jackkuo commited on
Commit
3fa6a84
·
verified ·
1 Parent(s): f0dc773

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. -9E1T4oBgHgl3EQfogQj/content/2301.03320v1.pdf +3 -0
  2. -9E1T4oBgHgl3EQfogQj/vector_store/index.faiss +3 -0
  3. -9E1T4oBgHgl3EQfogQj/vector_store/index.pkl +3 -0
  4. .gitattributes +53 -0
  5. 09FLT4oBgHgl3EQfpS-5/content/tmp_files/2301.12135v1.pdf.txt +1531 -0
  6. 09FLT4oBgHgl3EQfpS-5/content/tmp_files/load_file.txt +0 -0
  7. 39FAT4oBgHgl3EQfEhxj/content/2301.08422v1.pdf +3 -0
  8. 3NE1T4oBgHgl3EQf5wVj/content/tmp_files/2301.03515v1.pdf.txt +1635 -0
  9. 3NE1T4oBgHgl3EQf5wVj/content/tmp_files/load_file.txt +0 -0
  10. 3dE2T4oBgHgl3EQfjgcV/content/tmp_files/2301.03968v1.pdf.txt +1082 -0
  11. 3dE2T4oBgHgl3EQfjgcV/content/tmp_files/load_file.txt +0 -0
  12. 3tE2T4oBgHgl3EQfOAZw/content/tmp_files/2301.03743v1.pdf.txt +636 -0
  13. 3tE2T4oBgHgl3EQfOAZw/content/tmp_files/load_file.txt +0 -0
  14. 4tAyT4oBgHgl3EQfcPfW/content/2301.00280v1.pdf +3 -0
  15. 4tAyT4oBgHgl3EQfcPfW/vector_store/index.faiss +3 -0
  16. 5dAyT4oBgHgl3EQf2fkR/vector_store/index.pkl +3 -0
  17. 5tFAT4oBgHgl3EQfmx2I/content/2301.08625v1.pdf +3 -0
  18. 5tFAT4oBgHgl3EQfmx2I/vector_store/index.pkl +3 -0
  19. 79AyT4oBgHgl3EQfQvZ4/content/tmp_files/2301.00052v1.pdf.txt +378 -0
  20. 79AyT4oBgHgl3EQfQvZ4/content/tmp_files/load_file.txt +317 -0
  21. 89FLT4oBgHgl3EQfti_6/content/tmp_files/2301.12152v1.pdf.txt +1524 -0
  22. 89FLT4oBgHgl3EQfti_6/content/tmp_files/load_file.txt +0 -0
  23. CNFKT4oBgHgl3EQfXy51/content/tmp_files/2301.11796v1.pdf.txt +521 -0
  24. CNFKT4oBgHgl3EQfXy51/content/tmp_files/load_file.txt +304 -0
  25. DNAyT4oBgHgl3EQf4frj/vector_store/index.faiss +3 -0
  26. F9AyT4oBgHgl3EQf5Pp6/vector_store/index.faiss +3 -0
  27. G9E0T4oBgHgl3EQfhgFS/content/tmp_files/2301.02432v1.pdf.txt +1388 -0
  28. G9E0T4oBgHgl3EQfhgFS/content/tmp_files/load_file.txt +0 -0
  29. GNE1T4oBgHgl3EQfXAT2/content/tmp_files/2301.03123v1.pdf.txt +1030 -0
  30. GNE1T4oBgHgl3EQfXAT2/content/tmp_files/load_file.txt +0 -0
  31. GtE0T4oBgHgl3EQfzgJq/content/tmp_files/2301.02673v1.pdf.txt +940 -0
  32. GtE0T4oBgHgl3EQfzgJq/content/tmp_files/load_file.txt +0 -0
  33. I9AzT4oBgHgl3EQfVPxY/content/tmp_files/load_file.txt +156 -0
  34. J9FJT4oBgHgl3EQfwy0B/content/2301.11631v1.pdf +3 -0
  35. J9FJT4oBgHgl3EQfwy0B/vector_store/index.pkl +3 -0
  36. JtE2T4oBgHgl3EQfUwcR/content/2301.03815v1.pdf +3 -0
  37. JtE2T4oBgHgl3EQfUwcR/vector_store/index.faiss +3 -0
  38. JtE2T4oBgHgl3EQfUwcR/vector_store/index.pkl +3 -0
  39. KNE3T4oBgHgl3EQfXwo4/content/tmp_files/2301.04481v1.pdf.txt +1513 -0
  40. KNE3T4oBgHgl3EQfXwo4/content/tmp_files/load_file.txt +398 -0
  41. KdE1T4oBgHgl3EQfGgND/vector_store/index.faiss +3 -0
  42. LNE4T4oBgHgl3EQfiQ3g/vector_store/index.faiss +3 -0
  43. MNAyT4oBgHgl3EQf6vpX/content/2301.00826v1.pdf +3 -0
  44. MNAyT4oBgHgl3EQf6vpX/vector_store/index.pkl +3 -0
  45. MtAzT4oBgHgl3EQfkv1s/content/tmp_files/2301.01536v1.pdf.txt +671 -0
  46. MtAzT4oBgHgl3EQfkv1s/content/tmp_files/load_file.txt +0 -0
  47. MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf +0 -0
  48. MtE2T4oBgHgl3EQfqgg6/content/tmp_files/2301.04039v1.pdf.txt +232 -0
  49. MtE2T4oBgHgl3EQfqgg6/content/tmp_files/load_file.txt +114 -0
  50. N9AzT4oBgHgl3EQfk_2S/content/2301.01541v1.pdf +3 -0
-9E1T4oBgHgl3EQfogQj/content/2301.03320v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:84c877b87a34a94d8423105c524818e74ff067e9bdb75286d4452c3480ce68bb
3
+ size 713341
-9E1T4oBgHgl3EQfogQj/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e12f4e5eb136ed8d53a197798e7745e057eef6ad0f1b93033153a636d2204442
3
+ size 1310765
-9E1T4oBgHgl3EQfogQj/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e65c8a84a4e1ff2ccedf8e0e4d0c32378b25223eadc4722a2ed4cffa287dee0
3
+ size 47661
.gitattributes CHANGED
@@ -9447,3 +9447,56 @@ G9E1T4oBgHgl3EQfrQVc/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -tex
9447
  CdE4T4oBgHgl3EQfeQ2g/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9448
  lNE5T4oBgHgl3EQfiQ-H/content/2301.05647v1.pdf filter=lfs diff=lfs merge=lfs -text
9449
  hdE1T4oBgHgl3EQfMgPQ/content/2301.02991v1.pdf filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9447
  CdE4T4oBgHgl3EQfeQ2g/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9448
  lNE5T4oBgHgl3EQfiQ-H/content/2301.05647v1.pdf filter=lfs diff=lfs merge=lfs -text
9449
  hdE1T4oBgHgl3EQfMgPQ/content/2301.02991v1.pdf filter=lfs diff=lfs merge=lfs -text
9450
+ aNAyT4oBgHgl3EQfifjx/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9451
+ PtAzT4oBgHgl3EQfIvs5/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9452
+ TtE3T4oBgHgl3EQfEAn3/content/2301.04293v1.pdf filter=lfs diff=lfs merge=lfs -text
9453
+ aNAyT4oBgHgl3EQfifjx/content/2301.00399v1.pdf filter=lfs diff=lfs merge=lfs -text
9454
+ ctAzT4oBgHgl3EQfnv1S/content/2301.01585v1.pdf filter=lfs diff=lfs merge=lfs -text
9455
+ -9E1T4oBgHgl3EQfogQj/content/2301.03320v1.pdf filter=lfs diff=lfs merge=lfs -text
9456
+ R9E2T4oBgHgl3EQfWgcY/content/2301.03833v1.pdf filter=lfs diff=lfs merge=lfs -text
9457
+ ctAzT4oBgHgl3EQfnv1S/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9458
+ -9E1T4oBgHgl3EQfogQj/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9459
+ LNE4T4oBgHgl3EQfiQ3g/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9460
+ JtE2T4oBgHgl3EQfUwcR/content/2301.03815v1.pdf filter=lfs diff=lfs merge=lfs -text
9461
+ ftA0T4oBgHgl3EQfHv83/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9462
+ rdE2T4oBgHgl3EQf1AiG/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9463
+ oNAyT4oBgHgl3EQflfjb/content/2301.00455v1.pdf filter=lfs diff=lfs merge=lfs -text
9464
+ JtE2T4oBgHgl3EQfUwcR/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9465
+ 4tAyT4oBgHgl3EQfcPfW/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9466
+ MNAyT4oBgHgl3EQf6vpX/content/2301.00826v1.pdf filter=lfs diff=lfs merge=lfs -text
9467
+ 5tFAT4oBgHgl3EQfmx2I/content/2301.08625v1.pdf filter=lfs diff=lfs merge=lfs -text
9468
+ vdAzT4oBgHgl3EQf7P4z/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9469
+ btE5T4oBgHgl3EQfEQ7d/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9470
+ hdE1T4oBgHgl3EQfMgPQ/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9471
+ OdAzT4oBgHgl3EQfIfsX/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9472
+ rtE5T4oBgHgl3EQfKA5U/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9473
+ idAyT4oBgHgl3EQfXvfg/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9474
+ DNAyT4oBgHgl3EQf4frj/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9475
+ btE5T4oBgHgl3EQfEQ7d/content/2301.05413v1.pdf filter=lfs diff=lfs merge=lfs -text
9476
+ zNAzT4oBgHgl3EQfC_oX/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9477
+ _tFKT4oBgHgl3EQfVC14/content/2301.11786v1.pdf filter=lfs diff=lfs merge=lfs -text
9478
+ bNE_T4oBgHgl3EQfzRxG/content/2301.08322v1.pdf filter=lfs diff=lfs merge=lfs -text
9479
+ 4tAyT4oBgHgl3EQfcPfW/content/2301.00280v1.pdf filter=lfs diff=lfs merge=lfs -text
9480
+ ZNE1T4oBgHgl3EQfKAM7/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9481
+ ptE0T4oBgHgl3EQfqwEz/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9482
+ rdE2T4oBgHgl3EQf1AiG/content/2301.04147v1.pdf filter=lfs diff=lfs merge=lfs -text
9483
+ VtE0T4oBgHgl3EQflwHV/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9484
+ KdE1T4oBgHgl3EQfGgND/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9485
+ WNE3T4oBgHgl3EQf0wtX/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9486
+ nNE1T4oBgHgl3EQfhQT4/content/2301.03240v1.pdf filter=lfs diff=lfs merge=lfs -text
9487
+ 39FAT4oBgHgl3EQfEhxj/content/2301.08422v1.pdf filter=lfs diff=lfs merge=lfs -text
9488
+ YtE2T4oBgHgl3EQfvAjJ/content/2301.04087v1.pdf filter=lfs diff=lfs merge=lfs -text
9489
+ xtAzT4oBgHgl3EQf7v7M/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9490
+ tNAzT4oBgHgl3EQfBfpr/content/2301.00943v1.pdf filter=lfs diff=lfs merge=lfs -text
9491
+ xtE4T4oBgHgl3EQfxg0u/content/2301.05258v1.pdf filter=lfs diff=lfs merge=lfs -text
9492
+ F9AyT4oBgHgl3EQf5Pp6/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9493
+ ztFIT4oBgHgl3EQf2Ssk/content/2301.11376v1.pdf filter=lfs diff=lfs merge=lfs -text
9494
+ htAzT4oBgHgl3EQfa_y7/content/2301.01379v1.pdf filter=lfs diff=lfs merge=lfs -text
9495
+ sdE1T4oBgHgl3EQf3QVe/content/2301.03487v1.pdf filter=lfs diff=lfs merge=lfs -text
9496
+ bNE_T4oBgHgl3EQfzRxG/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9497
+ xdE4T4oBgHgl3EQfYAwp/content/2301.05045v1.pdf filter=lfs diff=lfs merge=lfs -text
9498
+ d9A0T4oBgHgl3EQfHP8Z/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
9499
+ fdAzT4oBgHgl3EQfMPug/content/2301.01129v1.pdf filter=lfs diff=lfs merge=lfs -text
9500
+ N9AzT4oBgHgl3EQfk_2S/content/2301.01541v1.pdf filter=lfs diff=lfs merge=lfs -text
9501
+ J9FJT4oBgHgl3EQfwy0B/content/2301.11631v1.pdf filter=lfs diff=lfs merge=lfs -text
9502
+ htAzT4oBgHgl3EQfa_y7/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
09FLT4oBgHgl3EQfpS-5/content/tmp_files/2301.12135v1.pdf.txt ADDED
@@ -0,0 +1,1531 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ AdaSfM: From Coarse Global to Fine Incremental Adaptive
2
+ Structure from Motion
3
+ Yu Chen1, Zihao Yu2, Shu Song2, Tianning Yu3, Jianming Li3, Gim Hee Lee1
4
+ Abstract— Despite the impressive results achieved by many
5
+ existing Structure from Motion (SfM) approaches, there is still
6
+ a need to improve the robustness, accuracy, and efficiency
7
+ on large-scale scenes with many outlier matches and sparse
8
+ view graphs. In this paper, we propose AdaSfM: a coarse-
9
+ to-fine adaptive SfM approach that is scalable to large-scale
10
+ and challenging datasets. Our approach first does a coarse
11
+ global SfM which improves the reliability of the view graph by
12
+ leveraging measurements from low-cost sensors such as Inertial
13
+ Measurement Units (IMUs) and wheel encoders. Subsequently,
14
+ the view graph is divided into sub-scenes that are refined in
15
+ parallel by a fine local incremental SfM regularised by the result
16
+ from the coarse global SfM to improve the camera registration
17
+ accuracy and alleviate scene drifts. Finally, our approach uses
18
+ a threshold-adaptive strategy to align all local reconstructions
19
+ to the coordinate frame of global SfM. Extensive experiments
20
+ on large-scale benchmark datasets show that our approach
21
+ achieves state-of-the-art accuracy and efficiency.
22
+ I. INTRODUCTION
23
+ Structure from Motion (SfM) is an important topic that
24
+ has been studied intensively over the past two decades. It
25
+ has wide applications in augmented reality and autonomous
26
+ driving for visual localization [1], [2], [3], and in multi-view
27
+ stereo [4], [5] and novel view synthesis [6] by providing
28
+ camera poses and optional sparse scene structures.
29
+ Despite the impressive results from many existing works,
30
+ SfM remains challenging in two aspects. The first challenge
31
+ is outlier feature matches caused by the diversity of scene
32
+ features, e.g. texture-less, self-similar, non-Lambertian, etc.
33
+ These diverse features impose challenges in sparse feature
34
+ extraction and matching which result in outliers that are detri-
35
+ mental to the subsequent reconstruction process. Incremental
36
+ SfM [7], [8] is notoriously known to suffer from drift due
37
+ to error accumulation, though is robust in handling outliers.
38
+ Global SfM methods [9], [10], [11] are proposed to handle
39
+ drift, but fail to solve the scale ambiguities [12] of camera
40
+ positions and are not robust to outliers [13], [14].
41
+ The second challenge is sparse view graphs from some
42
+ large-scale datasets. Incremental SfM is known to be ineffi-
43
+ cient on large-scale datasets. Several works [16], [17], [18],
44
+ [19] have been proposed to handle millions of images. These
45
+ are divide-and-conquer SfM methods that deal with very
46
+ large-scale datasets by grouping images into partitions. Each
47
+ partition is processed by a cluster of servers that concurrently
48
+ 1School of Computing, National University of Singapore, {chenyu,
49
+ gimhee.lee}@comp.nus.edu.sg
50
+ 2Segway-Ninebot Robotics Co., Ltd, yuzihao@buaa.edu.cn,
51
+ songshu0905@gmail.com
52
+ 3Navimow B.V. Co., Ltd, tianning.yu@rlm.segway.com,
53
+ jianming.li@ninebot.com
54
+ Fig. 1. When combining with global SfM, our AdaSfM is more robust than
55
+ traditional incremental SfM (tested on the public 4Seasons dataset [15]).
56
+ circumvents the memory limitation. However, these methods
57
+ [16], [17], [18], [19] are often limited to internet datasets
58
+ or aerial images where the view graphs are very densely
59
+ connected. The dense connections in the view graph ensure
60
+ that there are sufficient constraints between the graph par-
61
+ titions. Nonetheless, divide-and-conquer methods often fail
62
+ in datasets with weak associations between images for local
63
+ reconstruction alignments or lack of visual constraints for
64
+ stable camera registration. An example of such a dataset
65
+ is autonomous self-driving cars where the interval between
66
+ consecutive images can be large.
67
+ In view of the challenges from the outlier feature matches
68
+ and sparse view graphs on the existing SfM approaches,
69
+ we propose AdaSfM: a coarse-to-fine adaptive SfM pipeline
70
+ to enhance the robustness of SfM in dealing with large-
71
+ scale challenging scenes. Specifically, we first solve the
72
+ global SfM at a coarse scale, and then the result of the
73
+ global SfM is used to enhance the scalability of the local
74
+ incremental reconstruction. Both the scale ambiguities and
75
+ outlier ratio in global SfM can be significantly reduced
76
+ by incorporating measurements from the IMU and wheel
77
+ encoder, which are often available in mobile devices or
78
+ autonomous self-driving cars. We preintegrate [20] the IMU
79
+ measurements to get the relative poses of consecutive frames
80
+ Pt
81
+ = {Pt0, Pt1, · · · }, and use the measurements from
82
+ the wheel encoder to constrain scale drifts of the IMU
83
+ preintegration [21]. We then replace the relative poses of the
84
+ consecutive frames in the view graph formed by two-view
85
+ geometry [22], [8] with Pt estimated by the IMU and wheel
86
+ encoder. This augmented view graph is then used to estimate
87
+ arXiv:2301.12135v1 [cs.CV] 28 Jan 2023
88
+
89
+ Fig. 2.
90
+ The pipeline of our proposed SfM method. Our method takes images and measurements from low-cost sensors as inputs. The view graph is
91
+ built after feature matching and refined by the result of global SfM. The absolute poses from the global SfM are used as priors in the subsequent local
92
+ SfM process. The final reconstruction result is merged into the global SfM reference frame.
93
+ the global poses. Consequently, we obtain a coarse scene
94
+ structure and camera poses, where the latter can be used
95
+ to filter wrong feature matches. Since that, we partition the
96
+ view graph with the existing graph cut method [23] and then
97
+ extend the sub-graphs with a novel adaptive flood-fill method
98
+ to enhance the constraints of separators [24]. We define
99
+ separators as images that connect different sub-graphs. For
100
+ each local SfM, the poses from the global SfM are used for
101
+ camera registration and to constrain the global refinement of
102
+ 3D points and camera poses. Finally, we design an adaptive
103
+ global alignment strategy to merge local reconstructions with
104
+ the coordinate frame of the global SfM set as the reference
105
+ frame. We illustrate the pipeline of our method in Fig. 2.
106
+ We evaluate our method extensively on large-scale chal-
107
+ lenging scenes. Experimental results show that our AdaSfM
108
+ is adaptive to different scene structures. Furthermore, we
109
+ achieve better robustness and comparable efficiency in com-
110
+ parison to existing state-of-the-art SfM methods.
111
+ II. RELATED WORK
112
+ Incremental SfM. Agarwal et al. [7] apply preconditioned
113
+ conjugate gradient [25] to accelerate large-scale BA [26].
114
+ The drift problem is alleviated in [27] with a re-triangulation
115
+ (RT) step before global BA. Sch¨onberger and Frahm [8]
116
+ augment the view graph by estimating multiple geometric
117
+ models in geometric verification and improve the image
118
+ registration robustness with next best view selection. In
119
+ addition to the RT before BA [27], RT is also performed
120
+ after BA in [8]. To reduce the time complexity of repetitive
121
+ image registration, Cui et al [28] select a batch of images
122
+ for registration, and select a subset of good tracks for BA.
123
+ Global SfM. The simplest configuration of a global SfM
124
+ method only requires 1) estimating the global rotations by
125
+ rotation averaging (RA), 2) obtaining the global positions
126
+ by TA, and 3) triangulating 3D points and performing a
127
+ final global BA. Govindu [29] represents rotations by lie-
128
+ algebra, and global rotations and global positions are es-
129
+ timated simultaneously. Chatterjee and Govindu [30], [31]
130
+ improve the rotation estimation of
131
+ [29] by a robust l1
132
+ initialization followed by a refinement of the rotations with
133
+ iteratively reweighted least-squares (IRLS) [32]. To solve the
134
+ TA problem, Wilson et al [33] project relative translations
135
+ onto the 1D space to identify outliers. Relative translations
136
+ that are inconsistent with the translation directions that have
137
+ the highest consensus are removed. A nonlinear least-squares
138
+ problem is then solved to get the global positions. Goldstein
139
+ et al. [34] relax the scale constraints of
140
+ [33] to linear
141
+ scale factors, and the convex linear programming problem is
142
+ solved by ADMM [35]. ¨Ozyesil and Singer [12] utilize the
143
+ parallel rigidity theory to select the images where positions
144
+ can be estimated uniquely and solved as a constrained
145
+ quadratic programming problem. By minimizing the sin θ
146
+ between two relative translations, Zhuang et al. [36] improve
147
+ the insensitivity to narrow baselines of TA. The robustness of
148
+ TA is also improved in [36] by incorporating global rotations.
149
+ Hybrid SfM. Cui et al. [37] obtain orientations by RA
150
+ and then register camera centers incrementally with the
151
+ perspective-2-point (P2P) algorithm. Bhomick et al. [16]
152
+ propose to divide the scene graph, where the graph is built
153
+ from the similarity scores between images. Feature matching
154
+ and local SfM can then be executed in parallel and local
155
+ reconstructions are merged [16]. Zhu et al. [18], [19] adopt
156
+ a similar strategy to divide the scene and the graph is
157
+ constructed after feature matching. The relative poses are
158
+ collected after merging all local incremental reconstruction
159
+ results. The outliers are filtered during local reconstruction,
160
+ global rotations are fixed by RA, and camera centers are reg-
161
+ istered with TA at the cluster level. Based on [18], Chen et al.
162
+ [17] find the minimum spanning tree (MST) to solve the final
163
+ merging step. The MST is constructed at the cluster level,
164
+ and the most accurate similarity transformations between
165
+ clusters are given by the MST. Locher et al. [38] filtered
166
+ wrong epipolar geometries by RA before applying the divide-
167
+ and-conquer method [18]. Jiang et al. [39] use a visual-
168
+ inertial navigation system (VINS)
169
+ [40] to first estimate
170
+ the camera trajectories with loop detection and loop closure
171
+ [41]. Images are then divided into sequences according to
172
+ timestamps. However, [39] requires two carefully designed
173
+ systems: one for VINS with loop detection and the other for
174
+ SfM. Loop detection is also a challenge in real-world scenes.
175
+ III. NOTATIONS
176
+ We denote the absolute camera poses as P = {Pi =
177
+ [Ri|ti]}, where Ri, ti are the rotation and translation of the
178
+ i-th image, respectively. The absolute camera poses project
179
+ 3D points X = {Xk} from the world frame to the camera
180
+ frame. The camera centers are denoted by {Ci}. The relative
181
+ pose from image i to image j are denoted as Pij = [Rij|tij],
182
+
183
+ IMU +
184
+ Images
185
+ Wheel Encoders
186
+ Global Alignment
187
+ Global SfM
188
+ Local SfM
189
+ Local SfM
190
+ Global
191
+ View Graph
192
+ Graph
193
+ Bundle Adjustment
194
+ :
195
+ Partition
196
+ Matches
197
+ Refinement
198
+ Local SfM
199
+ Retriangulationwhere Rij, tij are the relative rotations and translations,
200
+ respectively. We define the view graph as G = {V, E}, where
201
+ V denotes the collection of images and E denotes the two
202
+ view geometries, i.e. the relative poses and inlier matches
203
+ between the image pairs. For two rotations Ri, Rj, we use
204
+ log(Ri, Rj) = log(RjR⊤
205
+ i ) to denote the angular error and
206
+ ∥Ri − Rj∥F to denote the chordal distance. Additionally,
207
+ the keypoints and the normalized keypoints after applying
208
+ the intrinsic matrix K are denoted by u and ˆu, respectively.
209
+ IV. COARSE GLOBAL TO FINE INCREMENTAL SFM
210
+ In this section, we introduce our method in detail. In
211
+ Sec. IV-A, we introduce our global SfM that can effectively
212
+ cope with outliers in challenging scenes. A refinement step
213
+ is also introduced to remove outlier matches after global
214
+ SfM. In Sec. IV-B, we describe our parallel incremental SfM
215
+ approach that utilizes the results from coarse global SfM to
216
+ mitigate the problems from sparse view graphs.
217
+ A. Coarse Global SfM
218
+ We first obtain the absolute rotations Ri by solving the
219
+ rotation averaging problem:
220
+ arg min
221
+ { ˆ
222
+ Ri}
223
+
224
+ i∈V,
225
+ (i,j)∈E
226
+ d( ˆRj ˆR⊤
227
+ i , Rij),
228
+ (1)
229
+ where ˆRi denotes the absolute poses obtained by rotation
230
+ averaging, and d(·) = ∥ · ∥F denotes the chordal distance.
231
+ Eq. (1) can be solved robustly and efficiently by [42]. We
232
+ then obtain the absolute camera positions by solving the
233
+ translation averaging problem. However, existing translation
234
+ averaging methods often fail to recover the camera positions
235
+ under challenging scenes due to two main factors: 1) The
236
+ high ratio of outliers in the relative translations. 2) The
237
+ view graph is solvable only when the parallel rigid graph
238
+ condition [12] is satisfied. To alleviate the first problem, we
239
+ first remove the erroneous matching pairs by checking the
240
+ discrepancy of relative rotations: log(R⊤
241
+ ij ˆRj ˆR⊤
242
+ i ) > ϵR, and
243
+ then the relative translations [12] are refined in parallel by:
244
+ arg min
245
+ tij
246
+ ∥ˆu′⊤([tij]×( ˆRj ˆR⊤
247
+ i ))ˆu∥,
248
+ s.t.
249
+ ∥tij∥ = 1.
250
+ (2)
251
+ We do not extract the rigid parallel graph [12] to solve the
252
+ scale ambiguities since it is time-consuming to solve poly-
253
+ nomial equations. Furthermore, the state-of-the-art method
254
+ to establish the solvability of a view graph is only limited
255
+ to 90 nodes [43]. We improve the solvability of the view
256
+ graph by augmenting the relative translations in Pt of the
257
+ consecutive frames from the IMU and wheel encoder. We
258
+ do not augment the relative rotations because they are more
259
+ accurate from the image-based two-view geometry. Note that
260
+ errors can accumulate increasingly in the augmented relative
261
+ poses during the motion of the devices due to the bias of
262
+ the accelerometers and gyroscopes in the IMU, and drifts in
263
+ the wheel encoder caused by friction and wheel slippages.
264
+ To circumvent this problem, we only use the relative poses
265
+ where the time difference is below a threshold ϵT .
266
+ Since we obtained the augmented view graph Gaug =
267
+ {V, Eaug}, the rigidity of the original view graph is aug-
268
+ mented and the scale ambiguities of some images can be
269
+ eliminated. We can then further solve the translation averag-
270
+ ing problem below:
271
+ arg min
272
+ ˆ
273
+ Ci,i∈V;
274
+ sij,(i,j)∈Eaug
275
+
276
+ (i,j)∈Eaug
277
+ ∥sij( ˆCi − ˆCj) − R⊤
278
+ j tij∥,
279
+ (3)
280
+ s.t.
281
+ sij ≥ 0,
282
+ ∀(i, j) ∈ Eaug;
283
+
284
+ i∈V
285
+ ˆCi = 0.
286
+ (3) can be solved efficiently and robustly under the l1-
287
+ norm by collecting all the constraints. Note all the relative
288
+ translations are normalized in Eaug. The right of Fig. 3 shows
289
+ our global SfM result by solving (3).
290
+ After translation averaging, we triangulate the 3D points
291
+ and perform an iterative global bundle adjustment to refine
292
+ camera poses. It is worth mentioning that, global SfM can
293
+ generate more tracks than incremental SfM, as its camera
294
+ poses are less accurate and thus it fails to merge some tracks
295
+ that are physically the same. Besides, according to [28],
296
+ tracks are redundant for optimisation. Therefore, we can
297
+ reduce the computation and memory burden with fewer
298
+ tracks. Though a well-designed algorithm may help with
299
+ the selection of tracks, we simply create tracks with a
300
+ stricter threshold: only when the angle between the two rays
301
+ respectively go through the 3D point and the two camera
302
+ centers are larger than 5 degrees, it is deemed as a valid
303
+ track. Note that for numerical stability during optimization,
304
+ the coordinates are normalized after each iteration.
305
+ Fig. 3.
306
+ Comparison of global SfM results. Results from [12] (left) and
307
+ Eq. (3) (right). Red and black colors respectively denote vehicle trajectories
308
+ and sparse point clouds.
309
+ 1) Matches Refinement: The correct camera poses recov-
310
+ ered by our global SfM with the relative poses from the
311
+ low-cost sensors to eliminate the wrong two-view geometry
312
+ estimates can be further utilized to filter out wrong image
313
+ feature matches. For a calibrated camera with known intrin-
314
+ sics, we can recover the essential matrix between images i
315
+ and j from ˆE = [ˆtij]× ˆRij with the absolute rotations ˆRi and
316
+ translations ˆti computed from rotation and translation aver-
317
+ aging. (ˆtij, ˆRij) are computed from ( ˆRi, ˆRj) and (ˆti,ˆtj).
318
+ The true matches ˆu′ ↔ ˆu must satisfy the check on the total
319
+ point-to-epipolar line distance [22] over the two views, i.e.
320
+ d⊥(ˆu, Eˆu′) + d⊥(ˆu′, Eˆu) ≤ ϵM.
321
+ (4)
322
+ d⊥(x, l) gives the shortest distance between a point x and
323
+ a line l. The epipolar lines on the two images are given by
324
+ l = Eˆu′ and l′ = Eˆu. ϵM is the threshold for the check.
325
+ The effectiveness of global SfM to filter wrong matches
326
+ can be seen in Fig. 7. We build a pseudo ground truth by
327
+
328
+ COLMAP [8] to evaluate the accuracy of the global SfM.
329
+ The ratio test is performed after NN by default. Fig. 4 shows
330
+ the inlier ratio distribution after NN+RANSAC and matches
331
+ refinement with relative poses obtained from global SfM
332
+ and incremental SfM, respectively. Table. I gives the relative
333
+ pose estimation AUC of NN+RANSAC and global SfM with
334
+ respect to incremental SfM. It can be seen that our coarse
335
+ global SfM can obtain comparable accuracy to COLMAP [8]
336
+ in the refinement of the matches.
337
+ Fig. 4.
338
+ Inlier ratio distribution of NN+RANSAC, global SfM and
339
+ incremental SfM (ground truth) on the 711 (left) and B6 (right) datasets.
340
+ AUC
341
+ NN+RANSAC
342
+ Global SfM
343
+ NN+RANSAC
344
+ Global SfM
345
+ R
346
+ t
347
+ R
348
+ t
349
+ R
350
+ t
351
+ R
352
+ t
353
+ @0.1◦
354
+ 1.52
355
+ 0.01
356
+ 6.67
357
+ 0.02
358
+ 2.14
359
+ 0.01
360
+ 8.41
361
+ 0.09
362
+ @0.5◦
363
+ 14.74
364
+ 0.25
365
+ 44.87
366
+ 0.48
367
+ 21.47
368
+ 0.36
369
+ 44.14
370
+ 1.96
371
+ @1.0◦
372
+ 28.92
373
+ 0.96
374
+ 64.15
375
+ 1.80
376
+ 40.99
377
+ 1.40
378
+ 64.48
379
+ 6.48
380
+ @3.0◦
381
+ 55.75
382
+ 5.85
383
+ 84.76
384
+ 9.60
385
+ 68.08
386
+ 9.18
387
+ 86.89
388
+ 24.00
389
+ @5.0◦
390
+ 68.27
391
+ 10.94
392
+ 90.34
393
+ 17.71
394
+ 77.39
395
+ 17.58
396
+ 92.06
397
+ 35.41
398
+ @10.0◦
399
+ 81.71
400
+ 20.21
401
+ 94.99
402
+ 33.03
403
+ 86.81
404
+ 32.46
405
+ 96.01
406
+ 51.07
407
+ @20.0◦
408
+ 90.29
409
+ 29.97
410
+ 97.48
411
+ 49.87
412
+ 92.90
413
+ 46.95
414
+ 98.00
415
+ 64.65
416
+ TABLE I
417
+ RELATIVE POSE ESTIMATION AUC OF NN+RANSAC AND GLOBAL
418
+ SFM WITH RESPECT TO INCREMENTAL SFM ON THE B6 (COLUMN 2-5)
419
+ AND 711 (COLUMN 6-9) DATASETS.
420
+ B. Finer Parallel Incremental SfM
421
+ Although we have obtained the absolute camera poses by
422
+ global SfM, these coarse poses are not accurate enough for
423
+ localization. To improve the accuracy, we propose to refine
424
+ the camera poses and scene structure with the divide-and-
425
+ conquer incremental SfM.
426
+ 1) Adaptive Graph Partition: Existing approaches [18],
427
+ [17] used a cut-and-expand schema to create overlapping
428
+ areas between partitions. However, these approaches have
429
+ two main drawbacks: : 1) The overlapping areas are not
430
+ enough for final merging when the view graph becomes too
431
+ sparse. This can be seen from Fig. 5(a). Edges (3, 20), (7,
432
+ 9), (8, 9), (8, 20), (16, 19), (17, 18) are collected after the
433
+ graph cut, and then the images on these edges are added
434
+ as separators of the partitions. In Fig. 5(a), only images
435
+ {3, 7, 8, 9, 16, 17, 18, 19, 20} can be used to create the over-
436
+ lapping areas (Fig. 5(b)). However, these separator images
437
+ are insufficient to compute the similarity transformations for
438
+ merging all local reconstructions due to the sparsity of the
439
+ view graph. 2) Graph cut tends to separate partitions along
440
+ edges with weak associations. This means the separators are
441
+ often weakly constrained during reconstruction and thus their
442
+ poses might not be accurate enough during reconstruction.
443
+ We propose a flood-fill graph partition algorithm to over-
444
+ come the above-mentioned disadvantages. We refer to the
445
+ added nodes in each cluster after an expansion operation as
446
+ a layer. The separators are collected to form a layer after
447
+ the graph cut on the complete view graph. Fig. 5(a) shows
448
+ examples of the separators marked green. We have separators
449
+ S1 = {{3, 7, 8}, {9, 16, 17}, {18, 19, 20}} in the first layer.
450
+ We then collect all the adjacent images of every separator
451
+ for each partition. We find one adjacent image that does
452
+ not belong to partition k, and add it to the second layer
453
+ of separators S2 in partition k. Adjacent images are sorted
454
+ in descending order according to the weights of the edges,
455
+ i.e. the number of inlier matches. Fig. 5(b) shows that the
456
+ separators S2 = {{9, 20}, {8, 18}, {8, 16}} at the second
457
+ layer after traversing all separators in S1. The expansion step
458
+ is repeated until the number of overlapping images reaches
459
+ the overlapping threshold τot (e.g. 30%).Fig. 5(c) shows the
460
+ separators S3 at the third layer.
461
+ 2) Local Incremental SfM: We perform incremental SfM
462
+ in parallel after graph partitioning. For local incremental
463
+ SfM, we utilize the result of global SfM ˆPglobal to improve
464
+ the robustness of the image registration step, and to further
465
+ constrain the camera poses during global optimization.
466
+ a) Image Registration: We follow [8] for the two-view
467
+ initialization. We then select a batch of the next-best images
468
+ to register, where any image that sees at least vp scene points
469
+ are put into one batch and sorted in descending order. For
470
+ each candidate image i, we first use the P3P [44] to compute
471
+ the initial pose Pp3p
472
+ i
473
+ . However, images can be registered
474
+ wrongly due to wrong matches or scene degeneration. We
475
+ propose to also compute the image pose Pgb
476
+ i
477
+ = [Rgb
478
+ i
479
+ | tgb
480
+ i ]
481
+ using ˆPglobal. We first collect the set of registered images that
482
+ are co-visible to image i, and then the rotation of image i
483
+ can be computed by a single rotation averaging [45]:
484
+ arg min
485
+ Rgb
486
+ i
487
+
488
+ k
489
+ ∥ log( ˆRkiRk, Rgb
490
+ i )∥,
491
+ where
492
+ ˆRki = ˆRi ˆR⊤
493
+ k , (5)
494
+ where k is the index of images that are co-visible to image
495
+ i. For image translation, we first compute the translation
496
+ of image i by each co-visible image and simply adopt the
497
+ median of each dimension in translations tgb
498
+ i :
499
+ tgb
500
+ i = median{ˆtki + ˆRkitk},
501
+ where
502
+ ˆtki = ˆti − ˆRkiˆtk.
503
+ (6)
504
+ To select the best initial pose, we reproject all visible 3D
505
+ points of image i to compute the reprojection errors and mark
506
+ the 3D point with the reprojection error less than 8px as an
507
+ inlier. Finally, we select the pose which has the most inliers.
508
+ b) Bundle Adjustment: To alleviate the drift problem
509
+ for local incremental SfM, we perform global optimization
510
+ using the classical bundle adjustment with the absolute
511
+ poses obtained from global SfM as the supervision for the
512
+ incrementally registered poses, i.e.
513
+ arg min
514
+ R,C,X
515
+ � �
516
+ i
517
+
518
+ k
519
+ ∥Π(Ri, Ci, Xk) − uik∥
520
+ +
521
+ (7)
522
+
523
+ (i,j)∈Eaug
524
+
525
+ ∥ log(Rij, ˆRij∥ + d∠(tij,ˆtij)
526
+ ��
527
+ ,
528
+
529
+ 30k
530
+ Ground Truth
531
+ GlobalSfM
532
+ NN ransac
533
+ 25k
534
+ 20k
535
+ 15k
536
+ 10k
537
+ 5k
538
+ Ok
539
+ 0.75
540
+ 0.80
541
+ 0.85
542
+ 0.90
543
+ 0.95
544
+ 1.00
545
+ inlier ratioGround Truth
546
+ 60k
547
+ GlobalSfM
548
+ NN ransac
549
+ 50k
550
+ 40k
551
+ 30k
552
+ 20k
553
+ 10k
554
+ ok
555
+ 0.75
556
+ 0.80
557
+ 0.85
558
+ 0.90
559
+ 0.95
560
+ 1.00
561
+ inlier ratio(a) Initial graph cut.
562
+ (b) The 1st graph expansion.
563
+ (c) The 2nd graph expansion.
564
+ Fig. 5.
565
+ Pipeline of adaptive flood-fill graph partition. In the view graph, nodes are denoted by blue circles, edges are denoted by blue solid lines.
566
+ Separators are marked by green circles.
567
+ Fig. 6.
568
+ Vehicle trajectories of different threshold trials when merging sub-reconstructions. The last figure is obtained by our method which starts
569
+ from an initial inlier threshold τinit. Others are the results of using a fixed threshold during the alignment to merge all local reconstructions.
570
+ where Π(·) reprojects a 3D point back to the image plane,
571
+ d∠(·) denotes the angle between two vectors. Note that we do
572
+ not make the hard constraint to force the translation part of
573
+ ˆP−1
574
+ ij Pij to be a zero-vector. Instead, we use d∠(tij,ˆtij) =
575
+ d∠(Ci−Cj, ˆCi− ˆCj) to constrain the translation direction of
576
+ camera poses. This is because the absolute positions obtained
577
+ from global SfM are not sufficiently accurate.
578
+ 3) Adaptive Global Alignment: The global alignment step
579
+ is crucial for the divide-and-conquer SfM since a wrong
580
+ similarity transformation can cause catastrophic failure of
581
+ the reconstruction. The difficulties in estimating a reliable
582
+ similarity transformation are due to 1) The existence of
583
+ outliers in registered camera poses. Although the outliers can
584
+ be identified by RANSAC [46], the threshold that indicates
585
+ outliers is hard to determine. This is due to the loss of the
586
+ absolute scale of the real world in SfM without additional
587
+ information such as GPS. It indicates that the optimal outlier
588
+ threshold varies for each cluster. 2) The estimated similarity
589
+ transformation can overfit wrongly with insufficient sample
590
+ points. Existing divide-and-conquer methods
591
+ [16], [18],
592
+ [19], [47], [17] suffer from the two issues because the
593
+ similarity transformations can only be estimated from the
594
+ overlapping areas between the pairwise local partitions.
595
+ To tackle the first issue, we propose an adaptive strategy
596
+ to determine the inlier threshold τinlier. Given an initial inlier
597
+ threshold τinit, we first estimate the similarity transformation
598
+ by RANSAC [46]. We then compute the inlier ratio rinlier and
599
+ increase the inlier threshold if rinlier < rmin. Furthermore,
600
+ we decrease the threshold if rinlier ≥ rmax to prevent the
601
+ threshold from becoming too large. A large threshold allows
602
+ more outliers to be falsely selected and thus harming the
603
+ similarity transformation estimation. The second issue can be
604
+ solved easily within our framework. We set the coordinate
605
+ frame of the global SfM as the reference frame, and align
606
+ each local SfM into the reference frame. Therefore, for each
607
+ partition, we can have as many sample points as the number
608
+ of common registered images between a global SfM and a
609
+ local partition to compute the similarity transformation. We
610
+ also show the effectiveness of the algorithm to merge local
611
+ reconstructions in Fig. 6. When zooming in, we can observe
612
+ that our adaptive strategy perfectly closed the loop while
613
+ other fixed threshold trials failed.
614
+ V. EXPERIMENTAL RESULTS
615
+ In this section, we perform extensive experiments to
616
+ demonstrate the accuracy, efficiency, and robustness of our
617
+ proposed methods.
618
+ A. Implementation Details
619
+ We use HFNet [48] as the default feature extractor and
620
+ use the NN search for matching. A maximum of 500 feature
621
+ points are extracted from each image and matched to the top
622
+ 30 most similar images based on the global descriptors from
623
+ HFNet. We assume cameras are pre-calibrated and use the
624
+ ceres-solver [49] for bundle adjustment. We did not compare
625
+ our method against [39], as VINs [40] fails to find the right
626
+ loops in our datasets. All methods are run on the same
627
+ computer with 40 CPU cores and 96 GB RAM.
628
+ Evaluation Datasets: We evaluate our method on our self-
629
+ collected outdoor datasets and the 4seasons [15] datasets.
630
+ Our self-collected datasets are collected by low-speed au-
631
+ tonomous mowers, of which the running environments have
632
+ many plants and texture-less areas. The 4seasons dataset is
633
+ a cross-season dataset that includes multi-sensor data such
634
+ as IMU, GNSS, and stereo images. It also provides camera
635
+ poses computed by VI-Stereo-DSO [50], [51] and ground-
636
+ truth camera poses by fusing multi-sensor data into a SLAM
637
+ system. See our attached video for a more qualitative and
638
+ quantitative evaluation of the 4Seasons dataset.
639
+
640
+ 1
641
+ 8
642
+ 9
643
+ 16
644
+ 21
645
+ 18
646
+ 231
647
+ 2
648
+ 8
649
+ 10
650
+ 16
651
+ 16
652
+ 2J
653
+ 13
654
+ 18
655
+ 231
656
+ 2
657
+ 8
658
+ 8
659
+ 10
660
+ 16
661
+ 15
662
+ 18
663
+ 13
664
+ 23
665
+ 18Tinlier = 0.5
666
+ Tinlier = 1.0
667
+ Tinlier = 1.5
668
+ Tinlier = 2.0
669
+ Tinit = 1.0Fig. 7.
670
+ Vehicle trajectories after match refinement on B6 dataset. In Fig.(a) and Fig.(b), the visual results are respectively reconstructed without (left)
671
+ and with (right) match refinement in each sub-figure. Fig.(c) shows some of the wrong matching pairs that are filtered by our method.
672
+ Dataset
673
+ N
674
+ COLMAP [8]
675
+ GraphSfM [17]
676
+ Ours(Global SfM)
677
+ Ours(Global+Inc.)
678
+ Nc
679
+ Np
680
+ ¯L
681
+ RMSE
682
+ T
683
+ Nc
684
+ Np
685
+ ¯L
686
+ RMSE
687
+ T
688
+ Nc
689
+ Np
690
+ ¯L
691
+ T
692
+ Nc
693
+ Np
694
+ ¯L
695
+ RMSE
696
+ T
697
+ high free
698
+ 48,753
699
+ 48,733
700
+ 567,030
701
+ 21.59
702
+ 1.47
703
+ 597,171
704
+ 48,491
705
+ 540,711
706
+ 22.73
707
+ 1.38
708
+ 88,896 (×6.7 ↑)
709
+ 48,758
710
+ 521,080
711
+ 14.51
712
+ 5,177
713
+ 48,694
714
+ 540,942
715
+ 22.79
716
+ 1.66
717
+ 105,163 (×5.7 ↑)
718
+ 711
719
+ 29,619
720
+ 27,175
721
+ 303,352
722
+ 25.35
723
+ 1.64
724
+ 160,322
725
+ 29,618
726
+ 259,292
727
+ 33.37
728
+ 1.46
729
+ 33,514 (×4.8 ↑)
730
+ 29,629
731
+ 249,673
732
+ 18.86
733
+ 3,499
734
+ 29,619
735
+ 256,495
736
+ 33.79
737
+ 1.61
738
+ 38,682 (×4.1 ↑)
739
+ yht
740
+ 7,472
741
+ 7,470
742
+ 90,437
743
+ 20.81
744
+ 1.16
745
+ 20,428
746
+ 6,709
747
+ 78,659
748
+ 20.58
749
+ 1.17
750
+ 7,526 (×2.7 ↑)
751
+ 7,472
752
+ 132,167
753
+ 13.67
754
+ 524
755
+ 7,472
756
+ 108,711
757
+ 17.35
758
+ 1.43
759
+ 9,778 (×2.1 ↑)
760
+ A4
761
+ 5,184
762
+ 5,132
763
+ 33,694
764
+ 41.92
765
+ 1.69
766
+ 18,104
767
+ 4,285
768
+ 28,726
769
+ 49.79
770
+ 1.55
771
+ 12,670 (×1.4 ↑)
772
+ 5,184
773
+ 24,193
774
+ 26.59
775
+ 1,349
776
+ 5,184
777
+ 34,007
778
+ 48.30
779
+ 1.43
780
+ 6,924 (×2.6 ↑)
781
+ Htbd
782
+ 14,651
783
+ 14,645
784
+ 231,870
785
+ 24.62
786
+ 1.30
787
+ 56,888
788
+ 14,645
789
+ 232,441
790
+ 24.25
791
+ 1.37
792
+ 17,187 (×3.3 ↑)
793
+ 14,646
794
+ 190,904
795
+ 23.47
796
+ 1,523
797
+ 14,646
798
+ 238,035
799
+ 23.76
800
+ 1.36
801
+ 16,852 (×3.4 ↑)
802
+ jy1
803
+ 32,484
804
+ 32,463
805
+ 534,117
806
+ 20.57
807
+ 1.44
808
+ 346,161
809
+ 32,466
810
+ 536,331
811
+ 20.18
812
+ 1.52
813
+ 28,673 (×12.1 ↑)
814
+ 32,484
815
+ 463052
816
+ 16.12
817
+ 3,077
818
+ 32,466
819
+ 621,437
820
+ 17.77
821
+ 1.53
822
+ 33,555 (×10.3 ↑)
823
+ TABLE II
824
+ COMPARISON OF RUNTIME AND ACCURACY ON REAL-WORLD DATASETS. FOR RUNTIME T (SECONDS), THE FIRST, SECOND AND THIRD THE BEST
825
+ RESULTS ARE HIGHLIGHTED IN COLOR. Nc, Np DENOTE THE NUMBER OF REGISTERED IMAGES AND 3D POINTS, RESPECTIVELY, ¯L DENOTES THE
826
+ AVERAGE TRACK LENGTH , AND RMSE DENOTES THE ROOT MEAN SQUARE ERROR IN PIXEL.
827
+ Running Parameters: Empirically, we use the time
828
+ threshold ϵT = 500 ms to adopt the fused relative poses
829
+ in Gaug, and ϵR = 5 degree to check to relative rotation
830
+ discrepancy. The point-to-epipolar line distance is ϵM =
831
+ 4 px. Besides, we set the overlapping ratio τot = 0.3 in
832
+ the graph partition, vp = 10 for an image to be a candidate
833
+ to register, and rmin = 0.7, rmax = 0.9, τinit = 1.0, αinc =
834
+ 0.2, αdec = 0.1 in global alignment.
835
+ B. How Matching Refinement Saves SfM?
836
+ In addition to running our experiments on HFNet, we
837
+ also do evaluations on different trials. We first show the
838
+ reconstruction results conducted on a challenging scene in
839
+ Fig. 7, which is difficult for visual methods to identify the
840
+ wrong feature matches due to specular issues.
841
+ We use two different combinations of methods for feature
842
+ extraction and matching in each scene. In the first combi-
843
+ nation, we use HFNet [48] for feature extraction and NN
844
+ search for feature matching. In the second combination, we
845
+ use Superpoint [52] for feature extraction and Superglue [53]
846
+ for feature matching. Both settings use RANSAC
847
+ [46] to
848
+ remove matching outliers that do not satisfy the point-to-
849
+ epipolar line constraint. In each sub-figure, the left and right
850
+ images are the results without and with matching refinement,
851
+ respectively. It can be seen that for HFNet + NN, while both
852
+ methods fail to reconstruct the two datasets, the result after
853
+ our result is visually better than without matches refinement.
854
+ For Superpoint + Superglue, the state-of-the-art methods
855
+ respectively on feature extraction and matching, also fails
856
+ on the dataset without refining matches. In contrast, our
857
+ method can correctly identify the wrong matching pairs and
858
+ then leverage the refined matchings to greatly improve the
859
+ reconstruction quality for both settings.
860
+ C. Qualitative Evaluation on Real-World Datasets
861
+ We evaluated our full pipeline on several outdoor datasets.
862
+ We use the registered images number Nc, the recovered 3D
863
+ points Np, the average track length ¯L, and the root mean
864
+ square error (RMSE) to evaluate the qualitative accuracy. As
865
+ shown in Table. II, our method shows the most number of
866
+ registered images in almost all the datasets, while [17] shows
867
+ the least number of registered images. In terms of efficiency,
868
+ our method is moderately slower than GraphSfM [17] in
869
+ most datasets since our method requires an additional global
870
+ SfM reconstruction step. Interestingly, GraphSfM [17] is
871
+ almost 1× slower than our method on the A4 dataset. We
872
+ conjecture that it is due to the frequent failure of GraphSfM
873
+ in selecting suitable images to register and therefore more
874
+ trials are required to register as many images as possible.
875
+ On the other hand, our method is robust enough to deal with
876
+ the case since we get the initial poses of the images from
877
+ P3P or global SfM. Our explanation is validated in Table. II
878
+ where GraphSfM [17] recovers only 4,235 poses out of 5,184
879
+ images, which is almost 20% less than our method. We can
880
+ further notice that the average track length of global SfM is
881
+ remarkably shorter than other methods, which means poses
882
+ from global SfM are not accurate.
883
+ VI. CONCLUSION
884
+ In this paper, we proposed a robust SfM method that
885
+ is adaptive to scenes in different scales and environments.
886
+ Integrating data from low-cost sensors, our initial global
887
+ SfM can benefit from the augmented view graph, where the
888
+ solvability of the original view graph is enhanced. The global
889
+ SfM result is used as a reliable pose prior to improve the
890
+ robustness of the subsequent local incremental SfM and the
891
+ final global alignment steps. Comprehensive experiments on
892
+ different challenging scenes demonstrated the robustness and
893
+ adaptivity of our method, whilst taking more computation
894
+ burden with an additional global SfM step.
895
+ Acknowledgement. This research/project is supported by
896
+ the National Research Foundation, Singapore under its AI
897
+ Singapore Programme (AISG Award No: AISG2-RP-2021-
898
+ 024), and the Tier 2 grant MOE-T2EP20120-0011 from the
899
+ Singapore Ministry of Education.
900
+
901
+ (a) HFNet + NN
902
+ (b) Superpoint + Sup
903
+ erglue
904
+ (c) Filtered matching pairsREFERENCES
905
+ [1] P. Sarlin, A. Unagar, M. Larsson, H. Germain, C. Toft, V. Larsson,
906
+ M. Pollefeys, V. Lepetit, L. Hammarstrand, F. Kahl, and T. Sattler,
907
+ “Back to the feature: Learning robust camera localization from pix-
908
+ els to pose,” in IEEE Conference on Computer Vision and Pattern
909
+ Recognition,, 2021, pp. 3247–3257.
910
+ [2] E. Brachmann, M. Humenberger, C. Rother, and T. Sattler, “On
911
+ the limits of pseudo ground truth in visual camera re-localisation,”
912
+ in IEEE/CVF International Conference on Computer Vision (ICCV),
913
+ 2021, pp. 6218–6228.
914
+ [3] M. Dusmanu, O. Miksik, J. L. Sch¨onberger, and M. Pollefeys, “Cross-
915
+ descriptor visual localization and mapping,” in IEEE/CVF Interna-
916
+ tional Conference on Computer Vision (ICCV), 2021, pp. 6058–6067.
917
+ [4] Y. Furukawa and C. Hern´andez, “Multi-view stereo: A tutorial,” Found.
918
+ Trends Comput. Graph. Vis., vol. 9, no. 1-2, pp. 1–148, 2015.
919
+ [5] Y. Yao, Z. Luo, S. Li, T. Fang, and L. Quan, “Mvsnet: Depth inference
920
+ for unstructured multi-view stereo,” in Computer Vision - ECCV 2018
921
+ - 15th European Conference, vol. 11212, 2018, pp. 785–801.
922
+ [6] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor-
923
+ thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields
924
+ for view synthesis,” in Computer Vision - ECCV 2020 - 16th European
925
+ Conference, vol. 12346, 2020, pp. 405–421.
926
+ [7] S. Agarwal, Y. Furukawa, N. Snavely, I. Simon, B. Curless, S. M.
927
+ Seitz, and R. Szeliski, “Building rome in a day,” Commun. ACM,
928
+ vol. 54, no. 10, pp. 105–112, 2011.
929
+ [8] J. L. Sch¨onberger and J. Frahm, “Structure-from-motion revisited,” in
930
+ IEEE Conference on Computer Vision and Pattern Recognition, 2016,
931
+ pp. 4104–4113.
932
+ [9] P. Moulon, P. Monasse, and R. Marlet, “Global fusion of relative
933
+ motions for robust, accurate and scalable structure from motion,” in
934
+ IEEE International Conference on Computer Vision, 2013, pp. 3248–
935
+ 3255.
936
+ [10] Z. Cui and P. Tan, “Global structure-from-motion by similarity aver-
937
+ aging,” in IEEE International Conference on Computer Vision, 2015,
938
+ pp. 864–872.
939
+ [11] C. Sweeney, T. Sattler, T. H¨ollerer, M. Turk, and M. Pollefeys,
940
+ “Optimizing the viewing graph for structure-from-motion,” in IEEE
941
+ International Conference on Computer Vision, 2015, pp. 801–809.
942
+ [12] O. ¨Ozyesil and A. Singer, “Robust camera location estimation by
943
+ convex programming,” in IEEE Conference on Computer Vision and
944
+ Pattern Recognition, 2015, pp. 2674–2683.
945
+ [13] V. M. Govindu, “Combining two-view constraints for motion es-
946
+ timation,” in IEEE Conference on Computer Vision and Pattern
947
+ Recognition, 2001, pp. 218–225.
948
+ [14] K. Wilson and N. Snavely, “Robust global translations with 1DSfM,”
949
+ in European Conference on Computer Vision, vol. 8691, 2014, pp.
950
+ 61–75.
951
+ [15] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg,
952
+ N. Zeller, and D. Cremers, “4seasons: A cross-season dataset for multi-
953
+ weather SLAM in autonomous driving,” in Pattern Recognition - 42nd
954
+ DAGM German Conference, DAGM GCPR 2020, ser. Lecture Notes
955
+ in Computer Science, vol. 12544.
956
+ Springer, 2020, pp. 404–417.
957
+ [16] B. Bhowmick, S. Patra, A. Chatterjee, V. M. Govindu, and S. Banerjee,
958
+ “Divide and conquer: Efficient large-scale structure from motion using
959
+ graph partitioning,” in Asian Conference on Computer Vision, 2014,
960
+ pp. 273–287.
961
+ [17] Y. Chen, S. Shen, Y. Chen, and G. Wang, “Graph-based parallel large
962
+ scale structure from motion,” Pattern Recognition, vol. 107, p. 107537,
963
+ 2020.
964
+ [18] S. Zhu, T. Shen, L. Zhou, R. Zhang, J. Wang, T. Fang, and L. Quan,
965
+ “Parallel structure from motion from local increment to global aver-
966
+ aging,” arXiv: 1702.08601, 2017.
967
+ [19] S. Zhu, R. Zhang, L. Zhou, T. Shen, T. Fang, P. Tan, and L. Quan,
968
+ “Very large-scale global SfM by distributed motion averaging,” in
969
+ IEEE Conference on Computer Vision and Pattern Recognition, 2018,
970
+ pp. 4568–4577.
971
+ [20] C. Forster, L. Carlone, F. Dellaert, and D. Scaramuzza, “IMU preinte-
972
+ gration on manifold for efficient visual-inertial maximum-a-posteriori
973
+ estimation,” in Robotics: Science and Systems XI,, L. E. Kavraki,
974
+ D. Hsu, and J. Buchli, Eds., 2015.
975
+ [21] K. Wu, C. X. Guo, G. A. Georgiou, and S. I. Roumeliotis, “VINS on
976
+ wheels,” in IEEE International Conference on Robotics and Automa-
977
+ tion,, 2017, pp. 5155–5162.
978
+ [22] R. Hartley and A. Zisserman, Multiple View Geometry in Computer
979
+ Vision.
980
+ Cambridge University Press, 2004.
981
+ [23] J. Shi and J. Malik, “Normalized cuts and image segmentation,” IEEE
982
+ Trans. Pattern Anal. Mach. Intell., vol. 22, no. 8, pp. 888–905, 2000.
983
+ [24] R. J. Lipton, D. J. Rose, and R. E. Tarjan, “Generalized nested
984
+ dissection,” SIAM Journal on Numerical Analysis, vol. 16, no. 2, pp.
985
+ 346—-358, 1979.
986
+ [25] S. Agarwal, N. Snavely, S. M. Seitz, and R. Szeliski, “Bundle
987
+ adjustment in the large,” in Computer Vision - ECCV 2010, 11th
988
+ European Conference on Computer Vision, 2010, pp. 29–42.
989
+ [26] B. Triggs, P. F. McLauchlan, R. I. Hartley, and A. W. Fitzgibbon,
990
+ “Bundle adjustment - A modern synthesis,” in Vision Algorithms:
991
+ Theory and Practice, International Workshop on Vision Algorithms,
992
+ 1999, pp. 298–372.
993
+ [27] C. Wu, “Towards linear-time incremental structure from motion,” in
994
+ International Conference on 3D Vision, 2013, pp. 127–134.
995
+ [28] H. Cui, S. Shen, X. Gao, and Z. Hu, “Batched incremental structure-
996
+ from-motion,” in International Conference on 3D Vision, 2017, pp.
997
+ 205–214.
998
+ [29] V. M. Govindu, “Lie-algebraic averaging for globally consistent mo-
999
+ tion estimation,” in IEEE Conference on Computer Vision and Pattern
1000
+ Recognition, 2004, pp. 684–691.
1001
+ [30] A. Chatterjee and V. M. Govindu, “Efficient and robust large-scale
1002
+ rotation averaging,” in IEEE International Conference on Computer
1003
+ Vision, 2013, pp. 521–528.
1004
+ [31] ——, “Robust relative rotation averaging,” IEEE Trans. Pattern Anal.
1005
+ Mach. Intell., vol. 40, no. 4, pp. 958–972, 2018.
1006
+ [32] H. P. W and W. R. E, “Robust regression using iteratively reweighted
1007
+ least-squares,” Communications in Statistics-theory and Methods,
1008
+ vol. 9, no. 6, pp. 813–827, 1977.
1009
+ [33] K. Wilson, D. Bindel, and N. Snavely, “When is rotations averaging
1010
+ hard?” in Computer Vision - ECCV 2016 - 14th European Conference,
1011
+ vol. 9911, 2016, pp. 255–270.
1012
+ [34] T. Goldstein, P. Hand, C. Lee, V. Voroninski, and S. Soatto, “Shapefit
1013
+ and shapekick for robust, scalable structure from motion,” in European
1014
+ Conference on Computer Vision, 2016, pp. 289–304.
1015
+ [35] S. P. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
1016
+ optimization and statistical learning via the alternating direction
1017
+ method of multipliers,” Foundations and Trends in Machine Learning,
1018
+ vol. 3, no. 1, pp. 1–122, 2011.
1019
+ [36] B. Zhuang, L. Cheong, and G. H. Lee, “Baseline desensitizing in
1020
+ translation averaging,” in IEEE Conference on Computer Vision and
1021
+ Pattern Recognition, 2018, pp. 4539–4547.
1022
+ [37] H. Cui, X. Gao, S. Shen, and Z. Hu, “Hsfm: Hybrid structure-
1023
+ from-motion,” in IEEE Conference on Computer Vision and Pattern
1024
+ Recognition, 2017, pp. 2393–2402.
1025
+ [38] A. Locher, M. Havlena, and L. V. Gool, “Progressive structure
1026
+ from motion,” in Computer Vision - ECCV 2018 - 15th European
1027
+ Conference, 2018, pp. 22–38.
1028
+ [39] Z. Jiang, H. Taira, N. Miyashita, and M. Okutomi, “Vio-aided structure
1029
+ from motion under challenging environments,” in 22nd IEEE Interna-
1030
+ tional Conference on Industrial Technology, 2021, pp. 950–957.
1031
+ [40] T. Qin, P. Li, and S. Shen, “Vins-mono: A robust and versatile monocu-
1032
+ lar visual-inertial state estimator,” IEEE Trans. Robotics, vol. 34, no. 4,
1033
+ pp. 1004–1020, 2018.
1034
+ [41] R. Mur-Artal, J. M. M. Montiel, and J. D. Tard´os, “ORB-SLAM:
1035
+ A versatile and accurate monocular SLAM system,” IEEE Trans.
1036
+ Robotics, vol. 31, no. 5, pp. 1147–1163, 2015.
1037
+ [42] Y. Chen, J. Zhao, and L. Kneip, “Hybrid rotation averaging: A fast and
1038
+ robust rotation averaging approach,” in IEEE Conference on Computer
1039
+ Vision and Pattern Recognition, 2021, pp. 10 358–10 367.
1040
+ [43] F. Arrigoni, A. Fusiello, E. Ricci, and T. Pajdla, “Viewing graph solv-
1041
+ ability via cycle consistency,” in IEEE/CVF International Conference
1042
+ on Computer Vision (ICCV), 2021, pp. 5540–5549.
1043
+ [44] L. Kneip, D. Scaramuzza, and R. Siegwart, “A novel parametrization
1044
+ of the perspective-three-point problem for a direct computation of ab-
1045
+ solute camera position and orientation,” in The 24th IEEE Conference
1046
+ on Computer Vision and Pattern Recognition, 2011, pp. 2969–2976.
1047
+ [45] R. I. Hartley, J. Trumpf, Y. Dai, and H. Li, “Rotation averaging,”
1048
+ International Journal of Computer Vision, vol. 103, no. 3, pp. 267–
1049
+ 305, 2013.
1050
+ [46] M. A. Fischler and R. C. Bolles, “Random sample consensus: A
1051
+ paradigm for model fitting with applications to image analysis and
1052
+ automated cartography,” Commun. ACM, vol. 24, no. 6, pp. 381–395,
1053
+ 1981.
1054
+
1055
+ [47] M. Fang, T. Pollok, and C. Qu, “Merge-sfm: Merging partial recon-
1056
+ structions,” in 30th British Machine Vision Conference 2019, 2019,
1057
+ p. 29.
1058
+ [48] P. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk, “From coarse
1059
+ to fine: Robust hierarchical localization at large scale,” in IEEE
1060
+ Conference on Computer Vision and Pattern Recognition, 2019, pp.
1061
+ 12 716–12 725.
1062
+ [49] S. Agarwal and K. Mierle, “Ceres solver,” http://ceres-solver.org.
1063
+ [50] R. Wang, M. Schw¨orer, and D. Cremers, “Stereo DSO: large-scale
1064
+ direct sparse visual odometry with stereo cameras,” in IEEE Inter-
1065
+ national Conference on Computer Vision.
1066
+ IEEE Computer Society,
1067
+ 2017, pp. 3923–3931.
1068
+ [51] L. von Stumberg, V. Usenko, and D. Cremers, “Direct sparse visual-
1069
+ inertial odometry using dynamic marginalization,” in 2018 IEEE
1070
+ International Conference on Robotics and Automation.
1071
+ IEEE, 2018,
1072
+ pp. 2510–2517.
1073
+ [52] D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint: Self-
1074
+ supervised interest point detection and description,” in 2018 IEEE
1075
+ Conference on Computer Vision and Pattern Recognition Workshops,
1076
+ 2018, pp. 224–236.
1077
+ [53] P. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superglue:
1078
+ Learning feature matching with graph neural networks,” in 2020
1079
+ IEEE/CVF Conference on Computer Vision and Pattern Recognition,
1080
+ 2020, pp. 4937–4946.
1081
+
1082
+ VII. APPENDIX
1083
+ A. Adaptive Flood-Fill Graph Partition Algorithm
1084
+ The pseudo-code of our adaptive flood-fill graph partition
1085
+ algorithm is given in Alg. 1.
1086
+ Algorithm 1 Adaptive Flood-Fill Graph Partition Algorithm
1087
+ Input: Initial view graph G = {V, E}, Overlapping thresh-
1088
+ old τot, Partition number K
1089
+ Output: Sub-graphs {Gk = {Vi, Ei} | i ∈ [0, K]}
1090
+ 1: Overlapping ratio τor := 0, Separators Vs := ∅, {Gk} :=
1091
+ GraphCut(G).
1092
+ 2: while τor < τot do
1093
+ 3:
1094
+ Update separators Vs = {Vs
1095
+ 0, · · · , Vs
1096
+ K}.
1097
+ 4:
1098
+ Edges Edis := ∅.
1099
+ 5:
1100
+ for k ∈ [0, K] do
1101
+ 6:
1102
+ Edis
1103
+ k
1104
+ = E − Ek and Edis
1105
+ k
1106
+ contains Vs
1107
+ k.
1108
+ 7:
1109
+ Edis+ = Edis
1110
+ k .
1111
+ 8:
1112
+ Sort Edis by descending order.
1113
+ 9:
1114
+ for Edge e ∈ Edis do
1115
+ 10:
1116
+ Select a partition Gk contains one of the nodes
1117
+ in e and has the smallest size.
1118
+ 11:
1119
+ Add e to Gk.
1120
+ 12:
1121
+ Update τor.
1122
+ B. Adaptive Global Alignment Algorithm
1123
+ The pseudo-code of our adaptive global alignment algo-
1124
+ rithm is given in Alg. 2.
1125
+ Algorithm 2 Adaptive Global Alignment Algorithm
1126
+ Input: Local
1127
+ reconstructions
1128
+ M
1129
+ =
1130
+ {Mi},
1131
+ τinit, rmin, rmax, iterNummax, αinc, αdec
1132
+ Output: Final reconstruction
1133
+ 1: for i < |M| do
1134
+ 2:
1135
+ τinlier := τinit, rinlier := 0, iterNum := 0
1136
+ 3:
1137
+ while rinlier < rmin & iterNum < iterNummax do
1138
+ 4:
1139
+ iterNum := iterNum + 1;
1140
+ 5:
1141
+ Compute sim3 by τinlier;
1142
+ 6:
1143
+ Compute rinlier by sim3;
1144
+ 7:
1145
+ if rinlier < rmin then
1146
+ 8:
1147
+ τinlier := τinlier + αinc;
1148
+ 9:
1149
+ else if rinlier ≥ rmax then
1150
+ 10:
1151
+ τinlier := τinlier − αdec;
1152
+ C. Visualization Results on Self-Collected Dataset
1153
+ The qualitative visualization results are shown in Fig. 8.
1154
+ We can see that our reconstruction results are better than
1155
+ COLMAP [8] and GraphSfM [17], especially when we zoom
1156
+ in to see the image poses. Moreover, GraphSfM [17] fails
1157
+ to correctly merge the sub-reconstructions. The misalignment
1158
+ can be observed from the zoom-in areas of side-view images,
1159
+ which further validates the robustness of our method.
1160
+ D. Ablations of Augmented View Graph
1161
+ We present more ablation of the augmented view graph
1162
+ on the 4Seasons dataset in Fig. 9. More visualization results
1163
+ on this dataset can be seen in our attached video.
1164
+ E. Quantitative Results on 4Seasons dataset.
1165
+ We present the quantitative results on the 4Seasons dataset
1166
+ in Table. III. The 4Seasons dataset provides ground truth
1167
+ camera poses and trajectories from VI-Stereo-DSO [50],
1168
+ [51]. The sensor data contain IMU, GNSS, and stereo
1169
+ images. In our experiment, we do not use the GNSS data.
1170
+ Besides, as this dataset does not provide wheel encode data,
1171
+ we perturb the VI-Stereo-DSO trajectories by Gaussian noise
1172
+ in the x-y-z axes to synthesize wheel encoder data. We
1173
+ strongly recommend readers refer to [15] for more details
1174
+ about the challenged dataset. As is expected, Our method
1175
+ outperforms COLMAP by a large margin in terms of both
1176
+ accuracy and efficiency. In the Old Town scene, COLMAP
1177
+ failed to reconstruct on sequence recording 2020-10-08 11-
1178
+ 53-41 and sequence recording 2021-02-25 12-34-08 (we use
1179
+ - to denote the failed cases). As the two sequences contain
1180
+ severe motion blur and tunnels in images, which makes them
1181
+ very challenging to reconstruct. However, our method is also
1182
+ robust to these scenes since it can robustly fuse different
1183
+ sensor data.
1184
+
1185
+ (a) Qualititve comparison on the 711 dataset.
1186
+ (b) Qualititve comparison on the A4 dataset.
1187
+ (c) Qualititve comparison on the high free dataset.
1188
+ Fig. 8.
1189
+ Reconstruction comparisons on our self-collected dataset. From left to right are the input images, top-view reconstruction, and side-view
1190
+ reconstruction.
1191
+
1192
+ COLMAP
1193
+ SJnoCOLMAI
1194
+ s.inoCOLMAP
1195
+ sJIn.Fig. 9.
1196
+ Ablations of our augmented view graph on the 4Seasons dataset.
