the nearsightedness principle of Kohn
3
+ Kiryl Piasotski1, 2, ∗
4
+ 1Institut f¨ur Theorie der Kondensierten Materie,
5
+ Karlsruher Institut f¨ur Technologie, 76131 Karlsruhe, Germany
6
+ 2Institut f¨ur QuantenMaterialien und Technologien,
7
+ Karlsruher Institut f¨ur Technologie, 76021 Karlsruhe, Germany†
8
+ (Dated: January 10, 2023)
9
+ We study the quantization of the excess charge on N localized (ultra-screened) impurities in d-
10
+ dimensional crystalline insulating systems. Solving Dyson’s equation, we demonstrate that such
11
+ charges are topological, by expressing them as winding numbers of appropriate functionals of bulk
12
+ position space Green’s functions. We discuss the ties of our topological invariant with the nearsight-
13
+ edness principle of W. Kohn, stating that the electronic charge density at fixed chemical potential
14
+ depends on the external field only locally, meaning that localized perturbations by external fields
15
+ may only result in localized charge redistributions. We arrive at the same conclusion by demonstrat-
16
+ ing that an adiabatic perturbation comprised of a variation of impurities’ positions and/or strengths
17
+ may only result in the change in the occupancy of impurity-localized bound states sitting, energy-
18
+ wise, close to the Fermi level. Finally, we conclude by discussing the relations of the nearsightedness
19
+ principle with the topological invariants characterizing the boundary charge.
20
+ I.
21
+ INTRODUCTION
22
+ With the discovery of the quantum Hall effect [1] and
23
+ its topological origins [2, 3] the study of the topological
24
+ structures in condensed matter systems became a style
25
+ rather than a fashion. The possibly best-known contem-
26
+ porary example of that is the field of topological insu-
27
+ lators, where the boundary-localized edge states, with
28
+ topologically granted existence and robustness, are be-
29
+ ing studied (popular reviews are [4, 5]). Despite being a
30
+ well-defined endeavor with its very own periodic table [6],
31
+ this discipline leaves a number of questions open. One
32
+ of these regards the direct experimental accessibility of
33
+ these topological surface states. In particular, aside from
34
+ the subspace of such states, their basis is incomplete for a
35
+ description of the physical system accommodating them
36
+ – a topological insulator, making it highly questionable
37
+ whether an actual physical observable may be expanded
38
+ into a basis of intra- and inter-surface state transition
39
+ operators. This is, for example, not true of the excess
40
+ charge density that these insulators accumulate at their
41
+ boundaries as it, as such, also features the exponentially
42
+ localized contributions of all of the occupied extended
43
+ states. Despite that, it is clear that as the surface states
44
+ also contribute to such an observable, the change in their
45
+ occupancy has to have an observable effect.
46
+ In a series of recent works [7, 8, 9, 10], the topologi-
47
+ cal properties of the boundary-localized electronic excess
48
+ charges (the boundary charges) in unidimensional crys-
49
+ tals were examined. In particular, a pair of topological
50
+ invariants characterizing the boundary charge upon two
51
+ bulk energy spectrum-preserving transformations of crys-
52
+ tal’s potential, translations and local inversions, were de-
53
+ vised. Specifically, it was demonstrated that upon local
54
+ inversion (inversion of coordinates within the unit cell),
55
+ the boundary charge maps to its negative, up to an inte-
56
+ gral topological quantum number known as the interface
57
+ invariant. Likewise, upon the lattice translation by xϕ,
58
+ the boundary charge was shown to grow linearly with
59
+ the shift variable xϕ (with the slope being the unit cell-
60
+ averaged average charge density in the bulk ¯ρ = ν
61
+ L, with
62
+ ν – filling factor and L-system’s period), whilst perform-
63
+ ing discontinuous downward jumps by a unit of the elec-
64
+ tron charge, as quantified by another topological quan-
65
+ tum number – the boundary invariant. These topologi-
66
+ cal invariants were shown to be generated by the spectral
67
+ flow of the energies corresponding to the edge states in-
68
+ side the energy gap that hosts the chemical potential, in
69
+ complete analogy with the integer quantum Hall effect
70
+ [3]. As opposed to the edge states in topological insu-
71
+ lators, the quantization of these invariants does not rely
72
+ on the internal symmetries of the bulk Bloch’s Hamilto-
73
+ nian (such as particle-hole or time-reversal symmetries)
74
+ and is instead guaranteed by a number of fundamental
75
+ physical principles, such as charge conservation, Pauli
76
+ principle, and the nearsightedness principle of W. Kohn
77
+ [11, 12, 13, 14] (to be discussed further on). Moreover,
78
+ these invariants are directly linked with the properties of
79
+ an experimental observable, a privilege shared by both
80
+ the quantum Hall effect and the topological defects (see
81
+ Ref. [15] for a review), while not being entirely clear in
82
+ the domain of the topological insulators.
83
+ Further, in a different paper [16], rational quantization
84
+ of boundary and interface charges was discussed. Partic-
85
+ ularly, with the aid of the aforementioned physical princi-
86
+ ples, a general framework for studying quantized charges
87
+ in one dimension was laid down, allowing us to quantify
88
+ all possible quantization patterns of the boundary charge
89
+ in terms of the non-symmorphic symmetries of the crys-
90
+ tal. The charges on the interfaces between pairs of in-
91
+ sulators sharing their bulk properties were demonstrated
92
+ to follow a lattice version of the Goldstone-Wilczek for-
93
+ mula [17], relating the interface charge to the sum of the
94
+ arXiv:2301.03305v1 [cond-mat.mes-hall] 9 Jan 2023
95
+
96
+ 2
97
+ boundary charges right and left to their septum, mod-
98
+ ulo an unknown integer generated by the local coupling
99
+ between the two subsystems.
100
+ The key feature of the method developed in Ref. [16]
101
+ is this “modulo an unknown integer” paradigm, arising
102
+ from the nearsightedness principle of the electronic mat-
103
+ ter. As such, the nearsightedness principle tells us that
104
+ (see Ref. [13, 14]), in insulators, localized perturbations
105
+ by external fields may result in localized charge redistri-
106
+ butions only. To be more specific, the corrections beyond
107
+ the characteristic length scale ξg = vF
108
+ Eg (where vF and Eg
109
+ are the Fermi velocity and the gap opening up at the
110
+ Fermi level, see Ref. [18] for example) are exponentially
111
+ suppressed.
112
+ An even further refinement of this state-
113
+ ment would be that such perturbations may only remove
114
+ or add an additional number of bound states whose wave
115
+ functions are localized around the corresponding pertur-
116
+ bations. One of the key purposes of the present paper
117
+ is to substantiate this claim mathematically, which turns
118
+ out to be possible in pretty general d-dimensional models.
119
+ To be more specific, this paper concerns the topolog-
120
+ ical properties of the electronic excess charges accumu-
121
+ lated around point-like defects in d-dimensional insula-
122
+ tors. Although we purposefully specify the Hamiltonian
123
+ of the crystal under consideration to make our exposi-
124
+ tion more transparent, the derivations presented in this
125
+ manuscript are shown to be independent of its choice.
126
+ What indeed matters is that the spectrum of the clean
127
+ system consists of the energy bands occasionally sepa-
128
+ rated by the energy gaps, that is, there exists at least
129
+ one bulk energy gap into which we can put the chemical
130
+ potential to promote the resulting statistical system into
131
+ an insulator.
132
+ Furthermore, neither we specify the internal structure
133
+ of the impurity vertices, nor do we assume any particu-
134
+ lar arrangement of them, making our analysis applicable
135
+ to a wide range of experimental setups. In particular,
136
+ quite conventionally, we may assume that a number of
137
+ randomly located point-like impurities exerting an ultra-
138
+ screened electrostatic force on the system’s electrons are
139
+ scattered through the charge sampling region of a crystal
140
+ under consideration. A slightly less familiar situation is
141
+ inspired by the work of Nomura and Nagaosa [19] and
142
+ may be formulated as follows. Assuming that a crystal is
143
+ further magnetic, we know that, in an insulating regime,
144
+ its ground state may accurately be described by a Heisen-
145
+ berg model that, by itself, features topological defects.
146
+ A familiar example of such a defect would be a magnetic
147
+ skyrmion or a hedgehog texture. Assuming that the total
148
+ spin of atoms comprising our crystal is large, these tex-
149
+ tures may be seen as an arrangement of classical magnetic
150
+ moments nailed down to the atomic positions. Their in-
151
+ teraction with the electron’s spin degree of freedom may
152
+ then be written as a sum of the Zeeman-like terms, each
153
+ weighted with the Dirac δ-function centered at the posi-
154
+ tion of the corresponding atom.
155
+ Quite generically, we show that the total electronic
156
+ excess charge accumulated around these defects is an
157
+ integer-valued topological invariant, which we express
158
+ as a contour integral winding number of an appropriate
159
+ functional of bulk position space Green’s functions. Fur-
160
+ ther analysis of this topological quantum number reveals
161
+ that upon an adiabatic modification of positions and/or
162
+ vertex functions of the localized scattering centers, the
163
+ value of the invariant may only be affected by the change
164
+ in the occupancy of the imperfection-localized bound
165
+ states in the process of the spectral flow of their eigenen-
166
+ ergies inside the chemical potential-accommodating en-
167
+ ergy gap. This observation allows for an immediate in-
168
+ terpretation in terms of the nearsightedness principle,
169
+ as well as for a direct read-off of the central memo of
170
+ Ref.
171
+ [16]: “localized perturbations in insulators result
172
+ in localized charge redistribution, leading to an addi-
173
+ tion/removal of the corresponding perturbation-localized
174
+ bound states to/from the occupied spectral region”. We
175
+ conclude our analysis by commenting on the relation be-
176
+ tween the nearsightedness principle and the topological
177
+ invariants characterizing the boundary charge.
178
+ In what follows, we set the reduced Plank’s constant ̵h
179
+ and the electron charge e equal to unity ̵h = e = 1.
180
+ II.
181
+ ADIABATIC RESPONSE OF THE EXCESS
182
+ CHARGE TO LOCALIZED PERTURBATIONS IN
183
+ AN INSULATING STATE
184
+ A.
185
+ A translationally invariant model
186
+ In the following, we shall specifically refer to an elec-
187
+ tronic system governed by the following Hamiltonian
188
+ H(0)
189
+ x
190
+ = p2
191
+ 2m + 1
192
+ 2m
193
+ d
194
+
195
+ j=1
196
+ { ˜Aj(x),pj} + V (x),
197
+ (1)
198
+ with V (x) and ˜Aj(x), j = 1, ..., d being the lattice
199
+ periodic Nc × Nc Hermitian matrices. More specifically,
200
+ {V (x)
201
+ ˜A(x)} = {V (x + Rm)
202
+ ˜A(x + Rm)},
203
+ ∀m ∈ Zd,
204
+ (2)
205
+ where Rm = ∑d
206
+ j=1 mjaj is a lattice vector characterized
207
+ by a d-dimensional vector of integers m = (m1 ⋯ md)
208
+ T ,
209
+ specifying its components in the basis of primitive vec-
210
+ tors {aj}j spanning the unit cell of a Bravais lattice.
211
+ Furthermore, p and x are vectorial momentum and po-
212
+ sition operators comprised of the individual components
213
+ pj = −i ∂
214
+ ∂xj and xj.
215
+ This model naturally generalizes the one recently stud-
216
+ ied in Ref. [10] in connection with the universal prop-
217
+ erties of one-dimensional boundary charge, to higher
218
+ dimensions.
219
+ We remark that other models of multi-
220
+ dimensional periodic structures [20] are expected to share
221
+ the same physics, as the effects we are about to describe
222
+ are rather generic to an insulating state.
223
+ Translationally invariant systems are characterized by
224
+ their band structure, comprised of the individual energy
225
+
226
+ 3
227
+ bands dispersing as ϵα,k, α = 1, 2, ..., as a function
228
+ of the vectorial quasimomentum variable k, confined to
229
+ the first Brillouin zone of the reciprocal space.
230
+ The
231
+ eigenstates of the Hamiltonian to which ϵα,k are the
232
+ corresponding eigenvalues are known as Bloch functions
233
+ ψα,k(x), and may be generically expressed as
234
+ ψα,k(x) = eik⋅xuα,k(x),
235
+ (3)
236
+ where uα,k(x) in the Nc-component object and is lattice
237
+ periodic in the same sense as vector and scalar potentials
238
+ are uα,k(x) = uα,k(x+Rm), ∀m ∈ Zd. The completeness
239
+ and identity resolution relations may be written as
240
+ VUC
241
+ (2π)d ∫Rd d(d)xψ†
242
+ α,k(x)ψα′,k′(x) = δα,α′δ(d)(k − k′),
243
+ (4)
244
+ VUC
245
+ (2π)d
246
+
247
+
248
+ α=1∫BZ d(d)kψα,k(x)ψ†
249
+ α,k(x′) = 1Ncδ(d)(x − x′),
250
+ (5)
251
+ where VUC is the volume of the unit cell, defined via
252
+ VUC = ∫UC d(d)x = det(a1∣ ⋯ ∣ad).
253
+ (6)
254
+ When studying charge, it is more convenient to intro-
255
+ duce the retarded single-particle Green’s function, con-
256
+ taining the information on both the eigenstates and the
257
+ energy spectrum.
258
+ In thermodynamic equilibrium, the
259
+ Laplace image of the latter is defined as the resolvent
260
+ of the single-particle Hamiltonian (1)
261
+ [z − H(0)
262
+ x ]G(0)(x,x′) = 1Ncδ(d)(x − x′),
263
+ (7)
264
+ where z is the complex energy variable, defined in terms
265
+ of the physical frequency variable ω as z = ω + iη, where
266
+ η → 0+. Owing to the identity resolution relation (5) we
267
+ can establish the conventional Lehmann representation
268
+ G(0)(x,x′) = VUC
269
+ (2π)d
270
+
271
+
272
+ α=1∫BZ d(d)k
273
+ ψα,k(x)ψ†
274
+ α,k(x′)
275
+ z − ϵα,k
276
+ . (8)
277
+ Further, using the completeness of the basis (4), in Ap-
278
+ pendix A, we establish the following important fusion
279
+ rule for the bare propagators
280
+ ∫Rd d(d)x′G(0)(x,x′)G(0)(x′,x′′) = − ∂
281
+ ∂ω G(0)(x,x′′).
282
+ (9)
283
+ As it is shown in Appendix A, this relation holds pretty
284
+ generally, without any reference to the Hamiltonian (1).
285
+ B.
286
+ Localized perturbations and Dyson’s equation
287
+ Now we perturb the translationally invariant (on the
288
+ scale of the unit cell) system by a finite number of point-
289
+ like impurities
290
+ ˜V (x) =
291
+ N
292
+
293
+ n=1
294
+ ˜V (n)
295
+ 0
296
+ δ(d)(x − xn),
297
+ (10)
298
+ where ˜V (n)
299
+ 0
300
+ are Nc × Nc matrices describing the action
301
+ of the nth impurity on the channel space. This action is
302
+ further assumed to be local as prescribed by Dirac delta-
303
+ function δ(d)(x − xn) centered at the impurity position
304
+ xn.
305
+ Let us remark that the problem of a Dirac delta-
306
+ function potential is well-known to be ill-defined in spa-
307
+ tial dimensions higher than d = 1.
308
+ In our analysis,
309
+ this is manifested in the ill-definiteness of the bulk po-
310
+ sition space Green’s function at equal spatial arguments
311
+ G(0)(x,x) due to the divergence of the defining integrals
312
+ (8) in the ultraviolet.
313
+ Such a divergence is not physi-
314
+ cal and has to be circumvented by an appropriate reg-
315
+ ularization scheme.
316
+ In particular, in the metallic case
317
+ ˜A(x) = 0, V (x) = 0, in d > 1 the problem of the delta-
318
+ potential has been extensively studied in both physical
319
+ [21, 22, 23] and mathematical [24] literature and several
320
+ meaningful regularization techniques were proposed and
321
+ shown to produce physically sensible results. Since the
322
+ presence of the energy gaps is of no importance in the
323
+ deep ultraviolet regime, the same methods may be ap-
324
+ plied in our case.
325
+ The Dyson’s equation for the full Green’s function of
326
+ the system is given by
327
+ G(x,x′) =G(0)(x,x′)
328
+ +
329
+ N
330
+
331
+ n=1
332
+ G(0)(x,xn) ˜V (n)
333
+ 0
334
+ G(xn,x′).
335
+ (11)
336
+ First we want to consistently solve for the functions
337
+ G(xn,x′), n = 1, ..., N. This problem is brought to
338
+ the solution of the following matrix equation
339
+ M(z)D(x′) = D(0)(x′),
340
+ (12)
341
+ where M(z) is the Nc ⋅ N × Nc ⋅ N block matrix defined
342
+ by
343
+ M(z) =1Nc⋅N − G(0)(z)˜V0,
344
+ (13)
345
+ (G(0)(z))n,n′ =G(0)(xn,xn′), (˜V0)n,n′ = δn,n′ ˜V (n)
346
+ 0
347
+ . (14)
348
+ Likewise, D(x′) and D(0)(x′) are the Nc ⋅ N × Nc matri-
349
+ ces comprised of the full G(xn,x′) and bare G(0)(xn,x′)
350
+ propagators, respectively. With these notations we ob-
351
+ tain
352
+ G(x,x′) =G(0)(x,x′) + D(0)†(x)˜V0D(x′)
353
+ =G(0)(x,x′) + D(0)†(x)˜V0M−1(z)D(0)(x′),
354
+ (15)
355
+ where in our definition the Hermitian conjugate does not
356
+ affect the z-variable, i.e.
357
+ (G(0)(x,x′))† = G(0)(x′,x).
358
+ (16)
359
+
360
+ 4
361
+ C.
362
+ Measuring the excess charge
363
+ We define the excess charge density operator in the
364
+ following manner:
365
+ δ̂ρ(x) = ̂ρ(x) − ¯ρ,
366
+ (17)
367
+ where
368
+ ̂ρ(x) = ̂ψ†(x)̂ψ(x),
369
+ (18)
370
+ is the density operator, expressed in terms of the Nc-
371
+ component fermionic field operators ̂ψ(x) and ̂ψ†(x).
372
+ The field operators ̂ψ(x) and ̂ψ†(x) are further assumed
373
+ to destroy/create excitations of the full Hamiltonian in-
374
+ cluding the effect of localized scattering centers in Eq.
375
+ (10). The constant contribution ¯ρ describes the unit cell-
376
+ averaged average charge density in the bulk:
377
+ ¯ρ = 1
378
+ VUC ∫VUC
379
+ d(d)xρ(0)(x),
380
+ (19)
381
+ ρ(0)(x) =⟨̂ψ(0)†(x)̂ψ(0)(x)⟩
382
+ = − 1
383
+ π Im∫
384
+ µ
385
+ −∞ dωtr{G(0)(x,x)},
386
+ (20)
387
+ where the field operators ̂ψ(0)(x) and ̂ψ(0)†(x) describe
388
+ the excitations of the translationally invariant system, µ
389
+ denotes the chemical potential, and G(0)(x, x′) is the
390
+ bare Green’s function defined by Eqs. (7) and (8).
391
+ We measure the excess charge with the help of the clas-
392
+ sical device, described by the envelope function f(x) (see
393
+ Refs. [7, 8, 9, 10] and Ref. [25] for similar definitions).
394
+ To be more specific, we define the excess charge operator
395
+ as
396
+ δ ̂Q = ∫Rd d(d)xf(x)δ̂ρ(x).
397
+ (21)
398
+ It is sensible to define the function f(x) relative to a
399
+ certain point xp, to which the charge probe is applied,
400
+ and further assume that the charge is sampled equiva-
401
+ lently in all directions f(x) = f(∣x − xp∣). Additionally,
402
+ we assume that all of the charge f(∣x − xp∣) ≈ 1 is sam-
403
+ pled in sufficiently large vicinity of the sampling point
404
+ xp, while the envelope function smoothly decays to zero
405
+ f(∣x − xp∣) → 0 far away from xp. For that matter, it is
406
+ convenient to choose
407
+ f(∣x − xp∣) = 1 − Θlp(∣x − xp∣ − Lp),
408
+ (22)
409
+ where Θlp(∣x − xp∣ − Lp) is some representation of the
410
+ Heaviside function broadened by lp. The length scales
411
+ characteristic of the charge probe are assumed to satisfy
412
+ Lp ≫ lp ≫ ξg,
413
+ (23)
414
+ where ξg ≃ vF
415
+ Eg is the charge localization length in an insu-
416
+ lator (also it is the charge correlation length, defining the
417
+ exponential decay length of the density-density correla-
418
+ tion function, see Ref. [18]), roughly defined as the ratio
419
+ between the Fermi velocity vF and size of the energy gap
420
+ at the Fermi level Eg.
421
+ D.
422
+ Topological invariant characterizing the excess
423
+ charge
424
+ Let us assume that N impurities, as characterized by
425
+ the potential (10), are placed in a region of a crystal
426
+ falling into the sampling district of the envelope function
427
+ ∣x∣ ≲ Lp. We define the total excess charge as the zero
428
+ temperature expectation value of the excess charge op-
429
+ erator in the grandcanonical equilibrium density matrix,
430
+ so that
431
+ δQ =⟨δ ̂Q⟩ = ∫Rd d(d)xf(x)(ρ(x) − ¯ρ),
432
+ (24)
433
+ ρ(x) = − 1
434
+ π Im∫
435
+ µ
436
+ −∞ dωtr{G(x,x)}.
437
+ (25)
438
+ With the help of the representation (15), we obtain
439
+ δQ = Q′ + QP ,
440
+ (26)
441
+ where Q′ contains the Friedel charge as well as the charge
442
+ due to the impurity-localized bound states
443
+ Q′ = ∫Rd d(d)xf(x)ρ′(x),
444
+ (27)
445
+ ρ′(x) = − 1
446
+ π Im∫
447
+ µ
448
+ −∞ dωtr{D(0)†(x)˜V0M−1(z)D(0)(x)},
449
+ (28)
450
+ while QP is the so-called polarization charge given by
451
+ QP = ∫Rd d(d)xf(x)(ρ(0)(x) − ¯ρ),
452
+ (29)
453
+ and, with the help of the properties of the envelope func-
454
+ tion, is shown to be zero QP = 0 in Appendix B. It hence
455
+ follows that
456
+ δQ =Q′ = − 1
457
+ π Im∫Rd d(d)xf(x)
458
+ × ∫
459
+ µ
460
+ −∞ dωtr{D(0)†(x)˜V0M−1(z)D(0)(x)}.
461
+ (30)
462
+ Due to the branch cuts and poles of the T-matrix
463
+ T(x,x′) = ∑n,n′[˜V0M−1(z)]n,n′δ(x − xn)δ(x′ − xn′), the
464
+ integrand of the outer integral is exponentially sup-
465
+ pressed ∼ e−∣x∣/ξg at large x, allowing us to set f(x) = 1.
466
+ Interchanging the order of the integrals, we consider
467
+ ∫Rd d(d)xtr{D(0)†(x)˜V0M−1(z)D(0)(x)}
468
+ = −
469
+ N
470
+
471
+ n,n′=1
472
+ tr{[M−1(z)]n,n′ ∂
473
+ ∂ω G(0)(xn′,xn) ˜V (n)
474
+ 0
475
+ }
476
+ =
477
+ N
478
+
479
+ n,n′=1
480
+ tr{[M−1(z)]n,n′ ∂
481
+ ∂ω [M(z)]n′,n}
482
+ =
483
+ N
484
+
485
+ n=1
486
+ tr{[M−1(z) ∂
487
+ ∂ω M(z)]
488
+ n,n
489
+ }
490
+ = ∂
491
+ ∂ω tr{log M(z)} = ∂
492
+ ∂ω log det{M(z)},
493
+ (31)
494
+
495
+ 5
496
+ where, in the last line, trace and determinant of the full
497
+ Nc ⋅N ×Nc ⋅N block matrix M(z) are understood. Using
498
+ the result in Eq. (31), we arrive at the following compact
499
+ formula for the total excess charge
500
+ δQ = − 1
501
+ π Im∫
502
+ µ
503
+ −∞ dω ∂
504
+ ∂ω log det{M(z)}.
505
+ (32)
506
+ To see why the integral in Eq. (32) may take on in-
507
+ tegral values only, in Appendix C we find an alternative
508
+ contour integral representation
509
+ δQ = − ∮C
510
+ dz
511
+ 2πi
512
+
513
+ ∂z log det{M(z)},
514
+ (33)
515
+ where C is an arbitrary non-self-intersecting curve that
516
+ crosses the real axis at two points only, below the low-
517
+ est eigenvalue of the full Hamiltonian and at the chemi-
518
+ cal potential µ, and the direction of C is assumed to be
519
+ clockwise.
520
+ In the representation (33), the excess charge δQ is nec-
521
+ essarily an integer as it is expressed as a contour integral
522
+ winding number and the chemical potential is by def-
523
+ inition inside one of the energy gaps (we focus on the
524
+ insulating systems solely). In other words, the integral
525
+ in Eq. (33) measures the degree of the mapping S1 → S1
526
+ and is thus a member of the only non-trivial homotopy
527
+ group of the unit circle π1(S1) = Z.
528
+ In particular, the integral in Eq.
529
+ (33), is a sum
530
+ of two distinct contributions:
531
+ the contribution of the
532
+ branch cuts corresponding to the extended or scatter-
533
+ ing states, and the contribution of poles corresponding
534
+ to the imperfection-localized bound states.
535
+ The bands in multidimensional (d > 1) and/or mul-
536
+ tichannel (Nc > 1) systems are typically composite, i.e.
537
+ overlapping with one another along the frequency axis.
538
+ For that matter, it is convenient to choose the branch
539
+ cuts to connect the bottom of the lowest sub-band with
540
+ the top of the highest one, within every patch of the en-
541
+ ergy bands surrounded by a pair of energy gaps.
542
+ The bound state poles, determined as a solution of
543
+ det{M(z)}∣z∈R = 0, are located on the complement of
544
+ the bare Hamiltonian’s spectrum, i.e. inside the energy
545
+ gaps and, in some cases (e.g. an attractive scalar impu-
546
+ rity), below the bottom of the lowest energy band of the
547
+ unperturbed Hamiltonian.
548
+ III.
549
+ RELATION WITH THE
550
+ NEARSIGHTEDNESS PRINCIPLE
551
+ A.
552
+ Discussion
553
+ Now we would like to discuss the topological invari-
554
+ ant (33) in greater detail. In what follows, we specify
555
+ the contour C as a rectangle of length µ − B in the real
556
+ direction and width 2η in the imaginary one.
557
+ Here B
558
+ is by definition an energy lying below the lowest eigen-
559
+ value of the total Hamiltonian Hx = H(0)
560
+ x
561
+ + ˜V (x) (i.e.
562
+ c)
563
+ b)
564
+ a)
565
+ Im{z}
566
+ Re{z}
567
+ C
568
+ FIG. 1. A schematic illustration of how the spectral flow of the
569
+ energies of the imperfection-circumscribing bound states sit-
570
+ ting inside the gap that accommodates the Fermi level affects
571
+ the total excess charge. The spectrum of the system is visu-
572
+ alized through the local spectral density as looked down on
573
+ the complex frequency plane. The occupied part of the spec-
574
+ trum is demonstrated in blue, while the yellow color marks its
575
+ complement (the states of the system that are unoccupied).
576
+ Panel a) shows a rectangular contour C encircling the occu-
577
+ pied spectral region. Panels b) and c) show the zoomed-in
578
+ vicinity of the chemical potential before and after the per-
579
+ turbation. As is demonstrated in panel c), the spectral flow
580
+ results in the removal of a single bound state, carrying away
581
+ a unity of the electron charge from the system (an inverse
582
+ process is of course also possible).
583
+ B ∈ (−∞,min{spec{Hx}})), and η is not necessarily an
584
+ infinitesimal positive but is rather a finite positive num-
585
+ ber (which is allowed as the integral is invariant under
586
+ such contour deformations (see Appendix C)). Further-
587
+ more, we assume that the chemical potential is located
588
+ above the νth bulk energy band.
589
+ Let us now consider making an adiabatic perturba-
590
+ tion to the system that is comprised of the change in the
591
+ positions {xn}n and/or vertex functions { ˜V (n)
592
+ 0
593
+ }n of the
594
+ impurities. As the span of the extended states’ energy
595
+ bands is unaffected by such adiabatic perturbations, the
596
+ branch cut contribution to the winding number remains
597
+ invariant (up to the cases when the bound state merges
598
+ with the band, as discussed below). This remark is essen-
599
+ tially true as such deformations of the parameter space
600
+ do not change the analytical structure of G(0)(xn, xn′),
601
+ through the functionals of which alone our topological
602
+ invariant is expressed.
603
+ We hence conclude that such
604
+ changes may only unleash themselves in the spectral flow
605
+ of the bound state energies.
606
+ As was anticipated in Section II D, the bound state
607
+ energies are energy-wise located inside the energy gaps
608
+ of the bulk system.
609
+ This assertion also regards the
610
+ energy gap below the bottom of the lowest band ω ∈
611
+ (−∞, mink ϵ1,k], which can accommodate the bound
612
+ states in the case of attractive impurities, for example.
613
+
614
+ 6
615
+ The energies of the bound states ϵbs inside the energy
616
+ gaps [maxk ϵα,k, mink ϵα+1,k] surrounded by a pair of
617
+ bands α, α + 1, (α = 1, ..., ν − 1), are solely character-
618
+ ized by their location within the gap. The same holds
619
+ true for the infinite gap below the bottom of the lowest
620
+ bulk energy band, with ϵbs now being energy-wise located
621
+ in (−∞, mink ϵ1,k]. This implies that the spectral flow
622
+ of these energies is constituted in the motion of ϵbs in be-
623
+ tween the top of ϵα,k and the bottom of ϵα+1,k, or between
624
+ the negative infinity and mink ϵ1,k shall some states be
625
+ also found in there. When merging with one of the energy
626
+ bands (either ϵα,k or ϵα+1,k, and ϵ1,k solely when consid-
627
+ ering the gap preceding the entire band structure), the
628
+ value of the contour integral winding number (33) relat-
629
+ ing to that band gets modified by unity26. It follows that
630
+ the motion of the bound state poles, inside such energy
631
+ gaps below the one hosting the chemical potential, has
632
+ absolutely no effect on the topological invariant (33) (one
633
+ may see this result as a form of charge conservation), as
634
+ B, by definition, resides below the lowest pole (effectively
635
+ meaning that none of the states are allowed to escape the
636
+ occupied spectral region from below).
637
+ The flow of the energies of the impurity-localized
638
+ bound states residing inside the gap separating the con-
639
+ duction and the valence bands apart (the gap where the
640
+ chemical potential is located), on the other hand, affects
641
+ the winding number in Eq. (33). When a bound state
642
+ crosses the chemical potential from above or below, the
643
+ number of poles encompassed by the integration contour
644
+ increases or decreases correspondingly. That means that
645
+ the unit of the electron charge gets either pumped in or
646
+ out of the system, modifying the topological invariant by
647
+ ±1. This discussion is summarized in Fig. 1.
648
+ The elaboration above allows us to draw the following
649
+ physical conclusion:
650
+ Localized adiabatic perturbations in insulators, may only
651
+ result in the localized charge redistributions, owing to
652
+ the change in the occupancy of the perturbation-localized
653
+ bound states at the Fermi level.
654
+ This intuitive result is nothing but a direct consequence
655
+ of the universal nearsightedness principle of W. Kohn
656
+ [12, 13, 14] stating that, at fixed chemical potential, the
657
+ electronic charge density depends on the external field
658
+ (in our case being an assembly of localized scattering
659
+ centers) only at nearby points.
660
+ Another conclusion drawn by E. Prodan and W. Kohn
661
+ in Ref.
662
+ [13] (see also Ref.
663
+ [14] for the fine details in
664
+ d = 1) is that the adiabatic perturbations to the exter-
665
+ nal potential, no matter how strong, have a negligible
666
+ effect on the local charge density beyond a certain char-
667
+ acteristic length scale, which, in the insulating regime,
668
+ is naturally provided by the charge correlation length
669
+ ξg. From the viewpoint of our topological invariant (33),
670
+ this means that in the case of well-separated impurities
671
+ ∣xn −xn′∣/ξg ≫ 1, the topological invariant is expected to
672
+ approach a sum of the individual single-impurity invari-
673
+ ants, as distant impurities are not supposed to be able
674
+ to “talk” with one another on such scales.
675
+ Indeed, in
676
+ an insulating state, it is well-known, that the two-point
677
+ correlation functions G(0)(xn, xn′) decay exponentially
678
+ at large distances ∼ e−∣Rmn−Rmn′ ∣/ξg (where mn labels
679
+ the unit cell accommodating the nth scattering center),
680
+ meaning that we can approximate
681
+ (G(0)(z))n,n′ ≃δn,n′G(0)(xn,xn),
682
+ (34)
683
+ implying that
684
+ M(z) ≃
685
+ N
686
+
687
+ n=1
688
+ (1Nc − G(0)(xn,xn)V (n)
689
+ 0
690
+ ),
691
+ (35)
692
+ and
693
+ δQ ≃ −
694
+ N
695
+
696
+ n=1∮C
697
+ dz
698
+ 2πi
699
+
700
+ ∂z log det{1Nc − G(0)(xn,xn)V (n)
701
+ 0
702
+ }.
703
+ (36)
704
+ This result may be seen as a form of the conven-
705
+ tional Born approximation of the linear transport theory,
706
+ whereby, to the lowest order in the impurity density, one
707
+ considers impurities as independent.
708
+ B.
709
+ An illustration: A pair of magnetic impurities
710
+ in an illuminated quantum wire
711
+ To illustrate some of the points highlighted in the
712
+ above discussion, we here consider a simple model of a
713
+ spin-orbit-interacting ballistic quantum wire, submersed
714
+ into the background of the spatially oscillating electro-
715
+ magnetic field. The bulk Hamiltonian assumes the form
716
+ of the Pauli Hamiltonian with an extra Rashba-like term:
717
+ H(0)
718
+ x
719
+ =
720
+ (p + e
721
+ cAx(x))
722
+ 2
723
+ 2m
724
+ + kR ⋅ σ
725
+ m
726
+ (p + e
727
+ cAx(x))
728
+ + µBge
729
+ 2
730
+ σ ⋅ B(x).
731
+ (37)
732
+ Above, kR = (kR,x, kR,y, kR,z) is the Rashba spin-orbit
733
+ vector, σ = (σx, σy, σz) is the vector of the Pauli spin
734
+ matrices, µB =
735
+ e
736
+ 2mc is the Bohr magneton, c is the speed
737
+ of light in vacuum, ge is the electron’s Land´e g-factor,
738
+ and
739
+ B(x) =∇ × A(x)∣
740
+ x=xˆex,
741
+ Ax(x) = ˆex ⋅ A(x)∣
742
+ x=xˆex, (38)
743
+ with ˆex being the ort in the x-direction, and A(x) be-
744
+ ing the electromagnetic vector potential of the monochro-
745
+ matic plane-wave form
746
+ A(x) = A0 cos(q ⋅ x + ϕ),
747
+ (39)
748
+ in the Coulomb gauge
749
+ ∇ ⋅ A(x) = 0 ⇐⇒ q ⋅ A0 = 0.
750
+ (40)
751
+
752
+ 7
753
+ ° º
754
+ L
755
+ ° º
756
+ 2L
757
+ 0
758
+ º
759
+ 2L
760
+ º
761
+ L
762
+ k
763
+ °2.00
764
+ °1.75
765
+ °1.50
766
+ °1.25
767
+ °1.00
768
+ °0.75
769
+ °0.50
770
+ ≤k
771
+ x
772
+ y
773
+ z
774
+ kR
775
+ m(1)
776
+ eff
777
+ m(2)
778
+ eff
779
+ R
780
+ q
781
+ B
782
+ E
783
+ λ
784
+ a)
785
+ b)
786
+ FIG. 2. Panel a): A schematic illustration of a ballistic quantum wire featuring Rashba-style spin-orbit coupling (defined by a
787
+ spin-orbit vector kR) and submersed into a spatially-periodic arrangement of electric E and magnetic B fields of wavelength λ.
788
+ The two impurity atoms, separated by distance R and carrying an effective magnetic moment of m(j)
789
+ eff , j = 1, 2, are schematically
790
+ shown by atomic symbols pierced with the magnetic moment-symbolizing arrows. Panel b): The bulk energy spectrum of the
791
+ two-impurity problem. The energy bands are shown in dark blue, the chemical potential located inside the second spectral gap
792
+ (above the fourth energy band) is depicted in orange, and the relevant spectral region is highlighted in light blue.
793
+ The wave vector of the background electromagnetic field
794
+ defines the fictitious lattice spacing
795
+ L =
796
+
797
+ ˆex ⋅ q,
798
+ (41)
799
+ where we have excluded the uninteresting case of the or-
800
+ thogonally propagating wave ˆex ⋅ q = 0.
801
+ We note that the Hamiltonian in Eq. (37) falls into
802
+ the class of systems defined by the Hamiltonian (1), with
803
+ d = 1 and
804
+ V (x) =µBge
805
+ 2
806
+ σ ⋅ B(x) + e2A2
807
+ x(x)
808
+ 2mc2
809
+ + ekR ⋅ σAx(x)
810
+ mc
811
+ , (42)
812
+ ˜Ax(x) =e
813
+ cAx(x) + kR ⋅ σ.
814
+ (43)
815
+ As this demonstration is assumed to be interpretative,
816
+ it suffices to consider the case of a pair of impurities,
817
+ which we assume to be separated by distance R:
818
+ ˜V (x) = ˜V (1)
819
+ 0
820
+ δ(x) + ˜V (2)
821
+ 0
822
+ δ(x − R).
823
+ (44)
824
+ Note that we can place the first impurity at x = 0 with-
825
+ out loss of generality, as its other positions inside the wire
826
+ may be achieved by appropriate tuning of the modula-
827
+ tion’s phase ϕ. Furthermore, we assume the impurities
828
+ to exert both the electrostatic and the exchange “force”
829
+ on the wire’s electrons, which we encode in the following
830
+ form of the impurities’ vertex functions
831
+ ˜V (j)
832
+ 0
833
+ = Ujσ0 + µBge
834
+ 2
835
+ σ ⋅ B(j)
836
+ eff ,
837
+ (45)
838
+ where B(j)
839
+ eff is the effective (also appropriately screened to
840
+ have a short-ranged effect only) magnetic field, produced
841
+ by the effective magnetic moment of the impurity atom
842
+ m(j)
843
+ eff =
844
+ qjgj
845
+ 2MjcS(j), with qj, gj, and Mj being the charge,
846
+ g-factor, and mass of the jth impurity. Furthermore, Uj
847
+ denotes the strength of the electrostatic potential, defin-
848
+ ing the corresponding force exerted by the impurity on
849
+ the electrons. Not going into much of the microscopic
850
+ details, in the following, we treat Uj and B(j)
851
+ eff as some
852
+ constant parameters.
853
+ The resulting setup is schemati-
854
+ cally illustrated in panel a) of Fig. 2.
855
+ Further, to illustrate our point, we assume that the
856
+ associated impurity parameters {R, {Uj}j, {B(j)
857
+ eff }j}
858
+ evolve with a fictitious “adiabatic time” τ ∈ [0,T], in
859
+ such a manner that their temporal derivatives remain
860
+ much smaller than the Fermi energy ϵF times their value,
861
+ for all τ ∈ [0,T].
862
+ The particular form of the pumping protocol used to
863
+ produce the numerical data and the concrete numerical
864
+ values of the free model parameters are provided in Ap-
865
+ pendix D. The resulting bulk energy spectrum is demon-
866
+ strated in panel b) of Fig. 2.
867
+ The
868
+ numerical
869
+ data
870
+ for
871
+ the
872
+ excess
873
+ charge-
874
+ characterizing topological invariant,
875
+ as well as the
876
+ spectral flow of bound state energies inside the chem-
877
+ ical potential-accommodating spectral gap, is shown
878
+ in Fig.
879
+ 3.
880
+ In particular, using the parametrization
881
+ R(τ) = (nR − 1)L + ¯R(τ), suggested in the Appendix
882
+ D, we present the data for five different values of
883
+ nR ∈ {1,...,5}, as shown in five different columns of
884
+ the corresponding figure, with upper and lower rows
885
+ corresponding to the spectral flow and the topolog-
886
+ ical invariant, respectively.
887
+ Solid black and dashed
888
+ burgundy lines mark the cases of independent and
889
+ “interacting”
890
+ impurities,
891
+ correspondingly.
892
+ By
893
+ in-
894
+ dependent impurities, we here understand that the
895
+ separation between them is effectively infinite, so that
896
+ the off-diagonal blocks of the M(z) matrix (see Eq.
897
+ (14) for the definition) may be completely ignored.