1197
+ Scene
1198
+ Sequence
1199
+ COLMAP [8]
1200
+ Ours (Global SfM)
1201
+ Ours (final)
1202
+ Nc
1203
+ Np
1204
+ ∆R
1205
+ ∆t
1206
+ T
1207
+ Nc
1208
+ Np
1209
+ ∆R
1210
+ ∆t
1211
+ T
1212
+ Nc
1213
+ Np
1214
+ ∆R
1215
+ ∆t
1216
+ T
1217
+ Neighborhood
1218
+ recording 2020-10-07 14-53-52
1219
+ 6,326
1220
+ 137,135
1221
+ 0.65
1222
+ 1.78
1223
+ 334.90
1224
+ 6,036
1225
+ 66,777
1226
+ 2.52
1227
+ 1.17
1228
+ 14.68
1229
+ 6,033
1230
+ 109,483
1231
+ 0.74
1232
+ 0.52
1233
+ 123.96
1234
+ recording 2020-12-22 11-54-24
1235
+ 6,518
1236
+ 127,892
1237
+ 0.55
1238
+ 3.68
1239
+ 354.35
1240
+ 6,144
1241
+ 64,405
1242
+ 1.10
1243
+ 0.86
1244
+ 15.83
1245
+ 6,144
1246
+ 102,857
1247
+ 0.51
1248
+ 0.62
1249
+ 151.88
1250
+ recording 2020-03-26 13-32-55
1251
+ 7,414
1252
+ 148,848
1253
+ 0.61
1254
+ 1.24
1255
+ 603.13
1256
+ 5,982
1257
+ 70,066
1258
+ 0.92
1259
+ 0.79
1260
+ 17.10
1261
+ 5,982
1262
+ 111,807
1263
+ 1.11
1264
+ 0.98
1265
+ 157.76
1266
+ recording 2020-10-07 14-47-51
1267
+ 6,688
1268
+ 152,307
1269
+ 0.56
1270
+ 1.67
1271
+ 359.03
1272
+ 6,248
1273
+ 76,305
1274
+ 2.20
1275
+ 1.17
1276
+ 15.70
1277
+ 6,248
1278
+ 121,657
1279
+ 0.75
1280
+ 0.74
1281
+ 152.85
1282
+ recording 2021-02-25 13-25-15
1283
+ 6,174
1284
+ 138,807
1285
+ 0.75
1286
+ 1.05
1287
+ 325.65
1288
+ 5,238
1289
+ 62,879
1290
+ 1.00
1291
+ 1.14
1292
+ 15.12
1293
+ 5,238
1294
+ 106,609
1295
+ 0.46
1296
+ 0.81
1297
+ 202.85
1298
+ recording 2021-05-10 18-02-12
1299
+ 7,784
1300
+ 149,528
1301
+ 3.04
1302
+ 9.57
1303
+ 444.85
1304
+ 5,834
1305
+ 61,889
1306
+ 1.49
1307
+ 1.38
1308
+ 12.76
1309
+ 5,834
1310
+ 101,102
1311
+ 0.47
1312
+ 0.59
1313
+ 153.36
1314
+ recording 2021-05-10 18-32-32
1315
+ 7,174
1316
+ 141,864
1317
+ 2.77
1318
+ 19.15
1319
+ 416.34
1320
+ 6,046
1321
+ 89,010
1322
+ 1.14
1323
+ 1.03
1324
+ 23.81
1325
+ 6,046
1326
+ 142,430
1327
+ 1.49
1328
+ 1.34
1329
+ 264.75
1330
+ Business Park
1331
+ recording 2021-01-07 13-12-23
1332
+ 8,016
1333
+ 109,399
1334
+ 0.72
1335
+ 0.75
1336
+ 643.22
1337
+ 9,010
1338
+ 72,096
1339
+ 1.76
1340
+ 1.60
1341
+ 56.16
1342
+ 9,010
1343
+ 100,057
1344
+ 0.66
1345
+ 0.51
1346
+ 465.34
1347
+ recording 2020-10-08 09-30-57
1348
+ 11,520
1349
+ 127,013
1350
+ 0.37
1351
+ 1.57
1352
+ 1284.44
1353
+ 8,278
1354
+ 66,087
1355
+ 1.59
1356
+ 1.51
1357
+ 48.72
1358
+ 8,278
1359
+ 108,000
1360
+ 0.63
1361
+ 0.45
1362
+ 366.81
1363
+ recording 2021-02-25 14-16-43
1364
+ 7,414
1365
+ 148,848
1366
+ 0.61
1367
+ 1.24
1368
+ 603.13
1369
+ 5,982
1370
+ 70,066
1371
+ 0.92
1372
+ 0.79
1373
+ 17.10
1374
+ 5,982
1375
+ 111,807
1376
+ 1.11
1377
+ 0.98
1378
+ 157.76
1379
+ Old Town
1380
+ recording 2020-10-08 11-53-41
1381
+ 19,332
1382
+ 279,989
1383
+ -
1384
+ -
1385
+ 2454
1386
+ 12,910
1387
+ 181,569
1388
+ 2.23
1389
+ 2.81
1390
+ 45.72
1391
+ 12,048
1392
+ 279,127
1393
+ 0.55
1394
+ 0.56
1395
+ 254.71
1396
+ recording 2021-01-07 10-49-45
1397
+ 16.420
1398
+ 307,383
1399
+ 8.63
1400
+ 360.51
1401
+ 1496.6
1402
+ 12,728
1403
+ 194,340
1404
+ 2.56
1405
+ 3.14
1406
+ 53.18
1407
+ 12,728
1408
+ 327,348
1409
+ 1.55
1410
+ 1.03
1411
+ 238.82
1412
+ recording 2021-02-25 12-34-08
1413
+ 18,950
1414
+ 305,461
1415
+ -
1416
+ -
1417
+ 2392.98
1418
+ 12,387
1419
+ 182,940
1420
+ 2.02
1421
+ 3.14
1422
+ 40.97
1423
+ 12,387
1424
+ 302,833
1425
+ 0.63
1426
+ 0.74
1427
+ 683.97
1428
+ Office Loop
1429
+ recording 2020-03-24 17-36-22
1430
+ 10,188
1431
+ 209,942
1432
+ 1.17
1433
+ 3.40
1434
+ 822.38
1435
+ 9,522
1436
+ 126,680
1437
+ 2.28
1438
+ 2.38
1439
+ 31.87
1440
+ 9,377
1441
+ 214,285
1442
+ 0.97
1443
+ 0.98
1444
+ 166.54
1445
+ recording 2020-03-24 17-45-31
1446
+ 8,582
1447
+ 195,738
1448
+ 0.92
1449
+ 3.04
1450
+ 865.48
1451
+ 9,186
1452
+ 122,713
1453
+ 2.79
1454
+ 2.20
1455
+ 33.91
1456
+ 8,940
1457
+ 205,790
1458
+ 0.84
1459
+ 0.85
1460
+ 209.06
1461
+ recording 2020-04-07 10-20-31
1462
+ 10,350
1463
+ 223.649
1464
+ 4.22
1465
+ 42.44
1466
+ 795.68
1467
+ 10,184
1468
+ 138,446
1469
+ 2.53
1470
+ 1.78
1471
+ 39.83
1472
+ 10,184
1473
+ 224,499
1474
+ 1.47
1475
+ 1.14
1476
+ 253.24
1477
+ recording 2020-06-12 10-10-57
1478
+ 9,990
1479
+ 236,593
1480
+ 18.97
1481
+ 83.94
1482
+ 705.93
1483
+ 10,150
1484
+ 164,062
1485
+ 1.92
1486
+ 1.61
1487
+ 37.32
1488
+ 10,150
1489
+ 246,516
1490
+ 0.76
1491
+ 0.87
1492
+ 206.48
1493
+ recording 2021-01-07 12-04-03
1494
+ 9,164
1495
+ 475,950
1496
+ 0.71
1497
+ 2.58
1498
+ 1000.75
1499
+ 10,300
1500
+ 143,715
1501
+ 3.32
1502
+ 2.39
1503
+ 48.68
1504
+ 10,300
1505
+ 223,676
1506
+ 1.08
1507
+ 0.67
1508
+ 249.42
1509
+ recording 2021-02-25 13-51-57
1510
+ 9,574
1511
+ 214,695
1512
+ 0.84
1513
+ 2.84
1514
+ 773.32
1515
+ 9,426
1516
+ 122,746
1517
+ 3.80
1518
+ 2.68
1519
+ 28.96
1520
+ 9,426
1521
+ 204,289
1522
+ 1.01
1523
+ 0.91
1524
+ 173.29
1525
+ TABLE III
1526
+ COMPARISON OF RUNTIME AND ACCURACY ON THE 4SEASONS DATASETS. T DENOTES THE RUNTIME (IN MINUTES), Nc, Np DENOTE THE NUMBER
1527
+ OF REGISTERED IMAGES AND 3D POINTS, RESPECTIVELY, ∆R, ∆t DENOTES THE MEAN ROTATION ERROR (IN DEGREES) AND TRANSLATION ERROR
1528
+ (IN METERS), RESPECTIVELY, AND WE HIGHLIGHT THE BEST RESULTS IN BOLD.
1529
+
1530
+ Global SfM from raw view graph
1531
+ gran
09FLT4oBgHgl3EQfpS-5/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
39FAT4oBgHgl3EQfEhxj/content/2301.08422v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9ea56b7095d238958fdc2250cc92d4c70547ffbd4e90730d69f7ba3cdf28f05
3
+ size 31803993
3NE1T4oBgHgl3EQf5wVj/content/tmp_files/2301.03515v1.pdf.txt ADDED
@@ -0,0 +1,1635 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MNRAS 000, 1–10 (2022)
2
+ Preprint 10 January 2023
3
+ Compiled using MNRAS LATEX style file v3.0
4
+ Identifying meteorite droppers among the population of bright ’sporadic’
5
+ bolides imaged by the Spanish Fireball Network during the spring of 2022
6
+ E. Peña-Asensio,1,2★ J. M. Trigo-Rodríguez,2,3 A. Rimola,1 M. Corretgé-Gilart,4 and D. Koschny5
7
+ 1Departament de Química, Universitat Autònoma de Barcelona 08193 Bellaterra, Catalonia, Spain
8
+ 2Institut de Ciències de l’Espai (ICE, CSIC), Campus UAB, C/ de Can Magrans s/n, 08193 Cerdanyola del Vallès, Catalonia, Spain
9
+ 3Institut d’Estudis Espacials de Catalunya (IEEC), 08034 Barcelona, Catalonia, Spain
10
+ 4Universitat Politècnica de Catalunya (UPC), Carrer de Jordi Girona, 31, 08034 Barcelona, Spain
11
+ 5TU Munich, Boltzmannstrasse 15, 85748 Garching, Germany
12
+ Accepted XXX. Received YYY; in original form ZZZ
13
+ ABSTRACT
14
+ We take advantage of the extraordinary weather conditions available between February and March 2022 over Spain to analyze
15
+ the brightest fireballs recorded by the monitoring stations of the Spanish Meteor Network (SPMN). We study the atmospheric
16
+ flight of 15 large meteoroids to determine if they are meteorite dropper events to prepare campaigns to search for freshly fallen
17
+ extraterrestrial material. We investigate their origins in the Solar System and their dynamic association with parent bodies and
18
+ meteoroid streams. Employing our Python pipeline 3D-FireTOC, we reconstruct the atmospheric trajectory utilizing ground-
19
+ based multi-station observations and compute the heliocentric orbit. In addition, we applied an ablation model to estimate the
20
+ initial and terminal mass of each event. Using a dissimilarity criterion and propagating backward in time, we check the connection
21
+ of these meteoroids with known complexes and near-Earth objects. We also calculate if the orbits are compatible with recent
22
+ meteoroid ejections. We find that ∼27% of these fireballs are dynamically associated with minor meteoroid streams and exhibit
23
+ physical properties of cometary bodies, as well as one associated with a near-Earth asteroid. We identify two meteorite-producing
24
+ events; however, the on-site search was unsuccessful. By considering that these fireballs are mostly produced by cm-sized rocks
25
+ that might be the fragmentation product of much larger meteoroids, our findings emphasize the idea that the population of
26
+ near-Earth objects is a source of near-term impact hazards, existing large Earth-colliding meteoroids in the known complexes.
27
+ Key words: meteorites, meteors, meteoroids – comets: general – minor planets, asteroids: general
28
+ 1 INTRODUCTION
29
+ The interplanetary medium is composed of countless millimeter- and
30
+ centimeter-sized objects called meteoroids, some of which eventually
31
+ cross the path of our planet (Brown et al. 2002; Murad & Williams
32
+ 2002; Trigo-Rodríguez 2022). These small bodies are fragments pro-
33
+ duced by the catastrophic disruption or collisions of comets, aster-
34
+ oids, or even impacts on planets (Chapman 2010; Tóth et al. 2011;
35
+ Gritsevich et al. 2012; Trigo-Rodriguez et al. 2014). Due to tidal
36
+ forces and sublimation by high temperatures of the Sun, cometary ag-
37
+ gregates and rubble pile asteroids with efficient disruption processes
38
+ suffer fragmentations in their passage through the perihelion, scat-
39
+ tering meteoroids throughout their orbit that constitute the so-called
40
+ meteoroid streams (also known as meteor showers) (Jenniskens 1994,
41
+ 1998, 2006; Vaubaillon et al. 2019). Some of these meteoroid streams
42
+ have Earth-intersecting orbits, so they are generally repeated in an-
43
+ nual cycles. After experiencing different physical phenomena such as
44
+ orbital perturbations, impacts with other objects, Yarkovsky, YORP,
45
+ or Poynting-Robertson effect, other meteoroids suffer time scale de-
46
+ coherence and end up their space travel impacting on our planet as
47
+ sporadic events, that is, apparently not associated with any known
48
+ ★ E-mail: eloy.pena@uab.cat, eloy.peas@gmail.com
49
+ complex (Olsson-Steel 1986; Bottke et al. 2000; Pauls & Gladman
50
+ 2005; Brož 2006; Koschny et al. 2019).
51
+ The impact of these objects at high velocity with the upper part of
52
+ our atmosphere produces a luminous phase in the visible range due
53
+ to the collision with the atoms of the air and the consequent melt-
54
+ ing, evaporation, and progressive ionization of the meteoroid mate-
55
+ rial (Ceplecha et al. 1998; Silber et al. 2018). This phenomenon is
56
+ known as a meteor and is called a fireball or bolide if its magnitude is
57
+ greater than that of Venus. From the observation and analysis of fire-
58
+ balls with ground-based multi-stations, more than 10 major showers
59
+ have been established (Quadrantids, April Lyrids, 𝜂-Aquarids, South-
60
+ ern Δ-Aquariids, Perseids, Orionids, Taurids, Leonids, Geminids and
61
+ Ursids), that is, meteoroid streams that present activity of more than
62
+ 10-15 meteors per hour (Bagnall 2021). However, there are hundreds
63
+ of minor showers with lower activities as well as near-Earth aster-
64
+ oids, many of them poorly studied, that can produce bright fireballs
65
+ and, therefore, potentially meteorite dropper events, just as being
66
+ a source of impact hazard to the Earth (Voloshchuk & Kashcheev
67
+ 1996; Halliday 1987; Madiedo & Trigo-Rodríguez 2008; Borovička
68
+ et al. 2015; Trigo-Rodríguez et al. 2017; Peña-Asensio et al. 2022).
69
+ The months between January and April are especially relevant
70
+ from the meteor science point of view as meteorite fall rates display
71
+ a peak during the beginning of spring in either hemisphere (Halliday
72
+ & Griffin 1982). Unfortunately, the weather during winter and spring
73
+ © 2022 The Authors
74
+ arXiv:2301.03515v1 [astro-ph.EP] 9 Jan 2023
75
+
76
+ 2
77
+ E. Peña-Asensio et al.
78
+ Table 1. Location of the fireball observation points involved in this work.
79
+ Station
80
+ Name
81
+ Long (◦)
82
+ Lat (◦)
83
+ Alt (m)
84
+ A
85
+ Alpicat
86
+ 0.5568
87
+ 41.6676
88
+ 252
89
+ B
90
+ Barx
91
+ -0.3041
92
+ 39.0146
93
+ 336
94
+ C
95
+ Benicàssim
96
+ 0.0386
97
+ 40.0342
98
+ 15
99
+ D
100
+ Calar Alto
101
+ -2.549
102
+ 37.2212
103
+ 2152
104
+ E
105
+ Cebreros
106
+ -4.3693
107
+ 40.4541
108
+ 700
109
+ F
110
+ Corbera
111
+ 1.8906
112
+ 41.4092
113
+ 501
114
+ G
115
+ Estepa
116
+ -4.8766
117
+ 37.2914
118
+ 537
119
+ H
120
+ GranTeCan
121
+ -17.8919
122
+ 28.7567
123
+ 2267
124
+ I
125
+ La Murta
126
+ -1.6756
127
+ 38.0967
128
+ 469
129
+ J
130
+ Monfragüe
131
+ -6.0108
132
+ 39.7736
133
+ 411
134
+ K
135
+ Morata de Jalón
136
+ -1.4821
137
+ 41.474
138
+ 415
139
+ L
140
+ Olocau
141
+ -0.5363
142
+ 39.6744
143
+ 225
144
+ M
145
+ Playa Blanca
146
+ -13.8241
147
+ 28.8747
148
+ 10
149
+ N
150
+ Puertollano
151
+ -4.1129
152
+ 38.7032
153
+ 697
154
+ O
155
+ Sant Mateu
156
+ 0.1758
157
+ 40.465
158
+ 349
159
+ is usually not helpful for fireball monitoring and clouds generally
160
+ prevent detailed trajectory reconstruction and strewn-field estimates.
161
+ In this sense, the months of February and March 2022 were especially
162
+ clement in the Spanish territory so the Spanish Meteor Network
163
+ (SPMN) has been able to record and analyze several spectacular
164
+ fireballs, many of them associated with minor meteoroid streams
165
+ rather than being sporadic.
166
+ In section 2, we first outline the SPMN network’s current in-
167
+ frastructure that has allowed recording these events with multiple
168
+ stations. We also mention the methodology applied for fireball anal-
169
+ ysis. In section 3, we describe the results of the atmospheric flight
170
+ reconstruction, terminal mass prediction, and heliocentric orbit cal-
171
+ culation. In section 4, we analyze the dynamic associations with
172
+ parent bodies, near-Earth asteroids and comets, and minor and major
173
+ meteoroid streams. In addition, we examined the compatibility of
174
+ these events being recently ejected meteoroids. Finally, we discuss
175
+ the results in section 5 and offer our conclusions in section 6.
176
+ 2 DATA COLLECTION AND METHODOLOGY
177
+ Since its creation in 2005, thanks to the operability of the SPMN net-
178
+ work, the whole sky of continental Spain is monitored full time, the
179
+ last decade also including the Balearic and Canary Islands. Currently,
180
+ a total of 34 stations with charged-coupled device (CCD) video and
181
+ all-sky cameras are operational, some of them equipped with spec-
182
+ trometers. In addition, three forward-scatter detectors monitor radio
183
+ meteors (Trigo-Rodríguez et al. 2004). The stations involved in the
184
+ events analyzed in this work are shown in Table 1, also incorporat-
185
+ ing the recently installed AllSky7 camera at European Space Agency
186
+ Cebreros’ station. This camera array allowed us to record 169 bright
187
+ meteors up to an apparent magnitude of -6 between February and
188
+ March of 2022, from which we selected the 15 largest multi-station
189
+ bolides for analysis.
190
+ New video processing and trajectory calculation techniques allow
191
+ the automation of the analysis process of meteors, bolides, and ar-
192
+ tificial fireballs produced by atmospheric re-entries of human-made
193
+ objects. We developed the 3D-FireTOC Python code that automates
194
+ this study allowing the reconstruction of atmospheric trajectories and
195
+ the calculation of heliocentric orbits from multiple recordings by us-
196
+ ing the intersection of planes method (Peña-Asensio et al. 2021b,a).
197
+ Unlike traditional analytical methods, which solve the orbit by cor-
198
+ recting for zenith attraction and diurnal aberration (Ceplecha 1987),
199
+ we have now implemented the accurate IAS15 high-order N-body
200
+ integrator with an adaptive time step included in the REBOUND
201
+ package to compute the heliocentric orbit (Rein & Spiegel 2015).
202
+ The integrator is based on the RADAU-15 developed in Everhart
203
+ (1985) and has a high performance resolving close encounters. We
204
+ account for the Earth’s and Moon’s oblateness by including the J2
205
+ and J4 gravitational harmonic coefficients thanks to the REBOUNDx
206
+ module (Tamayo et al. 2020).
207
+ For most cases, we performed the astrometric calibration by solv-
208
+ ing the polynomial modification of Borovička (1992) proposed by
209
+ Bannister et al. (2013), which exhibits a better convergence while
210
+ ensuring a very excellent level of uncertainty. To achieve the best fit,
211
+ we use a simplicial homology global optimization algorithm to find
212
+ the absolute minimum (Endres et al. 2018). For recordings with suf-
213
+ ficient background stars, we apply the method proposed in Borovicka
214
+ et al. (1995), which produces even lower errors down to 0.01◦ for
215
+ azimuth and elevation. All calibrations are also cross-checked with
216
+ the quadratic model described in Peña-Asensio et al. (2021b).
217
+ With the mean uncertainties obtained in the astrometry for the
218
+ camera calibration fit, we generate 1,000 clones to perform a Monte
219
+ Carlo simulation following a Gaussian distribution applied to each
220
+ detected point. We propagate every clone backward starting with its
221
+ pre-atmospheric velocity from the beginning of the detected lumi-
222
+ nous phase until they are outside the Earth’s influence, specifically, at
223
+ 10 times the Earth Hill sphere. We then integrate forward to the date
224
+ of impact but without taking into account the gravitational attraction
225
+ of the Earth-Moon system to obtain the osculating orbital elements
226
+ at the time of the detection (referred to the J2000 equinox).
227
+ We further perform a backward integration over 10,000 years eval-
228
+ uating the evolution of an orbital dissimilarity criterion to test the
229
+ dynamic association with parent body candidates. This is necessary
230
+ as the most favorable candidate at the time of impact is not always
231
+ the most reliable because it may be the result of a coincidence at
232
+ that precise date. The meteoroid is integrated with its correspond-
233
+ ing 1,000 clones generated from the uncertainties and the meteoroid
234
+ streams are modeled by 18 equally spaced distributed particles over
235
+ the true anomaly. Based on the orbital dissimilarity criterion, we
236
+ assume that an association is robust enough if it remains below the
237
+ cutoff for 5,000 years, minimizing the probability of being a random
238
+ association (Porubčan et al. 2004).
239
+ Different techniques have been developed and discussed to estab-
240
+ lish the association between meteors and meteor showers or parent
241
+ bodies, and they are still a source of debate today. One of the most
242
+ established and widely used criteria is 𝐷𝐷 (Drummond 1981), which
243
+ is a semi-quantitative approach to measure the dissimilarity of two
244
+ orbits as a function of their orbital parameters in the five-dimensional
245
+ phase.
246
+ Based on the 𝐷𝑆𝐻 criterion (Southworth & Hawkins 1963), the
247
+ 𝐷𝐷 criterion was defined as:
248
+ 𝐷2
249
+ 𝐷 =
250
+ � 𝑒𝐵 − 𝑒𝐴
251
+ 𝑒𝐵 + 𝑒𝐴
252
+ �2
253
+ +
254
+ � 𝑞𝐵 − 𝑞𝐴
255
+ 𝑞𝐵 + 𝑞𝐴
256
+ �2
257
+ +
258
+ � 𝐼𝐵𝐴
259
+ 𝜋
260
+ �2
261
+ +
262
+ +
263
+ � 𝑒𝐵 + 𝑒𝐴
264
+ 2
265
+ �2 � 𝜃𝐵𝐴
266
+ 𝜋
267
+ �2
268
+ ,
269
+ (1)
270
+ where 𝑒 is the eccentricity, 𝑞 is the perihelion distance, 𝐼𝐵𝐴 is
271
+ the angle between the orbital planes, 𝜋𝐵𝐴 is the difference between
272
+ longitudes of perihelia measured from the intersection of both orbits,
273
+ and 𝜃𝐵𝐴 is the orbit angle between the lines of apsides.
274
+ The thresholds of the dissimilarity functions, far from defining an
275
+ exact barrier, offer an approximation with fair statistical significance,
276
+ which, in addition, may vary depending on the inclination of the orbits
277
+ MNRAS 000, 1–10 (2022)
278
+
279
+ Meteorite dropper spring 2022
280
+ 3
281
+ and the population size. Therefore, they are not a defining indicator,
282
+ and it is also necessary to verify that the orbits are not only similar
283
+ at a given time but also that this similarity lasts over time. In this
284
+ sense, we use 0.18 as a cut-off for 𝐷𝐷 (Galligan 2001). Although
285
+ this threshold value is high, we use it as a first filter, but not as the
286
+ only association condition as we also check its evolution over time.
287
+ In addition, we evaluate if the separation of the meteoroid from
288
+ its possible parent body could have occurred in relatively short
289
+ timescales. For this purpose, during the orbital integration, we mon-
290
+ itor the minimum distance between the objects and the change in
291
+ the velocity vector that would be needed to move from one orbit to
292
+ the other one. In this way, we can observe if the velocity change is
293
+ compatible with typical collisional ejection processes between small
294
+ bodies.
295
+ We also examined Tisserand’s parameter with respect to Jupiter
296
+ 𝑇𝑗, which is helpful to determine the evolution of small bodies since
297
+ it remains broadly constant for long periods. It is used to classify
298
+ planet-crossing objects, usually, as Jupiter-family comets (JFCs) if
299
+ 2 < 𝑇𝑗 < 3 and asteroidal when 𝑇𝑗 > 3.
300
+ We evaluate the catastrophic disruption for each event by obtaining
301
+ the ram pressure at peak brightness, that is, the bulk aerodynamic
302
+ strength (𝑠 = 𝜌·𝑣2) accordingly to the U.S. standard atmosphere 1976
303
+ (Bronshten 1981). This parameter is typically used to mechanically
304
+ characterize the meteoroid and to classify the material regarding the
305
+ bulk density. For events that do not present an explosion, we evaluate
306
+ the peak of maximum brightness, thus obtaining only an estimate of
307
+ the lower limit for the composition.
308
+ Additionally, assuming an isothermal atmosphere and applying
309
+ the dynamic third-order time-dependent system for characterizing
310
+ meteor deceleration based on the velocity (𝑣) and the height (ℎ),
311
+ we compute the ballistic coefficient (𝛼) and mass loss parameter (𝛽)
312
+ (Gritsevich & Stulov 2006; Gritsevich 2008, 2009; Gritsevich et al.
313
+ 2012; Turchak & Gritsevich 2014):
314
+ 𝐹𝑖(ℎ𝑖, 𝑣𝑖, 𝛼, 𝛽) = 2𝛼𝑒−ℎ𝑖 − Δ𝑖����−𝛽,
315
+ (2)
316
+ with Δ𝑖 = 𝐸𝑖(𝛽) − 𝐸𝑖(𝛽𝑣2
317
+ 𝑖 ), 𝑖 = 1, 2, ..., 𝑛, where
318
+ 𝐸𝑖(𝑥) =
319
+ ∫ 𝑥
320
+ −∞
321
+ 𝑒𝑡𝑑𝑡
322
+ 𝑡
323
+ 𝑑𝑥.
324
+ These adimensional parameters are defined as
325
+ 𝛼 = 1
326
+ 2𝑐𝑑
327
+ 𝜌0ℎ0𝑆0
328
+ 𝑀0 sin 𝛾 ,
329
+ (3)
330
+ and
331
+ 𝛽 = (1 − 𝜇)
332
+ 𝑐ℎ𝑣2
333
+ 0
334
+ 2𝑐𝑑𝐻∗ ,
335
+ (4)
336
+ where 𝑐𝑑 is the drag coefficient, 𝜌0 is the atmospheric density
337
+ at sea level, ℎ0 is the scale height for a homogeneous atmosphere
338
+ and 𝛾 is the slope of the fireball to the local horizon, 𝑀0 is the
339
+ meteoroid mass before impacting the top of the atmosphere, 𝜇 is
340
+ the dimensionless shape change parameter, 𝑐ℎ is the heat transfer
341
+ coefficient, 𝑣0 is the entry velocity, and 𝐻∗ is the sublimation heat. 𝜇
342
+ is a constant value that relates the cross-sectional area 𝑆 with the mass
343
+ as follows: 𝑆/𝑆0 = (𝑀/𝑀0)𝜇 (Lyytinen & Gritsevich 2016). Note
344
+ that as it is an atmospheric flight dynamics model with an asymptotic
345
+ solution, the minimization problem itself yields an initial velocity at
346
+ infinity that corresponds to the pre-atmospheric velocity.
347
+ These parameters allow properly describing the atmospheric flight
348
+ and estimating the meteor fate based on the so-called 𝛼 − 𝛽 criterion
349
+ (Sansom et al. 2019). The boundaries that delimit the fall likelihood
350
+ (with a terminal mass threshold of 50 g) are determined by the two
351
+ extreme values of the shape change coefficient: 𝜇 = 0 when the
352
+ meteoroid is not spinning and 𝜇 = 2/3 when the meteoroid surface
353
+ is equally ablated due to the rotation.
354
+ From the aerodynamic strength values, we assign a mete-
355
+ oroid bulk density based on Chyba et al. (1993): cometary if
356
+ 𝑠 < 105 𝑃𝑎; carbonaceous if 105 𝑃𝑎 < 𝑠 < 106 𝑃𝑎; rocky if
357
+ 106 𝑃𝑎 < 𝑠 < 107 𝑃𝑎; and rocky-iron if its aerodynamic strength
358
+ is greater than 107 𝑃𝑎. This allows us to fit the object size 𝐷, the
359
+ pre-atmospheric mass 𝑀0, and the terminal mass 𝑀𝑡 (the final mass
360
+ at the end of the luminous atmospheric phase), being
361
+ 𝑀0 =
362
+
363
+ 1
364
+ 2
365
+ 𝑐𝑑 𝐴0𝜌0ℎ0
366
+ 𝛼𝜌2/3
367
+ 𝑚 sin 𝛾
368
+ �3
369
+ ,
370
+ (5)
371
+ where 𝐴0 is the pre-atmospheric shape coefficient.
372
+ The terminal mass can be computed using the last observed veloc-
373
+ ity in the following instant mass equation
374
+ 𝑀(𝑡) = 𝑀0𝑒
375
+
376
+ 𝛽
377
+ 1−𝜇
378
+
379
+ 1−
380
+
381
+ 𝑣 (𝑡)
382
+ 𝑣0
383
+ �2�
384
+ ,
385
+ (6)
386
+ where 𝑣(𝑡) is the instantaneous velocity.
387
+ 3 ATMOSPHERIC FLIGHT AND HELIOCENTRIC ORBIT
388
+ Once the most suitable recordings of each event have been selected,
389
+ and the lenses of each camera have been calibrated to correct distor-
390
+ tions and found the transformation between pixel and position in the
391
+ sky, we can apply the triangulation using the weighted method of the
392
+ intersection of planes for multiple stations to obtain the real position
393
+ of the meteoroid in each frame. Each station recorded the events in a
394
+ single shot, except for the grazing meteoroid SPMN080322, which
395
+ moved out of the field of view. Therefore, we had to combine the
396
+ recordings from two cameras to obtain the complete luminous trail.
397
+ Figure 1 shows a composite of overlapping images of some of the
398
+ events recorded and analyzed in the following section.
399
+ In some images, like the one of the SPMN060222 fireball captured
400
+ in color from Corbera, an intense reddish tone due to the glowing ion-
401
+ ized air can be seen, although further color calibrations are necessary
402
+ for a precise determination of the tone. In the trace drawn during the
403
+ atmospheric flights, it can be seen how several of them show multiple
404
+ brightness peaks, as a result of the rapid rotation and differentiated
405
+ ablation, while others only exhibited a large final flare due to the
406
+ catastrophic disruption. The beginning and ending position, distance
407
+ flight, and direction of the luminous phase for each event are shown
408
+ in Table 2. The initial heights range from ∼ 120 to 83 km and terminal
409
+ heights (before starting the dark flight) range from ∼ 80 to 13 km.
410
+ As expected, the azimuth and slope have a random distribution, with
411
+ the average slope being around 45◦. Note that the slope is measured
412
+ with respect to the local horizon, 0◦ corresponding to a fully grazing
413
+ meteor. In this regard, we see how the event SPMN010322A traveled
414
+ through the atmosphere a notably greater distance than the rest (∼198
415
+ km), its slope being close to 10◦. Event SPMN080322A, although
416
+ also with a shallow slope, underwent a rapid disruption at 70 km
417
+ altitude, which did not allow it to cover a long distance.
418
+ MNRAS 000, 1–10 (2022)
419
+
420
+ 4
421
+ E. Peña-Asensio et al.
422
+ Figure 1. Selection of blended frames of some of the events analyzed in this work: a) SPMN090322C from Calar Alto by José M. Serna García, b) SPMN060222 from Corbera, c) SPMN080222B from Barx, d)
423
+ SPMN220222 from Alpicat, e) SPMN180222 from Estepa, and f) SPMN110222 from Madrid.
424
+ Table 2. Recorded fireballs with the beginning and ending position, flight distance traveled, and direction of the atmospheric flight.
425
+ SPMN code
426
+ Datetime (UTC)
427
+ Stations
428
+ Long0 (◦)
429
+ Lat0 (◦)
430
+ h0 (km)
431
+ Long𝑡 (◦)
432
+ Lat𝑡 (◦)
433
+ h𝑡 (km)
434
+ Distance (km)
435
+ Azimuth (◦)
436
+ Slope (◦)
437
+ 060222
438
+ 2022-02-06 23:03:20
439
+ A,F
440
+ 4.324±0.011
441
+ 42.848±0.004
442
+ 91.3±0.4
443
+ 4.392±0.009
444
+ 42.8570±0.0031
445
+ 69.16±0.26
446
+ 22.9±0.5
447
+ 80±5
448
+ 75±4
449
+ 080222A
450
+ 2022-02-08 01:09:54
451
+ A,B,K
452
+ -2.5529±0.0029
453
+ 41.2470±0.0012
454
+ 101.594±0.028
455
+ -2.5542±0.0029
456
+ 41.6429±0.0014
457
+ 41.06±0.08
458
+ 77.54±0.32
459
+ 359.4±0.5
460
+ 51.33±0.11
461
+ 080222B
462
+ 2022-02-08 23:31:00
463
+ A,B
464
+ 1.1353±0.0010
465
+ 38.9446±0.0008
466
+ 89.16±0.07
467
+ 1.1055±0.0009
468
+ 39.1885±0.0007
469
+ 36.134±0.024
470
+ 60.68±0.22
471
+ 353.97±0.32
472
+ 60.940±0.032
473
+ 110222
474
+ 2022-02-11 02:26:30
475
+ B,E,O
476
+ -3.625±0.009
477
+ 39.702±0.007
478
+ 89.8±0.8
479
+ -3.731±0.005
480
+ 39.455±0.006
481
+ 37.50±0.21
482
+ 65.5±0.5
483
+ 198.7±2.8
484
+ 52.9±1.2
485
+ 140222B
486
+ 2022-02-14 20:59:07
487
+ G,I,N
488
+ -3.5864±0.0014
489
+ 37.8739±0.0004
490
+ 94.646±0.025
491
+ -3.2628±0.0009
492
+ 37.78175±0.00030
493
+ 49.547±0.018
494
+ 60.51±0.22
495
+ 109.69±0.05
496
+ 48.18±0.06
497
+ 180222
498
+ 2022-02-18 01:02:45
499
+ I,J,O
500
+ -6.1642±0.0023
501
+ 39.380±0.004
502
+ 88.79±0.06
503
+ -6.0776±0.0034
504
+ 39.5080±0.0034
505
+ 12.87±0.15
506
+ 82.9±0.8
507
+ 26.8±1.3
508
+ 66.43±0.18
509
+ 220222
510
+ 2022-02-22 04:34:24
511
+ A,K
512
+ -0.5435±0.0010
513
+ 42.3780±0.0005
514
+ 83.77±0.08
515
+ 0.1736±0.0004
516
+ 42.2556±0.0005
517
+ 38.92±0.04
518
+ 80.57±0.13
519
+ 102.42±0.08
520
+ 33.814±0.015
521
+ 010322A
522
+ 2022-03-01 00:48:01
523
+ A,C,L
524
+ 2.6121±0.0034
525
+ 41.3954±0.0020
526
+ 95.79±0.07
527
+ 1.4258±0.0018
528
+ 39.9335±0.0015
529
+ 50.499±0.024
530
+ 197.0±0.4
531
+ 211.99±0.07
532
+ 13.293±0.024
533
+ 010322B
534
+ 2022-03-01 01:43:57
535
+ B,O
536
+ -2.793±0.008
537
+ 39.9817±0.0019
538
+ 101.07±0.30
539
+ -3.258±0.010
540
+ 39.5159±0.0019
541
+ 71.70±0.24
542
+ 74.2±0.7
543
+ 217.7±1.0
544
+ 23.3±0.5
545
+ 080322A
546
+ 2022-03-08 00:36:59
547
+ A,F
548
+ 0.8633±0.0008
549
+ 40.6421±0.0005
550
+ 96.82±0.08
551
+ 1.7423±0.0006
552
+ 41.00590±0.00030
553
+ 80.13±0.05
554
+ 87.00±0.12
555
+ 60.904±0.032
556
+ 11.06±0.06
557
+ 080322B
558
+ 2022-03-08 19:26:22
559
+ A,L
560
+ 1.8383±0.0005
561
+ 40.4211±0.0005
562
+ 83.58±0.05
563
+ 1.8210±0.0005
564
+ 40.4374±0.0005
565
+ 36.786±0.021
566
+ 57.70±0.18
567
+ 320.613±0.024
568
+ 54.09±0.10
569
+ 090322B
570
+ 2022-03-09 03:01:46
571
+ A,B,C
572
+ -1.5107±0.0010
573
+ 39.8088±0.0004
574
+ 120.65±0.07
575
+ -2.0243±0.0011
576
+ 39.94857±0.00035
577
+ 77.071±0.035
578
+ 71.19±0.13
579
+ 289.42±0.04
580
+ 37.75±0.12
581
+ 090322C
582
+ 2022-03-09 04:25:38
583
+ D,I
584
+ -2.0192±0.0007
585
+ 36.9452±0.0009
586
+ 92.94±0.15
587
+ -2.1597±0.0006
588
+ 36.4849±0.0017
589
+ 58.54±0.10
590
+ 64.80±0.05
591
+ 193.73±0.15
592
+ 32.07±0.23
593
+ 100322
594
+ 2022-03-10 01:38:19
595
+ H,M
596
+ -15.540±0.014
597
+ 30.0550±0.0034
598
+ 85.4±0.8
599
+ -15.600±0.022
600
+ 29.689±0.005
601
+ 29.2±0.6
602
+ 82.5±0.6
603
+ 188±5
604
+ 42.94±0.17
605
+ 120322
606
+ 2022-03-12 22:15:53
607
+ A,L
608
+ 1.1473±0.0004
609
+ 40.7151±0.0007
610
+ 94.21±0.09
611
+ 1.09818±0.00035
612
+ 40.7597±0.0006
613
+ 67.85±0.05
614
+ 27.40±0.16
615
+ 319.4±0.4
616
+ 74.30±0.25
617
+ MNRAS 000, 1–10 (2022)
618
+
619
+ a
620
+ b
621
+ ldaia(
622
+ d)
623
+ ESTEPA-SEVILLA-SPAIN-@AJ_ROBLES
624
+ NORTEUTC2022-02-18 01:02:5Meteorite dropper spring 2022
625
+ 5
626
+ 0 °
627
+ 60 °
628
+ 120 °
629
+ 180 °
630
+ 240 °
631
+ 300 °
632
+ -90 °
633
+ 90 °
634
+ -60 °
635
+ -30 °
636
+ 0 °
637
+ 30 °
638
+ 60 °
639
+ 20
640
+ 40
641
+ 60
642
+ Geocentric velocity (km/s)
643
+ Figure 2. Sinusoidal projection of the geocentric (diamond) and apparent
644
+ (gray cross) radiants. Radiant pairs are connected with a light blue line.
645
+ Geocentric radiants are color-coded according to their geocentric velocity.
646
+ Using the height at which the brightest flare occurs, the air density,
647
+ and the velocity at that point, we calculate the aerodynamic strength.
648
+ According to the value of this dynamic pressure, we estimate the
649
+ bulk density as explained in Section 2, which is used to calculate
650
+ the pre-atmospheric diameter assuming a perfect sphere. To obtain
651
+ the ballistic coefficient and the mass loss parameter, we assume an
652
+ aerodynamic drag coefficient of 1.3 and a shape change coefficient
653
+ of 2/3 (Gritsevich & Koschny 2011). The geocentric velocities range
654
+ from ∼ 63 to 11 km/s, and most of the radiants are in the northern
655
+ hemisphere, as depicted in Figure 2 in sinusoidal projection. All the
656
+ computed parameters are shown in Table 3 and 4.
657
+ Two meteoroids penetrate up to ∼ 30 and 13 km altitude starting
658
+ the dark flight at a velocity of ∼ 8 and 20 km/s, respectively. As can
659
+ be seen in Figure 3, from the application of the 𝛼 − 𝛽 criterion and
660
+ assuming 50 g as the minimum terminal mass to produce a recov-
661
+ erable fall, event SPMN100322 had some possibility of generating
662
+ a meteorite with a mass of ∼140 g, and event SPMN180222 was
663
+ likely to be a ∼430 g meteorite dropper. Unfortunately, a field search
664
+ campaign was prepared but no fragments were recovered.
665
+ The computed osculating orbital elements at the time of impact
666
+ of the analyzed fireballs are compiled in Table 5. As an example of
667
+ the Monte Carlo simulation, Figure 4 shows a heat map of the semi-
668
+ major axis and inclination distribution for the 1,000 clones of event
669
+ SPMN010322A at the time of impact (t=0 year without Earth-Moon
670
+ gravitational focusing correction) and at the end of the backward
671
+ orbital integration (t=-10,000 year).
672
+ Four orbits present very high eccentricity values with large semi-
673
+ major axes, five can be classified as Jupiter-family comets, while
674
+ four are asteroid-like orbits. As expected, the orbits tend to be of
675
+ low inclination, with the exception of SPMN090322B which has an
676
+ inclination of 122◦. None of the meteoroids had close encounters
677
+ with the Moon prior to the impact.
678
+ 4 DYNAMIC ASSOCIATION WITH METEOROID
679
+ STREAMS AND PARENT BODIES
680
+ The study of the associations of meteoroids that impact our planet
681
+ with parent bodies or meteoroid streams is not a trivial task. There
682
+ 1
683
+ 2
684
+ 3
685
+ 4
686
+ 5
687
+ 6
688
+ 7
689
+ 8
690
+ ln( sin )
691
+ 4
692
+ 2
693
+ 2
694
+ 4
695
+ 6
696
+ ln( )
697
+ Likely fall
698
+ Possible fall
699
+ Unlikely fall
700
+ 20
701
+ 30
702
+ 40
703
+ 50
704
+ 60
705
+ 70
706
+ 80
707
+ Terminal height (km)
708
+ Figure 3. Distribution of the 15 fireballs analyzed over the Spanish territory
709
+ during February and March 2022 according to the 𝛼 − 𝛽 criterion. The color
710
+ bar shows the terminal height, the gray solid curve the boundary for a 50
711
+ g meteorite assuming no spin of the meteoroid, and the black solid curve
712
+ the boundary for a 50 g meteorite assuming equal ablation over the entire
713
+ meteoroid surface. We assume 𝜇 = 2/3 for all meteoroids.
714
+ are numerous mechanisms that prevent the correct linking of meteors
715
+ with their origins, from the intrinsically chaotic behavior of plane-
716
+ tary systems to non-gravitational effects and sporadic collisions and
717
+ interactions (Trigo-Rodríguez et al. 2005). Because of the high prob-
718
+ ability that two orbits are randomly associated (Wiegert & Brown
719
+ 2004), we have not only analyzed the similarity of the orbits at the
720
+ time of impact but also studied their robustness over time. From the
721
+ time evolution of the parent body dissimilarity criterion, we found
722
+ some dynamic associations. Figure 5 shows the evolution of the dis-
723
+ similarity criterion during the orbital integration of the 15 events
724
+ analyzed in this work, along with their most favorable parent body
725
+ candidates or meteor shower. Table 6 shows each event with its most
726
+ likely association, along with the years of time it lasts under the 𝐷𝐷
727
+ threshold, the minimum encounter distance, the required ejection ve-
728
+ locity at the time of minimum distance, and the minimum required
729
+ ejection velocity.
730
+ 5 out of 15 events, that is, about 30% of the bright fireballs, are
731
+ below the cut-off for at least 5,000 years. 4 events would be associated
732
+ with minor showers (∼27%) and 1 fireball associated with a near-
733
+ Earth asteroid (∼7%). In all the associated cases, the required ejection
734
+ velocity needed to transform the parent orbit into the meteoroid orbit
735
+ is in good agreement with the estimated range for collisions between
736
+ objects, which can produce a kick of a few kilometers per second
737
+ (Melosh 1984).
738
+ 5 DISCUSSION
739
+ In relation to the various ablation behaviors observed, it is impor-
740
+ tant to note that this could be the result of the differences between
741
+ chondritic meteoroid and cometary aggregate bulk properties. The
742
+ low density and high porosity of the latter are directly related to
743
+ their aerodynamic strengths (Blum et al. 2006). Cometary streams
744
+ typically produce centimeter-sized projectiles causing fireballs with
745
+ disruptive flares, and multiple sudden brightness increases or a catas-
746
+ trophic final flare. Due to the heterogeneity of the meteoroid compo-
747
+ nents, the evaporation temperature of each one is reached at different
748
+ MNRAS 000, 1–10 (2022)
749
+
750
+ 6
751
+ E. Peña-Asensio et al.
752
+ Table 3. Recorded fireballs with aerodynamic strength, ballistic coefficient, mass loss parameter, pre-atmospheric diameter, pre-atmospheric mass, and terminal
753
+ mass.
754
+ SPMN code
755
+ s (kPa)
756
+ 𝛼
757
+ 𝛽
758
+ D (cm)
759
+ M0 (g)
760
+ M𝑡 (g)
761
+ 060222
762
+ 18.9±0.4
763
+ (8.6±0.7)·102
764
+ 10.6±1.0
765
+ 1.17±0.08
766
+ 0.83±0.16
767
+ <1
768
+ 080222A
769
+ 724±7
770
+ 195.1±3.4
771
+ 1.023±0.031
772
+ 6.35±0.10
773
+ 134±7
774
+ 11.3±0.5
775
+ 080222B
776
+ 501.06±0.25
777
+ 79.72±0.27
778
+ 2.244±0.004
779
+ 13.90±0.04
780
+ 1405±13
781
+ 5.363±0.025
782
+ 110222
783
+ 361.6±3.5
784
+ 16.1±1.9
785
+ 11.3±1.7
786
+ 76±8
787
+ (2.3±0.7)·105
788
+ <1
789
+ 140222B
790
+ 78.16±0.15
791
+ 253.7±1.5
792
+ 2.917±0.017
793
+ 5.121±0.027
794
+ 70.3±1.1
795
+ <1
796
+ 180222
797
+ 1107±21
798
+ 11.94±0.23
799
+ 4.70±0.07
800
+ 25.3±0.5
801
+ (2.96±0.16)·104
802
+ 432±34
803
+ 220222
804
+ 283.6±1.8
805
+ 102.2±0.6
806
+ 1.690±0.022
807
+ 17.02±0.10
808
+ (2.58±0.05)·103
809
+ 33.9±1.2
810
+ 010322A
811
+ 209.4±1.1
812
+ 387±6
813
+ 2.13±0.04
814
+ 10.87±0.16
815
+ 673±29
816
+ 2.93±0.17
817
+ 010322B
818
+ 17.17±0.33
819
+ (1.38±0.28)·103
820
+ 18±4
821
+ 1.8±0.4
822
+ 3.1±1.8
823
+ <1
824
+ 080322A
825
+ 2.935±0.019
826
+ 6898±28
827
+ 3.60±0.07
828
+ 0.732±0.006
829
+ 0.205±0.005
830
+ <1
831
+ 080322B
832
+ 364.7±2.2
833
+ 82.9±0.8
834
+ 1.709±0.026
835
+ 14.42±0.16
836
+ (1.57±0.05)·103
837
+ 25.3±0.8
838
+ 090322B
839
+ 25.45±0.10
840
+ 4831±29
841
+ 5.974±0.015
842
+ 0.3274±0.0013
843
+ 0.01838±0.00022
844
+ <1
845
+ 090322C
846
+ (4.185±0.017)·105
847
+ 255±4
848
+ 9.11±0.17
849
+ 7.16±0.08
850
+ 192±7
851
+ 106±20
852
+ 100322
853
+ (1.30±0.14)·103
854
+ 40.3±3.4
855
+ 1.03±0.33
856
+ 10.1±0.8
857
+ (1.9±0.5)·103
858
+ (1.4±0.8)·102
859
+ 120322
860
+ 19.55±0.26
861
+ 82±22
862
+ 140±34
863
+ 12±4
864
+ (1.0±1.1)·103
865
+ <1
866
+ Table 4. Recorded fireballs with right ascension and declination of the radiant, apparent, geocentric, and heliocentric velocities.
867
+ SPMN code
868
+ RA𝑎 (◦)
869
+ Dec𝑎 (◦)
870
+ RA𝑔 (◦)
871
+ Dec𝑔 (◦)
872
+ RAℎ (◦)
873
+ Decℎ (◦)
874
+ V𝑎,0 (km/s)
875
+ V𝑎,𝑡 (km/s)
876
+ V𝑔 (km/s)
877
+ Vℎ (km/s)
878
+ 060222
879
+ 108±4
880
+ 38.4±2.2
881
+ 106±4
882
+ 37.4±2.5
883
+ 65.6±0.9
884
+ 5.8±1.0
885
+ 19.61±0.09
886
+ 11.229±0.033
887
+ 16.32±0.09
888
+ 41.5±0.5
889
+ 080222A
890
+ 153.14±0.33
891
+ 2.69±0.11
892
+ 152.86±0.34
893
+ 1.70±0.12
894
+ 106.92±0.06
895
+ -7.911±0.031
896
+ 37.17±0.24
897
+ 16.329±0.029
898
+ 35.46±0.25
899
+ 39.91±0.33
900
+ 080222B
901
+ 135.60±0.15
902
+ 10.09±0.04
903
+ 135.18±0.16
904
+ 8.23±0.05
905
+ 82.10±0.05
906
+ -4.635±0.004
907
+ 23.475±0.005
908
+ 9.7592±0.0022
909
+ 20.666±0.006
910
+ 37.79±0.04
911
+ 110222
912
+ 211.58±0.34
913
+ 71.8±2.0
914
+ 217.9±1.1
915
+ 74.3±2.3
916
+ 60.9±0.8
917
+ 27.0±1.7
918
+ 20.1±0.4
919
+ 15.89±0.20
920
+ 16.8±0.4
921
+ 35.24±0.31
922
+ 140222B
923
+ 41.35±0.08
924
+ 39.312±0.028
925
+ 30.63±0.09
926
+ 35.822±0.023
927
+ 51.850±0.018
928
+ 5.641±0.017
929
+ 15.068±0.013
930
+ 8.905±0.004
931
+ 10.536±0.020
932
+ 39.904±0.019
933
+ 180222
934
+ 146.3±0.5
935
+ 17.90±0.28
936
+ 145.2±0.5
937
+ 16.39±0.29
938
+ 91.44±0.10
939
+ 1.29±0.07
940
+ 23.95±0.08
941
+ 20.045±0.013
942
+ 21.32±0.09
943
+ 38.79±0.23
944
+ 220222
945
+ 149.508±0.030
946
+ 30.58±0.06
947
+ 140.248±0.028
948
+ 23.63±0.11
949
+ 80.31±0.07
950
+ 2.559±0.022
951
+ 15.60±0.04
952
+ 5.9502±0.0031
953
+ 11.41±0.06
954
+ 35.225±0.034
955
+ 010322A
956
+ 299.19±0.11
957
+ 50.410±0.029
958
+ 304.55±0.12
959
+ 47.034±0.012
960
+ 49.777±0.031
961
+ 34.32±0.04
962
+ 25.56±0.05
963
+ 9.799±0.015
964
+ 22.97±0.05
965
+ 36.477±0.009
966
+ 010322B
967
+ 288.24±0.33
968
+ 54.1±1.1
969
+ 295.0±0.4
970
+ 51.9±1.1
971
+ 56.2±0.5
972
+ 35±4
973
+ 23.83±0.06
974
+ 19.43±0.13
975
+ 21.01±0.07
976
+ 34.99±0.34
977
+ 080322A
978
+ 113.58±0.07
979
+ -13.672±0.026
980
+ 104.83±0.12
981
+ -21.95±0.07
982
+ 83.91±0.06
983
+ -13.79±0.07
984
+ 17.157±0.030
985
+ 13.560±0.006
986
+ 13.40±0.04
987
+ 39.542±0.029
988
+ 080322B
989
+ 121.84±0.05
990
+ 10.43±0.09
991
+ 123.60±0.04
992
+ 6.96±0.07
993
+ 91.87±0.07
994
+ -4.541±0.026
995
+ 18.72±0.06
996
+ 8.266±0.013
997
+ 14.90±0.08
998
+ 41.49±0.06
999
+ 090322B
1000
+ 259.88±0.09
1001
+ 10.98±0.09
1002
+ 260.24±0.09
1003
+ 10.72±0.09
1004
+ 259.93±0.27
1005
+ 57.545±0.015
1006
+ 63.937±0.007
1007
+ 40.138±0.033
1008
+ 62.749±0.008
1009
+ 41.32±0.04
1010
+ 090322C
1011
+ 340.5±1.2
1012
+ 77.58±0.05
1013
+ 6.2±0.9
1014
+ 73.90±0.14
1015
+ 73.418±0.026
1016
+ 18.82±0.17
1017
+ 18.159±0.018
1018
+ 17.96±0.07
1019
+ 14.397±0.024
1020
+ 38.81±0.05
1021
+ 100322
1022
+ 200±12
1023
+ 75.3±1.7
1024
+ 205±17
1025
+ 79.1±1.8
1026
+ 84.42±0.31
1027
+ 24.54±0.33
1028
+ 20.2±0.7
1029
+ 8.358±0.030
1030
+ 17.0±0.9
1031
+ 38.3±1.4
1032
+ 120322
1033
+ 156.87±0.09
1034
+ 28.29±0.26
1035
+ 157.00±0.10
1036
+ 27.05±0.26
1037
+ 104.92±0.23
1038
+ 7±34
1039
+ 21.39±0.15
1040
+ 14.27±0.06
1041
+ 18.27±0.18
1042
+ 40.48±0.09
1043
+ Table 5. Recorded fireballs with semi-major axis, eccentricity, inclination, perihelion distance, argument of the perihelion, ascending node, and Tisserand
1044
+ parameter (referred to the J2000 equinox). Uncertainty for the ascending node is 0.0001◦.
1045
+ SPMN code
1046
+ a (au)
1047
+ e
1048
+ i (◦)
1049
+ q (au)
1050
+ 𝜔 (◦)
1051
+ Ω (◦)
1052
+ T 𝑗
1053
+ 060222
1054
+ 11±5
1055
+ 0.92±0.04
1056
+ 6.1±1.4
1057
+ 0.892±0.011
1058
+ 216.823±0.011
1059
+ 317.8516
1060
+ 1.60±0.21
1061
+ 080222A
1062
+ 4.3±0.6
1063
+ 0.938±0.007
1064
+ 14.69±0.10
1065
+ 0.2658±0.0030
1066
+ 120.6728±0.0030
1067
+ 138.9337
1068
+ 1.77±0.14
1069
+ 080222B
1070
+ 2.395±0.018
1071
+ 0.7285±0.0015
1072
+ 5.47±0.05
1073
+ 0.6504±0.0015
1074
+ 78.9636±0.0015
1075
+ 139.8736
1076
+ 3.015±0.014
1077
+ 110222
1078
+ 1.60±0.06
1079
+ 0.403±0.020
1080
+ 27.2±1.0
1081
+ 0.954±0.006
1082
+ 208.156±0.006
1083
+ 322.0380
1084
+ 4.02±0.11
1085
+ 140222B
1086
+ 4.344±0.033
1087
+ 0.7739±0.0017
1088
+ 5.655±0.010
1089
+ 0.98217±0.00005
1090
+ 170.87497±0.00005
1091
+ 325.8715
1092
+ 2.320±0.008
1093
+ 180222
1094
+ 3.05±0.18
1095
+ 0.783±0.011
1096
+ 1.53±0.13
1097
+ 0.662±0.005
1098
+ 255.517±0.005
1099
+ 329.0900
1100
+ 2.60±0.09
1101
+ 220222
1102
+ 1.605±0.007
1103
+ 0.4698±0.0026
1104
+ 2.68±0.05
1105
+ 0.8510±0.0005
1106
+ 235.6854±0.0005
1107
+ 333.2960
1108
+ 4.089±0.013
1109
+ 010322A
1110
+ 1.9276±0.0027
1111
+ 0.5469±0.0006
1112
+ 36.06±0.10
1113
+ 0.87346±0.00008
1114
+ 131.68867±0.00008
1115
+ 340.1080
1116
+ 3.413±0.004
1117
+ 010322B
1118
+ 1.57±0.06
1119
+ 0.411±0.020
1120
+ 35.36±0.14
1121
+ 0.922±0.007
1122
+ 139.445±0.007
1123
+ 340.1481
1124
+ 4.00±0.12
1125
+ 080322A
1126
+ 3.96±0.04
1127
+ 0.7532±0.0026
1128
+ 13.885±0.012
1129
+ 0.97727±0.00023
1130
+ 15.37103±0.00023
1131
+ 167.1195
1132
+ 2.392±0.012
1133
+ 080322B
1134
+ 13.5±1.0
1135
+ 0.931±0.005
1136
+ 4.68±0.04
1137
+ 0.9330±0.0004
1138
+ 28.9213±0.0004
1139
+ 167.8948
1140
+ 1.570±0.025
1141
+ 090322B
1142
+ 11.2±0.5
1143
+ 0.911±0.004
1144
+ 122.44±0.13
1145
+ 0.99250±0.00009
1146
+ 178.06764±0.00009
1147
+ 348.2148
1148
+ 1.108±0.020
1149
+ 090322C
1150
+ 3.16±0.04
1151
+ 0.688±0.004
1152
+ 18.90±0.07
1153
+ 0.98494±0.00007
1154
+ 168.69848±0.00007
1155
+ 348.2915
1156
+ 2.665±0.019
1157
+ 100322
1158
+ 2.8±0.9
1159
+ 0.65±0.12
1160
+ 24.60±0.17
1161
+ 0.98426±0.00012
1162
+ 192.22661±0.00012
1163
+ 349.1653
1164
+ 2.8±0.6
1165
+ 120322
1166
+ 6.05±0.30
1167
+ 0.862±0.008
1168
+ 7.92±0.04
1169
+ 0.8339±0.0023
1170
+ 229.2663±0.0023
1171
+ 352.0240
1172
+ 1.93±0.04
1173
+ altitudes, giving rise to the so-called differential ablation (Gómez
1174
+ Martín et al. 2017). The aerodynamic overpressure experienced by
1175
+ meteoroids when they fragment allows for estimating their aerody-
1176
+ namic strength. This, in turn, allows for deducing the bulk properties
1177
+ of their meteoroid stream (Kresak 1982; Trigo-Rodríguez & Llorca
1178
+ 2006). These types of large fireballs associated with cometary ves-
1179
+ tiges are the result of rapid disruption in micrometric grains and
1180
+ the sudden ablation of volatile mineral phases driven by the thermal
1181
+ wave in the meteoroid head (Trigo-Rodríguez et al. 2019).
1182
+ Even in such circumstances, it is remarkable that the sporadic con-
1183
+ MNRAS 000, 1–10 (2022)
1184
+
1185
+ Meteorite dropper spring 2022
1186
+ 7
1187
+ Table 6. Most likely parent body and meteoroid stream candidates for each event with the minimum 𝐷𝐷 value, the years that fulfill the 𝐷𝐷 criterion threshold,
1188
+ the minimum encounter distance, the required ejection velocity at the time of minimum distance, and the minimum required ejection velocity during the orbital
1189
+ integration.
1190
+ SPMN code
1191
+ Association
1192
+ D𝑚𝑖𝑛
1193
+ t𝐷 (y)
1194
+ S𝑚𝑖𝑛 (au)
1195
+ V𝑆,𝑚𝑖𝑛 (km/s)
1196
+ V𝑚𝑖𝑛 (km/s)
1197
+ 060222
1198
+ 𝜌 Geminids
1199
+ 0.176
1200
+ 180
1201
+ 0.186
1202
+ 4.7
1203
+ 4.7
1204
+ 080222A
1205
+ o Leonids
1206
+ 0.174
1207
+ 90
1208
+ 0.231
1209
+ 4.6
1210
+ 0.9
1211
+ 080222B
1212
+ Southern 𝛿 Leonids
1213
+ 0.018
1214
+ 8720
1215
+ 0.129
1216
+ 0.8
1217
+ 0.4
1218
+ 110222
1219
+ 𝜔 Cassiopeiids
1220
+ 0.101
1221
+ 10000
1222
+ 0.087
1223
+ 9.6
1224
+ 1.4
1225
+ 140222B
1226
+ March Cassiopeiids
1227
+ 0.121
1228
+ 1610
1229
+ 0.145
1230
+ 10.2
1231
+ 0.5
1232
+ 180222
1233
+ Southern 𝛿 Leonids
1234
+ 0.07
1235
+ 240
1236
+ 0.278
1237
+ 13.2
1238
+ 2.0
1239
+ 220222
1240
+ Northern 𝛼 Leonids
1241
+ 0.09
1242
+ 10000
1243
+ 0.05
1244
+ 5.8
1245
+ 1.4
1246
+ 010322A
1247
+ 2019 CV2
1248
+ 0.099
1249
+ 2640
1250
+ 0.264
1251
+ 6.0
1252
+ 1.7
1253
+ 010322B
1254
+ 2017 FM91
1255
+ 0.092
1256
+ 9990
1257
+ 0.104
1258
+ 6.6
1259
+ 2.3
1260
+ 080322A
1261
+ 2007 DZ40
1262
+ 0.073
1263
+ 800
1264
+ 0.144
1265
+ 3.1
1266
+ 1.1
1267
+ 080322B
1268
+ February Hydrids
1269
+ 0.168
1270
+ 600
1271
+ 0.37
1272
+ 15.9
1273
+ 3.0
1274
+ 090322B
1275
+ 72 Ophiuchids
1276
+ 0.136
1277
+ 9990
1278
+ 0.811
1279
+ 14.3
1280
+ 0.4
1281
+ 090322C
1282
+ March Cassiopeiids
1283
+ 0.084
1284
+ 110
1285
+ 0.34
1286
+ 10.9
1287
+ 0.8
1288
+ 100322
1289
+ 𝜓 Draconids
1290
+ 0.106
1291
+ 2080
1292
+ 0.37
1293
+ 5.1
1294
+ 2.0
1295
+ 120322
1296
+ 𝜆 Leonids
1297
+ 0.125
1298
+ 1300
1299
+ 0.083
1300
+ 7.4
1301
+ 2.4
1302
+ 2.590
1303
+ 2.595
1304
+ 2.600
1305
+ 2.605
1306
+ 2.610
1307
+ 2.615
1308
+ 2.620
1309
+ Semi-major axis (au)
1310
+ 39.8
1311
+ 39.9
1312
+ 40.0
1313
+ 40.1
1314
+ 40.2
1315
+ Inclination (°)
1316
+ 1,000 clones heatmap
1317
+ t=0 year (impact)
1318
+ 1.86
1319
+ 1.88
1320
+ 1.90
1321
+ 1.92
1322
+ 1.94
1323
+ 1.96
1324
+ 1.98
1325
+ Semi-major axis (au)
1326
+ 43.4
1327
+ 43.6
1328
+ 43.8
1329
+ 44.0
1330
+ 44.2
1331
+ 44.4
1332
+ 44.6
1333
+ Inclination (°)
1334
+ 1,000 clones heatmap
1335
+ t=-10,000 year
1336
+ Figure 4. Typical heatmap of the inclination and semi-major axis distribution
1337
+ of the 1,000 clones for the SPMN010322A in the Monte Carlo simulation.
1338
+ The top figure corresponds to the time of impact (t=0 year) without Earth-
1339
+ Moon gravitational focusing correction. The bottom figure corresponds to the
1340
+ end of the backward orbital integration (t=-10,000 years).
1341
+ tribution is not dominant at all. We found a very significant percent-
1342
+ age of bright fireballs dynamically associated with minor showers.
1343
+ Although during the orbital integration there are no very close en-
1344
+ counters despite the reasonable ejection velocities, we must point
1345
+ out that we have propagated 18 particles distributed in true anomaly
1346
+ throughout the orbit of the meteoroid streams, but at their nominal
1347
+ values for the rest of the orbital elements. Due to the orbital perturba-
1348
+ tions accumulated over time and their violent origin, either by tidal
1349
+ forces disruption or catastrophic collisions, the meteoroid streams
1350
+ spread toroidally along their orbit and gradually disperse. Some re-
1351
+ gions even undergo more pronounced decoherence than others due
1352
+ to the gravitational influence of the Earth-Moon system or nearby
1353
+ planets.
1354
+ The minimum ejection velocities calculated to produce the me-
1355
+ teoroid orbit from the parent body have a standard deviation range
1356
+ between 0.16 and 1.4 km/s (with an average standard deviation of
1357
+ 0.4 km/s) for the studied events. Although the ejection velocities
1358
+ found are compatible with collisions of small objects in the inner
1359
+ Solar System, this does not necessarily mean that these meteoroids
1360
+ have separated from their meteoroid stream or parent body recently;
1361
+ we just note it as a feasible possibility due to the usual disruption
1362
+ behavior of crumbling asteroids and comets.
1363
+ Although remarkable, the high number of minor showers produc-
1364
+ ing fireballs should not come as a surprise as such a percentage of me-
1365
+ teors associated with meteoroid streams is not unusual. For example,
1366
+ percentages up to 80% between November and January were already
1367
+ reported belonging to meteor showers (Rao & Murthy 1974). On the
1368
+ other hand, among the 2,401 records studied by Lindblad (1971),
1369
+ apparently, 37% were associated with meteoroid streams. A similar
1370
+ percentage (41%) was found by Southworth & Hawkins (1963). Of
1371
+ the orbits analyzed by Jacchia & Whipple (1961), 65% were linked to
1372
+ a meteor shower. Regarding the Meteorite Observation and Recovery
1373
+ Project (MORP) database, 37% of the fireballs could be associated
1374
+ with meteoroid stream (Halliday et al. 1996). Terentjeva (1990) per-
1375
+ formed a grouping according to event candidates to produce mete-
1376
+ orites, finding that 68% of 554 fireballs studied could be part of a
1377
+ shower. And also in good agreement with the results of this work,
1378
+ Babadjanov (1963) reported that of the 185 meteors studied, 73%
1379
+ appeared to be of cometary origin. Recent studies also show large
1380
+ percentages of meteors associated with meteor showers, for example,
1381
+ MNRAS 000, 1–10 (2022)
1382
+
1383
+ 8
1384
+ E. Peña-Asensio et al.
1385
+ Figure 5. Evolution of the dissimilarity function 𝐷𝐷 of the 15 meteoroids with their most favorable candidates during the orbital backward integration over
1386
+ 10,000. The 1,000 clones of each event are also shown.
1387
+ MNRAS 000, 1–10 (2022)
1388
+
1389
+ 0.6
1390
+ 0.5
1391
+ 0.4
1392
+ D0.3
1393
+ 0.2
1394
+ 0.1
1395
+ SPMN060222 - p Geminids
1396
+ SPMN080222A-
1397
+ 0.0 -
1398
+ 0.6
1399
+ SPMN110222 - w Cassiopeiids
1400
+ 0.5
1401
+ 0.4
1402
+ β0.3
1403
+ 0.2
1404
+ 0.1
1405
+ SPMN140222B - March Cassiopeids
1406
+ SPMN180222 - Southern 6 Leonid
1407
+ 0.0
1408
+ 0.6
1409
+ SPMN220222 - 209 Northern α Leonids
1410
+ SPMN010322A - 2019 CV2
1411
+ SPMN010322B - 2017 FM91
1412
+ 0.5
1413
+ 0.4 -
1414
+ B0.3
1415
+ 0.2
1416
+ 0.1
1417
+ 0.0
1418
+ 0.6 7
1419
+ SPMN090322B - 72 Ophiuchids
1420
+ 0.5
1421
+ 0.4
1422
+ β0.3
1423
+ 0.1
1424
+ SPMN080322A - 2007 DZ40
1425
+ SPMN080322B - February Hydrids
1426
+ 0.0
1427
+ 0.6
1428
+ 2222727
1429
+ 0.5
1430
+ 0.4
1431
+ D0.3
1432
+ 0.2
1433
+ 0.1
1434
+ SPMN090322C - March Cassiopeids
1435
+ SPMN100322 - Draconids
1436
+ SPMN120322 - 入 Leonids
1437
+ 0.0
1438
+ -10000
1439
+ -8000
1440
+ -6000
1441
+ -4000
1442
+ -2000
1443
+ 10000 -8000
1444
+ -6000
1445
+ -4000
1446
+ -2000
1447
+ -10000
1448
+ -8000
1449
+ -6000
1450
+ -4000
1451
+ -2000
1452
+ Years
1453
+ Years
1454
+ YearsMeteorite dropper spring 2022
1455
+ 9
1456
+ 45% in Colas et al. (2020) and 35% in Drolshagen et al. (2021). Re-
1457
+ garding superbolides detected from space, 23% could be associated
1458
+ with meteoroid streams or near-Earth objects (Peña-Asensio et al.
1459
+ 2022).
1460
+ Therefore, as previously studied, it is reasonable to expect that a
1461
+ large percentage of the meteors belong to minor meteoroid streams,
1462
+ but also, as we show in this work, some meteor showers can be a
1463
+ significant source of large projectiles for the Earth and the Moon.
1464
+ 6 CONCLUSION
1465
+ The extraordinary meteorological conditions in Spain during the
1466
+ spring of 2022 have made it possible to obtain high-quality data re-
1467
+ lated to the fireball activity produced, to a large extent, by minor mete-
1468
+ oroid streams. Ground-based multi-station recordings were possible
1469
+ thanks to the ever-increasing atmospheric volume monitored by the
1470
+ SPMN network throughout Spain. We reported 15 bright bolides in
1471
+ February and March, two of them being potential meteorite dropper
1472
+ events. By applying novel computer vision techniques and improved
1473
+ methods of trajectory reconstruction and heliocentric orbit calcula-
1474
+ tion implemented in our software 3D-FireTOC, we have been able
1475
+ to study in detail the atmospheric flight and dynamic association of
1476
+ large cometary and asteroidal projectiles impacting our planet. Based
1477
+ on the trajectory data, we computed the initial and terminal mass, the
1478
+ aerodynamic strength, and the bulk density by means of an ablation
1479
+ model. In consequence, we claim that:
1480
+ • Among the 169 bright meteors recorded during the spring of
1481
+ 2022 in Spain, 2 of them were potentially meteorite dropper events.
1482
+ • We identify the minor showers o Leonids, Southern 𝛿 Leonids,
1483
+ 𝜔 Cassiopeiids, Northern 𝛼 Leonids, and 72 Ophiuchids, and the
1484
+ asteroid 2017 FM91 as sources of large projectiles during February
1485
+ and March.
1486
+ • Nearby meteoroid streams can be efficient producers of large
1487
+ projectiles as they account for the ∼27% of the fireballs.
1488
+ • Near-Earth objects may be a greater source of impact risk than
1489
+ previously thought.
1490
+ • It is needed to extend the study and cataloguing of minor show-
1491
+ ers, since, although they are not very active in terms of the number
1492
+ of meteors, our work indicates that they also produce large bolides
1493
+ annually.
1494
+ • These findings support the idea that certain meteoroid streams
1495
+ associated with comets or asteroids may represent a short-term im-
1496
+ pact hazard.
1497
+ Finally, we think that understanding the origin and mechanisms by
1498
+ which large meteoroids reach the Earth is of great scientific interest
1499
+ due to the possibility of associating complexes and parent bodies with
1500
+ fireballs and, ultimately, meteorites found on Earth and the Moon.
1501
+ The relevance of associations also reverts in outreach, as we can
1502
+ quickly inform the public about the origin of the fireballs reported
1503
+ by eyewitnesses.
1504
+ ACKNOWLEDGEMENTS
1505
+ This project has received funding from the European Research
1506
+ Council (ERC) under the European Union’s Horizon 2020 re-
1507
+ search and innovation programme (grant agreement No. 865657)
1508
+ for the project “Quantum Chemistry on Interstellar Grains”
1509
+ (QUANTUMGRAIN). JMT-R and E.P-A. acknowledge finan-
1510
+ cial support from project PID2021-128062NB-I00 funded by
1511
+ MCIN/AEI/10.13039/501100011033. AR acknowledge financial
1512
+ support from the FEDER/Ministerio de Ciencia e Innovación – Agen-
1513
+ cia Estatal de Investigación (PID2021-126427NB-I00, PI: AR). AR is
1514
+ indebted to DIUE (project 2017SGR1323). Cebreros #AMS81 ESA
1515
+ Ground station belongs to the AllSky7 fireball monitoring project).
1516
+ We also thank all station operators whose continuous dedication
1517
+ have allowed to record these bolides from multiple stations: Jordi
1518
+ Donet Donet, Vicent Ibáñez, Jose M. Serna, Rainer Kresken, Pablo
1519
+ Ramirez Moreta, Carlos Alcaraz, Antonio J. Robles, Ramón López,
1520
+ Agustín Núñez, José A. de los Reyes, Sensi Pastor, Antonio Fernán-
1521
+ dez Sánchez, Antonio Lasala, Álex Gómez, Juan Gómez, Ramón
1522
+ López, Francisco José García Rodríguez and Cesar Guasch Besal-
1523
+ duch.
1524
+ DATA AVAILABILITY
1525
+ The data underlying this article will be shared on reasonable request
1526
+ to the corresponding author.
1527
+ REFERENCES
1528
+ Babadjanov P., 1963, Smithsonian Contributions to Astrophysics, 7, 287
1529
+ Bagnall P. M., 2021, Atlas of Meteor Showers; A Practical Workbook for
1530
+ Meteor Observers. Springer Nature, doi:10.1007/978-3-030-76643-6
1531
+ Bannister S. M., Boucheron L. E., Voelz D. G., 2013, PASP, 125, 1108
1532
+ Blum J., Schräpler R., Davidsson B. J. R., Trigo-Rodríguez J. M., 2006, ApJ,
1533
+ 652, 1768
1534
+ Borovicka J., Spurny P., Keclikova J., 1995, A&AS, 112, 173
1535
+ Borovička J., 1992, Publications of the Astronomical Institute of the
1536
+ Czechoslovak Academy of Sciences, 79
1537
+ Borovička J., Spurný P., Brown P., 2015, in , Asteroids IV. University of
1538
+ Arizona Press, pp 257–280, doi:10.2458/azu_uapress_9780816532131-
1539
+ ch014
1540
+ Bottke William F. J., Rubincam D. P., Burns J. A., 2000, Icarus, 145, 301
1541
+ Bronshten V., 1981, Moscow, Izdatel’stvo Nauka, 1981. 416
1542
+ Brož M., 2006, PhD thesis, Charles University in Prague
1543
+ Brown P., Spalding R. E., ReVelle D. O., Tagliaferri E., Worden S. P., 2002,
1544
+ Nature, 420, 294
1545
+ Ceplecha Z., 1987, Bulletin of the Astronomical Institutes of Czechoslovakia,
1546
+ 38, 222
1547
+ Ceplecha Z., Borovička J., Elford W. G., Revelle D. O., Hawkes R. L.,
1548
+ Porubčan V., Šimek M., 1998, Space Sci. Rev., 84, 327
1549
+ Chapman C. R., 2010, Nature, 463, 305
1550
+ Chyba C. F., Thomas P. J., Zahnle K. J., 1993, Nature, 361, 40
1551
+ Colas F., et al., 2020, A&A, 644, A53
1552
+ Drolshagen E., et al., 2021, A&A, 652, A84
1553
+ Drummond J. D., 1981, Icarus, 45, 545
1554
+ Endres S. C., Sandrock C., Focke W. W., 2018, Journal of Global Optimiza-
1555
+ tion, 72, 181
1556
+ Everhart E., 1985, in Carusi A., Valsecchi G. B., eds, Astrophysics and Space
1557
+ Science Library Vol. 115, IAU Colloq. 83: Dynamics of Comets: Their
1558
+ Origin and Evolution. p. 185, doi:10.1007/978-94-009-5400-7_17
1559
+ Galligan D. P., 2001, MNRAS, 327, 623
1560
+ Gómez Martín J. C., Bones D. L., Carrillo-Sánchez J. D., James A. D.,
1561
+ Trigo-Rodríguez J. M., Fegley B. J., Plane J. M. C., 2017, ApJ, 836, 212
1562
+ Gritsevich M. I., 2008, Solar System Research, 42, 372
1563
+ Gritsevich M. I., 2009, Advances in Space Research, 44, 323
1564
+ Gritsevich M., Koschny D., 2011, Icarus, 212, 877
1565
+ Gritsevich M. I., Stulov V. P., 2006, Solar System Research, 40, 477
1566
+ Gritsevich M. I., Stulov V. P., Turchak L. I., 2012, Cosmic Research, 50, 56
1567
+ Halliday I., 1987, Icarus, 69, 550
1568
+ Halliday I., Griffin A. A., 1982, Meteoritics, 17, 31
1569
+ Halliday I., Griffin A. A., Blackwell A. T., 1996, Meteoritics & Planetary
1570
+ Science, 31, 185
1571
+ MNRAS 000, 1–10 (2022)
1572
+
1573
+ 10
1574
+ E. Peña-Asensio et al.