898
+ This means that the bound state spectrum of the
899
+ independent impurities is provided by the solution of
900
+ det(12 − G(0)(0,0) ˜V (1)
901
+ 0
902
+ )det(12 − G(0)(R,R) ˜V (2)
903
+ 0
904
+ )∣z∈R =
905
+
906
+ 8
907
+ 0
908
+ T/4
909
+ T/2
910
+ 3T/4
911
+ T
912
+ τ
913
+ max ϵk,4
914
+ min ϵk,5
915
+ nR = 1
916
+ 0
917
+ T/4
918
+ T/2
919
+ 3T/4
920
+ T
921
+ τ
922
+ nR = 2
923
+ 0
924
+ T/4
925
+ T/2
926
+ 3T/4
927
+ T
928
+ τ
929
+ nR = 3
930
+ 0
931
+ T/4
932
+ T/2
933
+ 3T/4
934
+ T
935
+ τ
936
+ nR = 4
937
+ 0
938
+ T/4
939
+ T/2
940
+ 3T/4
941
+ T
942
+ τ
943
+ nR = 5
944
+ 0
945
+ T/4
946
+ T/2
947
+ 3T/4
948
+ T
949
+ τ
950
+ −2
951
+ −1
952
+ 0
953
+ 1
954
+ 2
955
+ δQ
956
+ 0
957
+ T/4
958
+ T/2
959
+ 3T/4
960
+ T
961
+ τ
962
+ 0
963
+ T/4
964
+ T/2
965
+ 3T/4
966
+ T
967
+ τ
968
+ 0
969
+ T/4
970
+ T/2
971
+ 3T/4
972
+ T
973
+ τ
974
+ 0
975
+ T/4
976
+ T/2
977
+ 3T/4
978
+ T
979
+ τ
980
+ FIG. 3. The figure demonstrates the adiabatic flows of both the bound state energy spectrum and the excess charge invariant
981
+ in the toy model proposed in Section III B. Specifically, the spectral flow of the impurity-localized bound state energies is
982
+ shown in the upper row, while the second row is dedicated to the invariant itself. As is explained in Appendix D, the position
983
+ of the second impurity is parametrized as R(τ) = (nR − 1)L + ¯R(τ), where nR denotes the number of the unit cell hosting
984
+ the second imperfection, and ¯R(τ) ∈ [0, L] describes its location within the unit cell. The five distinct columns in the above
985
+ figure correspond to five choices of nR = 1, 2, 3, 4, 5. In all of the panels, red dashed lines correspond to the actual solution,
986
+ while black solid lines relate to the case of two independent impurities (see the approximate formula (36)). As the separation
987
+ between the impurities becomes of the order of the charge localization length ξg = O(L) (see Appendix D), both adiabatic flows
988
+ approach the limit of two independent impurities.
989
+ 0, while the topological invariant is given by the ap-
990
+ proximation (36).
991
+ By the “interacting” impurities, on
992
+ the other hand, we understand that the exact relations
993
+ were used to produce the numerical data. The numerical
994
+ technique for evaluation of bulk position space Green’s
995
+ functions, as well as the topological indices of the form
996
+ (33), is outlined in Ref.
997
+ [10].
998
+ In our calculations,
999
+ the values of the contour parameters were chosen as
1000
+ η = 1, B = −30 (such a choice of B is motivated by
1001
+ the presence of the bound states below the lowest band
1002
+ ω ∈ (−∞, mink ϵk,1] in our model (37)).
1003
+ The central purpose of our demonstration is to show
1004
+ that upon the increase in the impurity’s separation be-
1005
+ yond the charge localization length ξg = O(L) (see Ap-
1006
+ pendix D), both the topological invariant and the bound
1007
+ state spectrum approach that of a pair of independent
1008
+ impurities.
1009
+ This effect is a direct consequence of the
1010
+ nearsightedness principle, telling us that a localized cause
1011
+ leads to a localized effect. Furthermore, as one may an-
1012
+ ticipate, the discontinuous jumps of the excess charge in-
1013
+ variant occur precisely at the points where bound states
1014
+ enter/leave the occupied part of the energy spectrum, as
1015
+ is explained in Section III A. Another interesting obser-
1016
+ vation is the non-zero value of the topological invariant
1017
+ at the beginning of the adiabatic evolution in τ, where
1018
+ the strengths of the electrostatic repulsion are the small-
1019
+ est 0 < Uj ≪ 1 (see Appendix D). This feature is a conse-
1020
+ quence of the presence of impurity-localized bound states
1021
+ below the bottom of the lowest energy band. Such an
1022
+ effect is well-known in the case of attractive scalar impu-
1023
+ rities, whereas here, it is generated by the non-Abelian
1024
+ structure of the model, and, to the best of our knowledge,
1025
+ was not reported previously in the literature.
1026
+ C.
1027
+ Topological invariants characterizing the
1028
+ boundary charge in unidimensional crystals
1029
+ In this section, we would like to comment on the topo-
1030
+ logical invariants characterizing boundary charges in uni-
1031
+ dimensional crystals, extensively discussed in Refs. [7,
1032
+ 8, 9, 10].
1033
+ In particular, let us consider a d = 1 semi-
1034
+ infinite system described by the Hamiltonian (1), with
1035
+ the boundary placed at x = xb. An appropriate restric-
1036
+ tion of x defines the respective right and left subsystems:
1037
+ x ∈ [xb,∞),
1038
+ right sub-system,
1039
+ (46)
1040
+ x ∈ (−∞,xb],
1041
+ left sub-system.
1042
+ (47)
1043
+ In our definition, the primitive unit cell is defined as the
1044
+ one starting at the boundary of the right semi-infinite
1045
+ system UC = [xb, xb+L], with L being the lattice period.
1046
+ In this definition, the left half-system is always obtained
1047
+ from the right one by a local inversion operation, which
1048
+ acts by the inversion of local coordinates within each unit
1049
+ cell.
1050
+ Now we define the boundary charge operators corre-
1051
+ sponding to right and left semi-infinite systems as the
1052
+ envelope-weighted integrals of the expectation values of
1053
+ the appropriate excess charge density operators:
1054
+ Q(R)
1055
+ B
1056
+ =∫
1057
+
1058
+ xb
1059
+ dxf(x)⟨δ̂ρR(x)⟩,
1060
+ (48)
1061
+ Q(L)
1062
+ B
1063
+ =∫
1064
+ xb
1065
+ −∞ dxf(x)⟨δ̂ρL(x)⟩,
1066
+ (49)
1067
+ where, in analogy with Eq.
1068
+ (17), δ̂ρS(x) = ̂ρS(x) − ¯ρ,
1069
+ and ̂ρS(x) is the density operator referring to the system
1070
+ S = R, L. Furthermore, the envelope function f(x) is
1071
+ chosen in accordance with Eq. (22), with xp = xb, and
1072
+
1073
+ 9
1074
+ the range of x being restricted according to Eqs. (46)
1075
+ and (47).
1076
+ Let us now consider measuring the total excess charge
1077
+ δQ accumulated around x = xb in a translationally in-
1078
+ variant system x ∈ (−∞, ∞). By the polarization charge
1079
+ neutrality condition QP = 0, demonstrated in Appendix
1080
+ B, the total excess charge also vanishes δQ = 0.
1081
+ On
1082
+ the other hand, we may consider a translationally invari-
1083
+ ant system as a sum of right and left semi-infinite sys-
1084
+ tems with a coupling corresponding to the bulk Hamilto-
1085
+ nian switched in between them. This coupling manifests
1086
+ itself as a local perturbation and, by the nearsighted-
1087
+ ness principle of Kohn, is capable of affecting the total
1088
+ charge locally by at most introducing or removing a num-
1089
+ ber of additional bound states, resulting in an integer
1090
+ contribution QI. In this connection, we conclude that
1091
+ δQ = Q(R)
1092
+ B
1093
+ + Q(L)
1094
+ B
1095
+ − QI = 0, where QI is known as the in-
1096
+ terface invariant. One of the central results of Ref. [10],
1097
+ was to demonstrate that
1098
+ QI =Q(R)
1099
+ B
1100
+ + Q(L)
1101
+ B
1102
+ = −∮C
1103
+ dz
1104
+ 2πi
1105
+
1106
+ ∂z log det{G(0)(xb,xb)}.
1107
+ (50)
1108
+ That is, the interface invariant, characterizing the bound-
1109
+ ary charge upon local inversions, is a topological quan-
1110
+ tum number given by the winding of the determinant
1111
+ of bulk position space Green’s function evaluated at the
1112
+ location of the boundary.
1113
+ Now let us proceed with the transformations of the
1114
+ boundary charge under translations. First, we consider
1115
+ the right boundary charge of the so-called reference sys-
1116
+ tem, starting at xb = 0:
1117
+ Q(R)
1118
+ B (0) = ∫
1119
+
1120
+ 0
1121
+ dxf(x)(ρ(x) − ¯ρ),
1122
+ (51)
1123
+ and we would like to analyze the changes in this quantity
1124
+ upon the translation of the boundary by xϕ ∈ [0, L]. In-
1125
+ stead of shifting the boundary, we consider adding the fol-
1126
+ lowing potential ˆV (x) = ˆV0Θ(x)Θ(xϕ − x), ˆV0 → ∞. By
1127
+ the Pauli principle, the charge density becomes zero for
1128
+ x ∈ [0,xϕ] as these states sit at infinite energy above the
1129
+ chemical potential µ. From the definition of the bound-
1130
+ ary charge, we are left with the following contribution:
1131
+ δQ(R)
1132
+ B (xϕ) = ∫
1133
+
1134
+ 0
1135
+ dxf(x)(0 − ¯ρ)
1136
+ mod 1
1137
+ = −¯ρxϕ
1138
+ mod 1,
1139
+ (52)
1140
+ where
1141
+ mod 1 contribution again comes from the near-
1142
+ sightedness principle. This analysis allows us to conclude
1143
+ that:
1144
+ Q(R)
1145
+ B (xϕ) − Q(R)
1146
+ B (0) = ¯ρxϕ + I(xϕ),
1147
+ (53)
1148
+ where I(xϕ) is known as the boundary invariant. An-
1149
+ other important result of Ref. [10] was to show that
1150
+ I(xϕ) = −∮C
1151
+ dz
1152
+ 2πi
1153
+
1154
+ ∂z lndetU(xϕ),
1155
+ (54)
1156
+ where U(xϕ) is defined via the path-ordered exponential
1157
+ U(xϕ) =Pexp{∫
1158
+ x
1159
+ 0
1160
+ dx′L(x′)},
1161
+ (55)
1162
+ L(x) =[G(0)(x,x)]−1G(0)
1163
+ 2 (x,x+) − iA(x),
1164
+ (56)
1165
+ and G(0)
1166
+ 2 (x,x′) = ∂x′G(0)(x,x′).
1167
+ In other words, the
1168
+ boundary invariant is also a topological quantum number
1169
+ expressed as a winding of the appropriate functional of
1170
+ bulk position space Green’s functions.
1171
+ In this way, we see that the quantization of the topo-
1172
+ logical invariants characterizing the boundary charge in
1173
+ one-dimensional insulators is a direct consequence of the
1174
+ nearsightedness principle. As this intuitive physical prin-
1175
+ ciple holds beyond the single spatial dimension, one ex-
1176
+ pects the excess charges accumulated on inhomogeneities
1177
+ of various spatial co-dimensions in d-dimensional crystals
1178
+ to possess similar topological characterization schemes.
1179
+ Indeed, linear scaling of the boundary charge, along with
1180
+ its discontinuous jumps by a unit of the electron charge
1181
+ at the bound state escape/entrance spectral points, was
1182
+ recently demonstrated in a two-dimensional system [27].
1183
+ IV.
1184
+ CONCLUSIONS AND OUTLOOK
1185
+ In this paper, the quantization of the excess charges on
1186
+ localized scattering centers in d-dimensional insulators
1187
+ was discussed. Our analysis reveals that an assembly of
1188
+ such imperfections accumulates an integral excess charge,
1189
+ given by a winding number expression. We find that an
1190
+ adiabatic perturbation (no matter how strong) comprised
1191
+ of either relocation of the impurities or a modification of
1192
+ their vertex functions (or both at the same time) results
1193
+ in the change of the total charge by an integer, deter-
1194
+ mined by the saldo of the imperfection-localized bound
1195
+ states that entered or escaped the occupied spectral re-
1196
+ gion, inside the chemical potential-hosting bulk spectral
1197
+ gap. The quantization of this topological invariant was
1198
+ shown to be a direct consequence of the nearsightedness
1199
+ principle of the electronic matter, limiting the range of
1200
+ the effect of a localized cause. Additionally, this local
1201
+ behavior of the electronic matter in the insulating state
1202
+ was shown to be responsible for the quantization of the
1203
+ topological invariants characterizing the unidimensional
1204
+ boundary charge studied in [7, 8, 9, 10]. Furthermore,
1205
+ our study confirms the central paradigm of Ref.
1206
+ [16],
1207
+ namely that localized perturbations in insulators specifi-
1208
+ cally lead to the change in occupancy of the correspond-
1209
+ ing perturbation-localized bound states, modifying the
1210
+ total charge, defined as the macroscopic average on the
1211
+ scales significantly exceeding both the unit cell size L and
1212
+ the charge correlation length ξg, by at most an integer.
1213
+ As is now obvious, the present paper is of conceptual
1214
+ value only as the evaluation of the suggested topologi-
1215
+ cal invariant (33) for a specific multi-impurity (N ≫ 1)
1216
+ system poses a challenge on its own. In particular, this
1217
+ concerns questions regarding the regularization schemes
1218
+
1219
+ 10
1220
+ for the higher-dimensional equal-argument Green’s func-
1221
+ tions, as well as the basic questions regarding the numer-
1222
+ ical feasibility of the problem. Furthermore, it would be
1223
+ of future interest to study the expansion of the topolog-
1224
+ ical invariant in the interaction between the individual
1225
+ impurities, as generated by the off-diagonal blocks of the
1226
+ M(z) matrix, and analyze its ties with the conventional
1227
+ Born series for the impurity-dressed T-matrix. As it is
1228
+ suggested in the present study, in the insulating state,
1229
+ the impurity density ρI has to be always contrasted with
1230
+ the inverse charge localization length ξg, in such a man-
1231
+ ner that the condition 1 ≫ ρIξd
1232
+ g implies the validity of the
1233
+ Born approximation, treating impurities as independent.
1234
+ V.
1235
+ ACKNOWLEDGMENTS
1236
+ The author gratefully acknowledges the durable ex-
1237
+ change of ideas with M. Pletyukhov and H. Schoeller.
1238
+ Further, the author generously thanks S. Miles and M.
1239
+ Pletyukhov for their valuable comments.
1240
+ Most of the present work was done at the Institut f¨ur
1241
+ Theorie der Statistischen Physik of RWTH Aachen and
1242
+ was financially supported by the Deutsche Forschungsge-
1243
+ meinschaft via RTG 1995.
1244
+ Appendix A: Contraction of two Green’s functions
1245
+ Quite generically we may represent
1246
+ G = ⨋s
1247
+ ∣s⟩⟨s∣
1248
+ z − ϵs
1249
+ ,
1250
+ (A1)
1251
+ where the meta-index s labels the eigenstates ∣s⟩ and
1252
+ eigenenergies ϵs of the Hamiltonian.
1253
+ Considering the
1254
+ product of the Green’s function with itself
1255
+ GG = ⨋s ⨋s′
1256
+ ∣s⟩⟨s′∣
1257
+ (z − ϵs)(z − ϵs′) ⟨s∣s′⟩
1258
+
1259
+ δ(s,s′)
1260
+ = ⨋s
1261
+ ∣s⟩⟨s∣
1262
+ (z − ϵs)2
1263
+ = − ∂
1264
+ ∂ω ⨋s
1265
+ ∣s⟩⟨s∣
1266
+ z − ϵs
1267
+ = − ∂
1268
+ ∂ω G.
1269
+ (A2)
1270
+ Taking the position space matrix elements
1271
+ ⟨x∣GG∣x′′⟩ = − ∂
1272
+ ∂ω G(x,x′′),
1273
+ and inserting
1274
+ 1 = ∫Rd d(d)x∣x⟩⟨x∣,
1275
+ (A3)
1276
+ we obtain the desired identity
1277
+ ∫Rd d(d)x′G(x,x′)G(x′,x′′) = − ∂
1278
+ ∂ω G(x,x′′).
1279
+ (A4)
1280
+ Appendix B: Polarization charge
1281
+ We consider
1282
+ ∫Rd d(d)xf(x)(ρ(0)(x) − ¯ρ)
1283
+ = ∑
1284
+ m ∫UC d(d)¯xf(¯x + Rm)(ρ(0)(¯x) − ¯ρ).
1285
+ (B1)
1286
+ Above we parametrized the position space variable x as
1287
+ x = Rm+¯x, for some vector of integers m, and ¯x is the lo-
1288
+ cal coordinate within the unit cell ¯x ∈ UC. Furthermore,
1289
+ we used the periodicity property of ρ(0)(x), implied by
1290
+ the periodicity of the equal-argument Green’s function
1291
+ G(0)(x, x) = G(0)(x + Rm, x + Rm),
1292
+ ∀m ∈ Zd. (B2)
1293
+ The envelope function varies significantly only in the
1294
+ crossover region ∣Rm∣ = O(Lp), allowing us to approxi-
1295
+ mate
1296
+ f(¯x + Rm) ≈ f(Rm) + ¯x ⋅ ∇f(Rm),
1297
+ (B3)
1298
+ leading to
1299
+ ∫Rd d(d)xf(x)(ρ(0)(x) − ¯ρ)
1300
+ = ∫UC d(d)¯x∑
1301
+ m
1302
+ (¯x ⋅ ∇f(Rm))(ρ(0)(¯x) − ¯ρ).
1303
+ (B4)
1304
+ Now approximating
1305
+
1306
+ m
1307
+ (¯x ⋅ ∇f(Rm)) ≈
1308
+ 1
1309
+ VUC ∫Rd d(d)y(¯x ⋅ ∇yf(y))
1310
+ =
1311
+ 1
1312
+ VUC ∫Rd d(d)y∇y ⋅ (¯xf(y)) = 0,
1313
+ (B5)
1314
+ where in the last step we used Gauss’ divergence theorem.
1315
+ Appendix C: Contour integral representation
1316
+ First, we rewrite Eq. (32) as
1317
+ δQ = − 1
1318
+ 2πi ∫
1319
+ µ
1320
+ −∞ dω ∂
1321
+ ∂ω log det{M(z)}
1322
+ + 1
1323
+ 2πi ∫
1324
+ µ
1325
+ −∞ dω ∂
1326
+ ∂ω (log det{M(z)})∗ .
1327
+ (C1)
1328
+ Now we remind ourselves that
1329
+ (log f(z))∗ = log (f(z))∗ ≡ log f ∗(z).
1330
+ (C2)
1331
+ Furthermore, one has
1332
+ (det{M(z)})∗ = det{M†(z∗)},
1333
+ (C3)
1334
+ where, as before, the Hermitian conjugate does not affect
1335
+ the z-variable. Now we have
1336
+ det{M†(z∗)} = det{(1 − G(0)(z∗)˜V(0) )†}
1337
+ = det{1 − ˜V(0)G(0)(z∗)}
1338
+ = det{M(z∗)},
1339
+ (C4)
1340
+
1341
+ 11
1342
+ where to get from the pre-last to the last lines we em-
1343
+ ployed the Weinstein–Aronszajn identity.
1344
+ It hence follows that
1345
+ δQ = − 1
1346
+ 2πi ∫
1347
+ µ
1348
+ −∞ dω ∂
1349
+ ∂ω log det{M(ω + iη)}
1350
+ − 1
1351
+ 2πi ∫
1352
+ −∞
1353
+ µ
1354
+ dω ∂
1355
+ ∂ω log det{M(ω − iη)}
1356
+ = − ∮C
1357
+ dz
1358
+ 2πi
1359
+
1360
+ ∂z log det{M(z)}.
1361
+ (C5)
1362
+ Above, C is the counterclockwise rectangular contour de-
1363
+ fined as a union of four segments:
1364
+ C =[B + iη,µ + iη) ∪ [µ + iη,µ − iη) ∪ [µ − iη,B − iη)
1365
+ ∪ [B − iη,B + iη),
1366
+ B → −∞,
1367
+ η → 0+.
1368
+ (C6)
1369
+ We note that the integral in (C5) remains unaffected
1370
+ under continuous contour deformations, so long as the
1371
+ analytic structure of the integrand within the patch of
1372
+ the complex plane enclosed by contour C remains intact.
1373
+ In this connection, we may replace C with an arbitrary
1374
+ non-self-intersecting curve crossing the real axis at two
1375
+ points only, at any energy below the lowest eigenvalue of
1376
+ the full Hamiltonian, and at the chemical potential.
1377
+ Appendix D: Parameters and protocols
1378
+ In the numerical example provided in Section III B, the
1379
+ parameters of the model were chosen according to
1380
+ q = 2π(ex + κey)
1381
+ λ
1382
+
1383
+ 1 + κ2
1384
+ ,
1385
+ A0 = A0(κex − ey)
1386
+
1387
+ 1 + κ2
1388
+ ,
1389
+ m = 1, (D1)
1390
+ κ = 1 +
1391
+
1392
+ 5
1393
+ 2
1394
+ ,
1395
+ λ = 4,
1396
+ e
1397
+ cA0 = 1.17,
1398
+ kR =
1399
+
1400
+
1401
+
1402
+ 0.32
1403
+ 1.39
1404
+ 1.24
1405
+
1406
+
1407
+
1408
+ . (D2)
1409
+ Note that as we have set the electron’s mass m = 1 to
1410
+ unity (in addition to the electric charge e = 1 and reduced
1411
+ Plank’s constant ̵h = 1), we work in Hartree’s atomic
1412
+ units.
1413
+ In this way, the electromagnetic wave is propagating
1414
+ in the x − y plane, with the corresponding magnetic field
1415
+ being
1416
+ B(x) = 2πA0
1417
+ λ
1418
+ ez sin(q ⋅ x + ϕ).
1419
+ (D3)
1420
+ By definition, the corresponding lattice period is given
1421
+ by
1422
+ L = λ
1423
+
1424
+ 1 + κ2 = 2
1425
+
1426
+ 2(5 +
1427
+
1428
+ 5).
1429
+ (D4)
1430
+ To produce the data, we used the following pumping
1431
+ protocol for the impurities’ separation
1432
+ R(τ) = (nR − 1)L + ¯R(τ),
1433
+ ¯R(τ) = L
1434
+ T τ,
1435
+ L
1436
+ T ≪ vF ,
1437
+ (D5)
1438
+ where vF is the Fermi velocity and nR is an integer spec-
1439
+ ifying the number of the unit cell hosting the second
1440
+ impurity. For the impurities’ vertex functions, we fur-
1441
+ ther make an assumption of the equivalent impurities:
1442
+ U (1)(τ) = U (2)(τ) =∶ U(τ) and ∣B(1)
1443
+ eff (τ)∣ = ∣B(2)
1444
+ eff (τ)∣ =∶
1445
+ BI(τ). The direction of the magnetic moments, on the
1446
+ other hand, is allowed to be different in two scattering
1447
+ centers and is parametrized in spherical polar coordinates
1448
+ B(j)
1449
+ eff (τ)
1450
+ BI(τ) =
1451
+
1452
+
1453
+
1454
+ cosφ(j)(τ)sinθ(j)(τ)
1455
+ sinφ(j)(τ)sinθ(j)(τ)
1456
+ cosθ(j)(τ)
1457
+
1458
+
1459
+
1460
+ .
1461
+ (D6)
1462
+ In the following, we assume that, as is the case with the
1463
+ location of the second impurity within the unit cell num-
1464
+ ber nR, the impurity strength also grows linearly with
1465
+ τ
1466
+ U(τ) = U0 + δU τ
1467
+ T ,
1468
+ δU
1469
+ T ≪ ϵ2
1470
+ F .
1471
+ (D7)
1472
+ On the other hand, we assume the effective magnetic field
1473
+ of the impurity to oscillate as
1474
+ BI(τ) = B0 + δB sin(6πτ
1475
+ T ).
1476
+ (D8)
1477
+ The direction of the spins is prescribed by
1478
+ φ(1)(τ) = φ(2)(τ) = 2π sin(8πτ
1479
+ T ),
1480
+ (D9)
1481
+ θ(j)(τ) = π
1482
+ 2 (1 + (−1)j τ
1483
+ T ).
1484
+ (D10)
1485
+ The rest of the parameters are chosen as
1486
+ U0 = 0,
1487
+ δU = 10,
1488
+ e
1489
+ cB0 = 3,
1490
+ e
1491
+ cδB = 1.5.
1492
+ (D11)
1493
+ Now let us estimate the charge localization length ξg
1494
+ for the second bulk spectral gap, where the chemical po-
1495
+ tential µ is assumed to be placed. According to Ref. [10],
1496
+ the Fermi velocity may roughly be estimated as vF ≈
1497
+ kF
1498
+ m ≈
1499
+
1500
+ mL ≈ 0.825816. The energy gap at the Fermi level
1501
+ was numerically computed to be roughly Eg ≈ 0.271394,
1502
+ leading to the following estimate ξg ≈ 3 = O(L).
1503
+
1504
+ 12
1505
+ ∗ Email: kiryl.piasotski@kit.edu
1506
+ † On the leave from Institut f¨ur Theorie der Statistischen
1507
+ Physik, RWTH Aachen, 52056 Aachen, Germany
1508
+ 1 K. von Klitzing, G. Dorda, and M. Pepper, “New method
1509
+ for high-accuracy determination of the fine-structure con-
1510
+ stant based on quantized Hall resistance”, Phys. Rev. Lett.
1511
+ 45, 494 (1980).
1512
+ 2 D. J. Thouless, M. Kohmoto, M. P. Nightingale, and
1513
+ M. den Nijs, “Quantized Hall conductance in a two-
1514
+ dimensional periodic potential”, Phys. Rev. Lett. 49, 405
1515
+ (1982).
1516
+ 3 Y. Hatsugai, “Chern number and edge states in the integer
1517
+ quantum Hall effect”, Phys. Rev. Lett. 71, 3697 (1993); Y.
1518
+ Hatsugai, “Edge states in the integer quantum Hall effect
1519
+ and the Riemann surface of the Bloch function”, Phys.
1520
+ Rev. B 48, 11851, (1993).
1521
+ 4 M. Z. Hasan and C. L. Kane, “Colloquium: topological
1522
+ insulators”, Rev. Mod. Phys. 82, 3045 (2010).
1523
+ 5 X.-L. Qi and S.-C. Zhang, “Topological insulators and su-
1524
+ perconductors”, Rev. Mod. Phys. 83, 1057 (2011).
1525
+ 6 A. Kitaev, “Periodic table for topological insulators and
1526
+ superconductors”, AIP Conf. Proc. 1134, 22 (2009).
1527
+ 7 M. Pletyukhov, D. M. Kennes, J. Klinovaja, D. Loss,
1528
+ and H. Schoeller, “Surface charge theorem and topologi-
1529
+ cal constraints for edge states: An analytical study of one-
1530
+ dimensional nearest-neighbor tight-binding models” Phys.
1531
+ Rev. B 101, 165304 (2020); M. Pletyukhov, D. M. Kennes,
1532
+ J. Klinovaja, D. Loss, and H. Schoeller, “Topological in-
1533
+ variants to characterize universality of boundary charge in
1534
+ one-dimensional insulators beyond symmetry constraints”,
1535
+ Phys. Rev. B 101, 161106(R) (2020).
1536
+ 8 N. M¨uller, K. Piasotski, D. M. Kennes, H. Schoeller, and
1537
+ M. Pletyukhov, “Universal properties of boundary and in-
1538
+ terface charges in multichannel one-dimensional models
1539
+ without symmetry constraints”, Phys. Rev. B 104, 125447
1540
+ (2021).
1541
+ 9 S. Miles, D. M. Kennes, H. Schoeller, and M. Pletyukhov,
1542
+ “Universal properties of boundary and interface charges in
1543
+ continuum models of one-dimensional insulators”, Phys.
1544
+ Rev. B 104, 155409 (2021).
1545
+ 10 K. Piasotski, N. M¨uller, D. M. Kennes, H. Schoeller, and
1546
+ M. Pletyukhov, “Universal properties of boundary and in-
1547
+ terface charges in multichannel one-dimensional continuum
1548
+ models”, Phys. Rev. B 106, 165405 (2022).
1549
+ 11 W. Kohn and A. Yaniv, “Locality principle in wave me-
1550
+ chanics”, PNAS 75(11), 5270 (1978).
1551
+ 12 W.
1552
+ Kohn,
1553
+ “Density
1554
+ Functional
1555
+ and
1556
+ Density
1557
+ Matrix
1558
+ Method Scaling Linearly with the Number of Atoms”,
1559
+ Phys. Rev. Lett. 76, 3168 (1996).
1560
+ 13 E. Prodan and W. Kohn, “Nearsightedness of electronic
1561
+ matter”, PNAS 102, 11635 (2005).
1562
+ 14 E. Prodan, “Nearsightedness of electronic matter in one
1563
+ dimension”, Phys. Rev. B 73, 085108 (2006).
1564
+ 15 H.-R. Trebin, “The topology of non-uniform media in con-
1565
+ densed matter physics”, Adv. Phys. 31, 195 (1982).
1566
+ 16 M. Pletyukhov, D. M. Kennes, K. Piasotski, J. Klinovaja,
1567
+ D. Loss, and H. Schoeller, “Rational boundary charge in
1568
+ one-dimensional systems with interaction and disorder”,
1569
+ Phys. Rev. Res. 2, 033345 (2020).
1570
+ 17 J. Goldstone and F. Wilczek, “Fractional Quantum Num-
1571
+ bers on Solitons”, Phys. Rev. Lett. 47, 986 (1981).
1572
+ 18 C. S. Weber, K. Piasotski, M. Pletyukhov, J. Klinovaja,
1573
+ D. Loss, H. Schoeller, and D. M. Kennes, “Universality
1574
+ of Boundary Charge Fluctuations”, Phys. Rev. Lett. 126,
1575
+ 016803 (2021).
1576
+ 19 K. Nomura and N. Nagaosa, “Electric Charging of Mag-
1577
+ netic Textures on the Surface of a Topological Insulator”,
1578
+ Phys. Rev. B 82, 161401(R) (2010).
1579
+ 20 This includes models featuring a momentum-cubic spin-
1580
+ orbit interaction, or even tight-binding models, describing
1581
+ crystals in terms of incomplete basis sets of localized Wan-
1582
+ nier orbitals.
1583
+ 21 D. K. Park, “Green’s-function approach to two- and three-
1584
+ dimensional delta-function potentials and application to
1585
+ the spin-1/2 Aharonov–Bohm problem”, J. Math. Phys.
1586
+ 36, 5453 (1995).
1587
+ 22 D. A. Atkinson, H. W. Crater, “An exact treatment of
1588
+ the Dirac delta function potential in the Schr¨odinger equa-
1589
+ tion”, Amer. Jour. Phys. 43, 301 (1975).
1590
+ 23 R. Jackiw, “Delta-function potentials in two- and three-
1591
+ dimensional quantum mechanics” MAB B´eg memorial vol-
1592
+ ume (1991).
1593
+ 24 C. N. Friedman, “Perturbations of the Schr¨odinger equa-
1594
+ tion by potentials with small support”, J. Funct. Anal. 10,
1595
+ 346 (1972).
1596
+ 25 S. Kivelson and J. R. Schrieffer, “Fractional charge, a sharp
1597
+ quantum observable”, Phys. Rev. B 25, 6447 (1982).
1598
+ 26 In the case of one-dimensional (d = 1) single-channel
1599
+ (Nc = 1) systems, it is not possible for the bound state
1600
+ poles to coexist with the continuum (or extended) states.
1601
+ As a result, the effect of the bound state pole merging with
1602
+ the band is to increase the order of the branching pole at
1603
+ the band edge. In multichannel (Nc > 1) and/or multidi-
1604
+ mensional (d > 1) systems, on the other hand, the poles
1605
+ may, in principle, coexist with the band continuum. In ei-
1606
+ ther case, as a result, the band contribution to the winding
1607
+ number gets modified by unity.
1608
+ 27 Z. Hou, C. S. Weber, D. M. Kennes, D. Loss, H. Schoeller,
1609
+ J. Klinovaja, M. Pletyukhov, “Realization of a three-
1610
+ dimensional quantum Hall effect in a Zeeman-induced sec-
1611
+ ond order topological insulator on a torus”, arXiv preprint
1612
+ arXiv:2212.09053, (2022)
1613
+
7NE1T4oBgHgl3EQfnATd/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
7dAyT4oBgHgl3EQf2vnp/content/tmp_files/2301.00758v1.pdf.txt ADDED
@@ -0,0 +1,1713 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2
+
3
+
4
+ Analysis of a HAPS-Aided GNSS in Urban Areas
5
+ using a RAIM Algorithm
6
+
7
+ Hongzhao Zheng, Member, IEEE, Mohamed Atia, Senior Member, IEEE, and Halim Yanikomeroglu, Fellow, IEEE
8
+ Abstract—The global averaged civilian positioning accuracy is
9
+ still at meter level for all existing Global Navigation Satellite
10
+ Systems (GNSSs), and the performance is even worse in urban
11
+ areas. At lower altitudes than satellites, high altitude platform
12
+ stations (HAPS) offer several benefits, such as lower latency, less
13
+ pathloss, and likely smaller overall estimation error for the
14
+ parameters associated in the pseudorange equation. HAPS can
15
+ support GNSSs in many ways, and in this paper we treat the HAPS
16
+ as another type of ranging source. In so doing, we examine the
17
+ positioning performance of a HAPS-aided GPS system in an urban
18
+ area using both a simulation and physical experiment. The HAPS
19
+ measurements are unavailable today; therefore, they are modeled
20
+ in a rather simple but logical manner in both the simulation and
21
+ physical experiment. We show that the HAPS can improve the
22
+ horizontal dilution of precision (HDOP), the vertical dilution of
23
+ precision (VDOP), and the 3D positioning accuracy of GPS in both
24
+ suburban and dense urban areas. We also demonstrate the
25
+ applicability of a RAIM algorithm for the HAPS-aided GPS
26
+ system, especially in the dense urban area.
27
+
28
+ Index Terms—High altitude platform station (HAPS), horizontal
29
+ dilution of precision (HDOP), pseudorange, receiver autonomous
30
+ integrity monitoring (RAIM), vertical dilution of precision (VDOP).
31
+ I. INTRODUCTION
32
+ ODAY, many countries and the European union have
33
+ their own global navigation satellite systems (GNSSs).
34
+ However, 95 percent of the time, the global averaged
35
+ horizontal positioning accuracy of existing GNSSs is still at the
36
+ meter level, and it is even worse for the vertical positioning
37
+ accuracy [1]-[4] due to the nature of the satellite geometry.
38
+ Although vertical positioning performance is less important
39
+ than horizontal positioning performance today, it might be very
40
+ important in the future, for instance, for unmanned aerial
41
+ vehicles (UAVs) flying in the 3D aerial highways [5]. Thanks
42
+ to ongoing research on localization and navigation fields, there
43
+ are a number of techniques developed which can bring the
44
+ positioning accuracy of systems involving satellites to the
45
+ centimeter level. For example, Li et al. have shown that
46
+ centimeter-level positioning accuracy can be achieved using the
47
+ multi-constellation GNSS consisting of Beidou, Galileo,
48
+ GLONASS and GPS with precise point positioning (PPP) [6].
49
+ Because most civilian applications use single-frequency, low-
50
+ cost receivers for localization and navigation, many advanced
51
+ positioning algorithms, including PPP that delivers centimeter
52
+
53
+ This paper was supported in part by Huawei Canada.
54
+ H. Zheng, M. Atia, and H. Yanikomeroglu are with the Department of
55
+ Systems and Computer Engineering, Carleton University, Ottawa, ON K1S
56
+ 5B6,
57
+ Canada
58
+ (e-mail:
59
+ hongzhaozheng@cmail.carleton.ca;
60
+ Mohamed.Atia@carleton.ca; halim@sce.carleton.ca).
61
+ level positioning accuracy, cannot be implemented. Therefore,
62
+ the single point positioning (SPP) is the most commonly used
63
+ algorithm in civilian applications. But this is poised to change.
64
+ As increasing numbers of low-Earth-orbit (LEO) satellites are
65
+ launched into space, researchers are investigating the feasibility
66
+ of utilizing LEO satellites to aid the positioning service. For
67
+ instance, Li et al. have shown that a centimeter level Signal-In-
68
+ Space Ranging Error (SISRE) in the real-time PPP application
69
+ can be achieved using a LEO enhanced GNSS [7]. In the event
70
+ that GNSS signals are unavailable in urban areas, researchers
71
+ are also interested in building navigation systems that
72
+ exclusively rely on LEO satellite signals. For example, a
73
+ position root mean squared error (RMSE) of 14.8 m for a UAV
74
+ has been proven feasible using only two Orbcomm LEO
75
+ satellites with the carrier phase differential algorithm [8].
76
+ Compared to medium-Earth-orbit (MEO) satellites, which are
77
+ typically used in GNSSs, LEO satellites offer several
78
+ advantages, such as lower latency and less pathloss due to
79
+ shorter distance to ground users. LEO satellites also offer
80
+ greater availability due to the large number of them.
81
+ To further enhance high bandwidth networking coverage in
82
+ areas with obstacles, such as urban areas, another option is the
83
+ use of high altitude platform stations (HAPS1), which refer to
84
+ aerial platforms positioned in the stratosphere with a typical
85
+ altitude of about 20 km. HAPS can be utilized for many
86
+ technologies coming in 5G even 6G and beyond such as
87
+ computation offloading [9], edge computing [10], and aerial
88
+ base station [11]. As urban areas are where GNSS positioning
89
+ performance degrades severely, while also being where most
90
+ people live, we could improve the positioning performance of
91
+ GNSS by placing several HAPS above metro cities and
92
+ equipping them with satellite-grade atomic clocks so that HAPS
93
+ can be deployed as another type of ranging source. Even though
94
+ atomic clocks on satellites are highly accurate, they are not
95
+ perfect due to the time dilation postulates made in both
96
+ Einstein’s special theory of relativity and the general theory of
97
+ relativity. According to Einstein’s special theory of relativity,
98
+ an atomic clock on a fast-moving satellite runs slower than a
99
+ clock on Earth by around 7 microseconds per day. On the other
100
+ hand, according to the general theory of relativity, an atomic
101
+ clock which experiences weaker gravity on a distant satellite
102
+ runs faster than a clock which experiences greater gravity on
103
+ 1 In this paper, the acronym "HAPS" is used to denote “high altitude
104
+ platforms station” in both singular and plural forms, in line with the convention
105
+ adopted in the ITU (International Telecommunications Union) documents.
106
+ T
107
+
108
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
109
+
110
+ Earth by about 45 microseconds per day [12]. As HAPS operate
111
+ at an altitude of around 20 km and can be quasi-stationary, the
112
+ time dilation is negligible from the perspective of special
113
+ relativity and greatly reduced from the perspective of general
114
+ relativity. Therefore, the atomic clocks on HAPS will likely be
115
+ more accurate than that on satellites, which can make the
116
+ estimation error of the HAPS clock offset smaller than that of
117
+ the satellite clock offset. Since HAPS are positioned much
118
+ closer to the Earth than satellites, the pathloss of a HAPS is
119
+ expected to be much less, which will likely make the received
120
+ signal power of a HAPS stronger than that of a satellite, thereby
121
+ reducing the estimation error of the parameters associated in the
122
+ pseudorange measurement of the HAPS signal. The movement
123
+ of a HAPS can be confined to a cylindrical region with a radius
124
+ of 400 m and a height of about 700 m [13], which can reduce
125
+ the number of handovers during the course of navigation and
126
+ increase the utilization efficiency during its operation life. As
127
+ HAPS are positioned in the stratosphere, which is below the
128
+ ionosphere, their signals will likely be free of the ionospheric
129
+ effect, which is known to be one of the major sources of error
130
+ in pseudorange measurements.
131
+ Therefore, the overall
132
+ estimation error for a HAPS will likely be smaller than that of
133
+ a satellite. Similar to the pseudorange measurement for a
134
+ satellite, which incorporates the satellite position error, we
135
+ should also consider the position error in the pseudorange
136
+ measurement for HAPS. Fortunately, researchers have been
137
+ investigating the positioning of HAPS and have demonstrated
138
+ that HAPS positioning errors are comparable to or lesser than
139
+ satellite orbit errors. For example, Dovis et al. prove that 0.5 m
140
+ positioning accuracy (circular error probable [CEP] 68 percent)
141
+ for a HAPS is achievable using the modified RTK method [14].
142
+ There are a handful of papers in the literature that have
143
+ investigated the HAPS-aided GNSS [15]-[18]; however, to the
144
+ best of our knowledge, this paper is the first to provide a
145
+ comprehensive study of the positioning performance of a
146
+ HAPS-aided GNSS in urban areas.
147
+ There are plenty of operational GPS satellites that could fail
148
+ due to the degraded signal quality for reasons such as
149
+ obstruction, multipath, intentional or unintentional attacks,
150
+ thereby impacting the positioning performance of the GNSS. In
151
+ this case, a signal selection algorithm like the receiver
152
+ autonomous integrity monitoring (RAIM) algorithm, which can
153
+ detect and exclude poor quality signals, can be helpful in
154
+ improving the positioning performance. For example, about 35
155
+ percent decrease in RMS positioning error of the GPS-only case
156
+ and 50 percent decrease in RMS positioning error of the
157
+ GPS/GLONASS case in a severe urban scenario have been
158
+ achieved on smartphone GNSS chips by using a RAIM
159
+ algorithm [19]. Moreover, Yang and Xu propose a robust
160
+ estimation-based RAIM algorithm that can detect and exclude
161
+ multiple faulty satellites effectively with efficiency higher than
162
+ the conventional least squares (LS)-based RAIM algorithm
163
+ [20]. In this paper, we make three postulations: 1) a HAPS
164
+ signal is free of the ionospheric effect; 2) the estimation error
165
+ of the HAPS clock offset is smaller than that of the satellite
166
+ clock offset; and 3) the received signal power of a HAPS is
167
+ higher than that of a satellite, all of which contribute toward the
168
+
169
+ Fig. 1. System model of the HAPS-aided GPS.
170
+
171
+
172
+ assumption that the overall estimation error of the parameters
173
+ associated in the pseudorange equation for the HAPS is smaller
174
+ than that for the satellite. Under this assumption, we use the SPP
175
+ algorithm developed in our prior work [21] to show that HAPS
176
+ can indeed improve the positioning performance of legacy
177
+ GNSSs in urban areas through both a simulation and a physical
178
+ experiment. We also demonstrate the applicability of the RAIM
179
+ algorithm to a HAPS-aided GPS system, especially in dense
180
+ urban areas. Since the HAPS measurements are unavailable so
181
+ far, they are simulated in a rather simple but logical way in both
182
+ the simulation and physical experiment. The contributions of
183
+ this paper are listed below.