1575
+ Jacchia L. G., Whipple F. L., 1961, Smithsonian Contributions to Astro-
1576
+ physics, 4, 97
1577
+ Jenniskens P., 1994, A&A, 287, 990
1578
+ Jenniskens P., 1998, Earth, Planets and Space, 50, 555
1579
+ Jenniskens P., 2006, Meteor Showers and their Parent Comets. Cambridge
1580
+ University Press
1581
+ Koschny D., et al., 2019, Space Sci. Rev., 215, 34
1582
+ Kresak L., 1982, in Wilkening L. L., ed., IAU Colloq. 61: Comet Discoveries,
1583
+ Statistics, and Observational Selection. pp 56–82
1584
+ Lindblad B. A., 1971, Smithsonian Contributions to Astrophysics, 12, 1
1585
+ Lyytinen E., Gritsevich M., 2016, Planet. Space Sci., 120, 35
1586
+ Madiedo J. M., Trigo-Rodríguez J. M., 2008, Earth Moon and Planets, 102,
1587
+ 133
1588
+ Melosh H. J., 1984, Icarus, 59, 234
1589
+ Murad E., Williams I. P., 2002, Meteors in the Earth’s Atmosphere. Cam-
1590
+ bridge University Press
1591
+ Olsson-Steel D., 1986, MNRAS, 219, 47
1592
+ Pauls A., Gladman B., 2005, Meteoritics & Planetary Science, 40, 1241
1593
+ Peña-Asensio E., Trigo-Rodríguez J. M., Langbroek M., Rimola A., J. Robles
1594
+ A., 2021a, Astrodynamics, 5, 347
1595
+ Peña-Asensio E., Trigo-Rodríguez J. M., Gritsevich M., Rimola A., 2021b,
1596
+ MNRAS, 504, 4829
1597
+ Peña-Asensio E., Trigo-Rodríguez J. M., Rimola A., 2022, AJ, 164, 76
1598
+ Porubčan V., Williams I. P., Kornoš L., 2004, Earth Moon and Planets, 95,
1599
+ 697
1600
+ Rao M. S., Murthy A. G., 1974, Australian Journal of Physics, 27, 679
1601
+ Rein H., Spiegel D. S., 2015, MNRAS, 446, 1424
1602
+ Sansom E. K., et al., 2019, ApJ, 885, 115
1603
+ Silber E. A., Boslough M., Hocking W. K., Gritsevich M., Whitaker R. W.,
1604
+ 2018, Advances in Space Research, 62, 489
1605
+ Southworth R. B., Hawkins G. S., 1963, Smithsonian Contributions to As-
1606
+ trophysics, 7, 261
1607
+ Tamayo D., Rein H., Shi P., Hernandez D. M., 2020, MNRAS, 491, 2885
1608
+ Terentjeva A. K., 1990, in Lagerkvist C. I., Rickman H., Lindblad B. A., eds,
1609
+ Asteroids, Comets, Meteors III. p. 579
1610
+ Tóth J., Vereš P., Kornoš L., 2011, MNRAS, 415, 1527
1611
+ Trigo-Rodríguez J. M., 2022, Asteroid Impact Risk: Impact Hazard from
1612
+ Asteroids and Comets. Springer Nature
1613
+ Trigo-Rodríguez J. M., Llorca J., 2006, MNRAS, 372, 655
1614
+ Trigo-Rodríguez J. M., et al., 2004, Earth Moon and Planets, 95, 553
1615
+ Trigo-Rodríguez J. M., Betlem H., Lyytinen E., 2005, ApJ, 621, 1146
1616
+ Trigo-Rodriguez J., Madiedo J., Williams I., 2014, in Muinonen K., Penttilä
1617
+ A., Granvik M., Virkki A., Fedorets G., Wilkman O., Kohout T., eds,
1618
+ Asteroids, Comets, Meteors 2014. p. 533
1619
+ Trigo-Rodríguez J. M., Gritsevich M., Palme H., eds, 2017, Dynamic Sources
1620
+ of Contemporary Hazard from Meteoroids and Small Asteroids Astro-
1621
+ physics and Space Science Proceedings Vol. 46, doi:10.1007/978-3-319-
1622
+ 46179-3_2.
1623
+ Trigo-Rodríguez J. M., Rimola A., Tanbakouei S., Soto V. C., Lee M., 2019,
1624
+ Space Sci. Rev., 215, 18
1625
+ Turchak L. I., Gritsevich M. I., 2014, Journal of Theoretical and Applied
1626
+ Mechanics, 44, 15
1627
+ Vaubaillon J., Neslušan L., Sekhar A., Rudawska R., Ryabova G. O., 2019, in
1628
+ Ryabova G. O., Asher D. J., Campbell-Brown M. J., eds, , Meteoroids:
1629
+ Sources of Meteors on Earth and Beyond. Cambridge University Press.,
1630
+ p. 161
1631
+ Voloshchuk Y. I., Kashcheev B. L., 1996, Solar System Research, 30, 480
1632
+ Wiegert P., Brown P., 2004, Earth Moon and Planets, 95, 19
1633
+ This paper has been typeset from a TEX/LATEX file prepared by the author.
1634
+ MNRAS 000, 1–10 (2022)
1635
+
3NE1T4oBgHgl3EQf5wVj/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
3dE2T4oBgHgl3EQfjgcV/content/tmp_files/2301.03968v1.pdf.txt ADDED
@@ -0,0 +1,1082 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Ballistic surface channels in fully in situ defined Bi4Te3 Josephson
2
+ junctions with aluminum contacts
3
+ D. Rosenbach,1, 2, ∗ A. R. Jalil,3, 4 T. W. Schmitt,1, 4 B. Bennemann,3, 2
4
+ G. Mussler,1, 2 P. Schüffelgen,1, 2 D. Grützmacher,1, 2 and Th. Schäpers1, 2
5
+ 1Peter Grünberg Institute (PGI-9), Forschungszentrum Jülich, 52425 Jülich, Germany
6
+ 2JARA-Fundamentals of Future Information Technology, Jülich-Aachen Research Alliance,
7
+ Forschungszentrum Jülich and RWTH Aachen University, Germany
8
+ 3Peter Grünberg Institute (PGI-10),
9
+ Forschungszentrum Jülich, 52425 Jülich, Germany
10
+ 4JARA-FIT Institute Green IT, RWTH Aachen University, 52062 Aachen, Germany
11
+ (Dated: January 11, 2023)
12
+ 1
13
+ arXiv:2301.03968v1 [cond-mat.mes-hall] 10 Jan 2023
14
+
15
+ Abstract
16
+ In this letter we report on the electrical transport properties of Bi4Te3 in a Josephson junction
17
+ geometry using superconducting Al electrodes with a Ti interdiffusion barrier. Bi4Te3 is proposed
18
+ to be a dual topological insulator, for which due to time-reversal and mirror symmetry both a strong
19
+ topological insulator phase as well as a crystalline topological phase co-exist. The formation of a
20
+ supercurrent through the Bi4Te3 layer is explained by a two-step process. First, due to the close
21
+ proximity of the Al/Ti electrodes a superconducting gap is induced within the Bi4Te3 layer right
22
+ below the electrodes. The size of this gap is determined by analysing multiple Andreev reflections
23
+ (MARs) identified within the devices differential resistance at low voltage biases. Second, based
24
+ on the Andreev reflection and reverse Andreev reflection processes a supercurrent establishes in
25
+ the weak link region in between these two proximity coupled regions. Analyses of the temperature
26
+ dependency of both the critical current as well as MARs indicate mostly ballistic supercurrent
27
+ contributions in between the proximitized Bi4Te3 regions even though the material is characterized
28
+ by a semi-metallic bulk phase.
29
+ The presence of these ballistic modes gives indications on the
30
+ topological nature of Bi4Te3.
31
+ I.
32
+ INTRODUCTION
33
+ Hybrid structures of three-dimensional topological insulators and superconductors are
34
+ considered promising building blocks for the realization of topological quantum circuits [1–
35
+ 3]. A crucial optimization parameter is a sufficiently large coupling of the superconductor
36
+ to the topological insulator. In order to probe the proximity coupling strength a Josephson
37
+ junction with a topological insulator weak link bridging two superconducting electrodes
38
+ can be employed [4–6]. By measuring the current-voltage characteristics of these junctions,
39
+ the interface transparency as well as the underlying mode of transport, i.e. diffusive or
40
+ ballistic, can be investigated [7, 8]. The supercurrent in a Josephson junction is carried by
41
+ electron-hole bound states. Based on the nature of these bound states their energy phase
42
+ relation (EΦR) has a fixed periodicity. Irradiating the junction with a radio frequency signal
43
+ allows to investigate the Shapiro step response. As the Josephson voltage V0 = hf/2e in
44
+ ∗ rosenbach@ph2.uni-koeln.de
45
+ present address: Physics Institute II, University of Cologne, 50937 Köln, Germany.
46
+ 2
47
+
48
+ between two Shapiro steps depends on the periodicity of the the bound states EΦR they
49
+ give indications on the nature of the bound states [4]. In junctions with topological weak
50
+ link both Andreev bound states (ABS; diffusive bulk and surface modes) carrying 2e charge
51
+ per cycle and Majorana bound states (MBS; ballistic, perfectly transmitted surface modes)
52
+ carrying only a single 1e charge per cycle coexist. Hence the periodicity of the bound states
53
+ EΦR and respective the Josephson voltage in between two Shapiro steps differ by a factor
54
+ of 2 comparing MBSs to ABSs [9]. MBSs are crucial to probe the existence of zero energy
55
+ states within topological Josephson junctions and are indicated by missing odd Shapiro
56
+ steps in experiments [5–7].
57
+ Topological Josephson junctions can be separated into two regions. The first region is the
58
+ topological matter underneath the superconducting electrodes. Here, the proximity effect
59
+ opens an effective induced superconducting gap within both the surface and bulk of the
60
+ topological matter. The second region is in between these two laterally separated proxim-
61
+ itized regions called the weak link defined by the non-proximitized part of the topological
62
+ matter. For the investigation of novel topological matter the question arises what relevant
63
+ transport channels exist and what is their main mode of transport, i.e. ballistic topological
64
+ surface states or diffusive bulk states.
65
+ For the weak link in between the superconducting electrodes we chose Bi4Te3, which is a
66
+ natural superlattice of alternating Bi2 bilayers [10] and Bi2Te3 quintuple layers. Initially,
67
+ Bi4Te3 has been reported to be a zero band gap semimetal, comprising a Dirac cone at
68
+ the Γ-point [11]. More recently, band structure calculations supplemented with scanning
69
+ tunneling spectroscopy and angular photoemission spectroscopy measurements showed that
70
+ Bi4Te3 is a semimetal with topological surface states [12–14]. In advanced GW-band struc-
71
+ ture calculations a band gap of about 0.2 eV was identified around the Γ-point. Owing to
72
+ time-reversal and mirror symmetries, Bi4Te3 is a strong topological insulator (STI) as well
73
+ as a topological crystalline insulators (TCI). Furthermore, it is predicted that it contains
74
+ higher order topological states [14, 15].
75
+ We report on the transport properties of Josephson junctions based on the Bi4Te3 ma-
76
+ terial system as weak link material and Al/Ti as the superconducting electrodes. For the
77
+ fabrication of the samples, we employed an all in situ method [7, 16, 17], meaning that
78
+ 3
79
+
80
+ the Bi4Te3 weak link layer is grown by selective-area molecular beam epitaxy, while for the
81
+ definition of the superconducting Al/Ti electrodes we use an in situ shadow evaporation
82
+ technique. This approach allows to achieve a clean interface between the Bi4Te3 layer and
83
+ the superconductor without any contamination. In our study the proximity strength of the
84
+ Al/Ti superconducting electrodes towards the underlying Bi4Te3 nanoribbon is examined in
85
+ low temperature transport experiments including current-voltage characteristics and differ-
86
+ ential resistance measurements. From multiple Andreev reflections (MARs) identified within
87
+ the differential resistance we gain information about the strength of the proximity effect in
88
+ Bi4Te3 and the size of the induced superconducting gap. Furthermore, from the excess cur-
89
+ rent and from the temperature dependence of both the junctions critical current and the
90
+ MARs we are able to specify the dominant transport regime of the Josephson supercurrent.
91
+ II.
92
+ EXPERIMENTAL SETUP
93
+ Nanoribbon Josephson junctions have been defined following an all in situ approach[7,
94
+ 16, 17]. Therefore, two independent masking techniques are used. The masks are defined
95
+ using four alternating layers of SiO2 and Si3N4 deposited on a highly resistive Si (111)
96
+ substrate (R > 2000 Ω·cm) [17]. The first two layers are 5 nm of oxidized SiO2 and 15 nm
97
+ of low pressure chemical vapor deposited (LPCVD) Si3N4.
98
+ They comprise the selective
99
+ area growth (SAG) mask. Narrow (w = 1000 nm down to 100 nm) nanotrenches are etched
100
+ into the top Si3N4 layer using a combination of electron beam lithography and reactive ion
101
+ etching. Afterwards, a 300-nm-thick SiO2 layer and a 100-nm-thick Si3N4 layer are deposited
102
+ using LPCVD. These layers comprise the stencil mask used to deposit the superconducting
103
+ electrodes in situ. Free-hanging Si3N4 bridge structures are defined, as previously reported
104
+ [7], by patterning the Si3N4 and subsequently removing the SiO2 underneath in hydrofluoric
105
+ acid (HF). The HF dip also locally removes the SiO2 of the first oxidized layer of the SAG
106
+ mask only within the Si3N4 nanotrenches. During molecular beam epitaxy the Bi4Te3 will
107
+ selectively grow within these nanotrenches on top of the Si(111) that is revealed during SiO2
108
+ removal. The free-hanging Si3N4 bridge structures will be used after the deposition of Bi4Te3
109
+ to define the superconducting electrodes, without breaking the vacuum [7].
110
+ Bi4Te3 is a stoichiometric state of the (Bi2Te3)m(Bi2)n family with (m : n) = (3 : 3).
111
+ 4
112
+
113
+ A unit cell comprises an alternating stacking sequence of Bi2Te3 quintuple layers and Bi
114
+ bilayers. The planar epitaxy of Bi4Te3 stoichiometric alloy is achieved via molecular beam
115
+ epitaxy (MBE) by precisely controlling the Bi:Te beam flux ratio to 1:2 while keeping TBi at
116
+ 490◦C and TTe at 280◦C [15]. In order to acquire Bi4Te3 nanostructures, the optimum growth
117
+ parameters are subjected to the pre-patterned substrates with combinational surfaces. The
118
+ substrate rotation ensures a homogeneous growth of the Bi4Te3 layer also underneath the
119
+ free-hanging Si3N4 bridges. The thickness of the Bi4Te3 nanoribbon depends on the geometry
120
+ and width of the nanotrenches [15]. This is as also adatoms impinging on the Si3N4 within
121
+ the limit of the adatom diffusion length can contribute to the growth of Bi4Te3 within the
122
+ trenches. For the junctions investigated here their respective thicknesses are given in Tab. I.
123
+ The superconducting electrodes are deposited within a nitrogen cooled chamber below 0◦C
124
+ by turning off the substrate heater. The free-hanging Si3N4 bridges are aligned perpendicular
125
+ to the effusion cells of the evaporated metal, such that the shadow defines the weak link area.
126
+ Si
127
+ Si3N4
128
+ SiO2
129
+ Ti
130
+ Al2O3
131
+ x
132
+ z
133
+ y
134
+ Al
135
+ Bi4Te3
136
+ Δ
137
+ Δ*
138
+ a)
139
+ b)
140
+ x
141
+ y
142
+ z
143
+ 500 nm
144
+ Si3N4
145
+ Si3N4
146
+ Bi4Te3
147
+ Al/Ti
148
+ Al/Ti
149
+ FIG. 1. In situ deposited Bi4Te3 nanoribbon Josephson junction. a) shows a false-colored
150
+ SEM graph of the top view of an Al/Ti - Bi4Te3 - Ti/Al Josephson junction as prepared in situ.
151
+ The Al/Ti superconducting electrodes are highlighted in cyan/brown, while the Bi4Te3 nanoribbon
152
+ is shown in green and the Si3N4 hard mask in blue. The cross section along the nanoribbon main
153
+ axis is schematically depicted in b). Here, the Ti interdiffusion layer (brown), the Al2O3 dielectric
154
+ capping layer (light grey), the Si substrate (dark grey) as well as the Si3N4/SiO2 (blue/yellow)
155
+ SAG mask layers are visible. The Al/Ti contacts are attributed a composite superconducting pair
156
+ parameter ∆ and the pair parameter of the proximity coupled region in the Bi4Te3 layer (dark
157
+ green) is denoted by ∆∗.
158
+ 5
159
+
160
+ # w [nm] L [nm] t [nm] Ic [nA] RN [Ω] IcRN [µeV] ∆∗ [µeV] Iexc [nA] IexcRN [µeV]
161
+ α
162
+ τ
163
+ γB
164
+ 1
165
+ 1000
166
+ 130
167
+ 8.6
168
+ 176
169
+ 120
170
+ 21.12
171
+ 82.5
172
+ 500
173
+ 60
174
+ 0.72 0.65 0.52
175
+ 2
176
+ 500
177
+ 130
178
+ 10
179
+ 35
180
+ 310
181
+ 10.85
182
+ 95
183
+ 159
184
+ 49.3
185
+ 0.5 0.57 0.36
186
+ 3
187
+ 100
188
+ 140
189
+ 16.5
190
+ 30
191
+ 744
192
+ 22.32
193
+ 115
194
+ 75
195
+ 55.8
196
+ 0.49 0.56 0.24
197
+ TABLE I. Overview of interface parameters of Josephson junctions with Bi4Te3 nanoribbon weak
198
+ link and Al/Ti (30 nm/3 nm) superconducting contacts. Given are the geometric dimensions, the
199
+ junction width w, the junction length/electrode separation length L and the mean thickness t of
200
+ the nanoribbon. The proximity induced gap below the superconducting electrodes ∆∗, the excess
201
+ current Ic as well as the dimensionless parameters α, τ and γB that describe the interfacial quality
202
+ of the junctions.
203
+ After electrode deposition devices are covered by a 5 nm thin Al2O3 dielectric layer electron
204
+ beam evaporated from a stoichiometric target. A false-colored scanning electron micrograph
205
+ of an as-prepared Josephson junction with aluminum superconducting contacts is shown in
206
+ Fig. 1 a). Aluminum has previously been reported to diffuse into (Bi0.06Sb0.94)2Te3 thin
207
+ films [18], which increases the interfacial resistance of junctions within the superconducting
208
+ regime of the Al electrodes. In order to prevent diffusion of the Al into the underlying Bi4Te3
209
+ layer, a 3 nm thin Ti layer is deposited first as an interdiffusion barrier, as depicted in the
210
+ schematics of the junction cross section shown in Fig. 1 b). The critical temperature of
211
+ the superconducting Al/Ti composite electrodes is determined to be Tc = 0.95 K from four-
212
+ terminal measurements of the differential resistance as a function of the temperature T down
213
+ to 23 mK base temperature of a dilution refrigerator. The magnitude of the superconducting
214
+ pair parameter has been determined to measure ∆ = 145 µeV, following Bardeen-Cooper-
215
+ Schrieffer theory [19].
216
+ The electrodes of the in situ defined nanoribbon Josephson junctions are wire bonded
217
+ to a chip carrier in a quasi-four terminal contact configuration. The junctions are cooled
218
+ to a base temperature of T = 23 mK using a dilution refrigerator. At base temperature
219
+ the sample resistance is measured using standard lock-in techniques and the potential drop
220
+ across the junction is determined using a voltmeter.
221
+ 6
222
+
223
+ III.
224
+ EXPERIMENTAL RESULTS
225
+ A.
226
+ Multiple Andreev reflections
227
+ We have measured three junctions of different width w = 100, 500, and 1000 nm (see
228
+ Tab. I). Both dV/dI(I) as well as V (I) as a function of a d.c. current bias applied for the
229
+ 500,nm wide junction (junction #2) are shown in Fig. 2 a). For an applied bias current
230
+ below the critical current of Ic = 35 nA (see also Tab. I) a Josephson supercurrent estab-
231
+ lishes. The differential resistance dV/dI is zero below I < Ic and reaches a finite value as
232
+ soon as the current bias exceeds I > Ic. The critical current is not hysteretic, whether the
233
+ current bias is swept from positive to negative current biases or vice versa. In Tab. I the
234
+ critical current Ic, the normal state resistance RN values are listed. The corresponding IcRN
235
+ product values are found to be in the range between 10.85 and 22.32 µV.
236
+ As mentioned before, we anticipate that establishing a supercurrent through the Bi4Te3
237
+ a)
238
+ -900
239
+ -600
240
+ -300
241
+ 0
242
+ 300
243
+ 600
244
+ 900
245
+ -200
246
+ 0
247
+ 200
248
+ 0
249
+ 200
250
+ 400
251
+ 600
252
+ V (μV)
253
+ I (nA)
254
+ dV/dI (Ω)
255
+ 0
256
+ 200
257
+ 400
258
+ 600
259
+ dV/dI (Ω)
260
+ Iexc
261
+ -200
262
+ -100
263
+ 0
264
+ 100
265
+ 200
266
+ V (µV)
267
+ b)
268
+ 0.0
269
+ 0.2
270
+ 0.4
271
+ 0.6
272
+ 0.8
273
+ 80
274
+ 160
275
+ (μV)
276
+ MAR ord. -1 (1/n)
277
+ n=1
278
+ 2
279
+ 3
280
+ 4
281
+ 57
282
+ Vn
283
+ FIG. 2. IV -characteristics and differential resistance dV/dI of Josephson junction #2.
284
+ a) IV -characteristics and differential resistance as a function of the applied d.c.
285
+ bias current
286
+ (dV/dI(I)).
287
+ A linear extrapolation from the IV-characteristics above 2∆∗ to V = 0 is shown
288
+ (red dashed line) to extract the excess current Iexc.
289
+ b) Differential resistance as a function of
290
+ the measured d.c. potential drop across the Josephson junction (dV/dI(V )), showing signatures
291
+ of multiple Andreev reflections (MARs). The inset shows the position (Vn) of the MARs plotted
292
+ against the inverse of the MAR order number (1/n). The linear fit is forced through the origin.
293
+ layer is a two step process. First, the proximity to the superconducting metallic Al/Ti elec-
294
+ 7
295
+
296
+ trodes induces a superconducting pair potential into the Bi4Te3 layer (dark green regions in
297
+ Fig. 1 b)), which decays over a length scale given by the superconducting coherence length ξN
298
+ within the Bi4Te3 layer. For the superconducting coherence length we have to consider two
299
+ different cases. In the ’dirty limit’, the elastic scattering in the dissipative state of the Bi4Te3
300
+ layer takes place on length scales smaller than the superconducting coherence length. When
301
+ the distance between two elastic scattering events exceeds the superconducting coherence
302
+ length, the transition is in the ’clean limit’. Using low-temperature magnetotransport data
303
+ on nano-Hall structures, we find that the Bi4Te3 layer is (semi)metallic, in agreement with
304
+ recent reports [12–14], with a carrier density of n2D ≈ 4 × 1014 cm−2 (see Supplementary
305
+ Sec. A) and an elastic mean free path length of only le ≈ 4 nm. Furthermore, the Hall bar
306
+ data does not show any significant increase of the magnetoresistance, as expected from a
307
+ Dirac semimetal [20]. For given reasons we therefore assume that the proximitized regions
308
+ of the Bi4Te3 film underneath the superconducting Al/Ti electrodes are in the dirty limit,
309
+ since the estimated superconducting coherence length of ξN =
310
+
311
+ ¯hDBulk/2πkBTc = 45 nm,
312
+ with DBulk the diffusion constant of the bulk and Tc the critical temperature of the Ti/Al
313
+ superconducting electrodes.
314
+ When proximitizing the regions of the Bi4Te3 underneath the Al/Ti superconducting
315
+ electrodes a Josephson supercurrent establishes in a next step between the two proximitized
316
+ layers based on electron-hole bound states.
317
+ When the applied current bias exceeds the
318
+ critical current I > Ic the junctions resistance is modulated by Andreev reflection processes
319
+ at the superconductor to normal conductor interface. Only beyond a current bias of about
320
+ |I| > 740 nA the junctions resistance is mostly constant. At this point the potential drop
321
+ across the junction measures 2∆∗, i.e. the size of the proximity induced superconducting
322
+ gap, as indicated in Fig. 1 b). In order to quantify the size of the induced superconducting
323
+ gap ∆∗ in the proximitized Bi4Te3 more precisely (cf.
324
+ Fig. 1 b)) we analyzed multiple
325
+ Andreev reflections (MARs) visible within the differential resistance dV/dI of the junction.
326
+ These MARs occur at bias voltages below the size of the induced superconducting gap at
327
+ voltages of V = 2∆∗/en, where n is an integer [21]. In junction #2 we observe MARs of
328
+ the order n = 1, 2, 3, 4, 5, 6, 9, 11, 13. Missing signatures of intermediate order MARs (e.g.
329
+ n = 7, 8) has been observed before in BiSbTeSe2 nanoribbon Josephson junctions [22] but
330
+ an explanation is missing until now.
331
+ 8
332
+
333
+ The size of the induced superconducting gap is determined by plotting the position (in volt)
334
+ of each MAR against the inverse of the MAR order number (1/n). The induced supercon-
335
+ ducting gap measures ∆∗ = 95 µeV (for n = 1, as indicated by a blue dot within Fig. 2
336
+ b) at T = 50 mK), which is smaller than the gap of the Al/Ti superconducting electrodes
337
+ (∆ = 145 µeV).
338
+ B.
339
+ Temperature dependency of Ic and MARs
340
+ Figure 2 b) shows the differential resistance of junction #2 at different temperatures.
341
+ The signatures of MARs vanish above the critical temperature of the Al/Ti superconducting
342
+ electrodes. The temperature dependency of MARs of order n = 1 (blue dots), n = 2 (orange
343
+ dots) and n = 3 (green dots) is shown in Figs. 2 b) and c). The temperature dependency of
344
+ the induced superconducting gap is given by [23]
345
+ ∆∗(T) =
346
+ ∆Al/Ti(T)
347
+ 1 + γB
348
+
349
+ ∆2
350
+ Al/Ti(T) − ∆∗2(T)/kBTc
351
+ ,
352
+ (1)
353
+ where γB is a measure of the interfacial barrier strength in between the Al/Ti superconduct-
354
+ ing electrodes and the Bi4Te3 nanoribbon layer. Above formula is fitted to the ∆∗(T) data
355
+ and a value for γB = 0.36 has been determined. The γB values of the other two junctions
356
+ are listed in Tab. I. The value of γB indicates that there is an effective barrier present
357
+ between the Al/Ti layer and Bi4Te3 despite the in situ fabrication. It has been identified
358
+ that the Bi4Te3 tends to be terminated by a Bi bi-layer underneath the Al/Ti layer while it
359
+ is terminated by a Bi2Te3 layer otherwise [15]. A possible reason for the barrier identified
360
+ might be the mismatch of Fermi energies in between these different regions on the surface
361
+ of the proximitized and the non-proximitized regions of Bi4Te3 resulting in a potential step
362
+ at their interface.
363
+ As a next step, the Josephson supercurrent between the the proximitized regions with
364
+ the superconducting gap ∆∗ is analyzed in detail. The supercurrent depends on the kind
365
+ of transport, i.e. ballistic or diffusive, and on the transparency between the proximitized
366
+ Bi4Te3 layers and the Bi4Te3 weak link (cf. Fig. 1 b). The transparency of the interfaces of
367
+ the lateral Josephson junction are analyzed in two ways. The first method uses the excess
368
+ 9
369
+
370
+ a)
371
+ T=50mK
372
+ T=800mK
373
+ -200
374
+ -100
375
+ 0
376
+ 100
377
+ 200
378
+ 200
379
+ 300
380
+ 400
381
+ 500
382
+ 600
383
+ dV/dI (Ω)
384
+ V (µV)
385
+ b)
386
+ c)
387
+ V (μV)
388
+ -200
389
+ -150
390
+ -100
391
+ -50
392
+ 0
393
+ 50
394
+ 100
395
+ 150
396
+ 200
397
+ -30
398
+ 0
399
+ 30
400
+ 60
401
+ 90
402
+ 120
403
+ 150
404
+ I (nA)
405
+ T = 50 mK
406
+ T = 500 mK
407
+
408
+ Ic (nA)
409
+ 30
410
+ 20
411
+ 10
412
+ 100
413
+ 200
414
+ 300
415
+ T (mK)
416
+ 0
417
+ 40
418
+ 80
419
+ 120
420
+ 160
421
+ 200
422
+ 0.2
423
+ 0.4
424
+ 0.6
425
+ 0.8
426
+ V (µV)
427
+ T (K)
428
+ 2Δ*
429
+ Δ*
430
+ 2/3Δ*
431
+ FIG. 3.
432
+ Temperature dependency of a) the differential resistance as a function of the bias
433
+ potential of Josephson junction #2, dV/dI(V, T), in between 50 mK and 800 mK. Each trace is
434
+ offset by 25 Ω. The size of the proximity induced superconducting gap (2∆∗ in µV) is highlighted
435
+ by blue dots, while MARs of order n = 2 and 3 are highlighted by orange and green dots respectively.
436
+ b) shows the temperature dependent position of 2∆∗ and MARs of order n = 2 and n = 3. c) shows
437
+ the temperature dependent IV -characteristics from 50 mK to 500 mK, where each trace is offset by
438
+ 5 µV. The red dashed line indicates the critical current of the junction Ic = 35 nA at T = 50 mK.
439
+ The temperature dependent critical current, Ic(T), is shown in the inset with the fit indicated by
440
+ the dashed line.
441
+ 10
442
+
443
+ 600
444
+ 550
445
+ 500
446
+ 450
447
+ 400
448
+ 350
449
+ 300
450
+ 250
451
+ 200
452
+ -200
453
+ -100
454
+ 100
455
+ 200current of Iexc = 159 nA, which is determined from the junctions IV -characteristics by linear
456
+ extrapolation above the superconducting gap V ≥ 2∆∗ (highlighted as dashed red line in
457
+ Fig. 2 a)) to V = 0. The excess current displays the additional current due to successful
458
+ Andreev reflections in the dissipative state of the junction and is directly related to the
459
+ transparency of the junction. An analytical expression following Niebler, Cuniberti, and
460
+ Novotny [24] is used to determine the junctions interfacial barrier strength Z = 0.86 using
461
+ the parameter α = eIexcRN/∆∗ (cf. Tab. I). The barrier strength is related to the trans-
462
+ parency via τ = 1/(1 + Z2) = 0.57. Note, that τ expresses the transparency between the
463
+ proximitized and the non-proximitized regions of the Bi4Te3 layer in contrast to γB which
464
+ quantifies the barrier between the Al/Ti superconducting electrodes and the proximitized
465
+ Bi4Te3 region. Similar to the observation from Kunakova et al.[25] we find that the values for
466
+ the interface transparency parameters τ and γB are interdependent. This effect can be at-
467
+ tributed to a metallization effect the electrodes have on the surface states of the Bi4Te3 layer.
468
+ An additional method to determine the interface transparency is by quasi-classical analy-
469
+ sis of the temperature dependent critical current [7, 8]. Using a voltage criterion the critical
470
+ current is extracted from the IV -characteristics at different temperatures shown in Fig. 2
471
+ d), with the inset showing Ic(T). We used a ballistic model fit (shown as black dashed line
472
+ in the inset of Fig. 1 b)) based on the Gor’kov equations with arbitrary junction length L
473
+ [26] and barrier transparency D [27]. For the fit a value for the critical temperature of the
474
+ gap within the Al/Ti electrodes Tc = 0.95 K and a Fermi velocity of vF = 3.8 × 105 m/s of
475
+ the Bi2Te3 surface layer are used [15]. The best fit results in an interface transparency of
476
+ D = 0.6, which is in good agreement with the transparency τ determined using the excess
477
+ current analysis described before. We also performed a quasi-classical fit using a diffusive
478
+ model based on the Usadel equations [28]. However, within a physically reasonable range of
479
+ values we did not get a decent fit. Our analysis indicates that the supercurrent is carried by
480
+ ballistic modes with increased superconducting coherence length rather than bulk modes.
481
+ The superconducting coherence length of these ballistic modes can be estimated within
482
+ the clean junction limit to measure ξN = ¯hvF/2πkBTc = 1.15 µm. The observation of a
483
+ dominating ballistic channel might be attributed to highly conductive surface states of the
484
+ Bi4Te3 layer, which overrules the diffusive transport in the bulk.
485
+ 11
486
+
487
+ C.
488
+ Shapiro steps
489
+ We also performed differential resistance measurements under the influence of an exter-
490
+ nally applied radio-frequency (rf) signal using a λ/4 antenna. In Figs. 4 a)-f) the differential
491
+ resistance is displayed as a function of the applied rf power and the junctions potential differ-
492
+ ence is scaled by hf/2e. Within the range of frequencies applied 1.7 GHz ≤ frf ≤ 14.25 GHz
493
+ we observe full integer Shapiro steps. At frequencies frf = 14.25 GHz and 8.25 GHz, how-
494
+ ever, additional sub-integer Shapiro steps have been measured. Fractional Shapiro steps can
495
+ be caused by phase-slip centers inside the junction [29], phase instabilities introduced by
496
+ Abrikosov vortices [30], magnetic disorder [31] or due to a non-sinusoidal or skewed current
497
+ phase relation (CΦR) [32, 33]. In junctions of high transparencies or very short ballistic
498
+ junctions the CΦR is expected to be non-sinusodial [30]. The relative measure of the junc-
499
+ tion length over the superconducting coherence length within the Bi4Te3 layer (d/ξN) has
500
+ influence on the maximum Josephson supercurrent and the shape of the CΦR. Already for
501
+ values of ξN/d > 3 the CΦR is skewed and the maximum current density lies above a value of
502
+ φ > π [30]. Skewed, non-sinusoidal CΦRs can be decomposed into sinusoidal components of
503
+ lower periodicity, which can explain the evolution of sub-integer Shapiro steps. For ballistic
504
+ modes of increased superconducting coherence length (ξN = 1.15 µm) this limit would need
505
+ to be considered as the junction length of junction #2 (L = 130 nm) is much smaller. For
506
+ the wider junction (junction #1, w = 1000 nm) as for the narrower junction (junction #3,
507
+ w = 100 nm, see supplementary Sec. B) the sub-integer Shapiro steps have been observed
508
+ as well, confirming the presence of ballistic modes independent of the junction geometry.
509
+ For the presence of ballistic modes one would expect the supercurrent to be partially carried
510
+ by MBSs, resulting in odd integer Shapiro steps to vanish [9]. Based on the fraction of the
511
+ supercurrent that is carried by MBSs compared to the supercurrent carried by ABSs the
512
+ cross-over frequency for the observation of missing odd integer Shapiro steps [4] should lie
513
+ below fMBSs < 5.25 GHz, which is the cross over frequency considering the whole supercur-
514
+ rent is carried only by MBSs. As no missing Shapiro steps have been recorded it is assumed
515
+ that less then one third of the supercurrent is carried by MBSs. Therefore, a possible reason
516
+ that we did not observe missing odd integer Shapiro steps as an indication of MBSs might
517
+ be that the irradiated frequency was too large [9].
518
+ 12
519
+
520
+ a)
521
+ c)
522
+ -4
523
+ -2
524
+ 0
525
+ 2
526
+ 4
527
+ 6
528
+ 0
529
+ 1
530
+ V (hf/2e)
531
+ RF power (dBm)
532
+ dV/dI (Ω)
533
+ -2
534
+ -1
535
+ 3
536
+ 2
537
+ -3
538
+ 8
539
+ 10
540
+ 320
541
+ 240
542
+ 160
543
+ 80
544
+ 0
545
+ -20
546
+ -10
547
+ 0
548
+ 2
549
+ V (hf/2e)
550
+ RF power (dBm)
551
+ dV/dI (Ω)
552
+ -4
553
+ -2
554
+ 6
555
+ 4
556
+ -6
557
+ 0
558
+ 10
559
+ 320
560
+ 240
561
+ 160
562
+ 80
563
+ 0
564
+ e)
565
+ -6
566
+ -4
567
+ 0
568
+ 2
569
+ V (hf/2e)
570
+ RF power (dBm)
571
+ dV/dI (Ω)
572
+ -4
573
+ -2
574
+ 4
575
+ -2
576
+ 0
577
+ 320
578
+ 240
579
+ 160
580
+ 80
581
+ 0
582
+ -8
583
+ f = 14.25 GHz
584
+ f = 8.25 GHz
585
+ f = 1.7 GHz
586
+ 1/2
587
+ 1/3
588
+ 1/5
589
+ b)
590
+ dV/dI (Ω)
591
+ f = 14.25 GHz
592
+ 0
593
+ 1
594
+ 2
595
+ 3
596
+ 100
597
+ 200
598
+ 300
599
+ V (hf/2e)
600
+ 400
601
+ 500
602
+ 1/2
603
+ 1/3
604
+ 1/5
605
+ d)
606
+ dV/dI (Ω)
607
+ f = 8.25 GHz
608
+ 0
609
+ 1
610
+ 2
611
+ 3
612
+ 100
613
+ 200
614
+ 300
615
+ V (hf/2e)
616
+ 400
617
+ 500
618
+ P = 3 dBm
619
+ P = -4 dBm
620
+ P = -17 dBm
621
+ P = -10 dBm
622
+ P = -8 dBm
623
+ P = -2 dBm
624
+ f)
625
+ dV/dI (Ω)
626
+ 0
627
+ 1
628
+ 2
629
+ 3
630
+ 200
631
+ 300
632
+ 400
633
+ V (hf/2e)
634
+ 100
635
+ f = 1.7 GHz
636
+ FIG. 4. Shapiro response of Josephson junction #2 at different radio-frequencies ap-
637
+ plied. a), c) and e) show the differential resistance as a function of the radio-frequency excitation
638
+ amplitude/radio-frequency power (P in dBm) and the potential bias (V in hf/2e) of the junction
639
+ (dV/dI(P, V )). The differential resistance is displayed in between values of dV/dI = 0 (red) and
640
+ dV/dI = 320 Ω. b), d) and f) show line traces of the differential resistance as a function of the
641
+ bias potential in a given range of radio-frequency powers, with the lowest power displayed in blue
642
+ and the highest in red. Next to Shapiro steps at full integer values of V = n · hf/2e, there are
643
+ sub-integer steps visible in line-cuts at f = 14.25 GHz and 8.25 GHz. The sub-integer steps are
644
+ highlighted with their given fractions of the first integer Shapiro step.
645
+ 13
646
+
647
+ 3
648
+ 2
649
+ 1
650
+ 0
651
+ -1
652
+ -2
653
+ 3
654
+ -4
655
+ -2
656
+ 0
657
+ 2
658
+ 4
659
+ 9
660
+ 8
661
+ 100.000
662
+ 80.00
663
+ 160.0
664
+ 240.0
665
+ 320.00.000
666
+ 80.00
667
+ 160.0
668
+ 240.0
669
+ 320.00.000
670
+ 80.00
671
+ 160.0
672
+ 240.0
673
+ 320.0IV.
674
+ CONCLUSIONS
675
+ By characterizing Bi4Te3-based Josephson junctions we obtained a detailed picture of
676
+ the different contributions taking part in establishing a supercurrent through the junctions
677
+ weak link. By analysing MARs we found that the intimate contact of the Al/Ti layer on top
678
+ of the Bi4Te3 layer results in an induced superconductive gap ∆∗ in the topological matter
679
+ due to the proximity effect. In order to establish robust proximitzed regions underneath
680
+ the Al/Ti electrodes the presence of bulk carriers are probably beneficial, if not essential.
681
+ The Bi4Te3 has been identified to carry a large amount of bulk charges. The proximitized
682
+ regions of the Bi4Te3 are coupled by the unproximitized Bi4Te3 weak link giving rise to
683
+ a Josephson supercurrent.
684
+ We anticipate that the Josephson supercurrent mainly flows
685
+ in the topological surface channel rather than in the bulk of the Bi4Te3 link, similar to
686
+ results obtained in junctions with a different topological insulator layer [7]. Analysing the
687
+ temperature dependency of the critical current we indeed identified the transport regime in
688
+ these junctions to be mainly ballistic. However, by analysing the temperature dependency
689
+ of the MARs an effective barrier in between these regions, probably due to a different surface
690
+ termination of both regions, has been identified. From our Shapiro step measurements we
691
+ came to the conclusion that the current-phase relationship is non-sinusoidal, i.e. supporting
692
+ our claim of ballistic modes in our junctions. However, we did not find a 4π contribution in
693
+ the Shapiro step measurements indicating the presence of Majorana zero modes. One reason
694
+ might be that our lowest rf frequency of f ≤ 1.7 GHz was too high, so that we could not
695
+ enter the regime where the 4π contributions are visible. For future material combinations
696
+ in hybrid Josephson junctions including topological matter it is important to consider our
697
+ findings.
698
+ ACKNOWLEDGEMENTS
699
+ This work was partly funded by the Deutsche Forschungsgemeinschaft (DFG, German
700
+ Research Foundation) under Germany’s Excellence Strategy - Cluster of Excellence Matter
701
+ and Light for Quantum Computing (ML4Q) EXC 2004/1 - 390534769.
702
+ This work was
703
+ financially supported by the German Federal Ministry of Education and Research (BMBF)
704
+ via the Quantum Futur project "MajoranaChips" (Grant No. 13N15264) within the funding
705
+ 14
706
+
707
+ program Photonic Research Germany.
708
+ Competing interests
709
+ The author(s) declare no competing interests.
710
+ [1] F. Hassler, A. R. Akhmerov,
711
+ and C. W. J. Beenakker, New Journal of Physics 13, 095004
712
+ (2011).
713
+ [2] A. M. Cook, M. M. Vazifeh, and M. Franz, Phys. Rev. B 86, 155431 (2012).
714
+ [3] T. Hyart, B. van Heck, I. C. Fulga, M. Burrello, A. R. Akhmerov, and C. W. J. Beenakker,
715
+ Phys. Rev. B 88, 035121 (2013).
716
+ [4] F. Dominguez, F. Hassler, and G. Platero, Phys. Rev. B 86, 140503 (2012).
717
+ [5] J. Wiedenmann, E. Bocquillon, R. S. Deacon, S. Hartinger, O. Herrmann, T. M. Klapwijk,
718
+ L. Maier, C. Ames, C. Brüne, C. Gould, A. Oiwa, K. Ishibashi, S. Tarucha, H. Buhmann, and
719
+ L. W. Molenkamp, Nat. Commun. 7, 10303 (2015).
720
+ [6] E. Bocquillon, R. S. Deacon, J. Wiedenmann, P. Leubner, T. M. Klapwijk, C. Brüne,
721
+ K. Ishibashi, H. Buhmann, and L. W. Molenkamp, Nat. Nanotechnol. 12, 137 (2016).
722
+ [7] P. Schüffelgen, D. Rosenbach, C. Li, T. W. Schmitt, M. Schleenvoigt, A. R. Jalil, S. Schmitt,
723
+ J. Kölzer, M. Wang, B. Bennemann, U. Parlak, L. Kibkalo, S. Trellenkamp, T. Grap,
724
+ D. Meertens, M. Luysberg, G. Mussler, E. Berenschot, N. Tas, A. A. Golubov, A. Brinkman,
725
+ T. Schäpers, and D. Grützmacher, Nat. Nanotechnol. 14, 825 (2019).
726
+ [8] D. Rosenbach, T. W. Schmitt, P. Schüffelgen, M. P. Stehno, C. Li, M. Schleenvoigt, A. R. Jalil,
727
+ G. Mussler, E. Neumann, S. Trellenkamp, A. A. Golubov, A. Brinkman, D. Grützmacher, and
728
+ T. Schäpers, Science Advances 7, eabf1854 (2021).
729
+ [9] F. Dominguez, O. Kashuba, E. Bocquillon, J. Wiedenmann, R. S. Deacon, T. M. Klapwijk,
730
+ G. Platero, L. W. Molenkamp, B. Trauzettel, and E. M. Hankiewicz, Phys. Rev. B 95, 195430
731
+ (2017).
732
+ [10] E. Aktürk, O. U. Aktürk, and S. Ciraci, Phys. Rev. B 94, 014115 (2016).
733
+ [11] Y. Saito, P. Fons, K. Makino, K. V. Mitrofanov, F. Uesugi, M. Takeguchi, A. V. Kolobov, and
734
+ J. Tominaga, Nanoscale 9, 15115 (2017).
735
+ 15
736
+
737
+ [12] T. Chagas, G. A. S. Ribeiro, P. H. R. Gonçalves, L. Calil, W. S. Silva, Â. Malachias, M. S. C.
738
+ Mazzoni, and R. Magalhães-Paniago, Electronic Structure 2, 015002 (2020).
739
+ [13] T. Chagas, O. A. Ashour, G. A. S. Ribeiro, W. S. Silva, Z. Li, S. G. Louie, R. Magalhães
740
+ Paniago, and Y. Petroff, Phys. Rev. B 105, L081409 (2022).
741
+ [14] D. Nabok, M. Tas, S. Kusaka, E. Durgun, C. Friedrich, G. Bihlmayer, S. Blügel, T. Hirahara,
742
+ and I. Aguilera, Phys. Rev. Materials 6, 034204 (2022).
743
+ [15] A. R. Jalil, Ph.D. thesis, RWTH Aachen University (2022).
744
+ [16] D. Rosenbach, N. Oellers, A. R. Jalil, M. Mikulics, J. Kölzer, E. Zimmermann, G. Mussler,
745
+ S. Bunte, D. Grützmacher, H. Lüth,
746
+ and T. Schäpers, Advanced Electronic Materials 6,
747
+ 2000205 (2020).
748
+ [17] T. W. Schmitt, M. R. Connolly, M. Schleenvoigt, C. Liu, O. Kennedy, J. M. Chávez-Garcia,
749
+ A. R. Jalil, B. Bennemann, S. Trellenkamp, F. Lentz, E. Neumann, T. Lindström, S. E.
750
+ de Graaf, E. Berenschot, N. Tas, G. Mussler, K. D. Petersson, D. Grützmacher, and P. Schüf-
751
+ felgen, Nano Lett. 22, 2595 (2022).
752
+ [18] P. Schüffelgen, D. Rosenbach, E. Neumann, M. P. Stehno, M. Lanius, J. Zhao, M. Wang,
753
+ B. Sheehan, M. Schmidt, B. Gao, A. Brinkman, G. Mussler, T. Schäpers, and D. Grützmacher,
754
+ Journal of Crystal Growth 477, 183 (2017), proceeding of the 19th International Conference
755
+ on Molecular Beam Epitaxy.
756
+ [19] J. Bardeen, L. N. Cooper, and J. R. Schrieffer, Phys. Rev. 108, 1175 (1957).
757
+ [20] T. Liang, Q. Gibson, M. N. Ali, M. Liu, R. J. Cava, and N. P. Ong, Nature Materials 14, 280
758
+ (2014).
759
+ [21] A. W. Kleinsasser, R. E. Miller, W. H. Mallison, and G. B. Arnold, Phys. Rev. Lett. 72, 1738
760
+ (1994).
761
+ [22] L. A. Jauregui, M. Kayyalha, A. Kazakov, I. Miotkowski, L. P. Rokhinson, and Y. P. Chen,
762
+ Applied Physics Letters 112, 093105 (2018), https://doi.org/10.1063/1.5008746.
763
+ [23] B. A. Aminov, A. A. Golubov, and M. Y. Kupriyanov, Phys. Rev. B 53, 365 (1996).
764
+ [24] G. Niebler, G. Cuniberti, and T. Novotný, Supercond. Sci. Technol. 22, 085016 (2009).
765
+ [25] G. Kunakova, T. Bauch, E. Trabaldo, J. Andzane, D. Erts, and F. Lombardi, Applied Physics
766
+ Letters 115, 172601 (2019), https://doi.org/10.1063/1.5123554.
767
+ [26] A. Brinkman and A. A. Golubov, Phys. Rev. B 61, 11297 (2000).
768
+ [27] A. V. Galaktionov and A. D. Zaikin, Phys. Rev. B 65, 184507 (2002).
769
+ 16
770
+
771
+ [28] K. D. Usadel, Phys. Rev. Lett. 25, 507 (1970).
772
+ [29] B. Ivlev and N. Kopnin, Advances in Physics 33, 47 (1984).
773
+ [30] A. A. Golubov, M. Y. Kupriyanov, and E. Il’ichev, Rev. Mod. Phys. 76, 411 (2004).
774
+ [31] H. Sellier, C. Baraduc, F. m. c. Lefloch,
775
+ and R. Calemczuk, Phys. Rev. Lett. 92, 257005
776
+ (2004).
777
+ [32] R. Panghotra, B. Raes, C. C. de Souza Silva, I. Cools, W. Keijers, J. E. Scheerder, V. V.
778
+ Moshchalkov, and J. Van de Vondel, Communications Physics 3, 53 (2020).
779
+ [33] B. Raes, N. Tubsrinuan, R. Sreedhar, D. S. Guala, R. Panghotra, H. Dausy, C. C.
780
+ de Souza Silva, and J. Van de Vondel, Phys. Rev. B 102, 054507 (2020).
781
+ 17
782
+
783
+ SUPPLEMENTARY INFORMATION
784
+ A.
785
+ Magnetotransport
786
+ Besides Josephson junctions with a Bi4Te3 weak-link we have fabricated nano Hall bars.
787
+ Therefore we have selectively deposited Bi4Te3 in nanotrenches that have been arranged in
788
+ a Hall bar layout with one main nanoribbon and three ribbons symmetrically on each side.
789
+ The fabricated device is shown as a false color scanning electron micrograph in Fig. 5 a).
790
+ The nanoribbons have a width of w = 100 nm and the spacing in between two side nanorib-
791
+ bons is L = 1000 nm.
792
+ Nano Hall bar devices have been cooled down to T = 1.5 K in a variable temperature in-
793
+ sert cryostate equipped with a superconducting magnet that can apply magnetic fields up
794
+ to Bmax = 13 T. The sample holder insert is equipped with an electromechanical stepper
795
+ motor. The relative orientation of the magnetic field to the nano Hall bars can be changed
796
+ from an alignment of the magnetic field parallel to the main nanoribbon axis to a mag-
797
+ netic field oriented perpendicular out-of-plane. From Hall measurements in an out-of-plane
798
+ applied magnetic field the Hall slope AH = dRxy/dB = 1.58 Ω/T and subsequently the two-
799
+ dimensional sheet carrier density n2D = (AHe)−1 = 3.9 × 1014 cm−2 have been determined.
800
+ The Bi4Te3 has a strong metallic character with high charge carrier density and low mobil-
801
+ ities µ = L · (WRxxn2De)−1 = 215 cm2(V · s)−1
802
+ Fig. 5 b) shows the longitudinal resistance of the nano Hall bar for different relative angles of
803
+ the magnetic field applied to the surface of the substrate. Next to the weak antilocalization
804
+ feature, typical for 3D bulk as well as 2D surfaces with strong spin-orbit coupling, the mag-
805
+ netoresistance does not change by more then 2.5% over the whole range of applied magnetic
806
+ fields B ≤ |±13T|. In a perpendicular applied magnetic field (Θ = 90◦, red curve) the mag-
807
+ netoresistance does show a spectrum of universal conductance fluctuations. In Fig. 5 c) the
808
+ amplitude of individual oscillations frequencies from a fast fourier transformation performed
809
+ on the data from b), shows a set of prominent frequencies, limited by the phase coherence
810
+ length (lφ = (φ0 · fmax) ≈ 25 nm).
811
+ The temperature dependent magnetoresistance data in a perpendicular applied magnetic
812
+ field for temperatures in between 2 K ≤ T ≤ 30 K is shown in Fig. 5 d). For each trace the
813
+ root mean square of the oscillation amplitude is computed rms(δGxx) and the values are
814
+ 18
815
+
816
+ shown in the inset as a function of temperature. The value for rms(δGxx) is constant up to
817
+ a temperature of 9 K. For higher temperatures the values follow a T −3/2 dependency.
818
+ -10
819
+ -5
820
+ 0
821
+ 5
822
+ 10
823
+ 745
824
+ 750
825
+ 755
826
+ 760
827
+ Rxx (Ω)
828
+ B (T)
829
+ -10
830
+ 10
831
+ 30
832
+ 50
833
+ 90
834
+ 70
835
+ θ (°)
836
+ Bi4Te3
837
+ B (�=90Deg)
838
+ B (�=0Deg)
839
+ I
840
+ 0
841
+ 20
842
+ 40
843
+ 60
844
+ 80
845
+ θ (°)
846
+ 100
847
+ 0
848
+ 2
849
+ 4
850
+ 6
851
+ 8
852
+ 10
853
+ 12
854
+ 1/B (1/T)
855
+ FFT Amp. (a.u.)
856
+ -10
857
+ -5
858
+ 0
859
+ 5
860
+ 10
861
+ B (T)
862
+ 745
863
+ 750
864
+ 755
865
+ 760
866
+ Rxx (Ω)
867
+ 0
868
+ 5
869
+ 10
870
+ 15
871
+ 20
872
+ 25
873
+ 30
874
+ T (K)
875
+ 10
876
+ 1
877
+ 0.01
878
+ 0.02
879
+ 0.03
880
+ 0.04
881
+ T (K)
882
+ rms(δGxx)(e2/h)
883
+ a)
884
+ b)
885
+ c)
886
+ d)
887
+ T-3/2
888
+ 1 μm
889
+ FIG. 5. Magnetotransport data on Bi4Te3 Hall bars. a) Layout of the selectively grown
890
+ Bi4Te3 nano Hall bar investigated. b) Longitudinal magnetoresistance as a function of magnetic
891
+ field for various tilt angles (Rxx(B, θ) of the devices main channel w.r.t the magnetic field. The
892
+ orientation of the sample is schematically depicted. c) Fast-Fourier-transformation amplitude of
893
+ the magnetoresistance traces from b) showing high frequent universal conductance fluctuations at
894
+ a large range of angles in between 15◦ ≤ θ ≤ 90◦ and low frequent oscillations from coherent states
895
+ within the nanoribbon cross section for a magnetic field applied parallel to the main axis of the
896
+ nanoribbon. d) Temperature dependency of the longitudinal magnetoresistance (Rxx(B, T)) for
897
+ temperatures in between 1.5 K≤ T ≤ 30 K. The inset shows the temperature dependency of the
898
+ root mean square of the conductance fluctuation amplitude rms(δGxx(T)
899
+ B.
900
+ Shapiro response measurements
901
+ Next to the 500 nm wide Bi4Te3 Josephson junction characterized in the main text, we
902
+ have additionally measured a wide junction (w = 1000 nm, Junction #1) and a narrow
903
+ junction (w = 100 nm, Junction #3). All the junction parameters are given within the table
904
+ 19
905
+
906
+ BST2318 Bi4Te3 100nm HB Rxx(Q)
907
+ 100.0
908
+ 93.9
909
+ 87.8
910
+ 760
911
+ 81.7
912
+ 75.6
913
+ 69.4
914
+ 63.3
915
+ 57.2
916
+ 755
917
+ 51.1
918
+
919
+ 45.0
920
+ 0
921
+ 38.9
922
+ 32.8
923
+ 750
924
+ 26.7
925
+ 20.6
926
+ 14.4
927
+ 8.3
928
+ 745
929
+ 2.2
930
+ 3.9
931
+ -10.0
932
+ -10
933
+ -5
934
+ 0
935
+ 5
936
+ 10
937
+ B (T)BST2318 Bi4Te3 100nm FFT Ampl.(°), L=1900nm
938
+ 12
939
+ 10
940
+ 10000.00
941
+ 8
942
+ 1000.00
943
+ frequency (1/B)
944
+ FFT Amp. (a.u)
945
+ 6
946
+ 100.00
947
+ 4
948
+ 10.00
949
+ 2
950
+ 1.00
951
+ 0
952
+ 0.10
953
+ 0
954
+ 20
955
+ 40
956
+ 60
957
+ 80
958
+ 100
959
+ Deg (°)Mag = 32.58 K X
960
+ 1 μm
961
+ WD = 3.4 mm
962
+ EHT = 5.00 kV
963
+ Date :27 Nov 2019
964
+ IBN
965
+ signal A = InLensa)
966
+ b)
967
+ -4
968
+ -2
969
+ 0
970
+ 2
971
+ 4
972
+ 6
973
+ 0
974
+ 1
975
+ V (hf/2e)
976
+ RF power (dBm)
977
+ dV/dI (Ω)
978
+ -2
979
+ -1
980
+ 3
981
+ 2
982
+ -3
983
+ 8
984
+ 10
985
+ 600
986
+ 450
987
+ 300
988
+ 150
989
+ 0
990
+ c)
991
+ -20
992
+ -16
993
+ 0
994
+ 2
995
+ V (hf/2e)
996
+ RF power (dBm)
997
+ dV/dI (Ω)
998
+ -2
999
+ -12
1000
+ -8
1001
+ d)
1002
+ -12
1003
+ -8
1004
+ 0
1005
+ V (hf/2e)
1006
+ RF power (dBm)
1007
+ dV/dI (Ω)
1008
+ -2
1009
+ 2
1010
+ -4
1011
+ 0
1012
+ -12
1013
+ -8
1014
+ 0
1015
+ 2
1016
+ V (hf/2e)
1017
+ RF power (dBm)
1018
+ dV/dI (Ω)
1019
+ -4
1020
+ -2
1021
+ 4
1022
+ -16
1023
+ f = 9.75 GHz
1024
+ f = 5.25 GHz
1025
+ f = 3.0 GHz
1026
+ f = 2.1 GHz
1027
+ 600
1028
+ 450
1029
+ 300
1030
+ 150
1031
+ 0
1032
+ 600
1033
+ 450
1034
+ 300
1035
+ 150
1036
+ 0
1037
+ 600
1038
+ 450
1039
+ 300
1040
+ 150
1041
+ 0
1042
+ -4
1043
+ 0
1044
+ -6
1045
+ 6
1046
+ FIG. 6. Shapiro response of Josephson junction #3 at different radio-frequencies ap-
1047
+ plied. a), b), c) and d) show the differential resistance as a function of the radio-frequency ex-
1048
+ citation amplitude/radio-frequency power (P in dBm) and the potential bias (V in hf/2e) of the
1049
+ junction (dV/dI(P, V )) at f = 9.75 GHz, f = 5.25 GHz, f = 3.0 GHz and f = 2.1 GHz, respectively.
1050
+ The differential resistance is displayed in between values of dV/dI = 0 (red) and dV/dI = 600 Ω.
1051
+ Next to Shapiro steps at full integer values of V = n · hf/2e, there are sub-integer steps visible for
1052
+ an applied radio-frequency of f = 9.75 GHz.
1053
+ in the main manuscript. Next to these standard junction characeristics we here show Shapiro
1054
+ step measurements of the narrow junction #3, shown in Fig. 6. Within a similar range of
1055
+ radiofrequencies applied, as for junction #2 in the main text, we observe a similar behavior
1056
+ w.r.t. the Shapiro step evolution in the differential resistance as a function of the applied RF
1057
+ power applied to and the d.c. potential bias applied across the junction (dV/dI(P, V )). For
1058
+ the largest frequency applied of f = 9.75 GHz, not only full integer Shapiro steps, but also
1059
+ half-integer Shapiro steps can be observed. This demonstrates that the existence of high
1060
+ coherent ballistic channels do not seem to change with the width of the nanoribbon, as they
1061
+ 20
1062
+
1063
+ 0.000
1064
+ 80.00
1065
+ 160.0
1066
+ 240.0
1067
+ 320.00.000
1068
+ 80.00
1069
+ 160.0
1070
+ 240.0
1071
+ 320.00.000
1072
+ 80.00
1073
+ 160.0
1074
+ 240.0
1075
+ 320.00.000
1076
+ 80.00
1077
+ 160.0
1078
+ 240.0
1079
+ 320.0would in topological insulator nanoribbons, where a quantization of transverse momentum
1080
+ states would alter the surface state dispersion.
1081
+ 21
1082
+
3dE2T4oBgHgl3EQfjgcV/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
3tE2T4oBgHgl3EQfOAZw/content/tmp_files/2301.03743v1.pdf.txt ADDED
@@ -0,0 +1,636 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.03743v1 [astro-ph.SR] 10 Jan 2023
2
+ 1
3
+ LS And: WZ Sge-type outburst first time since the 1971 discovery
4
+ Taichi Kato1
5
+ tkato@kusastro.kyoto-u.ac.jp
6
+ 1 Department of Astronomy, Kyoto University, Sakyo-ku, Kyoto 606-8502, Japan
7
+ Abstract
8
+ LS And was a transient discovered in 1971 in the M 31 region and it has been argued whether it could
9
+ be an intergalactic nova or a dwarf nova. Using the Zwicky Transient Facility (ZTF) data, I found that the
10
+ object underwent the second known outburst in 2022 April. The behavior was that of a WZ Sge-type dwarf
11
+ nova with a long fading tail and the light curves of the 1971 and 2022 outbursts matched very well. The
12
+ light curves suggest that LS And is a typical WZ Sge-type dwarf nova near (but before reaching) the period
13
+ minimum of cataclysmic variables. The true observed peak of the 1971 outburst was likely 12.2 mag. The
14
+ outburst parameters were similar to those of other WZ Sge-type dwarf novae. The fading tail lasts more than
15
+ a year and the object is still currently on this tail. There was a hint of 0.5-mag temporary brightening on the
16
+ fading tail and the object appears still active after the outburst.
17
+ LS And was discovered by van den Bergh et al. (1973) in the region of M 31 (named “m” in their paper).
18
+ van den Bergh et al. (1973) stated that the object was visible only on a blue and on a yellow plate taken in
19
+ immediate succession on 1971 August 26.
20
+ van den Bergh et al. (1973) suggested that the variable might be
21
+ either a supernova or a flare star. Although van den Bergh et al. (1973) did not give the brightness of this object,
22
+ it was estimated to be 12.5 from their figure by Romano (1977).
23
+ Sharov (1973) examined plates taken in the Crimean Station of Sternberg Astronomical Institute and Latvian
24
+ Radio Astrophysical Observatory. Sharov (1973) succeeded in obtaining one observation near the maximum and
25
+ the light curve of the fading part. Sharov (1973) noted the presence of a star of 21–22 mag on Palomar Observatory
26
+ Sky Survey (POSS). Based on the large amplitude exceeding 8 mag, rapid fading (0.2 mag d−1) in the early fading
27
+ part and the very slow (less than 0.001 mag d−1) fading rate in the late fading part, Sharov (1973) stated that
28
+ the star was unlikely a supernova or a flare star. The light curve, however, did not resemble those of typical
29
+ novae or dwarf novae and Sharov (1973) suggested that it might be a very distant nova (i.e. intergalactic nova)
30
+ if it was indeed a nova.
31
+ Romano (1977) examined Asiago plates and presented a rough light curve of the outburst (probably unaware
32
+ of the work by Sharov 1973). Romano (1977) indicated that the variable was at the limit of visibility (∼20.5 mag)
33
+ on POSS and that color was almost white. Romano (1977) excluded a flare star based on the light curve and
34
+ also a supernova based on the absence of a galaxy near the star. Romano (1977) concluded that this object is
35
+ probably a dwarf nova of UV Per type.1
36
+ Following Romano (1977), Meinunger (1977) studied Sonneberg plates (probably also unaware of the work
37
+ by Sharov 1973) and constructed a light curve. Meinunger (1977) concluded that the star was clearly a fast
38
+ nova and could not be a supernova due to the absence of a galaxy near the star. Meinunger (1977) excluded a
39
+ long-period dwarf nova (like UV Per) based on the facts: (1) the amplitude was larger than 8 mag [Meinunger
40
+ (1977) even suggested that the object on POSS was a unrelated one], (2) the decline after the maximum was too
41
+ fast and (3) no further outbursts were observed. Meinunger (1977) suggested that this object was probably a
42
+ very bright nova in the halo of M 31.
43
+ Sharov and Karimova (1978) and his colleagues examined materials and found new records during the out-
44
+ burst close to the maximum in the collection of Odessa Observatory. Precise astrometry of the outbursting object
45
+ using the materials at Latvian Radio Astrophysical Observatory indicated the identity with the object on POSS.
46
+ Based on the large (9 mag) amplitude, exceeding those of dwarf novae, Sharov and Karimova (1978) considered
47
+ that the object should be regarded as a fast nova despite its small amplitude for a nova. Sharov and Karimova
48
+ (1978) also remarked that the supposed nova did not follow the maximum magnitude relation with decline time
49
+ for M 31 novae (Sharov 1989), and suggested that either the relation was broken or the object was an intergalactic
50
+ nova 100–150 kpc from the Sun. This classification by Sharov and Karimova (1978) was adopted in Duerbeck
51
+ (1987) and LS And was classified as a fast nova in General catalogue of variable stars (GCVS: Kholopov et al.
52
+ 1985).
53
+ In GCVS version 4.2 for extragalatic variables, LS And was also given a name M31V0002 probably
54
+ reflecting the possibility of an object in M 31.
55
+ 1UV Per was considered to be the prototype of dwarf novae with large-amplitude and rare outbursts at that time (cf. Petit 1960).
56
+ WZ Sge was considered as a recurrent nova and the concept of WZ Sge-type dwarf novae was not present. See Kato (2015) for a
57
+ modern review of WZ Sge-type dwarf novae.
58
+
59
+ 2
60
+ Table 1: Observations of the 1971 outburst of LS And.
61
+ JD∗
62
+ mag†
63
+ source‡
64
+ JD∗
65
+ mag†
66
+ source‡
67
+ JD∗
68
+ mag†
69
+ source‡
70
+ 179
71
+ [19.0
72
+ 3
73
+ 223.497
74
+ 18.30
75
+ 2
76
+ 292
77
+ [19.0
78
+ 3
79
+ 183.468
80
+ [20.0
81
+ 2
82
+ 224.511
83
+ 18.56
84
+ 2
85
+ 294
86
+ [19.0
87
+ 3
88
+ 183.508
89
+ [13.6
90
+ 5
91
+ 225.541
92
+ 18.30
93
+ 2
94
+ 296
95
+ [19.0
96
+ 3
97
+ 187.479
98
+ 12.7:
99
+ 5
100
+ 235
101
+ 18.5
102
+ 3
103
+ 298
104
+ [19.0
105
+ 3
106
+ 187.508
107
+ 11.7:
108
+ 5
109
+ 236.248
110
+ 18.80
111
+ 2
112
+ 300
113
+ 19.0:
114
+ 3
115
+ 190
116
+ 12.5
117
+ 1
118
+ 237.261
119
+ 18.83
120
+ 2
121
+ 302
122
+ 19.0
123
+ 3
124
+ 191
125
+ 13.8*
126
+ 4
127
+ 238.405
128
+ 18.83
129
+ 2
130
+ 304
131
+ 19.0
132
+ 3
133
+ 191.504
134
+ 13.60
135
+ 2
136
+ 239.392
137
+ 18.83
138
+ 2
139
+ 305.304
140
+ 18.83
141
+ 2
142
+ 193
143
+ 14.0*
144
+ 4
145
+ 240
146
+ 18.7
147
+ 3
148
+ 308
149
+ 19.0
150
+ 3
151
+ 193.476
152
+ 14.1:
153
+ 5
154
+ 240.407
155
+ 18.83
156
+ 2
157
+ 320
158
+ 19.0:
159
+ 3
160
+ 193.507
161
+ 14.1:
162
+ 5
163
+ 242
164
+ 18.7
165
+ 3
166
+ 324
167
+ 19.0
168
+ 3
169
+ 195.492
170
+ 14.5::
171
+ 5
172
+ 245
173
+ 18.5
174
+ 3
175
+ 332
176
+ [19.0
177
+ 3
178
+ 195.515
179
+ 14.5::
180
+ 5
181
+ 245.339
182
+ 19.0
183
+ 2
184
+ 335.238
185
+ 18.8:
186
+ 2
187
+ 208
188
+ 15.85
189
+ 4
190
+ 246.254
191
+ 18.83
192
+ 2
193
+ 353.24
194
+ 19:
195
+ 2
196
+ 209
197
+ 14.9
198
+ 3
199
+ 249
200
+ 18.5
201
+ 3
202
+ 570.392
203
+ 19.2:
204
+ 2
205
+ 209.359
206
+ 15.80
207
+ 2
208
+ 249.276
209
+ [18.8
210
+ 2
211
+ 575.408
212
+ 19.2:
213
+ 2
214
+ 210
215
+ 16.25
216
+ 4
217
+ 252.434
218
+ 18.8:
219
+ 2
220
+ 655.286
221
+ [18.3
222
+ 2
223
+ 210.499
224
+ 15.98
225
+ 2
226
+ 254.519
227
+ [18.3
228
+ 2
229
+ 681.291
230
+ 19.0
231
+ 2
232
+ 212
233
+ 16.4
234
+ 3
235
+ 263
236
+ 18.5:
237
+ 3
238
+ 682.168
239
+ [19.2
240
+ 2
241
+ 213.486
242
+ 17.54
243
+ 2
244
+ 266.367
245
+ 18.8
246
+ 2
247
+ 684.233
248
+ 19.4
249
+ 2
250
+ 214
251
+ 17.6*
252
+ 4
253
+ 268.427
254
+ 18.83
255
+ 2
256
+ 685.201
257
+ [19.2
258
+ 2
259
+ 215
260
+ 17.8*
261
+ 4
262
+ 271
263
+ 19.0
264
+ 3
265
+ 688.219
266
+ 19.4
267
+ 2
268
+ 217
269
+ 18.0*
270
+ 4
271
+ 276
272
+ [19.0
273
+ 3
274
+ 983
275
+ 20.0:
276
+ 5
277
+ 217.367
278
+ 18.30
279
+ 2
280
+ 276.284
281
+ 18.83
282
+ 2
283
+ 987
284
+ 20.0:
285
+ 5
286
+ 220.358
287
+ 18.33
288
+ 2
289
+ 277.396
290
+ 18.8:
291
+ 2
292
+ 2105
293
+ 20:
294
+ 5
295
+ 221.545
296
+ 18.38
297
+ 2
298
+ 278.308
299
+ 18.9
300
+ 2
301
+ 222.4
302
+ 18.43
303
+ 2
304
+ 280
305
+ [19.0
306
+ 3
307
+ ∗ JD−2441000.
308
+ † [ upper limits. : uncertain. * eye estimate from the published figure.
309
+ ‡ 1: van den Bergh et al. (1973), 2: Sharov (1973), 3: Romano (1977),
310
+ 4: Meinunger (1977), 5: Sharov and Karimova (1978).
311
+
312
+ 3
313
+ 41180
314
+ 41200
315
+ 41220
316
+ 41240
317
+ 12
318
+ 14
319
+ 16
320
+ 18
321
+ 20
322
+ vdB73
323
+ S73
324
+ R77
325
+ M77
326
+ S78
327
+ Figure 1:
328
+ Light curve of the 1971 outburst of LS And using the data in table 1.
329
+ The sources are
330
+ vdB73 (van den Bergh et al. 1973), S73 (Sharov 1973), R77 (Romano 1977), M77 (Meinunger 1977) and S78
331
+ (Sharov and Karimova 1978). The “v” symbols represent upper limits.
332
+ Although most professional astronomers considered or treated LS And as a nova (Downes and Shara 1993;
333
+ Szkody 1994; Collazzi et al. 2009; Evans et al. 2014; Özdönmez et al. 2018), and some suspected to be an X-
334
+ ray nova (Rosenbush 1999) or a recurrent nova (Duerbeck 1988; Pagnotta and Schaefer 2014), I may have been
335
+ the first to become confident that this should be a large-amplitude dwarf nova after knowing this object in
336
+ the freshly published work by Duerbeck (1987). A part of the atmosphere in the late 1980s among amateur
337
+ astronomers was already told in Kato (2022a). Visual monitoring of LS And for a new outburst already started
338
+ in 1987 by VSOLJ members, and then by observers around the world. Although results have not been fruitful
339
+ for decades [now exceeding 6000 observations without detecting an outburst in the American Association of
340
+ Variable Stars (AAVSO)2; I myself had more than 200 non-detection visual observations when I was an amateur
341
+ astronomer], I consistently considered LS And as a candidate WZ Sge star (Kato et al. 2001, 2002). I expected
342
+ that the Gaia satellite would clarify the nature of LS And, but there was no parallax information in Gaia DR2
343
+ (Gaia Collaboration et al. 2018). The blue color (Gaia B − R=+0.25) and a large proper motion were, however,
344
+ sufficient to convince me of the dwarf nova-type nature. The parallax in Gaia EDR3 (Gaia Collaboration et al.
345
+ 2021) was not conclusive, probably due to the faintness of this object. The color in Gaia EDR3 was even bluer
346
+ (B − R=−0.06).
347
+ The “moment” arrived like lightening when I was examining light curves obtained by the Zwicky Transient
348
+ Facility (ZTF: Masci et al. 2019)3. It was when I started examining light curves of recent ZTF data. As usual,
349
+ I was looking at the table of dwarf novae listed in alphabetical order, and almost unconsciously typed LS And
350
+ (as a matter of fact, I already did not pay special attention to this object regularly since I knew that it had
351
+ been well monitored by amateur observers and considered that no missed outburst would be expected in the
352
+ ZTF data). The reason why I specially selected LS And was unknown, but the light curve on the display was a
353
+ familiar one of a WZ Sge star. I initially considered that I entered a name of a different well-known WZ Sge star
354
+ (almost unconsciously as a routine work), but realized that it was “LS And”. Unthinkable! I initially could not
355
+ 2<http://www.aavso.org/data-download>.
356
+ 3The
357
+ ZTF
358
+ data
359
+ can
360
+ be
361
+ obtained
362
+ from
363
+ IRSA
364
+ <https://irsa.ipac.caltech.edu/Missions/ztf.html>
365
+ using
366
+ the
367
+ inter-
368
+ face
369
+ <https://irsa.ipac.caltech.edu/docs/program_interface/ztf_api.html>
370
+ or
371
+ using
372
+ a
373
+ wrapper
374
+ of
375
+ the
376
+ above
377
+ IRSA
378
+ API
379
+ <https://github.com/MickaelRigault/ztfquery>.
380
+
381
+ 4
382
+ 59700
383
+ 59720
384
+ 59740
385
+ 59760
386
+ 12
387
+ 14
388
+ 16
389
+ 18
390
+ 20
391
+ ZTF r
392
+ ZTF g
393
+ ATLAS o
394
+ ATLAS c
395
+ ASN g
396
+ Figure 2:
397
+ Light curve of the 2022 outburst of LS And using ZTF, ATLAS and ASAS-SN data. There were no
398
+ upper limit observations before the initial detection.
399
+ 41180
400
+ 41200
401
+ 41220
402
+ 41240
403
+ 12
404
+ 14
405
+ 16
406
+ 18
407
+ 20
408
+ vdB73
409
+ S73
410
+ R77
411
+ M77
412
+ S78
413
+ 2022
414
+ Figure 3:
415
+ Comparison of light curves of the 1971 and 2022 outburst of LS And. The symbols for the 1971
416
+ observations are the same as in figure 1. The 2022 data (ZTF r magnitudes) were shifted by 18503 d.
417
+
418
+ 5
419
+ 59500
420
+ 59600
421
+ 59700
422
+ 59800
423
+ 59900
424
+ 12
425
+ 14
426
+ 16
427
+ 18
428
+ 20
429
+ 22
430
+ ZTF r
431
+ ZTF g
432
+ ATLAS o
433
+ ATLAS c
434
+ ASN g
435
+ Figure 4:
436
+ Long-term light curve of the 2022 outburst of LS And. The symbols are the same as in figure 2.
437
+ believe my eyes, but it was indeed LS And and I almost automatically issued vsnet-alert 272674, even without
438
+ sufficient patience for waiting the result of a query to the All-Sky Automated Survey for Supernovae (ASAS-SN)
439
+ Sky Patrol data (Shappee et al. 2014; Kochanek et al. 2017). My emotion at that time may have been similar to
440
+ a situation when I encountered a rare bird which I could not believe (cf. Kato 2022a). Birders will agree.
441
+ In the world of birders, it must have become the busiest moment after any discovery — one needs to locate
442
+ the bird and take images or recordings sufficient for a proof of the existence of a rare bird. The case for the
443
+ detection of the 2022 outburst of LS And was different. There was no special care for preserving the data shown
444
+ on the display, and I went to the library (fortunately very close) to search the light curve of the 1971 outburst,
445
+ which still stayed deep in my memory even after decades.
446
+ So it’s time to return to science. In table 1, I summarized photometric data for the 1971 outburst. The
447
+ magnitudes were all photographic (equivalent to B). Magnitudes with * were estimated by my eyes from the
448
+ figure in Meinunger (1977), which are probably correct to ±1 d and ±0.1 mag. The magnitude for JD=190 was
449
+ similarly estimated from a published figure by Romano (1977). Meinunger (1977) claimed that the object was
450
+ estimated too bright by Romano (1977). The light curve drawn from these data is presented in figure 1. This
451
+ is not much different from the one published in Sharov and Karimova (1978), but is worth presenting here since
452
+ Sharov and Karimova (1978) is difficult to reach.
453
+ The 2022 light curve is shown in figure 2. It is very clear that the 1971 and 2022 light curves are very
454
+ similar: plateau-type fading lasting for ∼20 d followed by rapid decline and subsequent slow fading. They are
455
+ typical WZ Sge-type outbursts without rebrightening (type D superoutburst in Kato 2015). It is also well-known
456
+ that the same WZ Sge star tends to repeat the same type of rebrightening (Kato 2015) and LS And is also
457
+ the case. Although the mechanism of rebrightening(s) is not yet well understood, empirical relationship shows
458
+ that WZ Sge stars without rebrightening are mostly objects near the period minimum of cataclysmic variables,
459
+ but before reaching it (figure 17 in Kato 2015). The orbital period of LS And is thus expected to be within
460
+ 0.053–0.060 d. The fading rate of the plateau phase (BJD 2459696–2459714.5) was 0.089(1) mag d−1, which
461
+ corresponds to log td=1.05, a typical value for a WZ Sge star without rebrightening and not resembling a period
462
+ bouncer (see figure 87 in Kato et al. 2014). A comparison between the 1971 and 2022 outbursts is shown in figure
463
+ 3 (from now on, I treat all photometric bands in visual wavelengths almost identical with V , which is a good
464
+ approximation for a WZ Sge star in outburst). These outbursts were almost exactly the same and the interval of
465
+ 4<http://ooruri.kusastro.kyoto-u.ac.jp/mailarchive/vsnet-alert/27267>.
466
+
467
+ 6
468
+ these two outburst was 18503 d (=50.66 yr). This comparison suggests that the 2022 outburst would not have
469
+ started before JD 2459682 (2022 April 12). Definitely a sigh! (particularly for amateur observers) considering
470
+ the almost no evening visibility of this object in mid-April.
471
+ People may wonder if these outburst could be those of an SU UMa star rather than a WZ Sge star, and how
472
+ I can be confident about the classification without observation of early superhumps (cf. Kato 2015). I show a
473
+ long-term light curve of the 2022 outburst in figure 2. The object was brighter by 1.5 mag after the outburst. The
474
+ post-outburst phenomenon is a long fading tail, which is characteristic to a WZ Sge-type outburst and not seen
475
+ in an SU UMa star. The presence of the same phenomenon was also reported after the 1971 outburst (Sharov
476
+ 1973).5 Before the outburst plateau, there was a phase with more rapid fading (more evident in the 1971 light
477
+ curve and only one day in the 2022 one). This feature is commonly seen in WZ Sge-type outbursts and is referred
478
+ to as a viscous decay phase. Early superhumps are expected during this phase if the binary has a sufficient
479
+ inclination (Kato 2015, 2022b).
480
+ The peak magnitude probably requires re-examination. Although most literature gives 11.7 mag as the
481
+ maximum for LS And, it is evident from table 1 that this magnitude was uncertain (“:” usually means that
482
+ the object is close to the limit of photographic materials or the quality of the photograph is poor) and was the
483
+ brighter one of two uncertain observations (11.7 and 12.7 mag) only 40 min apart. It looks more likely that the
484
+ true brightest observation was close to their average (12.2 mag). The outburst amplitude based on this value is
485
+ 8.8 mag using the ZTF data before the 2022 outburst. The true peak would have been brighter, though, since
486
+ there was a 4 d observational gap before the first observation of the outburst (but see the discussion below).
487
+ As seen from the 2022 observations, the magnitude when ordinary superhumps should appear following the
488
+ viscous decay phase was 14.3 mag. In ordinary WZ Sge stars, the absolute magnitude (MV ) when ordinary
489
+ superhumps appear is +5.4 (for an average inclination of 1 radian) (Kato 2022b).
490
+ Using this value as the
491
+ standard candle, the distance modulus of LS And is estimated to be 8.9. The observed peak (12.2 mag) in
492
+ 1971 corresponds to MV =+3.3. The quiescent magnitude (21.0 mag, ZTF data) corresponds to MV =+12.1.
493
+ The difference (6.7 mag) between quiescent magnitude and the magnitude when ordinary superhumps appear is
494
+ typical for a (non-period bouncer) WZ Sge star (see fig. 23 in Kato 2015; Tampo et al. 2020). Other properties
495
+ of LS And are expected to be similar to those of typical WZ Sge stars.
496
+ The detection of the 2022 outburst of LS And brought a some kind of despair to observers who had been
497
+ expecting to see a fresh outburst for decades.
498
+ Could there be a possibility that LS And silently underwent
499
+ outbursts more frequently only around solar conjunctions? This was indeed the case of the SU UMa star VY Aqr
500
+ located close to the ecliptic.
501
+ Despite the mean interval of superoutbursts of less than 2 yr, this object was
502
+ not recorded in superoutburst between 1994 and 2006, and between 2008 and 2020. It was most likely that
503
+ superoutbursts in this object occurred around solar conjunctions and were not recorded. Although similar things
504
+ may have happened in LS And at least in the past, modern deep observations such as ZTF should have detected
505
+ the object during a fading tail if there was a missed superoutburst. There was no indication of such a detection in
506
+ the ZTF data since 2018, and the outburst interval should be longer than 5 yr. The fading tail lasted more than a
507
+ year (Sharov 1973). Sharov and Karimova (1978) described that the object returned to practically the same level
508
+ before the outburst after 5.5 yr, although this description may have assumed a nova-type light curve and could
509
+ have overestimated the duration of the fading tail. Considering these values and considering that parameters of
510
+ LS And are similar to those of typical WZ Sge stars, the next major outburst would be expected after a decade or
511
+ even more [see figure 5 in Kato (2015) for the distribution of outburst intervals in WZ Sge stars]. By comparing
512
+ the recorded peak MV =+3.3 (in 1971) with the statistics of known WZ Sge stars (figure 10 in Tampo et al. 2020),
513
+ it appears that the true peak in 1971 was not missed after a considerable delay (i.e. the object was unlikely to
514
+ have reached 11.0 mag even at the true peak). The next superoutburst would also be around 12.2 mag. There
515
+ are, however, exceptional objects like V3101 Cyg (Tampo et al. 2020; Hameury and Lasota 2021) and there may
516
+ be an unexpected phenomenon even after the outburst. In the post-outburst data of LS And, 0.5 mag brightening
517
+ lasting for 10–20 d and starting around JD 2459852 was present (figure 3). This might suggest that LS And is
518
+ still active in the post-superoutburst phase and would be worth observing before it finally returns to quiescence.
519
+ Acknowledgements
520
+ This work was supported by JSPS KAKENHI Grant Number 21K03616. The author is grateful to the ZTF,
521
+ ATLAS and ASAS-SN teams for making their data available to the public. I am grateful to VSOLJ, AAVSO and
522
+ 5It might be interesting to leave a remark that the figure in Sharov (1973) dealt with this phenomenon rather than the shape of
523
+ the outburst. Please have a look at his figure if you have a chance too see this reference.
524
+
525
+ 7
526
+ VSNET observers for reporting observations and to Naoto Kojiguchi for helping downloading the ZTF data.
527
+ Based on observations obtained with the Samuel Oschin 48-inch Telescope at the Palomar Observatory as
528
+ part of the Zwicky Transient Facility project. ZTF is supported by the National Science Foundation under Grant
529
+ No. AST-1440341 and a collaboration including Caltech, IPAC, the Weizmann Institute for Science, the Oskar
530
+ Klein Center at Stockholm University, the University of Maryland, the University of Washington, Deutsches
531
+ Elektronen-Synchrotron and Humboldt University, Los Alamos National Laboratories, the TANGO Consortium
532
+ of Taiwan, the University of Wisconsin at Milwaukee, and Lawrence Berkeley National Laboratories. Operations
533
+ are conducted by COO, IPAC, and UW.
534
+ The ztfquery code was funded by the European Research Council (ERC) under the European Union’s Horizon
535
+ 2020 research and innovation programme (grant agreement n◦759194 – USNAC, PI: Rigault).
536
+ This work has made use of data from the Asteroid Terrestrial-impact Last Alert System (ATLAS) project.
537
+ The Asteroid Terrestrial-impact Last Alert System (ATLAS) project is primarily funded to search for near earth
538
+ asteroids through NASA grants NN12AR55G, 80NSSC18K0284, and 80NSSC18K1575; byproducts of the NEO
539
+ search include images and catalogs from the survey area. This work was partially funded by Kepler/K2 grant
540
+ J1944/80NSSC19K0112 and HST GO-15889, and STFC grants ST/T000198/1 and ST/S006109/1. The ATLAS
541
+ science products have been made possible through the contributions of the University of Hawaii Institute for
542
+ Astronomy, the Queen’s University Belfast, the Space Telescope Science Institute, the South African Astronomical
543
+ Observatory, and The Millennium Institute of Astrophysics (MAS), Chile.
544
+ List of objects in this paper
545
+ LS And, VY Aqr, V3101 Cyg, UV Per, WZ Sge, SU UMa, M 31, M31V0002
546
+ References
547
+ We provide two forms of the references section (for ADS and as published) so that the references can be easily
548
+ incorporated into ADS.
549
+ References (for ADS)
550
+ Collazzi, A. C., Schaefer, B. E., Xiao, L., Pagnotta, A., Kroll, P., Löchel, K., & Henden, A. A. 2009, AJ, 138,
551
+ 1846 (arXiv:0909.4289)
552
+ Downes, R. A., & Shara, M. M. 1993, PASP, 105, 127 (https://doi.org/10.1086/133139)
553
+ Duerbeck, H. W. 1988, A&A, 197, 148
554
+ Duerbeck, H. W. 1987, Space Sci. Rev., 45, 1 (https://doi.org/10.1007/BF00187826)
555
+ Evans, A., Gehrz, R. D., Woodward, C. E., & Helton, L. A. 2014, MNRAS, 444, 1683 (arXiv:1407.5570)
556
+ Gaia Collaboration, et al. 2018, A&A, 616, A1 (arXiv:1804.09365)
557
+ Gaia Collaboration, et al. 2021, A&A, 649, A1 (arXiv:2012.01533)
558
+ Hameury, J.-M., & Lasota, J.-P. 2021, A&A, 650, A114 (arXiv:2104.02952)
559
+ Kato, T. 2015, PASJ, 67, 108 (arXiv:1507.07659)
560
+ Kato, T. 2022a, VSOLJ Variable Star Bull., 89, (arXiv:2201.02945)
561
+ Kato, T. 2022b, VSOLJ Variable Star Bull., 90, (arXiv:2202.02956)
562
+ Kato, T., et al. 2014, PASJ, 66, 30 (arXiv:1310.7069)
563
+ Kato, T., Sekine, Y., & Hirata, R. 2001, PASJ, 53, 1191 (arXiv:astro-ph/0110207)
564
+ Kato, T., Uemura, M., Matsumoto, K., Kinnunen, T., Garradd, G., Masi, G., & Yamaoka, H. 2002, PASJ, 54,
565
+ 999 (arXiv:astro-ph/0209283)
566
+
567
+ 8
568
+ Kholopov, P. N., et al. 1985, General Catalogue of Variable Stars, fourth edition (Moscow: Nauka Publishing
569
+ House)
570
+ Kochanek, C. S., et al. 2017, PASP, 129, 104502 (arXiv:1706.07060)
571
+ Masci, F.-J., et al. 2019, PASP, 131, 018003 (arXiv:1902.01872)
572
+ Meinunger, L. 1977, IBVS, 1331, 1
573
+ Özdönmez, A., Ege, E., Güver, T., & Ak, T. 2018, MNRAS, 476, 4162 (arXiv:1802.05725)
574
+ Pagnotta, A., & Schaefer, B. E. 2014, ApJ, 788, 164 (arXiv:1405.0246)
575
+ Petit, M. 1960, Journal des Observateurs, 43, 17
576
+ Romano, G. 1977, AJ, 82, 319 (https://doi.org/10.1086/112052)
577
+ Rosenbush, A. E. 1999, Astrophysics, 42, 270 (https://doi.org/10.1007/BF02700757)
578
+ Shappee, B. J., et al. 2014, ApJ, 788, 48 (arXiv:1310.2241)
579
+ Sharov, A. S. 1973, Astron. Tsirk., 793, 1
580
+ Sharov, A. S. 1989, Soviet Astronomy Letters, 15, 5
581
+ Sharov, A. S., & Karimova, D. K. 1978, Astron. Tsirk., 998, 1
582
+ Szkody, P. 1994, AJ, 108, 639 (https://doi.org/10.1086/117098)
583
+ Tampo, Y., et al. 2020, PASJ, 72, 49 (arXiv:2004.10508)
584
+ van den Bergh, S., Herbst, E., & Pritchet, C. 1973, AJ, 78, 375 (https://doi.org/10.1086/111426)
585
+ References (as published)
586
+ Collazzi, A. C., Schaefer, B. E., Xiao, L., Pagnotta, A., Kroll, P., Löchel, K., & Henden, A. A. (2009) The
587
+ behavior of novae light curves before eruption. AJ 138, 1846
588
+ Downes, R. A., & Shara, M. M. (1993) A catalog and atlas of cataclysmic variables. PASP 105, 127
589
+ Duerbeck, H. W. (1988) V394 CrA – outburst light curves and notes on its position among the recurrent novae.
590
+ A&A 197, 148
591
+ Duerbeck, H. W. (1987) A reference catalogue and atlas of galactic novae. Space Sci. Rev. 45, 1
592
+ Evans, A., Gehrz, R. D., Woodward, C. E., & Helton, L. A. (2014) A WISE view of novae – I. the data. MNRAS
593
+ 444, 1683
594
+ Gaia Collaboration et al. (2018) Gaia Data Release 2. Summary of the contents and survey properties. A&A
595
+ 616, A1
596
+ Gaia Collaboration et al. (2021) Gaia Early Data Release 3. Summary of the contents and survey properties.
597
+ A&A 649, A1
598
+ Hameury, J.-M., & Lasota, J.-P. (2021) Modelling rebrightenings, reflares, and echoes in dwarf nova outbursts.
599
+ A&A 650, A114
600
+ Kato, T. (2015) WZ Sge-type dwarf novae. PASJ 67, 108
601
+ Kato, T. (2022a) Evolution of short-period cataclysmic variables: implications from eclipse modeling and stage
602
+ a superhump method (with New Year’s gift). VSOLJ Variable Star Bull. 89, (arXiv:2201.02945)
603
+ Kato, T. (2022b) Emerging ordinary superhumps as the standard candle for WZ Sge stars. VSOLJ Variable Star
604
+ Bull. 90, (arXiv:2202.02956)
605
+
606
+ 9
607
+ Kato, T. et al. (2014) Survey of period variations of superhumps in SU UMa-type dwarf novae. V: The fifth year
608
+ (2012–2013). PASJ 66, 30
609
+ Kato, T., Sekine, Y., & Hirata, R. (2001) HV Vir and WZ Sge-type dwarf novae. PASJ 53, 1191
610
+ Kato, T., Uemura, M., Matsumoto, K., Kinnunen, T., Garradd, G., Masi, G., & Yamaoka, H. (2002) WZ Sge-type
611
+ star V592 Herculis. PASJ 54, 999
612
+ Kholopov, P. N. et al. (1985) General Catalogue of Variable Stars, fourth edition (Moscow: Nauka Publishing
613
+ House)
614
+ Kochanek, C. S. et al. (2017) The All-Sky Automated Survey for Supernovae (ASAS-SN) light curve server v1.0.
615
+ PASP 129, 104502
616
+ Masci, F.-J. et al. (2019) The Zwicky Transient Facility: Data processing, products, and archive. PASP 131,
617
+ 018003
618
+ Meinunger, L. (1977) A bright nova in the surroundings of the Andromeda Nebula. IBVS 1331, 1
619
+ Özdönmez, A., Ege, E., Güver, T., & Ak, T. (2018) A new catalogue of Galactic novae: investigation of the
620
+ MMRD relation and spatial distribution. MNRAS 476, 4162
621
+ Pagnotta, A., & Schaefer, B. E. (2014) Identifying and quantifying recurrent novae masquerading as classical
622
+ novae. ApJ 788, 164
623
+ Petit, M. (1960) Catalogue des Étoiles variables du type U Geminorum. Journal des Observateurs 43, 17
624
+ Romano, G. (1977) Some variable stars in the field of M31. AJ 82, 319
625
+ Rosenbush, A. E. (1999) X-ray nova candidates among old classical novae. Astrophysics 42, 270
626
+ Shappee, B. J. et al. (2014) The man behind the curtain: X-rays drive the UV through NIR variability in the
627
+ 2013 AGN outburst in NGC 2617. ApJ 788, 48
628
+ Sharov, A. S. (1973) Interesting variable star. Astron. Tsirk. 793, 1
629
+ Sharov, A. S. (1989) The bright novae in M31 and the relation between peak brightness and rate of decline. Soviet
630
+ Astronomy Letters 15, 5
631
+ Sharov, A. S., & Karimova, D. K. (1978) New data on an interesting variable star. Astron. Tsirk. 998, 1
632
+ Szkody, P. (1994) BVRJK observations of northern hemisphere old novae. AJ 108, 639
633
+ Tampo, Y. et al. (2020) First detection of two superoutbursts during the rebrightening phase of a WZ Sge-type
634
+ dwarf nova: TCP J21040470+4631129. PASJ 72, 49
635
+ van den Bergh, S., Herbst, E., & Pritchet, C. (1973) A search for faint variable objects. AJ 78, 375
636
+
3tE2T4oBgHgl3EQfOAZw/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
4tAyT4oBgHgl3EQfcPfW/content/2301.00280v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6b2b026c5aa88f16d68d19fe2deeb0a9df90ce97f9f5a8875833c2518dbf8093
3
+ size 1099662
4tAyT4oBgHgl3EQfcPfW/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:40d71cf89bff1d34dd8674c7b62ce11b1b6f55243c1f0734f99ee5f26ff4c77d
3
+ size 4980781
5dAyT4oBgHgl3EQf2fkR/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:89e229e02e8500d248d0c1c9315df4363189d584ab4e1041c3b23162bbd215d0
3
+ size 197256
5tFAT4oBgHgl3EQfmx2I/content/2301.08625v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5a40f37a3774ed5d3f8bee06fb9af892f03c6d88dac384c815d7e82a33bb85f3
3
+ size 392088
5tFAT4oBgHgl3EQfmx2I/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fff9a77556c6c87b76079f113dbd599ea02f50a142ee7564a39b76c2dfb79aad
3
+ size 100783
79AyT4oBgHgl3EQfQvZ4/content/tmp_files/2301.00052v1.pdf.txt ADDED
@@ -0,0 +1,378 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.00052v1 [math.GR] 30 Dec 2022
2
+ Examples of left-orderable and
3
+ non-left-orderable HNN extensions
4
+ Azer Akhmedov, Cody Martin
5
+ ABSTRACT: We prove that an HNN extension of a torsion-free nilpotent
6
+ group is left-orderable. We also construct examples of non-left-orderable
7
+ HNN extensions of left-orderable groups.
8
+ 1. Non-left-orderable HNN extensions of left-orderable
9
+ groups
10
+ It is well-known that an HNN extension of a torsion-free group is still
11
+ torsion-free ([3], [1]). On the hand, for many classes of groups, existence
12
+ of a torsion element is the only obstruction to left-orderability; for
13
+ example, this is the case for the classes of one-relator groups, nilpotent
14
+ groups, etc. Hence it is natural to study how left-orderability behaves
15
+ under an HNN extension.
16
+ In [2] (see Example 6.2 there), an example is constructed to show
17
+ that left-orderability is not preserved under the HNN extension. In
18
+ this section, we present systematic ways of producing non-left-orderable
19
+ HNN extensions of left-orderable groups. The example of [2] is built
20
+ as an HNN extension of a direct product of a free nilpotent group of
21
+ class two with the fundamental group of Klein bottle.
22
+ We produce
23
+ examples of HNN extensions of groups such as non-Abelian free groups
24
+ and virtually Abelian groups.
25
+ We rely on the following well-known
26
+ criterion about left-orderability of groups [4]
27
+ Proposition 1.1. A group G is left-orderable if and only if for all
28
+ k ≥ 1 and for all g1, . . . , gk ∈ G\{1}, there exist ǫ1, . . . , ǫk ∈ {−1, 1}
29
+ such that the semigroup of G generated by gǫ1
30
+ 1 , . . . , gǫk
31
+ k does not contain
32
+ the identity element.
33
+ Let us emphasize that we use the obvious “only if part” of this propo-
34
+ sition; the harder “if part” is not needed.
35
+ Given a group G, and subgroups A, B ≤ G with an isomorphism
36
+ φ : A → B, the HNN extension (G, A, B, t, φ) is defined as the quo-
37
+ tient of the free product G ∗ ⟨t⟩ by the normal closure of the subset
38
+ {tat−1φ(a)−1 | a ∈ A}. We also write this HNN extension as (G, A, B, t)
39
+ when φ is given in the context.
40
+ Theorem 1.2. A free group of rank bigger than one admits a non-left-
41
+ orderable HNN extension.
42
+ 1
43
+
44
+ 2
45
+ Proof. By Britton’s Lemma, it suffices to prove the theorem for the
46
+ group F2. Let a, b be the generators of F2. We can find positive expo-
47
+ nents pi, qi, ri, si, 1 ≤ i ≤ 8 such that the elements
48
+ u1 = ap1bq1, u2 = ap2bq2, u3 = ap3bq3, u4 = ap4bq4,
49
+ u5 = ap5b−q5, u6 = ap6b−q6, u7 = ap7b−q7, u8 = ap8b−q8
50
+ generate a free group of rank 8, and so do the elements
51
+ v1 = ar1bs1, v2 = ar2b−s2, v3 = a−r3bs3, v4 = a−r4b−s4,
52
+ v5 = ar5bs5, v6 = ar6b−s6, v7 = a−r7bs7, v8 = a−r8b−s8.
53
+ (It suffices to take the sequences (pi)1≤i≤8, (qi)1≤i≤8, (ri)1≤i≤8, (si)1≤i≤8
54
+ to be strictly increasing.) Let A, B be these free groups generated by
55
+ u1, . . . , u8 and v1, . . . , v8 respectively, and φ : A → B be the isomor-
56
+ phism such that φ(ui) = vi, 1 ≤ i ≤ 8.
57
+ Then, by Proposition 1.1, the HNN extension (G, A, B, t) where
58
+ t(a) = φ(a) for all a ∈ A is not left-orderable.
59
+
60
+ Remark 1.3. Let us remind that in the case of rank = 1, the claim
61
+ does not hold anymore since any HNN extension of Z is isomorphic
62
+ ⟨t, a | tamt−1 = an⟩ for some non-zero integers m, n. All these groups
63
+ (which include Z2, π1(Klein bottle) = ⟨a, b | aba−1 = b−1⟩, and the
64
+ solvable Baumslag-Solitar group BS(1, n) ∼= Z ⋉ Z[ 1
65
+ n]), are all left-
66
+ orderable as torsion-free one-relator groups.
67
+ Using similar ideas, we build a non-left-orderable HNN extension
68
+ of a left-orderable solvable group. We again rely on the criterion of
69
+ Proposition 1.1.
70
+ Let n ≥ 2 and Γn be a group given by the presentation
71
+ ⟨s, x | [sn, x] = 1, [x, sixs−i] = 1, 1 ≤ i ≤ n − 1⟩.
72
+ Let xi = sixs−i, i ∈ Z. Notice that xi = xj iff i ≡ j( mod n). The
73
+ elements xi, 0 ≤ i ≤ n−1 generate a normal subgroup Nn isomorphic to
74
+ Zn and the quotient by this subgroup is isomorphic to Z. Any element
75
+ g of Γn can be written uniquely as siw(x0, . . . , xn−1) where i ∈ Z and
76
+ w(x0, . . . , xn−1) = xp0
77
+ 0 . . . xpn−1
78
+ n−1 for some integer exponents p0, . . . , pn−1.
79
+ siw(x0, . . . , xn−1) will be called the canonical form of g. We also write
80
+ Σ(g) = i + p0 + · · · + pn.
81
+ Let us observe that Γn is torsion-free. Indeed, if g is a torsion element
82
+ with a canonical form siw(x0, . . . , xn−1) as above then for all k ≥ 1,
83
+ gk = sikw0(x0, . . . , xn−1)wi(x0, . . . , xn−1) . . . w(k−1)i(x0, . . . , xn−1)
84
+
85
+ 3
86
+ where wj(x0, . . . , xn−1) = w(xj, xj+1, . . . , xn−1+j) hence it follows im-
87
+ mediately that either i = 0; then, since Nn ∼= Zn, we obtain that
88
+ w = 1.
89
+ It turns out Γn is left-orderable (which also implies that it is torsion-
90
+ free). We introduce a left order < on Γn as follows: An element g with
91
+ the canonical form siw(x0, . . . , xn−1) as above will be called positive
92
+ if either Σ(w) > 0 or Σ(w) = 0 and i > 0. If Σ(w) = 0 and i = 0,
93
+ then we are in the group Nn ∼= Zn and there the order can be defined
94
+ lexicographically. Then we see that a product of two positive elements
95
+ is always positive and the inverse of a positive element is not positive.
96
+ Hence < is a left-order.
97
+ To state our next proposition we need to introduce some (well-
98
+ known) terminology.
99
+ Definition 1.4. Let G be a group generated by a subset S ⊆ G\{1}
100
+ such that for all x ∈ G, if x ∈ S, then x−1 /∈ S (in particular, 1 /∈ S).
101
+ We say that a non-trivial reduced word W(x1, . . . , xk) = xn1
102
+ 1 . . . xnk
103
+ k
104
+ is
105
+ positive in the alphabet S if x1, . . . , xk ∈ S and all exponents ni, 1 ≤
106
+ i ≤ k are positive.
107
+ Proposition 1.5. In the group Γn let S1 = {s, x}, S2 = {s−1, x}, S3 =
108
+ {s, x−1}, S4 = {s−1, x−1}. For n ≥ 12, there exists elements f1, . . . , f4,
109
+ g1, . . . , g4 ∈ Γn such that the following conditions hold:
110
+ i) ⟨f1, f2, f3, f4⟩ ∼= ⟨g1, g2, g3, g4⟩ ∼= Z4,
111
+ ii) The elements f1, f2, f3, f4 can be represented with positive words
112
+ in the alphabet S1,
113
+ iii) For all 1 ≤ i ≤ 4, the element gi can be represented with a
114
+ positive word in the alphabet Si.
115
+ Proof. We define f1 = sn−1xs, f2 = sn−2(xs)2, f3 = sn−4(xs)4, f4 =
116
+ sn−8(xs)8. Then f1, f2, f3, f4 belong to Nn and generate a subgroup
117
+ isomorphic to Z4. We also define g1 = sn−1xs, g2 = sn−2(x−1s)2, g3 =
118
+ s4−n(xs−1)4, f4 = s8−n(x−1s−1)8. The elements g1, g2, g3, g4 also belong
119
+ to Nn and generate a subgroup isomorphic to Z4.
120
+
121
+ In the above proposition, the n ≥ 12 is not necessarily the best possi-
122
+ ble. Using Proposition 1.5, we can now prove the following proposition
123
+ which establishes the existence of a non-left-orderable HNN extension
124
+ of a left-orderable virtually Abelian group.
125
+
126
+ 4
127
+ Proposition 1.6. For all n ≥ 12, Γn admits a non-left-orderable HNN
128
+ extension.
129
+ Proof. Let f1, . . . , f4, g1, . . . , g4 ∈ Γn be elements satisfying conditions
130
+ 1)-3) of Proposition 1.5. Let φ : ⟨f1, f2, f3, f4⟩ → ⟨g1, g2, g3, g4⟩ be an
131
+ isomorphism such that φ(fi) = gi, 1 ≤ i ≤ 4.
132
+ We consider an HNN extension
133
+ G := (Γn, ⟨f1, f2, f3, f4⟩, ⟨g1, g2, g3, g4⟩, t)
134
+ by letting txt−1 = φ(x) for all x ∈ ⟨f1, f2, f3, f4⟩
135
+ For any left-order on G, notice that the elements tfit−1, 1 ≤ i ≤ 4
136
+ are either all positive or all negative. On the other hand, among the
137
+ elements gi, 1 ≤ i ≤ 4 at least one is positive and one is negative. This
138
+ is a contradiction. Hence G is not left-orderable.
139
+
140
+ 2. HNN extensions of nilpotent groups
141
+ The aim of this section is to prove that unlike solvable groups,
142
+ an HNN extension of a left-orderable nilpotent group is always left-
143
+ orderable. Let us recall that a nilpotent group is left-orderable iff it is
144
+ torsion-free; this claim too does not hold for solvable groups.
145
+ Let us first observe that, since a direct limit of left-orderable groups
146
+ is left-orderable, an HNN extension an HNN extension (G, A.B, t) is
147
+ left-orderable if for all finitely generated subgroups A0 and G0 of G,
148
+ where A0 ≤ A, G0 ⊇ ⟨A0, B0⟩ and B0 = tA0t−1, the HNN extension
149
+ (G0, A0, B0, t) is left-orderable. We will use this observation repeatedly
150
+ in this section.
151
+ We already observed that by classification, an HNN extension of
152
+ an infinite cyclic group is left-orderable. The same holds for an HNN
153
+ extension of any torsion free Abelian group.
154
+ Indeed, it suffices to consider finitely generated Abelian groups so
155
+ let G be a finitely generated torsion-free Abelian group, A, B ≤ G, φ :
156
+ A → B be an isomorphism, and (G, A, B, t) be the HNN extension with
157
+ respect to the isomorphism φ. Let G ∼= Zd and r = rankA = rankB.
158
+ We will assume that G = Zd.
159
+ Then for some linearly independent
160
+ vectors u1, . . . , ur we have A = {c1u1 + · · · + crur : ci ∈ Z, 1 ≤ i ≤ r}
161
+ and similarly for some linearly independent vectors v1, . . . , vr we have
162
+ B = {c1v1 + · · · + crvr
163
+ :
164
+ ci ∈ Z, 1 ≤ i ≤ r}. We let G = R, A =
165
+ {c1u1 + · · · + crur : ci ∈ R, 1 ≤ i ≤ r}, B = {c1v1 + · · · + crvr : ci ∈
166
+ R, 1 ≤ i ≤ r} and φ : A → B be the extension of φ : A → B defined as
167
+ φ(c1u1 + · · · + crur) = c1φ(u1) + · · · + crφ(ur) for all c1, . . . , cr ∈ R.
168
+
169
+ 5
170
+ A key observation here is that even though the isomorphism φ :
171
+ A → B cannot necessarily be extended to G, but one can extend the
172
+ isomorphism φ : A → B to some automorphism F : G → G. Then the
173
+ HNN extension (G, A, B, t) with respect to the isomorphism φ : A → B
174
+ has a quotient isomorphic to the semidirect product Z⋉F G by a normal
175
+ subgroup N ≤ G. Since N and Z ⋉F G are left-orderable we obtain
176
+ that (G, A, B, t) is left-orderable (as an extension of a left-orderable
177
+ group by a left-orderable group). By Britton’s Lemma, (G, A, B, t) is
178
+ a subgroup of (G, A, B, t) hence it is also left-orderable.
179
+ We now would like to carry the same argument for any torsion-free
180
+ nilpotent group. The main issue here is that given a finitely generated
181
+ torsion-free nilpotent group Γ, one needs to construct a completion Γ
182
+ which would resemble the operation Zd → Rd so we can try to use the
183
+ argument in the Abelian case.
184
+ Let R be a commutative ring with identity and n ≥ 1.
185
+ We let
186
+ Un(R) be the group of n × n upper-triangular matrices with 1’s on the
187
+ diagonal. The cases R = R and R = Z will be the most interesting to
188
+ us.
189
+ It is well-known that any finitely generated torsion-free nilpotent
190
+ group Γ embeds in Un(Z) for some n ≥ 1. The Mal’cev completion of
191
+ Un(Z) is Un(R) (and the Mal’cev completion of Zn is Rn) 1, however,
192
+ given an isomorphism φ : A → B of subgroups of Un(Z), although it
193
+ induces an isomorphism φ : A → B but one cannot necessarily extend
194
+ this isomorphism to the entire G. For example, for n = 3, the group
195
+ U3(Z) is isomorphic to the Heisenberg group
196
+ ⟨x, y, z | z = [x, y], [x, z] = [y, z] = 1⟩
197
+ and if we let A = ⟨x⟩, B = ⟨z⟩ and φ(x) = z, then this isomorphism
198
+ cannot be extended to the isomorphism of U3(Z) (or U3(R)). Thus we
199
+ need to define a completion of Γ other than the Mal’cev completion.
200
+ Let Xn,i, 1 ≤ i ≤ n − 1 be the matrix of Un(Z) where all off-diagonal
201
+ entries are zero except the (i + 1, i)-th entry is equal to 1. In order to
202
+ define a more suitable completion of Un(Z) we will extend it first, and at
203
+ the end we will obtain a completion which is ”infinite-dimensional”. Let
204
+ U∞(Z) be a group generated by xk, k ∈ Z such that for all k ∈ Z, n ≥ 1
205
+ the subgroup generated by xk+1, . . . , xk+n−1 is isomorphic to Un(Z)
206
+ through the isomorphism f(xk+j) = Xn,j, 1 ≤ j ≤ n − 1. Notice that
207
+ U∞(Z) is well-defined this way and it contains isomorphic copies of all
208
+ 1in the literature, the term Mal’cev completion is used for some other related
209
+ operations as well.
210
+
211
+ 6
212
+ Un(Z), n ≥ 2. This group can be viewed as the group of infinite sized
213
+ integral unipotent matrices. But to achieve our goal we extend U∞(Z)
214
+ further as follows.
215
+ Let us first observe that in the group Un(Z) viewed as the group of
216
+ upper triangular unipotent integral matrices, [xi, xj] = 1 if |i − j| ≥ 2
217
+ and for all 1 ≤ i ≤ n − 2, [xi, xi+1] is a unipotent matrix with all the
218
+ off-diagonal entries zero, except the (i + 2, i + 1)-entry equals 1. Thus
219
+ the elements [xi, xi+1], 1 ≤ i ≤ n−2 generate a subgroup isomorphic to
220
+ Un−1(Z) with an isomorphism xi → [xi, xi+1], 1 ≤ i ≤ n − 2. Similarly,
221
+ in the group U∞(Z), the elements [xi, xi+1], i ∈ Z generate a subgroup
222
+ isomorphic to U∞(Z), and the homomorphism f : U∞(Z) → U∞(Z)
223
+ defined as f(xi) = [xi, xi+1], i ∈ Z (it is sufficient to define it on the
224
+ generators) establishes this isomorphism.
225
+ The group U∞(Z) is a direct limit of the groups Un(Z), n ≥ 1.
226
+ More precisely, let Hn, n ≥ 1 be the subgroup of U∞(Z) generated
227
+ by x−n, x−n+1, . . . , xn−1, xn. Then Hn is isomorphic to U2n+1(Z), and
228
+ U∞(Z) is a direct limit of the sequence Hn, n ≥ 1.
229
+ In our construction of the completion, we will use a direct limit of
230
+ groups each isomorphic to U∞(Z).
231
+ Let Γk, k ∈ Z be a group gen-
232
+ erated by zk,n, n ∈ Z with an isomorphism gk : Γk → U∞(Z) such
233
+ that gk(zk,n) = xn. We have · · · ≤ Γ−1 ≤ Γ0 ≤ Γ1 ≤ Γ2 ≤ . . . and
234
+ [zk,n, zk,n+1] = zk−1,n for all k, n ∈ Z. This defines an isomorphic em-
235
+ bedding gk,k+1 : Γk → Γk+1, k ∈ Z where gk,k+1(zn,k) = zn,k+1. These
236
+ inclusions define a direct limit U of Γk, k ∈ Z. The maps gk,k+1 induce
237
+ a shift isomorphism θ : U → U, so, in particular, θ(x) = gk,k+1(x) for
238
+ all x ∈ Γk, k ∈ Z
239
+ In defining the completion U, first, let us recall the following facts
240
+ about lattices of simply connected nilpotent Lie groups [5].
241
+ Proposition 2.1. Let G be simply connected nilpotent Lie group, Γ be
242
+ a discrete subgroup of G. The following are equivalent:
243
+ (i) Γ is a lattice of G;
244
+ (ii) Γ is Zariski dense in G;
245
+ (iii) Γ is not contained in any proper connected closed subgroup of
246
+ G;
247
+ (iv) Γ is co-compact in G.
248
+ Definition 2.2. Let m ≥ 2. For any subset Ω ⊆ Um(Z), we define
249
+ Span(Ω) = ⟨Ω⟩Z where the latter denotes the Zariski closure.
250
+ For
251
+ example, Span(Um(Z)) = Um(R). Then, for any subset Ω ⊆ U∞(Z) we
252
+ let
253
+ Span(Ω) = ∪
254
+ n≥1 Span(Ω ∩ Hn).
255
+
256
+ 7
257
+ Then, for any subset Ω ⊆ U we define Span(Ω) = ∪
258
+ k≥1 Span(Ω ∩ Γk).
259
+ Finally, we define U = Span(U).
260
+ The Lie subgroups of Un(R) (hence of U) are simply connected (in-
261
+ deed contractible, as the exponential map determines a homeomor-
262
+ phism to Rd with d being the dimension of the group) thus its iso-
263
+ morphism type can be determined at the level of Lie algebras. The
264
+ Lie algebra of every Lie subgroup of U is a finite-dimensional nilpotent
265
+ Lie algebra. On the other hand, by Engel’s Theorem, for every finite-
266
+ dimensional nilpotent Lie algebra g with the underlying vector space
267
+ V , there exists an associated flag F(g) in the form {0} = V0 ≤ V1 ≤
268
+ · · · ≤ Vn = V where dimVi = i, 0 ≤ i ≤ n and for all x ∈ g, 1 ≤ i ≤ n,
269
+ ad(x)(Vi) ⊆ Vi−1.
270
+ Thus ð can be faithfully represented by strictly
271
+ upper-triangular matrices with respect to some basis of V . If g, h are
272
+ finite-dimensional nilpotent Lie algebras and φ : g → h a Lie algebra
273
+ isomorphism, then H = f(F) will be an associated flag of h. On the
274
+ other hand, if g is a finite-dimensional nilpotent Lie algebra with un-
275
+ derlying vector space V and I is an ideal of g faithfully represented
276
+ in gl(V0) with strictly upper triangular matrices with respect to a ba-
277
+ sis of a proper subspace V0, then by inductive process as in the proof
278
+ of Engels’ Theorem, it follows that we can extend the basis of V0 to
279
+ a basis of V such that g is faithfully represented with strictly upper
280
+ triangular matrices. By this observation, any Lie group isomorphism
281
+ Φ : G → H between finite-dimensional nilpotent Lie subgroups of U
282
+ can be extended to the group automorphism of U, since for any Lie
283
+ subgroups G1, G2 of U, G1 belongs to a Lie subgroup G3 which con-
284
+ tains θk(G2) as a normal subgroup for some integer k (thus the Lie
285
+ algebra g3 of G3 contains the Lie algebra of θk(G2) as an ideal.)
286
+ We can now state and prove the following
287
+ Proposition 2.3. An HNN extension of a torsion-free nilpotent group
288
+ is left-orderable.
289
+ Proof. Let Γ be a torsion-free nilpotent group. It is well-known that Γ
290
+ is left-orderable (in fact, bi-orderable). Indeed, it suffices to prove this
291
+ only for finitely generated subgroups, and any such subgroup embeds
292
+ into Um(Z) for some m ≥ 2. The latter admits an easy bi-order. Indeed,
293
+ more generally, we define a matrix A = (ai,j)1≤i,j≤n ∈ Um(R) as positive
294
+ if d is the smallest positive integer such that ai,j ̸= 0, for some i, j ≥ 1
295
+ with i + j = d, moreover, for this d, if p is the smallest positive integer
296
+ with p + q = d and ap,q ̸= 0, then ap,q > 0. One easily checks that this
297
+
298
+ 8
299
+ is in fact a genuine left-order (and even a bi-order). Then U∞(R) is
300
+ also bi-orderable as a direct limit of Um(R), m ≥ 1 and so is U.
301
+ To show that an HNN extension of Γ is also left-orderable, it again
302
+ suffices to consider HNN extensions of finitely generated subgroups. So
303
+ let us assume that Γ is also finitely generated, A, B ≤ Γ and φ : A → B
304
+ be an isomorphism. Γ embeds in U∞(Z) and the latter is a subgroup
305
+ of G = U∞(R).
306
+ The isomorphism φ : A → B cannot necessarily be extended to
307
+ G, but one can extend the isomorphism φ : Span(A) → Span(B)
308
+ to some F : U → U where φ is an extension of φ by Mostow Strong
309
+ Rigidity Theorem for lattices in solvable Lie groups [5]. Then the HNN
310
+ extension (U, Span(A), Span(B), t) with respect to the isomorphism
311
+ φ : Span(A) → Span(B) has a quotient isomorphic to the semidirect
312
+ product Z⋉F U by a normal subgroup N ≤ U. Since N and Z⋉F U are
313
+ left-orderable we obtain that (U, Span(A), Span(B), t) is left-orderable
314
+ (as an extension of a left-orderable group by a left-orderable group). By
315
+ Britton’s Lemma, (Γ, A, B, t) is a subgroup of (U, Span(A), Span(B), t)
316
+ hence it is also left-orderable.
317
+
318
+ We would like to end this section with a torsion-free non-left-orderable
319
+ example which will contain a class two nilpotent group as an index two
320
+ subgroup; indeed, it will contain a subgroup of Heisenberg group H of
321
+ 3 × 3 integral unipotent matrices. This might be one of the simplest
322
+ (smallest) examples of a torsion-free non-left-orderable group in litera-
323
+ ture. It also shows that torsion-freeness does not imply left-orderability
324
+ in the class of polycyclic groups.
325
+ Let
326
+ Γ = ⟨t, u, v | [u, v] = t4, tut−1 = u−1, tvt−1 = v−1⟩.
327
+ The group Γ is related to the Heisenberg group
328
+ H = ⟨x, y, z | [x, y] = z, [z, x] = [z, y] = 1⟩.
329
+ Any element of H can be written uniquely as xmynzk where m, n, k ∈
330
+ Z.
331
+ H is bi-orderable.
332
+ The elements x2, y, z generate an index two
333
+ subgroup H0 of H.
334
+ The group Γ will have an index two subgroup isomorphic to H0. We
335
+ let u = x2, v = y, t2 = z. Let also G be a group given by the following
336
+ presentation
337
+ G = ⟨t, x, y, z | [x, y] = z, [z, x] = [z, y] = 1, t2 = z, txt−1 = x−1, tyt−1 = y−1⟩.
338
+ Then Γ is a subgroup of G generated by t, x2, y.
339
+
340
+ 9
341
+ Proposition 2.4. Γ is torsion-free and non-left-orderable.
342
+ Proof. Assume that < is a left-order on Γ. Without loss of generality
343
+ we may assume that t > 1. Then 1 < t < z thus z is also a positive
344
+ element. On the other hand, let us observe that for all n ∈ 2Z, we have
345
+ (txn)2 = t2 thus the element txn is positive for every even integer n.
346
+ Let m be positive if y > 1 and negative if y < 1. Then, for all m ∈ 2Z,
347
+ txnym is positive as a product of two positive elements txn and ym.
348
+ Then (txnym)2 > 1. However,
349
+ (txnym)2 = t2x−ny−mxnym = t2zmn = zmn+1.
350
+ We can choose n such that mn+1 < 0. This yields that (txnym)2 < 1.
351
+ Contradiction.
352
+ To see torsion-freeness let us observe that any element g ∈ G can be
353
+ written as g = x2pyqzr or g = tx2pyqzr. The element x2pyqzr is not a
354
+ torsion since H is torsion-free. As for the element tx2pyqzr, we have
355
+ (tx2pyqzr)2 = x−2py−qx2pyqz2r+1 = z±2pqz2r+1 ̸= 1. Thus Γ is torsion-
356
+ free.
357
+
358
+ Acknowledgement:
359
+ We are very thankful to Zipei Nie for reading
360
+ the draft of this paper and correcting errors.
361
+ References
362
+ [1] J. Button, Topics in infinite groups, Lecture Notes.
363
+ [2] V.V. Bludov and A.M.W.Glass, Word problems, embeddings, and free prod-
364
+ ucts of right-ordered groups with amalgamated subgroup, Proceedings of the
365
+ London Mathemtical Society, vol. 99, issue 3, (2009), 585-608
366
+ [3] R. Lyndon and P. Schupp. Combinatorial Group Theory, Volume 89 of Ergeb-
367
+ nisse der Mathematik und ihrer Grenzgebiete, Springer-Verlaq, 1977.
368
+ [4] A. Navas, Groups of Circle Diffeomorphisms. The University of Chicago Press,
369
+ 2011.
370
+ [5] M.S.Raghunathan, Discrete Subgroups of Lie Groups, Springer Berlin Heidel-
371
+ berg, Nov 16, 1972
372
+ Azer Akhmedov, Department of Mathematics, North Dakota State
373
+ University, Fargo, ND, 58102, USA
374
+ Email address: azer.akhmedov@ndsu.edu
375
+ Cody Martin, Department of Mathematics, North Dakota State
376
+ University, Fargo, ND, 58102, USA
377
+ Email address: cody.martin@ndsu.edu
378
+
79AyT4oBgHgl3EQfQvZ4/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,317 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf,len=316
2
+ page_content='arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
3
+ page_content='00052v1 [math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
4
+ page_content='GR] 30 Dec 2022 Examples of left-orderable and non-left-orderable HNN extensions Azer Akhmedov, Cody Martin ABSTRACT: We prove that an HNN extension of a torsion-free nilpotent group is left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
5
+ page_content=' We also construct examples of non-left-orderable HNN extensions of left-orderable groups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
6
+ page_content=' 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
7
+ page_content=' Non-left-orderable HNN extensions of left-orderable groups It is well-known that an HNN extension of a torsion-free group is still torsion-free ([3], [1]).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
8
+ page_content=' On the hand, for many classes of groups, existence of a torsion element is the only obstruction to left-orderability;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
9
+ page_content=' for example, this is the case for the classes of one-relator groups, nilpotent groups, etc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
10
+ page_content=' Hence it is natural to study how left-orderability behaves under an HNN extension.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
11
+ page_content=' In [2] (see Example 6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
12
+ page_content='2 there), an example is constructed to show that left-orderability is not preserved under the HNN extension.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
13
+ page_content=' In this section, we present systematic ways of producing non-left-orderable HNN extensions of left-orderable groups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
14
+ page_content=' The example of [2] is built as an HNN extension of a direct product of a free nilpotent group of class two with the fundamental group of Klein bottle.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
15
+ page_content=' We produce examples of HNN extensions of groups such as non-Abelian free groups and virtually Abelian groups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
16
+ page_content=' We rely on the following well-known criterion about left-orderability of groups [4] Proposition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
17
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
18
+ page_content=' A group G is left-orderable if and only if for all k ≥ 1 and for all g1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
19
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
20
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
21
+ page_content=' , gk ∈ G\\{1}, there exist ǫ1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
22
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
23
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
24
+ page_content=' , ǫk ∈ {−1, 1} such that the semigroup of G generated by gǫ1 1 , .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
25
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
26
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
27
+ page_content=' , gǫk k does not contain the identity element.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
28
+ page_content=' Let us emphasize that we use the obvious “only if part” of this propo- sition;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
29
+ page_content=' the harder “if part” is not needed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
30
+ page_content=' Given a group G, and subgroups A, B ≤ G with an isomorphism φ : A → B, the HNN extension (G, A, B, t, φ) is defined as the quo- tient of the free product G ∗ ⟨t⟩ by the normal closure of the subset {tat−1φ(a)−1 | a ∈ A}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
31
+ page_content=' We also write this HNN extension as (G, A, B, t) when φ is given in the context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
32
+ page_content=' Theorem 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
33
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
34
+ page_content=' A free group of rank bigger than one admits a non-left- orderable HNN extension.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
35
+ page_content=' 1 2 Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
36
+ page_content=' By Britton’s Lemma, it suffices to prove the theorem for the group F2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
37
+ page_content=' Let a, b be the generators of F2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
38
+ page_content=' We can find positive expo- nents pi, qi, ri, si, 1 ≤ i ≤ 8 such that the elements u1 = ap1bq1, u2 = ap2bq2, u3 = ap3bq3, u4 = ap4bq4, u5 = ap5b−q5, u6 = ap6b−q6, u7 = ap7b−q7, u8 = ap8b−q8 generate a free group of rank 8, and so do the elements v1 = ar1bs1, v2 = ar2b−s2, v3 = a−r3bs3, v4 = a−r4b−s4, v5 = ar5bs5, v6 = ar6b−s6, v7 = a−r7bs7, v8 = a−r8b−s8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
39
+ page_content=' (It suffices to take the sequences (pi)1≤i≤8, (qi)1≤i≤8, (ri)1≤i≤8, (si)1≤i≤8 to be strictly increasing.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
40
+ page_content=') Let A, B be these free groups generated by u1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
41
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
42
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
43
+ page_content=' , u8 and v1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
44
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
45
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
46
+ page_content=' , v8 respectively, and φ : A → B be the isomor- phism such that φ(ui) = vi, 1 ≤ i ≤ 8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
47
+ page_content=' Then, by Proposition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
48
+ page_content='1, the HNN extension (G, A, B, t) where t(a) = φ(a) for all a ∈ A is not left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
49
+ page_content=' □ Remark 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
50
+ page_content='3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
51
+ page_content=' Let us remind that in the case of rank = 1, the claim does not hold anymore since any HNN extension of Z is isomorphic ⟨t, a | tamt−1 = an⟩ for some non-zero integers m, n.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
52
+ page_content=' All these groups (which include Z2, π1(Klein bottle) = ⟨a, b | aba−1 = b−1⟩, and the solvable Baumslag-Solitar group BS(1, n) ∼= Z ⋉ Z[ 1 n]), are all left- orderable as torsion-free one-relator groups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
53
+ page_content=' Using similar ideas, we build a non-left-orderable HNN extension of a left-orderable solvable group.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
54
+ page_content=' We again rely on the criterion of Proposition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
55
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
56
+ page_content=' Let n ≥ 2 and Γn be a group given by the presentation ⟨s, x | [sn, x] = 1, [x, sixs−i] = 1, 1 ≤ i ≤ n − 1⟩.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
57
+ page_content=' Let xi = sixs−i, i ∈ Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
58
+ page_content=' Notice that xi = xj iff i ≡ j( mod n).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
59
+ page_content=' The elements xi, 0 ≤ i ≤ n−1 generate a normal subgroup Nn isomorphic to Zn and the quotient by this subgroup is isomorphic to Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
60
+ page_content=' Any element g of Γn can be written uniquely as siw(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
61
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
62
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
63
+ page_content=' , xn−1) where i ∈ Z and w(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
64
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
65
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
66
+ page_content=' , xn−1) = xp0 0 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
67
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
68
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
69
+ page_content=' xpn−1 n−1 for some integer exponents p0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
70
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
71
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
72
+ page_content=' , pn−1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
73
+ page_content=' siw(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
74
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
75
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
76
+ page_content=' , xn−1) will be called the canonical form of g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
77
+ page_content=' We also write Σ(g) = i + p0 + · · · + pn.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
78
+ page_content=' Let us observe that Γn is torsion-free.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
79
+ page_content=' Indeed, if g is a torsion element with a canonical form siw(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
80
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
81
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
82
+ page_content=' , xn−1) as above then for all k ≥ 1, gk = sikw0(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
83
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
84
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
85
+ page_content=' , xn−1)wi(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
86
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
87
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
88
+ page_content=' , xn−1) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
89
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
90
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
91
+ page_content=' w(k−1)i(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
92
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
93
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
94
+ page_content=' , xn−1) 3 where wj(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
95
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
96
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
97
+ page_content=' , xn−1) = w(xj, xj+1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
98
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
99
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
100
+ page_content=' , xn−1+j) hence it follows im- mediately that either i = 0;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
101
+ page_content=' then, since Nn ∼= Zn, we obtain that w = 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
102
+ page_content=' It turns out Γn is left-orderable (which also implies that it is torsion- free).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
103
+ page_content=' We introduce a left order < on Γn as follows: An element g with the canonical form siw(x0, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
104
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
105
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
106
+ page_content=' , xn−1) as above will be called positive if either Σ(w) > 0 or Σ(w) = 0 and i > 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
107
+ page_content=' If Σ(w) = 0 and i = 0, then we are in the group Nn ∼= Zn and there the order can be defined lexicographically.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
108
+ page_content=' Then we see that a product of two positive elements is always positive and the inverse of a positive element is not positive.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
109
+ page_content=' Hence < is a left-order.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
110
+ page_content=' To state our next proposition we need to introduce some (well- known) terminology.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
111
+ page_content=' Definition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
112
+ page_content='4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
113
+ page_content=' Let G be a group generated by a subset S ⊆ G\\{1} such that for all x ∈ G, if x ∈ S, then x−1 /∈ S (in particular, 1 /∈ S).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
114
+ page_content=' We say that a non-trivial reduced word W(x1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
115
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
116
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
117
+ page_content=' , xk) = xn1 1 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
118
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
119
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
120
+ page_content=' xnk k is positive in the alphabet S if x1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
121
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
122
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
123
+ page_content=' , xk ∈ S and all exponents ni, 1 ≤ i ≤ k are positive.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
124
+ page_content=' Proposition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
125
+ page_content='5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
126
+ page_content=' In the group Γn let S1 = {s, x}, S2 = {s−1, x}, S3 = {s, x−1}, S4 = {s−1, x−1}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
127
+ page_content=' For n ≥ 12, there exists elements f1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
128
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
129
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
130
+ page_content=' , f4, g1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
131
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
132
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
133
+ page_content=' , g4 ∈ Γn such that the following conditions hold: i) ⟨f1, f2, f3, f4⟩ ∼= ⟨g1, g2, g3, g4⟩ ∼= Z4, ii) The elements f1, f2, f3, f4 can be represented with positive words in the alphabet S1, iii) For all 1 ≤ i ≤ 4, the element gi can be represented with a positive word in the alphabet Si.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
134
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
135
+ page_content=' We define f1 = sn−1xs, f2 = sn−2(xs)2, f3 = sn−4(xs)4, f4 = sn−8(xs)8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
136
+ page_content=' Then f1, f2, f3, f4 belong to Nn and generate a subgroup isomorphic to Z4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
137
+ page_content=' We also define g1 = sn−1xs, g2 = sn−2(x−1s)2, g3 = s4−n(xs−1)4, f4 = s8−n(x−1s−1)8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
138
+ page_content=' The elements g1, g2, g3, g4 also belong to Nn and generate a subgroup isomorphic to Z4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
139
+ page_content=' □ In the above proposition, the n ≥ 12 is not necessarily the best possi- ble.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
140
+ page_content=' Using Proposition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
141
+ page_content='5, we can now prove the following proposition which establishes the existence of a non-left-orderable HNN extension of a left-orderable virtually Abelian group.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
142
+ page_content=' 4 Proposition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
143
+ page_content='6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
144
+ page_content=' For all n ≥ 12, Γn admits a non-left-orderable HNN extension.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
145
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
146
+ page_content=' Let f1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
147
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
148
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
149
+ page_content=' , f4, g1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
150
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
151
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
152
+ page_content=' , g4 ∈ Γn be elements satisfying conditions 1)-3) of Proposition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
153
+ page_content='5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
154
+ page_content=' Let φ : ⟨f1, f2, f3, f4⟩ → ⟨g1, g2, g3, g4⟩ be an isomorphism such that φ(fi) = gi, 1 ≤ i ≤ 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
155
+ page_content=' We consider an HNN extension G := (Γn, ⟨f1, f2, f3, f4⟩, ⟨g1, g2, g3, g4⟩, t) by letting txt−1 = φ(x) for all x ∈ ⟨f1, f2, f3, f4⟩ For any left-order on G, notice that the elements tfit−1, 1 ≤ i ≤ 4 are either all positive or all negative.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
156
+ page_content=' On the other hand, among the elements gi, 1 ≤ i ≤ 4 at least one is positive and one is negative.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
157
+ page_content=' This is a contradiction.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
158
+ page_content=' Hence G is not left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
159
+ page_content=' □ 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
160
+ page_content=' HNN extensions of nilpotent groups The aim of this section is to prove that unlike solvable groups, an HNN extension of a left-orderable nilpotent group is always left- orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
161
+ page_content=' Let us recall that a nilpotent group is left-orderable iff it is torsion-free;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
162
+ page_content=' this claim too does not hold for solvable groups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
163
+ page_content=' Let us first observe that, since a direct limit of left-orderable groups is left-orderable, an HNN extension an HNN extension (G, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
164
+ page_content='B, t) is left-orderable if for all finitely generated subgroups A0 and G0 of G, where A0 ≤ A, G0 ⊇ ⟨A0, B0⟩ and B0 = tA0t−1, the HNN extension (G0, A0, B0, t) is left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
165
+ page_content=' We will use this observation repeatedly in this section.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
166
+ page_content=' We already observed that by classification, an HNN extension of an infinite cyclic group is left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
167
+ page_content=' The same holds for an HNN extension of any torsion free Abelian group.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
168
+ page_content=' Indeed, it suffices to consider finitely generated Abelian groups so let G be a finitely generated torsion-free Abelian group, A, B ≤ G, φ : A → B be an isomorphism, and (G, A, B, t) be the HNN extension with respect to the isomorphism φ.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
169
+ page_content=' Let G ∼= Zd and r = rankA = rankB.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
170
+ page_content=' We will assume that G = Zd.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
171
+ page_content=' Then for some linearly independent vectors u1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
172
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
173
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
174
+ page_content=' , ur we have A = {c1u1 + · · · + crur : ci ∈ Z, 1 ≤ i ≤ r} and similarly for some linearly independent vectors v1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
175
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
176
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
177
+ page_content=' , vr we have B = {c1v1 + · · · + crvr : ci ∈ Z, 1 ≤ i ≤ r}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
178
+ page_content=' We let G = R, A = {c1u1 + · · · + crur : ci ∈ R, 1 ≤ i ≤ r}, B = {c1v1 + · · · + crvr : ci ∈ R, 1 ≤ i ≤ r} and φ : A → B be the extension of φ : A → B defined as φ(c1u1 + · · · + crur) = c1φ(u1) + · · · + crφ(ur) for all c1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
179
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
180
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
181
+ page_content=' , cr ∈ R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
182
+ page_content=' 5 A key observation here is that even though the isomorphism φ : A → B cannot necessarily be extended to G, but one can extend the isomorphism φ : A → B to some automorphism F : G → G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
183
+ page_content=' Then the HNN extension (G, A, B, t) with respect to the isomorphism φ : A → B has a quotient isomorphic to the semidirect product Z⋉F G by a normal subgroup N ≤ G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
184
+ page_content=' Since N and Z ⋉F G are left-orderable we obtain that (G, A, B, t) is left-orderable (as an extension of a left-orderable group by a left-orderable group).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
185
+ page_content=' By Britton’s Lemma, (G, A, B, t) is a subgroup of (G, A, B, t) hence it is also left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
186
+ page_content=' We now would like to carry the same argument for any torsion-free nilpotent group.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
187
+ page_content=' The main issue here is that given a finitely generated torsion-free nilpotent group Γ, one needs to construct a completion Γ which would resemble the operation Zd → Rd so we can try to use the argument in the Abelian case.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
188
+ page_content=' Let R be a commutative ring with identity and n ≥ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
189
+ page_content=' We let Un(R) be the group of n × n upper-triangular matrices with 1’s on the diagonal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
190
+ page_content=' The cases R = R and R = Z will be the most interesting to us.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
191
+ page_content=' It is well-known that any finitely generated torsion-free nilpotent group Γ embeds in Un(Z) for some n ≥ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
192
+ page_content=' The Mal’cev completion of Un(Z) is Un(R) (and the Mal’cev completion of Zn is Rn) 1, however, given an isomorphism φ : A → B of subgroups of Un(Z), although it induces an isomorphism φ : A → B but one cannot necessarily extend this isomorphism to the entire G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
193
+ page_content=' For example, for n = 3, the group U3(Z) is isomorphic to the Heisenberg group ⟨x, y, z | z = [x, y], [x, z] = [y, z] = 1⟩ and if we let A = ⟨x⟩, B = ⟨z⟩ and φ(x) = z, then this isomorphism cannot be extended to the isomorphism of U3(Z) (or U3(R)).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
194
+ page_content=' Thus we need to define a completion of Γ other than the Mal’cev completion.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
195
+ page_content=' Let Xn,i, 1 ≤ i ≤ n − 1 be the matrix of Un(Z) where all off-diagonal entries are zero except the (i + 1, i)-th entry is equal to 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
196
+ page_content=' In order to define a more suitable completion of Un(Z) we will extend it first, and at the end we will obtain a completion which is ”infinite-dimensional”.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
197
+ page_content=' Let U∞(Z) be a group generated by xk, k ∈ Z such that for all k ∈ Z, n ≥ 1 the subgroup generated by xk+1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
198
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
199
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
200
+ page_content=' , xk+n−1 is isomorphic to Un(Z) through the isomorphism f(xk+j) = Xn,j, 1 ≤ j ≤ n − 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
201
+ page_content=' Notice that U∞(Z) is well-defined this way and it contains isomorphic copies of all 1in the literature, the term Mal’cev completion is used for some other related operations as well.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
202
+ page_content=' 6 Un(Z), n ≥ 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
203
+ page_content=' This group can be viewed as the group of infinite sized integral unipotent matrices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
204
+ page_content=' But to achieve our goal we extend U∞(Z) further as follows.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
205
+ page_content=' Let us first observe that in the group Un(Z) viewed as the group of upper triangular unipotent integral matrices, [xi, xj] = 1 if |i − j| ≥ 2 and for all 1 ≤ i ≤ n − 2, [xi, xi+1] is a unipotent matrix with all the off-diagonal entries zero, except the (i + 2, i + 1)-entry equals 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
206
+ page_content=' Thus the elements [xi, xi+1], 1 ≤ i ≤ n−2 generate a subgroup isomorphic to Un−1(Z) with an isomorphism xi → [xi, xi+1], 1 ≤ i ≤ n − 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
207
+ page_content=' Similarly, in the group U∞(Z), the elements [xi, xi+1], i ∈ Z generate a subgroup isomorphic to U∞(Z), and the homomorphism f : U∞(Z) → U∞(Z) defined as f(xi) = [xi, xi+1], i ∈ Z (it is sufficient to define it on the generators) establishes this isomorphism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
208
+ page_content=' The group U∞(Z) is a direct limit of the groups Un(Z), n ≥ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
209
+ page_content=' More precisely, let Hn, n ≥ 1 be the subgroup of U∞(Z) generated by x−n, x−n+1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
210
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
211
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
212
+ page_content=' , xn−1, xn.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
213
+ page_content=' Then Hn is isomorphic to U2n+1(Z), and U∞(Z) is a direct limit of the sequence Hn, n ≥ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
214
+ page_content=' In our construction of the completion, we will use a direct limit of groups each isomorphic to U∞(Z).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
215
+ page_content=' Let Γk, k ∈ Z be a group gen- erated by zk,n, n ∈ Z with an isomorphism gk : Γk → U∞(Z) such that gk(zk,n) = xn.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
216
+ page_content=' We have · · · ≤ Γ−1 ≤ Γ0 ≤ Γ1 ≤ Γ2 ≤ .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
217
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
218
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
219
+ page_content=' and [zk,n, zk,n+1] = zk−1,n for all k, n ∈ Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
220
+ page_content=' This defines an isomorphic em- bedding gk,k+1 : Γk → Γk+1, k ∈ Z where gk,k+1(zn,k) = zn,k+1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
221
+ page_content=' These inclusions define a direct limit U of Γk, k ∈ Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
222
+ page_content=' The maps gk,k+1 induce a shift isomorphism θ : U → U, so, in particular, θ(x) = gk,k+1(x) for all x ∈ Γk, k ∈ Z In defining the completion U, first, let us recall the following facts about lattices of simply connected nilpotent Lie groups [5].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
223
+ page_content=' Proposition 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
224
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
225
+ page_content=' Let G be simply connected nilpotent Lie group, Γ be a discrete subgroup of G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
226
+ page_content=' The following are equivalent: (i) Γ is a lattice of G;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
227
+ page_content=' (ii) Γ is Zariski dense in G;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
228
+ page_content=' (iii) Γ is not contained in any proper connected closed subgroup of G;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
229
+ page_content=' (iv) Γ is co-compact in G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
230
+ page_content=' Definition 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
231
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
232
+ page_content=' Let m ≥ 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
233
+ page_content=' For any subset Ω ⊆ Um(Z), we define Span(Ω) = ⟨Ω⟩Z where the latter denotes the Zariski closure.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
234
+ page_content=' For example, Span(Um(Z)) = Um(R).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
235
+ page_content=' Then, for any subset Ω ⊆ U∞(Z) we let Span(Ω) = ∪ n≥1 Span(Ω ∩ Hn).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
236
+ page_content=' 7 Then, for any subset Ω ⊆ U we define Span(Ω) = ∪ k≥1 Span(Ω ∩ Γk).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
237
+ page_content=' Finally, we define U = Span(U).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
238
+ page_content=' The Lie subgroups of Un(R) (hence of U) are simply connected (in- deed contractible, as the exponential map determines a homeomor- phism to Rd with d being the dimension of the group) thus its iso- morphism type can be determined at the level of Lie algebras.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
239
+ page_content=' The Lie algebra of every Lie subgroup of U is a finite-dimensional nilpotent Lie algebra.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
240
+ page_content=' On the other hand, by Engel’s Theorem, for every finite- dimensional nilpotent Lie algebra g with the underlying vector space V , there exists an associated flag F(g) in the form {0} = V0 ≤ V1 ≤ · · ≤ Vn = V where dimVi = i, 0 ≤ i ≤ n and for all x ∈ g, 1 ≤ i ≤ n, ad(x)(Vi) ⊆ Vi−1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
241
+ page_content=' Thus ð can be faithfully represented by strictly upper-triangular matrices with respect to some basis of V .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
242
+ page_content=' If g, h are finite-dimensional nilpotent Lie algebras and φ : g → h a Lie algebra isomorphism, then H = f(F) will be an associated flag of h.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
243
+ page_content=' On the other hand, if g is a finite-dimensional nilpotent Lie algebra with un- derlying vector space V and I is an ideal of g faithfully represented in gl(V0) with strictly upper triangular matrices with respect to a ba- sis of a proper subspace V0, then by inductive process as in the proof of Engels’ Theorem, it follows that we can extend the basis of V0 to a basis of V such that g is faithfully represented with strictly upper triangular matrices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
244
+ page_content=' By this observation, any Lie group isomorphism Φ : G → H between finite-dimensional nilpotent Lie subgroups of U can be extended to the group automorphism of U, since for any Lie subgroups G1, G2 of U, G1 belongs to a Lie subgroup G3 which con- tains θk(G2) as a normal subgroup for some integer k (thus the Lie algebra g3 of G3 contains the Lie algebra of θk(G2) as an ideal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
245
+ page_content=') We can now state and prove the following Proposition 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
246
+ page_content='3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
247
+ page_content=' An HNN extension of a torsion-free nilpotent group is left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
248
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
249
+ page_content=' Let Γ be a torsion-free nilpotent group.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
250
+ page_content=' It is well-known that Γ is left-orderable (in fact, bi-orderable).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
251
+ page_content=' Indeed, it suffices to prove this only for finitely generated subgroups, and any such subgroup embeds into Um(Z) for some m ≥ 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
252
+ page_content=' The latter admits an easy bi-order.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
253
+ page_content=' Indeed, more generally, we define a matrix A = (ai,j)1≤i,j≤n ∈ Um(R) as positive if d is the smallest positive integer such that ai,j ̸= 0, for some i, j ≥ 1 with i + j = d, moreover, for this d, if p is the smallest positive integer with p + q = d and ap,q ̸= 0, then ap,q > 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
254
+ page_content=' One easily checks that this 8 is in fact a genuine left-order (and even a bi-order).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
255
+ page_content=' Then U∞(R) is also bi-orderable as a direct limit of Um(R), m ≥ 1 and so is U.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
256
+ page_content=' To show that an HNN extension of Γ is also left-orderable, it again suffices to consider HNN extensions of finitely generated subgroups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
257
+ page_content=' So let us assume that Γ is also finitely generated, A, B ≤ Γ and φ : A → B be an isomorphism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
258
+ page_content=' Γ embeds in U∞(Z) and the latter is a subgroup of G = U∞(R).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
259
+ page_content=' The isomorphism φ : A → B cannot necessarily be extended to G, but one can extend the isomorphism φ : Span(A) → Span(B) to some F : U → U where φ is an extension of φ by Mostow Strong Rigidity Theorem for lattices in solvable Lie groups [5].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
260
+ page_content=' Then the HNN extension (U, Span(A), Span(B), t) with respect to the isomorphism φ : Span(A) → Span(B) has a quotient isomorphic to the semidirect product Z⋉F U by a normal subgroup N ≤ U.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
261
+ page_content=' Since N and Z⋉F U are left-orderable we obtain that (U, Span(A), Span(B), t) is left-orderable (as an extension of a left-orderable group by a left-orderable group).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
262
+ page_content=' By Britton’s Lemma, (Γ, A, B, t) is a subgroup of (U, Span(A), Span(B), t) hence it is also left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
263
+ page_content=' □ We would like to end this section with a torsion-free non-left-orderable example which will contain a class two nilpotent group as an index two subgroup;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
264
+ page_content=' indeed, it will contain a subgroup of Heisenberg group H of 3 × 3 integral unipotent matrices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
265
+ page_content=' This might be one of the simplest (smallest) examples of a torsion-free non-left-orderable group in litera- ture.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
266
+ page_content=' It also shows that torsion-freeness does not imply left-orderability in the class of polycyclic groups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
267
+ page_content=' Let Γ = ⟨t, u, v | [u, v] = t4, tut−1 = u−1, tvt−1 = v−1⟩.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
268
+ page_content=' The group Γ is related to the Heisenberg group H = ⟨x, y, z | [x, y] = z, [z, x] = [z, y] = 1⟩.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
269
+ page_content=' Any element of H can be written uniquely as xmynzk where m, n, k ∈ Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
270
+ page_content=' H is bi-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
271
+ page_content=' The elements x2, y, z generate an index two subgroup H0 of H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
272
+ page_content=' The group Γ will have an index two subgroup isomorphic to H0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
273
+ page_content=' We let u = x2, v = y, t2 = z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
274
+ page_content=' Let also G be a group given by the following presentation G = ⟨t, x, y, z | [x, y] = z, [z, x] = [z, y] = 1, t2 = z, txt−1 = x−1, tyt−1 = y−1⟩.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
275
+ page_content=' Then Γ is a subgroup of G generated by t, x2, y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
276
+ page_content=' 9 Proposition 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
277
+ page_content='4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
278
+ page_content=' Γ is torsion-free and non-left-orderable.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
279
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
280
+ page_content=' Assume that < is a left-order on Γ.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
281
+ page_content=' Without loss of generality we may assume that t > 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
282
+ page_content=' Then 1 < t < z thus z is also a positive element.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
283
+ page_content=' On the other hand, let us observe that for all n ∈ 2Z, we have (txn)2 = t2 thus the element txn is positive for every even integer n.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
284
+ page_content=' Let m be positive if y > 1 and negative if y < 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
285
+ page_content=' Then, for all m ∈ 2Z, txnym is positive as a product of two positive elements txn and ym.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
286
+ page_content=' Then (txnym)2 > 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
287
+ page_content=' However, (txnym)2 = t2x−ny−mxnym = t2zmn = zmn+1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
288
+ page_content=' We can choose n such that mn+1 < 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
289
+ page_content=' This yields that (txnym)2 < 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
290
+ page_content=' Contradiction.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
291
+ page_content=' To see torsion-freeness let us observe that any element g ∈ G can be written as g = x2pyqzr or g = tx2pyqzr.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
292
+ page_content=' The element x2pyqzr is not a torsion since H is torsion-free.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
293
+ page_content=' As for the element tx2pyqzr, we have (tx2pyqzr)2 = x−2py−qx2pyqz2r+1 = z±2pqz2r+1 ̸= 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
294
+ page_content=' Thus Γ is torsion- free.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
295
+ page_content=' □ Acknowledgement: We are very thankful to Zipei Nie for reading the draft of this paper and correcting errors.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
296
+ page_content=' References [1] J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
297
+ page_content=' Button, Topics in infinite groups, Lecture Notes.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
298
+ page_content=' [2] V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
299
+ page_content='V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
300
+ page_content=' Bludov and A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
301
+ page_content='M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
302
+ page_content='W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
303
+ page_content='Glass, Word problems, embeddings, and free prod- ucts of right-ordered groups with amalgamated subgroup, Proceedings of the London Mathemtical Society, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
304
+ page_content=' 99, issue 3, (2009), 585-608 [3] R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
305
+ page_content=' Lyndon and P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
306
+ page_content=' Schupp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
307
+ page_content=' Combinatorial Group Theory, Volume 89 of Ergeb- nisse der Mathematik und ihrer Grenzgebiete, Springer-Verlaq, 1977.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
308
+ page_content=' [4] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
309
+ page_content=' Navas, Groups of Circle Diffeomorphisms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
310
+ page_content=' The University of Chicago Press, 2011.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
311
+ page_content=' [5] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
312
+ page_content='S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
313
+ page_content='Raghunathan, Discrete Subgroups of Lie Groups, Springer Berlin Heidel- berg, Nov 16, 1972 Azer Akhmedov, Department of Mathematics, North Dakota State University, Fargo, ND, 58102, USA Email address: azer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
314
+ page_content='akhmedov@ndsu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
315
+ page_content='edu Cody Martin, Department of Mathematics, North Dakota State University, Fargo, ND, 58102, USA Email address: cody.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
316
+ page_content='martin@ndsu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
317
+ page_content='edu' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/79AyT4oBgHgl3EQfQvZ4/content/2301.00052v1.pdf'}
89FLT4oBgHgl3EQfti_6/content/tmp_files/2301.12152v1.pdf.txt ADDED
@@ -0,0 +1,1524 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Layout-aware Webpage Quality Assessment
2
+ Anfeng Cheng∗, Yiding Liu∗, Weibin Li∗, Qian Dong, Shuaiqiang Wang, Zhengjie Huang, Shikun
3
+ Feng, Zhicong Cheng and Dawei Yin§
4
+ Baidu Inc., Beijing, China
5
+ {chenganfeng01,liweibin02,wangshuaiqiang,huangzhengjie,fengshikun01,chengzhicong01}@baidu.com
6
+ liuyiding.tanh@gmail.com,dq22@mails.tsinghua.edu.cn,yindawei@acm.org
7
+ ABSTRACT
8
+ Identifying high-quality webpages is fundamental for real-world
9
+ search engines, which can fulfil users’ information need with the
10
+ less cognitive burden. Early studies of webpage quality assessment
11
+ usually design hand-crafted features that may only work on par-
12
+ ticular categories of webpages (e.g., shopping websites, medical
13
+ websites). They can hardly be applied to real-world search engines
14
+ that serve trillions of webpages with various types and purposes.