184
+
185
+ First, using a commercial GNSS simulator, we
186
+ simulate the GPS pseudorange signals and generate
187
+ the positions of HAPS. By using the default system
188
+ parameters as well as a proper manipulation of the
189
+ number of visible satellites, we show that the
190
+ positioning performance of the GPS-only system in
191
+ both the suburban and dense urban areas are close to
192
+ the real scenario. Moreover, we show the positioning
193
+ performance of different systems where different
194
+ numbers of HAPS are used with or without the GPS
195
+ system. The issue of the ranging source geometry is
196
+ revealed from the simulation results.
197
+
198
+ Next, we apply the SPP algorithm to the real GPS data
199
+ collected using two commercial GNSS receivers as
200
+ well as the HAPS data generated using the commercial
201
+ GNSS software. In so doing, we show the advantage
202
+ of the HAPS-aided GPS system in the sense of the
203
+ horizontal dilution of precision (HDOP) and the
204
+ vertical dilution of precision (VDOP).
205
+
206
+ Finally, we implement a RAIM algorithm and
207
+ demonstrate its effectiveness in improving the 3D
208
+ positioning performance of the HAPS-aided GPS
209
+ system, especially in dense urban areas.
210
+ The rest of the paper is organized as follows: in Section Ⅱ, the
211
+ system model, the SPP algorithm, and the RAIM algorithm are
212
+ described. In Section Ⅲ, the simulation setup of the HAPS-
213
+
214
+ Ionosphere
215
+ HAPS
216
+ HAPS
217
+ 20km
218
+ HAPSfootprint
219
+ HAPSfootprint
220
+ 15°
221
+ cell.
222
+ cell> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
223
+
224
+ aided GPS system and the simulation results are presented. In
225
+ Section Ⅳ, the physical experiment setup and results, including
226
+ both the DOP analysis and the 3D positioning accuracy
227
+ analysis, are provided. Finally, Section V offers some
228
+ conclusions and a discussion of future research directions. For
229
+ simplicity, the GNSS signal only involves the GPS C/A L1
230
+ signal.
231
+ II. SYSTEM MODEL
232
+ The system model of the HAPS-aided GPS system is
233
+ depicted in Fig. 1. There are four satellites shown in Fig. 1, this
234
+ is just a reminder that at least four satellites are required to
235
+ perform precise 3D localization using GNSS. The typical
236
+ choice for the elevation mask is 10 degrees. However, we use
237
+ 15 degrees as the elevation mask for the satellites and HAPS
238
+ due to the following reasons: 1) the atmospheric error owing to
239
+ the signal refraction can be neglected if the elevation of a
240
+ satellite is greater than 15 degrees [22], which is likely true for
241
+ a HAPS as well; 2) As there is a higher chance of ensuring the
242
+ required number of ranging source with HAPS, we can improve
243
+ the positioning performance further by only using those satellite
244
+ signals with better quality. The pseudorange equation for
245
+ satellite is given by
246
+
247
+
248
+ 𝑝𝑆𝐴𝑇 = 𝜌𝑆𝐴𝑇 + 𝑑𝑆𝐴𝑇 + 𝑐(𝑑𝑡 − 𝑑𝑇𝑆𝐴𝑇) + 𝑑𝑖𝑜𝑛,𝑆𝐴𝑇 + 𝑑𝑡𝑟𝑜𝑝,𝑆𝐴𝑇
249
+ + 𝜖𝑚𝑝,𝑆𝐴𝑇 + 𝜖𝑝
250
+
251
+
252
+ (1)
253
+ where 𝑝𝑆𝐴𝑇 denotes the satellite pseudorange measurement,
254
+ 𝜌𝑆𝐴𝑇 is the geometric range between the satellite and receiver,
255
+ 𝑑𝑆𝐴𝑇 represents the satellite orbit error, 𝑐 is the speed of light,
256
+ 𝑑𝑡 is the receiver clock offset from GPS time, 𝑑𝑇𝑆𝐴𝑇 is the
257
+ satellite clock offset from GPS time, 𝑑𝑖𝑜𝑛,𝑆𝐴𝑇 denotes the
258
+ ionospheric delay for satellite signals, 𝑑𝑡𝑟𝑜𝑝,𝑆𝐴𝑇 denotes the
259
+ tropospheric delay for satellite signals, 𝜖𝑚𝑝,𝑆𝐴𝑇 is the delay
260
+ caused by the multipath for satellite signals, and 𝜖𝑝 is the delay
261
+ caused by the receiver noise. The pseudorange equation for
262
+ HAPS can be expressed as follows:
263
+
264
+
265
+ 𝑝𝐻𝐴𝑃𝑆 = 𝜌𝐻𝐴𝑃𝑆 + 𝑑𝐻𝐴𝑃𝑆 + 𝑐(𝑑𝑡 − 𝑑𝑇𝐻𝐴𝑃𝑆) + 𝑑𝑡𝑟𝑜𝑝,𝐻𝐴𝑃𝑆
266
+ + 𝜖𝑚𝑝,𝐻𝐴𝑃𝑆 + 𝜖𝑝
267
+
268
+ (2)
269
+ where 𝑝𝐻𝐴𝑃𝑆 denotes the HAPS pseudorange measurement,
270
+ 𝜌𝐻𝐴𝑃𝑆 represents the geometric range between the HAPS and
271
+ the receiver, 𝑑𝐻𝐴𝑃𝑆 represents the HAPS position error, 𝑑𝑇𝐻𝐴𝑃𝑆
272
+ is the HAPS clock offset from GPS time, 𝑑𝑡𝑟𝑜𝑝,𝐻𝐴𝑃𝑆 denotes the
273
+ tropospheric delay for HAPS signals, 𝜖𝑚𝑝,𝐻𝐴𝑃𝑆 is the delay
274
+ caused by the multipath for HAPS signals. The simulated
275
+ vehicle trajectory originates at Carleton University, which is in
276
+ a suburban area, and ends at Rideau Street, which is in a dense
277
+ urban part of Ottawa (see Fig. 2). There are six simulated HAPS
278
+ shown as transmitters on Fig. 3. As we can see, one HAPS is
279
+ positioned over downtown Ottawa; the other five HAPS are
280
+ positioned nearby, over populated areas and conservation areas.
281
+ HAPS is quasi-stationary, meaning that it will still be moving
282
+ in a variety of manners. In this work, all the HAPS are
283
+
284
+ Fig. 2. Vehicle trajectory.
285
+
286
+
287
+ Fig. 3. Locations of the simulated HAPS.
288
+
289
+
290
+ simulated to be following a circular trajectory with a radius of
291
+ 300 m. The elevation and azimuth angles of all the HAPS at the
292
+ beginning of the simulation are listed in Table I. The positions
293
+ of HAPS were chosen to provide a rich diversity in azimuth
294
+ angles. With one HAPS at the zenith and the others having
295
+ relatively low elevation angles, this constitutes a near Zenith +
296
+ Horizon (ZH) geometry, which can deliver a reasonably good
297
+ DOP [23]. To make sure the entire urban area is well covered,
298
+ HAPS are placed not too far away from the urban area. To better
299
+ understand the concept of DOP, the visual illustrations of the
300
+ HDOP and VDOP of the simulated HAPS constellation are
301
+ provided in Fig. 4 and Fig. 5, respectively. Due to various errors
302
+ impacting the pseudorange measurement, the estimated
303
+ distance between a HAPS and a receiver can be smaller or
304
+ larger than the geometric range. Objects with a higher elevation
305
+ angle will likely result in more uncertainty for the vertical
306
+
307
+ OSM + relief shading
308
+ Tracks:
309
+ Ottawa,On to Rideau Centre,
310
+ Ottawa,ON
311
+ Lyon
312
+ Denseurban
313
+ Areas
314
+ yview
315
+ Fost
316
+ onburg
317
+ Suburban
318
+ Areas
319
+ Ottawa
320
+ Cene:45.40751.75.0087m
321
+ Google
322
+ Map created at CSVisualizer.com
323
+ 一净用多款Transmitter5
324
+ deroure
325
+ Buckinohan
326
+ Transmitter6
327
+ arcde
328
+ otineau
329
+ Transmitter3
330
+ 174
331
+ Transmitter1
332
+ Simulator
333
+ 417
334
+ Embrun
335
+ Russell
336
+ Metcalfe
337
+ Transmitter2
338
+ Transmitter4> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
339
+
340
+
341
+ Fig. 4. HDOP of the simulated HAPS constellation (top view).
342
+
343
+
344
+ Fig. 5. VDOP of the simulated HAPS constellation (front view).
345
+
346
+
347
+ component and less uncertainty for the horizontal component
348
+ from the point of view of geometry, and vice versa. The shaded
349
+ area is where the receiver is estimated to be.
350
+
351
+ A. The Single Point Positioning (SPP) Algorithm
352
+ The single point positioning algorithm is implemented on the
353
+ basis of the SPP package developed by Napat Tongkasem [24]
354
+ with proper modifications [21] so that HAPS can be
355
+ incorporated in the SPP algorithm. Fig. 6 shows the flowchart
356
+ of the single point positioning algorithm. We should point out
357
+ that the implemented single point positioning algorithm is not
358
+ the best positioning algorithm, and that the objective of this
359
+ work is to show the significance of HAPS in aiding the
360
+ positioning performance of a legacy GNSS. The implemented
361
+ SPP algorithm can be improved in many ways. For example, if
362
+ the knowledge of the measurement error variance is available,
363
+ we can apply the weighted least squares (WLS) algorithm to
364
+ enhance the positioning performance of the SPP algorithm by
365
+ lowering the weights of those observations with higher
366
+ variances [25]. If the knowledge of the measurement error
367
+ variance is unavailable, the computational complexity of the
368
+ SPP algorithm can be reduced by imposing the Cholesky
369
+ decomposition for the matrix inversion in (9) [26]. We can also
370
+ TABLE I
371
+ ELEVATION AND AZIMUTH OF THE HAPS AT THE START OF THE SIMULATION
372
+
373
+ HAPS index
374
+ Elevation angle
375
+ Azimuth angle
376
+ HAPS #1
377
+ 81.087°
378
+ -14.210°
379
+ HAPS #2
380
+ 24.054°
381
+ -128.878°
382
+ HAPS #3
383
+ 27.952°
384
+ 68.022°
385
+ HAPS #4
386
+ 32.450°
387
+ 171.477°
388
+ HAPS #5
389
+ 36.554°
390
+ 2.204°
391
+ HAPS #6
392
+ 33.805°
393
+ -57.884°
394
+
395
+
396
+ use the carrier phase measurement to enhance the positioning
397
+ performance of the HAPS-aided GPS system, since carrier
398
+ performance of the HAPS-aided GPS system, since carrier
399
+ phase measurements come with much higher precision, which
400
+ usually delivers a more accurate position solution. Since the
401
+ HAPS clock offset in this work is not explicitly simulated, we
402
+ simply use 𝑑𝑇 to denote the satellite clock offset. From the data
403
+ collected by the GNSS receiver, we shall obtain both the
404
+ receiver independent exchange (RINEX) format observation
405
+ file and the RINEX navigation file, from which we can obtain
406
+ satellite information, such as the satellite pseudorange 𝒑𝑺𝑨𝑻, the
407
+ ionospheric parameters 𝜶 , the Keplerian parameters, the
408
+ pseudo-random noise (𝑷𝑹𝑵) code, which represents the unique
409
+ number of each satellite, the day of year ( 𝐷𝑂𝑌 ) which
410
+ represents the day of year at the time of measurement. We write
411
+ 𝑷𝑹𝑵 in bold to represent a vector containing the pseudo-
412
+ random noise code of all visible satellites at the current epoch.
413
+ We are able to compute the satellite positions, 𝑷𝑺𝑨𝑻 , and
414
+ satellite clock offset, 𝒅𝑻 , using the Keplerian parameters
415
+ contained in the navigation file. 𝑷𝑯𝑨𝑷𝑺 denotes a vector
416
+ containing the positions of all HAPS, which are generated using
417
+ the Skydel GNSS simulator [27], and 𝒑𝑯𝑨𝑷𝑺 denotes a vector
418
+ containing the HAPS pseudorange, which will be explained in
419
+ Section III. To compute the position solution 𝒙 , we first
420
+ initialize the receiver position to the center of the Earth; then
421
+ we initialize the receiver clock offset to zero and the change in
422
+ estimates 𝒅𝒙 to infinity. For each epoch of measurement, we
423
+ first check if the number of available ranging sources is more
424
+ than three, as at least four ranging sources are required to
425
+ perform precise 3D localization. Since the receiver position is
426
+ iteratively estimated, we calculate the elevation angles for both
427
+ satellites and HAPS with respect to the recently estimated
428
+ receiver position. Since both the tropospheric delay and the
429
+ ionospheric delay are functions of the receiver position, these
430
+ two atmospheric delays are estimated iteratively as well. The
431
+ elevation angle, satellite pseudorange, HAPS pseudorange,
432
+ satellite position, satellite clock offset, tropospheric delay
433
+ 𝒅𝒕𝒓𝒐𝒑 , ionospheric delay 𝒅𝒊𝒐𝒏 , and pseudo-random noise
434
+ (𝑷𝑹𝑵) code are modified iteratively on the basis of the re-
435
+ computed elevation angles for both satellites and HAPS. To
436
+ prepare the parameters needed for the least square method, the
437
+ pseudorange needs to be corrected as follows:
438
+
439
+
440
+ 𝒑𝑺𝑨𝑻
441
+ 𝒄
442
+ = 𝒑𝑺𝑨𝑻 + 𝑐 ∙ 𝒅𝑻 − 𝒅𝒕𝒓𝒐𝒑,𝑺𝑨𝑻 − 𝒅𝒊𝒐𝒏,𝑺𝑨𝑻
443
+ (3)
444
+
445
+ where 𝒑𝑺𝑨𝑻
446
+ 𝒄
447
+ represents the corrected pseudorange for the
448
+
449
+ Transmitter 5
450
+ Transmitter 6
451
+ Transmitter 3
452
+ Transmitter 2
453
+ Transmitter 4Transmitter 5
454
+ Transmitter 6
455
+ Transmitter 1
456
+ Transmitter 3
457
+ Transmitter 2
458
+ Transmitter 4> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
459
+
460
+
461
+ Fig. 6. Flow chart of the single point positioning algorithm.
462
+
463
+
464
+ satellites, and 𝒑𝑺𝑨𝑻 represents the uncorrected pseudorange for
465
+ the satellites. In this work, the HAPS pseudorange is modeled
466
+ as the sum of the geometric range and the pseudorange
467
+ error,which represents the overall estimation error of the
468
+ parameters in the HAPS pseudorange equation. Accordingly,
469
+ the HAPS pseudorange does not need to be corrected. Due to
470
+ the Earth’s rotation, the positions of satellites and HAPS at the
471
+ signal emission time are different from their positions at the
472
+ signal reception time; this is known as the Sagnac effect [28].
473
+ The coordinates of satellites/HAPS can be transformed from the
474
+ signal emission time to the signal reception time by [28]
475
+
476
+
477
+ ∆𝑡𝑅𝑂𝑇 = 𝑡𝑟𝑥 − 𝑡𝑡𝑥
478
+ (4)
479
+
480
+
481
+ 𝑃𝑖,𝑟𝑥 = 𝑀𝑅𝑂𝑇(𝜔𝐸 × ∆𝑡𝑅𝑂𝑇)𝑃𝑖,𝑡𝑥
482
+ (5)
483
+
484
+ where ∆𝑡𝑅𝑂𝑇 denotes the signal propagation time, 𝑡𝑟𝑥
485
+ represents the signal reception time, 𝑡𝑡𝑥 represents the signal
486
+ emission time, 𝑃𝑖,𝑟𝑥 is the 𝑖𝑡ℎ satellite/HAPS coordinates at the
487
+ signal reception time, 𝑃𝑖,𝑡𝑥 is the 𝑖𝑡ℎ satellite/HAPS coordinates
488
+ at the signal emission time, 𝜔𝐸 denotes the Earth’s rotation rate,
489
+ and 𝑀𝑅𝑂𝑇(𝜔𝐸 × ∆𝑡𝑅𝑂𝑇) is known as the rotation matrix, which
490
+ is described as follows:
491
+
492
+
493
+ 𝑀𝑅𝑂𝑇(𝜔𝐸 × ∆𝑡𝑅𝑂𝑇)
494
+ = [
495
+ cos(𝜔𝐸 × ∆𝑡𝑅𝑂𝑇)
496
+ sin(𝜔𝐸 × ∆𝑡𝑅𝑂𝑇)
497
+ 0
498
+ − sin(𝜔𝐸 × ∆𝑡𝑅𝑂𝑇)
499
+ 0
500
+ cos(𝜔𝐸 × ∆𝑡𝑅𝑂𝑇)
501
+ 0
502
+ 0
503
+ 1
504
+ ] .
505
+
506
+
507
+ (6)
508
+ The line-of-sight vector 𝒗, and the true range between ranging
509
+ sources and receiver 𝝆, are then calculated to compute the a
510
+ priori range residual vector 𝒃 and the design matrix 𝑯, where
511
+
512
+
513
+ 𝒃 = 𝒑𝒄 − 𝝆
514
+ (7)
515
+
516
+
517
+ 𝑯 = [𝒗, 𝟏𝑙𝑒𝑛𝑔𝑡ℎ(𝑷𝐜)×1]
518
+ (8)
519
+
520
+ where 𝒑𝒄 is the corrected satellite pseudorange combined with
521
+ the corrected HAPS pseudorange, 𝟏𝑙𝑒𝑛𝑔𝑡ℎ(𝑷𝐜)×1 denotes a
522
+ column vector of length being the total number of visible
523
+ ranging sources, and 𝑷𝐜 is a vector containing the corrected
524
+ positions of the visible ranging sources (i.e., satellite + HAPS).
525
+ Finally, the least square solution is computed as follows:
526
+
527
+
528
+ 𝑸 = (𝑯′𝑯)−1
529
+ (9)
530
+
531
+
532
+ 𝒅𝒙 = 𝑸𝑯′𝒃
533
+ (10)
534
+
535
+
536
+ 𝑑𝑡 = 𝒅𝒙(4)/𝑐
537
+ (11)
538
+
539
+ where Q is known as the covariance matrix, and 𝒅𝒙(4) denotes
540
+ the fourth element in the vector 𝒅𝒙. To prevent the algorithm
541
+ from getting numerical issues, we should ensure the term being
542
+ inversed in (9) is non-singular; in other words, the design matrix
543
+ 𝑯 should be non-singular. With the extra observations by
544
+ utilizing HAPS as additional ranging sources, the chance of 𝑯
545
+ being singular is likely reduced; the non-singular design matrix
546
+ can be ensured by avoiding the use of collinear observations,
547
+ which means that two or more observations have about the
548
+ same azimuth and elevation angle. We observe that the term
549
+ being inversed in (9), 𝑯′𝑯, is a Hermitian, positive definite
550
+ matrix, therefore the Cholesky decomposition can be imposed
551
+ to reduce the computational complexity [26]. The covariance
552
+ matrix, 𝑸, is described by
553
+
554
+
555
+ 𝑸 =
556
+ [
557
+
558
+
559
+
560
+ 𝜎𝑥
561
+ 2
562
+ 𝜎𝑥𝑦
563
+ 𝜎𝑥𝑧
564
+ 𝜎𝑥𝑡
565
+ 𝜎𝑥𝑦
566
+ 𝜎𝑦
567
+ 2
568
+ 𝜎𝑦𝑧
569
+ 𝜎𝑦𝑡
570
+ 𝜎𝑥𝑧
571
+ 𝜎𝑦𝑧
572
+ 𝜎𝑧
573
+ 2
574
+ 𝜎𝑧𝑡
575
+ 𝜎𝑥𝑡
576
+ 𝜎𝑦𝑡
577
+ 𝜎𝑧𝑡
578
+ 𝜎𝑡
579
+ 2 ]
580
+
581
+
582
+
583
+
584
+
585
+ (12)
586
+
587
+ Initialization
588
+ PsAT-PHAPS,PRN,DOY
589
+ x = 04x1
590
+ Input
591
+ dt = x(4)/c
592
+ PHAPS, PsAT, dT,α
593
+ dx = x + Inf
594
+ stop = 0
595
+ Exit
596
+ No
597
+ NsAr + NHAPs ≥ 4
598
+ IYes
599
+ No
600
+ dx(1:3)|> 0.01
601
+ Yes
602
+ Finding parameters
603
+ For sotellites
604
+ For HAPS
605
+ OsAT,dtrop.dion
606
+ HAPS
607
+ Applying elevation mask
608
+ For sotellites
609
+ For HAPS
610
+ sAT,dtrep,dion,PRN,dT,PsAr
611
+ HAPS,PHAPS
612
+ Pseudorange correction
613
+ PSAT-PHAPS
614
+ Combining the
615
+ corrected pseudoranges
616
+ p' = [psar.phaes]
617
+
618
+ Correcting for the Sagnaceffect
619
+ (i.e., Earth rotation)
620
+ PSAT,PHAPS
621
+ Combining the corrected
622
+ ranging source positions
623
+ P° = [PSAT.PHAPs]
624
+ Finding parameters
625
+ V,p,b,H,Q
626
+ andno
627
+ Computing the position solution
628
+ using the Least Square method
629
+ 3p'x
630
+ x,dt> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
631
+
632
+ where receiver coordinates x, y, z in the Earth-centered Earth-
633
+ fixed (ECEF) coordinate frame and the receiver clock offset,
634
+ respectively. The least square solution will be found when the
635
+ norm of the change in receiver position 𝒅𝒙(1: 3) is sufficiently
636
+ small. In this work, this threshold is set as 0.01 m. We use the
637
+ HDOP, the VDOP and the 3D positioning accuracy as the
638
+ metrics to show the advantage of the proposed HAPS-aided
639
+ GPS system; the 3D positioning accuracy is used to show the
640
+ applicability of the RAIM to the HAPS-aided GPS system. To
641
+ compute the HDOP, we need to convert the covariance matrix
642
+ into the local north-east-down (NED) coordinate frame, which
643
+ can be done with the following equations [29]:
644
+
645
+
646
+ 𝑸𝑵𝑬𝑫 = 𝑹′𝑸̃𝑹 = [
647
+ 𝜎𝑛
648
+ 2
649
+ 𝜎𝑛𝑒
650
+ 𝜎𝑛𝑑
651
+ 𝜎𝑛𝑒
652
+ 𝜎𝑒
653
+ 2
654
+ 𝜎𝑒𝑑
655
+ 𝜎𝑛𝑑
656
+ 𝜎𝑒𝑑
657
+ 𝜎𝑑
658
+ 2
659
+ ]
660
+ (13)
661
+
662
+
663
+ 𝑸̃ = [
664
+ 𝜎𝑥
665
+ 2
666
+ 𝜎𝑥𝑦
667
+ 𝜎𝑥𝑧
668
+ 𝜎𝑥𝑦
669
+ 𝜎𝑦
670
+ 2
671
+ 𝜎𝑦𝑧
672
+ 𝜎𝑥𝑧
673
+ 𝜎𝑦𝑧
674
+ 𝜎𝑧
675
+ 2
676
+ ]
677
+ (14)
678
+
679
+
680
+ 𝑹 = [
681
+ −sin 𝜆
682
+ cos 𝜆
683
+ 0
684
+ − cos 𝜆 sin 𝜑
685
+ − sin 𝜆 sin 𝜑
686
+ cos 𝜑
687
+ cos 𝜆 cos 𝜑
688
+ sin 𝜆 cos 𝜑
689
+ sin 𝜑
690
+ ]
691
+ (15)
692
+
693
+ where 𝜎𝑛, 𝜎𝑒, and 𝜎𝑑 represent the receiver position errors in
694
+ the local north, east, and down directions, respectively. 𝜆 and 𝜑
695
+ represent the longitude and latitude of the receiver,
696
+ respectively. Then, the HDOP is described by
697
+
698
+
699
+ 𝐻𝐷𝑂𝑃 = √𝜎𝑛2 + 𝜎𝑒2
700
+ (16)
701
+
702
+ and the VDOP is described by
703
+
704
+
705
+ 𝑉𝐷𝑂𝑃 = √𝜎𝑑
706
+ 2.
707
+ (17)
708
+
709
+ B. The Receiver Autonomous Integrity Monitoring (RAIM)
710
+ Algorithm
711
+ The RAIM algorithm is a signal selection algorithm that can
712
+ detect and even exclude abnormal observations using redundant
713
+ measurements. It can detect an abnormal observation when the
714
+ number of observations is at least five; it can exclude this
715
+ abnormal observation when the number of observations is at
716
+ least six. The RAIM algorithm is typically applied to multi-
717
+ constellation GNSSs where the number of ranging sources is
718
+ more than enough to perform precise 3D localization, and it is
719
+ typically applied to cases where there likely exists at least one
720
+ observation that differs from the expected value significantly.
721
+ Such cases include urban areas, where the pseudorange
722
+
723
+
724
+
725
+
726
+
727
+
728
+
729
+
730
+ measurement is highly subject to the multipath effect. With the
731
+ assistance from HAPS, the chance of enabling the RAIM
732
+ function will likely increase. Typical RAIM algorithms tend to
733
+ use the standard deviation of the target observable, which is the
734
+ pseudorange measurement in our work. As knowledge of the
735
+ standard deviation of the satellite pseudorange is unavailable on
736
+ the receivers we use, in this work the RAIM algorithm is
737
+ implemented on the basis of [30], which considers a 𝐶/𝑁0-
738
+ based variance model and a computationally efficient method,
739
+ namely the modified Danish estimation method. 2 The
740
+ implemented 𝐶/𝑁0-based RAIM algorithm is given in Alg. 1,
741
+ where 𝑁 denotes the number of visible ranging sources. The
742
+ input to this algorithm consists of the position fix computed
743
+ using the SPP algorithm 𝒙, and the 𝑪/𝑵𝟎 of the ranging source
744
+ signal. Since HAPS are located at much lower altitudes than
745
+ 2 To the best of our knowledge, RAIM is the most common algorithm used
746
+ for integrity monitoring. Since we only have the 𝐶/𝑁0 data which can be
747
+ utilized for the integrity monitoring, we could not identify in the literature any
748
+ other appropriate RAIM-like algorithm for comparison. However, we believe
749
+ that the other RAIM algorithms would also be applicable if the knowledge of
750
+ the standard deviation of the satellite pseudorange happens to be available.
751
+ Algorithm 1 The 𝐶/𝑁0-based RAIM Algorithm
752
+ Input: The SPP estimated position solution 𝒙 and 𝑪/𝑵𝟎;
753
+ Output: The SPP and RAIM jointly estimated position
754
+ solution 𝒙̂.
755
+ 1:
756
+ Initialize the parameters 𝑠𝑡𝑜𝑝 and 𝒅𝒙;
757
+ 2:
758
+ while |𝒅𝒙(1: 3)| > 0.01 do
759
+ 3:
760
+
761
+ Same procedures as the SPP algorithm until
762
+ “Finding parameters” after correcting for the
763
+ Sagnac effect;
764
+ 4:
765
+
766
+ for 𝑖 = 1 to 𝑁 do
767
+ 5:
768
+
769
+
770
+ if 𝑠𝑡𝑜𝑝 == 1 do
771
+ 6:
772
+
773
+
774
+
775
+ Find the variance of the observation 𝑖, 𝑠𝑖,
776
+ according to (19);
777
+ 7:
778
+
779
+
780
+ end if
781
+ 8:
782
+
783
+ end for
784
+ 9:
785
+
786
+ Find the weight matrix 𝑾 and the design matrix
787
+ 𝑯 , and calculate the covariance matrix 𝑸
788
+ according to (20);
789
+ 10:
790
+
791
+ Calculate the change in estimates 𝒅𝒙 according
792
+ to (21), and update the position solution 𝒙;
793
+ 11:
794
+
795
+ Calculate the pseudorange residual 𝒗̂ according
796
+ to (22), and the covariance matrix of the residuals
797
+ 𝑪𝒗̂ according to (24);
798
+ 12:
799
+
800
+ for 𝑖 = 1 to 𝑁 do
801
+ 13:
802
+
803
+
804
+ Find the normalized residual of observation 𝑖
805
+ at the current iteration 𝑘, 𝑤̅𝑖,𝑘 according to
806
+ (26);
807
+ 14:
808
+
809
+
810
+ if |𝑤̅𝑖,𝑘| > 𝑛1−(𝛼0/2) do
811
+ 15:
812
+
813
+
814
+
815
+ Update the variance of the observation 𝑖
816
+ for the next iteration 𝑘 + 1 , 𝜎𝑖,𝑘+1
817
+ 2
818
+ ,
819
+ according to (25);
820
+ 16:
821
+
822
+
823
+ end if
824
+ 17:
825
+
826
+ end for
827
+ 18:
828
+ end while
829
+
830
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
831
+
832
+ satellites, in practice the 𝐶/𝑁0 value of the HAPS might be
833
+ higher than that for any satellite. As it is possible that a handful
834
+ of HAPS signals might suffer from severe multipath effects, we
835
+ can exclude those HAPS signals whose 𝐶/𝑁0 values are much
836
+ lower than the higher ones. In this work, the multipath effect is
837
+ not explicitly simulated for the HAPS signal; therefore, we
838
+ assume that the 𝐶/𝑁0 of each HAPS is equal to the maximum
839
+ 𝐶/𝑁0 value of the available satellites at each epoch, meaning
840
+ that the signal quality for a HAPS will always be better than
841
+ that for any satellites. The variance covariance matrix (VCM)
842
+ 𝜮 of the observations (pseudoranges) 𝒑 is defined as follows:
843
+
844
+
845
+ 𝜮 = 𝑑𝑖𝑎𝑔(𝑠1, 𝑠2, … , 𝑠𝑛)
846
+ (18)
847
+
848
+
849
+ 𝑠𝑖 = 10 + 1502 ∗ 10(−𝐶/𝑁0,𝑖)/10
850
+ (19)
851
+
852
+ where 𝑠𝑖 denotes the variance of the observation 𝑖. We assume
853
+ that the observations are uncorrelated, and that the errors follow
854
+ the normal distribution with 𝑁(𝟎, 𝜮). The weight matrix, 𝑾, is
855
+ defined as the inverse of the VCM, 𝜮−1 . The least square
856
+ equations become
857
+
858
+
859
+ 𝑸 = (𝑯′𝑾𝑯)−1
860
+ (20)
861
+
862
+
863
+ 𝒅𝒙 = 𝑸𝑯′𝑾𝑷.
864
+ (21)
865
+
866
+ The least square residuals of the pseudorange 𝒗̂ can be obtained
867
+ as follows:
868
+
869
+
870
+ 𝒗̂ = 𝑯 ∙ 𝒅𝒙 − 𝑷
871
+ (22)
872
+
873
+
874
+ 𝑷 = 𝒑𝒄 − 𝝆
875
+ (23)
876
+
877
+ where 𝑯 represents the design matrix, 𝒅𝒙 represents the
878
+ change in estimates, 𝒑𝒄 denotes the corrected pseudoranges,
879
+ and 𝝆 denotes the geometric range between ranging sources
880
+ and the receiver. The covariance matrix of the residuals, 𝑪𝒗̂, is
881
+ computed as
882
+
883
+
884
+ 𝑪𝒗̂ = 𝜮 − 𝑯(𝑯𝑇𝜮−1𝑯)−1𝑯𝑇.
885
+ (24)
886
+
887
+ To detect and exclude the abnormal observations, we follow the
888
+ modified Danish estimation method proposed in [30].
889
+
890
+
891
+ 𝜎𝑖,𝑘+1
892
+ 2
893
+ = 𝜎𝑖,0
894
+ 2 ∙ {exp (
895
+ |𝑤̅𝑖,𝑘|
896
+ 𝑇 ) , |𝑤̅𝑖,𝑘| > 𝑛1−(𝛼0/2)
897
+ 1, |𝑤̅𝑖,𝑘| ≤ 𝑛1−(𝛼0/2)
898
+
899
+ (25)
900
+
901
+ with
902
+
903
+
904
+ 𝑤̅𝑖,𝑘 =
905
+ 𝒗̂𝑖,𝑘
906
+ √(𝑪𝒗̂𝒊,𝟏)𝑖𝑖
907
+
908
+ (26)
909
+
910
+ where 𝜎𝑖,0
911
+ 2 denotes the a priori variance of the observation 𝑖
912
+ (i.e., s𝑖), 𝑤̅𝑖,𝑘 denotes the normalized residual of observation 𝑖
913
+ at iteration 𝑘,√(𝑪𝒗̂𝒊,𝟏)𝑖𝑖 represents the standard deviation of
914
+ observation 𝑖 from the first iteration, 𝑛1−(𝛼0/2) denotes the 𝛼0-
915
+ quantile of the standard normal distribution, which is also called
916
+ the critical value, 𝛼0 is the predetermined false alarm rate
917
+ which is 0.5 % in this work. The modified Danish method is an
918
+ iteratively reweighted LS algorithm that implements a robust
919
+ estimator. This method detects and excludes abnormal
920
+ observations by comparing the absolute value of each
921
+ normalized pseudorange residual, |𝑤̅𝑖,𝑘|, with the critical value,
922
+ 𝑛1−(𝛼0/2), in each iteration. The variances for the observations
923
+ whose normalized residuals are greater than the critical value
924
+ are multiplied with exponential terms, making the variances of
925
+ those observations larger, hence lowering the weight of those
926
+ observations. By iteratively multiplying the variances of the
927
+ abnormal observations by exponential terms, the weight of the
928
+ abnormal observations will likely become much smaller than
929
+ that of the normal observations; therefore, the abnormal
930
+ observations can be considered as being excluded.
931
+ III. SIMULATION OF THE HAPS-AIDED GPS SYSTEM
932
+ In this section, we will describe the simulation setup used in
933
+ the Skydel GNSS software [27] and present the simulation
934
+ results in terms of 3D positioning accuracy for several hybrid
935
+ systems and two standalone systems in both a suburban
936
+ scenario and a dense urban scenario.
937
+
938
+ A. Simulation Setup
939
+ The system model is established using the default Earth
940
+ orientation parameters of the Skydel GNSS simulation software
941
+ [27], which considers all GPS satellites orbiting around the
942
+ Earth and transmitting the L1 C/A code. The Saastamoinen
943
+ model is chosen to emulate the tropospheric effect, and the
944
+ Klobuchar model is chosen to emulate the ionospheric effect
945
+ using the default Klobuchar parameters that come with the
946
+ software. The output from Skydel contains the ECEF
947
+ coordinates of satellites at the signal emission time, the
948
+ ionospheric corrections, the tropospheric corrections, the
949
+ satellite clock offsets, the ECEF coordinates of the receiver, the
950
+ signal emission time, and so forth, at each time stamp from the
951
+ start of the simulation. The receiver clock offset in the
952
+ simulation is zero by default. The correction terms in the
953
+ pseudorange equation of satellite including the satellite orbit
954
+ error, the multipath error, and the receiver noise are not
955
+ separately considered in the simulation; instead, a pseudorange
956
+ error is introduced to reflect the presence of those effects. The
957
+ pseudorange error of satellite is featured using the built-in first
958
+ order Gauss-Markov process with a default time constant of 10
959
+ s, and the standard deviation of 6 m. The continuous model for
960
+ the first order Gauss-Markov process is described by [31]:
961
+
962
+
963
+ ������̇ = −
964
+ 1
965
+ 𝑇𝑐 ������̇ + 𝑤
966
+ (27)
967
+
968
+ where ������̇ represents a random process with zero mean,
969
+ correlation time 𝑇𝑐, and noise 𝑤. The autocorrelation of the first
970
+
971
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
972
+
973
+ TABLE Ⅱ
974
+ DETAILS OF THE SIMULATION SETUP
975
+
976
+ Item
977
+ Processing strategy
978
+ Earth orientation parameter
979
+ Software default Earth orientation parameters
980
+ Satellite signal
981
+ GPS L1 C/A
982
+ Tropospheric model
983
+ Saastamoinen model
984
+ Ionospheric model
985
+ Klobuchar model
986
+ Sampling rate
987
+ 12.5 MS/s
988
+ Satellite pseudorange error
989
+ 1st order Gauss-Markov process (time constant = 10 s; standard deviation = 6 m)
990
+ HAPS pseudorange error
991
+ Gaussian noise (mean = 0 m; std = 2 m for the suburban scenario; std = 5 m for the dense urban
992
+ scenario)
993
+ Number of GPS satellites (dense urban
994
+ scenario)
995
+ 8-10
996
+ Number of GPS satellites (suburban
997
+ scenario)
998
+ 4
999
+ Total number of HAPS
1000
+ 6
1001
+ order Gauss-Markov process is described by [32]:
1002
+
1003
+ 𝑅(𝛥𝑡) = 𝜎2𝑒−|𝛥𝑡|
1004
+ 𝜏 (28)
1005
+
1006
+ where 𝛥𝑡 represents the sampling interval, 𝜎 and 𝜏 denote the
1007
+ standard deviation and the time constant of the first order
1008
+ Gauss-Markov process, respectively. The characteristics of the
1009
+ pseudorange errors for satellites are set to be the same in both
1010
+ the suburban scenario and the dense urban scenario. However,
1011
+ we randomly select four satellites in the dense urban scenario
1012
+ in order to emulate the dense urban area in a rather simple way.
1013
+ We verify that by doing so, the standard deviation of the 3D
1014
+ positioning accuracy for the GPS-only system in the simulation
1015
+ is close to that in the physical experiment. The pseudorange
1016
+ error for the HAPS is modeled using the Gaussian noise with
1017
+ standard deviations of 2 m and 5 m, representing the suburban
1018
+ and the dense urban scenario, respectively. Under the
1019
+ assumption that the overall estimation error of the HAPS is less
1020
+ than that of the satellite, the standard deviation for the HAPS
1021
+ pseudorange error is deliberately set to be smaller than that of
1022
+ the satellite pseudorange error in the suburban and dense urban
1023
+ scenarios. To investigate the impact of the number of HAPS on
1024
+ the positioning performance of the HAPS-aided GPS system,
1025
+ we consider four hybrid systems with different numbers of
1026
+ randomly selected HAPS at each epoch. We also consider the
1027
+ HAPS-only system for the completeness of a research problem.
1028
+ Under this setting, we examine the 3D positioning performance
1029
+ of different systems in the suburban and dense urban scenarios.
1030
+ In the suburban scenario, the number of visible satellites varies
1031
+ between eight and ten, while in the dense urban scenario the
1032
+ number of visible satellites is set to four. The details of the
1033
+ simulation setup are given in Table Ⅱ.
1034
+
1035
+ B. Simulation Results
1036
+ Fig. 7 shows the cumulative distribution function (CDF) of
1037
+ the 3D positioning accuracy for different positioning systems in
1038
+ the suburban scenario. With the assumption that the
1039
+ pseudorange error for a HAPS is smaller than that of a satellite,
1040
+ we can see from Fig. 7 that all the hybrid systems (HAPS +
1041
+
1042
+ Fig. 7. CDF of the 3D positioning accuracy for different systems (suburban
1043
+ scenario).
1044
+
1045
+
1046
+ Fig. 8. CDF of the 3D positioning accuracy for different systems (dense urban
1047
+ scenario).
1048
+
1049
+
1050
+ GPS) outperform the GPS-only system; the more HAPS, the
1051
+ better the positioning performance of the HAPS-aided GPS
1052
+
1053
+ Suburbanscenario(stdofHAPSpseudorangeerror=2m)
1054
+ 0.9
1055
+ 0.8
1056
+ 0.7
1057
+ 0.6
1058
+ D
1059
+ 0.5
1060
+ 0.4
1061
+ 0.3
1062
+ 0.2
1063
+ GPS-only system
1064
+ 1-HAPS with GPS system
1065
+ 2-HAPS with GPS system
1066
+ 3-HAPS with GPS system
1067
+ 0.1
1068
+ 4-HAPS with GPS system
1069
+ 4-HAPS-only system
1070
+ 0
1071
+ 5
1072
+ 10
1073
+ 15
1074
+ 20
1075
+ 25
1076
+ 30
1077
+ 3D positional accuracy (m) in local NED frameDenseurbanscenario(stdofHAP
1078
+ pseudorangeerror=5m)
1079
+ 0.9
1080
+ 0.8
1081
+ 0.7
1082
+ 0.6
1083
+ 0.5
1084
+ 0.4
1085
+ 0.3
1086
+ 0.2
1087
+ 4-GPS-only system
1088
+ 1-HAPS with 4-GPS system
1089
+ 2-HAPS with 4-GPS system
1090
+ 0.1
1091
+ 3-HAPS with 4-GPS system
1092
+ 4-HAPS with 4-GPS system
1093
+ 4-HAPS-only system
1094
+ 0
1095
+ 10
1096
+ 20
1097
+ 30
1098
+ 40
1099
+ 50
1100
+ 60
1101
+ 70
1102
+ 80
1103
+ 90
1104
+ 100
1105
+ 3D positionalaccuracy (m)inlocal NEDframe> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1106
+
1107
+
1108
+ Fig. 9. Positioning performance of M8T and M8U for horizontal and vertical
1109
+ planes.
1110
+
1111
+
1112
+ system. Nevertheless, we observe that the positioning
1113
+ performance of the 4-HAPS-only system, where all ranging
1114
+ sources have a much smaller pseudorange error than the
1115
+ satellite, is not the best and can occasionally be very poor. This
1116
+ may be due to the following reasons: 1) the 4-HAPS-only
1117
+ system has much fewer ranging sources for computing receiver
1118
+ positions; 2) the HAPS geometry can be poor occasionally since
1119
+ we are randomly selecting four HAPS at each epoch. There are
1120
+ several cases where the HAPS geometry is considered poor. For
1121
+ example, when the four randomly selected HAPS are on the
1122
+ same side of the receiver. The CDF of the 3D positioning
1123
+ accuracy for different positioning systems in the dense urban
1124
+ scenario is shown in Fig. 8, from which we can see a similar
1125
+ trend that the more HAPS the better the positioning
1126
+ performance of the HAPS-aided GPS system. In the dense
1127
+ urban scenario, where only four GPS satellites are selected for
1128
+ positioning, the 4-HAPS-only system achieves better
1129
+ positioning performance than the 4-GPS-only system due to the
1130
+ better signal quality for the HAPS. However, we should
1131
+ consider using the HAPS-aided GPS system for the best
1132
+ positioning performance.