15
+ In this paper, we propose a novel layout-aware webpage quality
16
+ assessment model currently deployed in our search engine. Intu-
17
+ itively, layout is a universal and critical dimension for the quality
18
+ assessment of different categories of webpages. Based on this, we
19
+ directly employ the meta-data that describes a webpage, i.e., Doc-
20
+ ument Object Model (DOM) tree, as the input of our model. The
21
+ DOM tree data unifies the representation of webpages with different
22
+ categories and purposes and indicates the layout of webpages. To
23
+ assess webpage quality from complex DOM tree data, we propose
24
+ a graph neural network (GNN) based method that extracts rich
25
+ layout-aware information that implies webpage quality in an end-
26
+ to-end manner. Moreover, we improve the GNN method with an
27
+ attentive readout function, external web categories and a category-
28
+ aware sampling method. We conduct rigorous offline and online
29
+ experiments to show that our proposed solution is effective in real
30
+ search engines, improving the overall usability and user experience.
31
+ KEYWORDS
32
+ Webpage Quality Models, Graph Neural Network, Information Re-
33
+ trieval, Search
34
+ ACM Reference Format:
35
+ Anfeng Cheng∗, Yiding Liu∗, Weibin Li∗, Qian Dong, Shuaiqiang Wang,
36
+ Zhengjie Huang, Shikun Feng, Zhicong Cheng and Dawei Yin§. 2023. Layout-
37
+ aware Webpage Quality Assessment. In SIGKDD ’23: ACM Special Interest
38
+ Group on Knowledge Discovery and Data Mining, August 06-10, 2023, Long
39
+ Beach, CA. ACM, New York, NY, USA, 11 pages. https://doi.org/XXXXXXX.
40
+ XXXXXXX
41
+ ∗ Co-first authors.
42
+ § Dawei Yin is the corresponding author.
43
+ Permission to make digital or hard copies of all or part of this work for personal or
44
+ classroom use is granted without fee provided that copies are not made or distributed
45
+ for profit or commercial advantage and that copies bear this notice and the full citation
46
+ on the first page. Copyrights for components of this work owned by others than ACM
47
+ must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
48
+ to post on servers or to redistribute to lists, requires prior specific permission and/or a
49
+ fee. Request permissions from permissions@acm.org.
50
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
51
+ © 2023 Association for Computing Machinery.
52
+ ACM ISBN 978-1-4503-XXXX-X/18/06...$15.00
53
+ https://doi.org/XXXXXXX.XXXXXXX
54
+ 1
55
+ INTRODUCTION
56
+ Search engines, such as Google and Baidu, plays an important
57
+ role in fulfilling users’ information need. Over the past decades,
58
+ relevance modeling is the main concern of search engines, dedicated
59
+ to putting the most relevant web content on top of the ranked
60
+ results [23, 28, 38, 42]. However, the very fact that not all relevant
61
+ contents are useful to users has become an increasingly serious
62
+ symptom, where relevant webpages with low quality would induce
63
+ a significant cognitive burden on the user. In such cases, useful
64
+ information is hard to be identified, and the users need to take extra
65
+ effort to understand the information to be conveyed. To reduce the
66
+ cognitive burden, measuring the quality of webpages has become a
67
+ critical concern, which can better benefit users with well-delivered
68
+ information and improve the overall usability of a search engine. For
69
+ example, given webpages with comparable relevance, high-quality
70
+ webpages should be ranked higher than its competitors.
71
+ Nevertheless, webpage quality assessment is a very important
72
+ yet challenging task in web search, due to the complexity and
73
+ diversity of webpages in the era of web 2.0. A webpage with a clear
74
+ structure, tidy organization and concentrated delivery of crucial
75
+ information is always preferable to one that only stacks content
76
+ without proper presentation, even though each may contain similar
77
+ information. Therefore, an accurate assessment of webpage quality
78
+ can facilitate a search engine to reduce the cognitive burden and
79
+ more effectively provide useful information for users.
80
+ Previous attempts at webpage quality assessment mainly aim
81
+ to manually design discriminative features [5, 14, 16, 24], where
82
+ classification algorithms [7, 14] are applied subsequently. How-
83
+ ever, modern search engines usually face trillions of webpages
84
+ with various categories, where simple hand-crafted features and
85
+ classification algorithms (e.g. Bayesian Networks [7]) can hardly
86
+ capture the in-depth information that reveals the webpage quality.
87
+ Moreover, most of them can only work on a particular category of
88
+ webpages, e.g., shopping websites [8], medical website [25], web
89
+ portals [6], and Wikipedia articles [17]. They are hard to be effec-
90
+ tively applied to real-world search engines that serve trillions of
91
+ heterogeneous webpages.
92
+ To address the aforementioned limitations, we conduct the first
93
+ work that investigates layout-aware webpage quality assessment on
94
+ real-world web data. The intuition is based on the findings that the
95
+ quality of a webpage is largely determined by its content layout [6,
96
+ 24], which is of a great influence on how users perceive textual and
97
+ multi-modal content [20, 31, 34, 35, 41]. Modeling in-depth layout
98
+ information is promising for webpage quality assessment in real-
99
+ world web search scenario. However, it is also very challenging,
100
+ where two crucial research questions need to be answered:
101
+ RQ1: How to capture the layout information of different categories
102
+ of webpages in a unified manner?
103
+ arXiv:2301.12152v1 [cs.IR] 28 Jan 2023
104
+
105
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
106
+ Cheng and Liu, et al.
107
+ <html>
108
+ <head>
109
+ <body>
110
+ <meta> <link>
111
+ <title>
112
+ <div>
113
+ <p>
114
+ <img>
115
+ <html>
116
+ <head>
117
+ <meta name="viewport" content="width=device-
118
+ width,initial-scale=1">
119
+ <link href=" style.css" rel=" stylesheet”>
120
+ <title>Critical Path</title>
121
+ </head>
122
+ <body>
123
+ <p>Hello <span>web performance</span>
124
+ students!</p>
125
+ <div><img src=" awesome-photo.jpg"></div>
126
+ </body>
127
+ </html>
128
+ Figure 1: A toy example that shows the webpage layout represented
129
+ by the DOM tree.
130
+ RQ2: How to encode webpage layout information for webpage quality
131
+ assessment on large-scale heterogeneous webpages?
132
+ To answer RQ1, we propose to extract layout information of
133
+ webpages from Document Object Model (DOM) data. Specifically,
134
+ DOM is a cross-platform and language-independent interface that
135
+ treats a webpage as a tree structure wherein each node is an object
136
+ representing a content piece of the webpage.
137
+ Figure 1 shows a toy example of the hierarchical structure of a
138
+ DOM tree, which is converted from its HTML source code. Each
139
+ node in the tree is an object that contains partial content of the
140
+ webpage and is associated with different attributes that describe the
141
+ content (e.g., type and size). Such data indicates rich hierarchical in-
142
+ formation on the content and its layout. And different categories of
143
+ webpages can be represented in a unified manner. It is indisputable
144
+ that inspecting the structure of DOM tree can help to measure the
145
+ quality.
146
+ For RQ2, it is very challenging as webpages in real search en-
147
+ gines are highly diverse, where the modeling of layout information
148
+ should be expressive to reveal the underlying patterns of hetero-
149
+ geneous DOM tree data. Recent advances in deep representation
150
+ learning [3, 22] have achieved great success on many web applica-
151
+ tions [9, 13, 28], and also sheds new light on our task at hand.
152
+ Notably, Graph Neural Networks [12, 15] has shown great per-
153
+ formance in modeling structured text (e.g., word interactions) [18,
154
+ 37, 41], yet they are unexplored for complex DOM tree structure.
155
+ Different from structure of textual document, the webpage layout
156
+ represented by DOM tree is more complicated, which usually has
157
+ hierarchical structure and the nodes usually have rich attributes.
158
+ Existing methods that designed for text structure usually lack spe-
159
+ cialized consideration for the problem of quality assessment on
160
+ DOM tree data, and thus might be unsatisfactory for real search
161
+ engines. To this end, we propose the first GNN-based method to
162
+ learn the underlying semantics of webpage layout in an end-to-end
163
+ manner, based on which we further make several improvements to
164
+ advance its performance on the task of webpage quality assessment.
165
+ To verify the effectiveness of our layout-aware webpage quality
166
+ assessment model, we perform offline experiments on the dataset
167
+ collected by the real-world search engine. Additionally, we deployed
168
+ our model in the online ranking system and achieve good improve-
169
+ ments. Last but not least, the proposed solution is currently fully
170
+ deployed in the online system of Baidu Search. To illustrate how
171
+ layout-aware webpage quality assessment facilitates the overall
172
+ usability of our search engine, we further present the details of the
173
+ model deployment.
174
+ Overall, our main contributions can be summarized as follows.
175
+ • We develop the largest application of deep learning for the
176
+ problem of webpage quality assessment, which significantly
177
+ improves the overall usability of real-world search engines.
178
+ • We leverage DOM tree data and propose a GNN-based so-
179
+ lution to learn the quality information of heterogeneous
180
+ webpages in an end-to-end fashion.
181
+ • We present the deployment of webpage quality assessment
182
+ model in the real production environment, which effectively
183
+ serves trillions of webpages with various categories and
184
+ purposes.
185
+ • We conduct rigorous offline and online experiments before
186
+ fully deploying the model online. The experimental results
187
+ show that the proposed solution is effective to be applied in
188
+ real-world search engines.
189
+ 2
190
+ RELATED WORK
191
+ 2.1
192
+ Graph Neural Network
193
+ Recent years have witnessed the success of graph neural networks
194
+ (GNNs) for relational data. For example, Graph Convolutional Net-
195
+ work (GCN) [21] is introduced to aggregate the one-hop neigh-
196
+ bours of each node in the graph, followed by a linear projection and
197
+ non-linear activation. GraphSAGE [15] is proposed to generalize
198
+ GCN’s aggregation operation from average to sum, max and a RNN
199
+ unit. Graph Attention Network (GAT) [30] employs the attention
200
+ mechanism into GNNs, which allows GAT to assign different im-
201
+ portance to nodes within the same neighbourhood. Generally, a
202
+ GNN can be regarded as using the input graph structure as the com-
203
+ putation graph for message passing [12], during which the local
204
+ neighbourhood information is aggregated to get a more contextual
205
+ representation. For more details, please refer to [32].
206
+ Moreover, there are many applications across various domains
207
+ that apply GNNs and achieve considerable improvements, such
208
+ as protein model quality assessment [2, 26], fuel ignition quality
209
+ assessment [27], advertising detection [36] and text classification
210
+ [11, 19, 37, 41]. For example, Sanyal et al. [26] explore an alterna-
211
+ tive approach and train a graph convolutional network with nodes
212
+ representing protein atoms and edges connecting spatially adja-
213
+ cent atom pairs. GraphQA [2] is a graph-based method to estimate
214
+ the quality of protein models, that possesses favorable properties
215
+ such as representation learning. Schweidtmann et al. [27] develop
216
+ GNN models for predicting three fuel ignition quality indicators
217
+ of oxygenated and non-oxygenated hydrocarbons. Yang et al. [36]
218
+ propose WTAGRAPH, a web tracking and advertising detection
219
+ framework based on graph neural networks.
220
+ Notably, a handful of researches [18, 37, 41] leverage GNN to
221
+ perceive text structure (e.g., word interactions) for down-stream
222
+ tasks, which are the closest research to our study. However, they
223
+ mainly consider the relationships between segments of text (e.g.,
224
+ words and paragraphs), and do not consider the overall structure
225
+ and layout of webpage, i.e., how the multi-modal content is orga-
226
+ nized and presented. In addition, the expressiveness of GNN is not
227
+ explored for the task of webpage quality assessment. In this paper,
228
+ we develop the first GNN-based method for webpage quality assess-
229
+ ment, which is further deployed in the real production environment
230
+ that facilitate the usability of our search engine.
231
+
232
+ Layout-aware Webpage Quality Assessment
233
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
234
+ 2.2
235
+ Layout
236
+ Layout, i.e., how the contents are organized and presented, is a
237
+ critical dimension for document generation [4], scene recognition
238
+ [1, 10] and webpage quality assessment [16, 24]. Substantial efforts
239
+ have been made to explore the importance of layout in many AI
240
+ areas. For example, Besides, Biswas et al. [4] design an automated
241
+ deep generative model using graph neural networks to generate
242
+ synthetic data with highly variable and plausible document layouts.
243
+ Avetisyan et al. [1] use a message-passing graph neural network to
244
+ model the inter-relationships between objects and layout, guiding
245
+ the generation of a global object alignment in a scene. Chen et
246
+ al. [10] build a Layout Graph Network (LGN) where regions in
247
+ PaSL are defined as nodes and two kinds of independent relations
248
+ between regions are encoded as edges. Zhang et al. [40] propose a
249
+ new kind of classification method for lithography layout patterns
250
+ based on graph convolution network.
251
+ Recently, with the development of pre-trained language models,
252
+ LayoutLM-style methods have achieved success in textual or multi-
253
+ model document understanding. For instance, LAMPreT [31] en-
254
+ codes each block with a multimodal transformer in the lower-level,
255
+ and aggregates the block-level representations and connections
256
+ utilizing a specifically designed transformer at the higher-level. Lay-
257
+ outLM [34] jointly model interactions between text and layout in-
258
+ formation across scanned document images, which is beneficial for
259
+ a great number of real-world document image understanding tasks
260
+ such as information extraction from scanned documents. LayoutLM-
261
+ v2 [35] further uses the new text-image alignment and text-image
262
+ matching tasks and integrates a spatial-aware self-attention mech-
263
+ anism into the Transformer architecture. LayoutLM-v3 [20] pre-
264
+ trains multi-modal Transformers for Document AI with unified text
265
+ and image masking.
266
+ Our work differs from the aforementioned studies, as our main
267
+ focus is to model the layout of webpages for quality assessment.
268
+ In particular, we model the layout of webpage with a graph con-
269
+ struction method that represents a DOM tree as a layout graph and
270
+ employ an expressive graph neural network to capture the underly-
271
+ ing semantics of the layout graphs for webpage quality assessment.
272
+ 3
273
+ PRELIMINARIES
274
+ In this section, we introduce the basic concepts and formalize the
275
+ problem of webpage quality assessment. We summarize the com-
276
+ monly used notations in Table 1.
277
+ 3.1
278
+ Layout-aware Webpage Quality
279
+ Intuitively, high-quality webpages are those that clearly provide
280
+ useful information for users in common. Specifically, given a set
281
+ of webpages with comparable relevance under the same query, we
282
+ consider the layout (i.e. structure design, content presentation) as
283
+ the key dimension of measuring webpage quality and improving
284
+ user experience [6, 24]. Based on this, we can construct a set of rules
285
+ and principles for annotating webpage quality and utilize human
286
+ annotation as the objective of our proposed method.
287
+ The considering aspects of rules and principles to score the
288
+ layout of a webpage are shown in table 2, including interactive
289
+ Experience, paragraph and layout design. We give the definition
290
+ Table 1: Commonly-used notations.
291
+ Notations
292
+ Descriptions
293
+ G𝑝 = {N, E}
294
+ A layout graph of webpage 𝑝
295
+ N
296
+ The node set of graph G𝑝
297
+ E
298
+ The edge set of graph G𝑝
299
+ F = {F𝑡 }𝑇
300
+ 𝑡=1
301
+ The layout-related feature sets
302
+ 𝑇
303
+ The number of node type
304
+ F𝑡 = {f𝑖}|F𝑡 |
305
+ 𝑖=1
306
+ The feature set of node type 𝑡
307
+ f𝑖
308
+ The layout-related feature
309
+ E(·)
310
+ The embedding of its input
311
+ �ℎ(0)
312
+ 𝑛
313
+ The initialized embedding of node 𝑛
314
+ 𝑡𝑛
315
+ The node type of node 𝑛
316
+ 𝜷𝑝
317
+ The category of webpage 𝑝
318
+ �ℎ(0)
319
+ 𝑣
320
+ The initialized embedding of virtual node 𝑛
321
+ 𝜎(·)
322
+ An activation function
323
+ 𝛼𝑛𝑚
324
+ The attention score between nodes 𝑚 and 𝑛
325
+ 𝑒𝑛𝑗
326
+ The attention coefficient between nodes 𝑚 and 𝑗
327
+ 𝑠𝑝
328
+ The predicted assessment score of webpage 𝑝
329
+ 𝑦𝑝
330
+ The manually assessment score of webpage 𝑝
331
+ and some examples for each aspect. The rules and principles for
332
+ annotators are defined as the following:
333
+ • 0 means poor layout. On the basis of ordinary pages, points
334
+ will be deducted for various flaws.
335
+ • 1 means ordinary layout. 1 point is common, and annotators
336
+ are required to add or deduct on this basis.
337
+ • 1.5 means better structure. a certain gain compared to ordi-
338
+ nary layout.
339
+ • 2 means gainful layout. The user experience of this layout is
340
+ significantly better than most layouts.
341
+ Based on these principle, bonus and deduction rules can be for-
342
+ mulated as: webpages with reasonable & beautiful layout or rich
343
+ information will have an extra bonus, on the contrary, unreason-
344
+ able & chaotic layout or valueless information will be deducted.
345
+ Finally, annotators are required to score the give webpage from 0
346
+ to 2 points based on the above rules and principles.
347
+ 3.2
348
+ Layout Graph
349
+ To extract quality information from webpage layout, we construct a
350
+ layout graph for each webpage based on its DOM tree. In particular,
351
+ a layout graph is denoted as G𝑝 = {N, E} that contains a node
352
+ set N and an edge set E. Each node has a specific type (e.g., text,
353
+ image and video), and is associated with several layout-related
354
+ features F𝑡 = {f𝑖}|F𝑡 |
355
+ 𝑖=1 . The features of different types of nodes are
356
+ denoted as F = {F𝑡 }𝑇
357
+ 𝑡=1, where𝑇 is the total number of node types.
358
+ Besides, each layout graph is also associated with the category of
359
+
360
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
361
+ Cheng and Liu, et al.
362
+ Table 2: The considering aspects of rules and principles to score the layout of a webpage.
363
+ Aspects
364
+ Definition
365
+ Examples
366
+ Interactive Experience
367
+ Whether the webpage has interactive function
368
+ Click to call, swipe to browse pictures
369
+ Paragraph
370
+ Ways to split the document into paragraphs
371
+ Using different heading, special font color to layering
372
+ Layout Design
373
+ The overall design of the webpage’s layout
374
+ many additional functions, various modules, font section size is appropriate
375
+ its webpage, which is denoted as 𝜷𝑝. The detailed construction
376
+ process of layout graph is depicted in Section 4.2.
377
+ 3.3
378
+ Webpage Quality Assessment
379
+ Given a layout graph G𝑝 = {N, E}, its features F = {F𝑡 }𝑇
380
+ 𝑡=1 cate-
381
+ gory 𝜷 and F𝑡 = {f𝑖}|F𝑡 |
382
+ 𝑖=1 , the task of webpage quality assessment is
383
+ to estimate a score 𝑠𝑝 for a given webpage 𝑝 w.r.t. its quality, i.e,
384
+ 𝑠𝑝 = 𝑓𝜃 (G𝑝, F𝑡, 𝜷𝑝),
385
+ (1)
386
+ where 𝑓 (·) represents the quality model, and 𝜃 denotes its param-
387
+ eters. The scores should be consistent with users’ perception of
388
+ webpage quality, and reflect the rules and principles as we described
389
+ above.
390
+ 4
391
+ METHOD
392
+ In this section, we first present the overview of our model. Then,
393
+ we describe the graph formulation process for a webpage, including
394
+ the construction of the layout graph and feature pre-processing.
395
+ After that, we present a GNN-based solution for webpage quality
396
+ assessment.
397
+ 4.1
398
+ Overview
399
+ Our solution mainly contains two components: layout graph for-
400
+ mulation (i.e., Section 4.2) and quality assessment model (i.e., Sec-
401
+ tion 4.3). In layout graph formulation, we first leverage the layout
402
+ information encoded in DOM tree to construct a layout graph
403
+ G𝑝 = {N, E} for every webpage 𝑝. Then, two types of features are
404
+ designed for the quality assessment, as depicted in Figure 2, includ-
405
+ ing those associated with each node in the graph, as well as the
406
+ category of the corresponding webpage (i.e., 𝜷𝑝).
407
+ Next, we propose a quality assessment model that leverages
408
+ Graph Attention Network (GAT) to perform expressive message
409
+ passing between nodes in the layout graph. Both local and global
410
+ structure information of the layout graph can be encoded in latent
411
+ representations, which are exploited for the quality assessment task.
412
+ Moreover, we improve the vanilla GAT model by 1) introducing an
413
+ attentive readout function via the virtual node, 2) incorporating
414
+ graph-level category information in the scoring function, and 3)
415
+ alleviating the data imbalance problem that is common in real-world
416
+ applications.
417
+ 4.2
418
+ Layout Graph Formulation
419
+ Graph construction. The content layout has been viewed as
420
+ one of the most critical dimensions for measuring webpage qual-
421
+ ity [24]. To formulate layout information for various categories of
422
+ webpages, we first construct layout graph based on DOM tree. In
423
+ Table 3: The summary of selected features for each node type in our
424
+ constructed graph.
425
+ Classification
426
+ Feature Name
427
+ Location
428
+ height, width, xpos, ypos, position type
429
+ Content
430
+ number of word, font size, font style,
431
+ line height, font weight, alignment
432
+ Layout
433
+ border, padding, margin, visibility,
434
+ display style, outline style, outline width
435
+ Others
436
+ tag name, webpage category
437
+ particular, we leverage HTML parser Beautiful Soup 1 to parse the
438
+ source code of a webpage, identifying the hierarchical structure
439
+ of the webpage. Then, Depth First Search (DFS) is used for exact-
440
+ ing adjacency relationships from the DOM tree. Specifically, we
441
+ recursively record the nodes and the corresponding edges between
442
+ parent and child nodes in the DOM tree, as shown in Algorithm 1.
443
+ The layout graph G𝑝 of webpage 𝑝 can be expressed by the exacted
444
+ nodes N and their relations E.
445
+ Virtual node. It is worth noting that, we also include a global
446
+ virtual node that connects to all the other nodes in the graph (as
447
+ shown in Figure 2). It can be viewed as a super-hub [39] of the layout
448
+ graph, which could be useful to aggregate the global information,
449
+ and serves as hyperlinks that connect any two nodes in the layout
450
+ graph. As such, we can capture global information of the given
451
+ graph via the virtual node.
452
+ Feature pre-processing. To capture the layout information of
453
+ the webpage, we design a series of features for each node type.
454
+ Taking the text node as an example, font style, font size, alignment
455
+ and position in webpage are all represented by learnable embedding.
456
+ The detailed list of features is presented in Table 3.
457
+ More specifically, for continuous features (e.g., height, line height
458
+ and margin), a non-uniform interval division strategy is employed
459
+ to divide the continuous interval into several buckets, which can
460
+ ensure that there are enough training samples in a single bucket.
461
+ The uniform division of the whole interval leads to the data sparse
462
+ issue since the continuous features typically obey a long-tail dis-
463
+ tribution. Discrete features (e.g., font style, display style and tag
464
+ name), are falling on a divided interval are mapped into a corre-
465
+ sponding bucket, and this bucket is assigned a learnable embedding
466
+ to represent the characteristics of its interval.
467
+ 1We parse webpages with the python library: https://www.crummy.com/software/
468
+ BeautifulSoup/bs4/doc/
469
+
470
+ Layout-aware Webpage Quality Assessment
471
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
472
+ Layout-Related Features
473
+ ...
474
+ Virtual Node
475
+ Layout Graph
476
+ Node Features
477
+ Graph Features
478
+ Projection
479
+ Page Score
480
+ Figure 2: The illustration of message passing in our model. The red node represents the virtual node in the constructed layout graph, which
481
+ is utilized for capture the graph-level information.
482
+ Algorithm 1: Layout Graph Construction
483
+ Input: HTML DOM tree R𝑝 of webpage 𝑝
484
+ Output: layout graph G𝑝 = {N, E} of webpage 𝑝
485
+ % recursive graph construction;
486
+ GraphConstruction(root) begin
487
+ N𝑟 = {𝑣𝑖𝑟𝑡𝑢𝑎𝑙_𝑛𝑜𝑑𝑒, root.𝑛𝑜𝑑𝑒};
488
+ E𝑟 = {(𝑣𝑖𝑟𝑡𝑢𝑎𝑙_𝑛𝑜𝑑𝑒, root.𝑛𝑜𝑑𝑒)};
489
+ for child ∈ root.𝑐ℎ𝑖𝑙𝑑𝑟𝑒𝑛 do
490
+ N𝑐 = child.𝑛𝑜𝑑𝑒;
491
+ N𝑟 = N𝑟 ∪ N𝑐;
492
+ E𝑐 = {(𝑣𝑖𝑟𝑡𝑢𝑎𝑙_𝑛𝑜𝑑𝑒, child.𝑛𝑜𝑑𝑒),
493
+ (child.𝑛𝑜𝑑𝑒, root.𝑛𝑜𝑑𝑒)};
494
+ E𝑟 = E𝑟 ∪ E𝑐;
495
+ N𝑐, E𝑐 = GraphConstruction(child);
496
+ end
497
+ return N𝑟, E𝑟
498
+ end
499
+ G𝑝 = {GraphConstruction(Rp)};
500
+ In addition to the node-level features, graph-level feature em-
501
+ bedding is introduced to the layout graph (i.e., webpage category)
502
+ to provide the model with the ability to perceive different cate-
503
+ gories of webpages, which is vital to the quality assessment. One
504
+ reason is that the same webpage category has a similar structure.
505
+ With the development of webpage makers (like Dreamweaver, and
506
+ Google Web Designer), large amounts of webpages are generated
507
+ from templates and almost in the same layout. Therefore, with this
508
+ graph-level embedding, the predicted assessment score shall be
509
+ more robust in the online search engine. Another reason lies in
510
+ that different webpage categories have different criteria for quality
511
+ assessment. For example, a succinct and well-organized document
512
+ layout without distracting pictures is preferred on a search page,
513
+ but for a portal, a document layout with pictures and text is con-
514
+ sidered to be better. In summary, it is meaningful and important to
515
+ take the graph-level feature embedding into account for the layout
516
+ graph.
517
+ 4.3
518
+ Quality Assessment Model
519
+ Given the constructed layout graph associated with rich features,
520
+ the key of webpage quality assessment is to expressively reveal
521
+ salient patterns underlying the graph. In particular, we consider
522
+ two types of relationships in the graph that could be discriminative
523
+ for the task:
524
+ • Local relationships. Intuitively, the relationships between
525
+ adjacent nodes in the layout graph are important to reveal
526
+ content quality. For example, a node with <image> tag is usu-
527
+ ally the illustration of its adjacent (e.g., parent) node with
528
+ <div> tag, which contains textual description. The interac-
529
+ tion of the two nodes indicates the web content has both
530
+ visual and textual presentation, forming a strong signal of
531
+ high-quality content.
532
+ • Global relationships. Another important insight is that
533
+ the relationships between local content and global layout
534
+ should also be considered. For example, a node with textual
535
+ description might be critical in a news article but is less
536
+ important in a video webpage, whose quality largely depends
537
+ on the node that contains the video.
538
+ Attentive message passing. To achieve this, we leverage graph
539
+ neural networks that are promising to capture such complicated
540
+ patterns. In particular, we utilize the Graph Attention Network
541
+ (GAT) [30] to model the interactions between nodes in the layout
542
+ graph, where the modeling of node relationships can be viewed as
543
+ message passing [12] among nodes.
544
+ In particular, the architecture of GAT is composed by stacking
545
+ multiple graph attention layers, each of which can be defined as
546
+ �ℎ(𝑘+1)
547
+ 𝑛
548
+ = 𝜎 ��
549
+
550
+ ∑︁
551
+ 𝑚∈N𝑛
552
+ 𝛼𝑛𝑚W(𝑘)
553
+ 1
554
+ �ℎ(𝑘)
555
+ 𝑚 ��
556
+
557
+ ,
558
+ (2)
559
+ where 𝜎(·) is an activation function and 𝛼𝑛𝑚 is the attention value
560
+ between node 𝑛 and node 𝑚. Here, �ℎ(𝑘)
561
+ 𝑛
562
+ represents the embedding
563
+ of node 𝑛 in the 𝑘-th layer. The attention value 𝛼𝑛𝑚 is learned
564
+ to selectively propagate information from neighbour node 𝑚 to
565
+ node 𝑛, and a node can attentively interact more with its important
566
+ neighbours than those trivial ones. Formally, the attention value
567
+
568
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
569
+ Cheng and Liu, et al.
570
+ can be defined as
571
+ 𝛼𝑛𝑚 = softmax𝑚 (𝑒𝑛𝑚) =
572
+ exp (𝑒𝑛𝑚)
573
+
574
+ 𝑗 ∈N𝑛 exp �𝑒𝑛𝑗
575
+ � ,
576
+ (3)
577
+ where the logits 𝑒𝑛𝑗 is computed as
578
+ 𝑒𝑛𝑗 = 𝜎
579
+
580
+ W(𝑘)
581
+ 3
582
+ [W(𝑘)
583
+ 2
584
+ �ℎ(𝑘)
585
+ 𝑛
586
+ ∥W(𝑘)
587
+ 2
588
+ �ℎ(𝑘)
589
+ 𝑗
590
+ ]
591
+
592
+ .
593
+ (4)
594
+ Here, we use ∥ to represent the concatenation operation, and W(𝑘)
595
+ 2
596
+ and W(𝑘)
597
+ 3
598
+ are the weight matrices of the linear transformations
599
+ at the 𝑘-th layer. Note that the weight matrices are shared across
600
+ different nodes in a single graph attention layer.
601
+ After 𝐾 times of message passing, the layout-aware patterns
602
+ could be captured by node interactions (as defined in Eq. (2)) within
603
+ 𝐾-hops. It is worth noting that the virtual node also plays an im-
604
+ portant role during the message passing process. The virtual node
605
+ offers a pathway for nodes’ interaction with considering the global
606
+ interactions in the graph, which is critical for the quality assess-
607
+ ment task. Overall, the GAT-based message passing framework is
608
+ able to comprehensively model both local and global relationships
609
+ for the final task.
610
+ Readout function. To compute the final quality score, we de-
611
+ fine the readout function as mean-pooling [15, 32] to summarize
612
+ all node representations as the final graph representation, and sub-
613
+ sequently adopt a linear layer as
614
+ 𝑠𝑝 = W mean_pooling( �𝐻 (𝐾)
615
+ N
616
+ ) + 𝑏,
617
+ (5)
618
+ where �𝐻 (𝐾)
619
+ N
620
+ is the set of node representations in 𝐾-th layer of GAT.
621
+ Alternatively, we can apply a more reasonable readout function,
622
+ which is to use the representation of the virtual node as the final
623
+ graph representation, and rewrite Eq. (5) as
624
+ 𝑠𝑝 = W�ℎ(𝐾)
625
+ 𝑣
626
+ + 𝑏,
627
+ (6)
628
+ where �ℎ(𝐾)
629
+ 𝑣
630
+ is the virtual node representation in 𝐾-th layer (i.e. the
631
+ last layer) of the model. In such case, the aggregation on the virtual
632
+ node can be viewed as an attentive readout function, which has
633
+ the capability of distinguishing the impact of different nodes in the
634
+ graph for the final task.
635
+ Category-aware quality assessment. The quality score defined
636
+ in Eq. (6) is based on rich information aggregated from nodes. How-
637
+ ever, graph-level information is critical yet not incorporated. There-
638
+ fore, we further improve Eq. (6) with the category information of
639
+ webpage. In particular, we denote the category embedding of a
640
+ given webpage 𝑝 as E(𝜷𝑝), and further rewrite Eq. (6) as
641
+ 𝑠𝑝 = W(�ℎ(𝐾)
642
+ 𝑣
643
+ + E(𝜷𝑝)) + 𝑏.
644
+ (7)
645
+ Note that the category embedding E(𝜷𝑝) has the same dimension-
646
+ ality as the graph embedding �ℎ(𝐾)
647
+ 𝑣
648
+ , such that the embeddings could
649
+ be summed for the final assessment.
650
+ Category-aware data sampling. As the graph-level category
651
+ embedding is introduced in Eq.(7) to perceive different categories
652
+ of webpages, the bias in different categories may affect the predic-
653
+ tion of models. In particular, some webpages are highly similar in
654
+ layout, such as some popular question-answering websites, which
655
+ are generated from templates. Such webpages typically have sim-
656
+ ilar layout scores. Consequently, the predicted assessment score
657
+ may be dominated by the category-aware embedding (i.e. graph
658
+ level embedding). To alleviate this issue, a category-aware sampling
659
+ strategy is employed. Up-sampling is utilized to balance the number
660
+ of two classes, based on which the bias could be mitigated and our
661
+ model could learn a distinguishable quality assessment score for a
662
+ single category of webpages.
663
+ Optimization objective. After up-sampling, the model could
664
+ be optimized through Mean Squared Error (MSE) loss. It can be
665
+ defined as
666
+ 𝐽 = 1
667
+ 𝑃
668
+ 𝑃
669
+ ∑︁
670
+ 𝑝=1
671
+ �𝑦𝑝 − 𝑠𝑝
672
+ �2 ,
673
+ (8)
674
+ where 𝑃 is the total number of training samples after up-sampling
675
+ and 𝑦𝑝 is the annotated layout score of webpage 𝑝.
676
+ 5
677
+ DEPLOYMENT
678
+ In this section, we show how the layout-aware webpage quality
679
+ assessment model be applied to our online ranking system. We
680
+ first introduce the input data construction process of the quality
681
+ assessment model and then present the general picture of the quality
682
+ score working in the ranking system. The overview of deployment
683
+ is shown in Figure 3.
684
+ 5.1
685
+ Offline Input Data Construction
686
+ In the left component of Figure 3, we present the process of input
687
+ data construction for our model. Firstly, each webpage on the world
688
+ wide web will be parsed through our HTML parser. All features
689
+ of the HTML are stored in a database. Secondly, we construct the
690
+ layout graph based on DOM tree and extract the features needed
691
+ for quality assessment model using the algorithm defined in Al-
692
+ gorithm 1. Note that this process runs offline, it can significantly
693
+ reduce the computing time of the online search system.
694
+ We also list the features which are used in our webpage quality
695
+ assessment model, details are shown in Table 3. We classify the
696
+ features into three main categories w.r.t., location, content, and
697
+ layout according to the different roles they play in building webpage.
698
+ Category location is the primarily feature that locates the position
699
+ of elements in the webpage e.g., height, width and position type.
700
+ Category content contains text-related features e.g., the number of
701
+ words, font style, and line height. Category layout is a feature that
702
+ controls the layout of elements, e.g., border, padding, and margin.
703
+ In addition, we add tag name, natural categorical information, and
704
+ webpage category, which is used to balance the distribution of train
705
+ data under different webpage forms.
706
+ 5.2
707
+ Online System Workflow
708
+ The online system workflow is presented in the right component of
709
+ Figure 3. Our ranking system contains a wide variety of webpage
710
+ features, where quality is one of the most important factors. To
711
+ apply our layout-aware webpage quality assessment model in our
712
+ online retrieval system, the new quality scores need to be loaded
713
+ into the retrieval feature list. The online ranking system only needs
714
+ to load the new quality assessment score and apply it to obtain the
715
+ new ranking results with respect to the new ranking webpage list,
716
+
717
+ Layout-aware Webpage Quality Assessment
718
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
719
+ HTML Parser
720
+ HTML
721
+ Database
722
+ Webpages
723
+ Layout Graph
724
+ Construction
725
+ Input Data
726
+ Database
727
+ Webpage
728
+ DOM Tree with Features
729
+ Input Data of Model
730
+ Virtual
731
+ Node
732
+ 𝑵𝟏
733
+ 𝑵𝟐
734
+ ...
735
+ height
736
+ width
737
+ paddling
738
+ margin
739
+ font border ...
740
+ Features
741
+ image
742
+ AirTag
743
+ head
744
+ html
745
+ body
746
+ title
747
+ iMac
748
+ p
749
+ div
750
+ height、width、margin、
751
+ border、padding、font
752
+ size、font style、xpos、
753
+ content length、ypos、
754
+ overflow、visibility ......
755
+ Layout-aware Quality
756
+ Assessment Model
757
+ Quality Score
758
+ Database
759
+ Online Ranking System
760
+ new feature ranking
761
+ list of each webpage
762
+ ......
763
+ Ranking
764
+ System
765
+ new ranking results
766
+ Input Data Construction
767
+ Online System Workflow
768
+ Figure 3: The overview of deployment in online ranking system.
769
+ which is shown in the lower left area of the online component. Note
770
+ that, the quality assessment scores of all webpages are calculated
771
+ offline and are independent of the online search query, thus are
772
+ inefficient for the online search query.
773
+ 6
774
+ OFFLINE EVALUATION
775
+ In this section, we conduct an offline evaluation of the proposed
776
+ layout-aware webpage quality assessment model on the manually-
777
+ labeled dataset from the search engine serves through the offline
778
+ experiments.
779
+ 6.1
780
+ Dataset
781
+ To evaluate the proposed method, we first collect a set of webpages
782
+ from our database, which stores the real webpages that our search
783
+ engine serves. Next, we manually label all the collected webpages on
784
+ our crowdsourcing platform, where a group of experts are required
785
+ to assign low-quality (0) or high-quality (1) to each of the given
786
+ webpage. In our experiments, we use 600,000 webpages for training
787
+ and 20,000 webpages for testing.
788
+ 6.2
789
+ Evaluation Metrics
790
+ Positive-Negative Ratio (PNR). We use PNR to measure the con-
791
+ sistency between manual quality labels and the scores estimated by
792
+ the model. In particular, by enumerating all the pairs of webpages
793
+ in the dataset (i.e., 𝐷), PNR can be formally defined as
794
+ 𝑃𝑁𝑅 =
795
+
796
+ 𝑑𝑖,𝑑𝑗 ∈𝐷 I �𝑦𝑖 > 𝑦𝑗
797
+ � · I �𝑓 (𝑑𝑖) > 𝑓 �𝑑𝑗
798
+ ��
799
+
800
+ 𝑑𝑖′,𝑑𝑗′ ∈𝐷 I �𝑦𝑖′ > 𝑦𝑗′� · I �𝑓 (𝑑𝑖′) < 𝑓 �𝑑𝑗′�� ,
801
+ (9)
802
+ where I is an indicator function, i.e., I (𝑎 > 𝑏) = 1, if 𝑎 > 𝑏, and 0
803
+ otherwise. Here, 𝑓 (𝑑𝑖) represents the quality score of a webpage 𝑑𝑖
804
+ estimated by the model. Higher PNR value indicates better perfor-
805
+ mance of the model.
806
+ Area Under Curve, Precision, Recall, F1-Score. We also report
807
+ Area Under Curve (AUC), Precision (P), Recall (R) and F1-Score (F1)
808
+ to evaluate our proposed model. Precision and recall are often in
809
+ tension, that is, improving precision typically reduces recall and
810
+ vice versa. F1-Score combines them to one performance metric. Area
811
+ under curve summarizes the trade-off between the true positive
812
+ rate and false positive rate for a predictive model using different
813
+ probability thresholds.
814
+ 6.3
815
+ Compared Baselines and Our Approach
816
+ To validate the effectiveness of our layout-aware webpage quality
817
+ model, we conduct experiments on several related baseline mod-
818
+ els: TreeLSTM [29], a standard LSTM architecture designed for
819
+ tree-structured network topologies. GIN [33] introduces a learnable
820
+ parameter to adjust the weight of the central node. GAT [30] lever-
821
+ ages the attention mechanism to improve neighbor aggregation
822
+ scheme. Our proposed models: Virt-GIN has a more expressive
823
+ readout mechanism by adding the virtual node �ℎ𝑣 to GIN model.
824
+ Virt-GAT is our approach similar to virt-GIN model, i.e., a GAT
825
+ model with virtual node. Models-NC: Note that all the above-
826
+ mentioned models use category information as proposed in Section
827
+ 4.3. To further clarify the influence of category in the model, we also
828
+ include four variants without using category information, which is
829
+ denoted with a suffix Non-Category (-NC).
830
+ In addition, we also compare our proposed method with Online
831
+ Baseline, which is the quality assessment model that was previ-
832
+ ously served online in our search engine. This can clearly illustrate
833
+ the improvement brought by the proposed solution for our search
834
+ engine.
835
+ 6.4
836
+ Experimental Settings
837
+ In our experiments, Adam is selected as the optimizer. We use the
838
+ following hyper-parameters: embedding size (64), number layers
839
+ (5), dropout probability (0.2), batch size (32), learning rate (0.0001)
840
+ for GNN models, train epochs (25). As for the TreeLSTM model,
841
+ we set the embedding size (64), dropout probability (0.5), batch size
842
+ (128), learning rate (0.0001), epochs (25) for it. We run 5 experiments
843
+ with different random seeds for all models mentioned above. The
844
+
845
+ <广告>
846
+ iMac
847
+ 新开篇
848
+ 进一步了解》
849
+ 购买>
850
+ AirTag
851
+ 丢三落四这门绝技,要失传了。
852
+ 进一步了解>
853
+ 购买>米非可酮片购买
854
+ QQ:
855
+ 你可能还想找:
856
+ 吃家非司限片有什么及应
857
+ 吃家非司限片有什么别作用
858
+ 吃家非司职片会出自码
859
+ 第一天吃来非可期片
860
+ 晚来非脲片
861
+ 晚完来非司片的反度
862
+ 晚来非司限片有什么用
863
+
864
+ 咨询药师
865
+ 吧药师微信用品益新技检影音乐/安全用品/电子电型改装用
866
+ 品/外维用品内信用品/养护用品自然范用品工投
867
+
868
+ 品牌特区:车墙土
869
+ 送进佳
870
+ Z室组调,今天小学生网小编竭据老师给大家整理了
871
+ 关于汉字(室》的绳调列表,基望下西整理的竞字
872
+ 组调资科及调语解择内容能够助到大家,
873
+ 室字简介
874
+ 首字母:y,群膏:yu,等声调拼音:yo,注音:
875
+ U,部首:穴,部首比划:5,比划:15,第体
876
+ 字:毫,字体结构:上下结构,第画顺序:擦擦折
877
+ PWRY,五第98编码: PWRY, Unicode :
878
+ 服擦操推所除择服摄折探除,五笔86编码
879
+ U+7AB3,双字编号:6008,
880
+ 基本解释
881
+ ●宝yo
882
+ Uo(事物)思务,租务:室务,室
883
+ 败(房效;数坏),室陷(雅务,质量根差),良
884
+ 室(优务),0量:室情,0蜜第
885
+ 京组调
886
+ 掌室(beny):掌重相劣,清线源(圣式记》
887
+ 卷/:“面官修战股,零意不能放洋,转座高组力
888
+ 剩摄之用。
889
+ 事室(bbye):(1).泄气;干事,如:气球欢得个
890
+ 头抵大,但用针一别就癌富了,(2).坑,童度,杨
891
+ 《麦子黄时》:“自卫队上操,有时练习石
892
+ 锁,他能单手掌置负子,一口气连孕十几下,后一
893
+ 敬手,稳出七八步运,肥场地打个大靠,SIGKDD ’23, August 06–10, 2023, Long Beach, CA
894
+ Cheng and Liu, et al.
895
+ Table 4: Offline experimental results of different models.
896
+ Model
897
+ PNR
898
+ AUC (%)
899
+ label 0
900
+ label 1
901
+ P (%)
902
+ R (%)
903
+ F1 (%)
904
+ P (%)
905
+ R (%)
906
+ F1 (%)
907
+ Online Baseline
908
+ 1.51
909
+ 60.10
910
+ 73.44
911
+ 63.99
912
+ 68.39
913
+ 40.09
914
+ 58.57
915
+ 47.60
916
+ TreeLSTM
917
+ 2.91 ± 0.01
918
+ 74.93 ± 0.07
919
+ 79.69 ± 0.01
920
+ 81.86 ± 0.07
921
+ 80.76 ± 0.03
922
+ 57.64 ± 0.06
923
+ 54.19 ± 0.06
924
+ 55.86 ± 0.05
925
+ GIN
926
+ 4.27 ± 0.05
927
+ 81.26 ± 0.20
928
+ 83.82 ± 0.39
929
+ 81.04 ± 1.17
930
+ 82.40 ± 0.44
931
+ 61.22 ± 0.96
932
+ 65.64 ± 1.45
933
+ 63.34 ± 0.26
934
+ GAT
935
+ 4.43 ± 0.06
936
+ 81.94 ± 0.23
937
+ 84.87 ± 1.03
938
+ 79.93 ± 2.21
939
+ 82.30 ± 0.69
940
+ 61.00 ± 1.52
941
+ 68.65 ± 3.40
942
+ 64.53 ± 0.75
943
+ Our Approach
944
+ Virt-GIN
945
+ 4.62 ± 0.03
946
+ 82.47 ± 0.10
947
+ 85.23 ± 0.41
948
+ 78.95 ± 1.49
949
+ 81.96 ± 0.61
950
+ 60.26 ± 1.20
951
+ 69.95 ± 1.56
952
+ 64.72 ± 0.14
953
+ Virt-GAT
954
+ 5.22 ± 0.10
955
+ 84.18 ± 0.24
956
+ 86.81 ± 0.57
957
+ 80.17 ± 1.13
958
+ 83.35 ± 0.35
959
+ 62.75 ± 0.79
960
+ 73.24 ± 1.71
961
+ 67.57 ± 0.29
962
+ Non-Category (-NC)
963
+ GIN-NC
964
+ 4,15 ± 0.07
965
+ 80.80 ± 0.30
966
+ 83.36 ± 1.42
967
+ 81.29 ± 3.57
968
+ 82.25 ± 1.19
969
+ 61.24 ± 2.62
970
+ 64.20 ± 5.15
971
+ 62.50 ± 1.17
972
+ GAT-NC
973
+ 4.26 ± 0.04
974
+ 81.27 ± 0.15
975
+ 83.85 ± 0.23
976
+ 81.11 ± 0.75
977
+ 82.46 ± 0.31
978
+ 61.32 ± 0.67
979
+ 65.70 ± 0.87
980
+ 63.43 ± 0.25
981
+ Virt-GIN-NC
982
+ 4.48 ± 0.04
983
+ 82.05 ± 0.14
984
+ 84.70 ± 0.53
985
+ 79.49 ± 1.27
986
+ 82.01 ± 0.44
987
+ 60.35 ± 0.86
988
+ 68.45 ± 1.77
989
+ 64.13 ± 0.33
990
+ Virt-GAT-NC
991
+ 5.03 ± 0.03
992
+ 83.66 ± 0.08
993
+ 85.99 ± 0.48
994
+ 81.40 ± 1.23
995
+ 83.62 ± 0.43
996
+ 63.47 ± 1.04
997
+ 70.86 ± 1.60
998
+ 66.94 ± 0.25
999
+ final result we reported is the mean test AUC, Precision, Recall,
1000
+ F1-Score and their corresponding standard deviation. All the above
1001
+ mentioned GNN models are implemented by Paddle Graph Learning
1002
+ (PGL)1, an efficient and flexible graph learning framework.
1003
+ 6.5
1004
+ Offline Experimental Results
1005
+ We report the offline experimental results of the proposed model
1006
+ and all baseline models. Besides, we also include a baseline method,
1007
+ i.e., the model that is used in the system before deploying the layout-
1008
+ aware webpage quality assessment model.
1009
+ All results are shown in Table 4, from where we have the follow-
1010
+ ing key findings:
1011
+ • We can clearly see that our layout-aware webpage qual-
1012
+ ity model can beat the online baseline by large margins on
1013
+ all metrics e.g., Δ𝐴𝑈𝐶 = 24.08, Δ𝐹1 = 14.96 (label0) and
1014
+ Δ𝐹1 = 19.97 (label1). Especially for PNR, where the value is
1015
+ improved from 1.51 to 5.22. These tell us that the proposed
1016
+ model prefers high-quality results.
1017
+ • By applying the proposed readout function, the model can
1018
+ have a significant improvement on all metrics. Especially,
1019
+ the new readout mechanism is able to improve PNR by a
1020
+ margin of 0.38 and 0.96 based on GIN and GAT, respectively.
1021
+ Moreover, we also observe that the relative improvement
1022
+ of both virt-GIN and virt-GAT over GIN and GAT is consid-
1023
+ erable for high-quality webpage (label1), in terms of recall
1024
+ (Δ(𝑉𝑖𝑟𝑡_𝐺𝐴𝑇,𝐺𝐴𝑇) = 4.59%, Δ(𝑉𝑖𝑟𝑡_𝐺𝐼𝑁,𝐺𝐼𝑁 ) = 2.22%). All
1025
+ these phenomena show that our readout mechanism is capa-
1026
+ ble of improving the model’s performance.
1027
+ • Comparing the results of the two models whether apply
1028
+ the category-aware optimization strategy (w,r,t., GIN-NC
1029
+ vs. GIN, Virt-GIN-NC vs. Virt-GIN, GAT-NC vs. GAT, Virt-
1030
+ GAT-NC vs. Virt-GAT), we can come to the conclusion that
1031
+ all methods with the proposed category-aware optimization
1032
+ have better performance than their backbone models, in
1033
+ terms of PNR and AUC. Although a few models obtain lower
1034
+ 1https://github.com/PaddlePaddle/PGL
1035
+ values on a few metrics (e.g., the F1-score of Virt-GAT-NC on
1036
+ label0 is 83.62 while Virt-GAT is 83.35, the precision of Virt-
1037
+ GAT-NC is 63.74% but Virt-GAT is 62.75%), the models with
1038
+ category-aware optimization show more robust performance
1039
+ considering all metrics.
1040
+ • The performance on different GNN models is better than
1041
+ TreeLSTM, model Virt_GAT is the most significant, Com-
1042
+ pare with Virt_GAT and TreeLSTM, Δ𝑃𝑁𝑅 = 2.31, Δ𝐴𝑈𝐶 =
1043
+ 9.25%. For high-quality webpage (label1) Δ𝑅 = 14.67%. These
1044
+ large margins suggest that our model is more expressive than
1045
+ TreeLSTM, although TreeLSTM is specifically designed for
1046
+ tree-structured network topologies.
1047
+ Overall, our proposed model is able to gain superior performance
1048
+ on webpage assessment task through the improved readout mech-
1049
+ anism and category-aware optimization and can beat the online
1050
+ baseline by a significant margin.
1051
+ 6.6
1052
+ Varying the number of GNN layer
1053
+ In general, a webpage is represented as a DOM tree. Its depth deter-
1054
+ mines how many layers of GNN are needed to obtain information
1055
+ from the root node to the leaf nodes. However, as the number of
1056
+ GNN layers increases, the computational efficiency will be lower.
1057
+ Therefore, we provide an experiment to verify the influence of the
1058
+ number of layers on the experimental results, as shown in Table
1059
+ 5. As seen from the table, the more layers, the higher the AUC
1060
+ score can be reached. However, compared with the 5-layer virt-
1061
+ GAT model, the improvement of 7-layer virt-GAT model is not
1062
+ significant. As it is important to trade off the efficiency and effec-
1063
+ tiveness for large search system, we use 5-layer GNN models on
1064
+ online evaluation which can maintain the experimental effect while
1065
+ reducing the amount of calculation.
1066
+ 7
1067
+ ONLINE EVALUATION
1068
+ To investigate the impact of our proposed quality assessment model
1069
+ to the search engine, we deploy the new model and conduct online
1070
+ experiments to compare it with the old retrieval system. Specifically,
1071
+
1072
+ Layout-aware Webpage Quality Assessment
1073
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
1074
+ Table 5: The influence of layer number on virt-GAT.
1075
+ #Layers
1076
+ AUC (%)
1077
+ label 0
1078
+ label 1
1079
+ P (%)
1080
+ R (%)
1081
+ F1 (%)
1082
+ P (%)
1083
+ R (%)
1084
+ F1 (%)
1085
+ 1
1086
+ 80.77 ± 0.23
1087
+ 83.38 ± 0.92
1088
+ 80.46 ± 2.49
1089
+ 81.87 ± 0.85
1090
+ 60.26 ± 1.82
1091
+ 64.72 ± 3.42
1092
+ 62.33 ± 0.66
1093
+ 3
1094
+ 83.80 ± 0.27
1095
+ 86.25 ± 0.64
1096
+ 79.80 ± 2.06
1097
+ 82.89 ± 0.86
1098
+ 61.98 ± 1.79
1099
+ 72.05 ± 2.16
1100
+ 66.59 ± 0.45
1101
+ 5
1102
+ 84.18 ± 0.24
1103
+ 86.81 ± 0.57
1104
+ 80.17 ± 1.13
1105
+ 83.35 ± 0.35
1106
+ 62.75 ± 0.79
1107
+ 73.24 ± 1.71
1108
+ 67.57 ± 0.29
1109
+ 7
1110
+ 84.25 ± 0.22
1111
+ 86.91 ± 0.80
1112
+ 80.53 ± 1.77
1113
+ 83.58 ± 0.61
1114
+ 63.23 ± 1.41
1115
+ 73.32 ± 2.43
1116
+ 67.86 ± 0.42
1117
+ we conduct a manual evaluation on the final ranking results with
1118
+ some real user-generated queries. This directly reflects the quality
1119
+ of the results exposed to the end users.
1120
+ We log a set of (million-scale) online queries and the correspond-
1121
+ ing final impressions, i.e., the top-ranked web documents in the
1122
+ final ranking stage, by individually using the layout-aware web-
1123
+ page quality assessment model and the old retrieval systems. Note
1124
+ that the data logging is conducted by multiple rounds to eliminate
1125
+ randomness. We filter out examples in which queries have identical
1126
+ impressions between the two systems, and then utilize the rest for
1127
+ the manual evaluation. Note that, considering the extremely high
1128
+ cost of the manual evaluation, we randomly generate thousands of
1129
+ data and eventually send it to experts for evaluation, so as to control
1130
+ costs while validating the effectiveness of the proposed model.
1131
+ 7.1
1132
+ Online Experimental Metrics
1133
+ As mentioned in Section 5, our proposed quality assessment model
1134
+ works in Baidu retrieval system. The online experiments major
1135
+ focus on the end-to-end evaluation, the metrics are often used to
1136
+ measure the effectiveness of information retrieval system. Details
1137
+ are as follows:
1138
+ Discounted Cumulative Gain (DCG). We first log a dataset
1139
+ and manually label the data with 0 to 4 grades, and then report
1140
+ the relative improvement w.r.t. the average DCG over the top-4
1141
+ final results of all queries. The formula of DCG accumulated at a
1142
+ particular rank position p is defined as
1143
+ DCGp =
1144
+ 𝑝
1145
+ ∑︁
1146
+ 𝑖=1
1147
+ 2𝑟𝑒𝑙𝑖 − 1
1148
+ log2(𝑖 + 1) ,
1149
+ (10)
1150
+ where 𝑟𝑒𝑙𝑖 indicates the manually label of 𝑖-th webpage.
1151
+ Additionally, we also report the relative improvement of DCG
1152
+ for the low quality ranking result w.r.t., manually label is 0/1.
1153
+ Side-by-side Comparison. Besides, we also conduct a side-by-
1154
+ side comparison between the two systems. We log another dataset
1155
+ and require the human experts to judge whether the new system or
1156
+ the base system gives better results that satisfy intentions of users.
1157
+ Here, the relative gain is measured Good vs. Same vs. Bad (GSB) as
1158
+ Δ𝐺𝑆𝐵 =
1159
+ #Good − #Bad
1160
+ #Good + #Same + #Bad,
1161
+ (11)
1162
+ where #Good (or #Bad) indicates the number of queries that the
1163
+ new system provides better (or worse) final results.
1164
+ Table 6: Discounted cumulative gain on manual evaluation.
1165
+ Rand-Query
1166
+ Tail-Query
1167
+ Same-Quality
1168
+ Δ𝐷𝐶𝐺
1169
+ +0.19%
1170
+ +0.42%
1171
+ -
1172
+ DCG_0/1 ratio
1173
+ -0.63%
1174
+ -0.56%
1175
+ -
1176
+ Table 7: Side-by-side comparison on manual evaluation.
1177
+ Rand-Query
1178
+ Tail-Query
1179
+ Same-Quality
1180
+ Δ𝐺𝑆𝐵
1181
+ +4.10%
1182
+ +0.52%
1183
+ +5.13%
1184
+ Node that we not only measure the final results but also measure
1185
+ the webpage quality when the relative result of two webpage is
1186
+ Same.
1187
+ 7.2
1188
+ Online Experimental Results
1189
+ The relative improvement validated by manual evaluation is given
1190
+ in Table 6 and 7, where we can summarize observations as below:
1191
+ • By applying our quality assessment model, the system can
1192
+ significantly outperform the base system. Especially for DCG_0/1
1193
+ ratio, the relative improvement values are respectively −0.63%,
1194
+ −0.56% for rand query and tail query. This shows that our
1195
+ proposed method can better filtrate retrieval results with
1196
+ low DCG scores, which is very helpful in improving the user
1197
+ experience for real-world search engine.
1198
+ • The conventional case-by-case comparison also has signifi-
1199
+ cant improvement over the base system, especially for the
1200
+ rand query (Δ𝐺𝑆𝐵 = +4.1%). This tells us that user experi-
1201
+ ence can be improved by taking into account the web page
1202
+ quality in search system.
1203
+ • In addition, we can observe that with comparable relevance,
1204
+ the GSB value of the quality improvement is Δ𝐺𝑆𝐵 = +5.13%.
1205
+ This intuitively shows that our new system can provide
1206
+ higher quality search results based on the guaranteed rele-
1207
+ vance of search results.
1208
+ Moreover, we perform the statistical test to estimate whether
1209
+ the experimental results is statistically significant. The p-value of
1210
+ DCG rand and tail query are 0.0613 and 0.1276, respectively. The p-
1211
+ value approximates the significance level that is set in our retrieval
1212
+
1213
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
1214
+ Cheng and Liu, et al.
1215
+ (a) Offline quality assessment
1216
+ (b) Online position changes
1217
+ Figure 4: The overview of case study.
1218
+ system, which can demonstrate that our experimental results are
1219
+ statistically significant.
1220
+ Overall, the online experimental results show that our proposed
1221
+ layout-aware quality assessment model can effectively improve the
1222
+ performance of real-world ranking system.
1223
+ 8
1224
+ CASE STUDY
1225
+ In this section, we present an illustration that includes the offline
1226
+ quality assessment score of webpage and online position changes
1227
+ of web pages. These typically cases are shown in Figure 4.
1228
+ 8.1
1229
+ Offline Quality Assessment
1230
+ In Figure 4(a), we present three webpages with different layout
1231
+ styles and their quality assessment scores.
1232
+ The first webpage has a chaotic layout, elements in this web-
1233
+ page are unreasonable. It affects the user’s normal browsing and
1234
+ is very difficult for user to obtain information from this webpage.
1235
+ Our quality assessment model marks this webpage as low quality
1236
+ (𝑠𝑐𝑜𝑟𝑒 = 0.0068). This extremely low score will be considered by
1237
+ the ranking system to lower its ranking position.
1238
+ The second webpage also has low quality, different with the
1239
+ chaotic layout of the first webpage, it has a normal layout. How-
1240
+ ever, considering that it contains very small amount of information
1241
+ (almost no valuable information), it should be presented to the user
1242
+ with a very small probability. The ranking system can judge this
1243
+ by our quality assessment model score 0.1653.
1244
+ Unlike the previous two webpages, the third one is high-quality.
1245
+ It is carefully laid out and informative, and quality score is 0.9788,
1246
+ which will help the ranking system raise its ranking position.
1247
+ 8.2
1248
+ Online Position Changes
1249
+ The case shown in Figure 4(b) comes from Section 7. Under the same
1250
+ query, these two webpages swapped positions in the new and old
1251
+ systems, The position of the left webpage in new system is 3-th but 4-
1252
+ th in the old system. Comparing the two webpages, we can observe
1253
+ that the left webpage (quality score is 0.5623) contains a rich amount
1254
+ of information but the right one (quality score is 0.2415) does not.
1255
+ This phenomenon demonstrates that online ranking system has
1256
+ adopted our model’s recommendations to provide users with higher
1257
+ quality webpage, which can greatly improve the user experience.
1258
+ 9
1259
+ CONCLUSION AND FUTURE WORK
1260
+ In this paper, we propose a layout-aware webpage assessment model
1261
+ to suggest ranking system providing webpages with higher quality.
1262
+ We not only enhance GAT with the read mechanism but also care-
1263
+ fully design the features for improving the quality assessment on
1264
+ the webpages. In addition, taking into account the particularity of
1265
+ real-world data, we utilize the category of webpage for optimiza-
1266
+ tion. Both input data construction and model calculation are offline,
1267
+ which guarantees the efficiency of the ranking system. We devel-
1268
+ oped and deployed the layout-aware webpage assessment model in
1269
+ Baidu Search, which is highly effective in conducting high-quality
1270
+ ranking for web search. Extensive offline and online experiments
1271
+ have shown that the ranking system can significantly improve the
1272
+ effectiveness and general usability of the search engine.
1273
+ In future work, we will explore the heterogeneous GNN architec-
1274
+ ture to model the multiple graph-based information of webpages.
1275
+ It is interesting to improve the construction method of layout and
1276
+ enhance the representation of nodes/edges with self-supervised
1277
+ contrastive pre-training techniques.
1278
+ REFERENCES
1279
+ [1] Armen Avetisyan, Tatiana Khanova, Christopher Bongsoo Choy, Denver Dash,
1280
+ Angela Dai, and Matthias Nießner. 2020. SceneCAD: Predicting Object Align-
1281
+ ments and Layouts in RGB-D Scans. ArXiv abs/2003.12622 (2020).
1282
+ [2] Federico Baldassarre, David Ménendez Hurtado, Arne Elofsson, and Hossein
1283
+ Azizpour. 2021. GraphQA: protein model quality assessment using graph convo-
1284
+ lutional networks. Bioinformatics 37 (2021), 360 – 366.
1285
+
1286
+ 安全国品电子电器政装用品
1287
+ 外饰用品内饰用品养护用品自空游用品正报价
1288
+ M
1289
+ Opma适8轮胎防摄膜章道询
1290
+ 昆聘特区:车博土
1291
+ 巴洛克
1292
+ 喜普风
1293
+ 香玉儿
1294
+ 迪佳
1295
+ 键科
1296
+ 暖忆
1297
+ H
1298
+ H
1299
+ H同酮片购买
1300
+ Q@:
1301
+ 你可自能摄线:
1302
+ 吃米非司酮片后流血
1303
+ 吃完米非司酮片的反应
1304
+ 吃了米非司酮片流血
1305
+ 吃米非司酮片有什么症状
1306
+ 服米非司酮片
1307
+ 吃米非司酮片第二天出血
1308
+ 吃米非司酮片的注意事项
1309
+ 吃米非司酮片会流血吗
1310
+ 开门红
1311
+ 咨询药师
1312
+ 药师微信什公值得买
1313
+ Q搜索分类/品牌/商品
1314
+ 打开
1315
+ 全部奶价
1316
+ 社区
1317
+ 商品百料
1318
+ 抗事等
1319
+ 要爽玩《魔兽世界:争霸艾泽拉斯》
1320
+ CPU
1321
+ 鲁118-88-50
1322
+ +美注
1323
+ RYZEN
1324
+ WORLDL
1325
+ 于一
1326
+
1327
+ 家宝业,额心后上靠玩安成的原国,可情不远
1328
+ 大本身视普世养料
1329
+ MAGB550MMORTARWiFI迫击炮+...
1330
+ ¥2392
1331
+ 1899
1332
+
1333
+ 市场价
1334
+ 值得买APP专享价代公值得买
1335
+ Q搜索分类/品牌/商品
1336
+ 打开
1337
+ 泡好价
1338
+ 全部好价
1339
+ 社区
1340
+ 商品百料
1341
+ 捷惠等
1342
+ 甲全部评论(119)
1343
+ 13心
1344
+ 3068-08-0
1345
+
1346
+ ALARE
1347
+
1348
+ 3008488-30
1349
+ 爱有年天
1350
+ 4心
1351
+ MAGB550MMORTARWiFI迫击炮+...
1352
+ X
1353
+ ¥2399
1354
+ 1899
1355
+
1356
+ 市场价
1357
+ 值得买APP专享价S搜狐网
1358
+ 晴/1°三
1359
+ 如何搭迪和自己不熟的女同事-如何搭让
1360
+ 如何搭训技巧之一:微笑地说出对方的名字
1361
+ 对于安生而言,如果一入男士非常绅士地对
1362
+ 她微笑,并具当着她的面,自然友好地叫出
1363
+ 了她的名字,她肯定会感到惊访,但随之而
1364
+ 来的更多是欢喜。这种搭训会让女生瞬时记
1365
+ 住自己,并且留下较好的印象。
1366
+ 因为你微笑对她,她也会回以礼貌的微笑
1367
+ 然后她会反问:“你怎么知道我的名学”。就
1368
+ 解中度过愉悦的时光。彼此都给对方以缸服
1369
+ 的感觉,这为下次的聊天或相聚做好
1370
+ 垫。
1371
+ 讨论5女相100男,徐州一相亲大会男女比例失调..>
1372
+
1373
+
1374
+ 我来说两句
1375
+ C知乎
1376
+ Q中国灵..
1377
+ 下载App
1378
+ 注册登录
1379
+ 不认识的同事(女)如何搭训?
1380
+ 关注问题
1381
+ 写回答
1382
+ 1个回答
1383
+ 小白兔
1384
+ 谢邀,建议先从她认识的人入手,比如拿联系方
1385
+ 式,然后可以先切入聊天
1386
+ 发布于2020-11-1421:17
1387
+ 一费同
1388
+ 评论
1389
+ 智能消费
1390
+ 新浪潮
1391
+ 创卷调研
1392
+ 广告
1393
+ 相关推荐
1394
+ 男朋友打游戏正确的处理方式,这个
1395
+ 女朋友不能要了
1396
+ 微博的广告
1397
+ 如何去搭陌生人
1398
+ haoyunlai2188的文章
1399
+ ApP内打开Layout-aware Webpage Quality Assessment
1400
+ SIGKDD ’23, August 06–10, 2023, Long Beach, CA
1401
+ [3] Yoshua Bengio, Aaron Courville, and Pascal Vincent. 2013.
1402
+ Representation
1403
+ learning: A review and new perspectives. IEEE transactions on pattern analysis
1404
+ and machine intelligence 35, 8 (2013), 1798–1828.
1405
+ [4] Sanket Biswas, Pau Riba, Josep Llad’os, and Umapada Pal. 2021. Graph-based Deep
1406
+ Generative Modelling for Document Layout Generation. In ICDAR Workshops.
1407
+ [5] A Caro, Coral Calero, Ismael Caballero, and Mario Piattini. 2005. Data qual-
1408
+ ity in web applications: A state of the art. In IADIS International Conference
1409
+ WWW/Internet, Vol. 2. 364–368.
1410
+ [6] Angélica Caro, Coral Calero, Ismael Caballero, and Mario Piattini. 2006. Defining a
1411
+ data quality model for web portals. In International Conference on Web Information
1412
+ Systems Engineering. Springer, 363–374.
1413
+ [7] Angelica Caro, Coral Calero, Houari A Sahraoui, and Mario Piattini. 2007. A
1414
+ Bayesian network to represent a data quality model. International Journal of
1415
+ Information Quality 1, 3 (2007), 272–294.
1416
+ [8] Selcuk Cebi. 2013. A quality evaluation model for the design quality of online
1417
+ shopping websites. Electronic Commerce Research and Applications 12, 2 (2013),
1418
+ 124–135.
1419
+ [9] Yingshan Chang, Mridu Baldevraj Narang, Hisami Suzuki, Guihong Cao, Jianfeng
1420
+ Gao, and Yonatan Bisk. 2021. WebQA: Multihop and Multimodal QA. ArXiv
1421
+ abs/2109.00590 (2021).
1422
+ [10] Gongwei Chen, Xinhang Song, Haitao Zeng, and Shuqiang Jiang. 2020. Scene
1423
+ Recognition With Prototype-Agnostic Scene Layout. IEEE Transactions on Image
1424
+ Processing 29 (2020), 5877–5888.
1425
+ [11] Kaize Ding, Jianling Wang, Jundong Li, Dingcheng Li, and Huan Liu. 2020. Be
1426
+ More with Less: Hypergraph Attention Networks for Inductive Text Classification.
1427
+ ArXiv abs/2011.00387 (2020).
1428
+ [12] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E
1429
+ Dahl. 2017. Neural message passing for quantum chemistry. In International
1430
+ conference on machine learning. PMLR, 1263–1272.
1431
+ [13] Jiafeng Guo, Yinqiong Cai, Yixing Fan, Fei Sun, Ruqing Zhang, and Xueqi Cheng.
1432
+ 2021. Semantic models for the first-stage retrieval: A comprehensive review.
1433
+ arXiv preprint arXiv:2103.04831 (2021).
1434
+ [14] Divya Gupta. 2021. An empirical study of predictive model for website quality
1435
+ analytics using dataset of different domains of websites. Recent Advances in
1436
+ Computer Science and Communications (Formerly: Recent Patents on Computer
1437
+ Science) 14, 4 (2021), 995–1007.
1438
+ [15] Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation
1439
+ learning on large graphs. Advances in neural information processing systems 30
1440
+ (2017).
1441
+ [16] Layla Hasan and Emad Abuelrub. 2011. Assessing the quality of web sites. Applied
1442
+ computing and informatics 9, 1 (2011), 11–29.
1443
+ [17] Daniel Hasan Dalip, Marcos André Gonçalves, Marco Cristo, and Pável Calado.
1444
+ 2009. Automatic quality assessment of content created collaboratively by web
1445
+ communities: a case study of wikipedia. In Proceedings of the 9th ACM/IEEE-CS
1446
+ joint conference on Digital libraries. 295–304.
1447
+ [18] Lianzhe Huang, Dehong Ma, Sujian Li, Xiaodong Zhang, and Houfeng Wang.
1448
+ 2019. Text level graph neural network for text classification. arXiv preprint
1449
+ arXiv:1910.02356 (2019).
1450
+ [19] Lianzhe Huang, Dehong Ma, Sujian Li, Xiaodong Zhang, and Houfeng Wang.
1451
+ 2019. Text Level Graph Neural Network for Text Classification. In EMNLP.
1452
+ [20] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. 2022. LayoutLMv3:
1453
+ Pre-training for Document AI with Unified Text and Image Masking. ArXiv
1454
+ abs/2204.08387 (2022).
1455
+ [21] Thomas Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph
1456
+ Convolutional Networks. ArXiv abs/1609.02907 (2017).
1457
+ [22] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. nature
1458
+ 521, 7553 (2015), 436–444.
1459
+ [23] Yiding Liu, Weixue Lu, Suqi Cheng, Daiting Shi, Shuaiqiang Wang, Zhicong
1460
+ Cheng, and Dawei Yin. 2021. Pre-trained language model for web-scale retrieval
1461
+ in baidu search. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge
1462
+ Discovery & Data Mining. 3365–3375.
1463
+ [24] Vassilis Moustakis, Charalambos Litos, Andreas Dalivigas, Loukas Tsironis, et al.
1464
+ 2004. Website Quality Assessment Criteria.. In ICIQ. 59–73.
1465
+ [25] Vahid Rafe and Maryam Monfaredzadeh. 2012. A qualitative framework to assess
1466
+ hospital/medical websites. Journal of medical systems 36, 5 (2012), 2927–2939.
1467
+ [26] Soumya Sanyal, Ivan V. Anishchenko, Anirudh Dagar, David Baker, and
1468
+ Partha Pratim Talukdar. 2020. ProteinGCN: Protein model quality assessment
1469
+ using Graph Convolutional Networks. bioRxiv (2020).
1470
+ [27] Artur M. Schweidtmann, Jan G. Rittig, Andrea König, Martin Grohe, Alexander
1471
+ Mitsos, and Manuel Dahmen. 2020. Graph Neural Networks for Prediction of
1472
+ Fuel Ignition Quality. Energy & Fuels (2020).
1473
+ [28] Yelong Shen, Xiaodong He, Jianfeng Gao, Li Deng, and Grégoire Mesnil. 2014.
1474
+ Learning semantic representations using convolutional neural networks for web
1475
+ search. In Proceedings of the 23rd international conference on world wide web.
1476
+ 373–374.
1477
+ [29] Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved
1478
+ semantic representations from tree-structured long short-term memory networks.
1479
+ arXiv preprint arXiv:1503.00075 (2015).
1480
+ [30] Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro
1481
+ Lio, and Yoshua Bengio. 2017.
1482
+ Graph attention networks.
1483
+ arXiv preprint
1484
+ arXiv:1710.10903 (2017).
1485
+ [31] Te-Lin Wu, Cheng Li, Mingyang Zhang, Tao Chen, Spurthi Amba Hombaiah, and
1486
+ Michael Bendersky. 2021. LAMPRET: Layout-Aware Multimodal PreTraining for
1487
+ Document Understanding. arXiv preprint arXiv:2104.08405 (2021).
1488
+ [32] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and
1489
+ S Yu Philip. 2020. A comprehensive survey on graph neural networks. IEEE
1490
+ transactions on neural networks and learning systems 32, 1 (2020), 4–24.
1491
+ [33] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful
1492
+ are graph neural networks? arXiv preprint arXiv:1810.00826 (2018).
1493
+ [34] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. 2020.
1494
+ LayoutLM: Pre-training of Text and Layout for Document Image Understanding.
1495
+ Proceedings of the 26th ACM SIGKDD International Conference on Knowledge
1496
+ Discovery & Data Mining (2020).
1497
+ [35] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan
1498
+ Lu, Dinei A. F. Florêncio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong
1499
+ Zhou. 2021. LayoutLMv2: Multi-modal Pre-training for Visually-rich Document
1500
+ Understanding. ArXiv abs/2012.14740 (2021).
1501
+ [36] Zhiju Yang, Weiping Pei, Mon-Chu Chen, and Chuan Yue. 2022. WTAGRAPH:
1502
+ Web Tracking and Advertising Detection using Graph Neural Networks. 2022
1503
+ IEEE Symposium on Security and Privacy (SP) (2022), 1540–1557.
1504
+ [37] Liang Yao, Chengsheng Mao, and Yuan Luo. 2019. Graph convolutional net-
1505
+ works for text classification. In Proceedings of the AAAI conference on artificial
1506
+ intelligence, Vol. 33. 7370–7377.
1507
+ [38] Dawei Yin, Yuening Hu, Jiliang Tang, Tim Daly, Mianwei Zhou, Hua Ouyang,
1508
+ Jianhui Chen, Changsung Kang, Hongbo Deng, Chikashi Nobata, et al. 2016.
1509
+ Ranking relevance in yahoo search. In Proceedings of the 22nd ACM SIGKDD
1510
+ International Conference on Knowledge Discovery and Data Mining. 323–332.
1511
+ [39] Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He,
1512
+ Yanming Shen, and Tie-Yan Liu. 2021. Do Transformers Really Perform Bad for
1513
+ Graph Representation?. In NeurIPS.
1514
+ [40] Junbi Zhang, Xu Ma, Shengen Zhang, Xianqiang Zheng, Rui Chen, Yihua Pan,
1515
+ Lisong Dong, Yayi Wei, and Gonzalo R. Arce. 2021. Lithography layout classifica-
1516
+ tion based on graph convolution network. In Advanced Lithography.
1517
+ [41] Yufeng Zhang, Xueli Yu, Zeyu Cui, Shu Wu, Zhongzhen Wen, and Liang Wang.
1518
+ 2020. Every document owns its structure: Inductive text classification via graph
1519
+ neural networks. arXiv preprint arXiv:2004.13826 (2020).
1520
+ [42] Lixin Zou, Shengqiang Zhang, Hengyi Cai, Dehong Ma, Suqi Cheng, Shuaiqiang
1521
+ Wang, Daiting Shi, Zhicong Cheng, and Dawei Yin. 2021. Pre-trained language
1522
+ model based ranking in Baidu search. In Proceedings of the 27th ACM SIGKDD
1523
+ Conference on Knowledge Discovery & Data Mining. 4014–4022.
1524
+
89FLT4oBgHgl3EQfti_6/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
CNFKT4oBgHgl3EQfXy51/content/tmp_files/2301.11796v1.pdf.txt ADDED
@@ -0,0 +1,521 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Call for Papers - The BabyLM Challenge: Sample-efficient pretraining
2
+ on a developmentally plausible corpus
3
+ https://babylm.github.io/
4
+ Alex Warstadt
5
+ ETH Zürich
6
+ Leshem Choshen
7
+ IBM Research
8
+ Aaron Mueller
9
+ Johns Hopkins University
10
+ Ethan Wilcox
11
+ ETH Zürich
12
+ Adina Williams
13
+ Meta AI
14
+ Chengxu Zhuang
15
+ MIT
16
+ Abstract
17
+ We present the call for papers for the BabyLM
18
+ Challenge: Sample-efficient pretraining on a
19
+ developmentally plausible corpus. This shared
20
+ task is intended for participants with an in-
21
+ terest in small scale language modeling, hu-
22
+ man language acquisition, low-resource NLP,
23
+ and cognitive modeling. In partnership with
24
+ CoNLL and CMCL, we provide a platform for
25
+ approaches to pretraining with a limited-size
26
+ corpus sourced from data inspired by the input
27
+ to children. The task has three tracks, two of
28
+ which restrict the training data to pre-released
29
+ datasets of 10M and 100M words and are ded-
30
+ icated to explorations of approaches such as
31
+ architectural variations, self-supervised objec-
32
+ tives, or curriculum learning. The final track
33
+ only restricts the amount of text used, allowing
34
+ innovation in the choice of the data, its domain,
35
+ and even its modality (i.e., data from sources
36
+ other than text is welcome). We will release a
37
+ shared evaluation pipeline which scores mod-
38
+ els on a variety of benchmarks and tasks, in-
39
+ cluding targeted syntactic evaluations and nat-
40
+ ural language understanding.
41
+ 1
42
+ Motivation
43
+ Huge efforts have been put into optimizing LM pre-
44
+ training at massive scales in the last several years
45
+ (Raffel et al., 2020; Brown et al., 2020; Chowdhery
46
+ et al., 2022; Hoffmann et al., 2022). While grow-
47
+ ing parameter counts often get the most attention,
48
+ datasets have also grown by orders of magnitude.
49
+ These increasingly larger pretraining datasets are
50
+ visualized, to scale, in Figure 1. At the same time,
51
+ there has been almost no progress in pretraining at
52
+ smaller human-like data scales.
53
+ Focusing on scaled-down pretraining has several
54
+ potential benefits: First, small-scale pretraining can
55
+ be a sandbox for developing novel techniques that
56
+ improve data efficiency. These techniques have the
57
+ potential to then scale up to larger datasets com-
58
+ monly seen in applied NLP, and could be used
59
+ Figure 1: Data Scale: Modern Language Models are
60
+ trained on data multiple orders of magnitude larger than
61
+ the amount available to a typical human child. Image
62
+ based off Fig. 1 from Warstadt and Bowman (2022)
63
+ to enhance current approaches to modeling low-
64
+ resource languages. Second, improving our ability
65
+ to train LMs on the same types and quantities of
66
+ data that humans learn from will give us greater
67
+ access to more plausible cognitive models of hu-
68
+ mans and help us understand what allows humans
69
+ to acquire language so efficiently (Keller, 2010;
70
+ Dupoux, 2018). That is, even model failure can
71
+ help in developing hypotheses about the differences
72
+ between human and LM language learning.
73
+ The goal of this shared task will be to incentivize
74
+ researchers with an interest in pretraining and/or
75
+ cognitive modeling to focus their efforts on opti-
76
+ mizing pretraining given data limitations inspired
77
+ by human development. Additionally, we hope to
78
+ democratize research on pretraining—which is typ-
79
+ ically thought to be practical only for large industry
80
+ groups—by drawing attention to open problems
81
+ that can be addressed on a university budget.
82
+ 2
83
+ Key Dates
84
+ • January 2023: Training data released
85
+ • March 2023: Evaluation pipeline released
86
+ • July 15, 2023: Results due
87
+ • August 1, 2023: Paper submissions due
88
+ • Date TBA: Presentation at CoNLL
89
+ arXiv:2301.11796v1 [cs.CL] 27 Jan 2023
90
+
91
+ 200
92
+ 1.4
93
+ Billion
94
+ Trillion
95
+ 30
96
+ 3
97
+ <100
98
+ Billion
99
+ Billion
100
+ Million
101
+ 13 y.0.
102
+ BERT
103
+ RoBERTa
104
+ GPT-3
105
+ Chinchilla
106
+ Human
107
+ (2018)
108
+ (2019)
109
+ (2020)
110
+ (2022)# Words
111
+ Dataset
112
+ Domain
113
+ STRICT-SMALL
114
+ STRICT
115
+ Proportion
116
+ CHILDES (MacWhinney, 2000)
117
+ Child-directed speech
118
+ 0.44M
119
+ 4.21M
120
+ 5%
121
+ British National Corpus (BNC),1 dialogue portion
122
+ Dialogue
123
+ 0.86M
124
+ 8.16M
125
+ 8%
126
+ Children’s Book Test (Hill et al., 2016)
127
+ Children’s books
128
+ 0.57M
129
+ 5.55M
130
+ 6%
131
+ Children’s Stories Text Corpus2
132
+ Children’s books
133
+ 0.34M
134
+ 3.22M
135
+ 3%
136
+ Standardized Project Gutenberg Corpus (Gerlach and Font-Clos, 2018)
137
+ Written English
138
+ 0.99M
139
+ 9.46M
140
+ 10%
141
+ OpenSubtitles (Lison and Tiedemann, 2016)
142
+ Movie subtitles
143
+ 3.09M
144
+ 31.28M
145
+ 31%
146
+ QCRI Educational Domain Corpus (QED; Abdelali et al., 2014)
147
+ Educational video subtitles
148
+ 1.04M
149
+ 10.24M
150
+ 11%
151
+ Wikipedia3
152
+ Wikipedia (English)
153
+ 0.99M
154
+ 10.08M
155
+ 10%
156
+ Simple Wikipedia4
157
+ Wikipedia (Simple English)
158
+ 1.52M
159
+ 14.66M
160
+ 15%
161
+ Switchboard Dialog Act Corpus (Stolcke et al., 2000)
162
+ Dialogue
163
+ 0.12M
164
+ 1.18M
165
+ 1%
166
+ Total
167
+
168
+ 9.96M
169
+ 98.04M
170
+ 100%
171
+ Table 1: The datasets we release for the STRICT and STRICT-SMALL tracks of the BabyLM Challenge. We present
172
+ the number of words in the training set of each corpus that we include. 1http://www.natcorp.ox.ac.uk
173
+ 2https:
174
+ //www.kaggle.com/datasets/edenbd/children-stories-text-corpus
175
+ 3https://dumps.wikimedia.org/
176
+ enwiki/20221220/
177
+ 4https://dumps.wikimedia.org/simplewiki/20221201/
178
+ 3
179
+ Tracks
180
+ This shared task includes three tracks: STRICT,
181
+ STRICT-SMALL, and LOOSE.
182
+ The STRICT and STRICT-SMALL tracks require
183
+ that submissions are trained exclusively on a fixed
184
+ dataset, which we provide. The main difference be-
185
+ tween these tracks is the size of the dataset (∼10M
186
+ words vs. ∼100M words). Both datasets con-
187
+ tain child-directed speech, transcribed speech from
188
+ multiple sources, children’s books, and Wikipedia,
189
+ among other datasets. The STRICT-SMALL dataset
190
+ is an approximately 10% uniform subsample of the
191
+ STRICT dataset. See §4 for a full description of the
192
+ fixed datasets. Winners will be determined based
193
+ on performance on the shared evaluation set.
194
+ The LOOSE track relaxes these restrictions. Sub-
195
+ missions must still be trained on a maximum of
196
+ 100M words, and will be tested on the shared eval-
197
+ uation set. However, they are permitted to use
198
+ unlimited non-linguistic data or text which differs
199
+ from the restricted shared task. Training on addi-
200
+ tional text is allowed without limits if that text is
201
+ generated by a model trained following the above
202
+ restrictions. For this track, winners will be selected
203
+ holistically based on evaluation performance, rel-
204
+ evance to the shared task goals, potential impact,
205
+ and novelty.
206
+ 4
207
+ Dataset
208
+ We distribute a developmentally plausible pretrain-
209
+ ing dataset inspired by the input to children.1 Sub-
210
+ missions must use only this training data to be con-
211
+ 1Clicking on the following link will download the dataset
212
+ (240MB zipped, 700MB unzipped): https://github.com/
213
+ babylm/babylm.github.io/raw/main/babylm_data.zip
214
+ sidered for the STRICT(-SMALL) tracks, but may
215
+ use different data for the LOOSE track. The dataset
216
+ has two key properties:
217
+ • Under 100M words: Children are exposed
218
+ to 2M-7M words per year (Gilkerson et al.,
219
+ 2017). Choosing the beginning of adolescence
220
+ (age 12) as a cutoff, the dataset should be
221
+ between 24M-84M words.
222
+ • Mostly transcribed speech: Most of the in-
223
+ put to children is spoken. Thus, we include a
224
+ higher proportion of transcribed speech in our
225
+ dataset.
226
+ The datasets we release are mixed domain, taken
227
+ from multiple sources. Table 1 summarizes the
228
+ composition of the datasets.
229
+ 5
230
+ Evaluation
231
+ We will distribute a shared evaluation pipeline
232
+ based in Google Colab. Colab provides access
233
+ to relatively small GPUs; this will allow users from
234
+ various research settings of varying resources to ef-
235
+ ficiently evaluate their submissions. Our evaluation
236
+ code will also be public, such that those wishing to
237
+ use their own computational resources may do so.
238
+ More details about the evaluation pipeline and the
239
+ set of tasks will be released subsequently.
240
+ The pipeline assumes all models can be loaded
241
+ and queried in HuggingFace’s transformers li-
242
+ brary (Wolf et al., 2020).2 Additionally, all mod-
243
+ els must be able to score a sequence—e.g., assign
244
+ 2While discouraged, participants whose models are not
245
+ compatible with the transformers library can still conduct
246
+ the necessary evaluation through their own pipeline.
247
+
248
+ a log-likelihood or pseudo log-likelihood (Wang
249
+ and Cho, 2019; Salazar et al., 2020)—and must
250
+ be able to be fine-tuned to perform classification
251
+ tasks. Models do not need to be able to generate se-
252
+ quences. Submissions must include model outputs
253
+ for each of the core evaluations in a format that we
254
+ specify in our evaluation pipeline.
255
+ We choose evaluations that represent the core
256
+ interests of this shared task, focusing on efficiency
257
+ and applied NLP, as well as cognitive science, lin-
258
+ guistics and language acquisition. Especially good
259
+ performance in one but not both of these areas may
260
+ be acknowledged with a special award.
261
+ 5.1
262
+ Baselines
263
+ We will also release a series of baseline models
264
+ with the evaluation pipeline. To train these, we sim-
265
+ ply take the hyperparameters from a series of estab-
266
+ lished large language models and train them from
267
+ scratch on our fixed datasets. We use hyperparame-
268
+ ters from OPT (decoder-only; Zhang et al., 2022),
269
+ RoBERTa (encoder-only; Liu et al., 2019), and T5
270
+ (encoder-decoder; Raffel et al., 2020). These are
271
+ not meant to be strong baselines, but rather to pro-
272
+ vide a naïve starting point for improving language
273
+ models for this domain.
274
+ 6
275
+ Submissions
276
+ What you Need to Submit
277
+ • A link where we can download the model
278
+ • A .zip of predictions (from our eval pipeline)
279
+ • A short description of the approaches taken
280
+ • If LOOSE track: a link where we can down-
281
+ load any additional data
282
+ Although scaled-down pretraining is more ac-
283
+ cessible to research groups with limited resources,
284
+ pretraining is still expensive from a computational,
285
+ energy, and financial perspective. To help groups
286
+ plan for total costs, we will release an estimate of
287
+ the resources required to pretrain on 10M words
288
+ and 100M words. For the LOOSE track, evaluation
289
+ of submissions may take into consideration compu-
290
+ tational efficiency as part of the holistic evaluation.
291
+ 7
292
+ FAQs
293
+ Can papers be submitted to multiple tracks?
294
+ Yes. For example, a single paper can describe mod-
295
+ els which are submitted separately to the STRICT
296
+ and STRICT-SMALL tracks.
297
+ Can I submit a paper about my work?
298
+ Yes,
299
+ we encourage all participants to submit their re-
300
+ ports, which will be published in the proceedings
301
+ of CoNLL. You may also describe any additional
302
+ experiments beyond those required for the shared
303
+ task evaluation.
304
+ Can I submit additional evaluation metrics?
305
+ Yes, if you wish to submit your own evaluation
306
+ metrics, along with model performance, alongside
307
+ our standardized evaluation results these can be
308
+ considered as part of the holistic evaluation in the
309
+ LOOSE track.
310
+ What training regimes are permitted?
311
+ For the
312
+ STRICT/STRICT-SMALL tracks, any kind of train-
313
+ ing objective/regime is permitted, as long as the
314
+ data restrictions are followed. Pretrained models
315
+ may not be used for any purpose such as reranking
316
+ or data augmentation.
317
+ We do however require for evaluation purposes
318
+ that the model provides a function to score a
319
+ sequence—e.g., log-likelihood for autoregressive
320
+ models or pseudo-log-likelihood for masked lan-
321
+ guage models—without the need for additional
322
+ fine-tuning.
323
+ Are there any limits on hyperparameters?
324
+ No.
325
+ In the LOOSE track, parameter efficiency and train-
326
+ ing efficiency may be considered along with other
327
+ factors in ranking submissions.
328
+ Are there any limits on the number of epochs?
329
+ No. We put no restrictions on the number of epochs,
330
+ for several reasons: First, from an engineering per-
331
+ spective, training LMs with SGD tends to require
332
+ multiple epochs at these scales to achieve peak per-
333
+ formance. Second, from a cognitive perspective,
334
+ humans have a memory of linguistic experience,
335
+ and can continue to access and learn from these
336
+ memories. Third, we try not to make a stand on
337
+ implementations to allow the most freedom for in-
338
+ novation.
339
+ 8
340
+ Organizing Committee
341
+ Leshem Choshen
342
+ Aaron Mueller
343
+ Ryan Cotterell
344
+ Alex Warstadt
345
+ Kundan Krishna
346
+ Ethan Wilcox
347
+ Tal Linzen
348
+ Adina Williams
349
+ Haokun Liu
350
+ Chengxu Zhuang
351
+
352
+ Questions? Feel free to contact us at the following
353
+ email addresses:
354
+ leshem.choshen@mail.huji.ac.il
355
+ haokunl@cs.unc.edu
356
+ amueller@jhu.edu
357
+ alexwarstadt@gmail.com
358
+ ewilcox@ethz.ch
359
+ chengxuz@mit.edu
360
+ References
361
+ Ahmed Abdelali, Francisco Guzman, Hassan Sajjad,
362
+ and Stephan Vogel. 2014.
363
+ The AMARA corpus:
364
+ Building parallel language resources for the educa-
365
+ tional domain.
366
+ In Proceedings of the Ninth Inter-
367
+ national Conference on Language Resources and
368
+ Evaluation (LREC’14), Reykjavik, Iceland. Euro-
369
+ pean Language Resources Association (ELRA).
370
+ Tom Brown, Benjamin Mann, Nick Ryder, Melanie
371
+ Subbiah,
372
+ Jared
373
+ D
374
+ Kaplan,
375
+ Prafulla
376
+ Dhariwal,
377
+ Arvind Neelakantan, Pranav Shyam, Girish Sastry,
378
+ Amanda Askell, Sandhini Agarwal, Ariel Herbert-
379
+ Voss, Gretchen Krueger, Tom Henighan, Rewon
380
+ Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu,
381
+ Clemens Winter, Chris Hesse, Mark Chen, Eric
382
+ Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess,
383
+ Jack Clark, Christopher Berner, Sam McCandlish,
384
+ Alec Radford, Ilya Sutskever, and Dario Amodei.
385
+ 2020. Language models are few-shot learners. In
386
+ Advances in Neural Information Processing Systems,
387
+ volume 33, pages 1877–1901. Curran Associates,
388
+ Inc.
389
+ Aakanksha Chowdhery, Sharan Narang, Jacob Devlin,
390
+ Maarten Bosma, Gaurav Mishra, Adam Roberts,
391
+ Paul Barham, Hyung Won Chung, Charles Sutton,
392
+ Sebastian Gehrmann, Parker Schuh, Kensen Shi,
393
+ Sasha Tsvyashchenko, Joshua Maynez, Abhishek
394
+ Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vin-
395
+ odkumar Prabhakaran, Emily Reif, Nan Du, Ben
396
+ Hutchinson, Reiner Pope, James Bradbury, Jacob
397
+ Austin, Michael Isard, Guy Gur-Ari, Pengcheng
398
+ Yin, Toju Duke, Anselm Levskaya, Sanjay Ghe-
399
+ mawat, Sunipa Dev, Henryk Michalewski, Xavier
400
+ Garcia, Vedant Misra, Kevin Robinson, Liam Fe-
401
+ dus, Denny Zhou, Daphne Ippolito, David Luan,
402
+ Hyeontaek Lim, Barret Zoph, Alexander Spiridonov,
403
+ Ryan Sepassi, David Dohan, Shivani Agrawal, Mark
404
+ Omernick, Andrew M. Dai, Thanumalayan Sankara-
405
+ narayana Pillai, Marie Pellat, Aitor Lewkowycz,
406
+ Erica Moreira, Rewon Child, Oleksandr Polozov,
407
+ Katherine Lee, Zongwei Zhou, Xuezhi Wang, Bren-
408
+ nan Saeta, Mark Diaz, Orhan Firat, Michele Catasta,
409
+ Jason Wei, Kathy Meier-Hellstern, Douglas Eck,
410
+ Jeff Dean, Slav Petrov, and Noah Fiedel. 2022.
411
+ PaLM: Scaling language modeling with pathways.
412
+ Computing Research Repository, arXiv:2204.02311.
413
+ Emmanuel Dupoux. 2018.
414
+ Cognitive science in the
415
+ era of artificial intelligence: A roadmap for reverse-
416
+ engineering the infant language-learner. Cognition,
417
+ 173:43–59.
418
+ Martin Gerlach and Francesc Font-Clos. 2018.
419
+ A
420
+ standardized Project Gutenberg corpus for statis-
421
+ tical analysis of natural language and quantita-
422
+ tive linguistics.
423
+ Computing Research Repository,
424
+ arXiv:1812.08092.
425
+ Jill Gilkerson, Jeffrey A Richards, Steven F Warren, Ju-
426
+ dith K Montgomery, Charles R Greenwood, D Kim-
427
+ brough Oller, John HL Hansen, and Terrance D
428
+ Paul. 2017.
429
+ Mapping the early language environ-
430
+ ment using all-day recordings and automated anal-
431
+ ysis. American journal of speech-language pathol-
432
+ ogy, 26(2):248–265.
433
+ Felix Hill, Antoine Bordes, Sumit Chopra, and Jason
434
+ Weston. 2016. The Goldilocks principle: Reading
435
+ children’s books with explicit memory representa-
436
+ tions.
437
+ Jordan Hoffmann, Sebastian Borgeaud, Arthur Men-
438
+ sch, Elena Buchatskaya, Trevor Cai, Eliza Ruther-
439
+ ford, Diego de Las Casas, Lisa Anne Hendricks,
440
+ Johannes Welbl, Aidan Clark, et al. 2022.
441
+ Train-
442
+ ing compute-optimal large language models. arXiv
443
+ preprint arXiv:2203.15556.
444
+ Frank Keller. 2010. Cognitively plausible models of
445
+ human language processing. In Proceedings of the
446
+ ACL 2010 Conference Short Papers, pages 60–67,
447
+ Uppsala, Sweden. Association for Computational
448
+ Linguistics.
449
+ Pierre Lison and Jörg Tiedemann. 2016.
450
+ OpenSub-
451
+ titles2016: Extracting large parallel corpora from
452
+ movie and TV subtitles. In Proceedings of the Tenth
453
+ International Conference on Language Resources
454
+ and Evaluation (LREC’16), pages 923–929, Por-
455
+ torož, Slovenia. European Language Resources As-
456
+ sociation (ELRA).
457
+ Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du,
458
+ Mandar Joshi, Danqi Chen, Omer Levy, Mike
459
+ Lewis, Luke Zettlemoyer, and Veselin Stoyanov.
460
+ 2019.
461
+ RoBERTa: A robustly optimized bert pre-
462
+ training approach. Computing Research Repository,
463
+ arXiv:1907.11692.
464
+ Brian MacWhinney. 2000. The CHILDES project: The
465
+ database, volume 2. Psychology Press.
466
+ Colin Raffel, Noam Shazeer, Adam Roberts, Kather-
467
+ ine Lee, Sharan Narang, Michael Matena, Yanqi
468
+ Zhou, Wei Li, and Peter J. Liu. 2020.
469
+ Exploring
470
+ the limits of transfer learning with a unified text-to-
471
+ text transformer. Journal of Machine Learning Re-
472
+ search, 21(140):1–67.
473
+ Julian Salazar, Davis Liang, Toan Q. Nguyen, and Ka-
474
+ trin Kirchhoff. 2020. Masked language model scor-
475
+ ing.
476
+ In Proceedings of the 58th Annual Meeting
477
+ of the Association for Computational Linguistics,
478
+ pages 2699–2712, Online. Association for Compu-
479
+ tational Linguistics.
480
+
481
+ Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth
482
+ Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Tay-
483
+ lor, Rachel Martin, Marie Meteer, and Carol Van
484
+ Ess-Dykema. 2000. Dialogue act modeling for au-
485
+ tomatic tagging and recognition of conversational
486
+ speech. Computational Linguistics, 26(3):339–371.
487
+ Alex Wang and Kyunghyun Cho. 2019.
488
+ BERT has
489
+ a mouth, and it must speak: BERT as a Markov
490
+ random field language model.
491
+ In Proceedings of
492
+ the Workshop on Methods for Optimizing and Eval-
493
+ uating Neural Language Generation, pages 30–36,
494
+ Minneapolis, Minnesota. Association for Computa-
495
+ tional Linguistics.
496
+ Alex Warstadt and Samuel R Bowman. 2022. What ar-
497
+ tificial neural networks can tell us about human lan-
498
+ guage acquisition. In Algebraic Structures in Natu-
499
+ ral Language, pages 17–60. CRC Press.
500
+ Thomas Wolf, Lysandre Debut, Victor Sanh, Julien
501
+ Chaumond, Clement Delangue, Anthony Moi, Pier-
502
+ ric Cistac, Tim Rault, Rémi Louf, Morgan Funtow-
503
+ icz, Joe Davison, Sam Shleifer, Patrick von Platen,
504
+ Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu,
505
+ Teven Le Scao, Sylvain Gugger, Mariama Drame,
506
+ Quentin Lhoest, and Alexander M. Rush. 2020.
507
+ Transformers: State-of-the-art natural language pro-
508
+ cessing. In Proceedings of the 2020 Conference on
509
+ Empirical Methods in Natural Language Processing:
510
+ System Demonstrations, pages 38–45, Online. Asso-
511
+ ciation for Computational Linguistics.
512
+ Susan Zhang, Stephen Roller, Naman Goyal, Mikel
513
+ Artetxe, Moya Chen, Shuohui Chen, Christopher De-
514
+ wan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mi-
515
+ haylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel
516
+ Simig, Punit Singh Koura, Anjali Sridhar, Tianlu
517
+ Wang, and Luke Zettlemoyer. 2022.
518
+ OPT: Open
519
+ pre-trained transformer language models. Comput-
520
+ ing Research Repository, arXiv:2205.01068.
521
+
CNFKT4oBgHgl3EQfXy51/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,304 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf,len=303
2
+ page_content='Call for Papers - The BabyLM Challenge: Sample-efficient pretraining on a developmentally plausible corpus https://babylm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
3
+ page_content='github.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
4
+ page_content='io/ Alex Warstadt ETH Zürich Leshem Choshen IBM Research Aaron Mueller Johns Hopkins University Ethan Wilcox ETH Zürich Adina Williams Meta AI Chengxu Zhuang MIT Abstract We present the call for papers for the BabyLM Challenge: Sample-efficient pretraining on a developmentally plausible corpus.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
5
+ page_content=' This shared task is intended for participants with an in- terest in small scale language modeling, hu- man language acquisition, low-resource NLP, and cognitive modeling.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
6
+ page_content=' In partnership with CoNLL and CMCL, we provide a platform for approaches to pretraining with a limited-size corpus sourced from data inspired by the input to children.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
7
+ page_content=' The task has three tracks, two of which restrict the training data to pre-released datasets of 10M and 100M words and are ded- icated to explorations of approaches such as architectural variations, self-supervised objec- tives, or curriculum learning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
8
+ page_content=' The final track only restricts the amount of text used, allowing innovation in the choice of the data, its domain, and even its modality (i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
9
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
10
+ page_content=', data from sources other than text is welcome).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
11
+ page_content=' We will release a shared evaluation pipeline which scores mod- els on a variety of benchmarks and tasks, in- cluding targeted syntactic evaluations and nat- ural language understanding.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
12
+ page_content=' 1 Motivation Huge efforts have been put into optimizing LM pre- training at massive scales in the last several years (Raffel et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
13
+ page_content=', 2020;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
14
+ page_content=' Brown et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
15
+ page_content=', 2020;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
16
+ page_content=' Chowdhery et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
17
+ page_content=', 2022;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
18
+ page_content=' Hoffmann et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
19
+ page_content=', 2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
20
+ page_content=' While grow- ing parameter counts often get the most attention, datasets have also grown by orders of magnitude.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
21
+ page_content=' These increasingly larger pretraining datasets are visualized, to scale, in Figure 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
22
+ page_content=' At the same time, there has been almost no progress in pretraining at smaller human-like data scales.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
23
+ page_content=' Focusing on scaled-down pretraining has several potential benefits: First, small-scale pretraining can be a sandbox for developing novel techniques that improve data efficiency.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
24
+ page_content=' These techniques have the potential to then scale up to larger datasets com- monly seen in applied NLP, and could be used Figure 1: Data Scale: Modern Language Models are trained on data multiple orders of magnitude larger than the amount available to a typical human child.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
25
+ page_content=' Image based off Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
26
+ page_content=' 1 from Warstadt and Bowman (2022) to enhance current approaches to modeling low- resource languages.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
27
+ page_content=' Second, improving our ability to train LMs on the same types and quantities of data that humans learn from will give us greater access to more plausible cognitive models of hu- mans and help us understand what allows humans to acquire language so efficiently (Keller, 2010;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
28
+ page_content=' Dupoux, 2018).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
29
+ page_content=' That is, even model failure can help in developing hypotheses about the differences between human and LM language learning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
30
+ page_content=' The goal of this shared task will be to incentivize researchers with an interest in pretraining and/or cognitive modeling to focus their efforts on opti- mizing pretraining given data limitations inspired by human development.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
31
+ page_content=' Additionally, we hope to democratize research on pretraining—which is typ- ically thought to be practical only for large industry groups—by drawing attention to open problems that can be addressed on a university budget.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
32
+ page_content=' 2 Key Dates January 2023: Training data released March 2023: Evaluation pipeline released July 15, 2023: Results due August 1, 2023: Paper submissions due Date TBA: Presentation at CoNLL arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
33
+ page_content='11796v1 [cs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
34
+ page_content='CL] 27 Jan 2023 200 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
35
+ page_content='4 Billion Trillion 30 3 <100 Billion Billion Million 13 y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
36
+ page_content='0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
37
+ page_content=' BERT RoBERTa GPT-3 Chinchilla Human (2018) (2019) (2020) (2022)# Words Dataset Domain STRICT-SMALL STRICT Proportion CHILDES (MacWhinney, 2000) Child-directed speech 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
38
+ page_content='44M 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
39
+ page_content='21M 5% British National Corpus (BNC),1 dialogue portion Dialogue 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
40
+ page_content='86M 8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
41
+ page_content='16M 8% Children’s Book Test (Hill et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
42
+ page_content=', 2016) Children’s books 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
43
+ page_content='57M 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
44
+ page_content='55M 6% Children’s Stories Text Corpus2 Children’s books 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
45
+ page_content='34M 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
46
+ page_content='22M 3% Standardized Project Gutenberg Corpus (Gerlach and Font-Clos, 2018) Written English 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
47
+ page_content='99M 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
48
+ page_content='46M 10% OpenSubtitles (Lison and Tiedemann, 2016) Movie subtitles 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
49
+ page_content='09M 31.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
50
+ page_content='28M 31% QCRI Educational Domain Corpus (QED;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
51
+ page_content=' Abdelali et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
52
+ page_content=', 2014) Educational video subtitles 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
53
+ page_content='04M 10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
54
+ page_content='24M 11% Wikipedia3 Wikipedia (English) 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
55
+ page_content='99M 10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
56
+ page_content='08M 10% Simple Wikipedia4 Wikipedia (Simple English) 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
57
+ page_content='52M 14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
58
+ page_content='66M 15% Switchboard Dialog Act Corpus (Stolcke et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
59
+ page_content=', 2000) Dialogue 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
60
+ page_content='12M 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
61
+ page_content='18M 1% Total – 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
62
+ page_content='96M 98.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
63
+ page_content='04M 100% Table 1: The datasets we release for the STRICT and STRICT-SMALL tracks of the BabyLM Challenge.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
64
+ page_content=' We present the number of words in the training set of each corpus that we include.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
65
+ page_content=' 1http://www.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
66
+ page_content='natcorp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
67
+ page_content='ox.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
68
+ page_content='ac.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
69
+ page_content='uk 2https: //www.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
70
+ page_content='kaggle.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
71
+ page_content='com/datasets/edenbd/children-stories-text-corpus 3https://dumps.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
72
+ page_content='wikimedia.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
73
+ page_content='org/ enwiki/20221220/ 4https://dumps.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
74
+ page_content='wikimedia.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
75
+ page_content='org/simplewiki/20221201/ 3 Tracks This shared task includes three tracks: STRICT, STRICT-SMALL, and LOOSE.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
76
+ page_content=' The STRICT and STRICT-SMALL tracks require that submissions are trained exclusively on a fixed dataset, which we provide.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
77
+ page_content=' The main difference be- tween these tracks is the size of the dataset (∼10M words vs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
78
+ page_content=' ∼100M words).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
79
+ page_content=' Both datasets con- tain child-directed speech, transcribed speech from multiple sources, children’s books, and Wikipedia, among other datasets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
80
+ page_content=' The STRICT-SMALL dataset is an approximately 10% uniform subsample of the STRICT dataset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
81
+ page_content=' See §4 for a full description of the fixed datasets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
82
+ page_content=' Winners will be determined based on performance on the shared evaluation set.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
83
+ page_content=' The LOOSE track relaxes these restrictions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
84
+ page_content=' Sub- missions must still be trained on a maximum of 100M words, and will be tested on the shared eval- uation set.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
85
+ page_content=' However, they are permitted to use unlimited non-linguistic data or text which differs from the restricted shared task.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
86
+ page_content=' Training on addi- tional text is allowed without limits if that text is generated by a model trained following the above restrictions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
87
+ page_content=' For this track, winners will be selected holistically based on evaluation performance, rel- evance to the shared task goals, potential impact, and novelty.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
88
+ page_content=' 4 Dataset We distribute a developmentally plausible pretrain- ing dataset inspired by the input to children.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
89
+ page_content='1 Sub- missions must use only this training data to be con- 1Clicking on the following link will download the dataset (240MB zipped, 700MB unzipped): https://github.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
90
+ page_content='com/ babylm/babylm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
91
+ page_content='github.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
92
+ page_content='io/raw/main/babylm_data.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
93
+ page_content='zip sidered for the STRICT(-SMALL) tracks, but may use different data for the LOOSE track.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
94
+ page_content=' The dataset has two key properties: Under 100M words: Children are exposed to 2M-7M words per year (Gilkerson et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
95
+ page_content=', 2017).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
96
+ page_content=' Choosing the beginning of adolescence (age 12) as a cutoff, the dataset should be between 24M-84M words.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
97
+ page_content=' Mostly transcribed speech: Most of the in- put to children is spoken.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
98
+ page_content=' Thus, we include a higher proportion of transcribed speech in our dataset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
99
+ page_content=' The datasets we release are mixed domain, taken from multiple sources.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
100
+ page_content=' Table 1 summarizes the composition of the datasets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
101
+ page_content=' 5 Evaluation We will distribute a shared evaluation pipeline based in Google Colab.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
102
+ page_content=' Colab provides access to relatively small GPUs;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
103
+ page_content=' this will allow users from various research settings of varying resources to ef- ficiently evaluate their submissions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
104
+ page_content=' Our evaluation code will also be public, such that those wishing to use their own computational resources may do so.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
105
+ page_content=' More details about the evaluation pipeline and the set of tasks will be released subsequently.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
106
+ page_content=' The pipeline assumes all models can be loaded and queried in HuggingFace’s transformers li- brary (Wolf et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
107
+ page_content=', 2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
108
+ page_content='2 Additionally, all mod- els must be able to score a sequence—e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
109
+ page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
110
+ page_content=', assign 2While discouraged, participants whose models are not compatible with the transformers library can still conduct the necessary evaluation through their own pipeline.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
111
+ page_content=' a log-likelihood or pseudo log-likelihood (Wang and Cho, 2019;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
112
+ page_content=' Salazar et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
113
+ page_content=', 2020)—and must be able to be fine-tuned to perform classification tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
114
+ page_content=' Models do not need to be able to generate se- quences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
115
+ page_content=' Submissions must include model outputs for each of the core evaluations in a format that we specify in our evaluation pipeline.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
116
+ page_content=' We choose evaluations that represent the core interests of this shared task, focusing on efficiency and applied NLP, as well as cognitive science, lin- guistics and language acquisition.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
117
+ page_content=' Especially good performance in one but not both of these areas may be acknowledged with a special award.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
118
+ page_content=' 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
119
+ page_content='1 Baselines We will also release a series of baseline models with the evaluation pipeline.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
120
+ page_content=' To train these, we sim- ply take the hyperparameters from a series of estab- lished large language models and train them from scratch on our fixed datasets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
121
+ page_content=' We use hyperparame- ters from OPT (decoder-only;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
122
+ page_content=' Zhang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
123
+ page_content=', 2022), RoBERTa (encoder-only;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
124
+ page_content=' Liu et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
125
+ page_content=', 2019), and T5 (encoder-decoder;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
126
+ page_content=' Raffel et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
127
+ page_content=', 2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
128
+ page_content=' These are not meant to be strong baselines, but rather to pro- vide a naïve starting point for improving language models for this domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
129
+ page_content=' 6 Submissions What you Need to Submit A link where we can download the model A .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
130
+ page_content='zip of predictions (from our eval pipeline) A short description of the approaches taken If LOOSE track: a link where we can down- load any additional data Although scaled-down pretraining is more ac- cessible to research groups with limited resources, pretraining is still expensive from a computational, energy, and financial perspective.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
131
+ page_content=' To help groups plan for total costs, we will release an estimate of the resources required to pretrain on 10M words and 100M words.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
132
+ page_content=' For the LOOSE track, evaluation of submissions may take into consideration compu- tational efficiency as part of the holistic evaluation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
133
+ page_content=' 7 FAQs Can papers be submitted to multiple tracks?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
134
+ page_content=' Yes.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
135
+ page_content=' For example, a single paper can describe mod- els which are submitted separately to the STRICT and STRICT-SMALL tracks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
136
+ page_content=' Can I submit a paper about my work?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
137
+ page_content=' Yes, we encourage all participants to submit their re- ports, which will be published in the proceedings of CoNLL.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
138
+ page_content=' You may also describe any additional experiments beyond those required for the shared task evaluation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
139
+ page_content=' Can I submit additional evaluation metrics?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
140
+ page_content=' Yes, if you wish to submit your own evaluation metrics, along with model performance, alongside our standardized evaluation results these can be considered as part of the holistic evaluation in the LOOSE track.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
141
+ page_content=' What training regimes are permitted?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
142
+ page_content=' For the STRICT/STRICT-SMALL tracks, any kind of train- ing objective/regime is permitted, as long as the data restrictions are followed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
143
+ page_content=' Pretrained models may not be used for any purpose such as reranking or data augmentation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
144
+ page_content=' We do however require for evaluation purposes that the model provides a function to score a sequence—e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
145
+ page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
146
+ page_content=', log-likelihood for autoregressive models or pseudo-log-likelihood for masked lan- guage models—without the need for additional fine-tuning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
147
+ page_content=' Are there any limits on hyperparameters?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
148
+ page_content=' No.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
149
+ page_content=' In the LOOSE track, parameter efficiency and train- ing efficiency may be considered along with other factors in ranking submissions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
150
+ page_content=' Are there any limits on the number of epochs?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
151
+ page_content=' No.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
152
+ page_content=' We put no restrictions on the number of epochs, for several reasons: First, from an engineering per- spective, training LMs with SGD tends to require multiple epochs at these scales to achieve peak per- formance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
153
+ page_content=' Second, from a cognitive perspective, humans have a memory of linguistic experience, and can continue to access and learn from these memories.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
154
+ page_content=' Third, we try not to make a stand on implementations to allow the most freedom for in- novation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
155
+ page_content=' 8 Organizing Committee Leshem Choshen Aaron Mueller Ryan Cotterell Alex Warstadt Kundan Krishna Ethan Wilcox Tal Linzen Adina Williams Haokun Liu Chengxu Zhuang Questions?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
156
+ page_content=' Feel free to contact us at the following email addresses: leshem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
157
+ page_content='choshen@mail.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
158
+ page_content='huji.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
159
+ page_content='ac.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
160
+ page_content='il haokunl@cs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
161
+ page_content='unc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
162
+ page_content='edu amueller@jhu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
163
+ page_content='edu alexwarstadt@gmail.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
164
+ page_content='com ewilcox@ethz.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
165
+ page_content='ch chengxuz@mit.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
166
+ page_content='edu References Ahmed Abdelali, Francisco Guzman, Hassan Sajjad, and Stephan Vogel.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
167
+ page_content=' 2014.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
168
+ page_content=' The AMARA corpus: Building parallel language resources for the educa- tional domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
169
+ page_content=' In Proceedings of the Ninth Inter- national Conference on Language Resources and Evaluation (LREC’14), Reykjavik, Iceland.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
170
+ page_content=' Euro- pean Language Resources Association (ELRA).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
171
+ page_content=' Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert- Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
172
+ page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
173
+ page_content=' Language models are few-shot learners.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
174
+ page_content=' In Advances in Neural Information Processing Systems, volume 33, pages 1877–1901.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
175
+ page_content=' Curran Associates, Inc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
176
+ page_content=' Aakanksha Chowdhery,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
177
+ page_content=' Sharan Narang,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
178
+ page_content=' Jacob Devlin,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
179
+ page_content=' Maarten Bosma,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
180
+ page_content=' Gaurav Mishra,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
181
+ page_content=' Adam Roberts,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
182
+ page_content=' Paul Barham,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
183
+ page_content=' Hyung Won Chung,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
184
+ page_content=' Charles Sutton,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
185
+ page_content=' Sebastian Gehrmann,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
186
+ page_content=' Parker Schuh,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
187
+ page_content=' Kensen Shi,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
188
+ page_content=' Sasha Tsvyashchenko,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
189
+ page_content=' Joshua Maynez,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
190
+ page_content=' Abhishek Rao,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
191
+ page_content=' Parker Barnes,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
192
+ page_content=' Yi Tay,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
193
+ page_content=' Noam Shazeer,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
194
+ page_content=' Vin- odkumar Prabhakaran,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
195
+ page_content=' Emily Reif,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
196
+ page_content=' Nan Du,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
197
+ page_content=' Ben Hutchinson,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
198
+ page_content=' Reiner Pope,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
199
+ page_content=' James Bradbury,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
200
+ page_content=' Jacob Austin,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
201
+ page_content=' Michael Isard,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
202
+ page_content=' Guy Gur-Ari,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
203
+ page_content=' Pengcheng Yin,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
204
+ page_content=' Toju Duke,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
205
+ page_content=' Anselm Levskaya,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
206
+ page_content=' Sanjay Ghe- mawat,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
207
+ page_content=' Sunipa Dev,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
208
+ page_content=' Henryk Michalewski,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
209
+ page_content=' Xavier Garcia,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
210
+ page_content=' Vedant Misra,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
211
+ page_content=' Kevin Robinson,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
212
+ page_content=' Liam Fe- dus,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
213
+ page_content=' Denny Zhou,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
214
+ page_content=' Daphne Ippolito,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
215
+ page_content=' David Luan,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
216
+ page_content=' Hyeontaek Lim,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
217
+ page_content=' Barret Zoph,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
218
+ page_content=' Alexander Spiridonov,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
219
+ page_content=' Ryan Sepassi,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
220
+ page_content=' David Dohan,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
221
+ page_content=' Shivani Agrawal,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
222
+ page_content=' Mark Omernick,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
223
+ page_content=' Andrew M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
224
+ page_content=' Dai, Thanumalayan Sankara- narayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Bren- nan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
225
+ page_content=' 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
226
+ page_content=' PaLM: Scaling language modeling with pathways.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
227
+ page_content=' Computing Research Repository, arXiv:2204.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
228
+ page_content='02311.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
229
+ page_content=' Emmanuel Dupoux.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
230
+ page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
231
+ page_content=' Cognitive science in the era of artificial intelligence: A roadmap for reverse- engineering the infant language-learner.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
232
+ page_content=' Cognition, 173:43–59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
233
+ page_content=' Martin Gerlach and Francesc Font-Clos.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
234
+ page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
235
+ page_content=' A standardized Project Gutenberg corpus for statis- tical analysis of natural language and quantita- tive linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
236
+ page_content=' Computing Research Repository, arXiv:1812.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
237
+ page_content='08092.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
238
+ page_content=' Jill Gilkerson, Jeffrey A Richards, Steven F Warren, Ju- dith K Montgomery, Charles R Greenwood, D Kim- brough Oller, John HL Hansen, and Terrance D Paul.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
239
+ page_content=' 2017.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
240
+ page_content=' Mapping the early language environ- ment using all-day recordings and automated anal- ysis.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
241
+ page_content=' American journal of speech-language pathol- ogy, 26(2):248–265.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
242
+ page_content=' Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
243
+ page_content=' 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
244
+ page_content=' The Goldilocks principle: Reading children’s books with explicit memory representa- tions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
245
+ page_content=' Jordan Hoffmann, Sebastian Borgeaud, Arthur Men- sch, Elena Buchatskaya, Trevor Cai, Eliza Ruther- ford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
246
+ page_content=' 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
247
+ page_content=' Train- ing compute-optimal large language models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
248
+ page_content=' arXiv preprint arXiv:2203.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
249
+ page_content='15556.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
250
+ page_content=' Frank Keller.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
251
+ page_content=' 2010.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
252
+ page_content=' Cognitively plausible models of human language processing.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
253
+ page_content=' In Proceedings of the ACL 2010 Conference Short Papers, pages 60–67, Uppsala, Sweden.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
254
+ page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
255
+ page_content=' Pierre Lison and Jörg Tiedemann.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
256
+ page_content=' 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
257
+ page_content=' OpenSub- titles2016: Extracting large parallel corpora from movie and TV subtitles.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
258
+ page_content=' In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 923–929, Por- torož, Slovenia.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
259
+ page_content=' European Language Resources As- sociation (ELRA).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
260
+ page_content=' Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
261
+ page_content=' 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
262
+ page_content=' RoBERTa: A robustly optimized bert pre- training approach.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
263
+ page_content=' Computing Research Repository, arXiv:1907.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
264
+ page_content='11692.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
265
+ page_content=' Brian MacWhinney.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
266
+ page_content=' 2000.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
267
+ page_content=' The CHILDES project: The database, volume 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
268
+ page_content=' Psychology Press.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
269
+ page_content=' Colin Raffel, Noam Shazeer, Adam Roberts, Kather- ine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
270
+ page_content=' Liu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
271
+ page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
272
+ page_content=' Exploring the limits of transfer learning with a unified text-to- text transformer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
273
+ page_content=' Journal of Machine Learning Re- search, 21(140):1–67.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
274
+ page_content=' Julian Salazar, Davis Liang, Toan Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
275
+ page_content=' Nguyen, and Ka- trin Kirchhoff.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
276
+ page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
277
+ page_content=' Masked language model scor- ing.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
278
+ page_content=' In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699–2712, Online.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
279
+ page_content=' Association for Compu- tational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
280
+ page_content=' Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Tay- lor, Rachel Martin, Marie Meteer, and Carol Van Ess-Dykema.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
281
+ page_content=' 2000.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
282
+ page_content=' Dialogue act modeling for au- tomatic tagging and recognition of conversational speech.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
283
+ page_content=' Computational Linguistics, 26(3):339–371.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
284
+ page_content=' Alex Wang and Kyunghyun Cho.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
285
+ page_content=' 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
286
+ page_content=' BERT has a mouth, and it must speak: BERT as a Markov random field language model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
287
+ page_content=' In Proceedings of the Workshop on Methods for Optimizing and Eval- uating Neural Language Generation, pages 30–36, Minneapolis, Minnesota.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
288
+ page_content=' Association for Computa- tional Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
289
+ page_content=' Alex Warstadt and Samuel R Bowman.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
290
+ page_content=' 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
291
+ page_content=' What ar- tificial neural networks can tell us about human lan- guage acquisition.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
292
+ page_content=' In Algebraic Structures in Natu- ral Language, pages 17–60.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
293
+ page_content=' CRC Press.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
294
+ page_content=' Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier- ric Cistac, Tim Rault, Rémi Louf, Morgan Funtow- icz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
295
+ page_content=' Rush.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
296
+ page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
297
+ page_content=' Transformers: State-of-the-art natural language pro- cessing.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
298
+ page_content=' In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
299
+ page_content=' Asso- ciation for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
300
+ page_content=' Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher De- wan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mi- haylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
301
+ page_content=' 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
302
+ page_content=' OPT: Open pre-trained transformer language models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
303
+ page_content=' Comput- ing Research Repository, arXiv:2205.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
304
+ page_content='01068.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/CNFKT4oBgHgl3EQfXy51/content/2301.11796v1.pdf'}
DNAyT4oBgHgl3EQf4frj/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e0432801c5bda9b00669fd7077c16a2b678dd9b29bce090783c85099224c7ea3
3
+ size 2359341
F9AyT4oBgHgl3EQf5Pp6/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:df9783bd87fd5b618f679de22e468c1a22b0c2e1cb3bb69cb5ccb8157442f50b
3
+ size 4063277
G9E0T4oBgHgl3EQfhgFS/content/tmp_files/2301.02432v1.pdf.txt ADDED
@@ -0,0 +1,1388 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Myths and Legends in High-Performance
2
+ Computing
3
+ arXiv preprints
4
+ ©The Author(s) 2023
5
+ Reprints and permission:
6
+ sagepub.co.uk/journalsPermissions.nav
7
+ DOI: 10.1177/ToBeAssigned
8
+ www.sagepub.com/
9
+ SAGE
10
+ Satoshi Matsuoka1, Jens Domke1, Mohamed Wahib1, Aleksandr Drozd1, Torsten Hoefler2
11
+ Abstract
12
+ In this humorous and thought provoking article, we discuss certain myths and legends that are folklore among members
13
+ of the high-performance computing community. We collected those myths from conversations at conferences and
14
+ meetings, product advertisements, papers, and other communications such as tweets, blogs, and news articles within
15
+ (and beyond) our community. We believe they represent the zeitgeist of the current era of massive change, driven by the
16
+ end of many scaling laws such as Dennard scaling and Moore’s law. While some laws end, new directions open up, such
17
+ as algorithmic scaling or novel architecture research. However, these myths are rarely based on scientific facts but often
18
+ on some evidence or argumentation. In fact, we believe that this is the very reason for the existence of many myths
19
+ and why they cannot be answered clearly. While it feels like there should be clear answers for each, some may remain
20
+ endless philosophical debates such as the question whether Beethoven was better than Mozart. We would like to see
21
+ our collection of myths as a discussion of possible new directions for research and industry investment.
22
+ Keywords
23
+ Quantum; zettascale; deep learning; clouds; HPC myths
24
+ This manuscript is intended for the “CCDSC Special Issue”.
25
+ Introduction
26
+ Any human society has their myths and legends—this also
27
+ applies to the high-performance computing (HPC) community.
28
+ HPC drives the largest and most powerful computers and
29
+ latest computing and acceleration technologies forward. One
30
+ may think that it’s scientific reasoning all the way down in
31
+ such an advanced field. Yet, we find many persistent myths
32
+ revolving around trends of the moment.
33
+ Since it’s late 2022, we started our analysis by asking the
34
+ all-knowing intelligence ChatGPT “Create myths or legends
35
+ in high performance computing”. In a HAL 9000 manner, it
36
+ refused to make up something for us: “I’m sorry [Dave], but
37
+ as an AI language model, I am not programmed to generate or
38
+ share myths or legends. My primary function is to assist users
39
+ with information and general knowledge, and I do not have
40
+ the ability to create or share fictional content.”. So, even the
41
+ smartest of internet parrots (Bender et al. 2021) that was itself
42
+ created with massive high-performace computation running
43
+ on a large accelerator system still has a long way to go. Thus,
44
+ we fall back to reasoning among the authors of this work.
45
+ We discuss 12 of today’s HPC myths, a number customary
46
+ in our community, similar to a panel statement where we
47
+ debate supporting and contradicting facts with a healthy
48
+ exaggeration in one of those directions. We attempt to neither
49
+ judge nor prove folklore right or wrong but instead try to
50
+ stipulate an intensive discussion in the community that drives
51
+ our future thinking.
52
+ Myth 1: Quantum Computing Will Take Over
53
+ HPC!
54
+ Numerous articles are hyping the quantum computing
55
+ revolution affecting nearly all aspects of life ranging from
56
+ quantum artificial intelligence to even quantum gaming.
57
+ The whole IT industry is following the quantum trend
58
+ and conceives quickly growing expectations. The actual
59
+ development of quantum technologies, algorithms, and use-
60
+ cases is on a very different time-scale. Most practitioners
61
+ would not expect quantum computers to outperform classical
62
+ computers within the next decade. Yet, we have constantly
63
+ been surprised by advances in device scaling as well as, more
64
+ recently, artificial intelligence. Thus, the fear of missing out
65
+ on getting rich is driving the industry to heavily invest into
66
+ quantum technologies pushing the technology forward.
67
+ With all this investment, it seems reasonable to expect that
68
+ quantum computation, which promises to deliver exponential
69
+ speedups, will replace high-performance computation as
70
+ we know it today with its meager linear speedup through
71
+ parallelism. Yet, the nature of quantum computation poses
72
+ some severe limitations: First, reading unstructured data into
73
+ a quantum state seems very challenging. Reasonable future
74
+ quantum computer designs can read in the order of Gigabit/s
75
+ while modern single-chip processors are already achieving
76
+ 1RIKEN Center for Computational Science, Japan
77
+ 2Eidgen¨ossische Technische Hochschule Z¨urich, Switzerland
78
+ Corresponding author:
79
+ Torsten
80
+ Hoefler,
81
+ ETH
82
+ Z¨urich,
83
+ Inst.
84
+ f.
85
+ Hochleistungsrechnersyst.,
86
+ Universit¨atstrasse 6, 8092 Z¨urich, Switzerland
87
+ Prepared using sagej.cls [Version: 2017/01/17 v1.20]
88
+ arXiv:2301.02432v1 [cs.DC] 6 Jan 2023
89
+
90
+ 2
91
+ arXiv preprints
92
+ Terabit/s—many orders of magnitude more (Hoefler et al.
93
+ 2023).
94
+ Furthermore, once a quantum state is constructed, it can
95
+ often be “used” only once because measurements destroy
96
+ superposition. A second limitation stems from the lack of
97
+ algorithms with high speedups. Most algorithms achieve
98
+ quadratic speedups for a wide range of use-cases using
99
+ amplitude amplification at their core. While this technique is
100
+ extremely versatile and can search any unstructured quantum
101
+ state (cf. Grover’s algorithm), its limited speedup is unlikely
102
+ to make it practical for quantum computers that may be
103
+ constructed in the next decades (Hoefler et al. 2023).
104
+ Thus, it seems unlikely that quantum computation is going
105
+ to replace a significant fraction of traditional HPC. It is more
106
+ likely that it will start as quantum acceleration with a small set
107
+ of use-cases that may grow in the future. To determine which
108
+ use-cases can realistically benefit from quantum acceleration,
109
+ resource estimation techniques (Beverland et al. 2022)
110
+ become crucial. But unlikely does not mean impossible—
111
+ we believe that now is the right time to begin a discussion
112
+ about the role of quantum computation in HPC. Furthermore,
113
+ it is crucial to guide the resources we invest into the right
114
+ directions.
115
+ We close with these questions. . .
116
+ x When will quantum computing be commercially
117
+ profitable? y What will be the first useful algorithm?
118
+ z What will be the next break-through area enabled by a
119
+ new quantum algorithm?
120
+ Myth 2: Everything Will Be Deep Learning!
121
+ Simultaneously with the quantum hype, we are in the midst
122
+ of the deep learning revolution. Indeed, in recent years
123
+ there has been a plethora of papers replacing traditional
124
+ simulation methods, or whole computational kernels with
125
+ data-driven models. Most of those employ deep neural
126
+ network architectures. Impressive results fire up expectations
127
+ equally high to the quantum world. Data-driven weather and
128
+ climate predictions apparently beat the best models (Pathak
129
+ et al. 2022; Bi et al. 2022) and output data can be compressed
130
+ by three orders of magnitude (Huang and Hoefler 2022).
131
+ Similar successes are touted in literally any application area.
132
+ There is no doubt that deep learning models can learn to
133
+ approximate complex functions used in scientific simulations
134
+ in a specific input domain. The issue is, as always, the trade-
135
+ offs: between speed on one hand, and accuracy on the other—
136
+ and we have to be very careful with these comparisons. In fact,
137
+ any result can be skewed into any of the extremes (Hoefler
138
+ 2022).
139
+ Sometimes even very simple models (and they have to be
140
+ simple to be compute-performance competitive) such as multi-
141
+ layer perceptrons (MLPs) can work well enough in place
142
+ of exact mathematical expression, e.g., Rasp et al. (2018);
143
+ Brenowitz and Bretherton (2018). One wonders sometimes
144
+ whether the latter could have been simplified in the first
145
+ place. A possible explanation is that neural nets, rather than
146
+ learning to approximate a given function in some abstract
147
+ sense, learn to decompose the input space into polyhedra
148
+ with corresponding simple mappings (Aytekin 2022). In other
149
+ words, neural nets can exploit the fact that typical input values
150
+ in many tasks are concentrated in particular ranges, which,
151
+ in turn, raises concerns about accuracy guarantees for out-of-
152
+ distribution inputs, and a possibility of some sort of hybrid /
153
+ fall-back mechanism.
154
+ An independent question is whether the architectures used
155
+ for machine learning tasks, like classification, are a good
156
+ match to serve as surrogate models in the first place? A new
157
+ line of research is addressing this by using neural architecture
158
+ search for such models (Kasim et al. 2021). In an extreme case,
159
+ the objective is to find a purely symbolic (and thus hopefully
160
+ more robust to out-of-distribution inputs) formulation for
161
+ cases where an exact mathematical expression for the problem
162
+ is not a-priori known (Liu and Tegmark 2021). Uncertainty
163
+ quantification and explainability are also two main aspects of
164
+ high importance in the scientific domain where DL is lacking
165
+ (due to its black-box optimization nature).
166
+ Overall the jury is still out as to which extent surrogate
167
+ models can replace first-principles simulations. However,
168
+ one thing is clear: there is a whole spectrum of simulation
169
+ tasks (Lavin et al. 2021)—ranging from ones where exact
170
+ mathematical expressions are not available in the first place
171
+ (e.g., contribution of specific vegetation to weather dynamics)
172
+ and learning it from data could not only be more efficient
173
+ but also more accurate; to those where utmost accuracy and
174
+ precision guarantees are required and can only be provided
175
+ by specialized error-controlling numerical methods.
176
+ We close with these questions. . .
177
+ x Will ML models replace or just augment traditional
178
+ simulations? y Where will ML models fail to deliver?
179
+ z How can we classify (pieces of) an application as ML-
180
+ acceleratable or not?
181
+ Myth 3: Extreme Specialization as Seen in
182
+ Smartphones Will Push Supercomputers
183
+ Beyond Moore’s Law!
184
+ AI, like Stable Diffusion, is now in the palm of everyone’s
185
+ hand. These modern smartphones typically are driven by a
186
+ System on Chip (SoC) that consists of a plethora of special
187
+ function units (SFUs) and/or special purpose processors that
188
+ accelerate various aspects of smartphone workloads. The main
189
+ purpose of such a composition is to achieve low power for
190
+ longer battery life while maintaining acceptable performance.
191
+ The success of GPUs, growing demands for lower power and
192
+ highest performance, and the end of Moore’s law created
193
+ a myth that future supercomputer architectures will be just
194
+ like smartphones in that there will be multitudes of hardware
195
+ customization per each facet of the entire workload.
196
+ However, such a claim misses the point in the analogy, and
197
+ entirely ignores multiple drawbacks of such an approach as
198
+ described below. In fact, the only successful “accelerator”
199
+ in the recent history of HPC is a GPU. The primary
200
+ reason for its success is high memory bandwidth, a feature
201
+ known since the vector supercomputer days, which is now
202
+ adopted by mainstream CPUs such as Fujitsu A64FX and
203
+ Intel Sapphire Rapids. The reason for the acceleration
204
+ is primarily that the majority of the HPC workloads are
205
+ memory bandwidth bound (Domke et al. 2021). Thus, modern
206
+ Prepared using sagej.cls
207
+
208
+ Matsuoka, Domke, Wahib, Drozd, Hoefler
209
+ 3
210
+ reincarnations of vector processors, such as vector units and
211
+ fast memory with HBM/GDDR variants, have been sufficient
212
+ to accelerate such workloads beyond CPUs with slower
213
+ DDR memory (Matsuoka 2008). So, to claim that multitudes
214
+ of special accelerators will constitute a supercomputers is
215
+ stretching the success of GPUs somewhat unfoundedly.
216
+ In fact, there are mainly three reasons why the plethora
217
+ of customized accelerated hardware approach would fail.
218
+ The first is the most important, in that acceleration via SoC
219
+ integration of various SFU is largely to enable strong scaling
220
+ at a compute node level, and will be subject to the limitations
221
+ of the Amdahl’s law, i.e., reducing the time to solution, the
222
+ potential speedup is bound by the ratio of accelerated and
223
+ non-accelerable fractions of the algorithm, which quickly
224
+ limits the speedup. Modern supercomputing is rather driven
225
+ by weak scaling as explained by Gustafson (1988), where the
226
+ speedup is based on how well the parallelizable or accelerable
227
+ fraction can be scaled on many nodes. This is often achieved
228
+ by linearly increasing the overall workload and maintaining
229
+ a constant amount of work per node, so the time to solution
230
+ remains constant but performance gain is proportional to
231
+ the number of nodes in an ideal case. This was exactly how
232
+ massive performance gain was obtained, despite skepticisms
233
+ from the then experts, towards massively parallel computing,
234
+ culminating in the first awarding of the Gordon Bell prize in
235
+ 1987 (Bell et al. 2017).
236
+ Combination of strong and weak scaling has been
237
+ instrumental in utilizing massive parallelism and performance
238
+ speedup in modern supercomputers such as Frontier and
239
+ Fugaku, but the contribution of the latter has been greater
240
+ in absolute speedup terms*. Now, weak scaling to large
241
+ number of nodes require that the workload can be subdivided
242
+ to achieve extremely good load balancing, i.e., (amount of
243
+ work) / (processing capability) is uniform among all nodes.
244
+ For homogeneous systems, if the workload domain is easily
245
+ compostable, then simple uniform partitioning will suffice,
246
+ and multitudes of studies have been conducted to achieve
247
+ proper domain decomposition for more complex algorithms.
248
+ Such load balancing work can be readily be applied even
249
+ for nodes that are composed of heterogeneous elements,
250
+ provided that (a) the architecture of the nodes are largely
251
+ uniform (homogeneous) across the entire machine, and (b)
252
+ during execution, the codes will be running simultaneous on
253
+ one of the processors within the node, all at the same time
254
+ within the machine. Practically all successful ‘accelerated’
255
+ supercomputers and their applications, e.g., GPU machines
256
+ such as Frontier, follow this pattern.
257
+ However, once the nodes would be composed of plethora
258
+ of customized hardware, and expected to be utilized in a
259
+ more random, heterogeneous fashion as in a smartphone, load
260
+ balancing becomes extremely difficult, and thus weak scaling
261
+ speedup will flatten quickly, especially in a large parallel
262
+ system. There have been efforts to alleviate this by creating
263
+ a task graph of the workload and conduct dynamic load
264
+ balancing, but have not really achieved success for very large
265
+ systems, let alone for numerous heterogeneous accelerators.
266
+ This is why, even for GPU-based machines, not only
267
+ the node architectures are homogeneous, but also, in any
268
+ given workload only GPUs or CPUs are used dominantly,
269
+ but not typically both. Contrastingly, that large parallel
270
+ program decomposed into a smaller task/dataflow graph and
271
+ executed on-demand basis heterogeneously on a plethora of
272
+ accelerators is only largely beneficial for small programs on
273
+ a small machine, but not for HPC where parallelism will
274
+ continue to increase to exploit weak scaling
275
+ The second reason is the increasing difficulty of dark
276
+ silicon being available in the system to be utilized for
277
+ heterogeneously specialized hardware, for cost reasons. In the
278
+ past, dark silicon was projected to be abundant with reduced
279
+ lithography, thus justifying the “plethora of accelerators” view,
280
+ as they were available for very low cost. However, with the
281
+ slowing down of Moore’s law, coupled with high cost of
282
+ manufacturing due to more advanced fab technologies such
283
+ as EUV, transistor cost over time is flattening, or may even
284
+ increase. Thus, the chip cost will become largely proportional
285
+ to the number of transistors irrespective of the lithography, so
286
+ every transistor has to contribute to the overall performance
287
+ improvements in a major fashion, turning dark silicon into
288
+ expensive unused silicon.
289
+ For smartphones, the major cost of the phone is not the
290
+ SoC but rather in the peripherals such as screen, camera, flash
291
+ memory, etc., and the battery life is premium in the cost metric
292
+ so extra cost incurred by dark silicon may be tolerable. For
293
+ supercomputers, however, the major cost of the machine is the
294
+ processors themselves, dominating over 50% of the overall
295
+ CapEx. So unless the acceleration could benefit some major
296
+ proportion of the workload, dark silicon would become an
297
+ intolerable waste. That is why, over generations, accelerators
298
+ such as GPUs tend to become more general purpose to
299
+ cover an increasing proportion of the workload, ultimately
300
+ becoming general purpose as the CPUs (or, GPGPUs).
301
+ The third reason is software and productivity. Unless the
302
+ accelerator usage is extremely easy, e.g., hidden under a set
303
+ of very simple APIs, expecting the programmers to adopt
304
+ an arcane programming model is not viable. In fact, this is
305
+ more serious for HPCs where the market for applications
306
+ is much smaller than major commodity ecosystems such
307
+ as smartphones, with a less performance-conscious but
308
+ extremely large market. Thus, for example, a large consumer-
309
+ oriented IT company such as Apple can afford to replace a
310
+ part of its API for a phone with hardware because it will sell
311
+ more than 100 million iPhones, but not for supercomputers
312
+ that have a much narrower market and thus do not warrant
313
+ such investment.
314
+ We close with these questions. . .
315
+ x Will extreme heterogeneity happen? y Are supercom-
316
+ puter workloads worth extreme specialization? z When
317
+ will we have production supercomputers with more than
318
+ one accelerator type?
319
+ ∗If one considers power efficiency for system scaling, massive weak
320
+ scaling would not have been possible without dramatic increase in
321
+ power/performance of compute nodes. However, such improvements usually
322
+ allow increase in the number of nodes and/or processor units, thus helping to
323
+ push weak scaling; as such, in terms of algorithmic scalability, weak scaling
324
+ is still the dominating factor.
325
+ Prepared using sagej.cls
326
+
327
+ 4
328
+ arXiv preprints
329
+ Figure 1. Classification of Compute Kernels and Supercomputing Architecture
330
+ Myth 4: Everything Will Run on Some
331
+ Accelerator!
332
+ Related to our previous myth, even if one accepts that there
333
+ will not be a plethora of accelerators, there could be a few
334
+ such as GPUs or FPGAs, where the dominant portion of
335
+ the workload will run. Indeed, for GPU-based machines
336
+ that would be an assumption, lest the extra investment will
337
+ not make sense. However, one could question, would some
338
+ superchip such as GPUs largely replace the CPUs, the latter
339
+ be degraded to second class citizens? It is not trivial as it may
340
+ seem, as such statements are rather dogmatic and not based
341
+ on candid analysis of the workloads. By proper analysis of
342
+ the workloads, we may find that CPUs may continue to play
343
+ a dominant role, with accelerator being an important but less
344
+ dominant sidekick.
345
+ From the hardware perspective, workloads can be largely
346
+ divided into three classes, (C) compute bound, (B) memory
347
+ bandwidth bound, and (L) memory latency bound. Any
348
+ application will be composed of multiple compute kernels,
349
+ each one being able to be largely classified into one of the
350
+ three in Figure 1. Over time, supercomputer architectures
351
+ have evolved in an attempt to cover all three in effective ways.
352
+ Up until the 90s, special-purpose vector machines such
353
+ as Cray and NEC SX accelerated largely (B), and (C) to
354
+ some extent. This was largely due to the dominant workload
355
+ that was CFD which was largely (B). Then in the 90s
356
+ the microprocessor evolution for HPC happened, utilizing
357
+ the commodity one-chip CPUs which had become very
358
+ powerful due to high end applications such as engineering
359
+ and multimedia needs, starting with workstation/server RISC
360
+ then later x86 processors in massively parallel fashion,
361
+ e.g., DoE ASCI Red. Individual processors were mediocre
362
+ in performance but attained performance via massive
363
+ parallelism, exercising weak-scaling, cf. Myth 3.
364
+ Then in the late 2000s, although achieving Petascale
365
+ performance was pioneered with the DoE Roadrunner and
366
+ Jaguar machines, there was an ambition to achieve exascale
367
+ by the late 2010s, achieving 1000x scaling in performance
368
+ in 10 years. The roadblock was power/performance
369
+ using conventional CPUs. However by the late 2000s
370
+ the GPUs were evolving from their graphics-specific
371
+ purpose to become general purpose compute processors,
372
+ as they were architectural descendents of classical vector
373
+ processors Matsuoka (2008). Different from classical vectors
374
+ were that the floating point performance had been significantly
375
+ enhanced, motivated by graphical workloads, and when
376
+ generalized, the GPUs were now covering (C) and (B), while
377
+ (L) was left for CPUs as GPU vector pipeline had very long
378
+ latency. CPUs that facilitated SIMD vector units with high
379
+ bandwidth memory such as the Intel Xeon Phi and Fujitsu
380
+ A64FX brought in classical vector properties back into the
381
+ CPUs, so in a sense homogeneous system composed of such
382
+ chips were not direct reincarnations of simple commodity
383
+ CPU based massively parallel machines, but rather, can be
384
+ more regarded as converging the GPU and CPU properties.
385
+ Circa 2022, the top machines are either homogeneously
386
+ configured heterogeneous CPU-GPU nodes, or ‘converged’
387
+ nodes such as RIKEN Fugaku or forthcoming machines with
388
+ Intel Sapphire Rapids CPUs with HBM. However, this is not
389
+ the only possible combination, and other configurations have
390
+ not been properly explored.. For example, one could conceive
391
+ of a machine with the latter configuration, with purpose built
392
+ matrix-based accelerators for compute intensive kernels as a
393
+ separate chip (or chiplet). In such a machine, the CPU would
394
+ cover workloads (B) and (L), while the matrix accelerator will
395
+ cover (C), . The benefit of such a machine would be ease of
396
+ programming of (B) workloads which often involve complex
397
+ memory access patterns, and thus porting to GPU codes has
398
+ proven to be challenging.
399
+ For further acceleration of (L) workloads, there is a limit to
400
+ acceleration, such as molecular dynamics that require strong
401
+ scaling. The best strategy seen for such workloads is fully
402
+ customized data pipelines such as Anton (Shaw et al. 2008)
403
+ with hardware design time synthesis. One could almost mimic
404
+ such customization with cost but make it programmable
405
+ by FPGAs or CGRAs. Such dataflow customization could
406
+ also be useful for compute bound workloads such as DL
407
+ Transformers, if small matrix engines as special function units
408
+ can be conjoined in a larger macro dataflow as seen in modern
409
+ FPGAs and CGRA chips. As such, in such a machine, (B)
410
+ will be covered by CPUs, while (C) and (L) will be covered
411
+ by a ‘strong scaling accelerator’.
412
+ As we observe here, we find that we have not even covered
413
+ the possible configurations of divergence/convergence of
414
+ Prepared using sagej.cls
415
+
416
+ Matsuoka, Domke, Wahib, Drozd, Hoefler
417
+ 5
418
+ processing units, as the only mainstream ‘accelerated’
419
+ machines are GPUs with the second property, while other
420
+ design spaces have not been properly explored.
421
+ We close with these questions. . .
422
+ x Will CPUs become pure “servants” to the accelerators?
423
+ y Are accelerators actually more than just better balanced
424
+ processors? z Will reconfigurable accelerators see a
425
+ renaissance?
426
+ Myth 5: Reconfigurable Hardware Will Give
427
+ You 100X Speedup!
428
+ In a “fool me once...” fashion, one accelerator in particular
429
+ has taken the HPC community by storm with lofty promises
430
+ of 100x speedup (Lee et al. 2010) ever since the first
431
+ ported matrix-multiplication by Larsen and McAllister (2001).
432
+ Fueled by NVIDIA’s gross margin of over 50% (Macrotrends
433
+ LLC 2022), and supported by billions of dollars from US
434
+ DOE for ECP and similar programs in other parts of the world,
435
+ the HPC community eventually migrated to a well designed
436
+ and broadly adopted GPU/CUDA ecosystem. Consequently,
437
+ 164 systems of the TOP500 list utilize accelerators from
438
+ NVIDIA. Nearly two decades later, Fugaku has shown that
439
+ it only took long vectors and high-bandwidth memory to
440
+ match GPU performance and energy-efficiency for many
441
+ workloads. One positive aspect is that that much code has
442
+ been “modernized”, i.e., rewritten in CUDA or languages and
443
+ frameworks promising portability to utilize new devices. But
444
+ the open question is how portable are these modernized codes
445
+ really? Can they run seamlessly on all new devices?
446
+ The global FPGA market was recently valued at about
447
+ one-third of the global GPU market (Allied Market Research
448
+ 2020, 2022). Major chip vendors buying the leading FPGA
449
+ hardware vendors, AMD acquired Xilinx and Intel bought
450
+ Altera, respectively, indicate an interest for FPGA integration
451
+ into future mainstream products. However, so far this has not
452
+ materialized. Whether FPGA can replace or complement the
453
+ mainstream GPUs in the HPC and data center market hinges
454
+ on the questions regarding the cost-to-performance ratio,
455
+ an existing software ecosystem, and most importantly the
456
+ productivity of programmers. Unfortunately, we see hurdles
457
+ in all these areas, which the community and industry might
458
+ be able to solve with enough time and money. Without
459
+ offering at least a factor of 10x performance gain at moderate
460
+ porting costs, “FPGAs are not a factor in our current planning,
461
+ because of their unprogrammability” (Sorensen et al. 2019).
462
+ The question whether reconfigurable logic can replace
463
+ or ament GPUs as accelerators is interesting. FPGAs will
464
+ certainly have a harder time due to their high flexibility that
465
+ comes at a cost. Units built from reconfigurable logic are
466
+ 10–20x less energy and performance efficient in silicon area.
467
+ This issue can be addressed by hardening certain blocks, e.g.,
468
+ floating point units, as some FPGA companies do. However,
469
+ even then, the whole control path would be much less efficient
470
+ and it is unclear whether program-driven execution is that
471
+ much less efficient compared to reconfigurable dataflow. A
472
+ new line of reconfigurable accelerators as materialized in
473
+ Xilinx’ adaptive compute acceleration platform are similar
474
+ to coarse-grained reconfigurable arrays (CGRAs) and offer
475
+ more programmable blocks with a configurable dataflow
476
+ interconnect. But if now 90% of the chip are hardened units,
477
+ are those devices just GPUs with a less mature ecosystem?