1133
+ IV. FIELD EXPERIMENTS
1134
+ To verify and support the simulation results, we process the
1135
+ raw GNSS data collected using two commercial GNSS
1136
+ receivers. In this section, we present the experiment setup, the
1137
+ modeling of HAPS signals, and the HAPS pseudorange error.
1138
+ We also provide an analysis of the DOP and the 3D positioning
1139
+ accuracy for both the GPS-only system and the HAPS-aided
1140
+ GPS system.
1141
+
1142
+ A. Experiment Setup
1143
+ The raw GNSS data is collected along a vehicle trajectory
1144
+ similar to the one shown in Fig. 2, except for a slight difference
1145
+ due to a partial road closure on the day of data collection. The
1146
+ raw GNSS data is collected using both the Ublox EVK-M8T
1147
+ TABLE Ⅲ
1148
+ EVK-M8T GNSS UNIT SPECIFICATIONS [35]
1149
+
1150
+ Parameter
1151
+ Specification
1152
+ Serial Interfaces
1153
+ 1 USB V2.0
1154
+ 1 RS232, max.baud rate 921,6 kBd
1155
+ DB9 +/- 12 V level
1156
+ 14 pin – 3.3 V logic
1157
+ 1 DDC (I2C compatible) max. 400 kHz
1158
+ 1 SPI-clock signal max. 5,5 MHz – SPI DATA
1159
+ max. 1 Mbit/s
1160
+ Timing Interfaces
1161
+ 2 Time-pulse outputs
1162
+ 1 Time-mark input
1163
+ Dimensions
1164
+ 105 × 64 × 26 mm
1165
+ Power Supply
1166
+ 5 V via USB or external powered via extra power
1167
+ supply pin 14 (V5_IN) 13 (GND)
1168
+ Normal Operating
1169
+ Temperature
1170
+ −40℃ to +65℃
1171
+
1172
+
1173
+ and the Ublox EVK-M8U GNSS units. The Ublox EVK-M8T
1174
+ unit is a timing product that can provide users with precise
1175
+ timing information for post-processing; the Ublox EVK-M8U
1176
+ unit is a dead reckoning product equipped with inertial
1177
+ measurement units (IMUs) such that the positioning
1178
+ performance of this product will not be degraded much even in
1179
+ the dense urban area. Fig. 9 shows the positioning performance
1180
+ along both horizontal and vertical planes for both M8T and
1181
+ M8U during the entire observation period. From Fig. 9, we can
1182
+ see that M8U outperforms M8T for both horizontal positioning
1183
+ accuracy and vertical positioning accuracy. As only M8T
1184
+ provides the timing information required for post-processing,
1185
+ the receiver positions computed by EVK-M8U are used as the
1186
+ ground truth data for the analysis of the positioning
1187
+ performance. The raw GNSS data is processed using the single
1188
+ point positioning algorithm described in Section Ⅱ. Table Ⅲ
1189
+ gives the specifications of the EVK-M8T GNSS unit. To
1190
+ emulate realistic LOS conditions for HAPS in the urban area,
1191
+ the LOS probability as a function of the HAPS elevation angle
1192
+ in the urban area is implemented on the basis of [33] and [34].
1193
+ It is worth mentioning that the LOS probability model for the
1194
+ HAPS provided by [33] is proposed based on the city of
1195
+ Chicago; imposing this LOS probability model for the dense
1196
+ urban area of Ottawa might be too harsh considering their
1197
+ incompatible city scales. Since there is no LOS probability
1198
+ corresponding to the suburban area in [34], the one for rural area
1199
+ in [34] is used as the LOS probability for the HAPS in the
1200
+ suburban area. The LOS probability for the HAPS in the rural
1201
+ area in [34] is verified as being consistent with the LOS
1202
+ probability for the HAPS in the suburban area in [33]. The
1203
+ pseudorange of the HAPS in the experiment is modeled as the
1204
+ addition of the geometric range between the satellite and
1205
+ receiver, the receiver clock offset multiplied by the speed of
1206
+ light, and the pseudorange error representing the sum of all
1207
+ kinds of estimation errors. The pseudorange errors for the
1208
+ HAPS in the suburban and dense urban areas are simulated as
1209
+ Gaussian noise with zero mean and standard deviations of 2 m
1210
+ and 5 m, respectively. Since the vehicle trajectory involves both
1211
+
1212
+ The positioning performance of M8T and M8U
1213
+ 12
1214
+ Horizontal positioning accuracy (M8T)
1215
+ Horizontalpositioningaccuracy(M8U)
1216
+ Vertical positioningaccuracy(M8T)
1217
+ 10
1218
+ Verticalpositioningaccuracy(M8U)
1219
+ g accuracy (m)
1220
+ 8
1221
+ 6
1222
+ Positioning
1223
+ 2
1224
+ 0
1225
+ 100
1226
+ 200
1227
+ 300
1228
+ 400
1229
+ 500
1230
+ 600
1231
+ 700
1232
+ epoch (s)> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1233
+
1234
+
1235
+
1236
+ Fig. 11. HDOP (top) and VDOP (bottom).
1237
+
1238
+
1239
+ Fig. 10. HAPS availability during the entire observation period.
1240
+
1241
+
1242
+ suburban and dense urban areas, the entire route is divided into
1243
+ two segments, where the first segment is considered as the
1244
+ suburban scenario while the second segment is considered as
1245
+ the dense urban scenario (see Fig. 2). By observing the
1246
+ positioning performance of the GPS-only system using the real
1247
+ GPS data, the LOS probability for the suburban area is applied
1248
+ to the HAPS for epochs less than 380 s, and the LOS probability
1249
+ for the dense urban area is applied to the HAPS for epochs
1250
+ greater than or equal to 380 s. Since the GNSS receivers we use
1251
+ do not provide an accurate receiver clock offset with respect to
1252
+ the GPS time, the receiver clock offset in each epoch is
1253
+ estimated by using the ground truth receiver positions provided
1254
+ from Ublox EVK-M8U and the precise timing information,
1255
+ such as the receiver clock drift and the receiver clock bias,
1256
+ provided from Ublox EVK-M8T. We should note that the
1257
+ pseudorange of the HAPS in the experiment is modeled as a
1258
+ function of the receiver clock offset, which is estimated with
1259
+ best effort. Nevertheless, additional errors should be expected
1260
+ in the pseudorange of the HAPS with the magnitude depending
1261
+ on the quality of all visible satellite signals and the ground truth
1262
+ receiver position. As we would expect the quality of the satellite
1263
+ signals in the suburban area to be better than that in the dense
1264
+ urban area, we should also expect the receiver clock offset to be
1265
+ estimated with higher accuracy in the suburban area than in the
1266
+ dense urban area.
1267
+
1268
+ B. Experiment Results
1269
+ With more ranging sources, we should expect the availability
1270
+ of the HAPS-aided GPS system to be higher than the GPS-only
1271
+ system. The availability of HAPS and GPS satellite during the
1272
+ entire course of observation is shown in Fig. 10. As we can see,
1273
+ the availability of the HAPS-aided GPS system during the
1274
+ entire observation period is 100 %, while the availability of the
1275
+ GPS-only system is 99.71 % as there are two epochs (circled in
1276
+ a black ellipse) where the number of GPS satellites is three.
1277
+ While the difference between the availability of the HAPS-
1278
+ aided GPS system and the GPS-only system is not significant,
1279
+ this is probably because Ottawa is a relatively small metro city
1280
+ compared to the metro cities like Chicago. In the following, we
1281
+ first present a comparison of the HDOP and VDOP between the
1282
+ HAPS-aided GPS system and the GPS-only system. Next, we
1283
+ analyze the 3D positioning performance for both the GPS-only
1284
+ system and the HAPS-aided GPS system. To show the
1285
+ improvements brought by the RAIM algorithm, we compare the
1286
+ RAIM-enabled positioning systems, where both the SPP and
1287
+ the RAIM algorithms are applied, with the RAIM-disabled
1288
+ positioning systems, where only the SPP algorithm is applied.
1289
+
1290
+ 1) Dilution of Precision Analysis
1291
+ Fig. 11 shows the HDOP and VDOP of the GPS-only system
1292
+ and the HAPS-aided GPS system. As we can see, both the
1293
+ HDOP and VDOP of the HAPS-aided GPS system are better
1294
+ than that of the GPS-only system. In particular, we notice that
1295
+ there are fewer spikes on the HDOP and VDOP performance of
1296
+ the HAPS-aided GPS system, which demonstrates that the
1297
+
1298
+ 15
1299
+ GPS-only system
1300
+ HAPS-aided GPS system
1301
+ HDOP
1302
+ 5
1303
+ 0
1304
+ 0
1305
+ 100
1306
+ 200
1307
+ 300
1308
+ 400
1309
+ 500
1310
+ 600
1311
+ 700
1312
+ epoch (s)20
1313
+ GPS-only system
1314
+ HAPS-aided GPS system
1315
+ (w)
1316
+ 15
1317
+ VDOI
1318
+ 10
1319
+ 5
1320
+ DAM
1321
+ 0
1322
+ 0
1323
+ 100
1324
+ 200
1325
+ 300
1326
+ 400
1327
+ 500
1328
+ 600
1329
+ 700
1330
+ epoch(s)11
1331
+ 10
1332
+ HAPSs/satellites
1333
+ 9
1334
+ 8
1335
+ visible
1336
+ 6
1337
+ 5
1338
+ of
1339
+ Number
1340
+ 4
1341
+ 3
1342
+ 2
1343
+ HAPS
1344
+ GPSsatellite
1345
+ 1
1346
+ 0
1347
+ 100
1348
+ 200
1349
+ 300
1350
+ 400
1351
+ 500
1352
+ 600
1353
+ 700
1354
+ epoch (s)> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1355
+
1356
+
1357
+ Fig. 12. CDF of the 3D positioning accuracy in the suburban area.
1358
+
1359
+
1360
+ Fig. 13. CDF of the 3D positioning accuracy in the dense urban area.
1361
+
1362
+
1363
+ HDOP and VDOP performance of the HAPS-aided GPS system
1364
+ is more stable than the GPS-only system.
1365
+
1366
+ 2) 3D Positioning Accuracy Analysis
1367
+ The CDFs of the 3D positioning accuracy for both the
1368
+ suburban area and the dense urban area are shown in Fig. 12
1369
+ and Fig. 13, respectively. As we can see, without enabling the
1370
+ RAIM, the 90-percentile 3D positioning accuracy of the GPS-
1371
+ only system can be improved by 36 % in the suburban area, and
1372
+ 33.64 % in the dense urban area with the assistance from the
1373
+ HAPS. With the RAIM turned on, we can observe that the
1374
+ positioning performance of both the GPS-only system and the
1375
+ HAPS-aided GPS system can be further improved. Yet we
1376
+ notice that the improvement brought by the RAIM in the
1377
+ suburban area for the GPS-only system is almost negligible,
1378
+ while it is more tangible for the HAPS-aided GPS system. This
1379
+ is because the quality of signals in the suburban area tends to be
1380
+ great, and the HAPS-aided GPS system has a relatively higher
1381
+ chance to enable the RAIM as there are more ranging sources
1382
+ in the system. This observation is also applicable to the dense
1383
+ urban scenario where the 90-percentile 3D positioning accuracy
1384
+ of the HAPS-aided GPS system is improved by 45.2 %, which
1385
+ is much more significant than that for the GPS-only system. For
1386
+ the dense urban scenario where the multipath is severe, the
1387
+ RAIM algorithm plays a more significant role, especially in the
1388
+ HAPS-aided GPS system. The reason behind this is that the
1389
+ number of visible satellites in the dense urban area is low,
1390
+ making the RAIM algorithm for the GPS-only system less
1391
+ effective. Since the implemented RAIM algorithm detects and
1392
+ excludes an abnormal observation by multiplying its variance
1393
+ with an exponential term if the absolute value of its normalized
1394
+ pseudorange residual surpasses the critical value, we count the
1395
+ number of times where the absolute value of the normalized
1396
+ pseudorange residuals surpass the critical value for both
1397
+ systems considered and for both the suburban scenario and the
1398
+ dense urban scenario. For convenience, we rephrase the number
1399
+ of times where the absolute value of the normalized
1400
+ pseudorange residuals surpass the critical value as the number
1401
+ of times the RAIM is enabled. With the system model
1402
+ considered in this work, we find that the number of times the
1403
+ RAIM is enabled for the GPS-only system is roughly 23.84 %
1404
+ as many as the number of times the RAIM is enabled for the
1405
+ HAPS-aided GPS system in the suburban area; and the number
1406
+ of times the RAIM is enabled for the GPS-only system is about
1407
+ 50.72 % as many as the number of times the RAIM is enabled
1408
+ for the HAPS-aided GPS system in the dense urban area. This
1409
+ demonstrates the applicability of the RAIM algorithm on the
1410
+ HAPS-aided GPS system, especially in the dense urban area.
1411
+ V. CONCLUSION
1412
+ HAPS have a number of advantages over satellites, including
1413
+ (but not limited to) lower latency, lower pathloss, smaller
1414
+ pseudorange errors, and HAPS can provide continuous
1415
+ coverage to reduce the number of handovers for users. This
1416
+ makes HAPS an excellent candidate to serve as another type of
1417
+ ranging source. Since urban areas are where GNSS positioning
1418
+ performance degrades severely, while also being where most
1419
+ people live, deploying several HAPS as additional ranging
1420
+ sources above metropolitan cities would improve GNSS
1421
+ positioning performance and maximize the profit of the extra
1422
+ payloads on HAPS. From both the simulation and physical
1423
+ experiment results, we observed that HAPS can indeed improve
1424
+ the HDOP, the VDOP, and the 3D positioning accuracy of a
1425
+ legacy GNSS. With the system model considered in this work,
1426
+ we showed that the 90-percentile 3D positioning accuracy of
1427
+ the GPS-only system can be improved by around 35 % in both
1428
+ suburban and dense urban areas. We demonstrated the
1429
+ applicability of the RAIM algorithm for the HAPS-aided GPS
1430
+ system, especially in the dense urban areas. To enhance the
1431
+ simulation of the HAPS-aided GPS, the receiver clock offset
1432
+ should be estimated with higher accuracy. We think the
1433
+ effectiveness of the RAIM algorithm can be improved if the
1434
+ standard deviation of the target observable is available. To
1435
+ further improve the positioning performance for urban areas,
1436
+ we can make use of terrestrial signals such as cellular network
1437
+ signals and multipath signals. This would constitute a vertical
1438
+
1439
+ Suburbanarea
1440
+ 0.9
1441
+ 0.8
1442
+ 0.7
1443
+ 0.6
1444
+ DF
1445
+ 0.5
1446
+ 0.4
1447
+ 0.3
1448
+ 0.2
1449
+ GPS-only system (RAIM off)
1450
+ GPS-only system (RAIM on)
1451
+ 0.1
1452
+ HAPS-aided GPS system (RAIM off)
1453
+ HAPS-aidedGPSsystem (RAIMon)
1454
+ 0
1455
+ 0
1456
+ 5
1457
+ 10
1458
+ 15
1459
+ 20
1460
+ 25
1461
+ 3Dpositioningerror(m)Dense urban area
1462
+ 0.9
1463
+ 0.8
1464
+ 0.7
1465
+ 0.6
1466
+ DF
1467
+ 0.5
1468
+ 0.4
1469
+ 0.3
1470
+ 0.2
1471
+ GPS-only system (RAIM off)
1472
+ GPS-only system (RAIM on)
1473
+ 0.1
1474
+ HAPS-aided GPS system (RAIM off)
1475
+ HAPS-aided GPS system (RAIM on)
1476
+ 0
1477
+ 0
1478
+ 20
1479
+ 40
1480
+ 60
1481
+ 80
1482
+ 100
1483
+ 120
1484
+ 140
1485
+ 160
1486
+ 180
1487
+ 3D positioning error (m)> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1488
+
1489
+ heterogeneous network (V-Het-Net) positioning system, which
1490
+ we believe will yield a lower VDOP based on the DOP
1491
+ illustration presented in this paper.
1492
+ ACKNOWLEDGMENT
1493
+ The Skydel software is a formal donation from Orolia to
1494
+ Carleton University.
1495
+ REFERENCES
1496
+ [1] “Global Positioning System standard positioning service performance
1497
+ standard,” GPS.GOV, Washington, DC, USA, Apr. 2020. Accessed on:
1498
+ Dec.
1499
+ 31,
1500
+ 2022.
1501
+ [Online].
1502
+ Available:
1503
+ https://www.gps.gov/technical/ps/2020-SPS-performance-standard.pdf
1504
+ [2] “BeiDou Navigation Satellite System open service performance
1505
+ standard,” China Satellite Navigation Office, China, May 2021. Accessed
1506
+ on:
1507
+ Dec.
1508
+ 31,
1509
+ 2022.
1510
+ [Online].
1511
+ Available:
1512
+ http://en.beidou.gov.cn/SYSTEMS/Officialdocument/202110/P0202110
1513
+ 14595952404052.pdf
1514
+ [3] Korolev, “Open service performance standard (OS PS),” GLOBAL
1515
+ NAVIGATION SATELLITE SYSTEM GLONASS, Russia, June 2020.
1516
+ Accessed on: Dec. 31, 2022. [Online]. Available: https://www.glonass-
1517
+ iac.ru/upload/docs/stehos/stehos_en.pdf
1518
+ [4] “Open service - service definition document,” EUROPEAN GNSS
1519
+ (GALILEO), May 2019. Accessed on: Dec. 31, 2022. [Online]. Available:
1520
+ https://galileognss.eu/wp-content/uploads/2020/08/Galileo-OS-
1521
+ SDD_v1.1.pdf
1522
+ [5] G. K. Kurt and H. Yanikomeroglu, “Communication, computing, caching,
1523
+ and sensing for next generation aerial delivery networks: Using a high-
1524
+ altitude platform station as an enabling technology”, IEEE Veh. Technol.
1525
+ Mag., vol. 16, no. 3, pp. 108-117, Sep. 2021.
1526
+ [6] X. Li et al., “Precise positioning with current multi-constellation Global
1527
+ Navigation Satellite Systems: GPS, GLONASS, Galileo and BeiDou,”
1528
+ Sci. Rep. 5, no. 8328, Feb. 2015.
1529
+ [7] B. Li et al., “LEO enhanced Global Navigation Satellite System
1530
+ (LeGNSS) for real-time precise positioning services,” Advances in Space
1531
+ Research, vol. 63, no. 1, pp. 73-93, ISSN 0273-1177, 2019.
1532
+ [8] J. Khalife, M. Neinavaie, and Z. M. Kassas, “Navigation with differential
1533
+ carrier phase measurements from megaconstellation LEO satellites,”
1534
+ IEEE/ION Position, Locat. Navig. Symp. (PLANS), Portland, Oregon,
1535
+ 2020, pp. 1393-1404.
1536
+ [9] Q. Ren, O. Abbasi, G. K. Kurt, H. Yanikomeroglu, and J. Chen, “Caching
1537
+ and computation offloading in high altitude platform station (HAPS)
1538
+ assisted intelligent transportation systems,” IEEE Trans. Wirel. Commun.,
1539
+ vol. 21, no. 11, pp. 9010-9024, Nov. 2022.
1540
+ [10] C. Ding, J. Wang, H. Zhang, M. Lin, and G. Y. Li, “Joint optimization of
1541
+ transmission and computation resources for satellite and high altitude
1542
+ platform assisted edge computing,” IEEE Trans. Wirel. Commun., vol. 21,
1543
+ no. 2, pp. 1362-1377, Feb. 2022.
1544
+ [11] M. S. Alam, G. K. Kurt, H. Yanikomeroglu, P. Zhu, and N. D. Dao, “High
1545
+ altitude platform station based super macro base station constellations”,
1546
+ IEEE Commun. Mag., vol. 59, no. 1, pp. 103-109, Jan. 2021.
1547
+ [12] J. V. Sickle and J. A. Dutton, “The satellite clock,” The Pennsylvania
1548
+ State University. Accessed on: Dec. 31, 2022. [Online]. Available:
1549
+ https://www.e-education.psu.edu/geog862/node/1714.
1550
+ [13] ITU-R, “Preferred characteristics of systems in the fixed service using
1551
+ high altitude platforms operating in the bands 47.2-47.5 GHz and 47.9-
1552
+ 48.2 GHz.” ITU, Geneva, Recommendation F.1500, Jan. 2000.
1553
+ [14] F. Dovis, L. Lo Presti, and P. Mulassano, “Support infrastructures based
1554
+ on high altitude platforms for navigation satellite systems,” IEEE Wirel.
1555
+ Commun., vol. 12, no. 5, pp. 106-121, Oct. 2005.
1556
+ [15] O. Kim et al., “Navigation augmentation in urban area by HALE UAV
1557
+ with onboard Pseudolite during multi-purpose missions,” Int. J. Aeronaut.
1558
+ Space Sci. (IJASS), vol. 18, no. 3, pp. 545-554, Sep. 2017.
1559
+ [16] G. Boiero, F. Dovis, P. Mulassano, and M. Mondin, “Increasing the
1560
+ spatial limits of LADGPS using stratospheric platform,” J. Navig., vol.
1561
+ 54, no. 2, pp. 255-267, May 2001.
1562
+ [17] F. Dovis, P. Mulassano, and M. Dumville, “The stratolite concept: Design
1563
+ of a stratospheric pseudo-satellite for Galileo,” in Proc. ION GPS 2002,
1564
+ Portland, OR, USA, 2002.
1565
+ [18] T. Tsujii, M. Harigae, and K. Okano, “A new positioning/navigation
1566
+ system based on Pseudolites installed on high altitude platforms systems
1567
+ (HAPS),” in Proc. 24th Int. Congr. Aeronaut. Sci. (ICAS), Yokohama,
1568
+ Japan, 2004.
1569
+ [19] A. Angrisano and S. Gaglione, “Smartphone GNSS performance in an
1570
+ urban scenario with RAIM application,” Sens., vol. 22, no. 3, Jan. 2022.
1571
+ [20] Y. Yang and J. Xu, “GNSS receiver autonomous integrity monitoring
1572
+ (RAIM) algorithm based on robust estimation,” Geod. Geodyn., vol. 7,
1573
+ no. 2, pp. 117-123, Mar. 2016.
1574
+ [21] H. Zheng, M. Atia, and H. Yanikomeroglu, “High altitude platform
1575
+ station (HAPS)-aided GNSS for urban areas”, in Proc. 10th Annual IEEE
1576
+ Int. Conf. Wirel. Space Extreme Environ. (WISEE), Winnipeg, Manitoba,
1577
+ Canada, 2022.
1578
+ [22] S. Bijjahalli, R. Sabatini, and A. Gardi, “GNSS performance modelling
1579
+ and augmentation for urban air mobility,” Sens., vol. 19, no. 19, Sep.
1580
+ 2019.
1581
+ [23] J. Jang, D. Park, S. Sung, and Y. J. Lee, “HDOP and VDOP analysis in
1582
+ an ideal placement environment for dual GNSSs,” Sens., vol. 22, no. 9,
1583
+ May 2022.
1584
+ [24] N. Tongkasem, S. Sopan, J. Budtho, and N. Wongthodsarat, “Calculate
1585
+ user position by using the single point method,” CSSRG Laboratory, King
1586
+ Mongkut's Institute of Technology Ladkrabang Bangkok, Thailand, Feb.
1587
+ 2019.
1588
+ Accessed
1589
+ on:
1590
+ Dec.
1591
+ 31,
1592
+ 2022.
1593
+ [Online].
1594
+ Available:
1595
+ https://github.com/cssrg-kmitl/single-positioning-MATLAB
1596
+ [25] A. Tian, D. Dong, D. Ning, and C. Fu, “GPS single point positioning
1597
+ algorithm based on least squares,” in Proc. 6th Int. Symp. Comput. Intell.
1598
+ Des. (ISCID), Hangzhou, China, 2013, pp. 16-19.
1599
+ [26] A. Krishnamoorthy and D. Menon, “Matrix inversion using Cholesky
1600
+ decomposition,” in Proc. Signal Processing: Algorithms, Architectures,
1601
+ Arrangements, and Applications (SPA), Poznan, Poland, 2013, pp. 70-72.
1602
+ [27] “Skydel
1603
+ GNSS
1604
+ Simulation
1605
+ Software,”
1606
+ Orolia,
1607
+ Available:
1608
+ https://www.orolia.com/product/skydel-simulation-engine/.
1609
+ [28] B. Bidikar, G. S. Rao, and L. Ganesh, “Sagnac effect and SET error based
1610
+ pseudorange modeling for GPS applications,” Procedia Comput. Sci., vol.
1611
+ 87, pp. 172-177, 2016.
1612
+ [29] T. Soler and M. Chin, “On transformation of covariance matrices between
1613
+ local Cartesian coordinate systems and commutative diagrams,” in Proc.
1614
+ 45th Annual Meeting ASP-ACSM Convention, Jan. 1985.
1615
+ [30] H. Kuusniemi, A. Wieser, G. Lachapelle, and J. Takala, “User-level
1616
+ reliability monitoring in urban personal satellite-navigation,” IEEE Trans.
1617
+ Aerosp. Electron. Syst., vol. 43, no. 4, pp. 1305-1318, Oct. 2007.
1618
+ [31] A. G. Quinchia, G. Falco, E. Falletti, F. Dovis, and C. Ferrer, “A
1619
+ comparison between different error modeling of MEMS applied to
1620
+ GPS/INS integrated systems,” Sens., vol. 13, no. 8, pp. 9549-9588, Jul.
1621
+ 2013.
1622
+ [32] O. G. Crespillo, M. Joerger, and S. Langel, “Overbounding GNSS/INS
1623
+ integration with uncertain GNSS Gauss-Markov error parameters,”
1624
+ IEEE/ION Position, Locat. Navig. Symp. (PLANS), Portland, Oregon,
1625
+ 2020, pp. 481-489.
1626
+ [33] F. Hsieh and M. Rybakowski, “Propagation model for high altitude
1627
+ platform systems based on ray tracing simulation,” in Proc. 13th Eur.
1628
+ Conf. Antennas Propag. (EuCAP), Krakow, Poland, 2019, pp. 1-5.
1629
+ [34] S. Alfattani, W. Jaafar, Y. Hmamouche, H. Yanikomeroglu, and A.
1630
+ Yongacoglu, “Link budget analysis for reconfigurable smart surfaces in
1631
+ aerial platforms,” IEEE Open J. Commun. Soc., vol. 2, pp. 1980-1995,
1632
+ 2021.
1633
+ [35] “EVK-M8T user guide,” Ublox, May 2018. Accessed on: Dec. 31, 2022.
1634
+ [Online].
1635
+ Available:
1636
+ https://content.u-
1637
+ blox.com/sites/default/files/products/documents/EVK-
1638
+ M8T_UserGuide_%28UBX-14041540%29.pdf
1639
+
1640
+
1641
+
1642
+ > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1643
+
1644
+ Hongzhao
1645
+ Zheng
1646
+ (Member,
1647
+ IEEE)
1648
+ received the B. Eng. (Hons.) degree in
1649
+ engineering physics from the Carleton
1650
+ University, Ottawa, ON, Canada, in 2019.
1651
+ He is currently a PhD student at Carleton
1652
+ University. His research interest is the
1653
+ urban positioning using sensor-enabled
1654
+ heterogeneous wireless infrastructure.
1655
+
1656
+
1657
+ Mohamed Atia (Senior Member, IEEE)
1658
+ received the B.S. and M.Sc. degrees in
1659
+ computer systems from Ain Shams
1660
+ University, Cairo, Egypt, in 2000 and
1661
+ 2006, respectively, and the Ph.D. degree
1662
+ in electrical and computer engineering
1663
+ from Queen’s University, Kingston, ON,
1664
+ Canada, in 2013. He is currently an
1665
+ Associate Professor with the Department of Systems and
1666
+ Computer Engineering, Carleton University. He is also the
1667
+ Founder and the Director of the Embedded and Multi-Sensory
1668
+ Systems Laboratory (EMSLab), Carleton University. His
1669
+ research interests include sensor fusion, navigation systems,
1670
+ artificial intelligence, and robotics.
1671
+
1672
+
1673
+ Halim Yanikomeroglu (Fellow, IEEE)
1674
+ received the BSc degree in electrical and
1675
+ electronics engineering from the Middle
1676
+ East Technical University, Ankara, Turkey,
1677
+ in 1990, and the MASc degree in electrical
1678
+ engineering (now ECE) and the PhD degree
1679
+ in electrical and computer engineering from
1680
+ the University of Toronto, Canada, in 1992
1681
+ and 1998, respectively. Since 1998 he has
1682
+ been with the Department of Systems and Computer
1683
+ Engineering at Carleton University, Ottawa, Canada, where he
1684
+ is now a Full Professor. His research interests cover many
1685
+ aspects of wireless communications and networks. He has given
1686
+ 110+ invited seminars, keynotes, panel talks, and tutorials in the
1687
+ last five years. Dr. Yanikomeroglu’s collaborative research
1688
+ with industry resulted in 39 granted patents. Dr. Yanikomeroglu
1689
+ is a Fellow of the IEEE, the Engineering Institute of Canada
1690
+ (EIC), and the Canadian Academy of Engineering (CAE). He is
1691
+ a Distinguished Speaker for the IEEE Communications Society
1692
+ and the IEEE Vehicular Technology Society, and an Expert
1693
+ Panelist of the Council of Canadian Academies (CCA|CAC).
1694
+ Dr. Yanikomeroglu is currently serving as the Chair of the
1695
+ Steering Committee of IEEE’s flagship wireless event,
1696
+ Wireless Communications and Networking Conference
1697
+ (WCNC). He is also a member of the IEEE ComSoc GIMS,
1698
+ IEEE ComSoc Conference Council, and IEEE PIMRC Steering
1699
+ Committee. He served as the General Chair and Technical
1700
+ Program Chair of several IEEE conferences. He has also served
1701
+ in the editorial boards of various IEEE periodicals.
1702
+ Dr. Yanikomeroglu received several awards for his research,
1703
+ teaching, and service, including the IEEE ComSoc Fred W.
1704
+ Ellersick Prize (2021), IEEE VTS Stuart Meyer Memorial
1705
+ Award (2020), and IEEE ComSoc Wireless Communications
1706
+ TC Recognition Award (2018). He received best paper awards
1707
+ at IEEE Competition on Non-Terrestrial Networks for B5G and
1708
+ 6G in 2022 (grand prize), IEEE ICC 2021, IEEE WISEE 2021
1709
+ and 2022.
1710
+
1711
+
1712
+
1713
+
7dAyT4oBgHgl3EQf2vnp/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
8dE2T4oBgHgl3EQf8Ait/content/tmp_files/2301.04215v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
8dE2T4oBgHgl3EQf8Ait/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
9dAyT4oBgHgl3EQfdPeW/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:478414d9a786eb50068fda9202f31cebdc3c2dec2b2604c1c5e7852824a11d23
3
+ size 113297
AtAyT4oBgHgl3EQfq_mZ/content/2301.00553v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:29998b6ce30d2730259993cf4c1f519322d5428b976cf89727c9d50316c06cd3
3
+ size 16723618
AtAyT4oBgHgl3EQfq_mZ/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b9988de6357a03f1f8ac4e91aad9076500f57d3318eed180ef871ce42aeabad
3
+ size 3145773
AtE1T4oBgHgl3EQfDQOp/content/2301.02875v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3112f4b4cc72765815457da2ebfad320326ef86ab5648e78e140117935b76d0f
3
+ size 207923
AtE1T4oBgHgl3EQfDQOp/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c7e0bfc59dd6e76cf5f868165b777afe3ece0588d88825acfd46961014e37444
3
+ size 3211309
AtE1T4oBgHgl3EQfDQOp/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:25e463f0959926703042e6e2f11752bb1ae7cde5774386a90bb2d60f4eb44928
3
+ size 113222
BNE1T4oBgHgl3EQf9Qa6/content/2301.03555v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad31ef6a7151f4f6e02dacda9974861ffa487cb8661fa954da9975fe80389d11
3
+ size 1138489
BNE1T4oBgHgl3EQf9Qa6/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f39453e6d5b4243e4ace2613541d19913cee9237d70ee9f7369763b40e89ceb6
3
+ size 2818093
BNE1T4oBgHgl3EQf9Qa6/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4745475dd7ca3e729ae688e2b680c0631e2acca911cb1d0e695deb0f6fe2837e
3
+ size 109662
BNFKT4oBgHgl3EQfWi5i/content/tmp_files/2301.11792v1.pdf.txt ADDED
@@ -0,0 +1,1203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Graph Attention with Hierarchies for Multi-hop Question Answering
2
+ Yunjie He∗
3
+ University College London
4
+ yunjie.he.17@ucl.ac.uk
5
+ Ieva Stali¯unait˙e†
6
+ Accelex Technology
7
+ ieva.staliunaite@gmail.com
8
+ Philip John Gorinski
9
+ Huawei Noah’s Ark Lab, London
10
+ philip.john.gorinski@huawei.com
11
+ Pontus Stenetorp
12
+ University College London
13
+ pontus@stenetorp.se
14
+ Abstract
15
+ Multi-hop QA (Question Answering) is the
16
+ task of finding the answer to a question across
17
+ multiple documents. In recent years, a number
18
+ of Deep Learning-based approaches have been
19
+ proposed to tackle this complex task, as well
20
+ as a few standard benchmarks to assess mod-
21
+ els’ Multi-hop QA capabilities. In this paper,
22
+ we focus on the well-established HotpotQA
23
+ benchmark dataset, which requires models to
24
+ perform answer span extraction as well as sup-
25
+ port sentence prediction. We present two ex-
26
+ tensions to the state-of-the-art Graph Neural
27
+ Network (GNN) based model for HotpotQA,
28
+ Hierarchical Graph Network (HGN): (i) we
29
+ complete the original hierarchical structure by
30
+ introducing new edges between the query and
31
+ context sentence nodes; (ii) in the graph prop-
32
+ agation step, we propose a novel extension
33
+ to Hierarchical Graph Attention Network –
34
+ GATH (Graph ATtention with Hierarchies) –
35
+ that makes use of the graph hierarchy to up-
36
+ date the node representations in a sequential
37
+ fashion.
38
+ Experiments on HotpotQA demon-
39
+ strate the efficiency of the proposed modifica-
40
+ tions and support our assumptions about the
41
+ effects of model-related variables.
42
+ 1
43
+ Introduction
44
+ Question Answering (QA) tasks can be classified
45
+ into single-hop and multi-hop ones, depending on
46
+ the complexity of the underlying reasoning. Dif-
47
+ ferent from single-hop QA (Rajpurkar et al., 2016;
48
+ Trischler et al., 2017; Lai et al., 2017), where ques-
49
+ tions can be answered given a single paragraph or
50
+ single sentence in the context, multi-hop QA re-
51
+ quires us to retrieve and reason over scattered infor-
52
+ mation from multiple documents, as demonstrated
53
+ in Figure 1. There are many methods proposed for
54
+ addressing the multi-hop QA problem. One type of
55
+ ∗Work carried out as part of MSc thesis supervised by
56
+ Huawei Noah’s Ark Lab, London
57
+ †Work carried out while working at Huawei Noah’s Ark
58
+ Lab, London
59
+ Question:
60
+ Where did the form of music played by Die
61
+ Rhöner Säuwäntzt originate?
62
+ Answer:
63
+ United States
64
+ Supports:
65
+ Document 9
66
+ s1:
67
+ Die Rhöner Säuwäntzt are a Skiffle-
68
+ Bluesband from Eichenzell-Lütter in Hessen,
69
+ Germany.
70
+ Document 4
71
+ s1: Skiffle is a music genre with jazz, blues,
72
+ folk and American folk influences [...]
73
+ s2: Originating as a term in the United States
74
+ in the first half of the 20th century [...]
75
+ Figure 1: Example of a multi-hop answer and support
76
+ prediction, as found in HotpotQA.
77
+ these recent approaches extends well-performing
78
+ single-hop machine reading comprehension mod-
79
+ els to be multi-hop, such as DecompRC (Min et al.,
80
+ 2019) and QFE (Nishida et al., 2019).
81
+ The other avenue is to develop models specifi-
82
+ cally aimed at multi-hop QA. Among those, Graph
83
+ Neural Networks (GNNs) have recently garnered a
84
+ lot of attention. In GNN-based approaches, gaphs
85
+ are employed to represent query and context con-
86
+ tents (nodes) and the underlying relationships be-
87
+ tween them (edges). Information between nodes is
88
+ simultaneously propagated via the edges with the
89
+ help of a variety of GNNs, such as Graph Convolu-
90
+ tional Network (GCN) (Kipf and Welling, 2017),
91
+ Graph Attention Network (GAT) (Veliˇckovi´c et al.,
92
+ 2017), or Graph Recurrent Network (GRN) (Song
93
+ et al., 2018b). With these GNNs, node representa-
94
+ tions are obtained conditioned on the question and
95
+ context documents, and used for the QA task.
96
+ In this paper, we focus on one particular GNN ap-
97
+ proach designed for the Hotpot QA benchmark, the
98
+ Hierarchical Graph Network (HGN) introduced in
99
+ Fang et al. (2020). HGN constructs a hierarchical
100
+ graph that integrates nodes from different granu-
101
+ larity levels (question/paragraph/sentence/entity).
102
+ The edges in the graph capture the interactions be-
103
+ tween the information from heterogeneous levels
104
+ of the hierarchy. This hierarchical graph structure
105
+ arXiv:2301.11792v1 [cs.CL] 27 Jan 2023
106
+
107
+ has been shown to be crucial to the model’s remark-
108
+ able performance1 on both finding scattered pieces
109
+ of supporting information across documents and
110
+ the answer span prediction.
111
+ The contribution of this work is three-fold: (i)
112
+ we extend the edges of HGN with a new edge type
113
+ between the query and sentences, completing its
114
+ original structure; (ii) we introduce a novel exten-
115
+ sion of the Graph Attention Network – Graph At-
116
+ tention with Hierarchies (GATH). GATH allows for
117
+ making use of the explicit hierarchical graph struc-
118
+ ture, by propagating information through the graph
119
+ in a sequential fashion based on the hierarchy’s
120
+ levels, rather than updating all nodes simultane-
121
+ ously. (iii) We perform initial experiments on the
122
+ HotpotQA benchmark, providing evidence of the
123
+ effectiveness of our proposed extensions.
124
+ Code related to graph completion and GATH
125
+ will be made publicly available at redacted.
126
+ 2
127
+ Background
128
+ To solve the multi-hop QA problem, two general
129
+ research paths have been studied. The first direc-
130
+ tion focuses on extending the successful single-
131
+ hop machine reading comprehension method to the
132
+ multi-hop QA. DecompRC (Min et al., 2019) de-
133
+ composes the multi-hop reasoning problem into
134
+ multiple single-hop sub-questions based on span
135
+ predictions and applied traditional machine reading
136
+ comprehension techniques on these sub-questions
137
+ to obtain answers to the question. Query-Focused
138
+ Extractor (QFE) (Nishida et al., 2019) reformulates
139
+ the multi-hop QA task as a query-focused summa-
140
+ rization task based on the extractive summarization
141
+ model (Chen and Bansal, 2018).
142
+ The second research direction natively addresses
143
+ the task as a multi-hop setting, and directly tries
144
+ to gather the information from all context doc-
145
+ uments in order to answer the question. Many
146
+ approaches based on the transformer architecture
147
+ (Vaswani et al., 2017) address the multi-hop QA
148
+ task as simply one of attention between all words
149
+ in all available documents. In such approaches,
150
+ the problem quickly becomes intractable due to
151
+ the long inputs involved, and they thus typically
152
+ focus on alleviating the problems of using a full
153
+ attention mechanism. The Longformer (Beltagy
154
+ et al., 2020), for example, introduces a windowed
155
+ attention mechanism to localise the problem, allow-
156
+ 1At the time of writing, HGN achieves SOTA results on
157
+ HotpotQA, for GNN-based approaches.
158
+ ing for much longer input sequences to be handled
159
+ than with standard BERT-based language models
160
+ (Devlin et al., 2018).
161
+ However, recently more research effort has been
162
+ put toward approaches that employ Graph Neural
163
+ Networks, which allow for organising information
164
+ from various sources into a graph structure before
165
+ addressing the core task of Question Answering,
166
+ mitigating the need for very-long-distance attention
167
+ functions.
168
+ Coref-GRU (Dhingra et al., 2018) integrates mul-
169
+ tiple evidence associated with each entity mention
170
+ by incorporating co-reference information using
171
+ a collection of GRU layers of a gated-attention
172
+ reader (Dhingra et al., 2017). However, Coref-
173
+ GRU only leverages co-references local to a sen-
174
+ tence but ignores other useful global information.
175
+ To address this problem, MHQA-GRN and MHQA-
176
+ GCN (Song et al., 2018a) integrate evidence in a
177
+ more complex entity graph, with edges that also
178
+ connect global evidence. Similarly, De Cao et al.
179
+ (2019) also encode different relations between en-
180
+ tity mentions in the documents and perform the
181
+ graph reasoning via Graph Convolutional Network
182
+ (GCN) (Kipf and Welling, 2017).
183
+ All of the above methods which involve Graph
184
+ Neural Networks only consider entity nodes and the
185
+ relations between them. The HDE-Graph (Tu et al.,
186
+ 2019) extends these works by creating a new type
187
+ of graph with nodes corresponding to answer candi-
188
+ dates, documents and entities. Different edges are
189
+ included into the graph to capture the interaction
190
+ between these nodes. DFGN (Qiu et al., 2019) con-
191
+ structs a dynamic entity graph and performs graph
192
+ reasoning with a fusion block. This fusion block in-
193
+ cludes iterative interactions between the graph and
194
+ the documents (Doc2Graph and Graph2Doc flows)
195
+ in the graph construction process. Hierarchical
196
+ Graph Network (Fang et al., 2020) proposes a hier-
197
+ archical graph that incorporates nodes on different
198
+ levels of a hierarchy, including query, paragraph,
199
+ sentence, and entity nodes. This hierarchical graph
200
+ allows the model to aggregate query-related data
201
+ from many sources at various granularities.