478
+ We close with these questions. . .
479
+ x Will the HPC community embrace FPGAs as
480
+ alternatives to GPUs in large-scale production systems?
481
+ y Can the community afford a “Fool me twice...”
482
+ moment? z Will CGRA-style reconfigurable dataflow
483
+ accelerators take the place of FPGAs to compete?
484
+ Myth 6: We Will Soon Run at Zettascale!
485
+ Maybe FPGAs are the way to zettascale. With Aurora still
486
+ under construction, Intel ignited the debate about zettascale
487
+ in late 2021. While the HPC community initially smirked
488
+ at their plans, Intel continued pushing the zettascale agenda,
489
+ culminating in the latest claims to achieve 1 zettaflop/s by the
490
+ end of the decade (Cutress 2022a). This proposition needs to
491
+ be addressed, and we try to put their claims into perspective
492
+ and predict a realistic timeline. Obviously, we cannot rule
493
+ out that Intel has a secret, revolutionary technology which
494
+ they plan to commercialize in due time, however let us not
495
+ speculate now and instead build on publicly available data.
496
+ But first we have to distinguish the terms. We assume
497
+ in the following, that (1) “zettaflop system” refers to
498
+ any computer capable of achieving over 1021 double-
499
+ precision floating-point operations (“FP64”) per second
500
+ on the Linpack benchmark; (2) “zettaop system” refers
501
+ to any computer theoretically capable of performing 1021
502
+ operations† per second, and (3) “zettascale system” denotes
503
+ any computer executing a scientific application with a
504
+ sustained performance of over 1 zettaflop/s in fp64.
505
+ Before we extrapolate, we look at historical trends
506
+ by Strohmaier et al. (2022). The HPC community achieved
507
+ 1.068 teraflop/s with Sandia/IBM’s ASCI Red in summer
508
+ 1997, 1.026 petaflop/s with Los Alamos/IBM’s Roadrunner
509
+ in summer 2008, and achieved (unofficially) 1.05 exaflop/s
510
+ in spring of 2021 with China’s OceanLight system and
511
+ 1.1 exaflop/s with OakRidge/HPE’s Frontier in summer
512
+ 2022. Not only do 11 and 13 years lie in between these
513
+ achievements, respectively, but also multiple megawatt. ASCI
514
+ Red consumed “only” 0.850 MW, Roadrunner increased that
515
+ to 2.35 MW, and OceanLight and Frontier now consume
516
+ 35 MW and 21.1 MW, respectively. This and Figure 2 show
517
+ that the energy efficiency of modern chips cannot keep up
518
+ with the demand for increasing compute.
519
+ Back to Intel claiming to manage 2x performance
520
+ improvements year-over-year which would yield zettaflop/s
521
+ by 2032—but at a power requirement of the entire
522
+ system of 50–100 MW (Cutress 2022b). Hence, this 1,000x
523
+ in performance comes at the cost of 3–5x in power;
524
+ and reformulated: the energy efficiency to perform fp64
525
+ operations needs to increase by 200–350x, from ≈50 to
526
+ over 10.000 Gflop/s
527
+ Watt . Even under idealized conditions and
528
+ using Frontier’s Rpeak as baseline, this goal requires a
529
+ †An exact and consistent definition of “operation” in this context is still
530
+ debated in the HPC community.
531
+ Prepared using sagej.cls
532
+
533
+ 6
534
+ arXiv preprints
535
+ Figure 2. Historical fp64 power efficiency [in Gflop/s
536
+ Watt ] extrapolated until 2038 to put Intel’s zettaflop/s claims into perspective.
537
+ 125x improvement in 10 years, and all of that while
538
+ other big players slowly acknowledge the end of practical
539
+ silicon scaling laws (White 2022). If we believe the IEEE
540
+ IRDS™ (2021) roadmap, we might gain 5x in power
541
+ density (optimistically rounded from 4.27x) by 2034 at 7 ˚A
542
+ compared to 5 nm. This leaves 25x, which we could split
543
+ into 5x from increased transistor count per chip and 5x from
544
+ increased node count per system. Can we cool the former,
545
+ yes (Wu et al. 2021), and can we interconnect the latter?
546
+ Sure, but doing so, at 2.5 GW, comes down to the will to
547
+ invest more than anything else, and without revolutions in
548
+ memory and interconnect technologies, we might see Linpack
549
+ transition into memory- or I/O-bound territory, nullifying any
550
+ computational advances.
551
+ On the other hand, a zettaop/s system at 100 MW in 2032
552
+ is far more likely, since low-precision units (such as tensor
553
+ cores) can boost the op/s
554
+ Watt metric, e.g., currently fp16 tensor
555
+ cores demonstrate an 8x advantage over fp64 vector units.
556
+ Lowering the precision further from fp16 to 3-bit operands
557
+ could allow for another 5x improvement (Frantar et al. 2022),
558
+ but only if the industry (and HPC community) sees the need
559
+ for adding these low-precision units, as we discuss in Myth 11.
560
+ Considering the above, our more realistic, yet optimistic,
561
+ timeline for zetta is zettaop/s in 2032 at 50 MW, zettaflop/s
562
+ in 2037 at 200 MW, and zettascale by 2038. Can Intel or
563
+ anybody else pull it off before then? Only time will tell.
564
+ We close with these questions. . .
565
+ x Will we reach zettaflop/s performance or will fp64
566
+ lose relevance before? y Will we continue to build
567
+ more power-hungry supercomputers as we did in the
568
+ past? z Which one will happen first: zettascale, practical
569
+ quantum advantage, or all internal combustion-based
570
+ engines cease to be produced?
571
+ Myth 7: Next-Generation Systems Need
572
+ More Memory per Core!
573
+ Before, on the road to peta- and exascale, application
574
+ scientists continuously raised alarms that the memory per
575
+ core is decreasing with each new computer generation.
576
+ This was mainly due to the quick growth in the number
577
+ of cores while the performance per core was stagnating.
578
+ Yet, many workloads can keep those cores utilized with a
579
+ relatively small working set while staging larger amounts of
580
+ data remotely and/or recomputing parts. Much of this large
581
+ memory requirement seemingly turns out to be legacy and
582
+ somewhat wasteful design from times where memory space
583
+ was abundant compared to other resources.
584
+ Simplistic arguments along the lines of “we need more
585
+ of X” seem to have a solid tradition in our community. For
586
+ example, the HPC community spent the first decades to hunt
587
+ more floating point computations per second. Recently, a
588
+ demand for larger and faster memory replaced this main goal.
589
+ The community nearly made a complete 360-degree turn,
590
+ with Haus (2021) saying “computation is free” and Ivanov
591
+ et al. (2021) showing “data movement is all you need”.
592
+ Some even argue that this turn was taken too late due
593
+ to the fixation on flop/s. While this was all true at the
594
+ time, the general discussion should really be about the
595
+ intricate relation between the application requirements and
596
+ the system capabilities in terms of balance, i.e., ratio between
597
+ the different resources such as memory size/bandwidth and
598
+ compute (Czechowski et al. 2011).
599
+ These ratios usually shift with chip technology and
600
+ architectural choices. For example, Moore’s law drove the
601
+ costs for compute on chip down over decades but off-chip
602
+ communication was limited by Rent’s rule. This led to the
603
+ recent data movement crisis. Newly emerging optical off-
604
+ chip connectivity, see Myth 8, as well as 3D integrated
605
+ memory (Domke et al. 2022) shifts the balance again and
606
+ may alleviate many of these aspects, at least at the scale of
607
+ a single chip. It seems key to understand the malleability of
608
+ application, i.e., which resources can be traded for which
609
+ other resources (e.g., memory capacity for computation
610
+ bandwidth using recomputation or caching as techniques).
611
+ Prepared using sagej.cls
612
+
613
+ Matsuoka, Domke, Wahib, Drozd, Hoefler
614
+ 7
615
+ Here, specifically I/O complexity analysis is a tool to deeply
616
+ understand this trade-off. Once all trade-offs are understood,
617
+ requirements models (Calotoiu et al. 2018) could be used to
618
+ fix trade-offs into designs. These models could then inform
619
+ architectural choices as well as hardware developments.
620
+ One area to highlight in this context is embedded design
621
+ where such trade-offs have long been used to build real
622
+ systems due to resource scarcity (e.g., battery). While those
623
+ designs were initially limited to very narrow application
624
+ domains (e.g., radio signal, audio, or video processing),
625
+ embedded devices have recently been expanded towards more
626
+ diverse workloads (“apps”). We believe that HPC can learn
627
+ from this field by defining clear system design methodologies
628
+ based on a solid combination of empirical and analytical
629
+ modeling. More particularly, systems design in HPC can
630
+ benefit from the embedded systems doctrine of accounting for
631
+ over-engineering just as one accounts for under-engineering.
632
+ We close with these questions. . .
633
+ x When will the current “data movement” focus end?
634
+ y What will be the next bottleneck resource? z Will
635
+ our community be able to adopt a performance modeling
636
+ discipline to discuss bottlenecks scientifically?
637
+ Myth 8: Everything Will Be Disaggregated!
638
+ To stop the waste of memory resources, the academic com-
639
+ munity is advancing on the Silicon Photonics front (Gonzalez
640
+ et al. 2022) and industry is pursuing scale-out technologies (Li
641
+ et al. 2022), such as Compute Express LinkTM (CXL), a
642
+ cache-coherent interconnect for data centers. But a few
643
+ players seem to push the idea over the edge with their
644
+ plans to disaggregate everything (NTT R&D 2020; Shan
645
+ et al. 2022). As Gonzalez et al. (2022) stated: “An optical
646
+ interconnect is more appealing than an electrical interconnect
647
+ for memory disaggregation due to three properties: its (1)
648
+ high bandwidth density significantly reduces the number of
649
+ IO lanes, (2) power consumption and crosstalk do not increase
650
+ with distance, and (3) propagation loss is low.” However,
651
+ several barriers remain before we can fully replace copper-
652
+ based interconnects in our supercomputers.
653
+ Generally, we see two remaining challenges for a broad
654
+ adoption of Silicon Photonics and all-optical interconnects:
655
+ low-cost manufacturing and optical switching. The former
656
+ is obvious, because after all, the data center and HPC
657
+ community relies on inexpensive components to optimize the
658
+ overall system performance-to-cost ratio. The latter challenge
659
+ is less obvious for the uninitiated. Current electrically
660
+ switched networks can operate in “packet switching” mode
661
+ to effectively lower the observable latency and utilize the
662
+ available link bandwidth. The alternative to this mode
663
+ is “circuit-switching” and it was abandoned by the HPC
664
+ community long ago in favor of packet-switching. However,
665
+ without (cost-)effective means to buffer light, process photon
666
+ headers in-flight, or reverting to electric switches with
667
+ expensive optical-electrical-optical conversions, we would
668
+ have to resort to circuit-switching (Bergman et al. 2022)
669
+ with all the inherent deficiencies: complex traffic steering
670
+ calculations, switching delays, latency increase due to lack of
671
+ available paths, under-utilization of links, just to name some.
672
+ For HPC, an extensive or extreme disaggregation yields
673
+ another challenge, specifically the speed of light. Photons
674
+ travel at a maximum speed of 3.3 ns/m in hollow fibers
675
+ (or slower in other transport media). This is equivalent to
676
+ a level-2 cache access of a modern CPU, but does not yet
677
+ include the disaggregation overhead, such as from the CXL
678
+ protocol itself, switching, or optical-electrical conversions at
679
+ the endpoints. At 3–4 m distance, the photon travel time alone
680
+ exceeds the first-word access latency of modern DDR memory.
681
+ Therefore, if main memory would be disaggregated beyond
682
+ rack boundaries, it will become noticeable for memory-
683
+ latency sensitive applications, cf. Myth 4. The more sensible
684
+ solution, in line with Myth 7, for future HPC systems are
685
+ smaller node-local memory configurations (e.g., HBM3)
686
+ paired with rack-local, CXL-based memory pools if the
687
+ capacity- and performance-to-cost ratios of the memory pool
688
+ plus required interconnect can outperform node-local SSD
689
+ solutions.
690
+ We close with these questions. . .
691
+ x Will CXL be deployed widely in HPC? y Will large-
692
+ scale supercomputers be disaggregated beyond rack-
693
+ scale? z Should we disaggregate main memory?
694
+ Myth 9: Applications Continue to Improve,
695
+ Even on Stagnating Hardware!
696
+ Modernizing hardware, with Silicon Photonics, Tensor Cores,
697
+ or simply shrinking transistors, has too long been the primary
698
+ method of accelerating legacy software. More than half
699
+ of this improvement was based on Moore’s law and its
700
+ observation that transistors will continue to become smaller
701
+ every few years (originally 18 months). The remaining
702
+ hardware improvements came from architectural innovations,
703
+ such as deeper cache hierarchies, the migration to more
704
+ specialized architectures (e.g., GPUs), or the utilization of
705
+ larger and wider vector-units (SIMD), as well as scaling the
706
+ HPC systems up by giving them more processors and cores.
707
+ Unfortunately, we are no longer seeing the consistent
708
+ technology scaling that Moore observed. Consequently, in
709
+ the so-called Post-Moore era, the “performance road” forks
710
+ three-ways, yielding the following options: (1) architectural
711
+ innovations will attempt to close the performance gap, and
712
+ an explosion of diverging architectures tailored for specific
713
+ science domains will emerge, or (2) alternative materials and
714
+ technologies (e.g., non-CMOS technologies) that allow the
715
+ spirit of Moore’s law to continue for a foreseeable future,
716
+ or (3) we abandon the von-Neumann paradigm together and
717
+ move to a neuromorphic or quantum-like computer (which,
718
+ in time, might or might not become practical as discussed in
719
+ Myth 1). One major aspect that reflects the uncertainty about
720
+ the future is the initiatives of unprecedented scale: CHIPS act
721
+ in the US and similar initiatives in other countries in the order
722
+ of 100s Billion USD, quantum computing initiatives in the
723
+ order of 10s Billion USD, etc.
724
+ But one point that is often overlooked is that algorithmic
725
+ improvements in HPC (dubbed as “Algorithmic Moore’s
726
+ Law” by Keyes (2022)) have over time provided exponential
727
+ improvement in key areas of HPC, see Figure 3. Similar
728
+ reports attribute a significant portion of the performance
729
+ Prepared using sagej.cls
730
+
731
+ 8
732
+ arXiv preprints
733
+ higher
734
+ order AMR
735
+ 1
736
+ 10
737
+ 100
738
+ 1000
739
+ 10000
740
+ 100000
741
+ 1000000
742
+ 10000000
743
+ 100000000
744
+ 1980
745
+ 1990
746
+ 2000
747
+ 2010
748
+ 2020
749
+ Effective Sustained Speedup
750
+ Algorithmic Moore's Law Examples
751
+ 100
752
+ 101
753
+ 102
754
+ 103
755
+ 104
756
+ 105
757
+ 106
758
+ 107
759
+ 108
760
+ Sustained Speed in Gflop/s
761
+ Combustion Simulation
762
+ (Complex Kinetics)
763
+ Combustion Simulation
764
+ (CFD)
765
+ COSMO Climate Model
766
+ Fusion Energy Simulation
767
+ (Global MHD)
768
+ Moore’s Law
769
+ Fusion Energy Simulation
770
+ (Micro-turbulence)
771
+ improved
772
+ linear solver
773
+ ARK integrator
774
+ complex chem
775
+ AMR
776
+ semi-implicit
777
+ high-order
778
+ elements
779
+ gyro-
780
+ kinetics
781
+ delta-f,
782
+ magnetic
783
+ coordinates
784
+ improved
785
+ electron
786
+ models
787
+ low Mach
788
+ auto-code
789
+ high order
790
+ improved
791
+ explicit/implicit
792
+ solvers
793
+ Figure 3. Examples of “Algorithmic Moore’s Law” for different areas in HPC; Fusion energy and combustion simulations data
794
+ by Keyes (2022) and climate simulation data by Schulthess (2016)
795
+ improvement in many legacy codes to be from numerical
796
+ solvers, algorithms, low-precision numerics, system software,
797
+ etc Schulthess (2016). However, we have to be cautious that—
798
+ just as hardware improvements have physics and engineering
799
+ limits—the “Algorithmic Moore’s Law” also has its own
800
+ limits: numerical stability, hitting asymptotic limits, etc. That
801
+ being said, those limits might not be as clear and quantifiable
802
+ as the limits on hardware. That is since even if one numerical
803
+ method hits its limit, domain experts can often reduce/pre-
804
+ condition their problem to another numerical method that is
805
+ more efficient.
806
+ We close with these questions. . .
807
+ x As the performance improvements from hardware
808
+ technologies drop, should the HPC community dramat-
809
+ ically increase the investment in software? y Will the
810
+ “Algorithmic Moore’s Law” end soon as well? z To what
811
+ extent is the HPC community willing to refactor/rewrite
812
+ legacy codebases when/if hardware stagnates?
813
+ Myth 10: Fortran Is Dead, Long Live the DSL!
814
+ Applications might have limits, but what about languages.
815
+ How often have we heard “Fortran is dead, long live X”?
816
+ Slogans like this have been resonating in the community for
817
+ nearly 40 years (Post 1982). X has been everything from
818
+ C to C++, and more recently Python or Domain-Specific
819
+ Languages (DSLs). Yet, Fortran remains in wide use in
820
+ important communities such as weather and climate even
821
+ for newly written codes. Other languages, such as COBOL
822
+ were indeed replaced with more modern alternatives such
823
+ as Java. Why is this? Are some parts of our community just
824
+ stubborn to follow the youngsters? Or are old languages not
825
+ necessarily bad for the task? Indeed, Fortran is a very well
826
+ designed language for its purpose of expressing mathematical
827
+ programs at highest performance. It seems hard to replace it
828
+ with C or other languages and outperform it or even achieve
829
+ the same baseline. This may be due to the highly optimized
830
+ Fortran compilers or the limited language features (e.g., no
831
+ pointer aliasing) that enable more powerful optimizations.
832
+ Fortran and other general-purpose languages remain
833
+ competitive with many DSLs on CPUs (Ben-Nun et al.
834
+ 2022) and are recently also adopted to GPUs, albeit often
835
+ less elegant. General-purpose portability approaches such as
836
+ SYCL (Keryell et al. 2015), also powering Intel’s oneAPI,
837
+ or OpenMP provide flexibility as well as some portability
838
+ across devices. High-productivity general-purpose languages
839
+ are hard to accelerate in practice. For example, Python’s
840
+ flexibility (e.g., monkey patching and flexible typing) disables
841
+ many static optimizations. However, when restricting the
842
+ syntax to high-performance Python (much of NumPy), then
843
+ optimizations become simpler (Ziogas et al. 2021). Any
844
+ language becomes more complex over time—Fortran 66
845
+ evolved into the complex Fortran 2018 language standard.
846
+ Similar trends affect DSLs that are widening their scope over
847
+ time. Do we require this generality? If yes, then DSLs are
848
+ doomed to fail or they morph into general-purpose languages.
849
+ Another argument is that the lower levels usually remain
850
+ C/C++ and programmers interested in highest performance
851
+ are often happy to dig into the lower levels. Then the question
852
+ remains—where should the portability layer be located? At a
853
+ (virtualized) Instruction Set Architecture (ISA) as in LLVM’s
854
+ IR (Lattner and Adve 2004), some lower-level language
855
+ such as C/C++ as in SYCL/oneAPI, or even dataflow graph
856
+ representations as in DaCe (Ben-Nun et al. 2019)?
857
+ We close with these questions. . .
858
+ x When will programmers stop using Fortran for new
859
+ applications? y Will we ever have more application codes
860
+ written in DSLs than general-purpose languages? z What
861
+ will be the next big DSL?
862
+ Myth 11: HPC Will Pivot to Low or Mixed
863
+ Precision!
864
+ A high-performance language is nothing without proper data
865
+ types, but high-precision operations such as fp64 come at a
866
+ significant cost in terms of silicon area, energy and speed,
867
+ according to Myth 6. Lowering this precision can save costs
868
+ Prepared using sagej.cls
869
+
870
+ Matsuoka, Domke, Wahib, Drozd, Hoefler
871
+ 9
872
+ but may reduce accuracy of the results and, in the worst case,
873
+ break the application (e.g., convergence). But there is more
874
+ to this trade-off: what if a more clever implementation could
875
+ maintain convergence properties of high precision numerics,
876
+ while enjoying computational efficiency of low precision?
877
+ One common trick is using mixed precision on the algorithmic
878
+ level, for example, using low precision for individual particles
879
+ and only using high precision for aggregated values (Kutzner
880
+ et al. 2019). Some processors offer mixed precision tricks
881
+ at the hardware level in the form of instructions with low
882
+ precision inputs but higher precision accumulations.
883
+ There is however more to reduced precision than using
884
+ fewer bits—the question is how to optimally distribute bits
885
+ between mantissa and exponent (Tesla, Inc. 2021), or even if
886
+ to use an entirely different (not IEEE-754) way to represent
887
+ numbers (Gustafson and Yonemoto 2017). The story of
888
+ reduced precision in AI hardware is quite telling: In early
889
+ days of the field, predominantly the IEEE fp32 format was
890
+ used, but knowing that in deep neural nets the weights and
891
+ activations are typically distributed on a small range of values,
892
+ researchers began to explore the fp16 format. Soon the Pascal
893
+ generation of GPUs with fp16 performance—at a factor of
894
+ two compared to fp32 was released—and the magic did not
895
+ happen by itself. Exploding and vanishing gradients, outlier
896
+ weights, etc., made training large deep neural nets require
897
+ extra effort to stabilize (incurring corresponding overhead) or
898
+ just did not converge at all. The next generation of devices
899
+ came with bfloat16 format—same 16 bits, but more bits
900
+ allocated to range, less for precision. It worked better, but
901
+ still once in a while a model would collapse. Finally, the
902
+ recent generation of GPUs came with a 19-bit numeric format,
903
+ misleadingly called TensorFloat-32. So far it seems to be at
904
+ the sweet spot for artificial intelligence workloads—allowing
905
+ for noticeably faster arithmetics than fp32, while maintaining
906
+ enough numeric stability for the models to reliably converge
907
+ without extra programming effort.
908
+ Now that mixed precision is a de-facto standard in the AI
909
+ domain, more hardware support is being implemented. So
910
+ far there is no general clarity on the limits—how few bits
911
+ can we get away with in different HPC areas. The following
912
+ factors in particular are important to consider as we move
913
+ forward. A fully transparent solution for the problem is to
914
+ simulate higher precision using low precision operations,
915
+ e.g., as shown by Ootomo and Yokota (2022). Our Myth 4’s
916
+ memory-bound problems in particular are good candidates
917
+ for exploiting “simulated” high precision, since the overhead
918
+ can be masked by data transfers. It is not clear however
919
+ if this incurred overhead is an acceptable price that HPC
920
+ agrees to pay for remaining in higher precision. A less
921
+ transparent method is to approach the problem as precision
922
+ auto-tuning task by adapting the precision to a minimum
923
+ while bounding the error, e.g., as demonstrated by Menon et al.
924
+ (2018). One main limitation of that method is the reliance
925
+ on automatic differentiation (AD) to track error propagation,
926
+ which is not practical for large codebases. Finally, the least
927
+ transparent approach requires domain experts in HPC to study
928
+ the numerical stability of solvers to identify, on a case-by-case
929
+ basis, the susceptibility of solvers to lower/mixed precision.
930
+ While this approach is viable for solvers that are wrapped in
931
+ libraries to be consumed by HPC domain experts, it is unclear
932
+ whether domain experts writing their own solvers (common
933
+ in HPC) would be willing to take on this burden.
934
+ We close with these questions. . .
935
+ x Is the HPC community ready (or already late?) to react
936
+ to the new low precision formats driven by deep learning?
937
+ y Will HPC navigate itself into a high-precision niche?
938
+ z When, if ever, will the industry drop fp64 support?
939
+ Myth 12: All HPC Will Be Subsumed by the
940
+ Clouds!
941
+ The rapidly advancing AI and new precision options has
942
+ reignited the cloud discussion. The question whether clouds
943
+ will subsume supercomputing has been ongoing for more
944
+ than a decade, since the late 2000s Deelman et al. (2008), but
945
+ remains inconclusive. Today’s cloud offerings offer a wide
946
+ spectrum for HPC customers, ranging from low-cost standard
947
+ virtual machines to specialized top-gear HPC equipment in
948
+ the cloud. It is not surprising that cloud providers offer exactly
949
+ the same performance as on-prem supercomputing centers
950
+ in practice De Sensi et al. (2022). After all, they simply buy
951
+ the same hardware! Thus, this discussion is more of a fiscal
952
+ argument with an interesting economy-of-scale twist.
953
+ There are actually bi-directional aspects to the cloud-vs-
954
+ supercomputer argument. One is the so-called “cloudification
955
+ of supercomputers”, and the latter being “supercomputifica-
956
+ tion of clouds”, but they often get mixed-up leading to the
957
+ confusions in the discussions. We must look at both aspects,
958
+ and it is in fact the latter where such subsumption may happen
959
+ or not.
960
+ The former, “cloudification of supercomputers”, is an
961
+ unmistakable trend, in that various software stack features
962
+ and APIs are added so that supercomputers effectively
963
+ become high end compute resources in the same manner as
964
+ commercial clouds. Indeed, many major supercomputers are
965
+ already facilitating cloud features, so that they are effectively
966
+ clouds themselves, and interoperable with commercial clouds.
967
+ However, this assumes that there is already a supercomputing
968
+ resource facilitated for themselves, and does not directly affect
969
+ the subsumption argument.
970
+ The latter, or “supercomputification of clouds”, is where
971
+ subsumption may happen, in that clouds nowadays can
972
+ support features as well as performances of dedicated
973
+ supercomputers directly, such that they are directly amenable
974
+ as their replacement. Certainly, there are now multiple
975
+ cloud services that facilitate virtual compute clusters in
976
+ the cloud. However, although Intersect 360 reports that
977
+ HPC-in-the-cloud CAGR has been dramatic, over 80% in
978
+ 2021 Intersect360 Research (2022), it also reports the overall
979
+ high growth in the HPC market, especially in the high end,
980
+ and also projects that, the growth in the cloud HPC market
981
+ will flatten over time to be consistent with the overall HPC
982
+ industry growth. Continued investments by all major global
983
+ regions in exascale machines and beyond, coupled with
984
+ companies facilitating their own top-ranked machines, will
985
+ likely continue to fuel the on-prem infrastructure growth.
986
+ In fact, for enterprise IT infrastructures, there has always
987
+ been a swing between on-prem and public clouds, largely
988
+ Prepared using sagej.cls
989
+
990
+ 10
991
+ arXiv preprints
992
+ driven by economics. While standing up comprehensive
993
+ internal IT has become less attractive with multitudes of
994
+ cloud services readily available in the cloud, so the CAPX for
995
+ clouds would be cheaper, especially for small enterprises and
996
+ startups, for large enterprises there is a tendency to move back
997
+ to on-prem infrastructures, as the OPEX of clouds could be
998
+ expensive. The same could be the case of HPC increasingly as
999
+ the whole field would pose continuous uprisings in economic
1000
+ viability for industry and societal benefits, thus being driven
1001
+ by economic metrics.
1002
+ However, the variant of the subsumption scenario is
1003
+ that, although on-prem supercomputers continue to exist,
1004
+ processors and other hardware developments will be largely
1005
+ driven by enterprise HPC needs, currently dominated by
1006
+ AI / deep learning workloads. The R&D expenditures of
1007
+ hyperscalers in IT now outclass the government investments,
1008
+ and increasingly the hyperscalers are investing in high end
1009
+ computing. If the commercial cloud hyperscalers can work
1010
+ out the scale of economy in their own hardware manufacturing
1011
+ to the extent that, it could build and operate large scale
1012
+ HPC infrastructures cheaper than on-prem supercomputers
1013
+ of any size, then the swing could totally happen towards
1014
+ full subsumption— although somewhat unlikely, this could
1015
+ compromise the ability to cover some of the traditional HPC
1016
+ workloads that do not meet main industrial needs, such as the
1017
+ requirement for dense 64 bit linear algebra capabilities.
1018
+ We close with these questions. . .
1019
+ x What could be a defining development to decide
1020
+ between cloud and on-prem HPC? y When will more
1021
+ than half of the HPC cycles be spent in the cloud? z Will
1022
+ on-prem systems be a niche or remain with a significant
1023
+ fraction of HPC cycles spent?
1024
+ Conclusions
1025
+ Many myths shape the discussions in the HPC community
1026
+ today—in this work, we debate some of those and hope to
1027
+ stir up arguments. While we present them in an exaggerated
1028
+ and humorous way, many of those myths form the core of
1029
+ thinking in our community. Some may be more divisive than
1030
+ others but it seems that many are hard to answer definitively.
1031
+ Maybe some points will settle in the future while others will
1032
+ not. Yet, their sheer importance mandates a serious treatment
1033
+ in order to help guide future directions for academic research
1034
+ but also industry and government investment.
1035
+ References
1036
+ Allied Market Research (2020) Graphic Processing Unit (GPU)
1037
+ Market by Type (Dedicated, Integrated, and Hybrid), Device
1038
+ (Computer, Tablet, Smartphone, Gaming Console, Television,
1039
+ and Others), Industry Vertical (Electronics, IT & Telecommu-
1040
+ nication, Defense & Intelligence, Media & Entertainment, and
1041
+ Others): Global Opportunity Analysis and Industry Forecast,
1042
+ 2020-2027.
1043
+ https://www.alliedmarketresearch.
1044
+ com/graphic-processing-unit-market.
1045
+ Allied
1046
+ Market
1047
+ Research
1048
+ (2022)
1049
+ Field
1050
+ Programmable
1051
+ Gate
1052
+ Array Market by Technology (EEPROM, Antifuse, SRAM,
1053
+ Flash, and Others (EPROM and PROM)), Application
1054
+ (Data Processing, Consumer Electronics, Industrial, Mil-
1055
+ itary & Aerospace, Automotive, Telecom, and Others),
1056
+ and
1057
+ Type
1058
+ (High-end,
1059
+ Mid-end,
1060
+ and
1061
+ Low-end
1062
+ FPGA):
1063
+ Global Opportunity Analysis and Industry Forecast, 2021-
1064
+ 2030. https://www.alliedmarketresearch.com/
1065
+ field-programmable-gate-array-market.
1066
+ Aytekin C (2022) Neural Networks are Decision Trees. https:
1067
+ //arxiv.org/abs/2210.05189.
1068
+ Bell G, Bailey DH, Dongarra J, Karp AH and Walsh K (2017) A look
1069
+ back on 30 years of the Gordon Bell Prize. The International
1070
+ Journal of High Performance Computing Applications 31(6):
1071
+ 469–484. DOI:10.1177/1094342017738610.
1072
+ Ben-Nun T, de Fine Licht J, Ziogas AN, Schneider T and Hoefler T
1073
+ (2019) Stateful Dataflow Multigraphs: A Data-Centric Model
1074
+ for Performance Portability on Heterogeneous Architectures.
1075
+ In: Proceedings of the International Conference for High
1076
+ Performance Computing, Networking, Storage and Analysis,
1077
+ SC ’19. New York, NY, USA: Association for Computing
1078
+ Machinery. ISBN 978-1-4503-6229-0, p. 14. DOI:10.1145/
1079
+ 3295500.3356173.
1080
+ Ben-Nun T, Groner L, Deconinck F, Wicky T, Davis E, Dahm
1081
+ J, Elbert OD, George R, McGibbon J, Tr¨umper L, Wu E,
1082
+ Fuhrer O, Schulthess T and Hoefler T (2022) Productive
1083
+ Performance Engineering for Weather and Climate Modeling
1084
+ with Python. In: Proceedings of the International Conference
1085
+ on High Performance Computing, Networking, Storage and
1086
+ Analysis, SC ’22. IEEE Press. ISBN 9784665454445, p. 14.
1087
+ DOI:10.5555/3571885.3571982.
1088
+ Bender EM, Gebru T, McMillan-Major A and Shmitchell S (2021)
1089
+ On the Dangers of Stochastic Parrots: Can Language Models
1090
+ Be Too Big? In: Proceedings of the 2021 ACM Conference
1091
+ on Fairness, Accountability, and Transparency, FAccT ’21.
1092
+ New York, NY, USA: Association for Computing Machinery.
1093
+ ISBN 978-1-4503-8309-7, pp. 610–623. DOI:10.1145/3442188.
1094
+ 3445922.
1095
+ Bergman K, Beausoleil R and Milojicic D (2022) Silicon Photonics.
1096
+ Computer 55(04): 78–81. DOI:10.1109/MC.2022.3148491.
1097
+ Beverland ME, Murali P, Troyer M, Svore KM, Hoefler T,
1098
+ Kliuchnikov V, Low GH, Soeken M, Sundaram A and
1099
+ Vaschillo A (2022) Assessing requirements to scale to practical
1100
+ quantum advantage.
1101
+ https://doi.org/10.48550/
1102
+ arXiv.2211.07629.
1103
+ Bi K, Xie L, Zhang H, Chen X, Gu X and Tian Q (2022) Pangu-
1104
+ Weather: A 3D High-Resolution Model for Fast and Accurate
1105
+ Global Weather Forecast.
1106
+ https://arxiv.org/abs/
1107
+ 2211.02556.
1108
+ Brenowitz ND and Bretherton CS (2018) Prognostic Validation
1109
+ of a Neural Network Unified Physics Parameterization.
1110
+ Geophysical Research Letters 45(12): 6289–6298.
1111
+ DOI:10.
1112
+ 1029/2018GL078510.
1113
+ Calotoiu A, Graf A, Hoefler T, Lorenz D, Rinke S and Wolf F
1114
+ (2018) Lightweight Requirements Engineering for Exascale
1115
+ Co-design.
1116
+ In: IEEE International Conference on Cluster
1117
+ Computing, CLUSTER 2018, Belfast, UK, September 10-13,
1118
+ 2018. IEEE. ISBN 978-1-5386-8319-4, pp. 201–211. DOI:
1119
+ 10.1109/CLUSTER.2018.00038.
1120
+ Cutress
1121
+ I
1122
+ (2022a)
1123
+ Interview
1124
+ with
1125
+ Intel’s
1126
+ Raja
1127
+ Koduri:
1128
+ Zettascale or ZettaFLOP? Metaverse what?
1129
+ https:
1130
+ //www.anandtech.com/show/17298/interview-
1131
+ with-intels-raja-koduri-zettascale-or-
1132
+ Prepared using sagej.cls
1133
+
1134
+ Matsuoka, Domke, Wahib, Drozd, Hoefler
1135
+ 11
1136
+ zettaflop-metaverse-what.
1137
+ Cutress I (2022b) The key to Intel’s ZF (ZettaFLOP, 1ZF FP64
1138
+ system), is the power. Trying to keep within the 40-100 MW per
1139
+ Supercomputer. https://twitter.com/IanCutress/
1140
+ status/1599452910743478273.
1141
+ Czechowski K, Battaglino C, McClanahan C, Chandramowlishwaran
1142
+ A and Vuduc R (2011) Balance Principles for Algorithm-
1143
+ Architecture Co-Design. In: Proceedings of the 3rd USENIX
1144
+ Conference on Hot Topic in Parallelism, HotPar’11. USA:
1145
+ USENIX Association, p. 9.
1146
+ De Sensi D, De Matteis T, Taranov K, Di Girolamo S, Rahn T and
1147
+ Hoefler T (2022) Noise in the Clouds: Influence of Network
1148
+ Performance Variability on Application Scalability. Proc. ACM
1149
+ Meas. Anal. Comput. Syst. 6(3). DOI:10.1145/3570609.
1150
+ Deelman E, Singh G, Livny M, Berriman B and Good J (2008) The
1151
+ cost of doing science on the cloud: The Montage example. In:
1152
+ SC ’08: Proceedings of the 2008 ACM/IEEE Conference on
1153
+ Supercomputing. pp. 1–12. DOI:10.1109/SC.2008.5217932.
1154
+ Domke J, Vatai E, Drozd A, Peng C, Oyama Y, Zhang L, Salaria
1155
+ S, Mukunoki D, Podobas A, Wahib M and Matsuoka S (2021)
1156
+ Matrix Engines for High Performance Computing: A Paragon
1157
+ of Performance or Grasping at Straws?
1158
+ In: 2021 IEEE
1159
+ International Parallel and Distributed Processing Symposium,
1160
+ IPDPS 2021, Portland, Oregon, USA, May 17-21, 2021.
1161
+ Portland, Oregon, USA: IEEE Press, pp. 1056–1065. DOI:
1162
+ 10.1109/IPDPS49936.2021.00114.
1163
+ Domke J, Vatai E, Gerofi B, Kodama Y, Wahib M, Podobas A,
1164
+ Mittal S, Peric`as M, Zhang L, Chen P, Drozd A and Matsuoka
1165
+ S (2022) At the Locus of Performance: A Case Study in
1166
+ Enhancing CPUs with Copious 3D-Stacked Cache. https:
1167
+ //arxiv.org/abs/2204.02235.
1168
+ Frantar E, Ashkboos S, Hoefler T and Alistarh D (2022) GPTQ:
1169
+ Accurate Post-Training Quantization for Generative Pre-
1170
+ trained Transformers. https://arxiv.org/abs/2210.
1171
+ 17323.
1172
+ Gonzalez J, G Palma M, Hattink M, Rubio-Noriega R, Orosa
1173
+ L, Mutlu O, Bergman K and Azevedo R (2022) Optically
1174
+ Connected Memory for Disaggregated Data Centers. Journal
1175
+ of Parallel and Distributed Computing 163(C): 300–312. DOI:
1176
+ 10.1016/j.jpdc.2022.01.013.
1177
+ Gustafson JL (1988) Reevaluating Amdahl’s Law. Communications
1178
+ of the ACM 31(5): 532–533. DOI:10.1145/42411.42415.
1179
+ Gustafson JL and Yonemoto IT (2017) Beating Floating Point at its
1180
+ Own Game: Posit Arithmetic. Supercomputing Frontiers and
1181
+ Innovations 4(2): 71–86. DOI:10.14529/jsfi170206.
1182
+ Haus
1183
+ UU
1184
+ (2021)
1185
+ The
1186
+ Brave
1187
+ New
1188
+ World
1189
+ of
1190
+ Exascale
1191
+ Computing:
1192
+ Computation
1193
+ Is
1194
+ Free,
1195
+ Data
1196
+ Movement
1197
+ Is
1198
+ Not.
1199
+ https://minoa-itn.fau.de/wp-content/
1200
+ uploads/2021/03/TRR154-MINOA-20210303.pdf.
1201
+ Hoefler T (2022) Benchmarking Data Science: 12 Ways to Lie With
1202
+ Statistics and Performance on Parallel Computers. Computer
1203
+ 55(8): 49–56. DOI:10.1109/MC.2022.3152681.
1204
+ Hoefler T, Haener T and Troyer M (2023) Disentangling Hype from
1205
+ Practicality: On Realistically Achieving Quantum Advantage.
1206
+ Communications of the ACM .
1207
+ Huang L and Hoefler T (2022) Compressing multidimensional
1208
+ weather and climate data into neural networks. https://
1209
+ arxiv.org/abs/2210.12538.
1210
+ IEEE IRDS™ (2021) International Roadmap for Devices and
1211
+ Systems (IRDS™) 2021 Edition – Executive Summary. IEEE
1212
+ IRDS™ Roadmap, IEEE.
1213
+ https://irds.ieee.org/
1214
+ images/files/pdf/2021/2021IRDS_ES.pdf.
1215
+ Intersect360
1216
+ Research
1217
+ (2022)
1218
+ Worldwide
1219
+ HPC
1220
+ and
1221
+ AI
1222
+ Training
1223
+ Market,
1224
+ 2021
1225
+ Actuals,
1226
+ 2022-26
1227
+ Forecast.
1228
+ https://www.intersect360.com/wp-content/
1229
+ uploads/Webinar-Intersect360-WW-HPC-AI-
1230
+ Unified-2021-market-size-and-2022-26-
1231
+ forecast.pdf.
1232
+ Ivanov A, Dryden N, Ben-Nun T, Li S and Hoefler T (2021) Data
1233
+ Movement Is All You Need: A Case Study on Optimizing
1234
+ Transformers.
1235
+ In: Proceedings of Machine Learning and
1236
+ Systems 3 (MLSys 2021).
1237
+ Kasim MF, Watson-Parris D, Deaconu L, Oliver S, Hatfield P, Froula
1238
+ DH, Gregori G, Jarvis M, Khatiwala S, Korenaga J, Topp-
1239
+ Mugglestone J, Viezzer E and Vinko SM (2021) Building high
1240
+ accuracy emulators for scientific simulations with deep neural
1241
+ architecture search. Machine Learning: Science and Technology
1242
+ 3(1): 015013. DOI:10.1088/2632-2153/ac3ffa.
1243
+ Keryell R, Reyes R and Howes L (2015) Khronos SYCL for OpenCL:
1244
+ A Tutorial. In: Proceedings of the 3rd International Workshop
1245
+ on OpenCL, IWOCL ’15. New York, NY, USA: Association
1246
+ for Computing Machinery. ISBN 978-1-4503-3484-6. DOI:
1247
+ 10.1145/2791321.2791345.
1248
+ Keyes D (2022) Efficient Computation through Tuned Approx-
1249
+ imation.
1250
+ https://siag-sc.org/media/files/DK-
1251
+ slides.pdf.
1252
+ Kutzner C, P´all S, Fechner M, Esztermann A, de Groot BL and
1253
+ Grubm¨uller H (2019) More bang for your buck: Improved use
1254
+ of GPU nodes for GROMACS 2018. Journal of Computational
1255
+ Chemistry 40(27): 2418–2431. DOI:10.1002/jcc.26011.
1256
+ Larsen ES and McAllister D (2001) Fast Matrix Multiplies Using
1257
+ Graphics Hardware. In: Proceedings of the 2001 ACM/IEEE
1258
+ Conference on Supercomputing, SC ’01. New York, NY, USA:
1259
+ Association for Computing Machinery. ISBN 1-58113-293-X,
1260
+ p. 55. DOI:10.1145/582034.582089.
1261
+ Lattner C and Adve V (2004) LLVM: A compilation framework for
1262
+ lifelong program analysis & transformation. In: International
1263
+ Symposium on Code Generation and Optimization, 2004. CGO
1264
+ 2004. pp. 75–86. DOI:10.1109/CGO.2004.1281665.
1265
+ Lavin A, Krakauer D, Zenil H, Gottschlich J, Mattson T, Brehmer
1266
+ J, Anandkumar A, Choudry S, Rocki K, Baydin AG, Prunkl
1267
+ C, Paige B, Isayev O, Peterson E, McMahon PL, Macke J,
1268
+ Cranmer K, Zhang J, Wainwright H, Hanuka A, Veloso M,
1269
+ Assefa S, Zheng S and Pfeffer A (2021) Simulation Intelligence:
1270
+ Towards a New Generation of Scientific Methods. https:
1271
+ //arxiv.org/abs/2112.03235.
1272
+ Lee VW, Kim C, Chhugani J, Deisher M, Kim D, Nguyen AD, Satish
1273
+ N, Smelyanskiy M, Chennupaty S, Hammarlund P, Singhal R
1274
+ and Dubey P (2010) Debunking the 100X GPU vs. CPU Myth:
1275
+ An Evaluation of Throughput Computing on CPU and GPU.
1276
+ In: Proceedings of the 37th Annual International Symposium
1277
+ on Computer Architecture, ISCA ’10. New York, NY, USA:
1278
+ Association for Computing Machinery. ISBN 978-1-4503-0053-
1279
+ 7, pp. 451–460. DOI:10.1145/1815961.1816021.
1280
+ Li H, Berger DS, Novakovic S, Hsu L, Ernst D, Zardoshti P, Shah
1281
+ M, Rajadnya S, Lee S, Agarwal I, Hill MD, Fontoura M
1282
+ and Bianchini R (2022) Pond: CXL-Based Memory Pooling
1283
+ Systems for Cloud Platforms. https://arxiv.org/abs/
1284
+ 2203.00241.
1285
+ Prepared using sagej.cls
1286
+
1287
+ 12
1288
+ arXiv preprints
1289
+ Liu Z and Tegmark M (2021) Machine Learning Conservation Laws
1290
+ from Trajectories. Physical Review Letters 126(18): 180604.
1291
+ DOI:10.1103/PhysRevLett.126.180604.
1292
+ Macrotrends
1293
+ LLC
1294
+ (2022)
1295
+ NVIDIA
1296
+ Profit
1297
+ Margin
1298
+ 2010-
1299
+ 2022.
1300
+ https://www.macrotrends.net/stocks/
1301
+ charts/NVDA/nvidia/profit-margins.
1302
+ Matsuoka S (2008) The Rise of the Commodity Vectors. In: Palma
1303
+ JMLM, Amestoy PR, Dayd´e M, Mattoso M and Lopes JC
1304
+ (eds.) High Performance Computing for Computational Science
1305
+ - VECPAR 2008. Berlin, Heidelberg: Springer Berlin Heidelberg.
1306
+ ISBN 978-3-540-92859-1, pp. 53–62. DOI:10.1007/978-3-540-
1307
+ 92859-1 7.
1308
+ Menon H, Lam MO, Osei-Kuffuor D, Schordan M, Lloyd S,
1309
+ Mohror K and Hittinger J (2018) ADAPT: Algorithmic
1310
+ Differentiation Applied to Floating-Point Precision Tuning.
1311
+ In: SC18: International Conference for High Performance
1312
+ Computing, Networking, Storage and Analysis. pp. 614–626.
1313
+ DOI:10.1109/SC.2018.00051.
1314
+ NTT R&D (2020) What is the All-Photonics Network? https:
1315
+ //www.rd.ntt/e/iown/0002.html.
1316
+ Ootomo H and Yokota R (2022) Recovering Single Precision
1317
+ Accuracy from Tensor Cores While Surpassing the FP32
1318
+ Theoretical Peak Performance. International Journal of High
1319
+ Performance Computing Applications 36(4): 475–491. DOI:
1320
+ 10.1177/10943420221090256.
1321
+ Pathak J, Subramanian S, Harrington P, Raja S, Chattopadhyay
1322
+ A, Mardani M, Kurth T, Hall D, Li Z, Azizzadenesheli
1323
+ K, Hassanzadeh P, Kashinath K and Anandkumar A (2022)
1324
+ FourCastNet: A Global Data-driven High-resolution Weather
1325
+ Model using Adaptive Fourier Neural Operators.
1326
+ https:
1327
+ //arxiv.org/abs/2202.11214.
1328
+ Post E (1982) Real Programmers Don’t Use PASCAL. http://
1329
+ www.ee.ryerson.ca/˜elf/hack/realmen.html.
1330
+ Rasp S, Pritchard MS and Gentine P (2018) Deep learning to
1331
+ represent subgrid processes in climate models. Proceedings
1332
+ of the National Academy of Sciences - PNAS 115(39): 9684–
1333
+ 9689. DOI:10.1073/pnas.1810286115.
1334
+ Schulthess TC (2016) Exascale computing: Endgame or new
1335
+ beginning for climate modelling.
1336
+ https://www.ecmwf.
1337
+ int/sites/default/files/elibrary/2016/
1338
+ 16804-exascale-computing-endgame-or-new-
1339
+ beginning-climate-modelling.pdf.
1340
+ Shan Y, Lin W, Guo Z and Zhang Y (2022) Towards a Fully
1341
+ Disaggregated and Programmable Data Center. In: Proceedings
1342
+ of the 13th ACM SIGOPS Asia-Pacific Workshop on Systems,
1343
+ APSys ’22. New York, NY, USA: Association for Computing
1344
+ Machinery.
1345
+ ISBN 978-1-4503-9441-3, pp. 18–28.
1346
+ DOI:
1347
+ 10.1145/3546591.3547527.
1348
+ Shaw DE, Deneroff MM, Dror RO, Kuskin JS, Larson RH, Salmon
1349
+ JK, Young C, Batson B, Bowers KJ, Chao JC, Eastwood MP,
1350
+ Gagliardo J, Grossman JP, Ho CR, Ierardi DJ, Kolossv´ary I,
1351
+ Klepeis JL, Layman T, McLeavey C, Moraes MA, Mueller R,
1352
+ Priest EC, Shan Y, Spengler J, Theobald M, Towles B and Wang
1353
+ SC (2008) Anton, a Special-Purpose Machine for Molecular
1354
+ Dynamics Simulation. Commun. ACM 51(7): 91–97. DOI:
1355
+ 10.1145/1364782.1364802.
1356
+ Sorensen B, Norton A, Joseph E and Conway S (2019) Special
1357
+ Report for NASA: Exploring Options for a Bespoke
1358
+ Supercomputer Targeted for Weather and Climate Workloads.
1359
+ Technical Report HR4.0046.10.01.2019, Hyperion Research,
1360
+ LLC.
1361
+ https://hec.nasa.gov/news/reports/
1362
+ NASA-Hyperion-Research-Bespoke-Weather-
1363
+ HPC-Full-Report.pdf.
1364
+ Strohmaier E, Dongarra J, Simon H and Meuer M (2022) TOP500.
1365
+ http://www.top500.org/.
1366
+ Tesla, Inc (2021) Whitepaper: Tesla Dojo Technology – A Guide
1367
+ to Tesla’s Configurable Floating Point Formats & Arith-
1368
+ metic.
1369
+ https://tesla-cdn.thron.com/static/
1370
+ MXMU3S_tesla-dojo-technology_1WDVZN.pdf.
1371
+ White MJ (2022) Nvidia says falling GPU prices are ’a story
1372
+ of the past’.
1373
+ https://www.digitaltrends.com/
1374
+ computing/nvidia-says-falling-gpu-prices-
1375
+ are-over/.
1376
+ Wu CJ, Hsiao ST, Wang JY, Lin WH, Chang CW, Shao TL, Tung
1377
+ CH and Yu DCH (2021) Ultra High Power Cooling Solution for
1378
+ 3D-ICs. In: 2021 Symposium on VLSI Technology. pp. 1–2.
1379
+ Ziogas AN, Schneider T, Ben-Nun T, Calotoiu A, De Matteis
1380
+ T, de Fine Licht J, Lavarini L and Hoefler T (2021)
1381
+ Productivity, Portability, Performance: Data-Centric Python.
1382
+ In: Proceedings of the International Conference for High
1383
+ Performance Computing, Networking, Storage and Analysis,
1384
+ SC ’21. New York, NY, USA: Association for Computing
1385
+ Machinery. ISBN 978-1-4503-8442-1, p. 13. DOI:10.1145/
1386
+ 3458817.3476176.
1387
+ Prepared using sagej.cls
1388
+
G9E0T4oBgHgl3EQfhgFS/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
GNE1T4oBgHgl3EQfXAT2/content/tmp_files/2301.03123v1.pdf.txt ADDED
@@ -0,0 +1,1030 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.03123v1 [math.AG] 9 Jan 2023
2
+ Lax colimits of posets with structure sheaves:
3
+ applications to descent
4
+ Javier Sánchez González
5
+ Universidad de Salamanca, Department of Mathematics
6
+ javier14sg@usal.es
7
+ January 10, 2023
8
+ Abstract
9
+ We consider categories of posets with C-valued structure sheaves for
10
+ any category C and see how they possess poset-indexed lax colimits that
11
+ are both easy to describe and "weakly equivalent" to their ordinary
12
+ colimits in a certain sense.
13
+ We employ this construction to study
14
+ descent problems on schematic spaces—a particular scheme-like kind
15
+ of ringed poset—, proving a general Seifert-Van Kampen Theorem
16
+ for their étale fundamental group that recovers and generalizes the
17
+ homonym result for schemes to the topology of flat monomorphisms.
18
+ The techniques are general enough to consider their applications in
19
+ many other frameworks.
20
+ 1
21
+ Introduction
22
+ In the geometric world, categorical colimits can be thought of as a general
23
+ way of expressing gluing of spaces. For example, any scheme is the colimit,
24
+ in the category of locally ringed spaces, of the components of any of its affine
25
+ coverings. Furthermore, for any reasonable scheme, one might assume that
26
+ these colimits are indexed by posets: the nerve of the corresponding affine
27
+ covering with its redundancies removed. This idea is simply a generalization
28
+ to locally ringed spaces of the construction of finite models of topological
29
+ spaces, an old technique of McCord to study homotopy types [2].
30
+ To better understand the situation, one may consider these recollement
31
+ data of schemes as functors from some poset X, understood as a category,
32
+ which we may assume finite under mild compactness hypothesis on the
33
+ original scheme S; to the category of commutative rings with unit, i.e.
34
+ 1
35
+
36
+ X → CRing. It is also very classical that the category of such functors
37
+ coincides with the category of sheaves of rings on the poset X, understood
38
+ as a topological space, so we may study these collections of data as (non-
39
+ locally) ringed spaces themselves. A major advantage of this point of view
40
+ is that many sheaf-theoretic notions admit a simple description, like quasi-
41
+ coherent modules or the Čech resolution of an abelian sheaf; and in many
42
+ cases they coincide with their scheme-theoretic analogues on S. From these
43
+ observations, Sancho first axiomatized in [6] a—non-full—subcategory of
44
+ finite ringed posets that behaves suitably well with respect to quasi-coherent
45
+ modules and that, not only contains all finite models of (quasi-compact and
46
+ quasi-separated) schemes, but also that non-trivially generalizes them: the
47
+ category of (finite) schematic spaces, SchFin. After spending some time
48
+ with these objects, it is easy to convince oneself of their geometric interest
49
+ compared to similarly-purposed constructions such as those of simplicial
50
+ schemes. A brief summary of this non-standard approach will be provided
51
+ in Section 2.
52
+ Taking a step back from the previous discussion, we also note that
53
+ computing colimits in the category of schemes—or simply determining if
54
+ they exist—is a very general and difficult problem: even when existence is
55
+ guaranteed—for example, colimits via open immersions—, obtaining explicit
56
+ expressions for them is no easy task. The situation is not much better if
57
+ we attempt to compute colimits of schematic spaces as described in the
58
+ previous paragraph; however, due to their combinatorial nature, there is
59
+ an alternative: given a poset-indexed "datum" U : P → SchFinop, we may
60
+ construct a space, which we call the cylinder of U, by simply turning the set
61
+
62
+ p∈P |U(p)| into a poset with the structure inherited from the underlying
63
+ posets of each |U(p)| and the transition morphisms |U(p)| → |U(q)| for q ≤ p.
64
+ If we endow the resulting poset with the structure sheaf—which is just a
65
+ functor—induced by the structure sheaves of each U(p), we obtain a ringed
66
+ poset—which will be another schematic space under certain conditions—that
67
+ represents—will be weakly equivalent to—the desired colimit in a precise
68
+ sense and which has been computed without performing any complicated
69
+ categorical operations on either commutative rings or posets.
70
+ In this paper, we will study this cylinder space in its most general
71
+ formulation, replacing CRing for any category C. The category of finite
72
+ posets with a C-values structure sheaf will be constructed and called the
73
+ category of C-data, denoted C -data. If C = pos is the category of posets,
74
+ it turns out that the cylinder space is just an incarnation of a poset-indexed
75
+ lax colimit, understanding pos as a strict 2-category in the natural way. We
76
+ prove this and see how it generalizes to posets admiting structure sheaves
77
+ 2
78
+
79
+ with values in C, detailing the different 2-categorical structures that one may
80
+ consider and how they interact with each other. For this purpose, we will
81
+ need to consider D -data with D being a strict 2-category, which also proves
82
+ useful in other applications. Our language of choice will keep things analogue
83
+ to the category of ringed posets that we use as a reference.
84
+ These constructions, of C -data and lax colimits on it, are general enough
85
+ to model many descent problems, which appear in a natural and functorial
86
+ way.
87
+ For the sake of keeping the discussion focused, we will specialize
88
+ everything to the previously-discussed schematic case: we will characterize
89
+ when a cylinder of schematic spaces remains schematic and will apply such
90
+ characterization to give a very general descent Theorem for data on schematic
91
+ spaces, which even admits a topos-theoretic interpretation—only sketched
92
+ here due to space limitations—.
93
+ We will see some examples, with the
94
+ main one being the Seifert-Van Kampen Theorem for the étale fundamental
95
+ group of schematic spaces, as constructed in [4].
96
+ It is worth noting that
97
+ a variation of the homonym result for schemes and their topology of flat
98
+ monomorphisms—rather than Zariski or étale—follows purely from the formal
99
+ descent result developed in this paper and the classical case, which exemplifies
100
+ how these techniques automatically extend any "reasonable" Zariski-local
101
+ statement to the aforementioned topology.
102
+ 2
103
+ Motivation: schematic spaces
104
+ Let us give a rather informal introduction to the objects of study for the
105
+ applications, which were the original motivation to develop the theory of
106
+ C -data that we will introduce in the following sections. In prose: schematic
107
+ spaces arise as the largest subcategory of ringed finite posets that behaves
108
+ "like quasi-compact and quasi-separated (qc-qs)" schemes with respect to
109
+ categories of quasi-coherent sheaves.
110
+ The basic example comes from the
111
+ construction of finite models of schemes, see [7], which is a generalization
112
+ of an earlier topological technique, see [2]: given a qc-qs scheme S and a
113
+ finite covering {Ui} of S, one may define a poset X as the T0-fication of the
114
+ topology generated by the covering. Explicitly, if for any s ∈ S we denote
115
+ U s = ∩s∈UiUi, one sets X = S/ ∼ with s ∼ s′ whenever U s = U s′ and
116
+ [s] ≤ [s′] if and only if U s′ ⊆ U s. The result is a morphism of ringed spaces
117
+ π: S → (X, π∗OS).
118
+ If the covering is chosen so that the U s are affine, π induces an adjoint
119
+ equivalence (π∗ ⊣ π∗): Qcoh(S) ∼
120
+ → Qcoh(X), so S can be studied from X.
121
+ 3
122
+
123
+ Schematic spaces were first introduced in [6] and studied in [5], [8] or [4].
124
+ In [9] one can find an extensive compilation of characterizations, and the
125
+ author of this paper has expanded upon this and exhaustively explored their
126
+ role in algebraic geometry in his PhD thesis, which includes the contents of
127
+ this paper and more to come. Morally speaking, we recommend to think
128
+ of a schematic space X as a model or "structured descent data" of the
129
+ locally affine locally ringed space Spec(X) = colimx∈X Spec(OX,x), but the
130
+ schematic condition forces such discrete incarnation X to be "nice enough"
131
+ to reflect the geometry of Spec(X) to some extent. One shows that these
132
+ Spec(X) have to be locally affine in the topology of flat monomorphisms of
133
+ affine schemes and, thus, contain all qc-qs schemes; but schematic spaces do
134
+ not model, for example, algebraic spaces, for which the associated locally
135
+ ringed space does not preserve enough useful algebraic information. There
136
+ are also "geometric" arguments for considering ringed posets as the basis
137
+ for our combinatorial models over other, more classical, alternatives like
138
+ simplicial schemes.
139
+ While perhaps not the most enlightening approach, it will be convenient
140
+ for our purposes to consider the following definitions. Assume for simplicity
141
+ that all stalk rings of our ringed posets are Noetherian.
142
+ Definition 2.1. A finite ringed poset X is a (finite) schematic space if
143
+ • For any x ≤ y, the morphism rxy : OX,x → OX,y is flat.
144
+ • For any t ≤ x, y, the morphism
145
+ OX,x ⊗OX,t OX,y →
146
+
147
+ z≥x,y
148
+ OX,z
149
+ is faithfully flat.
150
+ A morphism f : X → Y between schematic spaces is schematic if
151
+ • For any x ∈ X and y ≥ f(x), the morphism
152
+ OX,x ⊗OY,f(x) OY,y →
153
+
154
+ z∈Ux∩f−1(Uy)
155
+ OX,z
156
+ induces a surjection between the prime spectra.
157
+ Remark 2.2. A simple descent argument for faithfully flat morphisms shows
158
+ that OX,x ⊗OX,y OX,y ≃ OX(Ux ∩Uy) for all t ≤ x, y. If x = y, this condition
159
+ implies that the restriction morphisms rxy of any schematic space are flat
160
+ epimorphisms of rings, hence local isomorphisms.
161
+ 4
162
+
163
+ Let SchFin denote the category of schematic spaces and morphisms.
164
+ All ringed posets and morphisms will be considered schematic unless stated
165
+ otherwise. The Spec construction outlined in the introduction of this section
166
+ defines a functor to the category of locally ringed spaces which is neither full
167
+ or faithful:
168
+ Spec: SchFin → LRS
169
+ X �→ Spec(X) := colimx∈X Spec(OX,x).
170
+ It can be shown that the schematic category has finite fibered products and
171
+ that are preserved by both the forgetful to CRing -data and by Spec.
172
+ Remark 2.3. Heuristically, the restriction maps of X being flat epimorphisms
173
+ implies that the information in X can be recovered from Spec(X). The other
174
+ schematicity conditions can be shown to be equivalent to the existence of a
175
+ certain map πX : Spec(X) → X, i.e. to X being essentially a "finite model"
176
+ of Spec(X) in a topological sense.
177
+ Definition 2.4. A morphism f : X → Y is said to be a qc-isomorphism if
178
+ Spec(f) is an isomorphism.
179
+ The class of qc-isomorphisms is a multiplicative system of arrows in
180
+ SchFin that is maximal by definition, so the corresponding localization—
181
+ Verdier quotient—defines a faithful—but not full—functor
182
+ Spec: SchFinqc → LRS.
183
+ To study properties P of schematic spaces we will ask for two requisites:
184
+ • A "rigorous" requisite: that P factors through the localization, i.e.
185
+ that any representative of the qc-isomorphism class of a space or arrow
186
+ determines is the whole class verifies the property or not.
187
+ In other
188
+ words, P is geometric.
189
+ • A "moral" requisite:
190
+ P can be studied in terms of finite models,
191
+ without applying the Spec functor. In other words, P is discretizable.
192
+ In certain cases, one can "rigidify" poorly-behaved properties by studying
193
+ them on certain reflective subcategories of SchFin that induce equivalences
194
+ after localizing by qc-isomorphisms. As an example of this, see the discussion
195
+ about connectedness in [4]. In this paper we will tacitly assume that all our
196
+ definitions work in this nice way, but let it be known that more technical
197
+ considerations are needed for a full exposition—and that is the reason why we
198
+ employ quotation marks so often to highlight seemingly ordinary notions—.
199
+ 5
200
+
201
+ Finally, the main result in [4]—which, with enough work, can be written
202
+ in much more geometric and elegant terms that the ones presented there—is
203
+ concerned with the existence of a Galois category of "finite étale covers" for
204
+ any "connected" schematic space X that, when X models a qc-qs scheme S,
205
+ is naturally equivalent to the homonym Galois category of S.
206
+ Theorem 2.5. [4] For a schematic space X and a schematic morphism
207
+ x: Spec(Ω) → X with Ω an algebraically closed field—a geometric point—,
208
+ there exists a category RÉt(X) and a functor Fibx : RÉt(X) → FinSet
209
+ such that, when X is "connected", the pair (RÉt(X), Fibx) is a Galois
210
+ category. We denote its fundamental group by πet
211
+ 1 (X, x). If S = Spec(X)
212
+ is a scheme, the Spec functor induces a equivalence of Galois categories
213
+ (RÉt(X), Fibx) ≃ (FEt(S), FibSpec(x)), where FEt(S) is the category of
214
+ finite étale covers of S, and thus πet
215
+ 1 (X, x) ≃ πet
216
+ 1 (S, Spec(x)).
217
+ Remark 2.6. In [4] we also showed that qc-isomorphic spaces have equivalent
218
+ Galois categories of finite étale covers: the construction is geometric.
219
+ Of course, for a general schematic space X, one can consider the set of
220
+ all its geometric points and define the étale fundamental (Stone) groupoid
221
+ Πet
222
+ 1 (X).
223
+ In its general version, the Galois Theorem states that the fiber
224
+ functors induce an equivalence of categories
225
+ RÉt(X) ≃ [Πet
226
+ 1 (X), FinSet] ≡ Πet
227
+ 1 (X)-FinSet
228
+ where the action of this groupoid is continuous. As always, this is just a
229
+ particular case of more general topos-theoretic results.
230
+ 2.1
231
+ The topology of flat immersions
232
+ We begin by introducing the natural (pre)topology on SchFin.
233
+ Definition 2.7. Let f : X → Y be a schematic morphism. We say that f
234
+ is flat if f ♯
235
+ x: OY,f(x) → OX,x is flat for all x ∈ X. Such flat morphism is a
236
+ flat immersion if its diagonal ∆f : X → X ×Y X is a qc-isomorphism. A flat
237
+ morphism f is faithfully flat if Spec(f) is surjective.
238
+ These three types of maps can be characterized in terms of the adjoint
239
+ pair (f ∗, f∗) for quasi-coherent sheaves. We remark that a flat immersion
240
+ is, by definition, a flat monomorphism in SchFinqc; and Spec(f) for such
241
+ f is a flat monomorphism of locally ringed spaces. One can show that qc-
242
+ isomorphisms are exactly faithfully flat immersions.
243
+ 6
244
+
245
+ Remark 2.8. It can be shown that a morphism Ux → Uy is a flat immersion
246
+ if and only if OY,y → OX,x is a flat epimorphism of rings. Since schematic
247
+ spaces have flat epimorphisms of rings as restriction maps, the restriction
248
+ morphisms between their basic open subsets are flat immersions—actually,
249
+ between all their open subsets—.
250
+ In other words, schematic spaces are
251
+ colimits of (certain) affine schematic spaces via flat immersions. This class
252
+ of morphisms was first shown to be important in the context of descent
253
+ problems in [3].
254
+ Lemma 2.9. If f : X → Y is a flat immersions, f ♯
255
+ x : OY,f(x) → OX,x are flat
256
+ epimorphisms of rings for all x ∈ X.
257
+ Proof. They are flat by definition and the condition on the diagonal trivially
258
+ translates to OY,f(x) ⊗OX,x OY,f(x) → OY,f(x) being an isomorphism.
259
+ Recall that an open immersion of schemes is a flat monomorphism (locally)
260
+ of finite presentation. As such, flat immersions are like "open immersions",
261
+ but without the finite presentation condition. The reader might notice the
262
+ analogy with the étale and pro-étale topologies for schemes. This justifies
263
+ the following notation:
264
+ Definition 2.10. Let X be a schematic space. We define XwZar to be the site
265
+ of flat immersions with target X, whose covers are given by finite and jointly
266
+ faithfully flat families of flat immersions. Similarly, we define SchFinwZar
267
+ to be the "big" site of flat immersions.
268
+ These sites present a number of interesting pathologies that we will
269
+ describe in more detail in future papers. We shall remark a few of them:
270
+ • The category XwZar is not small, only its localization (XwZar)qc. Each
271
+ qc-isomorphism class of open immersions is identified with a subset of
272
+ Spec(X), but before localization, the collection of representatives is as
273
+ large as the entire class of finite posets.
274
+ • Since qc-isomorphisms are both flat immersions and covers, yet they
275
+ are not isomorphisms, a standard descent argument shows that sheaves
276
+ map qc-isomorphisms to isomorphisms—Category theorists sometimes
277
+ call morphisms with such property local isomorphisms—. In particular,
278
+ functors of points are not sheaves, because they determine spaces up
279
+ to isomorphism, so the site XwZar and its bigger analogue are not
280
+ 7
281
+
282
+ subcanonical. However, one can show that, for any Y ∈ XwZar and
283
+ sheaf F ∈ XwZar, there are natural bijections
284
+ HomPSh(XwZar)(F, HomXwZar(−, Y )) ≃
285
+ ≃ HomSh(XwZar)(F, Hom(XwZar)qc(−, Y ))
286
+ In other words, the functor of points in the localization satisfies the
287
+ universal property of sheafification.
288
+ • We have avoided talking about sheafification in the previous points
289
+ because, due to potential size issues, we cannot guarantee that such
290
+ functor exists in XwZar—this is related to the inability to find bounds
291
+ for refinements of covers, which may lead to pathologies, as it happens
292
+ with the fpqc topology of schemes, see [12, Theorem 5.5]—; but the
293
+ good news is that it does exist (XwZar)qc. I.e. we can sheafify presheaves
294
+ that factor through qc-isomorphism, which will be enough in all natural
295
+ situations.
296
+ • Endowing XwZar with the natural sheaf of rings, it is possible to show
297
+ that Qcoh(XwZar) ≃ Qcoh(X).
298
+ As it happens with schemes and open immersions, it is obvious that if
299
+ X = {Xi}i∈I is a diagram of schematic spaces and the transition morphisms
300
+ Xi → Xj (for any i → j) are flat immersions, taking colimi Xi in the
301
+ category of ringed posets yields Spec(colimi Xi) = colimi Spec(Xi) and the
302
+ resulting space is a gluing of affine schemes via flat monomorphisms of affine
303
+ schemes. The problem is that, in general, it is very difficult to determine
304
+ if colimi Xi is schematic or not, due to the combinatorial nature of the
305
+ definition of schematicity and the surprisingly subtle description of colimits
306
+ of finite posets (see [1, Proposition 2.4]).
307
+ Our solution will be defining an object "equivalent" to colimi Xi in the
308
+ sense of representing the same locally ringed space, but whose combinatorial
309
+ nature is elementary. This will be done in Sections 5 and 6. The result will
310
+ be called cylinder space, denoted Cyl(X).
311
+ This construction is central in the theory of schematic spaces will have
312
+ applications that are beyond our purposes here, but the goal for this paper is
313
+ to study descent properties with respect to the topology of flat immersions.
314
+ For instance, let us consider the case of the étale fundamental groupoid. It
315
+ clearly defines a functor
316
+ Πét
317
+ 1 : SchFin → GpdStone
318
+ 8
319
+
320
+ valued in the strict 2-category of Stone groupoids. Proving the Seifert-Van
321
+ Kampen Theorem in its general form—for the topology of flat immersions—
322
+ essentially amounts to saying that Πét
323
+ 1 maps colimits to 2-colimits.
324
+ This
325
+ will be the same as saying that (Πét
326
+ 1 )op is a 2-sheaf —thus it maps qc-
327
+ isomorphisms to equivalences—.
328
+ By the properties of these sites, this is
329
+ equivalent to proving that it maps objects "qc-equivalent to colimits"—our
330
+ cylinder spaces—to 2-colimits. However, our abstract descent result for the
331
+ topology of flat immersions and cylinders will show that it is enough to prove
332
+ that it is a 2-sheaf in the combinatorial topology. Such statement amounts
333
+ to showing that Πét
334
+ 1 maps a very specific kind of cylinders to 2-colimits; and
335
+ in some particular cases, this will even be formal.
336
+ 3
337
+ Categories of C -data
338
+ Without further ado, let C be a 1-category and pos be the category of finite
339
+ posets—or arbitrary posets, being careful in that case with set-theoretic size
340
+ considerations—. For a given poset X and x ∈ X, let Ux = {x′ ≥ x} denote
341
+ the minimal open neighborhood of the point X. The following is well-known:
342
+ Lemma 3.1. If C has finite limits and X ∈ pos, there is an equivalence
343
+ Sh(X, C) ≃ [X, C]
344
+ between the categories of C-valued sheaves on X and functors X → C.
345
+ Proof. Each sheaf gives a functor defined by its stalks—sections at the
346
+ minimal open neighborhoods—and restrictions morphisms.
347
+ The converse
348
+ follows from the sheaf condition and the fact that the {Ux}x∈X are a basis
349
+ for the topology, so for any open U ⊆ X and functor F : X → C, one defines
350
+ its "sections" on U as F(U) = limx∈U F(x).
351
+ Now let us consider the functor to the 1-category of categories—big
352
+ enough so that C ∈ Cat—
353
+ C -data: pos → Cat
354
+ X �→ [X, C]
355
+ f �→ f −1.
356
+ Definition 3.2. For any C, the cateory of C -data is the fibered category
357
+ over C defined by the Grothendieck construction applied to the previous
358
+ functor. Explicitly:
359
+ 9
360
+
361
+ • Ob(C -data) = {F
362
+ not
363
+ ≡ (X, F) : X ∈ pos and F ∈ [X, C]};
364
+ • HomC -data((X, F), (Y, G)) = {f : X → Y and f ♯: f −1G → F};
365
+ • | − |: C -data → pos is the "underlying poset" structure functor.
366
+ Notation 3.3. We will usually denote F
367
+ not
368
+ ≡ (X, F) and X = |F|, unless
369
+ C = CRing is the category of commutative rings, in which case C -data
370
+ is the category of ringed posets and we will keep the traditional notation
371
+ (X, OX).
372
+ Furthermore, for any F and x ≤ y ∈ |F|, we will denote its
373
+ "restriction morphisms" by Fxy : F(x) → F(y).
374
+ Remark 3.4. Note that the construction of C -data is functorial on the
375
+ category: if Φ: C → D is a functor, we have Φ∗ : C -data → D -data induced
376
+ by post-composition.
377
+ This category comes with a natural inclusion functor
378
+ iC : Cop → C -data
379
+ c �→ (⋆, c)
380
+ analogue to the "diagonal inclusion" in categories of diagrams of a fixed
381
+ shape. Due to the choice of ⋆ as the final object in pos, we have the following:
382
+ Lemma 3.5. If C has finite limits (resp. colimits), the functor iC has a left
383
+ (resp. right) adjoint Γ ≡ ΓC : C -data → Cop (resp. L) called the sections
384
+ (resp. cosections) functor. Explicitly, Γ(F) = lim F (resp. Γ(F) = colim F).
385
+ Remark 3.6. The terminology of Lemma 3.5 comes from the equivalence
386
+ of Lemma 3.1. Of course, one may assume no hypothesis on C and define
387
+ sections via Yoneda at the level of [Cop, Set] -data, only to ask if these
388
+ "sheaves of sections" are representable on a case-by-case basis. One may
389
+ also interpret sections via projections to the terminal poset π: X → ⋆ by
390
+ constructing π∗ right adjoint to π−1.
391
+ Example 3.7 (Locally representable functors). As a simple application of
392
+ this terminology, we will give a "structured" interpretation of the concept of
393
+ locally representable functor. Indeed, let Y : C → [Cop, Set] be the Yoneda
394
+ embedding for C and Y∗ : C -data → [Cop, Set] -data the—fully faithful—
395
+ induced functor. One may think of an object in the image of Y∗ as a "locally
396
+ representable functor".
397
+ Note that, if C has finite limits, the sections of
398
+ such an object are representable by the sections of the original C-datum.
399
+ Additionally, we shall consider the Yoneda embedding for C -data, that is
400
+ Y ′ : C -data → [C -dataop, Set]. At this stage, we define a third functor
401
+ D: [Cop, Set] -data → [C -dataop, Set]
402
+ X �→ Hom[Cop,Set] -data(Y∗(−), X)
403
+ 10
404
+
405
+ such that D ◦ Y∗ = Y ′—since Y∗ is fully faithful—. We leave as an exercise
406
+ to the reader checking that D is fully faithful itself—recall that categories
407
+ of presheaves are compactly generated by their representable functors—. In
408
+ particular, if X is such that D(X) is representable by some F ∈ C -data,
409
+ one has that Y∗(F) ≃ X, in other words, "representing each X(p) by some
410
+ Fp ∈ C for each p ∈ |X| in a compatible way is equivalent to representing X
411
+ by a C-datum F with F(p) = Fp".