202
+ One limitation that all of the above conventional
203
+ QA graph neural networks share is that their in-
204
+ formation propagation mechanisms do not directly
205
+ utilise the (explicit or implicit) hierarchical prop-
206
+ erty of the graph structure. In fields outside of
207
+ Natural Language Processing, recent studies on hi-
208
+ erarchical graph neural networks focus on passing
209
+
210
+ information on each hierarchical level to the node
211
+ at different attention weights.
212
+ In multi-agent reinforcement learning, HGAT
213
+ (Ryu et al., 2020) generates hierarchical state-
214
+ embedding of agents. This HGAT model stacks
215
+ inter-agent and inter-group graph attention net-
216
+ works hierarchically to capture inter-group node
217
+ interaction. A two-level graph attention mechanism
218
+ (Zhang et al., 2020) was developed for propagating
219
+ information in the close neighborhood of each node
220
+ in the constructed hierarchical graph. HATS (Kim
221
+ et al., 2019) predicts stock trends using relational
222
+ data on companies in the stock market. HATS
223
+ selectively aggregates information from different
224
+ relation types with a hierarchically designed atten-
225
+ tion mechanism. By maintaining only important
226
+ information at each level, HATS efficiently filters
227
+ out relations (edges) not useful for trend prediction.
228
+ However, all previous studies on hierarchical
229
+ graph neural networks only exploit the possible
230
+ hierarchical structure on the graph node itself. Dif-
231
+ ferent from the above methods, our proposed hi-
232
+ erarchical graph attention mechanism allows the
233
+ graph node embeddings to be updated in the or-
234
+ der of the hierarchical granularity level, instead of
235
+ simultaneously.
236
+ 3
237
+ Model
238
+ As our proposed improvements are largely aimed at
239
+ the established Hierarchical Graph Network (HGN)
240
+ model (Fang et al., 2020) for HotpotQA, we briefly
241
+ describe the original architecture. HGN builds a
242
+ hierarchical graph with nodes from several granu-
243
+ larity levels (question/paragraph/sentence/entity).
244
+ This hierarchical graph structure is good at captur-
245
+ ing the interaction between nodes from different
246
+ granularity levels and has been shown beneficial
247
+ to the model’s remarkable performance on both
248
+ finding scattered pieces of supporting information
249
+ across documents, and to answer span prediction.
250
+ The full HGN model pipeline consists of four
251
+ modules: (i) the Graph Construction Module se-
252
+ lects query-related paragraphs and builds a hier-
253
+ archical graph that contains edges between nodes
254
+ from different granularity levels within the para-
255
+ graphs; (ii) the Context Encoding Module gives
256
+ an initial representation/embeddings for nodes in
257
+ the graph via encoding layers that consist of a
258
+ RoBERTa (Liu et al., 2019) encoder and a bi-
259
+ attention layer; (iii) the Graph Reasoning Mod-
260
+ ule updates the initial representation of all nodes
261
+ via reasoning over the hierarchical graph; (iv) the
262
+ Multi-task Prediction Module performs multiple
263
+ sub-tasks including paragraph selection, support-
264
+ ing facts prediction, entity prediction and answer
265
+ span extraction, based on the representation of all
266
+ nodes. This process is summarized in Figure 2,
267
+ as presented by the original authors of the HGN
268
+ model.
269
+ We note that HGN still has limitations on its
270
+ graph structure and the graph reasoning step, and
271
+ in this work introduce according changes. Our
272
+ proposed extensions aim to further improve HGN
273
+ through a more complete graph structure, and a
274
+ novel hierarchical graph nodes update mechanism.
275
+ As such, our method mainly targets the Graph Con-
276
+ struction and Graph Reasoning Modules, described
277
+ in more detail below, while we leave the Context
278
+ Encoding and Multi-task Prediction Modules un-
279
+ changed.
280
+ Graph Construction Module
281
+ The Hierarchical Graph is built based on the given
282
+ HotpotQA question-context pair. This construc-
283
+ tion process consists of two steps: (i) multi-hop
284
+ reasoning paragraph retrieval from Wikipedia, i.e.
285
+ selecting candidate paragraphs with potential multi-
286
+ hop relationship to the question as paragraph nodes;
287
+ (ii) adding edges between question, sentence and
288
+ entity nodes within the retrieved paragraphs.
289
+ In particular, the first step consists of retriev-
290
+ ing “first-hop” paragraphs, that is, paragraphs of
291
+ Wikipedia entries that belong to entities mentioned
292
+ in the question. After this, a number of “second-
293
+ hop” paragraphs is selected, from Wikipedia arti-
294
+ cles that are hyper-linked from these first hops.
295
+ Our work keeps the original paragraph selection
296
+ method, but introduces novel meaningful edges
297
+ between graph nodes.
298
+ Context Encoding Module
299
+ With the hierarchical graph structure in place, rep-
300
+ resentations of the nodes within the graph are ob-
301
+ tained via the Context Encoding Module. In this
302
+ encoder, query and context are concatenated and
303
+ fed into a pretrained RoBERTa (Liu et al., 2019).
304
+ The obtained representations are further passed into
305
+ a bi-attention layer (Seo et al., 2018) to enhance
306
+ the cross interactions between the question and the
307
+ context. Through this encoding mechanism, the
308
+ question node is finally represented as q ∈ Rd and
309
+ the i-th paragraph/sentence/entity nodes are repre-
310
+ sented by pi, si and ei ∈ Rd respectively.
311
+
312
+ Figure 2: Model architecture of Hierarchical Graph Network (HGN). This illustration was originally introduced in
313
+ Fang et al. (2020). We include it here for completion, to provide an overview of HGN.
314
+ Graph Reasoning Module
315
+ Intuitively, the initial representations of the graph
316
+ nodes only carry the contextualized information
317
+ contained within their local contexts. To benefit
318
+ from the hierarchy and information across differ-
319
+ ent contexts, the Graph Reasoning Module further
320
+ propagates information between the graph nodes
321
+ using a single-layered Multi-head Graph Attention
322
+ Network (GAT) (Veliˇckovi´c et al., 2017). How-
323
+ ever, we believe the simultaneous node-update per-
324
+ formed by standard GAT can be improved, in the
325
+ presence of the explicitly given hierarchical prop-
326
+ erty of the graph. We therefore propose a novel
327
+ hierarchical graph reasoning method that performs
328
+ node updates sequentially, for different levels of
329
+ the hierarchy. In this manner, nodes on certain
330
+ granularity levels of the graph are allowed to first
331
+ aggregate some information, before passing it on
332
+ to their neighbours on other levels. We speculate
333
+ that this staggered information passing paradigm
334
+ can be beneficial to the multi-hop Question An-
335
+ swering task, by passing on more question-specific
336
+ contextualized information to relevant nodes.
337
+ Multi-task Prediction Module
338
+ The final step of the HGN model is to jointly pre-
339
+ dict answer and supporting facts for the question
340
+ via multi-task learning based on the updated graph
341
+ node representations. This is decomposed into five
342
+ sub-tasks: (i) paragraph selection determines if a
343
+ paragraph contains the ground truth; (ii) sentence
344
+ selection determines if a sentence from the selected
345
+ paragraph is a supporting fact; (iii) answer span
346
+ prediction finds the start and end indices of the
347
+ ground-truth span; (iv) answer type prediction pre-
348
+ dicts the type of the question; (v) entity prediction
349
+ determines if the answer can be found among the
350
+ selected entities. The above sub-tasks are jointly
351
+ trained through multi-task learning with the final
352
+ objective of the total loss from these sub-tasks:
353
+ Ljoint =Lstart + Lend + λ1Lpara+
354
+ λ2Lsent + λ3Lentity + λ4Ltype
355
+ (1)
356
+ With HGN re-introduced for completeness, we
357
+ describe our proposed extensions to the original
358
+ architecture in the subsequent sections.
359
+ 3.1
360
+ Completion of the graph structure
361
+ HGN constructs a hierarchical graph connecting
362
+ the query node with the selected multi-hop para-
363
+ graphs. Each selected paragraph contains sentences
364
+ and entities which are also encoded as nodes in the
365
+ hierarchical graph. The graph not only incorpo-
366
+ rates the natural hierarchy existing in paragraphs,
367
+ sentences and entities, but also includes helpful
368
+ connections between them to faciliate the structual
369
+ information propagation within the graph. Specif-
370
+ ically, the graph consists of seven types of edges,
371
+ which link the nodes in the graph. These edges
372
+ are (i) edges between the question node and first-
373
+ hop paragraph nodes; (ii) edges between paragraph
374
+ nodes; (iii) edges between sentences in the same
375
+ paragraph; (iv) edges between paragraph nodes
376
+ and the corresponding within-paragraph sentence
377
+ nodes; (v) edges between second-hop paragraphs
378
+ and the hyperlinked sentences; (vi) edges between
379
+
380
+ Multi-task Prediction Module
381
+ Graph Construction Module
382
+ Paragraph
383
+ Supporting Facts
384
+ Entity
385
+ Answer Span
386
+ Q
387
+ Selection
388
+ Prediction
389
+ Prediction
390
+ Extraction
391
+
392
+
393
+
394
+
395
+ Paragraph
396
+ (P1
397
+ P2
398
+ Updated:
399
+ Gated Attention
400
+ Level
401
+ hyperlink
402
+
403
+ Sentence
404
+ S1
405
+ S2
406
+ S3
407
+ S4
408
+ S5
409
+ Graph Reasoning Module
410
+ Level
411
+ Initial Representations:
412
+ Entity
413
+ E2
414
+ E3
415
+ E4
416
+ E1
417
+ Level
418
+ Adriana
419
+
420
+ New York
421
+ Virginia
422
+ Greenwich
423
+ Trigiani
424
+ Village
425
+ City
426
+ Context Encoding Module
427
+
428
+ Q
429
+ P1
430
+ P2Figure 3: Hierarchical Graph with (orange-colored)
431
+ new question_sentence edges added.
432
+ the question node and its matching entity nodes;
433
+ (vii) edges between sentence nodes and their corre-
434
+ sponding within-sentence entity nodes.
435
+ We note that the only type of edge that seems to
436
+ be missing from the graph are question-sentence
437
+ edges.
438
+ Hence, we first complete the hierarchi-
439
+ cal graph by introducing novel question_sentence
440
+ edges which connect the question node with all
441
+ sentence nodes of selected paragraphs. Such new
442
+ connections are introduced as edge (viii) in the hier-
443
+ archical graph. The constructed hierarchical graph
444
+ with novel edges added is illustrated in Figure 3.
445
+ We reason that this more complete graph might
446
+ help the model to learn more useful embedding
447
+ because of the modification in the graph topology,
448
+ which facilitates the information transmission be-
449
+ tween the question and sentences.
450
+ 3.2
451
+ Graph Attention with Hierarchies
452
+ The Graph Reasoning Module updates the contex-
453
+ tualized representations of graph nodes to capture
454
+ the information aggregated from topological neigh-
455
+ bours such that the local structures of these nodes
456
+ can be included. In HGN, this process is realized
457
+ by the Graph Attention Network (GAT) (Veliˇckovi´c
458
+ et al., 2017), a well-established GNN approach.
459
+ However, we note that in the specific setting
460
+ of Multi-hop QA with the presence of an explicit
461
+ hierarchical graph structure, GAT might not be
462
+ able to make full use of the information encoded
463
+ in the graph, as it will not directly capture the
464
+ crucial dependencies between “levels” of the hi-
465
+ erarchical. To address this problem, we propose a
466
+ novel Graph Attention Network with Hierarchies
467
+ (GATH) which updates nodes sequentially condi-
468
+ tioned on an imposed order over the hierarchy lev-
469
+ els. This is expected to help the model more effec-
470
+ tively processes the local observation of each node
471
+ into an information-condensed and contextualized
472
+ state representation for individual nodes on specific
473
+ levels, e.g. for paragraphs, before passing their in-
474
+ formation on to their neighbours on other levels,
475
+ such as to entity nodes. We expect this staggered
476
+ flow of information might help the model aggregate
477
+ information that is more useful and conditioned on
478
+ the task at hand.
479
+ The nodes in the graph are split into four cate-
480
+ gories, and can be represented by q, P, S and E:
481
+ P = {pi}np
482
+ i=1
483
+ S = {si}ns
484
+ i=1
485
+ E = {ei}ne
486
+ i=1
487
+ with each node embedded with an embedding func-
488
+ tion as described above, into a d-dimensional vec-
489
+ tor. These node representations are jointly repre-
490
+ sent the graph nodes as
491
+ H = {q, P, S, E} ∈ Rg×d, g = 1 + np + ns + ne
492
+ GATH updates all initial node embedding H to
493
+ H
494
+ ′ through hierarchical graph updates. Different
495
+ from GAT, GATH updates the nodes representation
496
+ sequentially, according to a pre-determined order
497
+ of hierarchical levels, instead of simultaneously. It
498
+ takes the initial node representations H as input,
499
+ but first only updates information of node features
500
+ of the first hierarchical level while keeping other
501
+ node embeddings unchanged.
502
+ For example, if the first level to be updated is
503
+ the paragraph level, we obtain the updated graph
504
+ representation
505
+ Hpara = {h1, h
506
+
507
+ 2, h
508
+
509
+ 3, . . . h
510
+
511
+ 1+np, h2+np, . . . , hg}
512
+ Specifically,
513
+ h
514
+
515
+ i = ∥K
516
+ k=1LeakyRelu(
517
+
518
+ j∈Ni
519
+ αk
520
+ ijhjWk)
521
+ (2)
522
+ where ∥K
523
+ k=1 represents concatenation of K heads,
524
+ Wk is the weight matrix to be learned, Ni repre-
525
+ sents the set of neighbouring nodes of node i and
526
+ αk
527
+ ij is the attention coefficient calculated by:
528
+ αk
529
+ ij =
530
+ exp(LeakyRelu([hi; hj]wk
531
+ eij))
532
+
533
+ t∈Ni exp(LeakyRelu([hi; ht]wkeit)) (3)
534
+ where [hi; hj] denotes the concatenation of hi and
535
+ hj, and wk
536
+ eij is the weight vector corresponding to
537
+ the edge between node i and j.
538
+ Based on the updated embeddings on the para-
539
+ graph level Hpara, we might next consider updating
540
+
541
+ Q
542
+ edge (vi)
543
+ edge (i)
544
+ Paragraph
545
+ edge (ii)
546
+ Level
547
+ P1
548
+ P2
549
+ édge (vili)
550
+ edge (v)
551
+ edge (iv)
552
+ Sentence
553
+ Level
554
+ S1
555
+ S2
556
+ S3
557
+ S4
558
+ S5
559
+ edge (ii)
560
+ edge (vii)
561
+ Entity Level
562
+ E1
563
+ E2
564
+ E3
565
+ E4the information on the sentence level. GATH propa-
566
+ gates information to all nodes on the sentence level
567
+ based on Hpara. This will output a further updated
568
+ graph representation
569
+ Hsent = {h1, h
570
+
571
+ 2, . . . h
572
+
573
+ 1+np+ns, h2+np+ns, . . . , hg}
574
+ with all nodes in P and S updated.
575
+ Continuing the process in this manner, we even-
576
+ tually will have updated all node representations
577
+ to obtain H
578
+ ′ {h
579
+
580
+ 1, h
581
+
582
+ 2, ..., h
583
+
584
+ g}. Algorithm 1 summa-
585
+ rizes the above procedures in pseudo code. Ad-
586
+ ditionally, these updating steps are combined and
587
+ illustrated in Figure 4.
588
+ 4
589
+ Experiments
590
+ In this section, we present experiments comparing
591
+ our extended HGN models with GATH with the
592
+ original one employing GAT, and provide a detailed
593
+ analysis of the proposed improvements and results.
594
+ For all experiments, we use RoBERTalarge as
595
+ the base embedding model. We train with a batch
596
+ size of 16 and a learning rate of 1e−5 over 5 epochs,
597
+ with λ1, λ3, λ4 = 1 and λ2 = 2, and we employ a
598
+ dropout rate of 0.2 on the transformer outputs, and
599
+ 0.3 throughout the rest of the model.
600
+ 4.1
601
+ Dataset
602
+ The effects of the above proposed improvements
603
+ are assessed based on HotpotQA (Yang et al., 2018).
604
+ It is a dataset with 113k English Wikipedia-based
605
+ question-answer pairs with two main features: (i) It
606
+ requires reasoning over multiple documents with-
607
+ out constraining itself to an existing knowledge
608
+ base or knowledge schema; (ii) Sentence-level sup-
609
+ porting facts are given for the answer to each ques-
610
+ tion, which explain the information sources that the
611
+ answer comes from. The performance of models
612
+ on HotpotQA is mainly assessed on two metrics,
613
+ exact match (EM) and F1 score. The model is ex-
614
+ pected to not only provide an accurate answer to the
615
+ question, but also to give supporting evidences for
616
+ its solution. Thus, EM and F1 score are calculated
617
+ for both answer spans and supporting facts.
618
+ HotpotQA has two settings: Distractor and Full-
619
+ wiki. In the distractor setting, context paragraphs
620
+ consist of 2 gold truth paragraphs containing in-
621
+ formation that is needed to solve the question, and
622
+ 8 paragraphs retrieved from Wikipedia based on
623
+ the question, serving as related yet uninformative
624
+ distractors for the question-answer pair. In the
625
+ Fullwiki setting, all context paragraphs come from
626
+ Wikipedia’s top search results, and they need to be
627
+ pre-ranked and selected in a first step. Compared
628
+ with the distractor setting, this setting requires us
629
+ to propose an additional paragraph selection model
630
+ concerned with information retrieval, before we
631
+ address multi-hop reasoning task. As all our pro-
632
+ posed extensions aim at the graph construction and
633
+ reasoning steps, we only perform these initial ex-
634
+ periments to assess the impact of our approach in
635
+ the distractor setting, where we are independent
636
+ from the influence of such a retrieval system.
637
+ 4.2
638
+ Experimental Results
639
+ Using the HotpotQA dataset, the models with our
640
+ extensions of graph completion and GATH are com-
641
+ pared against the baseline model of HGN with stan-
642
+ dard GAT. Since it could reasonably be argued
643
+ that GATH “simulates” a (partially) multi-layered
644
+ GAT in the sense that some nodes are updated only
645
+ after others have already been able to incorporate
646
+ neighbouring information – which in standard GAT
647
+ requires at least two full layers – we also include
648
+ an HGN trained with a two-layer network rather
649
+ than the single layer used in the original paper.
650
+ Table 1 summarizes the results on the dev set of
651
+ HotpotQA2.
652
+ 2Authors’ note: unfortunately, despite our best efforts we
653
+ were not able to reproduce the numbers reported for HGN
654
+ in Fang et al. (2020), even with their original, open-sourced
655
+ code. We tried both the hyper-parameters as published in the
656
+ paper, and the ones shipped with the code release however, the
657
+ RoBERTalarge performance when training from scratch was
658
+ consistently much lower than expected on dev (∼ 74 vs ∼ 70
659
+ joint F1). We contacted the original authors, who were not
660
+ able to help out with this. In light of these discrepancies, we
661
+
662
+ Algorithm 1
663
+ Graph Attention Network with Hierarchies
664
+ (GATH)
665
+ Input: H = hi, h2, .., hg}
666
+ Output: H' = {hi, h2, ..,h.]
667
+ for t in 1 : total number of levels do
668
+ create Ht = [ ]
669
+ foriin 1 : g do
670
+ if node i belongs to level t then
671
+ ht ← IIK=LeakyRelu(ZjeN; Qtit)
672
+ k;(t)ht-1wk,(t)
673
+ k,(t)
674
+ EsE N, exp( LeakyRelu([ht-1;h-1]we;(t))
675
+ else
676
+ h ←ht-1
677
+ end if
678
+ Append h, to Ht
679
+ end for
680
+ end for
681
+ return H' - Htotal number of levelsFigure 4: Hierarchical node representation update process. The grey-colored graph nodes are initial contextualized
682
+ embedding given by the Context Encoding Layer. Through the paragraph level message passing layer, only the
683
+ neighboring information of all paragraph nodes can be passed and renewed on them. Similar steps repeat for
684
+ sentence level and entity level. For convenience of labeling indices, we set np = 2, ns = 5, ne = 4
685
+ Completion of graph structure
686
+ The HGN with
687
+ new query-sentence edges improves over the base-
688
+ line by 0.7/0.4 on Joint EM and F1 scores. This
689
+ supports our intuition that the the missing question-
690
+ to-sentence edges can indeed bring advantages to
691
+ the model’s abilities of both answer span extraction
692
+ and supporting facts prediction.
693
+ Graph Attention with Hierarchies
694
+ GATH al-
695
+ lows for pre-defining the order of level updates in
696
+ the model. Given that the order in which the hierar-
697
+ chy levels are updated is likely to affect the model’s
698
+ performance, we perform experiments with dif-
699
+ ferent orders (P/S/E3,4, E/S/P, S/E/P and S/P/E)
700
+ and compare them to the baseline models with
701
+ one and two-layer GAT. All the GATH-based ex-
702
+ tended models outperform the baseline model on
703
+ the answer-span extraction by an absolute gain of
704
+ 1.6 to 2.4 points on the answer extraction metrics.
705
+ On the other hand, the order of hierarchical lev-
706
+ els does show an influence on the model’s evi-
707
+ decided to focus only on dev set performance when assessing
708
+ the impact of our extensions against re-trained vanilla HGN,
709
+ as a fair comparison to the original model on test was not
710
+ possible at this time.
711
+ 3P/S/E abbreviates Paragraph/Sentence/Entity
712
+ 4We exclude the query level update to make it more com-
713
+ parable to the baseline model, which also excludes this update.
714
+ dence collection ability. The “wrong” order leads
715
+ to worse performance of the extended model, such
716
+ as in the E/S/P and S/P/E cases.
717
+ On most metrics, but specifically on joint F1
718
+ score, the extended GATH-based model with the or-
719
+ der S/E/P outperforms not only the baseline model,
720
+ but also the other GATH models. It achieves a Joint
721
+ EM/F1 score of 43.9/71.5, exceeding the baseline
722
+ model’s performance by 1.2 each.
723
+ Interestingly, the 2-layer GAT version of HGN
724
+ slightly under-performs when compared to the orig-
725
+ inal HGN setup. While gaining 0.3 points in sup-
726
+ port prediction F1, it loses the same amount of
727
+ performance in answer prediction and joint scores.
728
+ We assume this is why the original HGN calls for
729
+ only one layer, when we could intuitively have ex-
730
+ pected multi-layered networks to perform better.
731
+ Combined query-sentence edges and GATH
732
+ The above experimental results demonstrate the
733
+ individual effectiveness of these two proposed im-
734
+ provements of graph completion and GATH. Nat-
735
+ urally, we are also interested in the performance
736
+ resulting from combining both. The “HGN (Com-
737
+ bined)” row in Table 1 represents the model com-
738
+ bining graph completion and GATH-S/E/P. This
739
+ combined model brings slight improvement over
740
+
741
+ H
742
+ Hsent
743
+ Hpara
744
+ h1
745
+ h1
746
+ edge (i)
747
+ Paragraph
748
+ edge (i)
749
+ Paragraph
750
+ Level
751
+ P1
752
+ P2
753
+ Level
754
+ P1
755
+ h2
756
+ h2
757
+ h2
758
+ edge (iv)
759
+ edge (v)
760
+ edge (iv) -
761
+ dge (v)
762
+ Sentence
763
+ Sentence
764
+ h's
765
+ Level
766
+ S2
767
+ 63
768
+ h?
769
+ h3
770
+ Level
771
+ S1
772
+ edge (ii)
773
+ edge (vii)i
774
+ h'4
775
+ Entity Level
776
+ h4
777
+ h4
778
+ E1
779
+ EntityLevel
780
+ hs
781
+ Paragraph level nodes updating
782
+ ..
783
+ Sentence level nodes updating
784
+ hg
785
+ hg
786
+ hg
787
+ H'
788
+ Hent
789
+ edge(vi)
790
+ hi
791
+ edge (vi),
792
+ h1
793
+ Paragraph
794
+ Paragraph
795
+ Level
796
+ P1
797
+ P2
798
+ Level
799
+ P1
800
+ h2
801
+ h2
802
+ Sentence
803
+ hs
804
+ Level
805
+ S2
806
+ S3
807
+ $4
808
+ S5
809
+ Sentence
810
+ S1
811
+ hs
812
+ Level
813
+ S3
814
+ SA
815
+ edge (vii) :
816
+ h4
817
+ h4
818
+ edge (vi) i
819
+ Entity Level
820
+ E2
821
+ E3
822
+ E4
823
+ EntityLevel
824
+ hs
825
+ E3
826
+ hs
827
+ 1
828
+ ..
829
+ Question level nodes updating
830
+ hg
831
+ hg
832
+ Entity level nodes updatingAnswer
833
+ Support
834
+ Joint
835
+ Model
836
+ EM
837
+ F1
838
+ P
839
+ R
840
+ EM
841
+ F1
842
+ P
843
+ R
844
+ EM
845
+ F1
846
+ P
847
+ R
848
+ HGN (baseline)
849
+ 64.5
850
+ 78.3
851
+ 81.6
852
+ 79.0
853
+ 60.4
854
+ 87.4
855
+ 89.6
856
+ 87.5
857
+ 42.7
858
+ 70.3
859
+ 75.0
860
+ 70.9
861
+ HGN (2-layer GAT)
862
+ 64.1
863
+ 78.0
864
+ 81.4
865
+ 78.9
866
+ 59.9
867
+ 87.7 89.5
868
+ 88.4
869
+ 41.6
870
+ 70.0
871
+ 74.5
872
+ 71.3
873
+ HGN (que_sent edge)
874
+ 65.0
875
+ 79.1
876
+ 82.2
877
+ 80.1
878
+ 60.9
879
+ 87.0
880
+ 89.9
881
+ 86.9
882
+ 43.4
883
+ 70.7
884
+ 75.7
885
+ 71.5
886
+ HGN with GATH(P/S/E)
887
+ 66.1
888
+ 80.1
889
+ 83.1
890
+ 81.1
891
+ 54.2
892
+ 80.1
893
+ 86.1
894
+ 79.5
895
+ 38.7
896
+ 66.5
897
+ 73.6
898
+ 66.9
899
+ HGN-GATH(E/S/P)
900
+ 66.4
901
+ 80.3
902
+ 83.6
903
+ 81.2
904
+ 38.3
905
+ 74.4
906
+ 70.0 90.0
907
+ 27.2
908
+ 61.2
909
+ 59.8
910
+ 74.1
911
+ HGN-GATH(S/E/P)
912
+ 67.0
913
+ 80.6 83.7 81.4
914
+ 60.5
915
+ 86.3 92.3 83.7
916
+ 43.9
917
+ 71.5 78.8 70.2
918
+ HGN-GATH(S/P/E)
919
+ 66.8
920
+ 80.7 83.8
921
+ 81.6
922
+ 38.7
923
+ 74.6
924
+ 70.3 90.0
925
+ 27.3
926
+ 61.5
927
+ 60.1 74.5
928
+ HGN (Combined)
929
+ 66.7
930
+ 80.7 83.7 81.7 61.4
931
+ 87.0
932
+ 91.2
933
+ 85.6
934
+ 43.9
935
+ 71.9 77.8
936
+ 71.8
937
+ Table 1: Performance of the proposed HGN with completed edges (HGN que_sent), GATH, and both extensions
938
+ combined on the development set of HotpotQA in distractor setting, against the baseline model HGN with GAT.
939
+ the other models on most metrics. This final model
940
+ sees further improvements, particularly in the an-
941
+ swer span prediction task, and achieves the overall
942
+ highest joint F1 score at 71.9, indicating that the
943
+ contributions of graph completion and GATH are
944
+ mutually benefitial.
945
+ 4.3
946
+ Error Analysis
947
+ In this section, we perform an error analysis on
948
+ the concrete influence of the proposed HGN (com-
949
+ bined) model based on question types. The major-
950
+ ity of questions in HotpotQA fall under the bridge5
951
+ and comparison reasoning categories.
952
+ As sug-
953
+ gested by Fang et al. (2020), we split comparison
954
+ questions into comp-yn and comp-span. The former
955
+ represents questions that should answer the compar-
956
+ ison between two entities with “yes” or “no”, e.g.
957
+ “Is Obama younger than Trump?”, while the latter
958
+ requires an answer span, e.g. “Who is younger,
959
+ Obama or Trump?”.
960
+ Table 2 shows the performance of the original
961
+ HGN model and the proposed model HGN-GATH
962
+ (combined) on various types of reasoning questions.
963
+ Results indicate that comp-yn questions are easiest
964
+ for both models, and the bridge type is the hardest
965
+ to solve. The analysis table shows that HGN (com-
966
+ bined) is more effective than the original model
967
+ on all of these reasoning kinds except support EM
968
+ for comp-yn, though even here the much improved
969
+ answer prediction leads to an overall improvement
970
+ of 2.42 on Joint EM.
971
+ 5requiring a bridging entity between support sentences,
972
+ needed to arrive at the answer
973
+ HGN-GAT
974
+ Question
975
+ Ans EM
976
+ Sup EM
977
+ Joint EM
978
+ Pct(%)
979
+ comp-yn
980
+ 81.22
981
+ 81.44
982
+ 68.34
983
+ 6.19
984
+ comp-span
985
+ 65.50
986
+ 71.04
987
+ 48.49
988
+ 13.90
989
+ bridge
990
+ 63.08
991
+ 57.01
992
+ 39.73
993
+ 79.91
994
+ HGN-GATH (Combined)
995
+ Question
996
+ Ans EM
997
+ Sup EM
998
+ Joint EM
999
+ Pct(%)
1000
+ comp-yn
1001
+ 85.81
1002
+ 80.79
1003
+ 70.96
1004
+ 6.19
1005
+ comp-span
1006
+ 68.42
1007
+ 71.53
1008
+ 50.34
1009
+ 13.90
1010
+ bridge
1011
+ 64.95
1012
+ 58.08
1013
+ 40.71
1014
+ 79.91
1015
+ Table 2: Original HGN (top) and HGN-GATH com-
1016
+ bined (bottom) model results for various reasoning
1017
+ types. ‘Pct’ signifies percentage of all questions per
1018
+ category.
1019
+ 5
1020
+ Conclusions and Future Work
1021
+ In this paper, we proposed two extensions to Hi-
1022
+ erarchical Graph Network (HGN) for the multi-
1023
+ hop Question Answering task on HotpotQA. First,
1024
+ we completed the hierarchical graph structure by
1025
+ adding new edges between the query and context
1026
+ sentence nodes. Second, we introduced GATH as
1027
+ the mechanism for neural node updates, a novel
1028
+ extension to GAT that can update node representa-
1029
+ tions sequentially, based on hierarchical levels. To
1030
+ the best of our knowledge, this is the first time the
1031
+ hierarchical graph structure is directly exploited in
1032
+ the update mechanism for information propagation.
1033
+ Experimental results indicate the validity of our
1034
+ approaches individually, as well as when used
1035
+ jointly for the multi-hop QA problem, outperform-
1036
+ ing the currently best performing graph neural net-
1037
+ work based model, HGN, on HotpotQA.
1038
+ In the future, we would particularly like to in-
1039
+ tegrate hierarchical graph attention weights into
1040
+
1041
+ GATH, as motivated by related research in Rein-
1042
+ forcement Learning.
1043
+ References
1044
+ Iz Beltagy, Matthew E Peters, and Arman Cohan.
1045
+ 2020. Longformer: The long-document transformer.
1046
+ arXiv preprint arXiv:2004.05150.
1047
+ Yen-Chun Chen and Mohit Bansal. 2018. Fast abstrac-
1048
+ tive summarization with reinforce-selected sentence
1049
+ rewriting. In Proceedings of the 56th Annual Meet-
1050
+ ing of the Association for Computational Linguis-
1051
+ tics (Volume 1: Long Papers), pages 675–686, Mel-
1052
+ bourne, Australia. Association for Computational
1053
+ Linguistics.
1054
+ Nicola De Cao, Wilker Aziz, and Ivan Titov. 2019.
1055
+ Question answering by reasoning across documents
1056
+ with graph convolutional networks.
1057
+ In Proceed-
1058
+ ings of the 2019 Conference of the North American
1059
+ Chapter of the Association for Computational Lin-
1060
+ guistics: Human Language Technologies, Volume 1
1061
+ (Long and Short Papers), pages 2306–2317, Min-
1062
+ neapolis, Minnesota. Association for Computational
1063
+ Linguistics.
1064
+ Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
1065
+ Kristina Toutanova. 2018. Bert: Pre-training of deep
1066
+ bidirectional transformers for language understand-
1067
+ ing. arXiv preprint arXiv:1810.04805.
1068
+ Bhuwan Dhingra, Qiao Jin, Zhilin Yang, William Co-
1069
+ hen, and Ruslan Salakhutdinov. 2018. Neural mod-
1070
+ els for reasoning over multiple mentions using coref-
1071
+ erence. In Proceedings of the 2018 Conference of
1072
+ the North American Chapter of the Association for
1073
+ Computational Linguistics: Human Language Tech-
1074
+ nologies, Volume 2 (Short Papers), pages 42–48,
1075
+ New Orleans, Louisiana. Association for Computa-
1076
+ tional Linguistics.
1077
+ Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William
1078
+ Cohen, and Ruslan Salakhutdinov. 2017.
1079
+ Gated-
1080
+ attention readers for text comprehension.
1081
+ In Pro-
1082
+ ceedings of the 55th Annual Meeting of the Associa-
1083
+ tion for Computational Linguistics (Volume 1: Long
1084
+ Papers), pages 1832–1846, Vancouver, Canada. As-
1085
+ sociation for Computational Linguistics.
1086
+ Yuwei Fang, Siqi Sun, Zhe Gan, Rohit Pillai, Shuo-
1087
+ hang Wang, and Jingjing Liu. 2020.
1088
+ Hierarchical
1089
+ graph network for multi-hop question answering. In
1090
+ Proceedings of the 2020 Conference on Empirical
1091
+ Methods in Natural Language Processing (EMNLP),
1092
+ pages 8823–8838, Online. Association for Computa-
1093
+ tional Linguistics.
1094
+ Raehyun Kim, Chan Ho So, Minbyul Jeong, Sanghoon
1095
+ Lee, Jinkyu Kim, and Jaewoo Kang. 2019. Hats: A
1096
+ hierarchical graph attention network for stock move-
1097
+ ment prediction. arXiv preprint arXiv:1908.07999.
1098
+ Thomas N. Kipf and Max Welling. 2017.
1099
+ Semi-
1100
+ supervised classification with graph convolutional
1101
+ networks.
1102
+ Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang,
1103
+ and Eduard Hovy. 2017. Race: Large-scale reading
1104
+ comprehension dataset from examinations.
1105
+ arXiv
1106
+ preprint arXiv:1704.04683.
1107
+ Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-
1108
+ dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
1109
+ Luke Zettlemoyer, and Veselin Stoyanov. 2019.
1110
+ Roberta: A robustly optimized bert pretraining ap-
1111
+ proach. arXiv preprint arXiv:1907.11692.
1112
+ Sewon Min, Victor Zhong, Luke Zettlemoyer, and Han-
1113
+ naneh Hajishirzi. 2019. Multi-hop reading compre-
1114
+ hension through question decomposition and rescor-
1115
+ ing.
1116
+ In Proceedings of the 57th Annual Meeting
1117
+ of the Association for Computational Linguistics,
1118
+ pages 6097–6109, Florence, Italy. Association for
1119
+ Computational Linguistics.
1120
+ Kosuke Nishida, Kyosuke Nishida, Masaaki Nagata,
1121
+ Atsushi Otsuka, Itsumi Saito, Hisako Asano, and
1122
+ Junji Tomita. 2019. Answering while summarizing:
1123
+ Multi-task learning for multi-hop QA with evidence
1124
+ extraction. In Proceedings of the 57th Annual Meet-
1125
+ ing of the Association for Computational Linguistics,
1126
+ pages 2335–2345, Florence, Italy. Association for
1127
+ Computational Linguistics.
1128
+ Lin Qiu, Yunxuan Xiao, Yanru Qu, Hao Zhou, Lei
1129
+ Li, Weinan Zhang, and Yong Yu. 2019.
1130
+ Dynami-
1131
+ cally fused graph network for multi-hop reasoning.
1132
+ In Proceedings of the 57th Annual Meeting of the
1133
+ Association for Computational Linguistics, pages
1134
+ 6140–6150, Florence, Italy. Association for Compu-
1135
+ tational Linguistics.
1136
+ Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and
1137
+ Percy Liang. 2016. SQuAD: 100,000+ questions for
1138
+ machine comprehension of text. In Proceedings of
1139
+ the 2016 Conference on Empirical Methods in Natu-
1140
+ ral Language Processing, pages 2383–2392, Austin,
1141
+ Texas. Association for Computational Linguistics.
1142
+ Heechang Ryu, Hayong Shin, and Jinkyoo Park. 2020.
1143
+ Multi-agent actor-critic with hierarchical graph at-
1144
+ tention network. In Proceedings of the AAAI Con-
1145
+ ference on Artificial Intelligence, volume 34, pages
1146
+ 7236–7243.
1147
+ Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and
1148
+ Hannaneh Hajishirzi. 2018. Bidirectional attention
1149
+ flow for machine comprehension.
1150
+ Linfeng Song, Zhiguo Wang, Mo Yu, Yue Zhang,
1151
+ Radu Florian, and D. Gildea. 2018a.
1152
+ Exploring
1153
+ graph-structured passage representation for multi-
1154
+ hop reading comprehension with graph neural net-
1155
+ works. ArXiv, abs/1809.02040.
1156
+ Linfeng Song, Yue Zhang, Zhiguo Wang, and Daniel
1157
+ Gildea. 2018b.
1158
+ A graph-to-sequence model for
1159
+ AMR-to-text generation. In Proceedings of the 56th
1160
+
1161
+ Annual Meeting of the Association for Computa-
1162
+ tional Linguistics (Volume 1: Long Papers), pages
1163
+ 1616–1626, Melbourne, Australia. Association for
1164
+ Computational Linguistics.
1165
+ Adam Trischler, Tong Wang, Xingdi Yuan, Justin Har-
1166
+ ris, Alessandro Sordoni, Philip Bachman, and Ka-
1167
+ heer Suleman. 2017. NewsQA: A machine compre-
1168
+ hension dataset.
1169
+ In Proceedings of the 2nd Work-
1170
+ shop on Representation Learning for NLP, pages
1171
+ 191–200, Vancouver, Canada. Association for Com-
1172
+ putational Linguistics.
1173
+ Ming Tu, Guangtao Wang, Jing Huang, Yun Tang, Xi-
1174
+ aodong He, and Bowen Zhou. 2019. Multi-hop read-
1175
+ ing comprehension across multiple documents by
1176
+ reasoning over heterogeneous graphs. In Proceed-
1177
+ ings of the 57th Annual Meeting of the Association
1178
+ for Computational Linguistics, pages 2704–2713,
1179
+ Florence, Italy. Association for Computational Lin-
1180
+ guistics.
1181
+ Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
1182
+ Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
1183
+ Kaiser, and Illia Polosukhin. 2017. Attention is all
1184
+ you need. In Advances in neural information pro-
1185
+ cessing systems, pages 5998–6008.
1186
+ Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova,
1187
+ Adriana Romero, Pietro Liò, and Yoshua Bengio.
1188
+ 2017. Graph attention networks. 6th International
1189
+ Conference on Learning Representations.
1190
+ Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Ben-
1191
+ gio, William W. Cohen, Ruslan Salakhutdinov, and
1192
+ Christopher D. Manning. 2018. Hotpotqa: A dataset
1193
+ for diverse, explainable multi-hop question answer-
1194
+ ing.
1195
+ Zhao Zhang, Fuzhen Zhuang, Hengshu Zhu, Zhiping
1196
+ Shi, Hui Xiong, and Qing He. 2020.
1197
+ Relational
1198
+ graph neural network with hierarchical attention for
1199
+ knowledge graph completion.
1200
+ In Proceedings of
1201
+ the AAAI Conference on Artificial Intelligence, vol-
1202
+ ume 34, pages 9612–9619.
1203
+
BNFKT4oBgHgl3EQfWi5i/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
C9AzT4oBgHgl3EQfTvzN/content/2301.01256v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a1df23e7f207db2f86bfcc845a2f0ccb429dffc4867b1d7e29cc5afb85bd964d
3
+ size 24150783
FNE1T4oBgHgl3EQf-gbR/content/tmp_files/2301.03571v1.pdf.txt ADDED
@@ -0,0 +1,1215 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Dipolar Spin Liquid Ending with Quantum Critical Point in a Gd-based Triangular Magnet
2
+ Junsen Xiang,1, ∗ Cheng Su,2, ∗ Ning Xi,3, ∗ Zhendong Fu,4 Zhuo Chen,5 Hai Jin,6 Ziyu Chen,2 Zhao-Jun Mo,7
3
+ Yang Qi,8, 9 Jun Shen,5, 10 Long Zhang,11, 12 Wentao Jin,2, † Wei Li,3, 12, 13, ‡ Peijie Sun,1, § and Gang Su11, 12, ¶
4
+ 1Beijing National Laboratory for Condensed Matter Physics,
5
+ Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China
6
+ 2School of Physics, Beihang University, Beijing 100191, China
7
+ 3CAS Key Laboratory of Theoretical Physics, Institute of Theoretical Physics, Chinese Academy of Sciences, Beijing 100190, China
8
+ 4Neutron Platform, Songshan Lake Materials Laboratory, Dongguan 523808, China
9
+ 5School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081, China
10
+ 6Department of Astronomy, Tsinghua University, Beijing 100084, China
11
+ 7Ganjiang Innovation Academy, Chinese Academy of Sciences, Ganzhou 341119, People’s Republic of China.
12
+ 8State Key Laboratory of Surface Physics, Fudan University, Shanghai 200433, China
13
+ 9Center for Field Theory and Particle Physics, Department of Physics, Fudan University, Shanghai 200433, China
14
+ 10Technical Institute of Physics and Chemistry, Chinese Academy of Sciences, Beijing 100190, China
15
+ 11Kavli Institute for Theoretical Sciences, and School of Physical Sciences,
16
+ University of Chinese Academy of Sciences, Beijng 100049, China
17
+ 12CAS Center for Excellence in Topological Quantum Computation,
18
+ University of Chinese Academy of Sciences, Beijng 100190, China
19
+ 13Peng Huanwu Collaborative Center for Research and Education, Beihang University, Beijing 100191, China
20
+ (Dated: January 10, 2023)
21
+ By performing experimental and model studies of a triangular-lattice dipolar magnet KBaGd(BO3)2 (KBGB),
22
+ we find the highly frustrated magnet with a planar anisotropy hosts a strongly fluctuating dipolar spin liquid
23
+ (DSL) originating from the intriguing interplay between dipolar and Heisenberg interactions. The DSL con-
24
+ stitutes an extended regime in the field-temperature phase diagram, which gets narrowed in temperature range
25
+ as field increases and eventually ends with a quantum critical point (QCP) at Bc ≃ 0.75 T. Based on dipolar
26
+ Heisenberg model calculations, we identify the DSL as a Berezinskii-Kosterlitz-Thouless (BKT) phase. Due to
27
+ the tremendous entropy accumulation that can be related to the strong BKT and quantum fluctuations, unprece-
28
+ dented magnetic cooling effects are observed in the DSL regime and particularly near the QCP, making KBGB
29
+ a superior dipolar coolant over commercial Gd-based refrigerants. We establish a universal phase diagram for
30
+ triangular-lattice dipolar quantum magnets where emergent symmetry plays an essential role, and lay down
31
+ foundations for their applications in sub-Kelvin refrigeration.
32
+ Introduction.— Triangular-lattice quantum antiferromag-
33
+ nets have raised great research interest recently, due to the
34
+ unusual quantum spin states and transitions therein [1, 2].
35
+ One prominent example is the quantum spin liquid (QSL) [3–
36
+ 5] and its possible materialization in organic compounds [6–
37
+ 8] and rare-earth triangular magnets [9–16]. Due to the in-
38
+ triguing spin frustration effects and two dimensionality (2D),
39
+ Berezinskii-Kosterlitz-Thouless (BKT) physics may appear in
40
+ the triangular quantum antiferromagnets. The Co-based quan-
41
+ tum antiferromagnet Na2BaCo(PO4)2 hosts persistent spin
42
+ fluctuations [17–20] till very low temperature, and is proposed
43
+ to posses spin supersolid state with BKT fluctuations of U(1)
44
+ phase [21]. Emergent symmetry, as a consequence of frustra-
45
+ tion, has also been disclosed on the triangular lattice, with a
46
+ recent example of rare-earth magnet TmMgGaO4 [22–27].
47
+ Recently, it has been proposed that the dipolar interactions
48
+ can give rise to QSL in triangular-lattice quantum spin sys-
49
+ tems [29]. Lately such dipolar system has been realized in Yb-
50
+ based triangular compounds [30–34]. However, the dipolar
51
+ ∗ These authors contributed equally to this work.
52
+ † wtjin@buaa.edu.cn
53
+ ‡ w.li@itp.ac.cn
54
+ § pjsun@iphy.ac.cn
55
+ ¶ gsu@ucas.ac.cn
56
+ interactions are rather weak and it is very challenging for con-
57
+ ventional thermodynamic and spectroscopic measurements to
58
+ probe the exotic spin states due to dipolar interactions. On
59
+ the contrary, the rare-earth dipolar magnets with larger mo-
60
+ ments, e.g., Gd-based compounds with µeff ≈ 8µB and high
61
+ spin S = 7/2, are much less explored both in experiments
62
+ and theories. It is expected that the dipolar frustration ef-
63
+ fects are a priori more evident in these systems. Moreover,
64
+ in sub-Kelvin refrigeration for space applications [35, 36] and
65
+ quantum computations [37], high-spin frustrated magnets, es-
66
+ pecially those with spin-liquid like behaviors [38], can have
67
+ great entropy densities and cooling capacity, holding thus
68
+ strong promise as superior coolants.
69
+ In this work, we perform low-temperature thermodynam-
70
+ ics and magnetocalorics measurements on single-crystal sam-
71
+ ples of Gd-based triangular-lattice compound KBaGd(BO3)2
72
+ (KBGB). The thermodynamic measurements suggest a dipo-
73
+ lar spin liquid state with no conventional ordering but strong
74
+ spin fluctuations, which are reflected in the algebraic specific
75
+ heat and imaginary dynamical susceptibility (χ
76
+ ′′
77
+ ac). We estab-
78
+ lish a dipolar Heisenberg model with both dipole-dipole and
79
+ Heisenberg interactions for KBGB. Monte Carlo (MC) sim-
80
+ ulations explain excellently the experimental measurements
81
+ and unveil the exotic spin states and transitions in the phase
82
+ diagram. In particular, the model simulations suggest a two-
83
+ step melting of the clock antiferromagnetic (AF) order via two
84
+ arXiv:2301.03571v1 [cond-mat.str-el] 9 Jan 2023
85
+
86
+ 2
87
+ -0.8
88
+ 0
89
+ 0.8
90
+ -0.8
91
+ 0
92
+ 0.8
93
+ Yy
94
+ Yx
95
+ -0.8
96
+ 0
97
+ 0.8
98
+ -0.8
99
+ 0
100
+ 0.8
101
+ Yy
102
+ Yx
103
+ -0.8
104
+ 0
105
+ 0.8
106
+ -0.8
107
+ 0
108
+ 0.8
109
+ Yy
110
+ Yx
111
+ 6-clock AF
112
+ DSL
113
+ PM
114
+ Temperature
115
+ (b)
116
+ (c)
117
+ (d)
118
+ (e)
119
+ (a)
120
+ b
121
+ a
122
+ c
123
+ K/Ba
124
+ Gd
125
+ B
126
+ O
127
+ b
128
+ a
129
+ a∗
130
+ Si
131
+ Sj
132
+ eij
133
+ FIG. 1. (a) shows the crystal structure of KBaGd(BO3)2, where (b)
134
+ triangular-lattice layers of GdO6 octahedra are separated by the Ba/K
135
+ layers with site mixing. The grey arrows refer to the spins on site i
136
+ and j, and the unit vector eij is also indicated. Dipole-dipole inter-
137
+ actions are bond-dependent and follow the ¯3m site symmetry. (c)-(e)
138
+ are histograms of the order parameter Ψxy ≡ Ψx + iΨy for the 6-
139
+ clock antiferromagnetic (AF) [28], dipolar spin liquid (DSL) with an
140
+ emergent U(1) symmetry, and the paramagnetic (PM) phases.
141
+ BKT transitions, between which a floating BKT phase emerge
142
+ with an emergent U(1) symmetry, well accounting for the ex-
143
+ perimentally observed spin liquid behaviors with enormous
144
+ low-temperature entropy. Consequently, giant magnetocaloric
145
+ effect (MCE) is observed in the quasi-adiabatic demagnetiza-
146
+ tion measurements, where we find a clear dip in temperature
147
+ which suggests the presence of quantum critical point (QCP)
148
+ near Bc ≃ 0.75 T. The lowest temperature of 70 mK clearly
149
+ surpasses that of commercial refrigerant Gd3Ga5O12 (GGG)
150
+ under similar conditions. Overall, the triangular-lattice rare-
151
+ earth dipolar magnets open an avenue for exploring exotic
152
+ spin states as well as finding superior sub-Kelvin coolants.
153
+ Crystal structure and effective model for KBaGd(BO3)2.—
154
+ Centimeter-sized single crystals of KBGB were synthesized
155
+ using the flux method as described in detail in Supplementary
156
+ Materials (SM) [28], and the X-ray diffraction measurements
157
+ suggest high quality of the single crystals. KBGB is found
158
+ to crystallize in a trigonal structure [40, 41] with space group
159
+ R-3m [c.f., Fig. 1(a)], and has a relatively high ionic density
160
+ of 6.4 nm−3. As shown in Fig. 1(b), magnetic Gd3+ ions with
161
+ 4f 7 electron configuration (L = 0, S = 7/2) form perfect
162
+ triangular lattice.
163
+ The dipolar interaction between magnetic ions Gd3+
164
+ has a characteristic energy Edp
165
+
166
+ 2µ0µ2
167
+ eff/4πa3
168
+
169
+ 0.05 meV (with µeff ≈ 8µB), which determines the low-
170
+ temperature spin states in KBGB. To simulate such Gd-
171
+ based dipolar magnet, we consider the following Hamil-
172
+ tonian, H
173
+ =
174
+ JH
175
+
176
+ ⟨i,j⟩NN Si · Sj + JD
177
+
178
+ i,j[Si · Sj −
179
+ 3(Si · eij)(Sj · eij)]/r3
180
+ ij, where eij(rij) refers to the unit vec-
181
+ tor(distance) between site i and j. JH and JD refer to the
182
+ nearest neighbor (NN) Heisenberg and dipole-dipole interac-
183
+ tions, respectively. As the dipolar interactions show rapid (cu-
184
+ bic) power-law decay and the longer range interactions can be
185
+ washed out, we keep only NN terms as
186
+ HDH =
187
+
188
+ ⟨i,j⟩NN
189
+ J Si · Sj − D (Si · eij)(Sj · eij),
190
+ (1)
191
+ where J = JH + JD/a3 is the NN isotropic coupling and
192
+ D = 3JD/a3 refers to the dipolar anisotropic term. We find
193
+ the NN dipolar Heisenberg (DH) model with couplings J =
194
+ 47 mK and D = 80 mK very well describe the compound
195
+ and accurately reproduce the experimental measurements on
196
+ KBGB. The MC simulations are performed on up to 60 × 60
197
+ triangular lattice. Due to the high-spin state with S = 7/2,
198
+ classical MC simulations capture well the finite-temperature
199
+ properties of KBGB [28]. We guarantee the error bars to be
200
+ always smaller than the symbol size in the presented data.
201
+ Magnetic specific heat, susceptibility, and dipolar spin
202
+ liquid.— In Fig. 2(a) we show the zero-field specific heat
203
+ Cm measured down to 65 mK. There exists a round peak at
204
+ T ∗ ≃ 218 mK, below which the system exhibits Cm ∼ T 2
205
+ with algebraic scaling, resembling that of 2D Heisenberg
206
+ or XY quantum spin model with U(1) symmetry [42, 43].
207
+ The dipolar anisotropy in Eq. (1), like in spin-orbit magnets,
208
+ leads to a discretized C3 rotational symmetry, and it gener-
209
+ ically corresponds to a divergent Cm peak when transition-
210
+ ing to low-T symmetry breaking phase.
211
+ The presence of
212
+ round peak and T 2 scaling in Cm is very remarkable, which
213
+ suggests a liquid-like and strongly fluctuating spin state. In
214
+ Fig. 2(a), when compared to the renowned Gd-based refrig-
215
+ erant GGG [36, 39, 44, 45], KBGB has tremendous low-
216
+ temperature specific heat, far surpassing that of GGG.
217
+ In Fig. 2(b), we apply out-of-plane fields (B//c) to the com-
218
+ pound, and find also round peaks in Cm curves, which move
219
+ towards lower temperature with heights slightly reduced. This
220
+ suggests that the spin liquid states constitute an extended
221
+ phase that we dub as dipolar spin liquid (DSL). As field fur-
222
+ ther increases and exceeds about 0.75 T, the DSL behaviors
223
+ disappears [c.f., the contour plot of Cm/T in Fig. 3(b)], and
224
+ the Cm peak moves now to high-temperature side, with the
225
+ low-T peak and low-energy fluctuations quickly suppressed.
226
+ In Fig. 2(c), we perform magnetization measurements
227
+ on single-crystal sample of KBGB, and find a clear mag-
228
+ netic anisotropy between the out-of-plane (//c axis) and in-
229
+ plane (//a) directions.
230
+ This anisotropy can be clearly rec-
231
+ ognized in the different saturation magnetization moments
232
+ and transition field values, i.e., 1 T(0.5 T) along c(a) axis.
233
+ In Fig. 2(d), we perform low-temperature dc susceptibility
234
+ (χdc) measurements, and find χdc also exhibits a clear easy-
235
+ plane anisotropy. In addition, a small but sensible in-plane
236
+ anisotropy between a and a∗ [see inset of Fig. 2(c)] is also
237
+ observed, consistent with the intrinsic anisotropy in bond-
238
+ dependent dipolar interaction [c.f., Eq. (1)].
239
+ To further explore the DSL, ac magnetic susceptibilities are
240
+ measured in Figs. 2(e,f), with χ′
241
+ ac and χ′′
242
+ ac for real and imag-
243
+ inary parts, respectively. The real χ′
244
+ ac exhibits a frequency-
245
+ independent maximum and remains large even below the char-
246
+ acteristic temperature scale T ∗.
247
+ Therefore, although there
248
+ exist K/Ba site mixing in the compound, the spin-glass sce-
249
+ nario can be excluded in KBGB. Interestingly, the imaginary
250
+
251
+ 3
252
+ 1
253
+ 10
254
+ 100
255
+ 0
256
+ 2
257
+ 4
258
+ 6
259
+ 8
260
+ 10
261
+ 12
262
+ Exp.
263
+ Model
264
+ χdc (emu·Oe-1·mol-1
265
+ Gd)
266
+ T (K)
267
+ a
268
+ a*
269
+ c
270
+ 0.1 T
271
+ -2 0 2 4 6 8 10
272
+ 0.0
273
+ 0.4
274
+ 0.8
275
+ 1.2
276
+ χdc
277
+ -1
278
+ θa -0.30 K
279
+ θa* -0.33 K
280
+ θc
281
+ -1.32 K
282
+ 0.0
283
+ 0.1
284
+ 0.2
285
+ 0.3
286
+ 0.4
287
+ 0.5
288
+ 0
289
+ 25
290
+ 50
291
+ 75
292
+ T (K)
293
+ Cm/T (J·mol-1
294
+ Gd·K-2)
295
+ KBGB
296
+ GGG
297
+ 0 T
298
+ Cm/T ~ T
299
+ T* 218 mK
300
+ 0.1
301
+ 1
302
+ 0.0
303
+ 0.5
304
+ 1.0
305
+ 1.5
306
+ 2.0
307
+ 4943 Hz
308
+ 6253 Hz
309
+ 9984 Hz
310
+ χac'' (a.u.)
311
+ T (K)
312
+ T*
313
+ 0
314
+ 1
315
+ 2
316
+ 3
317
+ 4
318
+ 0
319
+ 2
320
+ 4
321
+ 6
322
+ 8
323
+ 10
324
+ B // a
325
+ B // c
326
+ Moment (µB/Gd)
327
+ B (T)
328
+
329
+ 0.4 K
330
+ 2.36
331
+ 2.49
332
+ Model
333
+ ga
334
+ gc
335
+ 0.1
336
+ 1
337
+ 1.2
338
+ 1.4
339
+ 1.6
340
+ 1.8
341
+ 2.0
342
+ 2.2
343
+ 2.4
344
+ 91 Hz
345
+ 955 Hz
346
+ 2439 Hz
347
+ 3087 Hz
348
+ 3910 Hz
349
+ T (K)
350
+ χac' (a.u.)
351
+ T*
352
+ 0.0
353
+ 0.1
354
+ 0.2
355
+ 0.3
356
+ 0.4
357
+ 0.5
358
+ 0
359
+ 25
360
+ 50
361
+ 75
362
+ 0.25 T
363
+ 0.5 T
364
+ 0.75 T
365
+ Cm/T (J·mol-1
366
+ Gd·K-2)
367
+ T (K)
368
+ 1 T
369
+ 2 T
370
+ 3 T
371
+ 4 T
372
+ (a)
373
+ (d)
374
+ (e)
375
+ (f)
376
+ (c)
377
+ (b)
378
+ a*
379
+ a
380
+ FIG. 2. Specific heat of KBGB under (a) zero and (b) finite fields along out-of-plane direction (B//c). An algebraic Cm ∼ T 2 scaling is
381
+ observed below the round peak temperature T ∗, and the Cm/T values far outweigh that of GGG [39] for T ≲ T ∗. In (b) we find the round
382
+ peak in Cm/T firstly moves towards lower temperature and later for B > Bc ≃ 0.75 T the low-temperature Cm quickly gets suppressed. (c)
383
+ shows the magnetization curves of the single-crystal KBGB sample for B//a and //c, and the results show excellent agreement with the DH
384
+ model calculations (solid lines). The saturation moments are µsat
385
+ a
386
+ ≃ 8.26µB and µsat
387
+ c
388
+ ≃ 8.72µB, from which we determine the Landé factors
389
+ ga ≃ 2.36 and gc ≃ 2.49, respectively. The as-grown KBGB single crystal is shown in the inset, with directions a and a∗ also indicated. (d)
390
+ shows the molar dc magnetic susceptibilities (χdc) measured along the a, a∗, and c axes, respectively, where the solid lines representing the
391
+ DH model calculations show excellent agreements. The inset shows the Curie-Weiss fittings in the paramagnetic regime 0.4 K ≤ T ≤ 10 K,
392
+ with the fitted Curie-Weiss temperatures θa,a∗,c also indicated. (e, f) present respectively the real and imaginary ac susceptibilities measured
393
+ with different frequencies.
394
+ ac susceptibility χ′′
395
+ ac(T), although being featureless for low
396
+ frequencies ω ≲ 4 kHz, show a clear temperature-dependent
397
+ behavior for higher frequencies in Fig. 2(f). Considering that
398
+ χ′′(ω) can be directly related to the dynamical correlation
399
+ S(ω) through the fluctuation-dissipation theorem, χ′′(ω) ∝
400
+ ω
401
+ T S(ω) (ω ≪ T), this clearly suggests the persistence of low-
402
+ energy spin fluctuations even below T ∗ and supports the spin-
403
+ liquid scenario.
404
+ Magnetocaloric effect and quantum critical point.— In
405
+ Fig. 3(a), we perform quasi-adiabatic demagnetization mea-
406
+ surements and obtain the isentropic curves. It is found that
407
+ KBGB clearly outperforms GGG in the minimal temperature,
408
+ i.e., Tm ≃ 70 mK (KBGB) vs. 322 mK (GGG), when starting
409
+ from the same initial condition of Ti = 2 K and Bi = 6 T.
410
+ In Fig. 3(b) we provide more of the isentropic lines from dif-
411
+ ferent initial conditions, and observe the highly asymmetric
412
+ isentropes, which “levels off” in the bright DSL regime as in-
413
+ dicated by large values of Cm/T.
414
+ For KBGB, the lowest temperature Tm is achieved at the
415
+ dip in isentropic lines and remains below 100 mK in the
416
+ small field side. This happens also for measurements starting
417
+ from rather low temperature Ti ≃ 95 mK, where the lowest
418
+ Tm ≃ 33 mK. Such unprecedented MCE response strongly
419
+ corroborates the existence of QCP at Bc ≃ 0.75 T. The mag-
420
+ netic Grüneisen ratio ΓB =
421
+ 1
422
+ T ( ∂T
423
+ ∂B )S has been widely used
424
+ in the studies of heavy fermion [46–50] and low-dimensional
425
+ quantum spin systems [51–54]. In the inset of Fig. 3 an ev-
426
+ ident peak-dip structure with sign change is observed [55–
427
+ 58], and the peak height exceeds 4 times that of GGG. Such a
428
+ prominent critical cooling effect provides valuable MCE evi-
429
+ dence for QCP in the compound KBGB.
430
+ Emergent symmetry in KBGB.— According to the magne-
431
+ tothermal and MCE measurements above, we arrive at the
432
+ phase diagram of KBGB in Fig. 3(b).
433
+ The two schematic
434
+ dashed lines, enclosing the DSL with large Cm/T, meet at a
435
+ QCP (Bc) where the demagnetization process reaches its low-
436
+ est temperature. Besides QCP, within the DSL regime we find
437
+ persistent spin fluctuations and cooling effects whose origin is
438
+ clarified by model calculations below.
439
+ We conduct MC calculations of the DH model [Eq. (1)]
440
+ for KBGB. As the model is highly frustrated in the out-of-
441
+ plane direction, the order parameter lies within the ab plane.
442
+ Note although the determined Landé factor gc ≃ 2.49 is
443
+ slightly larger than ga ≃ 2.36, the intrinsic planar anisotropy
444
+ of dipolar interaction leads to larger in-plane χdc (along a
445
+ and a∗ axes) than that along the c axis. The negative Curie-
446
+ Weiss temperatures fitted from the dc susceptibility reflect the
447
+ AF nature, and the slightly different θa ≃ −300 mK and
448
+ θa∗ ≃ −330 mK shows the in-plane anisotropy. In Fig. 2(d),
449
+ we find the anisotropic susceptibility and magnetization mea-
450
+
451
+ 4
452
+ 0
453
+ 1
454
+ 2
455
+ 3
456
+ 4
457
+ 5
458
+ 6
459
+ 0.1
460
+ 1
461
+ T (K)
462
+ B (T)
463
+ 70 mK
464
+ KBGB
465
+ GGG
466
+ B // c
467
+ 322 mK
468
+ 2 K
469
+ 33 mK
470
+ 0.0
471
+ 0.4
472
+ 0.8
473
+ 1.2
474
+ 0.0
475
+ 0.1
476
+ 0.2
477
+ 0.3
478
+ B (T)
479
+ T (K)
480
+ 0
481
+ 20
482
+ 40
483
+ 60
484
+ Cm/T
485
+ DSL
486
+ QCP
487
+ Quasi-Adiabatic
488
+ PM
489
+ 6-clock AF
490
+ (a)
491
+ (b)
492
+ ΓΒ (T-1)
493
+ 0 1 2 3 4
494
+ -1
495
+ 0
496
+ 1
497
+ 2
498
+ 3
499
+ 4×GGG
500
+ Bc
501
+ FIG. 3. (a) shows the quasi-adiabatic isentropes measured in KBGB
502
+ under out-of-plane field (see details in SM [28]). The KBGB curve
503
+ exhibits a clear dip at the lowest temperature Tm ≃ 70 mK, much
504
+ lower than that of GGG (Tm ≃ 322 mK). Starting from Ti ≃ 95 mK,
505
+ KBGB is observed to cool down to remarkably low temperature
506
+ Tm ≃ 33 mK in the dip (blue dotted line). The inset shows the mag-
507
+ netic Grüneisen ratio ΓB deduced from the curves in (a). (b) shows
508
+ the phase diagram of KBGB with the contour plot of Cm/T in the
509
+ background. The bright regime with large spin fluctuations represent
510
+ the DSL, with schematic dashed line boundaries, ending up with a
511
+ QCP at Bc ≃ 0.75 T.
512
+ sured along a and c axes can be well captured by the DH
513
+ model. Besides, the model calculations of specific heat also
514
+ obtain a round peak at about 270 mK, which again gets sup-
515
+ pressed as field increases (see SM [28]), very much resem-
516
+ bling the experimental data in Figs. 2(a,b). The comparisons
517
+ confirm that the compound KBGB can indeed be accurately
518
+ described by the DH model.
519
+ To characterize the spin states in the phase diagram, we in-
520
+ troduce the order parameter Ψxy ≡ meiθ = �
521
+ j eiQrj(mx
522
+ j +
523
+ imy
524
+ j ),
525
+ where j runs over the lattice sites and Q
526
+ =
527
+ ± 1
528
+ 2a∗, ± 1
529
+ 2b∗, ± 1
530
+ 2(a∗ − b∗) [28]. Histogram of the complex
531
+ order parameter Ψxy at various temperature are shown in
532
+ Figs. 1(c-e). At low temperature, the dipolar system exhibits
533
+ a 6-clock AF order corresponding to θ = 0, ±π/3, ±2π/3,
534
+ and π [28]. As temperature ramps up, the six points in the
535
+ histogram prolong and merge into a circle with emergent
536
+ U(1) symmetry, where the angle θ can choose arbitrary angle.
537
+ As temperature further enhances, eventually the amplitude m
538
+ vanishes and the system enters the conventional PM phase.
539
+ Recall that the 6-state clock model with an anisotropic term
540
+ ∼ cos (6θ) undergoes two successive BKT transitions [59],
541
+ between which the anisotropic term becomes irrelevant per-
542
+ turbation. Based on this symmetry argument, we consider the
543
+ intermediate DSL in the system as BKT phase with emer-
544
+ gent U(1) symmetry and effectively described by 2D XY
545
+ model [60–63]. The emergent symmetry extends also to the
546
+ zero-temperature QCP as the clock term is dangerously irrele-
547
+ vant [60], and the transition directly between the 6-fold clock
548
+ symmetry broken and PM phases belong to the 3D XY univer-
549
+ sality class. Therefore, the emergent symmetry constitutes a
550
+ key for demystifying spin-liquid state and quantum criticality
551
+ in the compound KBGB.
552
+ 0
553
+ 60
554
+ 120
555
+ 180
556
+ 0
557
+ 1
558
+ 2
559
+ 3
560
+ 4
561
+ T (K)
562
+ Time (min)
563
+ 4 T
564
+ 6 T
565
+ 0.1
566
+ 1
567
+ 0
568
+ 5
569
+ 10
570
+ 15
571
+ 20
572
+ Sm (×10·J·Kg-1·K-1)
573
+ T (K)
574
+ ∆Q
575
+ 4 T
576
+ ∆Sm
577
+ 0 T
578
+ (a)
579
+ 0.1
580
+ 1
581
+ 10
582
+ 0
583
+ 10
584
+ 20
585
+ 30
586
+ 40
587
+ � � m (J·Kg-1·K-1)
588
+ T (K)
589
+ KBGB
590
+ GGG
591
+ 4 T
592
+ 2 T
593
+ 1 T
594
+ (b)
595
+ FIG. 4. (a) The quasi-adiabatic demagnetization cooling curves of
596
+ KBGB, starting from two different initial conditions (Ti = 4 K,
597
+ Bi = 4 T) and (Ti = 2 K, Bi = 6 T), with reached lowest tem-
598
+ perature Tm ≃ 205 mK and 70 mK, respectively. Parasitic heat
599
+ loads are estimated to be 0.2 µW for Ti = 4 K environment and
600
+ 0.05 µW for Ti = 2 K. The inset shows magnetic entropy under
601
+ zero and 4 T fields, with the shaded area representing the absorbed
602
+ heat ∆Q = 47.44 J·Kg−1 in the hold process. (b) plots the entropy
603
+ change ∆Sm vs. T, for fields decreasing from 1 T, 2 T, and 4 T to
604
+ zero, respectively. Comparisons to GGG are also presented [39, 44].
605
+ Superior cooling performance.— Starting from 2 K en-
606
+ vironment, KBGB can reach as low as 70 mK as shown
607
+ in Fig. 4(a), such a low cooling temperature far surpasses
608
+ other Gd-based refrigerants, e.g., GGG (322 mK) and GdLiF4
609
+ (480 mK) [64]. Besides, KBGB also exhibits long hold time
610
+ and large isothermal entropy change ∆Sm. In Fig. 4(a) we
611
+ show that KBGB remains in low temperature for a long period
612
+ after the field is exhausted. In the environment temperature of
613
+ 2 K, 0.5 g KBGB remains below 140 mK for th ≈ 2 h under
614
+ 0.05 µW heat leak, which can be ascribed to the large heat
615
+ absorption ∆Q depicted in the inset of Fig. 4(a).
616
+ The isothermal entropy change ∆Sm characterizes the
617
+ cooling capacity of refrigerants. In Fig. 4(b), we compare
618
+ ∆Sm of KBGB with that of GGG, and find that in the whole
619
+ temperature range concerned KBGB has significantly larger
620
+ ∆Sm for 1 T field. Moreover, the maximal ∆Sm of KBGB lo-
621
+ cates below 1 K [shaded regime in Fig. 4(b)], and the entropy
622
+ change in KBGB exceeds that of GGG in this sub-Kelvin
623
+ regime of central interest. Overall, the low cooling temper-
624
+ ature Tm, long hold time th, and enormous entropy change
625
+ ∆Sm in the sub-Kelvin regime lead to the conclusion that
626
+ KBGB serves a superior quantum magnet coolant.
627
+ Discussions and outlook.— The pursue for high entropy
628
+ density and low ordering temperature constitutes two oppos-
629
+ ing factors hard to fulfill simultaneously in optimizing sub-
630
+ Kelvin refrigerants. Here the spin frustration and quantum
631
+ criticality in the dipolar system come to the rescue. We show
632
+ that the compound KBaGd(BO3)2 with high Gd3+ ion density
633
+ yet form a disordered and strongly fluctuating spin liquid till
634
+ extremely low temperature, giving rise to the superior cooling
635
+ capacity due to the entropy accumulation near QCP. We use
636
+ the DH model within NN interactions to describe KBGB and
637
+ find it well reproduces the experimental results. Inclusion of
638
+ further neighboring dipolar couplings will not change the con-
639
+
640
+ 5
641
+ clusion here, as it has been shown to maintain the universality
642
+ class of BKT transitions in planar dipolar models [63, 65].
643
+ The scenario of DSL ending up with emergent U(1) QCP
644
+ may also be applicable to other dipolar quantum magnets. Re-
645
+ cent progress in experimental studies reveal several families of
646
+ rare-earth triangular quantum dipolar antiferromagnets, e.g.,
647
+ Ba3REB3O9/Ba3REB9O18 (with RE a rare-earth ion) [32, 33]
648
+ and ABaRE(BO3)2 (with A an alkali ion) [66, 67]. It has been
649
+ observed that in Ba3YbB3O9 that 80% entropy remain below
650
+ 56 mK [31], despite a dipolar energy scale of about 160 mK,
651
+ suggesting that the DSL may also play a role in the Yb-based
652
+ dipolar compounds. Therefore, this work opens a venue for
653
+ hunting exotic spin states as well as superior quantum coolants
654
+ in triangular dipolar magnets.
655
+ Note added.— Upon finishing the present work, we are
656
+ aware of a recent work [68] also conducting MCE study of
657
+ KBGB with however polycrystalline samples, where they find
658
+ strong cooling effect down to 121 mK.
659
+ Acknowledgements.— W.L. is indebted to Yuan Wan and
660
+ Tao Shi for helpful discussions. W.J. and C.S. acknowledge
661
+ the support from the beamline 1W1A of the Beijing Syn-
662
+ chrotron Radiation Facility.
663
+ This work was supported by
664
+ the National Natural Science Foundation of China (Grant
665
+ Nos. 12222412, 11834014, 11974036, 12047503, 12074023,
666
+ 12074024, 12174387, and 12141002), National Key R
667
+ & D Program of China (Grant No. 2018YFA0305800),
668
+ Strategic
669
+ Priority
670
+ Research
671
+ Program
672
+ of
673
+ CAS
674
+ (Grant
675
+ No. XDB28000000), and CAS Project for Young Scien-
676
+ tists in Basic Research (Grant No. YSBR-057). We thank the
677
+ HPC-ITP for the technical support and generous allocation
678
+ of CPU time. This work was supported by the Synergetic
679
+ Extreme Condition User Facility (SECUF).
680
+ [1] M. F. Collins and O. A. Petrenko, Review/synthèse: Triangular
681
+ antiferromagnets, Can. J. Phys. 75, 605 (1997).
682
+ [2] O. A. Starykh, Unusual ordered phases of highly frustrated
683
+ magnets: a review, Rep. Prog. Phys. 78, 052502 (2015).
684
+ [3] P. W. Anderson, Resonating valence bonds: A new kind of in-
685
+ sulator?, Mater. Res. Bull. 8, 153 (1973).
686
+ [4] Y. Zhou, K. Kanoda, and T.-K. Ng, Quantum spin liquid states,
687
+ Rev. Mod. Phys. 89, 025003 (2017).
688
+ [5] L. Balents, Spin liquids in frustrated magnets, Nature 464, 199
689
+ (2010).
690
+ [6] Y. Shimizu, K. Miyagawa, K. Kanoda, M. Maesato, and
691
+ G. Saito, Spin liquid state in an organic Mott insulator with a
692
+ triangular lattice, Phys. Rev. Lett. 91, 107001 (2003).
693
+ [7] M. Yamashita, N. Nakata, Y. Senshu, M. Nagata, H. M. Ya-
694
+ mamoto, R. Kato, T. Shibauchi, and Y. Matsuda, Highly mo-
695
+ bile gapless excitations in a two-dimensional candidate quan-
696
+ tum spin liquid, Science 328, 1246 (2010).
697
+ [8] K. Kanoda and R. Kato, Mott physics in organic conductors
698
+ with triangular lattices, Annu. Rev. Condens. Matter Phys. 2,
699
+ 167 (2011).
700
+ [9] Y. Li, H. Liao, Z. Zhang, S. Li, F. Jin, L. Ling, L. Zhang, Y. Zou,
701
+ L. Pi, Z. Yang, J. Wang, Z. Wu, and Q. Zhang, Gapless quantum
702
+ spin liquid ground state in the two-dimensional spin-1/2 trian-
703
+ gular antiferromagnet YbMgGaO4, Sci. Rep. 5, 16419 (2015).
704
+ [10] Y. Li, G. Chen, W. Tong, L. Pi, J. Liu, Z. Yang, X. Wang, and
705
+ Q. Zhang, Rare-earth triangular lattice spin liquid: A single-
706
+ crystal study of YbMgGaO4, Phys. Rev. Lett. 115, 167203
707
+ (2015).
708
+ [11] Y. Shen, Y.-D. Li, H. Wo, Y. Li, S. Shen, B. Pan, Q. Wang, H. C.
709
+ Walker, P. Steffens, M. Boehm, Y. Hao, D. L. Quintero-Castro,
710
+ L. W. Harriger, M. D. Frontzek, L. Hao, S. Meng, Q. Zhang,
711
+ G. Chen, and J. Zhao, Evidence for a spinon fermi surface in
712
+ a triangular-lattice quantum-spin-liquid candidate, Nature 540,
713
+ 559 (2016).
714
+ [12] J. A. M. Paddison, M. Daum, Z. Dun, G. Ehlers, Y. Liu, M. B.
715
+ Stone, H. Zhou, and M. Mourigal, Continuous excitations of the
716
+ triangular-lattice quantum spin liquid YbMgGaO4, Nat. Phys.
717
+ 13, 117 (2017).
718
+ [13] Y. Shen, Y.-D. Li, H. C. Walker, P. Steffens, M. Boehm,
719
+ X. Zhang, S. Shen, H. Wo, G. Chen, and J. Zhao, Fractional-
720
+ ized excitations in the partially magnetized spin liquid candi-
721
+ date YbMgGaO4, Nat. Commun. 9, 4138 (2018).
722
+ [14] W. Liu, Z. Zhang, J. Ji, Y. Liu, J. Li, X. Wang, H. Lei, G. Chen,
723
+ and Q. Zhang, Rare-earth chalcogenides: A large family of
724
+ triangular lattice spin liquid candidates, Chin. Phys. Lett. 35,
725
+ 117501 (2018).
726
+ [15] M. M. Bordelon, E. Kenney, C. Liu, T. Hogan, L. Posthuma,
727
+ M. Kavand, Y. Lyu, M. Sherwin, N. P. Butch, C. Brown, M. J.
728
+ Graf, L. Balents, and S. D. Wilson, Field-tunable quantum dis-
729
+ ordered ground state in the triangular-lattice antiferromagnet
730
+ NaYbO2, Nat. Phys. 15, 1058 (2019).
731
+ [16] P.-L. Dai, G. Zhang, Y. Xie, C. Duan, Y. Gao, Z. Zhu, E. Feng,
732
+ Z. Tao, C.-L. Huang, H. Cao, A. Podlesnyak, G. E. Granroth,
733
+ M. S. Everett, J. C. Neuefeind, D. Voneshen, S. Wang, G. Tan,
734
+ E. Morosan, X. Wang, H.-Q. Lin, L. Shu, G. Chen, Y. Guo,
735
+ X. Lu, and P. Dai, Spinon fermi surface spin liquid in a triangu-
736
+ lar lattice antiferromagnet NaYbSe2, Phys. Rev. X 11, 021044
737
+ (2021).
738
+ [17] R. Zhong, S. Guo, G. Xu, Z. Xu, and R. J. Cava, Strong quan-
739
+ tum fluctuations in a quantum spin liquid candidate with a
740
+ Co-based triangular lattice, Proc. Natl. Acad. Sci. U.S.A. 116,
741
+ 14505 (2019).
742
+ [18] N. Li, Q. Huang, X. Y. Yue, W. J. Chu, Q. Chen, E. S. Choi,
743
+ X. Zhao, H. D. Zhou, and X. F. Sun, Possible itinerant excita-
744
+ tions and quantum spin state transitions in the effective spin-1/2
745
+ triangular-lattice antiferromagnet Na2BaCo(PO4)2, Nat. Com-
746
+ mun. 11, 4216 (2020).
747
+ [19] S. Lee, C. H. Lee, A. Berlie, A. D. Hillier, D. T. Adroja,
748
+ R. Zhong, R. J. Cava, Z. H. Jang, and K.-Y. Choi, Temporal and
749
+ field evolution of spin excitations in the disorder-free triangu-
750
+ lar antiferromagnet Na2BaCo(PO4)2, Phys. Rev. B 103, 024413
751
+ (2021).
752
+ [20] C. Wellm, W. Roscher, J. Zeisner, A. Alfonsov, R. Zhong, R. J.
753
+ Cava, A. Savoyant, R. Hayn, J. van den Brink, B. Büchner,
754
+ O. Janson, and V. Kataev, Frustration enhanced by Kitaev ex-
755
+ change in a jeff =
756
+ 1
757
+ 2 triangular antiferromagnet, Phys. Rev. B
758
+ 104, L100420 (2021).
759
+ [21] Y. Gao, Y.-C. Fan, H. Li, F. Yang, X.-T. Zeng, X.-L. Sheng,
760
+ R. Zhong, Y. Qi, Y. Wan, and W. Li, Spin supersolidity
761
+ in nearly ideal easy-axis triangular quantum antiferromagnet
762
+ Na2BaCo(PO4)2, npj Quantum Mater. 7, 89 (2022).
763
+ [22] F. A. Cevallos, K. Stolze, T. Kong, and R. J. Cava, Anisotropic
764
+ magnetic properties of the triangular plane lattice material
765
+ TmMgGaO4, Mater. Res. Bull. 105, 154 (2018).
766
+
767
+ 6
768
+ [23] Y. Shen, C. Liu, Y. Qin, S. Shen, Y.-D. Li, R. Bewley,
769
+ A. Schneidewind, G. Chen, and J. Zhao, Intertwined dipolar and
770
+ multipolar order in the triangular-lattice magnet TmMgGaO4,
771
+ Nat. Commun. 10, 4530 (2019).
772
+ [24] Y. Li, S. Bachus, H. Deng, W. Schmidt, H. Thoma, V. Hutanu,
773
+ Y. Tokiwa, A. A. Tsirlin, and P. Gegenwart, Partial up-up-
774
+ down order with the continuously distributed order parameter
775
+ in the triangular antiferromagnet TmMgGaO4, Phys. Rev. X 10,
776
+ 011007 (2020).
777
+ [25] H. Li, Y. D. Liao, B.-B. Chen, X.-T. Zeng, X.-L. Sheng, Y. Qi,
778
+ Z. Y. Meng, and W. Li, Kosterlitz-Thouless melting of magnetic
779
+ order in the triangular quantum Ising material TmMgGaO4,
780
+ Nat. Commun. 11, 1111 (2020).
781
+ [26] Z. Hu, Z. Ma, Y.-D. Liao, H. Li, C. Ma, Y. Cui, Y. Shangguan,
782
+ Z. Huang, Y. Qi, W. Li, Z. Y. Meng, J. Wen, and W. Yu, Ev-
783
+ idence of the Berezinskii-Kosterlitz-Thouless phase in a frus-
784
+ trated magnet, Nat. Commun. 11, 5631 (2020).
785
+ [27] Z. Dun, M. Daum, R. Baral, H. E. Fischer, H. Cao, Y. Liu,
786
+ M. B. Stone, J. A. Rodriguez-Rivera, E. S. Choi, Q. Huang,
787
+ H. Zhou, M. Mourigal, and B. A. Frandsen, Neutron scatter-
788
+ ing investigation of proposed Kosterlitz-Thouless transitions in
789
+ the triangular-lattice ising antiferromagnet TmMgGaO4, Phys.
790
+ Rev. B 103, 064424 (2021).
791
+ [28] Supplementary Sec. 1 describes the KBGB sample prepara-
792
+ tion and their XRD charactherization. The magnetothermal and
793
+ magnetocaloric measurements are elaborated in Secs. 2 and 3,
794
+ respectively. Sec. 4 is devoted to additional model calculations.
795
+ [29] N. Y. Yao, M. P. Zaletel, D. M. Stamper-Kurn, and A. Vish-
796
+ wanath, A quantum dipolar spin liquid, Nat. Phys. 14, 405
797
+ (2018).
798
+ [30] K. Y. Zeng, L. Ma, Y. X. Gao, Z. M. Tian, L. S. Ling, and
799
+ L. Pi, NMR study of the spin excitations in the frustrated anti-
800
+ ferromagnet Yb(BaBO3)3 with a triangular lattice, Phys. Rev.
801
+ B 102, 045149 (2020).
802
+ [31] R. Bag, M. Ennis, C. Liu, S. E. Dissanayake, Z. Shi, J. Liu,
803
+ L. Balents, and S. Haravifard, Realization of quantum dipoles
804
+ in triangular lattice crystal Ba3Yb(BO3)3, Phys. Rev. B 104,
805
+ L220403 (2021).
806
+ [32] H. Cho,
807
+ S. J. Blundell,
808
+ T. Shiroka,
809
+ K. MacFarquhar-
810
+ son, D. Prabhakaran, and R. Coldea, Studies on Novel
811
+ Yb-based Candidate Triangular Quantum Antiferromagnets:
812
+ Ba3YbB3O9 and Ba3YbB9O18, arXiv:2104.01005 (2021).
813
+ [33] J. Khatua, M. Pregelj, A. Elghandour, Z. Jagliˇcic, R. Klingeler,
814
+ A. Zorko, and P. Khuntia, Magnetic properties of triangular lat-
815
+ tice antiferromagnets Ba3RB9O18 (R = Yb, Er), Phys. Rev. B
816
+ 106, 104408 (2022).
817
+ [34] C. Y. Jiang, Y. X. Yang, Y. X. Gao, Z. T. Wan, Z. H. Zhu, T. Shi-
818
+ roka, C. S. Chen, Q. Wu, X. Li, J. C. Jiao, K. W. Chen, Y. Bao,
819
+ Z. M. Tian, and L. Shu, Spin excitations in the quantum dipolar
820
+ magnet Yb(BaBO3)3, Phys. Rev. B 106, 014409 (2022).
821
+ [35] C. Hagmann and P. L. Richards, Adiabatic demagnetization re-
822
+ frigerators for small laboratory experiments and space astron-
823
+ omy, Cryogenics 35, 303 (1995).
824
+ [36] P. J. Shirron, Applications of the magnetocaloric effect in
825
+ single-stage, multi-stage and continuous adiabatic demagneti-
826
+ zation refrigerators, Cryogenics 62, 130 (2014).
827
+ [37] A. E. Jahromi, P. J. Shirron, and M. J. DiPirro, Sub-Kelvin Cool-
828
+ ing Systems for Quantum Computers, Tech. Rep. (NASA God-
829
+ dard Space Flight Center Greenbelt, MD, United States, 2019).
830
+ [38] X.-Y. Liu, Y. Gao, H. Li, W. Jin, J. Xiang, H. Jin, Z. Chen,
831
+ W. Li, and G. Su, Quantum spin liquid candidate as superior re-
832
+ frigerant in cascade demagnetization cooling, Commun. Phys.
833
+ 108, 233 (2022).
834
+ [39] P. Schiffer, A. P. Ramirez, D. A. Huse, and A. J. Valentino, In-
835
+ vestigation of the field induced antiferromagnetic phase transi-
836
+ tion in the frustrated magnet: Gadolinium gallium garnet, Phys.
837
+ Rev. Lett. 73, 2500 (1994).
838
+ [40] M. B. Sanders, F. A. Cevallos, and R. J. Cava, Magnetism in
839
+ the KBaRE(BO3)2(RE = Sm, Eu, Gd, Tb, Dy, Ho, Er, Tm, Yb,
840
+ Lu) series: materials with a triangular rare earth lattice, Mater.
841
+ Res. Express 4, 036102 (2017).
842
+ [41] S. Guo, T. Kong, F. A. Cevallos, K. Stolze, and R. Cava, Crystal
843
+ growth, crystal structure and anisotropic magnetic properties of
844
+ KBaR(BO3)2 (R=Y, Gd, Tb, Dy, Ho, Tm, Yb and Lu) triangular
845
+ lattice materials, J. Magn. Magn. Mater. 472, 104 (2019).
846
+ [42] P. Hasenfratz and F. Niedermayer, Finite size and temperature
847
+ effects in the AF heisenberg model, Z. Phys., B Condens. matter
848
+ 92, 91 (1993).
849
+ [43] A. W. Sandvik and C. J. Hamer, Ground-state parameters,
850
+ finite-size scaling, and low-temperature properties of the two-
851
+ dimensional S = 1
852
+ 2 XY model, Phys. Rev. B 60, 6588 (1999).
853
+ [44] T. Numazawa, K. Kamiya, P. Shirron, M. DiPirro, and K. Mat-
854
+ sumoto, Magnetocaloric effect of polycrystal GdLiF4 for adia-
855
+ batic magnetic refrigeration, AIP Conference Proceedings 850,
856
+ 1579 (2006).
857
+ [45] J. A. Paddison, H. Jacobsen, O. A. Petrenko, M. T. Fernández-
858
+ Díaz, P. P. Deen, and A. L. Goodwin, Hidden order in spin-
859
+ liquid Gd3Ga5O12, Science 350, 179 (2015).
860
+ [46] Y. Tokiwa, T. Radu, C. Geibel, F. Steglich, and P. Gegen-
861
+ wart, Divergence of the Magnetic Grüneisen Ratio at the Field-
862
+ Induced Quantum Critical Point in YbRh2Si2, Phys. Rev. Lett.
863
+ 102, 066401 (2009).
864
+ [47] D. Jang, T. Gruner, A. Steppke, K. Mitsumoto, C. Geibel, and
865
+ M. Brando, Large magnetocaloric effect and adiabatic demag-
866
+ netization refrigeration with YbPt2Sn, Nat. Commun. 6, 8680
867
+ (2015).
868
+ [48] Y. Tokiwa, B. Piening, H. S. Jeevan, S. L. Budko, P. C. Can-
869
+ field, and P. Gegenwart, Super-heavy electron material as metal-
870
+ lic refrigerant for adiabatic demagnetization cooling, Sci. Adv.
871
+ 2, e1600835 (2016).
872
+ [49] P. Gegenwart, Grüneisen parameter studies on heavy fermion
873
+ quantum criticality, Rep. Prog. Phys. 79, 114502 (2016).
874
+ [50] Y. Shimura, K. Watanabe, T. Taniguchi, K. Osato, R. Ya-
875
+ mamoto, Y. Kusanose, K. Umeo, M. Fujita, T. Onimaru, and
876
+ T. Takabatake, Magnetic refrigeration down to 0.2 K by heavy
877
+ fermion metal YbCu4Ni, J. of Appl. Phys. 131, 013903 (2022).
878
+ [51] A. Honecker and S. Wessel, Magnetocaloric effect in quantum
879
+ spin-s chains, Condens. Matter Phys. 12, 399 (2009).
880
+ [52] B. Wolf, Y. Tsui, D. Jaiswal-Nagar, U. Tutsch, A. Honecker,
881
+ K. Removi´c-Langer, G. Hofmann, A. Prokofiev, W. Assmus,
882
+ G. Donath, and M. Lang, Magnetocaloric effect and magnetic
883
+ cooling near a field-induced quantum-critical point, Proc. Natl.
884
+ Acad. Sci. 108, 6862 (2011).
885
+ [53] M. Lang, B. Wolf, A. Honecker, Y. Tsui, D. Jaiswal-Nagar,
886
+ U. Tutsch, G. Hofmann, A. Prokofiev, P. T. Cong, N. Krüger,
887
+ F. Ritter, and W. Assmus, Magnetic cooling through quantum
888
+ criticality, J. Phys.: Conf. Series 400, 032043 (2012).
889
+ [54] S. Bachus, D. A. S. Kaib, Y. Tokiwa, A. Jesche, V. Tsurkan,
890
+ A. Loidl, S. M. Winter, A. A. Tsirlin, R. Valentí, and P. Gegen-
891
+ wart, Thermodynamic perspective on field-induced behavior of
892
+ α RuCl3, Phys. Rev. Lett. 125, 097203 (2020).
893
+ [55] L. J. Zhu, M. Garst, A. Rosch, and Q. M. Si, Universally Di-
894
+ verging Grüneisen Parameter and the Magnetocaloric Effect
895
+ Close to Quantum Critical Points, Phys. Rev. Lett. 91, 066404
896
+ (2003).
897
+ [56] J.-S. Xiang, C. Chen, W. Li, X.-L. Sheng, N. Su, Z.-H. Cheng,
898
+ Q. Chen, and Z.-Y. Chen, Criticality-enhanced magnetocaloric
899
+ effect in quantum spin chain material copper nitrate, Sci. Rep.
900
+
901
+ 7
902
+ 7, 44643 (2017).
903
+ [57] M. Garst and A. Rosch, Sign change of the Grüneisen parameter
904
+ and magnetocaloric effect near quantum critical points, Phys.
905
+ Rev. B 72, 205129 (2005).
906
+ [58] T. Liu, X.-Y. Liu, Y. Gao, H. Jin, J. He, X.-L. Sheng, W. Jin,
907
+ Z. Chen, and W. Li, Significant inverse magnetocaloric effect
908
+ induced by quantum criticality, Phys. Rev. Research 3, 033094
909
+ (2021).
910
+ [59] J. V. José, L. P. Kadanoff, S. Kirkpatrick, and D. R. Nelson,
911
+ Renormalization, vortices, and symmetry-breaking perturba-
912
+ tions in the two-dimensional planar model, Phys. Rev. B 16,
913
+ 1217 (1977).
914
+ [60] R. Moessner, S. L. Sondhi, and P. Chandra, Two-dimensional
915
+ periodic frustrated ising models in a transverse field, Phys. Rev.
916
+ Lett. 84, 4457 (2000).
917
+ [61] R. Moessner and S. L. Sondhi, Ising models of quantum frus-
918
+ tration, Phys. Rev. B 63, 224401 (2001).
919
+ [62] S. V. Isakov and R. Moessner, Interplay of quantum and thermal
920
+ fluctuations in a frustrated magnet, Phys. Rev. B 68, 104409
921
+ (2003).
922
+ [63] S. K. Baek, P. Minnhagen, and B. J. Kim, Kosterlitz-Thouless
923
+ transition of magnetic dipoles on the two-dimensional plane,
924
+ Phys. Rev. B 83, 184409 (2011).
925
+ [64] P. Wikus, E. Canavan, S. T. Heine, K. Matsumoto, and T. Nu-
926
+ mazawa, Magnetocaloric materials and the optimization of
927
+ cooling power density, Cryogenics 62, 150 (2014).
928
+ [65] A. Y. Vasiliev, A. E. Tarkhov, L. I. Menshikov, P. O.
929
+ Fedichev, and U. R. Fischer, Universality of the Berezin-
930
+ skii–Kosterlitz–Thouless type of phase transition in the dipolar
931
+ XY-model, New J. Phys. 16, 053011 (2014).
932
+ [66] S. Guo, A. Ghasemi, C. L. Broholm, and R. J. Cava, Mag-
933
+ netism on ideal triangular lattices in NaBaYb(BO2)2, Phys.
934
+ Rev. Mater. 3, 094404 (2019).
935
+ [67] Y. Tokiwa, S. Bachus, K. Kavita, A. Jesche, A. A. Tsirlin, and
936
+ P. Gegenwart, Frustrated magnet for adiabatic demagnetization
937
+ cooling to milli-kelvin temperatures, Commun. Mater. 2, 42
938
+ (2021).
939
+ [68] A. Jesche, N. Winterhalter-Stocker, F. Hirschberger, A. Bel-
940
+ lon, S. Bachus, Y. Tokiwa, A. A. Tsirlin, and P. Gegen-
941
+ wart, Adiabatic demagnetization cooling well below the mag-
942
+ netic ordering temperature in the triangular antiferromagnet
943
+ KBaGd(BO3)2, arXiv:2212.12483 (2022).
944
+ [69] C. Hagmann and P. Richards, Two-stage magnetic refrigerator
945
+ for astronomical applications with reservoir temperatures above
946
+ 4 K, Cryogenics 34, 221 (1994).
947
+ [70] A. W. Sandvik, Computational studies of quantum spin sys-
948
+ tems, AIP Conf. Proc. 1297, 135 (2010).
949
+ [71] M. Creutz, Overrelaxation and monte carlo simulation, Phys.
950
+ Rev. D 36, 515 (1987).
951
+
952
+ 8
953
+ Intensity (arb. units)
954
+ Intensity (arb. units)
955
+ (003)
956
+ (006)
957
+ (009)
958
+ (0012)
959
+ FIG. S1. (a) shows the powder XRD pattern of KBGB measured at room temperature and corresponding Rietveld refinement. The open circle
960
+ and red solid line represent the observed and calculated intensities, respectively, while the blue solid line shows their difference. The olive
961
+ vertical bars mark the expected reflections for KBGB. (b) Single-crystal XRD scan along the (0,0,L) direction for one representative crystal,
962
+ revealing only peaks that are well indexed by (0,0,3n). The insets show the image of the as-grown KBGB crystals and the rocking-curve
963
+ scan of the (0,0,12) reflection fitted by a Gaussian profile. The very narrow peak width of FWHM = 0.041◦ indicates excellent quality of the
964
+ crystals.
965
+ Supplementary Materials
966
+ Dipolar Spin Liquid Ending with Quantum Critical Point in a Gd-based Triangular Magnet
967
+ Xiang et al.
968
+ Section 1.
969
+ SAMPLE PREPARATION AND STRUCTURE CHARACTERIZATION
970
+ Polycrystalline samples of KBGB were firstly prepared by standard solid-state reaction method as reported in Ref. 40. Sto-
971
+ ichiometric mixtures of K2CO3 (99.99%), BaCO3 (99.95%), H3BO3 (99.99%) and Gd2O3 (99.99%) (with 6% excess H3BO3
972
+ and 5% excess of K2CO3 and BaCO3) were thoroughly ground and pelletized. Then the pellet was placed into an aluminum
973
+ crucible and sintered at 900◦C in air for 10 h. This sintering process was repeated for several times to minimize possible
974
+ impurities.
975
+ Single-crystal samples of KBGB were grown using the flux method as reported in Ref. 41. The pre-obtained polycrystalline
976
+ KBGB with high purity was mixed with the H3BO3 (99.99%) and KF (99.9%) fluxes in a molar ratio of 2:3:[2-3], and thoroughly
977
+ ground. The mixture was transferred into a Pt crucible, heated up to 980◦C in air for 24 h, and then slowly cooled to 790◦C with
978
+ a rate of 2◦C/h. After the furnace cooling, centimeter-sized crystals were obtained on top of the fluxes.
979
+ The phase purity of the polycrystalline KBGB sample was confirmed by powder XRD at room temperature, performed
980
+ on a Bruker D8 ADVANCE diffractometer in Bragg-Brentano geometry with Cu-Kα radiation (λ = 1.5406 Å). As shown in
981
+ Fig. S1(a), the powder XRD pattern can be well fitted with the previously reported trigonal phase of KBGB [40] (a = b =
982
+ 5.4676(1) Å, c = 17.9514(3) Å), without any visible impurity peaks, indicating high purity of the synthesized KBGB powders.
983
+ The quality of the single-crystal KBGB sample was checked by high-resolution synchrotron XRD (λ = 1.54564 Å) measure-
984
+ ments at room temperature, performed on the 1W1A beamline at the Beijing Synchrotron Radiation Facility (BSRF), China. As
985
+ shown in Fig. S1(b), a long L scan, equivalent to a θ-2θ scan with respect to the normal direction of the plate-like KBGB crystal,
986
+ only shows Bragg reflections well indexed by (0, 0, 3n) as expected for the R-3m space group. The peak width (full width at half
987
+ maximum, FWHM) observed in the rocking-curve scan of the (0, 0, 12) peak is very small, 0.041(2)◦, as shown in the inset of
988
+ Fig. S1(b), which suggests excellent crystal quality. KBGB is relatively easy to synthesize and has excellent chemical stability,
989
+ paving its viable way for applications in advanced cryogenics.
990
+
991
+ 9
992
+ 0.0
993
+ 0.5
994
+ 1.0
995
+ 1.5
996
+ 2.0
997
+ 2.5
998
+ 3.0
999
+ 0.1
1000
+ 430 mK
1001
+ 280 mK
1002
+ 195 mK
1003
+ 150 mK
1004
+ 95 mK
1005
+ T (K)
1006
+ B (T)
1007
+ 0.4
1008
+ 0.03
1009
+ 0.06 - 0.09 T·min-1
1010
+ Bi = 3 T
1011
+ 0
1012
+ 1
1013
+ 2
1014
+ 3
1015
+ 4
1016
+ 5
1017
+ 6
1018
+ 0
1019
+ 1
1020
+ 2
1021
+ 3
1022
+ 4
1023
+ KBGB
1024
+ GGG
1025
+ T (K)
1026
+ B (T)
1027
+ 0.15 T·min-1
1028
+ (a)
1029
+ (c)
1030
+ (b)
1031
+ 0.0
1032
+ 0.6
1033
+ 1.2
1034
+ 0.1
1035
+ 0.2
1036
+ 0.3
1037
+ 0.4
1038
+ 0.5
1039
+ 3.0
1040
+ 2.5
1041
+ 2.0
1042
+ 0
1043
+ 1
1044
+ 2
1045
+ 3
1046
+ -1
1047
+ 0
1048
+ 1
1049
+ 2
1050
+ ΓB (T-1)
1051
+ Bc
1052
+ FIG. S2. (a) Illustration of the two-stage quasi-adiabatic demagnetization cooling device for the measurements of 0.5 g KBGB single crystals.
1053
+ (b) shows the measured isentropic curves of KBGB starting from various initial conditions (Ti = 2 K, Bi = 4 T), (2 K, 6 T), and (4 K, 4 T),
1054
+ respectively, where the lowest temperature are found to be significantly lower than those of GGG. The inset zooms in the small-field range
1055
+ (B ≤ 1.2 T). (c) The DR-based measurements with an initial temperature Ti ≤ 430 mK and field Bi = 3 T, where the lowest achieved
1056
+ temperature is Tm ≃ 33 mK. The inset shows the magnetic Grüneisen ratio ΓB deduced from the low-temperature isentropic T-B lines in the
1057
+ main plot, where the sign change is evident and the peak becomes more and more pronounced as the initial temperature Ti lowers.
1058
+ Section 2.
1059
+ MAGNETOTHERMAL MEASUREMENTS
1060
+ Comprehensive magnetothermal measurements were performed on single-crystal samples of KBGB. The low-temperature
1061
+ specific heat (Cp, T ≥ 50 mK) and ac susceptibility (χac, T ≥ 50 mK) measurements were conducted using the Quantum
1062
+ Design Physical Property Measurement System (PPMS) equipped with a 3He–4He dilution refrigerator (DR) insert. The specific
1063
+ heat data were measured under various out-of-plane fields (B//c) with the semi-adiabatic heat pulse method. The phonon
1064
+ contributions are negligible below 2 K as estimated via a Debye T 3 analysis of high-temperature Cp data. The ac susceptibility
1065
+ (χac), as a function of temperature, was measured in zero dc field under different ac frequencies, with the amplitude of the ac
1066
+ excitation field set as 3 Oe. The dc magnetic susceptibility χdc, as a function of temperature down to 0.4 K, was measured
1067
+ using a Quantum Design Magnetic Property Measurement System (MPMS) equipped with a 3He insert. The isothermal dc
1068
+ magnetization curves in the field up to 7 T applied along the a and c axes were measured at 0.4 K with the same setup.
1069
+ Section 3.
1070
+ MAGNETOCALORIC MEASUREMENTS
1071
+ Magnetocaloric effect (MCE) of the frustrated dipolar magnet KBGB was characterized using a homemade setup integrated
1072
+ into the PPMS, for initial temperature 2 K ≤ Ti ≤ 4 K. A DR-based setup is also exploited for MCE measurements with low
1073
+ initial temperature Ti ≤ 500 mK.
1074
+ A.
1075
+ PPMS-based setup for quasi-adiabatic demagnetization measurements
1076
+ As shown in Fig. S2, a homemade PPMS-based construction for quasi-adiabatic demagnetization process is set up, inspired
1077
+ by the Hagmann-Richards design for space applications [69]. An additional guard stage consisting of copper cylinders and
1078
+ Gd3Ga5O12 (GGG) crystals (20 g), a conventional coolant, offer thermal intercepts between the sample stage and the PPMS
1079
+ chamber. In experiments, plate-like KBGB single crystals (with a total mass of 0.5 g) are stacked along the c-axis and fixed
1080
+ on a silver foil by cryogenic glue. To improve the thermal insulation, a Vespel straw is used to support the sample pillar inside
1081
+ the copper cylinder. The guard stage is suppported by PEEK tubes to reduce the thermal exchange with the environment. The
1082
+ electrical connection of the thermometer (a field-calibrated RuO2 chip) on top of the pillar is made by two pairs of twisted
1083
+
1084
+ bnck
1085
+ COWWGLCISI
1086
+ bbW2
1087
+ bEEK
1088
+ Ccc (To a)
1089
+ Ws12
1090
+ N6abel
1091
+ blgid2
1092
+ CCC (To a)
1093
+ KBCB 2luajG
1094
+ p 2.0-10
1095
+ manganese wires (25 µm in diameter and approximately 60 cm in length) to reduce the heat leak. A thermal shield protects
1096
+ the sample from radiant heating and reduce other parasitic heat loads from the PPMS chamber. Demagnetization cooling
1097
+ measurements are performed by gradually decreasing the fields from the initial field Bi at a rate of ˙B = 0.15 T·min−1.
1098
+ The parasitic heat load can be estimated from the temperature change rate of sample after the magnet field being exhausted,
1099
+ i.e., in the hold process with B = 0. To be specific, the parasitic heat load is estimated by ˙Q = C0 ˙T, where C0 is heat capacity
1100
+ of the sample and ˙T is the temperature change rate. For example, when starting from an initial condition of 2 K, it is found
1101
+ that ˙T ≈ 5 × 10−6 K/s. Considering C0 ≃ 0.01 J/K for 0.5 g KBGB samples, we thus figure out the parasitic heat load as
1102
+ ˙Q ≈ 0.05 µW.
1103
+ In Fig. S2 we show the isentropic lines of KBGB obtained through the quasi-adiabatic demagnetization measurements, and
1104
+ make a comparison with the widely used refrigerant GGG. The results with different initial conditions lead to the same conclusion
1105
+ that KBGB clearly outperforms GGG in the lowest cooling temperature.
1106
+ B.
1107
+ The DR-based quasi-adiabatic demagnetization measurements
1108
+ To perform MCE measurements from a lower initial temperature below 500 mK, a standard DR heat capacity sample mount
1109
+ is used, which provides a quasi-adiabatic condition with high vacuum in the 3He–4He dilution insert of PPMS. The thermometer
1110
+ used is a RuO2 semiconductor. It has been carefully calibrated as functions of temperature (50 mK-4 K) and magnetic field
1111
+ (0-5 T), and also extrapolated to 30 mK according to the scaling behavior ln(R − R0) ∼ T −1/4 [67].
1112
+ The polymer strips are used to support the sample platform. A KBGB single crystal with a much smaller mass of 2.3 mg is
1113
+ used here, to avoid large magnetic torque that may break the suspended lines in the sample mount. To decrease the irreversible
1114
+ heating effect on the DR mount, the field sweep rate ˙B has been reduced to 0.06 - 0.09 T·min−1. Due to the small mass of
1115
+ the sample, the parasitic heat loads have a stronger influence in the MCE measurements. However, a prominent dip can still be
1116
+ observed in the quasi-adiabatic cooling curve that clearly signals the existence of a QCP in Fig. S2(c).
1117
+ (a) θ = 0
1118
+ (b) θ = Τ
1119
+ π 3
1120
+ (c) θ = 2 Τ
1121
+ π 3
1122
+ (d) θ = π
1123
+ (e) θ = 4 Τ
1124
+ π 3
1125
+ (f) θ = 5 Τ
1126
+ π 3
1127
+ Q1 = ±b*/2
1128
+ Q3 = ±b*/2
1129
+ Q2 = ±(b*-a*)/2
1130
+ b*
1131
+ a*
1132
+ Q1
1133
+ Q3
1134
+ Q2
1135
+ (g) First Brillouin Zone
1136
+ FIG. S3. (a)-(f) show the magnetic configurations of the stripe order with 6-fold degeneracy, i.e., 6-clock AF, which can be labeled with angle
1137
+ θ (in complex order parameter Ψxy), and also by ordering vector Q1 (blue dots), Q2 (yellow dots), and Q3 (red dots) shown in (g).
1138
+ Section 4.
1139
+ MONTE CARLO SIMULATIONS
1140
+ As the spin quantum number S = 7/2 is large in KBGB, here we use the classical Monte Carlo simulations with standard
1141
+ Metropolis algorithm and single spin update and [70, 71]. The largest system size is 60 × 60, and we calculate the snapshots of
1142
+
1143
+ 11
1144
+ the ground-state spin configurations in Figs. S3(a-f). The corresponding ordering wave vectors Q = ± 1
1145
+ 2a∗, ± 1
1146
+ 2b∗, ± 1
1147
+ 2(a∗ −b∗),
1148
+ with a∗, b∗ are the three ordering wave vectors for the 6-clock AF order shown in Fig. S3(g). The phase angle θ of the complex
1149
+ order parameter Ψxy can only take 6 discretized values that correspond to the 6-fold degenerate ground states (corresponding to
1150
+ Q1, Q2, Q3, see below).
1151
+ In Figs. 1(c-e) of the main text, we show histograms of the complex order parameter Ψxy ≡ meiθ under magnetic field
1152
+ B = 0.68 T and at different temperature, i.e., (c) T = 0.05 K (6-clock AF), (d) T = 0.14 K (DSL), and (e) T = 0.25 K (PM),
1153
+ respectively. To count the histograms, we collect 5 × 106 MC samples on a L = 12 × 12 lattice for statistics.
1154
+ The MC simulation results of specific heat are shown in Fig. S4, where the contour plot in Fig. S4(a) resembles the experi-
1155
+ mental data in Fig. 3(b) of the main text. The round peak in Cm is located at T ∗ ≃ 270 mK, and the peak heights are converged
1156
+ with system sizes, as indicated in the inset of Fig. S4(b). As magnetic fields are applied along the out-of-plane direction, similar
1157
+ to the experiments, we also observe that the Cm peaks move towards low temperature side, with heights lowered, in Fig. S4(c).
1158
+ 0.1
1159
+ 0.2
1160
+ 0.3
1161
+ 0.4
1162
+ 0.5
1163
+ 1
1164
+ 2
1165
+ B= 0 T
1166
+ B= 0.39 T
1167
+ B= 0.68 T
1168
+ Cm
1169
+ T (K)
1170
+ L = 60
1171
+ 0.2
1172
+ 0.4
1173
+ 1
1174
+ 2
1175
+ L=36
1176
+ L=24
1177
+ L=60
1178
+ L=54
1179
+ L=48
1180
+ Cm
1181
+ T (K)
1182
+ 0.0
1183
+ 0.5
1184
+ 1.0
1185
+ 0.0
1186
+ 0.2
1187
+ 0.4
1188
+ 0.6
1189
+ B (T)
1190
+ T(K)
1191
+ 0.4
1192
+ 0.8
1193
+ 1.2
1194
+ 1.6
1195
+ 2
1196
+ (a)
1197
+ (b)
1198
+ (c)
1199
+ Cm
1200
+ 0.25
1201
+ 0.29
1202
+ 1.5
1203
+ 2.5
1204
+ Cm
1205
+ T (K)
1206
+ FIG. S4. The calculated results of specific heat Cm. (a) shows the contour plot of Cm data under out-of-plane field B. The (b) zero-field Cm
1207
+ curves for different system sizes and (c) Cm curves for different fields are also presented. The inset in (b) compares the Cm data near crossover
1208
+ temperature T ∗ ≃ 270 mK. The MC simulations are performed on the HD model [Eq. (1) in the main text] with couplings J = 47 mK and
1209
+ D = 80 mK. (c) compares the specific heat curves under zero and finite magnetic fields.
1210
+ In the simulations, we use the natural unit (J = 1) in the MC calculations and thus the following process is required for
1211
+ comparing the model calculations to experimental data in SI units: (1) We replace the Si operators in Eq. (1) of the main text
1212
+ by classical vectors, Si → Sni ≡ 7/2 ni, where ni is a unit vector; (2) The value of temperature T in natural unit should
1213
+ be multiplied by a factor of J = 0.047 K; (3) Multiply the magnetic field B in natural unit (i.e., B/JS = 1) by a factor of
1214
+ JkB/(gcµB) ≃ 0.028 T.
1215
+
FNE1T4oBgHgl3EQf-gbR/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
FdAzT4oBgHgl3EQfUfww/content/tmp_files/2301.01266v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
FtE3T4oBgHgl3EQftQtg/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a2fb58b792a33a2739fec2086ac1cb3359f8a514e171ed3dd6383e1bf31d7cfe
3
+ size 3604525
GtE1T4oBgHgl3EQf_QYB/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8dfae2eef69dc0b06b146767bf008623e2b6716325a20c92b8fbc668cbe5f8e5
3
+ size 105686
H9E1T4oBgHgl3EQfFgPu/content/tmp_files/2301.02904v1.pdf.txt ADDED
@@ -0,0 +1,1040 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Sensitivity analysis for transportability in multi-study,
2
+ multi-outcome settings
3
+ Ngoc Q. Duong, Amy J. Pitts, Soohyun Kim, Caleb H. Miles
4
+ Department of Biostatistics, Mailman School of Public Health, Columbia University
5
+ Abstract
6
+ Existing work in data fusion has covered identification of causal estimands when
7
+ integrating data from heterogeneous sources. These results typically require additional
8
+ assumptions to make valid estimation and inference. However, there is little literature
9
+ on transporting and generalizing causal effects in multiple-outcome setting, where the
10
+ primary outcome is systematically missing on the study level but for which other
11
+ outcome variables may serve as proxies. We review an identification result developed
12
+ in ongoing work that utilizes information from these proxies to obtain more efficient
13
+ estimators and the corresponding key identification assumption. We then introduce
14
+ methods for assessing the sensitivity of this approach to the identification assumption.
15
+ Keywords: Causal inference, Data fusion, External validity, Generalizability, Missing data,
16
+ Proxy variable
17
+ arXiv:2301.02904v1 [stat.ME] 7 Jan 2023
18
+
19
+ 1
20
+ Introduction
21
+ Research in clinical medicine and public health is often concerned with estimating the effect
22
+ of some treatment in a specific target population. However, even in a randomized clinical
23
+ trial, which is considered the gold-standard study design, ensuring external validity remains
24
+ a challenge. This can be due to a variety of reasons, including non-random sampling, overly
25
+ stringent exclusion criteria, or an ill-defined target population of interest (Tan et al., 2022;
26
+ Kennedy-Martin et al., 2015).
27
+ Meta-analysis of summary statistics is a commonly used
28
+ tool to synthesize and generalize findings from published study-level summary statistics,
29
+ but tends to rely on strong, often implausible assumptions. An alternative approach that
30
+ allows for more control over the nuances and heterogeneity across studies is to combine
31
+ individual-level data, when available, from multiple studies, each of which may contain
32
+ insufficient information to address a given scientific question by itself, but which collectively
33
+ have the power to do so. There has been a growing body of work on generalizability and
34
+ transportability methods, which can help address the problem of external validity of the
35
+ effect estimates from integrating individual level data across studies.
36
+ Generalizability concerns the setting where the study population is a subset of the target
37
+ population of interest while transportability addresses the setting where the study popula-
38
+ tion is partially or completely external to the target population (Degtiar and Rose, 2023).
39
+ Specifically, generalizability typically involves extending the causal effect estimate derived
40
+ from a study as long as the covariates in the study population and the target population
41
+ have common support (Gechter, 2015; Tipton, 2014). On the other hand, transportability
42
+ entails extrapolating the effect estimated from a study in which some primary outcome of
43
+ interest is observed to a population represented by a sample in which the outcome is not
44
+ measured.
45
+ Existing methodologies involve directly transporting some estimated causal effect, e.g.,
46
+ the average treatmemt effect (ATE), from studies where the outcomes are observed to other
47
+ studies with missing outcomes or across heterogeneous study designs and settings (Barein-
48
+ 1
49
+
50
+ boim and Pearl, 2016; Dong et al., 2020; Pearl and Bareinboim, 2014; H¨unermund and
51
+ Bareinboim, 2019), or to some broader target population (Dahabreh et al., 2020a,b; Lesko
52
+ et al., 2017; Westreich et al., 2017). When considering multiple studies, it is often the case
53
+ that one will observe different outcomes at follow up. However, existing methods do not
54
+ take advantage of these other potentially correlated and informative outcome variables mea-
55
+ sured at follow-up, which could potentially be leveraged to achieve large efficiency gains.
56
+ Existing outcome proxy-blind methods typically rely on an assumption of homogeneous con-
57
+ ditional potential outcome means for valid transportation of estimation from one population
58
+ to another. Sensitivity analysis strategies have been proposed to study the extent to which
59
+ the violation of these assumptions will affect the estimations and inferences drawn (Nguyen
60
+ et al., 2017; Dahabreh and Hern´an, 2019; Dahabreh et al., 2022).
61
+ In ongoing work, we have developed a new strategy to more efficiently estimate the
62
+ ATE from integrated data across multi-outcome studies, with inconsistent availability of the
63
+ primary outcome of interest at the study level. The proposed methodology takes advantage
64
+ of the availability of follow-up measurements of potential correlates of the main outcome
65
+ to yield more precise estimate of the causal effects.
66
+ In this article, we consider the key
67
+ common outcome regression (or conditional exchangeability for study selection) assumption
68
+ for transportability while leveraging these outcome proxies, which differs slightly from the
69
+ common outcome regression assumption that has been traditionally used for transportability.
70
+ We discuss the resulting bias when this assumption is not met, and develop methodology for
71
+ sensitivity analysis to the violation of this assumption.
72
+ The remainder of the article is organized as follows. In Section 2, we discuss identification
73
+ of the average treatment effect in the multi-study, multi-outcome setting. In Section 3, we
74
+ discuss the bias incurred by violations of the key conditional exchangeability assumption. In
75
+ Section 4, we compare the conditional exchangeability assumption in our setting with that
76
+ used in settings that do not leverage outcome proxies. In Section 5, we develop methods
77
+ for sensitivity analysis for when our assumption is violated. We demonstrate the empirical
78
+ 2
79
+
80
+ performance of our proposed methods in a simulation study in Section 6, and conclude with
81
+ a discussion in Section 7.
82
+ 2
83
+ Data integration for studies with primary outcome
84
+ missing systematically
85
+ 2.1
86
+ Study and data setting
87
+ In this setting, we let A be the treatment indicator, W be a set of covariates that are
88
+ commonly observed across studies, Y be the primary outcome variable, the set {T1, . . . , Tk}
89
+ be all the potential outcome proxies measured at follow-up in any study, and Js be the
90
+ study-specific subset of {T1, . . . , Tk} that is measured in study s.
91
+ Suppose there are S
92
+ studies that are ordered such that for each s in the first s∗ studies, we observe the set of
93
+ variables (Y, A, Js, W), while for each s in the remaining S − s∗ studies, only the subset
94
+ (A, Js, W) are observed. In other words, Y is systematically missing in the latter set of
95
+ studies. Unlike the standard setup in other works concerning effect transportability that
96
+ only involves (Y, A, W), we introduced the use of Ts, where Ts ⊂ Js is some user-specified
97
+ subset of Js for each study s. Ts could be chosen based on availability and subject matter
98
+ knowledge and must be chosen such that they are observed in at least one of the studies
99
+ {1, 2, . . . , s∗}.
100
+ Studies can be randomized experiments or observational; however, we will not consider
101
+ scenarios in which some studies are randomized experiments and others are observational in
102
+ this work. Then the study-specific average treatment effect and conditional average treat-
103
+ ment effect can be written as:
104
+ ATE(s) = E(Y1 − Y0 | S = s)
105
+ CATE(w, s) = E(Y1 − Y0 | W = w, S = s).
106
+ 3
107
+
108
+ Accordingly, we can define the overall average treatment effect and conditional average treat-
109
+ ment effect as:
110
+ ATE =
111
+ S
112
+
113
+ s=1
114
+ ATE(s)
115
+ CATE(w) = E(Y1 − Y0 | W = w)
116
+ where the weights can be user-specified such that �
117
+ s πs = 1. For instance, one can choose
118
+ πs = P(S = s), or the marginal probability of being in each study. Alternatively, we could
119
+ define ATE = EQW,SCATE(W, S) for a user-specified, known distribution QW,S of W and
120
+ S.
121
+ Since Y is not measured in s ∈ {s∗ +1, . . . , S}, we cannot directly estimate the ATE and
122
+ CATE using data from these studies alone. Our purpose is to transport the ATE from the
123
+ first s∗ studies where Y is observed, to the remaining S − s∗ studies while also leveraging
124
+ the information from the outcome proxy set Ts to improve efficiency. For ease of notation,
125
+ let σs be a subset of the first s∗ studies in which both Y and Ts are observed. We can then
126
+ use this information from the studies that form σs to estimate the outcome regression that
127
+ will allow us to transport the causal effects to study s. In this setting, we have shown in
128
+ ongoing, not-yet-published work that the ATE can be nonparametrically identified as:
129
+ ΨATE =
130
+ s∗
131
+
132
+ s=1
133
+ πsE{E(Y | W, A = 1, S = s) − E(Y | W, A = 0, S = s) | S = s}
134
+ +
135
+ S
136
+
137
+ s=s∗+1
138
+ πsE[E{E(Y | Ts, W, A = 1, S ∈ σs) | W, A = 1, S = s}
139
+ (1)
140
+ − E{E(Y | Ts, W, A = 0, S ∈ σs) | W, A = 0, S = s} | S = s].
141
+ The terms in the first sum are simply the standard identification formula for the (study-
142
+ specific) average treatment effects when Y is observed. The second sum is identified since
143
+ it only depends on the distribution of Y in the studies in σs, i.e., in which Y is actually
144
+ 4
145
+
146
+ observed.
147
+ Here, we introduced a modification to how transportability has traditionally been done by
148
+ incorporating information from a set of outcomes measured at follow-up that are correlated
149
+ with the main outcome of interest.
150
+ 2.2
151
+ Assumptions for Identification of the ATE
152
+ This derivation ATE can be nonparametrically identified given the assumptions that are
153
+ standard for identification for ATE when outcomes are all observed:
154
+ Assumption 1 (Positivity). P(A = 1 | W = w) > 0 for all w with positive probability.
155
+ Assumption 2 (Consistency). Y = AY1 + (1 − A)Y0.
156
+ Assumption 3 (Within-study conditional exchangeability).
157
+ E[Y a | W, A, S = s] = E[Y a | W, S = s] for all s.
158
+ The validity of our estimator relies on a fourth assumption that allows for the transporta-
159
+ tion of the effect across studies:
160
+ Assumption 4 (Common outcome regression (proxy-aware version)).
161
+ E(Y | Ts, W, A = a, S = s) = E(Y | Ts, W, A = a, S ∈ σs) for all s.
162
+ This is a missing at random (MAR)-type assumption, where S can in a sense be thought
163
+ of as a missingness indicator, since missingness is systematic by study.
164
+ We can also introduce a fifth assumption that is not necessary for identification, but
165
+ allows for more borrowing of information across studies, which can help with efficiency:
166
+ Assumption 5 (Common distribution of outcome proxies). Ts ⊥ S | W, A for all s.
167
+ 5
168
+
169
+ This implies the distribution of Ts conditional on treatment assignment and baseline
170
+ covariates is the same across studies. Under this additional assumption, the identification
171
+ result simplifies to:
172
+ ATE =
173
+ S
174
+
175
+ s=1
176
+ πsE[E{E(Y | Ts, W, A = 1, S ∈ σs) | W, A = 1}
177
+ − E{E(Y | Ts, W, A = 0, S ∈ σs) | W, A = 0} | S = s].
178
+ In ongoing work, we have developed a simple substitution estimator that involves replac-
179
+ ing each expectation with a regression-based estimate and the outer expectation with an
180
+ empirical mean.
181
+ For the outcome proxy-blind approach, in addition to the first three standard internal
182
+ validity assumptions, Assumption 4 is replaced by a slightly different mean outcome ex-
183
+ changeability assumption: across studies assumption (exchangeability over S) (Dahabreh
184
+ and Hern´an, 2019; Lesko et al., 2017):
185
+ Assumption 6 (Common outcome regression (proxy-blind version)).
186
+ E(Y | W, A = a, S = s) = E(Y | W, A = a, S ∈ σs) for all s.
187
+ Assumption 4 differs from Assumption 6 by additionally conditioning on Ts for each study
188
+ s. Assumptions 4 and 5 together imply Assumption 6. In this article, we will only consider
189
+ sensitivity analysis for the violation of Assumption 4.
190
+ When Assumption 5 is violated,
191
+ the ATE estimator based on Assumption 4 (i.e., the substitution estimator based on the
192
+ identification formula (1)) will remain consistent.
193
+ 6
194
+
195
+ 3
196
+ Characterizing the bias resulting from violation of
197
+ the identification assumption
198
+ The validity of ΨATE is dependent on the key assumption 4. This assumption requires no
199
+ heterogeneity in the conditional outcome means given treatment, covariates, and outcomes
200
+ proxies between studies with and without missing outcome (Y ) data. This allows for trans-
201
+ portation of the conditional outcome means, and correspondingly, the ATE and CATE,
202
+ estimable from one study to others.
203
+ In practice, this could be a strong assumption to make while also untestable using ob-
204
+ served data. For instance, in previous unpublished work, we estimated the average treatment
205
+ effect of cognitive remediation (CR) therapy on Social Behavioral Scale (SBS) score, a mea-
206
+ sure for social functioning, using harmonized data from three trials in the NIMH Database
207
+ of Cognitive Training and Remediation Studies (DoCTRS) database. However, the degree
208
+ of effectiveness of CR, especially on functional and occupational outcomes, was less evident
209
+ and has been suggested to vary depending on the setting in which the treatment was admin-
210
+ istered (Barlati et al., 2013; Combs et al., 2008; McGurk et al., 2007; Wykes et al., 2007,
211
+ 2011). When this assumption is violated, the substitution estimators described in the pre-
212
+ vious section will be biased. Therefore, we examine two strategies for sensitivity analysis in
213
+ order to examine the robustness of estimates under varying degrees of assumption violation.
214
+ To quantify the degree of violation, let the bias functions be defined as:
215
+ u(A = 1, Ts, W) = E(Y | Ts, W, A = 1, S = s) − E(Y | Ts, W, A = 1, S ∈ σs),
216
+ u(A = 0, Ts, W) = E(Y | Ts, W, A = 0, S = s) − E(Y | Ts, W, A = 0, S ∈ σs)
217
+ (2)
218
+ 7
219
+
220
+ Then, equation (1) when assumption 4 is violated instead becomes:
221
+ ATE =
222
+ s∗
223
+
224
+ s=1
225
+ πsE{E(Y | W, A = 1, S = s) − E(Y | W, A = 0, S = s) | S = s)
226
+ +
227
+ S
228
+
229
+ s=s∗+1
230
+ πsE [E {E (Y | Ts, W, A = 1, S ∈ σs) | W, A = 1, S = s)}
231
+ −E {E (Y | Ts, W, A = 0, S ∈ σs) | W, A = 0, S = s)} | S = s]
232
+ +
233
+ S
234
+
235
+ s=s∗+1
236
+ πsE[E {u (A = 1, Ts, W) | W, A = 1, S = s}
237
+ − E {u (A = 0, TS, W) | W, A = 0, S = s} | S = s],
238
+ where the last sum is not identified. Then, the study-specific bias for study s is:
239
+ E [E {u (A = 1, Ts, W) | W, A = 1, S = s} − E {u (A = 0, Ts, W) | W, A = 0, S = s} | S = s]
240
+ = E[δ∗(W)|S = s].
241
+ (3)
242
+ By rearranging terms, δ∗(W) can be alternatively written as:
243
+ E [E (Y | Ts, W, A = 1, S = s) − E (Y | TS, W, A = 1, s ∈ σs) | W, A = 1, S = s]
244
+ − E [E (Y | Ts, W, A = 0, S = s) − E (Y | Ts, W, A = 0, s ∈ σs) | W, A = 0, S = s]
245
+ = E(Y | W, A = 1, S = s) − E(Y | W, A = 0, S = s)
246
+ − {E [E (Y | Ts, W, A = 1, s ∈ σs) | W, A = 1, S = s]
247
+ − E [E (Y | Ts, W, A = 0, s ∈ σs) | W, A = 0, S = s]}.
248
+ (4)
249
+ The latter term cannot be simplified unless Assumption 5 holds.
250
+ 8
251
+
252
+ 4
253
+ Comparison with bias functions in settings without
254
+ incorporation of follow-up surrogate outcomes
255
+ In recent work, Dahabreh and Hern´an (2019) developed sensitivity analysis for transportabil-
256
+ ity considering a similar setting of two types of studies with and without missing outcomes.
257
+ In the base case, there are two studies considered (missingness of the outcome variable de-
258
+ noted by a binary indicator S). To describe this setting using our notation, we simply have
259
+ σ0 = σ1 = {1} (i.e., study S = 1 with the observed outcome of interest is used to impute
260
+ the conditional outcome means for study S = 0). Equivalently, for ease of interpretation in
261
+ the base case, let S = 1 and S = 0 denote the study where the primary outcome of interest
262
+ is observed and not observed, respectively.
263
+ In the setting where the model used to impute conditional potential outcomes does not
264
+ utilize information from Ts, Dahabreh and Hern´an (2019) define:
265
+ u(A = a, W) = E[Y | A = a, W, S = 1] − E[Y | A = a, W, S = 0].
266
+ The difference between these bias functions can then be obtained as:
267
+ δ(W) = u(A = 1, W) − u(A = 0, W)
268
+ = E[Y 1 − Y 0 | W, S = 1] − E[Y 1 − Y 0 | W, S = 0]
269
+ This expression can be qualitatively expressed as the difference in the conditional average
270
+ treatment effects between the two studies. This qualitative interpretation can aid in concep-
271
+ tualizing and thinking about more appropriate values and range for sensitivity parameters
272
+ when examining robustness of the results. More specifically, assuming higher levels of the
273
+ outcome are preferred, if we believe the participants in studies with missing outcomes benefit
274
+ less from treatment, then true δ can be assumed to be positive and vice versa (Dahabreh and
275
+ Hern´an, 2019). Since our bias functions are conditional on the set of proxy outcomes, the
276
+ 9
277
+
278
+ term δ∗(W) in (4) unfortunately cannot be reduced further to a more interpretable statistical
279
+ entity. When we take Ts to be the empty set, the bias function δ∗(W) reduces to the same
280
+ expression.
281
+ 5
282
+ Accounting for violation of the common outcome re-
283
+ gression assumption through sensitivity analyses
284
+ We consider two scenarios in which we assume the bias terms u(A = 1, Ts, W) and u(A =
285
+ 0, Ts, W) to be 1) constants and 2) bounded functions of the outcome proxies and/or baseline
286
+ covariates. The first scenario involves making a stronger assumption about the bias terms.
287
+ On the other hand, the second scenario requires weaker assumptions but allow them to be
288
+ non-constant.
289
+ 5.1
290
+ Bias functions assumed to be some fixed values
291
+ Although it might be more reasonable to assume that the bias functions are dependent on
292
+ some baseline covariates, for ease of implementation of sensitivity analysis, one can also
293
+ suppose they are constant. When u(A = 1, Ts, W) and u(A = 0, Ts, W) are independent of
294
+ the baseline covariates W and the outcome proxy set Ts, the conditional expectations of the
295
+ bias functions, and in turn, the term δ∗(W) in (3), reduce to:
296
+ δ = u1 − u0, where δ, u1, and u0 ∈ R
297
+ (5)
298
+ The sensitivity analysis involves correcting for the above-mentioned bias term by adding it
299
+ back to the identification formula ΨATE, which relies on the common outcome regression
300
+ assumption.
301
+ 10
302
+
303
+ ATE =
304
+ s∗
305
+
306
+ s=1
307
+ πsE {E (Y | W, A = 1, S ∈ σs) − E (Y | W, A = 0, S ∈ σs) | S = s}
308
+ +
309
+ S
310
+
311
+ s=s∗+1
312
+ πsE [E {E (Y | Ts, W, A = 1, S ∈ σs) | W, A = 1, S = s}
313
+ − E {E (Y | Ts, W, A = 0, S ∈ σs) | W, A = 0, S = s} | S = s] +
314
+ S
315
+
316
+ s=s∗+1
317
+ πs (u1 − u0)
318
+ =ΨATE +
319
+ S
320
+
321
+ s=s∗+1
322
+ πs (u1 − u0)
323
+ (6)
324
+ where u1 and u0 are scalars.
325
+ In practice, the true bias term would be unknown. Thus, one strategy is to propose a
326
+ grid of sensitivity parameters that covers the potential range of values in which the true bias
327
+ term might fall. This grid of sensitivity parameters can be specified using subject-matter
328
+ knowledge. We can then adjust for the bias term in the estimation step by adding back
329
+ the different sensitivity parameters to the estimated ATE using our proposed method. This
330
+ also allows for observation of the behavior of the estimated ATE as we vary the sensitivity
331
+ parameters.
332
+ 5.2
333
+ Bounded covariate-dependent bias functions
334
+ One might also believe that the bias term is not constant at all levels of the baseline covariates
335
+ and/or the outcome proxies. When the assumption of fixed-value bias terms is considered
336
+ too strong, but the functional forms for bias terms cannot be confidently determined from
337
+ existing knowledge of the data mechanism (as will typically be the case), one can still recover
338
+ some information about the true ATE without having to correctly specify the bias terms. If
339
+ we instead assume the bias terms to be some bounded functions, we can compute a bound
340
+ around the (na¨ıve) ATE estimate that contains the true ATE by varying the bounds of these
341
+ functions. This provides information on how far away the true ATE can be from the estimate
342
+ 11
343
+
344
+ obtained constrained by the bounds of the bias term.
345
+ Identifying the bounds for the bias term can be expressed as maximizing and minimizing
346
+ the objective function:
347
+ E[E[u(A = 1, Ts, W) | W, A = 1, S = s] − E[u(A = 0, Ts, W) | W, A = 0, S = s] | S = s]
348
+ subject to the following constraints:
349
+ |u(A = 1, Ts = ts, W = w)| ≤ γ1
350
+ |u(A = 0, Ts = ts, W = w)| ≤ γ0
351
+ for all ts and w, which implies |E[u(A = 1, Ts, W) | W, A = 1, S = s]| ≤ γ1 and |E[u(A =
352
+ 0, Ts, W) | W, A = 0, S = s]| ≤ γ0 where γ1, γ1 ∈ R+.
353
+ Then we have −(γ1 + γ0) ≤ u(A = 1, Ts, W) − u(A = 0, Ts, W) ≤ γ1 + γ0. If we have no
354
+ reason to suspect we know more about the bounds of one bias function than the other (as
355
+ will typically be the case), we may simply choose to specify a scalar sensitivity parameter γ
356
+ to be the maximum of γ1 and γ2, in which case we have −2γ ≤ u(A = 1, Ts, W) − u(A =
357
+ 0, Ts, W) ≤ 2γ.
358
+ By equation (6) even though we do not know the form of the bias functions u(A =
359
+ 1, Ts, W) and u(A = 0, Ts, W), we can partially recover the true ATE using the bounds
360
+ around the na¨ıve estimate:
361
+ ΨATE − 2 max(γ1, γ0) ≤ ATE ≤ ΨATE + 2 max(γ1, γ0)
362
+ ΨATE − 2γ ≤ ATE ≤ ΨATE + 2γ
363
+ (7)
364
+ If the bias functions are in fact bounded by some value smaller than or equal to our specified
365
+ values for the sensitivity bounds, the true ATE would fall between [ΨATE − 2γ, ΨATE + 2γ].
366
+ Then, the true ATE is partially identified without assumptions about the functional form
367
+ 12
368
+
369
+ of u(A = 1, Ts, W) and u(A = 0, Ts, W). One can then use the bootstrap standard error for
370
+ the substitution estimator of the identification formula (1) to determine the amount to add
371
+ and subtract from the upper and lower bounds, respectively, in order to produce confidence
372
+ intervals for the partial identification sets for each value of the sensitivity parameter. Since
373
+ the sensitivity bounds are a deterministic function of the sensitivity parameter, bootstrapping
374
+ need only be done once.
375
+ 6
376
+ Simulations
377
+ 6.1
378
+ Data generating mechanism
379
+ We consider the setting of two studies, with S = 1 indicating the study where the primary
380
+ outcome is available.
381
+ We generate random sample draws with sample size n = 100 for
382
+ both studies. The data generating mechanism is as follows. W, T0 come from independent
383
+ standard normal distributions, and T1 comes from a normal distribution with mean and
384
+ variance of 1. Then
385
+ T = I(A = 1) × T1 + I(A = 0) × T0
386
+ Y 0 = −4T0 + W + ϵ0
387
+ Y 1 = 4T1 + W + ϵ1
388
+ Y = I(A = 1) × Y1 + I(A = 0) × Y0
389
+ where ϵ1, ϵ0 ∼ N(0, 1).
390
+ Via these specifications, T fully mediates the relationship between A and Y (direct effect
391
+ from A to Y is constrained to be 0). As a result, the true ATE = 4. This is also a more
392
+ basic setting in which the vector T is observed in all studies.
393
+ Due to the nature of the DoCTRS database, which is comprised of randomized clinical
394
+ trials, in our base setting, we specified the marginal probability P(A = 1) = 0.5, represent-
395
+ 13
396
+
397
+ ing random treatment assignment. This treatment assignment satisfies the positivity and
398
+ exchangeability assumption.
399
+ Specifically, to incorporate the difference in conditional outcome means between the
400
+ two types of studies, in studies missing the outcome, we added constant bias terms to
401
+ the counterfactual outcomes Y0 and Y1. Similar to the data generating step, we preserved
402
+ the observed counterfactual outcome from the corresponding treatment assignment, which
403
+ satisfies the consistency assumption. By (5), we have:
404
+ Y 0
405
+ S=1 = Y 0
406
+ S=0 + u0
407
+ Y 1
408
+ S=1 = Y 1
409
+ S=0 + u0 + δ
410
+ (8)
411
+ for u0 ∈ {−3, 0, 3}, δ ∈ {−2, 0, 2}.
412
+ Then the bias reduces to a single parameter δ, since it is no longer a function of u0 when
413
+ computing the ATE:
414
+ E(Y 1 − Y 0 | S = 1) = E(Y 1 − Y 0 | S = 0) + δ
415
+ (9)
416
+ In the case where the bias term is a function of baseline covariates and surrogate outcome,
417
+ we had the following specification for the true bias:
418
+ u0 = b0 × sin (Ts + W)
419
+ u1 = b1 ×
420
+ exp(Ts+W)
421
+ 1+exp(Ts+W)
422
+ for b0 ∈ {2, 3, 4} and b1 ∈ {1, 2, 3}.
423
+ 6.2
424
+ Adjusting for sensitivity parameter in estimation step
425
+ In the presence of non-zero bias, when the value of the sensitivity parameter δ is specified
426
+ such that it is equal to true δ, the ATE estimate after bias adjustment tends to be closer to
427
+ 14
428
+
429
+ the true ATE after compared to before. In addition, the corresponding 95% CIs are expected
430
+ to cover the true ATE 95% of the times. Although coverage probability can be examined
431
+ more in a more robust fashion using bootstrapped confidence intervals across all simulations,
432
+ in Fig. 1, 2, and A.1-A.4, the 95% CIs covers the true ATE at the value of the sensitivity
433
+ parameter that reflects the degree of assumption violation all but one instance, which is in
434
+ line with our expectations.
435
+ Scenario 1. When the bias terms are assumed to be constants, a natural approach
436
+ would be to specify a two-dimensional grid of sensitivity parameters for both scalars u0 and
437
+ u1. However, by (8), it is equivalent to specifying u0 (or u1) and δ. In fact, since the u0 (or
438
+ u1) as constant terms cancel out during adjustment, it is sufficient to specify one sensitivity
439
+ parameter δ (9). We also note that δ being 0 does not necessarily imply assumption 4 is
440
+ met, since the bias terms u0 and u1 could cancel exactly.
441
+ To implement sensitivity analysis, we follow the steps:
442
+ 1. Specify a grid of sensitivity parameters δ.
443
+ The grid should be reasonably wide to
444
+ contain true δ.
445
+ 2. Estimate the na¨ıvely transported ATE using the identification result in (1)
446
+ 3. Sequentially add the values in the sensitivity parameter grid to the na¨ıvely estimated
447
+ ATE, using the result in (6) to obtain the bias-corrected ATE estimates.
448
+ We then plotted the bias-corrected estimates under different sensitivity parameters against
449
+ the true ATE. Additionally, we bootstrapped the bias-corrected estimates to obtain the 95%
450
+ confidence intervals and explore coverage across different values of u0 and δ.
451
+ Scenario 2. When we want to make minimal assumptions about the functional form of
452
+ the bias, we can still perform sensitivity analysis on the true ATE using the following steps:
453
+ 1. Specify a grid of sensitivity parameters called γ that potentially include the upper and
454
+ lower bounds of the true bias functions
455
+ 15
456
+
457
+ 2. Computed the “na¨ıve” ATE estimate using the identification result in (1)
458
+ 3. Construct the upper and lower bound around the estimated ATE using (7) where γ is
459
+ replaced with the sensitivity parameters.
460
+ We also plot the na¨ıve ATE estimates and the bounds around these estimates at each value
461
+ of the sensitivity parameters. In practice, the bias functions are of course unknown and
462
+ cannot be estimated from observed data. Therefore, when specifying the grid of sensitivity
463
+ parameters, the analyst needs to employ subject matter knowledge about the data generating
464
+ mechanism to select values of δ and γ.
465
+ We then explore the behavior of the adjusted estimators via simulations. In the first case,
466
+ we focused on the general unbiasedness of the correctly-adjusted point estimate for both the
467
+ overall ATE and ATE among studies with missing outcomes, as well as the 95% CI coverage
468
+ across degrees of assumption violation (i.e., across values of true u0 and δ). In the second
469
+ case, we looked for correct bounding of the true ATE.
470
+ 6.3
471
+ Simulation Results
472
+ 6.3.1
473
+ Bias terms as constants
474
+ We examine the estimates produced by our method under the different degrees of violation
475
+ of assumption 4, before and after taking into account the specified sensitivity parameter.
476
+ Figure 1 shows the estimates (95% CI) for the true overall ATE using our method under
477
+ varying magnitudes and directions of the bias terms from one single simulation.
478
+ 16
479
+
480
+ Figure 1: Sensitivity-parameter-adjusted ATE estimate shown against the true overall ATE
481
+ across values of the true bias and sensitivity parameter; n=100 for each study, 95% CI
482
+ constructed from 1000 bootstrap samples. When sensitivity parameter δ = 0, the adjusted
483
+ estimate corresponds to the unadjusted estimate. Horizontal dotted line shows the true ATE
484
+ given true δ; vertical dotted line indicates sensitivity parameter δ equals true δ
485
+ In the presence of non-zero bias, when the value of the sensitivity parameter δ is specified
486
+ such that it is equal to true δ, the ATE estimate after bias adjustment tends to be closer to
487
+ the true ATE after compared to before. In addition, the corresponding 95% CIs are expected
488
+ to cover the true ATE 95% of the times. Although coverage probability can be examined
489
+ more in a more robust fashion using bootstrapped confidence intervals across all simulations,
490
+ in Fig. 1, 2, and A.1-A.4, the 95% CIs covers the true ATE at the value of the sensitivity
491
+ parameter that reflects the degree of assumption violation all but one instance, which is in
492
+ 17
493
+
494
+ True = -2
495
+ True = 0
496
+ True θ = 2
497
+ UU
498
+ 2
499
+ Estimate
500
+ 4
501
+ JU
502
+ -2
503
+ 2
504
+ Sensitivity parameterline with our expectations.
505
+ Figure 2:
506
+ Sensitivity-parameter-adjusted ATE estimates shown against the true study-
507
+ specific ATE in the study in which the outcome is unobserved across values of the true
508
+ bias and sensitivity parameter; n=100 for each study, 95% CI constructed from 1000 boot-
509
+ strap samples. When sensitivity parameter δ = 0, the adjusted estimate corresponds to the
510
+ unadjusted estimate. Horizontal dotted line shows the true study-specific ATE given true δ;
511
+ vertical dotted line indicates sensitivity parameter δ equals true δ
512
+ Figure 2 shows similar results for the study-specific ATE estimates in the study with
513
+ missing outcomes (before and after bias adjustment) from the same simulated data. Com-
514
+ pared to the results in Figure 1, after adjustment using the correct sensitivity parameters,
515
+ the 95% CIs contain the true ATE more frequently than the CIs of the unadjusted estimates
516
+ in the study with missing primary outcome. Figure 2 also shows an example where infer-
517
+ 18
518
+
519
+ True = -2
520
+ True = 0
521
+ True = 2
522
+ 10.0 -
523
+ 7.5
524
+ 5.0
525
+ uo
526
+ =
527
+ 2.5
528
+ -3
529
+ 0.0
530
+ 10.0
531
+ 7.5
532
+ Estimate
533
+ rue
534
+ 5.0
535
+ uo
536
+ 2.5
537
+ II
538
+ -
539
+ 0.0
540
+ 7.5
541
+ 5.0 -
542
+ 2.5
543
+ II
544
+ 3
545
+ 0.0
546
+ -2.5
547
+ 2
548
+ 0
549
+ 2
550
+ 2
551
+ 0
552
+ 2
553
+ 2
554
+ Sensitivity parameterence is sensitive to the violation of our assumption at a magnitude of δ between -1 and -2
555
+ (u0 = −3, bottom left panel), between which the 95% CI changes from not containing to
556
+ zero to containing zero.
557
+ When we increased the sample size (n=200 and n=500), we saw general reductions in the
558
+ errors of these single estimates (Figures A.1, A.3). In most cases, even when there is error
559
+ in the adjusted estimates, the 95% CI bootstrap confidence intervals provide good coverage
560
+ (Figures 1, A.1, A.3). The reduction in error and improved coverage are more pronounced
561
+ when estimating the study-specific effect in the study with missing outcomes than in the
562
+ overall ATE combining the two studies (Figures A.2, A.4).
563
+ We also ran 1000 simulations under the same data generating mechanism and obtained
564
+ the unadjusted and sensitivity-parameter-adjusted estimates for each simulation. We then
565
+ showed the mean and 2.5th and 97.5th quantiles of these estimates under each combination
566
+ of the true bias values. We can see that when averaged across 1000 simulations, the adjusted
567
+ estimates closely approximate the true ATE (Figures 3, 4) when the true value of δ is used
568
+ for the sensitivity parameter.
569
+ 19
570
+
571
+ Figure 3: Sensitivity-parameter-adjusted ATE estimates shown against the true overall ATE
572
+ across values of the true bias sensitivity parameter; mean, 2.5th and 97.5th quantiles obtained
573
+ from 1000 simulations. When sensitivity parameter δ = 0, the adjusted estimate corresponds
574
+ to the unadjusted estimate. Horizontal dotted line shows the true overall ATE given true δ;
575
+ vertical dotted line indicates sensitivity parameter δ equals true δ
576
+ 20
577
+
578
+ True = -2
579
+ True = 0
580
+ True = 2
581
+ 6
582
+ 5
583
+ .
584
+ True
585
+ 4
586
+ uO = 3
587
+ 2
588
+ 6
589
+ Estimate
590
+ 150
591
+ True uO = 0
592
+ 4
593
+ 2
594
+ 6
595
+ 5
596
+ .
597
+ .
598
+ True uO = 3
599
+ 4
600
+ 2
601
+ 2
602
+ 1
603
+ 0
604
+ 2
605
+ -2
606
+ -1
607
+ 0
608
+ "
609
+ 2
610
+ -2
611
+ 1
612
+ 0
613
+ 2
614
+ Sensitivity parameter Figure 4: Sensitivity-parameter-adjusted ATE estimates shown against the true ATE in the
615
+ study with missing outcome across values of the true bias and sensitivity parameter; mean,
616
+ 2.5th and 97.5th quantiles obtained from 1000 simulations. When sensitivity parameter δ
617
+ = 0, the adjusted estimate corresponds to the unadjusted estimate. Horizontal dotted line
618
+ shows the true study-specific ATE given true δ; vertical dotted line indicates sensitivity
619
+ parameter δ equals true δ
620
+ When approximate sensitivity parameters δ are used (δ ∈ {−1, 1} when true δ ∈ {−2, 2}),
621
+ the middle 95% values of adjusted estimates also cover the true ATE whereas those of
622
+ unadjusted estimates do not (Figure 4).
623
+ Figure 5 compares the errors in the estimates and sensitivity of associated inferences
624
+ between the outcome proxy-blind method of Dahabreh et al. (2020b); Lesko et al. (2017)
625
+ and our proposed method across 1000 simulations.
626
+ 21
627
+
628
+ True = -2
629
+ True = 0
630
+ True = 2
631
+ 8
632
+ 6
633
+ True uO = -3
634
+ 4
635
+ 2
636
+ 0
637
+ 8
638
+ 6
639
+ Estimate
640
+ True uO = 0
641
+ 4
642
+ 2
643
+ 0
644
+ 8
645
+ 6
646
+ True uO = 3
647
+ 4
648
+ 2
649
+ 0
650
+ 2
651
+ U
652
+ 2
653
+ -2
654
+ 1
655
+ 2
656
+ 0
657
+ 2
658
+ Sensitivity parameterFigure 5: Sensitivity-parameter-adjusted ATE estimates obtained from our proposed method
659
+ and the outcome proxy-blind method; mean, 2.5th and 97.5th quantiles obtained from 1000
660
+ simulations. When sensitivity parameter δ = 0, the adjusted estimate corresponds to the
661
+ unadjusted estimate. Horizontal dotted line shows the true overall ATE given true δ; vertical
662
+ dotted line indicates sensitivity parameter δ equals true δ
663
+ The distributions of the estimates from both methods are centered on the true parameter.
664
+ However, the estimates tend to be more precise when we utilize the information from the
665
+ outcome proxy (as demonstrated through the narrower 2.5th-97.5th quantile range). The
666
+ efficiency gains have implications for the sensitivity analysis, since resulting inferences are
667
+ not as sensitive given analogous magnitude in violation of the identification assumption 4.
668
+ Assumption 4 implies both u0 and u1 equal 0. As a result, the true δ also equals 0.
669
+ This suggests transportation of the conditional potential outcome means, and in turn, the
670
+ 22
671
+
672
+ True = -2
673
+ True = 0
674
+ True θ = 2
675
+ 6
676
+ 5
677
+ 3
678
+ 6
679
+ 4
680
+ 3
681
+ 6
682
+ 5
683
+ 2
684
+ 2
685
+ 2
686
+ -2
687
+ 2
688
+ Sensitivityparameter
689
+ Outcome regression method
690
+ Proposed methodconditional average treatment effects, can be done without incurring bias (vertical middle
691
+ panes, figure 3). We also observed that, when δ is 0, regardless of the values of u0 (and
692
+ u1), there is also no bias (vertical middle panes, figure 3) in the unadjusted estimator. In
693
+ both cases, no bias correction would be necessary, and incorporating a non-zero δ sensitivity
694
+ parameter will actually introduce bias to the estimate.
695
+ 6.3.2
696
+ Bias terms as bounded functions
697
+ When the sensitivity parameter γ is greater or equal to max{γ0, γ1} for the true function
698
+ bounds γ0 and γ1, the bounds always include the true ATE when the bias functions are
699
+ bounded by γ0 and γ1 (Figure 6).
700
+ 23
701
+
702
+ Figure 6: ATE estimates with sensitivity bounds shown against the true overall ATE across
703
+ values of the true bias and sensitivity parameter. When sensitivity parameter γ = 0, the
704
+ bounds collapse to a point estimate. Blue horizontal dotted line shows the true study-specific
705
+ ATE given true bias functions
706
+ Although this approach requires minimal assumptions about the bias functional form, it
707
+ can also be conservative since the true bias functions are unlikely to evaluate to the bounds
708
+ across the domain of the functions. For instance, the bottom three panels of Figure 6 show
709
+ that when the sensitivity parameter γ is greater than or equal to max(true γ0, true γ1),
710
+ while the bounds on the estimate contain the true ATE, they also contains the null value
711
+ zero as well. On the other hand, these bounds do not rely on an assumption of constant bias
712
+ functions, which we may often have no reason to believe. Here, we demonstrated through
713
+ 24
714
+
715
+ True y 1 = 1
716
+ True 1 = 2
717
+ True y 1 = 3
718
+ U
719
+ Estimated ATEwith
720
+ 4
721
+ n
722
+ 8.
723
+ Sensitivity parameter ysimulations that sensitivity analysis with relaxed and more credible assumptions can still
724
+ provide helpful information about the parameter of interest. However, when the bounds
725
+ are too narrow or too wide, sensitivity analysis using bounded bias functions might not be
726
+ accurate (i.e., not containing the true parameter) or useful (i.e., containing the null value
727
+ when the truth is non-null), respectively.
728
+ 7
729
+ Discussion
730
+ In this paper, we discussed a data integrative method that utilizes information from avail-
731
+ able proxies of the outcome of interest measured at follow-up for efficiency gains. We then
732
+ presented two sensitivity analysis strategies specific to this approach for causal effect trans-
733
+ portation when the identification assumption is violated. Our modification to the identifica-
734
+ tion of the ATE in (1) allows for more efficient estimators given sufficiently strong outcome
735
+ proxies. As a result, our bias functions also have similar, yet distinct interpretations than
736
+ the bias functions of Dahabreh and Hern´an (2019).
737
+ When the bias terms are assumed to be constants, we can obtain different bias-adjusted
738
+ point estimates based on our specification of the sensitivity parameters. Additionally, via
739
+ obtaining the 95% bootstrap confidence interval for the bias-adjusted estimates, we can
740
+ examine the robustness of inferences made using our method under varying magnitudes of
741
+ assumption violation. Specifically, beyond certain values of the sensitivity parameters, the
742
+ 95% CI will cross the null value 0. These are the degrees of violation that can affect inferences
743
+ (where the 95% CI suggest a change from significant results to non-significant results).
744
+ We also proposed sensitivity analysis using bounded bias functions as an alternative when
745
+ one believes the assumption of a fixed-value bias term is too strong. This approach allows
746
+ for inferences with minimal assumptions about the unobserved bias functions but can still
747
+ provide useful information about the parameter of interest. Due to fewer assumptions being
748
+ made, the results are more conservative and robust, hence more reasonable and credible.
749
+ 25
750
+
751
+ Specifically, although we are unable to obtain a point estimate, sensitivity analysis using
752
+ bounded bias functions can still be informative in the sense of providing information about
753
+ the general direction of the parameter of interest (beneficial or harmful). This method is
754
+ generally more conservative if the bounds on the functions are not close to their extreme
755
+ values, if the bias functions are generally not close to their extreme values, or if there is a
756
+ large difference between the extrema of the two bias functions.
757
+ Correct specification of the bias functions would allow for more precise and informative
758
+ estimation of the true ATE. However, since they are generally unknown and non-estimable
759
+ from observed data, sensitivity analysis will typically be the realistic course of action.
760
+ When conducting sensitivity analysis, the analyst can start off by specifying a wide grid
761
+ of the sensitivity parameter and examining the behaviors of the point estimates and 95%
762
+ CI (first approach) as well as bounds around the estimates (second approach). They can
763
+ then search for the “critical” sensitivity parameters that still suggest rejection of the null
764
+ hypothesis, i.e., the 95% CI (in the first case) and bounds around the estimate (in the
765
+ second case) that do not contain 0. It can be determined if greater bias is plausible by using
766
+ background knowledge of the data generating mechanism or further hypothesizing about
767
+ such mechanism. If there is little or no evidence that the true bias functions exceed these
768
+ critical sensitivity parameters, one can be more comfortable in concluding that the observed
769
+ effect and associated inferences are robust to violation of the transportability assumption
770
+ (Ding and VanderWeele, 2016; Cornfield et al., 1959).
771
+ References
772
+ Bareinboim, E. and Pearl, J. (2016). Causal inference and the data-fusion problem. Pro-
773
+ ceedings of the National Academy of Sciences, 113(27):7345–7377. 1
774
+ Barlati, S., Deste, G., De Peri, L., Ariu, C., and Vita, A. (2013). Cognitive remediation
775
+ 26
776
+
777
+ in schizophrenia: Current status and future perspectives.
778
+ Schizophrenia Research and
779
+ Treatment, 2013:156084. 7
780
+ Combs, D. R., Tosheva, A., Penn, D. L., Basso, M. R., Wanner, J. L., and Laib, K. (2008).
781
+ Attentional-shaping as a means to improve emotion perception deficits in schizophrenia.
782
+ Schizophrenia Research, 105(1-3):68–77. 7
783
+ Cornfield, J., Haenszel, W., Hammond, E. C., Lilienfeld, A. M., Shimkin, M. B., and Wynder,
784
+ E. L. (1959). Smoking and lung cancer: Recent evidence and a discussion of some questions.
785
+ Journal of the National Cancer Institute, 22(1):173–203. 26
786
+ Dahabreh, I., Robins, J., Haneuse, S., Robertson, S., Steingrimsson, J., and Hern´an, M.
787
+ (2022). Global sensitivity analysis for studies extending inferences from a randomized
788
+ trial to a target population. arXiv preprint arXiv:2207.09982. 2
789
+ Dahabreh, I. J. and Hern´an, M. A. (2019). Extending inferences from a randomized trial to
790
+ a target population. European Journal of Epidemiology, 34(8):719–722. 2, 6, 9, 25
791
+ Dahabreh, I. J., Petito, L. C., Robertson, S. E., Hern´an, M. A., and Steingrimsson, J. A.
792
+ (2020a). Toward causally interpretable meta-analysis: Transporting inferences from mul-
793
+ tiple randomized trials to a new target population. Epidemiology, 31(3):334–344. 2
794
+ Dahabreh, I. J., Robertson, S. E., Steingrimsson, J. A., Stuart, E. A., and Hernan, M. A.
795
+ (2020b). Extending inferences from a randomized trial to a new target population. Statis-
796
+ tics in Medicine, 39(14):1999–2014. 2, 21
797
+ Degtiar, I. and Rose, S. (2023). A review of generalizability and transportability. Annual
798
+ Review of Statistics and Its Application, 10(1). 1
799
+ Ding, P. and VanderWeele, T. J. (2016). Sensitivity analysis without assumptions. Epidemi-
800
+ ology (Cambridge, Mass.), 27(3):368. 26
801
+ 27
802
+
803
+ Dong, L., Yang, S., Wang, X., Zeng, D., and Cai, J. (2020). Integrative analysis of ran-
804
+ domized clinical trials with real world evidence studies. arXiv preprint arXiv:2003.01242.
805
+ 2
806
+ Gechter, M. (2015). Generalizing the results from social experiments: Theory and evidence
807
+ from Mexico and India.
808
+ Boston Univ., Department of Economics, Inst. for Economic
809
+ Development. 1
810
+ H¨unermund, P. and Bareinboim, E. (2019). Causal inference and data fusion in econometrics.
811
+ arXiv preprint arXiv:1912.09104. 2
812
+ Kennedy-Martin, T., Curtis, S., Faries, D., Robinson, S., and Johnston, J. (2015). A litera-
813
+ ture review on the representativeness of randomized controlled trial samples and implica-
814
+ tions for the external validity of trial results. Trials, 16(495). 1
815
+ Lesko, C., Buchanan, A., Westreich, D., Edwards, J., Hudgens, M., and Cole, S. (2017).
816
+ Generalizing study results: A potential outcomes perspective. Epidemiology (Cambridge,
817
+ Mass.), 28(4):553. 2, 6, 21
818
+ McGurk, S., Twamley, E., Sitzer, D., McHugo, G., and Mueser, K. (2007).
819
+ A meta-
820
+ analysis of cognitive remediation in schizophrenia. The American Journal of Psychiatry,
821
+ 164(12):1791–1802. 7
822
+ Nguyen, T., Ebnesajjad, C., Cole, S., and Stuart, E. (2017). Sensitivity analysis for an
823
+ unobserved moderator in RCT-to-target-population generalization of treatment effects.
824
+ The Annals of Applied Statistics, 11(1):225 – 247. 2
825
+ Pearl, J. and Bareinboim, E. (2014). External validity: From do-calculus to transportability
826
+ across populations. In Probabilistic and Causal Inference: The Works of Judea Pearl,
827
+ pages 451–482. 2
828
+ 28
829
+
830
+ Tan, Y., Papez, V., Chang, W., Mueller, S., Denaxas, S., and Lai, A. (2022). Comparing
831
+ clinical trial population representativeness to real-world populations: An external validity
832
+ analysis encompassing 43,895,trials and 5,685,738 individuals across 989 unique drugs and
833
+ 286 conditions in England. The Lancet Healthy Longevity, 3(10):e674–e689. 1
834
+ Tipton, E. (2014).
835
+ How generalizable is your experiment?
836
+ An index for comparing ex-
837
+ perimental samples and populations. Journal of Educational and Behavioral Statistics,
838
+ 39(6):478–501. 1
839
+ Westreich, D., Edwards, J. K., Lesko, C. R., Stuart, E., and Cole, S. R. (2017). Trans-
840
+ portability of trial results using inverse odds of sampling weights. American Journal of
841
+ Epidemiology, 186(8):1010–1014. 2
842
+ Wykes, T., Huddy, V., Cellard, C., McGurk, S., and Czobor, P. (2011). A meta-analysis
843
+ of cognitive remediation for schizophrenia: Methodology and effect sizes. The American
844
+ Journal of Psychiatry, 168(5):472–485. 7
845
+ Wykes, T., Reeder, C., Landau, S., Everitt, B., Knapp, M., Patel, A., and Romeo, R. (2007).
846
+ Cognitive remediation therapy in schizophrenia: Randomised controlled trial. The British
847
+ Journal of Psychiatry, 190(5):421–427. 7
848
+ 29
849
+
850
+ A
851
+ Additional figures
852
+ Figure A.1: Sensitivity-parameter-adjusted ATE estimates shown against the true overall
853
+ ATE across values of the true bias and sensitivity parameter; n=200 for each study, 95% CI
854
+ constructed from 1000 bootstrap samples. When sensitivity parameter δ = 0, the adjusted
855
+ estimate corresponds to the unadjusted estimate.
856
+ Horizontal dotted line shows the true
857
+ overall ATE given true δ; vertical dotted line indicates sensitivity parameter δ equals true δ
858
+ 30
859
+
860
+ True = -2
861
+ True = 0
862
+ True θ = 2
863
+ 6
864
+ 6
865
+ stimate
866
+ 5
867
+ rue
868
+ uo
869
+ 4
870
+ =
871
+ -
872
+ 3
873
+ 6
874
+ 5
875
+ 4
876
+ 3
877
+ 3
878
+ 2
879
+ 0
880
+ -2
881
+ 0
882
+ 2
883
+ 2
884
+ 2
885
+ Sensitivity parameterFigure A.2: Sensitivity-parameter-adjusted ATE estimates shown against the true ATE in
886
+ the study with missing outcome across values of the true bias and sensitivity parameter;
887
+ n=200 for each study, 95% CI constructed from 1000 bootstrap samples. When sensitivity
888
+ parameter δ = 0, the adjusted estimate corresponds to the unadjusted estimate. Horizontal
889
+ dotted line shows the true study-specific ATE given true δ; vertical dotted line indicates
890
+ sensitivity parameter δ equals true δ
891
+ 31
892
+
893
+ True = -2
894
+ True = 0
895
+ True θ = 2
896
+ 7.5
897
+ 5.0
898
+ rue
899
+ uo
900
+ =
901
+ 2.5
902
+ 0.0
903
+ 8
904
+ 6
905
+ Estimate
906
+ True
907
+ 4
908
+ uo
909
+ II
910
+ -
911
+ 0
912
+ 8
913
+ 6.
914
+ True uO =
915
+ 4
916
+ 2.*
917
+ 3
918
+ 0
919
+ 2
920
+ 0
921
+ 2
922
+ -2
923
+ 0
924
+ 2
925
+ Sensitivity parameterFigure A.3: Sensitivity-parameter-adjusted ATE estimates shown against the true overall
926
+ ATE across values of the true bias and sensitivity parameter; n=500 for each study, 95% CI
927
+ constructed from 1000 bootstrap samples. When sensitivity parameter δ = 0, the adjusted
928
+ estimate corresponds to the unadjusted estimate.
929
+ Horizontal dotted line shows the true
930
+ overall ATE given true δ; vertical dotted line indicates sensitivity parameter δ equals true δ
931
+ 32
932
+
933
+ True = -2
934
+ True = 0
935
+ True θ = 2
936
+ 6
937
+ 5
938
+ 4
939
+ 3
940
+ =3
941
+ 2
942
+ 6
943
+ 5
944
+ Estimate
945
+ True
946
+ 4
947
+ uo
948
+ =
949
+ 3
950
+ 2
951
+ 6
952
+ True uO
953
+ 4
954
+ II
955
+ 3
956
+ 0
957
+ -2
958
+ 2
959
+ Sensitivity parameter Figure A.4: Sensitivity-parameter-adjusted ATE estimates shown against the true ATE in
960
+ the study with missing outcome across values of the true bias and sensitivity parameter;
961
+ n=500 for each study, 95% CI constructed from 1000 bootstrap samples. When sensitivity
962
+ parameter δ = 0, the adjusted estimate corresponds to the unadjusted estimate. Horizontal
963
+ dotted line shows the true study-specific ATE given true δ; vertical dotted line indicates
964
+ sensitivity parameter δ equals true δ
965
+ 33
966
+
967
+ True = -2
968
+ True = 0
969
+ True θ = 2
970
+ 6
971
+ True
972
+ 4
973
+ u0 = -3
974
+ 2
975
+ 0.
976
+ 8
977
+ Estimate
978
+ 6
979
+ True uO = 0
980
+ 4
981
+ -T
982
+ -
983
+ 0
984
+ 8
985
+ I
986
+ 6.
987
+ True uO =
988
+ 4
989
+ 2
990
+ 2
991
+ 0
992
+ -2
993
+ 0
994
+ 2
995
+ 2
996
+ 2
997
+ Sensitivity parameterB
998
+ Derivation of the sensitivity analysis formula when
999
+ Assumption 4 is violated
1000
+ ATE =
1001
+ s∗
1002
+
1003
+ s=1
1004
+ πsE{E(Y | W, A = 1, S = s) − E(Y | W, A = 0, S = s) | S = s}
1005
+ +
1006
+ S
1007
+
1008
+ s=s∗+1
1009
+ πsE [E {E (Y | Ts, W, A = 1, S = s) | W, A = 1, S = s}
1010
+ − E {E (Y | Ts, W, A = 0, S = s) | W, A = 0, S = s} | S = s]
1011
+ =
1012
+ s∗
1013
+
1014
+ s=1
1015
+ πsE{E(Y | W, A = 1, S = s) − E(Y | W, A = 0, S = s) | S = s}
1016
+ +
1017
+ S
1018
+
1019
+ s=s∗+1
1020
+ πsE [E {E (Y | Ts, W, A = 1, S ∈ σs) + u (A = 1, Ts, W) | W, A = 1, S = s}
1021
+ − E {E (Y | Ts, W, A = 0, S ∈ σs) + u (A = 0, Ts, W) | W, A = 0, S = s} | S = s]
1022
+ =
1023
+ s∗
1024
+
1025
+ s=1
1026
+ πsE{E(Y | W, A = 1, S = s) − E(Y | W, A = 0, S = s) | S = s)
1027
+ +
1028
+ S
1029
+
1030
+ s=s∗+1
1031
+ πsE [E {E (Y | Ts, W, A = 1, S ∈ σs) | W, A = 1, S = s)}
1032
+ −E {E (Y | Ts, W, A = 0, S ∈ σs) | W, A = 0, S = s)} | S = s]
1033
+ +
1034
+ S
1035
+
1036
+ s=s∗+1
1037
+ πsE[E {u (A = 1, Ts, W) | W, A = 1, S = s}
1038
+ − E {u (A = 0, Ts, W) | W, A = 0, S = s} | S = s]
1039
+ 34
1040
+