412
+ One of the main advantages of considering C -data over categories of
413
+ diagrams of fixed shape is that it inherits the natural 2-categorical structure
414
+ of pos.
415
+ More precisely, recall that pos is a strict 2-category with its 2-
416
+ morphisms being, for each X, Y ∈ pos,
417
+ HomHompos(X,Y )(f, g) =
418
+
419
+ ⋆ if f ≤ g
420
+ ∅ otherwise.
421
+ If f, g: F → G are morphisms in C -data and |f| ≤ |g| in pos, we
422
+ have a natural transformation rfg : f −1G → g−1G given, at each x ∈ |F|,
423
+ by the restriction morphisms of G. We simply ask this arrow to induce a
424
+ commutative triangle, i.e. we define our 2-morphisms to be:
425
+ HomHomC -data(F,G)(f, g) =
426
+
427
+ ⋆ if f ≤ g and g♯ = rfg ◦ f ♯
428
+ ∅ otherwise.
429
+ We note that this structure generalizes the partial order defined in [7] to
430
+ study naif homotopy types of ringed posets. We also remark that C -data is
431
+ actually a pos-enriched category.
432
+ It is easy to check that, if C has finite limits (resp. colimits), then C -data
433
+ has finite colimits (resp. limits), described in an analogous way as in the
434
+ category of ringed posets (or spaces). To approach descent problems, we are
435
+ interested in computing colimits of C -data, or in other words, describing
436
+ the sections functor of the inclusion
437
+ iC -dataop : C -data → (C -data)op -data;
438
+ but it turns out that we can obtain, up to a certain to-be-introduced notion
439
+ of weak equivalence, a more explicit description of these colimits that does
440
+ not require us to perform any 1-categorical operations on either pos or C.
441
+ We will call this construction the "cylinder functor". The context in which
442
+ it arises naturally employs the 2-categorical structure of C -data, hence, for
443
+ this and other reasons, we shall devote the next section to briefly describe
444
+ D -data for D a strict 2-category.
445
+ 11
446
+
447
+ 4
448
+ The 2-categorical case
449
+ Let D be a strict 2-category and endow posets with the trivial 2-categorical
450
+ structure. Among other possibilities, we shall consider the categories of
451
+ • pseudofunctors X → D and pseudonatural transformations, [X, D];
452
+ • pseudofunctors X → D and lax natural transformations, [X, D]Lax.
453
+ The Grothendieck construction for each of these possibilities now yields,
454
+ as in Definition 3.2, two different 1-categories, denoted for emphasis as
455
+ D -data and D -dataLax respectively. In both cases, their objects are pairs
456
+ (X, F) of a finite poset and a pseudofunctor, with the only difference being
457
+ that a morphism (f, f ♯): F → G is defined by a pseudonatural transformation
458
+ f ♯ when considering it in D -data and by a Lax natural transformation
459
+ when considering it in D -dataLax.
460
+ Note that, if D is pos-enriched—as
461
+ is the case when D = C -data for a 1-category C—, defining such a lax
462
+ natural transformation amounts to giving, for each p ≤ q ∈ |F|, 1-morphisms
463
+ αp : F(p) → G(f(p)) such that
464
+ Gf(p)f(q) ◦ αp ≤ αq ◦ Fpq,
465
+ rather than asking for strict equality.
466
+ Furthermore, in order to turn the inclusion functors
467
+ iD : Dop → D -data,
468
+ iLax
469
+ D
470
+ : Dop → D -dataLax,
471
+ D -data → D -dataLax
472
+ into pseudofunctors, we need to endow both categories of data with the same
473
+ lax 2-categorical structure, whose 2-morphisms are:
474
+ HomHomD -data(F,G)(f, g) =
475
+
476
+ η: rfg ◦ g♯ → f ♯ when |f| ≤ |g|
477
+ ∅ otherwise.
478
+ Again, if D is pos-enriched, giving this lax natural transformation amounts
479
+ to asking that, for |f| ≤ |g|, we only have
480
+ rfg ◦ f ♯ ≤ g♯.
481
+ With this structure, iD and iLax
482
+ D
483
+ are pseudofunctors that map any 2-morphism
484
+ η: s → t in D to the 2-morphism defined by the natural transformation η,
485
+ since riD(s)iD(t) is the identity and the underlying posets are singletons.
486
+ Finally, as in the 1-categorical case, and almost by definition, we have:
487
+ 12
488
+
489
+ Proposition 4.1. The left 2-adjoint of iD (resp. iLax
490
+ D ) is, if it exists, the
491
+ pseudolimit (resp.
492
+ lax limit) of the structure pseudofunctor.
493
+ We call it
494
+ sections (resp. lax sections) functor and denote it by Γ ≡ ΓD (resp. LaxΓ).
495
+ 5
496
+ The Cylinder Functor
497
+ Now we construct the lax sections functor for the 2-category D = C -dataop
498
+ with C a 1-category, that is, the lax colimit functor in C -data. We begin
499
+ with the explicit description:
500
+ Definition 5.1. For any X ∈ (C -data)op -data, we define the cylinder of
501
+ X as the C-datum Cyl(X) such that:
502
+ • As a set, |Cyl(X)| = �
503
+ p∈|X |X(p)|. We endow it with the partial order
504
+ induced by those of |X(p)| and setting that xp ≤ yq—with xp ∈ |X(p)|
505
+ and yq ∈ |X(q)|—whenever xp ≤ Xpq(yq).
506
+ • The structure functor is Cyl(X)(xp) = X(p)(xp) on objects, and its
507
+ restriction morphisms are given by X(p)xpx′p in each X(p) and by
508
+ (Xpq)♯
509
+ yq : Cyl(X)(yq) → Cyl(X)(Xpq(yq))
510
+ when p ≤ q.
511
+ It is easy to check that this construction is functorial, thus we have
512
+ Cyl: (C -data)op -data → C -data .
513
+ Lemma 5.2. If C = ⋆, hence C -data = pos, the functor Cyl coincides up to
514
+ natural isomorphism with the lax sections functor of the inclusion i⋆ -dataop.
515
+ In other words, pos has pos-indexed lax colimits, described by Cyl.
516
+ Proof. We will check that, for any Y ∈ C -data and X ∈ (C -data)op -data,
517
+ there are functorial isomorphisms of categories
518
+ Hompos(Cyl(X), Y ) ∼
519
+ → Homposop -dataLax(X, Y ).
520
+ Since Y ≡ iC -dataop(Y ) has the terminal category ⋆ as underlying poset, there
521
+ is an isomorphism Homposop -dataLax(X, Y ) ≃ Hom[X,pos]Lax(X, Y ), where
522
+ X ≡ |X|. Now, given a morphism f : Cyl(X) → Y , we have, by construction,
523
+ a family of morphisms {fp : X(p) → Y }p∈X that verify fp◦Xpq ≤ fq for p ≥ q.
524
+ This is exactly the information that defines a lax natural transformation
525
+ X → Y : giving, for each p ∈ X, an arrow X(p) → Y (p) = Y in pos and,
526
+ 13
527
+
528
+ for each p ≤ q, a 2-morphism on the corresponding diagram, which amounts
529
+ to asking that the previous inequalities hold. The converse follows from the
530
+ same argument: given g: X → Y , the g♯
531
+ p are exactly the morphisms fp.
532
+ Finally, saying that two morphisms f, g: Cyl(X) → Y verify f ≤ g is
533
+ just saying that fp ≤ gq for all p ∈ X—with the previous notations—. This
534
+ is precisely the notion of 2-morphism in posop -dataLax.
535
+ Proposition 5.3. For any category C, the functor Cyl coincides up to
536
+ natural isomorphism with the lax sections functor of the inclusion iC -dataop.
537
+ I.e. C -data has pos-indexed lax colimits and they are described by Cyl.
538
+ Proof. Again, we check that for Y ∈ C -data and X ∈ (C -data)op -data,
539
+ there are functorial isomorphisms of categories
540
+ HomC -data(Cyl(X), Y ) ∼
541
+ → Hom(C -data)op -dataLax(X, Y ).
542
+ The topological part of the proof has been taken care of in Lemma 5.2, so we
543
+ only need to check that such isomorphism extends to the level of C-valued
544
+ functors.
545
+ Given f : Cyl(X) → Y , using the same notations as in the aforementioned
546
+ Lemma, we have morphisms fp such that fp◦Xpq ≤ fq : X(q) → Y topologically.
547
+ This is a 2-morphism of C -data because, for each yq ∈ |X(q)|,
548
+ Y(fp◦Xpq)(yq) ◦ (fp ◦ Xpq)♯
549
+ yq = Y(fp◦Xpq)(yq) ◦ (Xpq)♯
550
+ yq ◦ (fp)♯
551
+ Xpq(yq) = (fq)♯
552
+ yq;
553
+ but by the definition of Cyl(X) and f, for all p ≤ q and xp = Xpq(yq),
554
+ Cyl(X)xpyq ◦ (fp)♯
555
+ xp = (fq)♯
556
+ yq,
557
+ (5.1)
558
+ where Cyl(X)xpyq = (Xpq)♯
559
+ yq, as desired. The converse follows from the same
560
+ relations.
561
+ At the level of morphisms, if we have arrows f, g: Cyl(X) → Y with
562
+ f ≤ g in C -data, they verify |f| ≤ |g| in pos and, for all xp ∈ Cyl(X),
563
+ g♯
564
+ xp ◦ Yf(xp)g(xp) = f ♯
565
+ xp.
566
+ (5.2)
567
+ If {fq : X(p) → Y } and {gp : X(p) → Y } are their corresponding families of
568
+ morphisms in [X, (C -data)op]Lax, there only remains to check that fp ≤ gp
569
+ for all p ∈ X. Once again, |fp| ≤ |gp| by Lemma 5.2, so we complete the
570
+ proof by remarking that, for each xp ∈ |X(p)|, the fact that the equation 5.2
571
+ holds is equivalent to fp ≤ gp in C -data.
572
+ 14
573
+
574
+ Note that C -data is actually a pos-enriched category, hence the universal
575
+ property of Cyl is necessarily given by an isomorphism of categories, rather
576
+ than an equivalence. This means that, provided that colimits of C -data also
577
+ exist, there is a natural transformation to the 1-categorical sections:
578
+ Cyl → ΓC -dataop.
579
+ One can make a case for this natural transformation being a "weak
580
+ equivalence" relative to certain descent problems for information codified
581
+ in a given collection of C-datum. We will not introduce the full terminology
582
+ here, since that would be a technical exercise far past our aim, but Sections
583
+ 2 and 6 will put us in a particular case that hints towards this direction.
584
+ Example 5.4. A very important remark is that, not only Cyl ◦ iC -dataop is
585
+ trivially the identity, but that every C-datum is the "cilinder of its points".
586
+ More precisely, for any C, there is a second "obvious" inclusion functor given
587
+ by post-composition with iop
588
+ C :
589
+ (iop
590
+ C )∗ : C -data → (C -data)op -data;
591
+ such that (iop
592
+ C )∗(F) has the same underlying poset as F, but we "replace"
593
+ each F(p) by the constant datum (⋆, F(p)). It is obvious that Cyl ◦ (iop
594
+ C )∗ is
595
+ also the identity. Furthermore, there is a natural transformation
596
+ ηC : (iop
597
+ C )∗ → iC -dataop
598
+ given by the natural projections to the terminal poset and identities in C,
599
+ which will be relevant when dealing with descent problems.
600
+ Proposition 5.5. The functor Cyl commutes with finite fibered products.
601
+ Proof. Exercise to the reader: it follows from the explicit construction.
602
+ 6
603
+ The schematic cylinder
604
+ The schematic category introduced in Section 2 is a non-full subcategory
605
+ of CRing -data, where CRing denotes the category of commutative rings
606
+ with unit. In particular, the cylinder functor restricts to
607
+ Cyl: SchFinop -data → CRing -data .
608
+ The next few pages are devoted to characterizing SchFinop-data whose
609
+ cylinder spaces are schematic. The first justification is that such lax colimit
610
+ represents up to "qc-isomorphism"—see discussion after the next Lemma—
611
+ the same locally ringed space:
612
+ 15
613
+
614
+ Lemma 6.1. Given X ∈ SchFinop -data, the natural morphism of ringed
615
+ spaces Cyl(X) → Γ(X) induces an isomorphism Spec(Cyl(X)) ∼
616
+ → Spec(Γ(X)).
617
+ Proof. This follows from the fact that colimits commute with colimits.
618
+ We would like to say that Cyl(X) → Γ(X) is a qc-isomorphism, but
619
+ note that we have not checked—and will not check—whether or not Γ(X) is
620
+ schematic. However, it will be sufficient to check schematicity of Cyl(X) for
621
+ our applications—and crucial, since we would not be able to guarantee the
622
+ stability under qc-isomorphisms of the properties and constructions we are
623
+ interested in dealing with otherwise—.
624
+ Definition 6.2. A ringed poset X is said to be pseudo-schematic if it has
625
+ flat epimorphisms of rings as restriction maps.
626
+ Definition 6.3. A ringed poset X is Mod-affine if π: X → (⋆, OX(X))
627
+ induces an adjoint equivalence (π∗ ⊣ π∗): Qcoh(X) → Mod(OX(X)). We
628
+ say that X is affine if it is schematic and Mod-affine.
629
+ Example 6.4. Any ringed poset with a minimum X = Ux is Mod-affine.
630
+ Remark 6.5. If X is pseudo-schematic, Qcoh(X) is a Grothendieck abelian
631
+ category. In particular, if X is also Mod-affine, π∗ is exact.
632
+ Lemma 6.6. If X is pseudo-schematic and Mod-affine, the natural morphism
633
+ OX(X) → �
634
+ x∈X OX,x is faithfully flat.
635
+ Proof. It suffices to see that �
636
+ x∈X Spec(OX,x) → Spec(OX(X)) is surjective.
637
+ Given a prime p ⊆ OX(X) with non-zero residue field κ(p), the equivalence
638
+ gives a non-zero module π∗κ(p) ̸= 0, thus there is some x ∈ X such that
639
+ (π∗κ(p))x ≃ κ(p)⊗OX(X) OX,x ̸= 0. Geometrically, this means that the fiber
640
+ of p via Spec(OX,x) → Spec(OX(X)) is non-empty, so we win.
641
+ Definition 6.7. A morphism of ringed spaces f : X → Y between pseudo-
642
+ schematic spaces will be called a qc-isomorphism if f −1(Uy) is Mod-affine
643
+ for all y ∈ Y and f♯: OY → f∗OX is an isomorphism.
644
+ Example 6.8. Any ringed poset with a minimum X = Ux is qc-isomorphic
645
+ to (⋆, OX,x) via the natural projection.
646
+ In the schematic category, Definition 6.7 restricts to the usual one. In
647
+ this generality, we cannot even guarantee that the notion is stable under
648
+ composition and base change, so the reader must think of it as an abbreviated
649
+ way of storing information whose purpose will soon become clear.
650
+ We
651
+ 16
652
+
653
+ would like to remark, however, that the notion of Mod-affinity and the
654
+ concept of qc-isomorphism it produces are particular cases of more abstract
655
+ constructions for C -data.
656
+ Lemma 6.9. Given X ∈ SchFinop -data whose restriction morphisms are
657
+ flat immersions, Cyl(X) is pseudo-schematic.
658
+ Proof. It follows from the construction, Lemma 2.9 and Remark 2.2.
659
+ Now, given X ∈ SchFinop -data and p ∈ |X|, denote by Up the datum
660
+ induced on the open subset Up ⊆ |X|. We have qc-isomorphisms of ringed
661
+ spaces
662
+ πp: Cyl(Up) → X(p).
663
+ In general, given an open subset U ⊂ |X| and endowing it with the induced
664
+ structure functor, we have open subsets
665
+ iU : Cyl(U) ֒→ Cyl(X);
666
+ so, for every p, q ∈ |X| and fixed t ≤ p, q, we have natural morphisms
667
+ ip
668
+ pq : Cyl(Up ∩ Uq) → Cyl(Up),
669
+ iq
670
+ pq : Cyl(Up ∩ Uq) → Cyl(Uq);
671
+ which, composing with the previous projections, induce
672
+ πt
673
+ pq : Cyl(Up ∩ Uq) → X(p) ×X(t) X(q).
674
+ Note that the space on the right hand side is always schematic and that, for
675
+ every (xp, yq) ∈ |X(p) ×X(t) X(q)|, we have
676
+ (πt
677
+ pq)−1(U(xp,yq)) = Uxp ∩ Uyq ⊆ |Cyl(Up ∩ Uq)| ⊆ |Cyl(X)|.
678
+ Theorem 6.10. Given X ∈ SchFinop -data whose restriction morphisms
679
+ are flat immersions, Cyl(X) is schematic if and only if for every t ≤ p, q in
680
+ |X|, the natural morphism πt
681
+ pq is a qc-isomorphism (a priori of ringed posets,
682
+ a posteriori of schematic spaces).
683
+ Proof. With the technology introduced in this paper, we can only prove the
684
+ "if" part, which will be the one used in our applications. Indeed, if πt
685
+ pq is
686
+ a qc-isomorphism, Uxp ∩ Uyq is Mod-affine for every (xp, yq) as before and
687
+ its global sections are isomorphic to OX(p),xp ⊗OX(t),zt OX(q),yq, with zt the
688
+ common image of xp and yq. Now, Lemma 6.6 translates exactly into the
689
+ conditions of Definition 2.1.
690
+ 17
691
+
692
+ For morphisms f : X → Y in SchFinop -data, we can modify the previous
693
+ construction to obtain, for each p ∈ |X| and q ≥ f(p),
694
+ ρf
695
+ pq : Cyl(Up ∩ f −1(Uq)) → X(p) ×Y(f(p)) Y(q).
696
+ Theorem 6.11. Given a morphism f : X → Y in SchFinop -data and such
697
+ that Cyl(X) and Cyl(Y) are schematic, Cyl(f) is schematic if and only if for
698
+ every p, q ≥ f(p), the map ρf
699
+ pq is a qc-isomorphism.
700
+ Proof. We only prove the "if" part, which follows from the same results as
701
+ Theorem 6.10 and the fact that, for (xp, yq) ∈ X(p) ×Y(f(p)) Y(q), one has
702
+ ρ−1
703
+ pq (U(xp,yq)) = Uxp ∩ Cyl(f)−1(Uyq).
704
+ Remark 6.12. Note that, applied to a datum X with X(p) = (⋆, Ap) for all
705
+ p, Theorems 6.10 and 6.11 restrict to the usual Definition of schematicity.
706
+ See this in view of Example 5.4.
707
+ Definition 6.13. Given a finite family of flat immersions {Ui → X}i∈I,
708
+ we define the Nerve datum associated to it as U ∈ SchFinop -data with
709
+ underlying poset |U| = P∗(I)—non-empty parts of I—and U(∆) = �
710
+ i∈∆ Ui
711
+ —fibered product over X—.
712
+ Note that U comes equipped with a morphism U → X ≡ iC -data(X).
713
+ Corollary 6.14. If {Ui → X} a finite family of flat immersions, Cyl(U)
714
+ is schematic and the morphism Cyl(U) → X is a schematic flat immersion,
715
+ which is a qc-isomorphism if and only if the family is a covering.
716
+ Proof. First, we check the condition of Theorem 6.10: for ∆1, ∆2 ∈ |U|,
717
+ U∆1 ∩ U∆2 = U∆1∪∆2; but Cyl(U∆1∪∆2) → U(∆1 ∪ ∆2) is a qc-isomorphism,
718
+ with U(∆1 ∪ ∆2) ≃ U(∆1) ×U(∆1∩∆2) U(∆2) by definition. Schematicity of
719
+ Cyl(f) follows from Theorem 6.11 and a similar argument.
720
+ The morphism Cyl(f) is flat by the local construction and its diagonal
721
+ is a qc-isomorphism because, by Proposition 5.5,
722
+ Cyl(U) → Cyl(U) ×X Cyl(U) ≃ Cyl(U ×X U),
723
+ and a morphism of SchFinop -data that is topologically the identity and a
724
+ qc-isomorphism at each point, induces a qc-isomorphism between cylinder
725
+ spaces (as shown by an easy computation).
726
+ Finally, Cyl(f) being faithfully flat (hence a qc-isomorphism) is clearly
727
+ equivalent to {f∆ : U(∆) → X}∆ being a covering family, which happens if
728
+ and only if the original family was a covering.
729
+ 18
730
+
731
+ 7
732
+ Descent and the topos of flat immersions
733
+ Now we use the technology of the previous section to describe colimits in
734
+ a sheaf-theoretic manner. In the following definition—if appropriate—, one
735
+ shall consider SchFin as a 1-category with the trivial 2-categorical structure.
736
+ Definition 7.1. Let C be a 1-category (resp. strict 2-category). A geometric
737
+ datum is a functor (resp. pseudofunctor) Dat: SchFin → C that maps qc-
738
+ isomorphisms to isomorphisms (resp. equivalences); in other words, one that
739
+ factors through SchFinqc.
740
+ Example 7.2. The functors Spec: SchFin → LRS, Qcoh: SchFin → Catop
741
+ —with values in the 2-category of categories—and Πét
742
+ 1 : SchFin → GpdStone
743
+ —with values in the 2-category of Stone groupoids—are all geometric data.
744
+ In the discussion that follows, let us assume that C is a 1-category; the
745
+ argument also works for 2-categories, replacing isomorphisms by equivalences.
746
+ In Example 5.4 we saw that there are two natural immersions of any category
747
+ of C-data into its category (C -data)op -data. In this case, there is a natural
748
+ transformation between functors in [SchFin, SchFinop
749
+ qc -data]:
750
+ ηC : (iop
751
+ C )∗ → iC -dataop.
752
+ Remark 7.3. For general ringed posets, this natural transformation is induced
753
+ by the morphisms (⋆, OX,x) → X, which are not schematic. That is one of
754
+ the reasons to consider the localized category, where it is induced by the
755
+ triangles (⋆, OX,x) ← Ux → X.
756
+ Now, given a geometric datum Dat: SchFinqc → C, we define
757
+ Dat := Dat∗ ◦ (iop
758
+ C )∗
759
+ Dat ≡ Dat ◦ iC -dataop;
760
+ where Dat(X) is the Cop-datum with |Dat(X)| = |X| and structure functor
761
+ Dat(X)(x) = Dat(⋆, OX,x). These induce a natural transformation
762
+ ηDat : Dat → Dat
763
+ between functors in [SchFin, Cop -data], given by the projection to the point
764
+ at the topological level and by the morphisms in C
765
+ Dat(⋆, OX,x) ∼
766
+ ← Dat(Ux) → Dat(X).
767
+ Composing with the sections functor Γ: Cop -data → C—always assuming
768
+ that C has enough limits—, one arrives to the following definition:
769
+ 19
770
+
771
+ Definition 7.4. We say that a geometric datum Dat satisfies internal descent
772
+ if Γ(ηDat): Γ∗ ◦ Dat → Γ∗ ◦ Dat ≡ Dat is an isomorphism in [SchFin, C].
773
+ Example 7.5. The datum Qcoh: SchFin → Catop satisfies internal descent.
774
+ Indeed, since Qcoh(⋆, OX,x) = Mod(OX,x), this amounts to proving that
775
+ the natural functor
776
+ Qcoh(X) → 2-limx∈X Mod(OX,x)
777
+ is an equivalence of categories. This holds because quasi-coherent modules
778
+ on ringed posets are collections of {Mx}x∈X with Mx an OX,x-module such
779
+ that, for all x ≤ y, the natural morphisms Mx ⊗OX,x OX,y → My are
780
+ isomorphisms; which coincides with the description of this pseudolimit in
781
+ Cat. The reader may notice that this result holds for arbitrary ringed posets,
782
+ but that it tacitly requires the tensor-Hom adjunction for modules to hold.
783
+ If one wants to extend the result to quasi-coherent sheaves of algebras, it is
784
+ necessary to assume that X is, at least, pseudo-schematic. This is because
785
+ base changes by flat epimorphisms of rings satisfy said adjunction (left as an
786
+ algebra exercise to the reader).
787
+ Proposition 7.6 (External descent for nerves). If X is a schematic space,
788
+ {fi : Ui → X} is a covering by flat immersions with associated nerve datum
789
+ U and Dat is a geometric datum satisfying internal descent, then there is a
790
+ natural isomorphism
791
+ colim∆∈|U| Dat(U(∆)) ∼
792
+ → Dat(X).
793
+ Proof. By Corollary 6.14, Cyl(U) → X is a qc-isomorphism, and since Dat is
794
+ geometric, one has that Dat(Cyl(U)) ≃ Dat(X). Since Dat satisfies internal
795
+ descent—applied at each U(∆)—and colimits commute with colimits,
796
+ Dat(Cyl(U)) ≃ colimx∆∈Cyl(U) Dat(⋆, OU(∆),x∆) ≃
797
+ ≃ colim∆∈|U| colimx∆∈|U(∆)| OU(∆),x∆) ≃ colim∆∈|U| Dat(U(∆)),
798
+ which completes the proof.
799
+ Example 7.7. In the situation of Proposition 7.6 and thanks to Example 7.5,
800
+ we obtain that Qcoh(X) ≃ 2-lim∆∈|U| Qcoh(U(∆)). In particular, being
801
+ quasi-coherent is local in the topology of flat immersions.
802
+ Theorem 7.8 (External descent for topoi). If SchFinτ denotes the (big)
803
+ site of schematic spaces with the combinatorial topology and SchFinwZar
804
+ 20
805
+
806
+ denotes the (big) site of flat immersions, the natural inclusion defines an
807
+ equivalence of topoi
808
+ Sh((SchFinqc)τ) ≃ Sh(SchFinwZar).
809
+ Similarly, it induces equivalences between respective categories of C-valued
810
+ sheaves (resp. stacks) for any 1-category (resp. 2-category) that has finite
811
+ poset-indexed colimits.
812
+ Proof. This is simply a reinterpretation of Proposition 7.6 in terms of the
813
+ language of Section 2.1: sheaves in Sh(SchFinwZar) map qc-isomorphisms
814
+ to isomorphisms, so they are geometric data in the sense fo this section. We
815
+ shall remark that the analogous equivalence between small topoi does not
816
+ hold—a priori—because cylinders change the base space.
817
+ Remark 7.9. Thanks to the sheaf condition, it can be shown that a sheaf F in
818
+ Sh(SchFinτ) maps qc-isomorphisms to isomorphisms if and only if, for every
819
+ affine schematic space X, the natural morphism F(⋆, OX(X)) → F(X) is
820
+ an isomorphism.
821
+ In other words, to prove that a presheaf in the schematic category is a
822
+ sheaf in the topology of flat immersions, it is enough to see that it maps
823
+ qc-isomorphisms to isomorphisms and that it is a sheaf in the combinatorial
824
+ topology for every poset. This is similar to what happens in the category of
825
+ schemes for the set-theoretic topology and the Zariski site. A consequence
826
+ for qc-qs schemes is the following slogan:
827
+ In the category of qc-qs schemes, any Zariski sheaf that can be
828
+ studied through finite models is a sheaf in the topology of flat
829
+ monomorphisms of schemes and finite coverings.
830
+ Theorem 7.8 makes the meaning of can be studied through precise: such a
831
+ sheaf F must induce a geometric datum on the schematic category that is a
832
+ sheaf in the combinatorial topology.
833
+ 8
834
+ Example: Seifert-Van Kampen Theorem
835
+ A less trivial application comes from the étale fundamental groupoid—and
836
+ group—, as promised. Let us consider the pseudofunctor
837
+ Πét
838
+ 1 : SchFin → GpdStone
839
+ 21
840
+
841
+ to the 2-category of Stone groupoids. By Remark 2.6 it is a geometric datum,
842
+ so to apply the results of the previous section it is enough to see that it is
843
+ a sheaf in the combinatorial topology. This follows quite easily in two steps.
844
+ Before that, we highlight that the category of finite étale covers defined
845
+ without any detail in Theorem 2.5 can be described in terms of quasi-coherent
846
+ sheaves of algebras, as done in [4]; more precisely:
847
+ RÉt(X) =
848
+
849
+ "opposite category of quasi-coherent algebras A
850
+ such that OX,x → Ax is a finite étale ring map".
851
+ (8.1)
852
+ Lemma 8.1. The pseudofunctor RÉt: SchFinqc → Cat defined on objects
853
+ by1 X �→ RÉt(X) satisfies internal descent.
854
+ Proof. We have to show that RÉt(X) ≃ 2-limx∈X RÉt(⋆, OX,x). Since it is
855
+ a subcategory of the category of quasi-coherent algebras, this follows from
856
+ Example 7.5—bearing in mind the remark at the end—and the fact that the
857
+ property of being finite étale at stalks is obviously local in this sense.
858
+ Proposition 8.2. The pseudofunctor Πét
859
+ 1 : SchFin → GpdStone satisfies
860
+ internal descent.
861
+ Proof. Following the notations of Section 7. By Lemma 8.1 we know that
862
+ the natural transformation ηRÉt : RÉt → RÉt induces an equivalence after
863
+ taking sections. Composing Πét
864
+ 1 with the 2-functor Φ: GpdStone → Catop
865
+ such that G �→ G-FinSet—with continuous action—, we obtain a commutative
866
+ square of functors [SchFin, Catop]
867
+ RÉt
868
+
869
+
870
+ RÉt
871
+
872
+ Πét
873
+ 1 -FinSet
874
+ ηΦ◦Πét
875
+ 1 � Πét
876
+ 1 -FinSet;
877
+ where the vertical arrows are isomorphisms after taking sections by the
878
+ Galois Theorem for fundamental groupoids, hence Γ(ηΦ◦Πét
879
+ 1 ) an isomorphism.
880
+ Finally, since Φ well known to commute with pseudocolimits, one has
881
+ that Γ(ηΦ◦Πét
882
+ 1 ) ≃ Φ ◦ Γ(ηΠét
883
+ 1 ); and since this map is an equivalence and Φ
884
+ is (2-)conservative by [10, 3.11], Γ(ηΠét
885
+ 1 ) is an equivalence, which proves the
886
+ statement.
887
+ 1On 1-morphisms, we send each f : X → Y to the inverse image functor; since SchFin
888
+ is considered as a 1-category, it only remains to specify invertible equivalences in Cat that
889
+ make all suitable diagrams commute, but we can and do choose those to be the ones given
890
+ by the universal property of tensor products.
891
+ 22
892
+
893
+ Example 8.3. If a schematic space X satisfies that Πét
894
+ 1 ((⋆, OX,x)) = {⋆}—the
895
+ trivial 2-category—for all x ∈ X, Proposition 8.2 yields
896
+ Πét
897
+ 1 (X) ≃ 2-colimx∈|X|{⋆} ≃
898
+
899
+ Π1(|X|),
900
+ where the hat denotes the profinite completion.
901
+ Theorem 8.4. Let X be a schematic space and {Ui → X} be a covering by
902
+ flat immersions with associated nerve datum U. Then, the natural morphism
903
+ 2-colim∆∈|U| Πét
904
+ 1 (U(∆)) → Πét
905
+ 1 (X)
906
+ is an equivalence of topological groupoids. In other words, the functor Πét
907
+ 1
908
+ is a (co)stack in the topology of flat immersions.
909
+ Proof. It follows from 8.2 and Theorem 7.8.
910
+ Remark 8.5. Note that the topological fundamental groupoid of |U| is always
911
+ trivial, since any space of parts has generic point and thus is contractible to
912
+ a point. One can give the statement of the Theorem in greater generality,
913
+ for any X ∈ SchFinop-datum such that Cyl(X) is schematic; and in that
914
+ case the topological fundamental groupoid of |X| plays a role.
915
+ Corollary 8.6. If S is a qc-qs scheme and {Vj → S}j∈J is a finite cover by
916
+ flat monomorphisms with associated nerve codatum V : P∗(J) → Schop—
917
+ with Sch the category of schemes—, the natural morphism
918
+ 2-colim∆∈|V| Πét
919
+ 1 (V(∆)) → Πét
920
+ 1 (S)
921
+ is an equivalence of Stone groupoids, i.e. the étale fundamental groupoid of
922
+ schemes is a costack in the topology of flat monomorphisms and finite covers.
923
+ Finally, we can very easily specialize this result to fundamental groups,
924
+ which a formulation that we deem more natural than that of [11].
925
+ Definition 8.7. Given a schematic space X and a cover by flat immersions
926
+ with associated nerve datum U extended to P(I) by U(∅) = X, a system of
927
+ base points x⋆ is an object
928
+ x⋆ ∈ Ob(2-lim∆∈|U|(Πét
929
+ 1 (U(∆)))).
930
+ In other words: x⋆ is given by a collection geometric points x∆ of U(∆)
931
+ for each ∆ and a collection of Tannaka paths
932
+ ϕ∆∆′ : Fibx∆ ◦ RÉt(X)(∆ → ∆′) ∼
933
+ → Fibx∆′
934
+ for each ∆ ≤ ∆′. Let us denote by x = x∅ the geometric point of X given
935
+ by this collection.
936
+ 23
937
+
938
+ Theorem 8.8. Let X be schematic and connected, U the nerve codatum
939
+ associated to some covering by flat immersions such that U(∆) is connected,
940
+ and x⋆ a system of base points. Then there is an isomorphism of topological
941
+ groups
942
+ colim∆∈|U| πét
943
+ 1 (U(∆), x∆) ∼
944
+ → πét
945
+ 1 (X, x)
946
+ induced by conjugation the ϕ∆∆′.
947
+ Proof. Since X is connected, the natural inclusion πét
948
+ 1 (X, x) → Πét
949
+ 1 (X) is an
950
+ equivalence. Let GrStone ⊆ GpdStone be the category of profinite groups,
951
+ which one may think set-theoretically or as Top-enriched categories. Define
952
+ the datum
953
+ πét
954
+ 1 (−, x⋆): |U| → Grop
955
+ Stone
956
+ ∆ → πét
957
+ 1 (U(∆), x∆)
958
+ whose restriction morphisms given by conjugation with the ϕ∆∆′.
959
+ Since
960
+ U(∆) is connected for every ∆, the natural transformation
961
+ πét
962
+ 1 (−, x⋆) → Πét
963
+ 1 ◦ U
964
+ is an isomorphism of GpdStone-valued pseudofunctors, hence it induces an
965
+ isomorphism after taking sections. From this fact and Theorem 8.4, there
966
+ are equivalences
967
+ 2-colim∆∈|U| πét
968
+ 1 (U(∆), x∆) ∼
969
+ → 2-colim∆∈|U| Πét
970
+ 1 (U(∆)) ≃ Πét
971
+ 1 (X);
972
+ where the first groupoid is identified with the 1-colimit of abstract profinite
973
+ groups colim∆∈|U| πét
974
+ 1 (U(∆), x∆) and the last one is equivalent to πét
975
+ 1 (X, x)
976
+ as remarked before. Since any equivalence between one-object categories is
977
+ an isomorphism, the proof ends.
978
+ Remark 8.9. Note that the topological Seifert-Van Kampen Theorem can be
979
+ written in terms of C -data: if S is a quasi-compact topological space and
980
+ π: S → X
981
+ is a finite model, we can turn X into a Topop-datum—with Top being the
982
+ category of topological spaces—by setting that X(x) = π−1(Ux). If each one
983
+ of these fibers is simply connected and we assume connectedness, the result
984
+ recovers the classical one of McCord for π1. For the higher homotopy groups,
985
+ we are positive that should be a consequence of a Seifert-Van Kampen
986
+ Theorem for fundamental homotopy groupoids thought as strict n-categories.
987
+ 24
988
+
989
+ References
990
+ [1] Codara, P. PhD thesis:
991
+ A theory of partitions of partially ordered
992
+ sets; O.M. D’Antona, V. Marra. Milano:
993
+ Università degli studi di
994
+ Milano. Dipartimento di Matematica, Dipartimento di Informatica e
995
+ Comunicazione, 2008 Nov 21. 20. ciclo, Anno Accademico 2006/2007.
996
+ [2] McCord, M. C. Singular homology groups and homotopy groups of finite
997
+ topological spaces, Duke Math. J. 33 (1966), 465-474.
998
+ [3] Raynaud, M. Un critère de effectivité de descente. In: Séminaire Samuel,
999
+ Algèbre Conmutative, vol. 2, pp. 1-22 (1967-1967).
1000
+ [4] Sánchez González, J.; Tejero Prieto, C. Étale Covers and Fundamental
1001
+ Groups of Schematic Finite Spaces. Mediterr. J. Math. 19 (2022), no. 5,
1002
+ 229.
1003
+ [5] Sancho de Salas, F.; Sancho de Salas, P. Affine ringed spaces and Serre’s
1004
+ criterion. Rocky Mountain J. Math. 47 (2017), no. 6, 2051–2081.
1005
+ [6] Sancho de Salas, F. Finite spaces and schemes. J. Geom. Phys. 122
1006
+ (2017), 3–27.
1007
+ [7] Sancho de Salas, F. Homotopy of finite ringed spaces. J. Homotopy Relat.
1008
+ Struct. 13 (2018), no. 3, 481–501.
1009
+ [8] Sancho de Salas, F.; Torres Sancho, J.F. Derived categories of finite
1010
+ spaces and Grothendieck duality. Mediterr. J. Math. 17 (2020), no. 3,
1011
+ Paper No. 80, 22 pp.
1012
+ [9] Sancho de Salas, F.; Sancho de Salas, P. Notes on schematic finite spaces.
1013
+ arXiv:2102.09263v1 [math.AG].
1014
+ [10] Pirashvili, I. The étale fundamental groupoid as a 2-terminal costack.
1015
+ Kyoto J. Math. 60 (2020), no. 1, 379–403.
1016
+ [11] Stix,
1017
+ J.
1018
+ A
1019
+ general
1020
+ Seifert-Van
1021
+ Kampen
1022
+ theorem
1023
+ for
1024
+ algebraic
1025
+ fundamental groups. Publ. Res. Inst. Math. Sci. 42 (2006), no. 3,
1026
+ 763–786.
1027
+ [12] Waterhourse, W.C. Basically bounded functors and flat sheaves. Pacific
1028
+ J. Math. 57 (1975), no. 2, 597-610.
1029
+ 25
1030
+
GNE1T4oBgHgl3EQfXAT2/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
GtE0T4oBgHgl3EQfzgJq/content/tmp_files/2301.02673v1.pdf.txt ADDED
@@ -0,0 +1,940 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Data-driven discovery and extrapolation of parameterized pattern-forming dynamics
2
+ Zachary G. Nicolaou,1 Guanyu Huo,1 Yihui Chen,1 Steven L. Brunton,2 and J. Nathan Kutz1
3
+ 1Department of Applied Mathematics, University of Washington, Seattle, WA 98195, USA
4
+ 2Department of Mechanical Engineering, University of Washington, Seattle, WA 98195, USA
5
+ Pattern-forming systems can exhibit a diverse array of complex behaviors as external parameters
6
+ are varied, enabling a variety of useful functions in biological and engineered systems. First-principle
7
+ derivations of the underlying transitions can be characterized using bifurcation theory on model sys-
8
+ tems whose governing equations are known. In contrast, data-driven methods for more complicated
9
+ and realistic systems whose governing evolution dynamics are unknown have only recently been de-
10
+ veloped. Here, we develop a data-driven approach sparse identification for nonlinear dynamics with
11
+ control parameters (SINDyCP) to discover dynamics for systems with adjustable control parameters,
12
+ such as an external driving strength. We demonstrate the method on systems of varying complexity,
13
+ ranging from discrete maps to systems of partial differential equations. To mitigate the impact of
14
+ measurement noise, we also develop a weak formulation of SINDyCP and assess its performance
15
+ on noisy data. We demonstrate applications including the discovery of universal pattern-formation
16
+ equations, and their bifurcation dependencies, directly from data accessible from experiments and
17
+ the extrapolation of predictions beyond the weakly nonlinear regime near the onset of an instability.
18
+ Data-driven approaches to system identification are
19
+ undergoing a revolution, spurred by the increasing avail-
20
+ ability of computational resources, data, and the develop-
21
+ ment of novel and reliable machine learning algorithms
22
+ [1–3].
23
+ The sparse identification of nonlinear dynamics
24
+ (SINDy) is a particularly simple and flexible mathemat-
25
+ ical approach that leverages efficient sparse optimization
26
+ algorithms in the automated discovery of complex sys-
27
+ tem dynamics and governing equations [4]. In this work,
28
+ we leverage the SINDy model discovery framework to
29
+ understand parametric dependencies and underlying bi-
30
+ furcations in pattern forming systems. Specifically, we
31
+ develop the SINDY with control parameters (SINDyCP)
32
+ to discover such parameterized dynamics.
33
+ It has been thirty years since Cross and Hohenberg’s
34
+ seminal and authoritative review consolidating an excep-
35
+ tionally large body of work on pattern formation across a
36
+ broad range of physical systems [5]. Universal equations
37
+ determined by normal forms of canonical bifurcations [6],
38
+ such as the complex Ginzburg-Landau equation [7], gov-
39
+ ern the formation of patterns near the onset of instabili-
40
+ ties across scientific disciplines. Such equations continue
41
+ to reveal insights into complex systems, including in the
42
+ study of, for example, synchronization, biophysics, active
43
+ matter, and quantum dynamics [8, 9].
44
+ Despite the success of pattern-formation theory in
45
+ modeling complex dynamics, ongoing challenges remain
46
+ in applying such model equations more broadly. First-
47
+ principle derivations and the computation of normal-
48
+ form parameters in terms of physical driving parameters
49
+ are tedious, costly, and error-prone.
50
+ Furthermore, the
51
+ normal-form approach is only theoretically justified in
52
+ the weakly-nonlinear regime near the onset of an insta-
53
+ bility, while interesting and important pattern-forming
54
+ processes often occur far from the instability threshold.
55
+ Recent advances in data-driven system identification are
56
+ opening new avenues of research to address these chal-
57
+ lenges, including a paradigm for modeling strongly non-
58
+ linear regimes beyond the asymptotic approximations re-
59
+ viewed by Cross and Hohenberg [5].
60
+ The SINDy model discovery framework is particularly
61
+ well-suited to the modern analysis of bifurcations and
62
+ normal forms, as it generates interpretable models that
63
+ have as few terms as possible, balancing model complex-
64
+ ity and descriptive capability. A variety of extensions of
65
+ the SINDy approach have been developed since its in-
66
+ troduction. For example, SINDYc enables discovery of
67
+ systems subject to external control signals [11, 12], while
68
+ PDEFind [13, 14] enables discovery of spatio-temporal
69
+ dynamics characterized by partial differential equations
70
+ (PDEs). SINDy can also learn to disambiguate between
71
+ parametric dependency and governing equations [15].
72
+ Model pattern formation equations typically encode the
73
+ effects of external drive through a number of driving pa-
74
+ rameters, which characterize the bifurcation leading to
75
+ the onset of instability. Several recent works establish
76
+ system identification on pattern-forming systems rang-
77
+ ing from closure models for fluid turbulence [16–18] to
78
+ biochemical reactions and active active matter systems
79
+ [19–21]. These approaches show promise, but crucially,
80
+ they have not demonstrated the ability to extrapolate
81
+ by detecting pattern-forming instabilities that may de-
82
+ velop when driving parameter differ the training data.
83
+ While there has been success for discrete maps and or-
84
+ dinary differential equations (ODEs) [4, 22], combining
85
+ the PDEFind and SINDYc approaches to discover pa-
86
+ rameterized spatio-temporal dynamics poses a significant
87
+ challenge, as we detail below.
88
+ The key insight underlying SINDyCP is recognizing
89
+ the need to introduce distinct libraries of possible de-
90
+ pendencies for the dependent variables and the control
91
+ parameters. Our approach is implemented in the open-
92
+ source PySINDy repository [23, 24], enabling other pow-
93
+ erful methods to be used in conjunction.
94
+ In particu-
95
+ arXiv:2301.02673v1 [nlin.PS] 6 Jan 2023
96
+
97
+ 2
98
+ Construct library
99
+ and derivatives
100
+ from samples
101
+ Sparse
102
+ regression
103
+ Parameterized
104
+ equation
105
+ Trajectories with
106
+ varying parameters
107
+ Feature
108
+ Library,
109
+ Parameter
110
+ Library,
111
+ Time
112
+ Derivatives,
113
+ FIG. 1. Schematic of the SINDyCP approach. Data collected from sample trajectories collected under various driving parame-
114
+ ters are processed to construct a matrix of time derivatives, a feature library Θfeat of possible governing terms, and a parameter
115
+ library Θpar of parametric dependencies. Sparse regression is applied on the library coefficients ξ to identify a parameterized
116
+ governing equation.
117
+ lar, we develop and assess a weak formulation [25–28] of
118
+ SINDyCP, which shows excellent performance on noisy
119
+ data. We demonstrate that the method can be easily and
120
+ effectively employed to discover accurate parameterized
121
+ models from the kind of data available in typical pat-
122
+ tern formation experiments and that these parameterized
123
+ models enable extrapolation beyond the conditions under
124
+ which they were developed.
125
+ Building
126
+ the
127
+ library.—Figure
128
+ 1
129
+ illustrates
130
+ the
131
+ SINDyCP approach applied to the spatio-temporal
132
+ evolution of four trajectories of the complex Ginzburg-
133
+ Landau equation
134
+ ˙A = A + (1 + ib)∇2A − (1 − ic)|A|2A,
135
+ (1)
136
+ which
137
+ is
138
+ described
139
+ by
140
+ a
141
+ complex
142
+ dependent
143
+ vari-
144
+ able A(x, t) in two spatial dimensions x
145
+ =
146
+ (x, y).
147
+ Ginzburg-Landau exhibits a stunning variety of pat-
148
+ terns, depending on the bifurcation parameters b and
149
+ c.
150
+ We generate four trajectories with parameters val-
151
+ ues (b, c) = (2.0, 1.0), (2.0, 0.75), (0.5, 0.5) and (1.0, 0.75),
152
+ which exhibit differing dynamical phases, corresponding
153
+ to amplitude turbulence, phase turbulence, stable waves,
154
+ and frozen spiral glasses, respectively [7]. Our goal is to
155
+ discover the partial differential equation for the real and
156
+ imaginary components A = X + iY parameterized by b
157
+ and c from time series data.
158
+ As with most SINDy algorithms, we first form a ma-
159
+ trix of the input data X, whose columns correspond to
160
+ the dependent variables and whose rows correspond to
161
+ the sample measurements of the dependent variables. In
162
+ the case of Fig. 1, for example, X consists of two columns
163
+ corresponding to the real and imaginary parts of A and
164
+ 4NxNyNt rows, where Nx, Ny, and Nt are the number
165
+ of sample points in the corresponding spatio-temporal di-
166
+ mensions; again, there are four trajectories. We then de-
167
+ termine the temporal derivative ˙X for each sample point,
168
+ either through numerical differentiation or through direct
169
+ measurements.
170
+ In basic SINDy, we define a matrix of library terms
171
+ Θ = Θ(X) depending on the input data, which includes
172
+ all possible terms that may be present in the differen-
173
+ tial equation that describes the temporal derivatives.
174
+ These terms may be built from polynomial combina-
175
+ tions of the dependent variables and their spatial deriva-
176
+ tives, for example, although more general libraries are
177
+ possible. In the SINDYc approach, we augment the li-
178
+ brary dependence with an external control signal U, i.e.,
179
+ Θ = Θ(X, U). The library terms are typically determined
180
+ by appending the control variables to the dependent vari-
181
+ ables and again forming polynomials and derivatives. In
182
+ the case in Fig. 1, we can treat the parameters as exter-
183
+ nal control signals, U = (b, c) and apply SINDYc, but
184
+ the traditional implementation of this approach will fail
185
+ for PDEs, as we show.
186
+ SINDYc aims to find a sparse linear combination of
187
+ the library terms determined by the vector of coefficients
188
+ ξ which minimizes the fit error
189
+ ξ∗ = argminξ
190
+ ��� ˙X − Θ(X, U)ξ
191
+ ��� + λ |ξ|0 .
192
+ (2)
193
+ Crucially, all SINDy methods employ sparse regression
194
+ (with appropriate regularization) to determine a sparse
195
+ set of nonzero coefficients ξ∗. Such sparsity is expected in
196
+ physically-relevant dynamics and produces parsimonious
197
+ and interpretable models.
198
+ A significant challenge arises when applying the tradi-
199
+ tional SINDYc to control parameters in PDEs with ex-
200
+ isting implementations such as PySINDy.
201
+ The matrix
202
+ of library terms Θ is traditionally formed by computing
203
+ all polynomial combinations of spatial derivatives of the
204
+ dependent and control variables. However, since the con-
205
+ trol parameters are spatially constant, the spatial deriva-
206
+ tives will vanish identically, leading to a singular matrix
207
+ Θ.
208
+ To overcome this challenge, we propose construct-
209
+ ing a more general library through products of a feature
210
+
211
+ 3
212
+ library Θfeat(X) and a parameter library Θpar(U), as
213
+ Θ(X, U) = Θfeat(X) ⊗ Θpar(U),
214
+ (3)
215
+ where the product ⊗ here is defined to give the ma-
216
+ trix consisting of all combinations of products of columns
217
+ (computed component-wise across the row elements) be-
218
+ tween the libraries.
219
+ By distinguishing the feature and
220
+ parameter library dependencies with this SINDyCP ap-
221
+ proach, we can construct much more targeted and well-
222
+ conditioned libraries.
223
+ Using a feature library consisting of spatial derivatives
224
+ up to third order and polynomials up to third order along
225
+ with a linear parameter library, the SINDyCP approach
226
+ easily discovers Eq. (1) in Cartesian coordinates. Details
227
+ of the numerical integration, an animation illustrating
228
+ the temporal evolution of the sample trajectories, and
229
+ additional demonstrations for maps and ODEs are avail-
230
+ able in the Supplemental Materials [29].
231
+ Beyond weakly nonlinear theory.—SINDyCP enables
232
+ discovery of nonlinear corrections to weakly nonlin-
233
+ ear theory directly from data that can be gathered in
234
+ pattern-formation experiments. To illustrate this result,
235
+ we implement an in silico experiment of the Belousov-
236
+ Zhabotinksy chemical reaction system. We numerically
237
+ integrate the Oregonator model [30],
238
+ ˙CX = k1CAC2
239
+ HCY − k2CHCXCY + k3CACHCX
240
+ − 2k4C2
241
+ X + DX∇2CX,
242
+ (4a)
243
+ ˙CY = −k1CAC2
244
+ HCY − k2CHCXCY + νk5CBCZ
245
+ + DY ∇2CY
246
+ (4b)
247
+ ˙CZ = 2k3CACHCX − k5CBCZ + DZ∇2CZ,
248
+ (4c)
249
+ which describes the evolution of oscillating chemical con-
250
+ centrations CX, CY , and CZ for given supplied concen-
251
+ trations CA, CB, and CH and stoichiometric coefficient
252
+ ν, which depends on the experimental setup. We vary
253
+ the concentration of CB and define a control parame-
254
+ ter µ ≡ CB − Cc
255
+ B, where Cc
256
+ B is the critical value where
257
+ the Hopf bifurcation occurs (see Supplementary Mate-
258
+ rials [29] for parameter values and other details in the
259
+ Oregonator model) to generate six trajectories with µ
260
+ ranging from 0.02 to 0.12.
261
+ We use a SINDyCP feature library with polynomial
262
+ terms up to fifth order and second order spatial deriva-
263
+ tives and a parameter library with polynomial terms up
264
+ to second order for the control parameter µ1/2 in con-
265
+ junction with implicit SINDy [31] to discovers a highly
266
+ nonlinear parameterized model from time-series measure-
267
+ ments of CX and CZ. Figure 2(a) shows the R2 score of
268
+ the model on test trajectories corresponding to the pa-
269
+ rameter values that the model was trained on (a value
270
+ of R2 = 1 means that the fit perfectly predicts the tem-
271
+ poral derivatives of the data). While the score decreases
272
+ modestly as µ increases, the model remains very accurate
273
+ (a)
274
+ (b)
275
+ (c)
276
+ FIG. 2.
277
+ Corrections to the weakly nonlinear theory of
278
+ the Oregonator model.
279
+ (a) R2 score for the parameterized
280
+ SINDyCP model on test trajectories collected at the param-
281
+ eter values used to train the model. (b) Corrected normal-
282
+ form parameter values relative to the weakly nonlinear values
283
+ b0 and c0 as a function of the bifurcation parameter µ1/2. (c)
284
+ Limit cycles in the homogeneous system exhibiting the highly-
285
+ nonlinear canard explosion with increasing µ. The pattern
286
+ formation above the canard explosion (upper inset) is quali-
287
+ tatively different than for smaller driving (lower inset), with
288
+ more extreme spatio-temporal variation that does not emerge
289
+ in the weakly nonlinear theory.
290
+ on all the testing trajectories, accounting for 99% of the
291
+ variation in the data in each case.
292
+ A nonlinear change of coordinates transforms the dis-
293
+ covered model into the normal-form in Eq. (1) with
294
+ parameter-dependent values of b(µ) and c(µ) and small
295
+ quintic corrections. These normal-form parameters agree
296
+ with the analytic values derived [32] from the original
297
+ model as µ → 0, but here we are able to discover them di-
298
+ rectly from data without any knowledge of the governing
299
+ equations. Furthermore, as shown in Fig. 2(b), the pa-
300
+ rameters vary with µ, representing additional corrections
301
+ to the weakly nonlinear theory. This variation becomes
302
+ extreme for µ1/2 > 0.35, which we were able to discover
303
+ via the implicit version of SINDy. In fact, as shown in
304
+ Fig. 2(c), the Oregonator model exhibits a canard explo-
305
+ sion (in which the limit cycle amplitude expands abruptly
306
+ due to highly nonlinear effects) [30] around µ1/2 ≈ 0.39,
307
+ where the weakly nonlinear theory breaks down.
308
+ The
309
+ SINDyCP model reflects this breakdown and enables the
310
+ development of higher-order corrections to account for it.
311
+ Weak formulation.—The weak formulation utilizes in-
312
+ tegration against compactly supported “test functions”
313
+ to defined the SINDy problem. The weak method shows
314
+ excellent performance for noisy data, owing to its ability
315
+ to minimize the need for computing numerical deriva-
316
+ tives.
317
+ Rather than forming samples (rows in Fig. 1)
318
+ from spatio-temporal points for each trajectory, the weak
319
+ method constructs the system rows by projecting the
320
+ data onto weak samples such as
321
+
322
+ ik ≡
323
+
324
+ Ωk
325
+ φ(x; t)X(ν)
326
+ i
327
+ (x; t) dDxdt,
328
+ (5)
329
+
330
+ 4
331
+ where Ωk is a compactly-supported spatio-temporal do-
332
+ main, φ is the test function, and X(ν)
333
+ i
334
+ denotes the νth
335
+ partial derivative the ith dependent variable. By moving
336
+ derivatives off of the data and onto the test functions via
337
+ integration by parts,
338
+
339
+ ik = (−1)|ν|
340
+
341
+ Ωk
342
+ φ(ν)(x; t)Xi(x; t) dDxdt,
343
+ (6)
344
+ the weak method significantly reduces the impact of mea-
345
+ surement noise on the SINDy library and improves the
346
+ fit results [33].
347
+ To maximize the performance for the weak method,
348
+ we have optimized and fully vectorized numerical inte-
349
+ gration for the weak formulation in PySINDy, which can
350
+ be easily combined with the SINDyCP library class. De-
351
+ tails about our efficient numerical implementation are
352
+ available in the Supplemental Material [29]. Products of
353
+ weak features do not generally form reasonable samples
354
+ for a SINDy model, since multiplication and integration
355
+ do not commute, so on first sight, it is not clear how to
356
+ combine weak form feature and parameter libraries with
357
+ SINDyCP. However, when computing the weak samples
358
+ corresponding to constant functions, such as those that
359
+ form the parameter library, the integrals simply repre-
360
+ sent the spatio-temporal volume of the domain Ωk. Our
361
+ implementation thus rescales the weak features for the
362
+ temporal derivatives by the same volumetric factors.
363
+ Performance.—Using 500 randomly distributed sam-
364
+ ple domains (measuring 1/10th the spatio-temporal do-
365
+ main size in each direction), the weak SINDyCP easily
366
+ identifies the complex Ginzburg-Landau equation using
367
+ the same data used for the traditional differential form
368
+ shown in Fig. 1. Furthermore, it can do so in just a few
369
+ seconds of run-time on a modern processor in this case
370
+ (over five times faster than the differential form).
371
+ To assess the impact of noise, we inject random Gaus-
372
+ sian noise of varying intensity [34] into the four trajecto-
373
+ ries used as the training data for the complex Ginzburg-
374
+ Landau equation. We then generate two new sample tra-
375
+ jectories to use as testing data, with b = 2.0, 1.5 and
376
+ c = 1.5, 1.0, respectively.
377
+ Using the training data, we
378
+ perform the SINDyCP fits using both the differential for-
379
+ mulation and the weak formulation and evaluate the R2
380
+ score on our test trajectories. Figure 3(a) shows the re-
381
+ sults for the R2 score on the test trajectories. While both
382
+ methods provide good fits for low noise intensity, only
383
+ the weak method exhibits a robust fit for parameterized
384
+ equations for large noise intensities.
385
+ The SINDyCP fit also requires a sufficient amount
386
+ of data to identify governing equations.
387
+ Figure 3(b)
388
+ shows the performance of SINDyCP on the testing data
389
+ for fits performed with a varying number of trajectories
390
+ nt = 2, 3, 4, 5 and of varying length corresponding to a
391
+ number of time samples Nt = 25, 50, 75, 100, with an in-
392
+ jected noise intensity of 10−3.
393
+ Unlike the trajectories
394
+ in Fig. 1, the parameters for trajectories were randomly
395
+ (a)
396
+ (b)
397
+ FIG. 3. Performance of SINDyCP for the fit of the complex
398
+ Ginzburg-Landau equation with noisy data. (a) Model score
399
+ vs noise intensity using the differential and weak forms of
400
+ SINDyCP with nt = 4 trajectories. (b) Model score vs num-
401
+ ber of samples for varying number of randomly generated tra-
402
+ jectories, varying trajectory length, and noise intensity 10−3.
403
+ generated, with (b, c) distributed as Gaussian random
404
+ variables with means (1.5, 1.0) and standard deviations
405
+ (0.5, 0.25).
406
+ For too little data, the fit fails to identify
407
+ the correct model, and the value of 1 − R2 is O(1). The
408
+ models improve moderately with an increasing number
409
+ of samples per trajectory (the product of Nt with the
410
+ number of spatial grid points). More importantly, a suf-
411
+ ficiently large number of trajectories nt is required to
412
+ achieve a good fit (at least 3 in this case). The amount
413
+ of data required will further increase when including a
414
+ larger number of possible library terms and when identi-
415
+ fying a larger number of parameters. These requirements
416
+ should be carefully assessed in order to achieve successful
417
+ SINDyCP fits for more general pattern forming systems.
418
+ Parameter extrapolation.—As a final demonstration
419
+ (Fig. 4), we consider the one-dimensional cubic-quintic
420
+ Swift-Hohenberg equation
421
+ ˙u = du − uxxxx − 2uxx − u + eu3 − fu5,
422
+ (7)
423
+ with parameters d, e, and f describing the linear, cu-
424
+ bic, and regularizing quintic terms, respectively.
425
+ This
426
+ model pattern formation equation has been used to study
427
+ defect dynamics incorporating corrections beyond the
428
+ weakly nonlinear approximation and has revealed uni-
429
+ versal snaking bifurcations leading to the formation of
430
+ localized states for e > 0 and d < 0 [35].
431
+ The parameters d, e and f are the minimal and natu-
432
+ ral set to describe the possible dynamics in the Swift-
433
+ Hohenberg equation derived from normal-form theory.
434
+ However, in typical pattern formation applications, one
435
+ does not have direct control over such parameters. In-
436
+ stead, experimentally accessible parameters will have a
437
+ complicated and nonlinear relationship with the normal-
438
+ form parameters, which requires detailed knowledge and
439
+
440
+ 5
441
+ (b)
442
+ (a)
443
+ FIG. 4. Extrapolation of localized states in the cubic-quintic
444
+ Swift-Hohenberg equation. (a) The randomly generated re-
445
+ lationships between the normal-form parameters (d, e, f) and
446
+ the experimental parameter ε (bottom panel) gives rise to
447
+ snaking bifurcations (top panel) near ε = 0. Red dotted lines
448
+ show the values used to train the SINDyCP fit, and dashed
449
+ colored lines show the coefficients derived from the fit. (b)
450
+ Localized states extrapolated from numerical simulations of
451
+ the SINDyCP fit with ε = 0.1, corresponding to the black
452
+ dotted line in (a).
453
+ tedious calculations to derive, e.g., an expansion and cen-
454
+ ter manifold transformation around a bifurcation point.
455
+ The SINDyCP approach enables an automated discovery
456
+ of such relationships, which can be used to extrapolate
457
+ system behavior beyond a set of measurements.
458
+ To illustrate this idea, we generate random quadratic
459
+ relationships between an experimental parameter ε and
460
+ the normal-form parameters (d, e, f), and we create three
461
+ training trajectories using random values of the param-
462
+ eter 1 < ε < 3 [Fig. 4(a)]. To determine the possible
463
+ dynamics, we numerically continue the solution branches
464
+ corresponding to the trivial state and localized and pe-
465
+ riodic states using the AUTO package [36].
466
+ For all of
467
+ the training trajectories, ε is sufficiently large that no
468
+ localized or periodic states are exist, and all trajecto-
469
+ ries decay to the trivial u = 0 solution.
470
+ We perform
471
+ the weak SINDyCP fit using these trajectories subject
472
+ to 1% injected noise with a quadradic parameter library.
473
+ To test the ability of SINDyCP to extrapolate beyond
474
+ the parameter regime given in the input data, we sim-
475
+ ulate the identified model for the experimental param-
476
+ eter value ε = 0.1. Remarkably, even with limited and
477
+ noisy training data, the method identifies an accurate
478
+ relationship between ε and the normal-form parameters.
479
+ Thus, simulations of the identified model with random
480
+ initial conditions converge to localized states [Fig. 4(b)]
481
+ for ε = 0.1 despite the significant extrapolation of the
482
+ parameter value beyond the input data.
483
+ Discussion—The SINDyCP approach represents a
484
+ simple but powerful generalization of SINDy with con-
485
+ trol. By disambiguating the feature and parameter com-
486
+ ponents of the SINDy libraries, the method enables dis-
487
+ covery of systems of partial differential equations param-
488
+ eterized by driving parameters. Such equations arise nat-
489
+ urally in the context of pattern formation, where the
490
+ normal forms of bifurcations lead to parameterized equa-
491
+ tions near the onset of instabilities. The approach can be
492
+ easily applied with the data available in typical pattern
493
+ formation experiments and promises to enable true ex-
494
+ trapolation beyond the regime that can be theoretically
495
+ described with weakly nonlinear theory. Combining the
496
+ SINDyCP approach with autoencoder-assisted discovery
497
+ of physical coordinates [37–39] will further enable re-
498
+ searchers to discover nonlinear equations governing com-
499
+ plex systems directly from data gathered through ex-
500
+ periments conducted under various driving parameters.
501
+ This approach may also help inform universal mecha-
502
+ nisms leading to the formation of localized states beyond
503
+ the snaking bifurcations of the Swift-Hohenberg equation
504
+ [40, 41].
505
+ This work benefited from insightful discussions with
506
+ Alan Kaptanoglu. Zachary G. Nicolaou is a WRF post-
507
+ doctoral fellow. We acknowledge support from the Na-
508
+ tional Science Foundation AI Institute in Dynamic Sys-
509
+ tems (grant number 2112085).
510
+ [1] S. L. Brunton, and J. N. Kutz. Data-driven science and
511
+ engineering: Machine learning, dynamical systems, and
512
+ control. (Cambridge University Press, 2022).
513
+ [2] S. M. Udrescu, and M. Tegmark. AI Feynman: A physics-
514
+ inspired method for symbolic regression. Sci. Adv. 6,
515
+ eaay2631 (2020).
516
+ [3] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris,
517
+ S. Wang, and L. Yang. Physics-informed machine learn-
518
+ ing. Nature Reviews Physics 3, 422-440 (2021).
519
+ [4] S. L. Brunton, J. L. Proctor, and J. N. Kutz, Discovering
520
+ governing equations from data by sparse identification
521
+ of nonlinear dynamical systems. Proc. Natl. Acad. Sci.
522
+ U.S.A. 113, 3932-3937 (2016).
523
+ [5] M. C. Cross and P. C. Hohenberg. Pattern formation
524
+ outside of equilibrium. Rev. Mod. Phys. 65, 851 (1993).
525
+ [6] Y. A. Kuznetsov, I. A. Kuznetsov, and Y. Kuznetsov.
526
+ Elements of applied bifurcation theory. Vol. 112. New
527
+ York: Springer, 1998.
528
+ [7] I. S. Aranson and L. Kramer. The world of the com-
529
+ plex Ginzburg-Landau equation. Rev. Mod. Phys. 74, 99
530
+ (2002).
531
+ [8] Z. G. Nicolaou, H. Riecke, and A. E. Motter. Chimera
532
+ states in continuous media: Existence and distinctness.
533
+ Phys. Rev. Lett. 119, 244101 (2017).
534
+ [9] V. Heinonen, A. J. Abraham, J. S�lomka, K. J. Burns, P.
535
+ J. S´aenz, and J. Dunkel. Emergent universal statistics in
536
+ nonequilibrium systems with dynamical scale selection.
537
+ arXiv preprint arXiv:2205.01627 (2022).
538
+ [10] S. L. Brunton, J. L. Proctor, and J. N. Kutz. Sparse iden-
539
+ tification of nonlinear dynamics with control (SINDYc).
540
+ IFAC-PapersOnLine 49, 710-715 (2016).
541
+ [11] E. Kaiser, J. N. Kutz, and S. L. Brunton, Sparse identifi-
542
+
543
+ 6
544
+ cation of nonlinear dynamics for model predictive control
545
+ in the low-data limit, Proc. Royal Soc. A 474, 20180335
546
+ (2018).
547
+ [12] U. Fasel, E. Kaiser, J. N. Kutz, B. W. Brunton, and S.
548
+ L. Brunton. Sindy with control: A tutorial. In 2021 60th
549
+ IEEE Conference on Decision and Control (CDC), pp.
550
+ 16-21. IEEE, 2021.
551
+ [13] S. H. Rudy, S. L. Brunton, J. L. Proctor, and J. N. Kutz.
552
+ Data-driven discovery of partial differential equations.
553
+ Sci. Adv. 3, e1602614 (2017).
554
+ [14] H. Schaeffer, Learning partial differential equations via
555
+ data discovery and sparse optimization, Proc. Royal Soc.
556
+ A 473, 20160446 (2017).
557
+ [15] S. H. Rudy, A. Alla, S. Brunton, and J. N. Kutz, Data-
558
+ driven identification of parametric partial differential
559
+ equations. SIAM J. App. Dyn. Sys. 18, 643-660 (2019).
560
+ [16] Schmelzer, Martin, Richard P. Dwight, and Paola Cin-
561
+ nella. ”Discovery of algebraic Reynolds-stress models us-
562
+ ing sparse symbolic regression.” Flow, Turbulence and
563
+ Combustion 104, no. 2 (2020): 579-603.
564
+ [17] L. Zanna and T. Bolton, Data-driven equation discov-
565
+ ery of ocean mesoscale closures, Geophys. Res. Lett. 47,
566
+ e2020GL088376 (2020).
567
+ [18] S. Beetham, R. O. Fox, and J. Capecelatro, Sparse iden-
568
+ tification of multiphase turbulence closures for coupled
569
+ fluid–particle flows, J. Fluid Mech. 914, A11 (2021).
570
+ [19] Z. Wang, B. Wu, K. Garikipati, and X. Huan, A per-
571
+ spective on regression and Bayesian approaches for sys-
572
+ tem identification of pattern formation dynamics, Theor.
573
+ Appl. Mech. Lett. 10, 188-194 (2020).
574
+ [20] N. Romeo, A. Hastewell, A. Mietke, and J. Dunkel.
575
+ Learning developmental mode dynamics from single-cell
576
+ trajectories. Elife 10 e68679 (2021).
577
+ [21] R. Supekar, B. Song, A. Hastewell, A. Mietke, and
578
+ J. Dunkel. Learning hydrodynamic equations for active
579
+ matter from particle simulations and experiments. arXiv
580
+ preprint arXiv:2101.06568 (2021).
581
+ [22] H. Schaeffer, G. Tran, and R. Ward, Learning dynam-
582
+ ical systems and bifurcation via group sparsity, arXiv
583
+ preprint arXiv:1709.01558 (2017).
584
+ [23] A. A. Kaptanoglu et al. PySINDy:
585
+ A comprehensive
586
+ Python package for robust sparse system identification.
587
+ J. of Open Source Softw. 7, 3994 (2022).
588
+ [24] The pysindy repository is available at https://github.
589
+ com/dynamicslab/pysindy.
590
+ [25] P. A. K. Reinbold, D. R. Gurevich, and R. O. Grigoriev.
591
+ Using noisy or incomplete data to discover models of spa-
592
+ tiotemporal dynamics. Phys. Rev. E 101, 010203 (2020).
593
+ [26] P. A. K. Reinbold, L. M. Kageorge, M. F. Schatz, and
594
+ R. O. Grigoriev. Robust learning from noisy, incom-
595
+ plete, high-dimensional experimental data via physically
596
+ constrained symbolic regression. Nat. Comm. 12, 3219
597
+ (2021).
598
+ [27] D. A. Messenger and D. M. Bortz. Weak SINDy for
599
+ partial differential equations. J. of Comput. Phys. 443,
600
+ 110525 (2021).
601
+ [28] D. A. Messenger and D. M. Bortz. Learning mean-field
602
+ equations from particle data using WSINDy. Physica D,
603
+ 133406 (2022).
604
+ [29] See Supplemental Material for details about numerical
605
+ integration, additional demonstrations, the oregonator
606
+ model, and the weak form implementation.
607
+ [30] M. Mazzotti, M. Morbidelli, and G. Serravalle. Bifurca-
608
+ tion analysis of the Oregonator model in the 3-D space
609
+ bromate/malonic acid/stoichiometric coefficient. J Phys
610
+ Chem. 99, 4501 (1995).
611
+ [31] N. M. Mangan, S. L. Brunton, J. L. Proctor, and J.
612
+ N. Kutz, Inferring biological networks by sparse iden-
613
+ tification of nonlinear dynamics. IEEE Trans. Mol. Biol.
614
+ Multi-Scale Commun. 2, 52. (2016).
615
+ [32] M. Ipsen, F. Hynne, and P. G. Sørensen, Amplitude equa-
616
+ tions for reaction–diffusion systems with a Hopf bifurca-
617
+ tion and slow real modes, Physica D 136, 66 (2000).
618
+ [33] It is not possible to remove all numerical derivatives in
619
+ the weak formulation, but the maximum order of deriva-
620
+ tives can generally be reduced to at most half the original
621
+ order for the library.
622
+ [34] Noise intensity here refers to the pointwise standard devi-
623
+ ation on the spatio-temporal grid employed in the simu-
624
+ lations. True white noise has a Dirac delta variance, and
625
+ intensity should thus scale with grid spacing and time
626
+ step to 1/2 power.
627
+ [35] J. Burke and E. Knobloch. Homoclinic snaking: structure
628
+ and stability. Chaos 17, 037102 (2007).
629
+ [36] E. J. Doedel, A. R. Champneys, F. Dercole, T. F. Fair-
630
+ grieve, Y. A. Kuznetsov, B. Oldeman, R. C. Paffenroth,
631
+ B. Sandstede, X. J. Wang, and C. H. Zhang. AUTO-
632
+ 07P: Continuation and bifurcation software for ordinary
633
+ differential equations. (2007).
634
+ [37] K. Champion, B. Lusch, J. N. Kutz, and S. L. Brunton.
635
+ Data-driven discovery of coordinates and governing equa-
636
+ tions. Proc. Natl. Acad. Sci. U.S.A. 116, 22445-22451
637
+ (2019).
638
+ [38] B. Chen, K. Huang, S. Raghupathi, I. Chandratreya, Q.
639
+ Du, and H. Lipson, Automated discovery of fundamental
640
+ variables hidden in experimental data, Nat. Comput. Sci.
641
+ 2, 433-442 (2022).
642
+ [39] J. Bakarji, K. Champion, J. N. Kutz, and S. L. Brun-
643
+ ton. Discovering governing equations from partial mea-
644
+ surements with deep delay autoencoders. arXiv preprint
645
+ arXiv:2201.05136 (2022).
646
+ [40] B. G. Chen, N. Upadhyaya, and V. Vitelli. Nonlinear con-
647
+ duction via solitons in a topological mechanical insulator.
648
+ Proc. Natl. Acad. Sci. U.S.A. 111, 13004-13009 (2014).
649
+ [41] Z. G. Nicolaou, D. J. Case, E. B. Wee, M. M. Driscoll,
650
+ and A. E. Motter. Heterogeneity-stabilized homogeneous
651
+ states in driven media. Nat. Comm. 12, 4486 (2021).
652
+
653
+ 1
654
+ Supplementary Material for “Data-driven discovery and extrapolation
655
+ of parameterized pattern-forming dynamics”
656
+ Zachary G. Nicolaou, Guanyu Huo,Yihui Chen, Steven L. Brunton, and J. Nathan Kutz
657
+ S1.
658
+ NUMERICAL INTEGRATION
659
+ For the complex Ginzburg-Landau equation, we use a pseudospectral integration method. We take a
660
+ periodic domain of size of size L = 32π in each direction and discretize using Nx = Ny = 128 grid points in
661
+ each spatial direction. Derivatives are calculated using fast Fourier transforms, and the discretized system
662
+ is integrated with a 4(5) Runge-Kutta-Fehlberg method (which is also used for the other equations, with
663
+ relative and absolute error tolerances of 10−6). To produce states in the dynamical phases of interest, we
664
+ take random initial conditions A0 = �
665
+ nm αnmeiknm·x + ϵeik2 2·x, where αnm are complex random Gaussian
666
+ amplitudes with mean zero and variance σ2/(1 + n2 + m2), knm = 2π(nˆx + mˆy)/L, the sum ranges over
667
+ −2 ≤ n, m ≤ 2, and ϵ is the scale of an initial plane wave perturbation with wavevector k2 2. The mode
668
+ amplitudes are determined by σ = 0.1, 0.1, 0.1, 1.0 and ϵ = 0.01, 0.01, 1.0, 0.01 for the four trajectories used
669
+ in the main text. The system is allowed to approach an attractor for the first 90 time units, then the
670
+ trajectory is formed by the next 10 time units, in steps of 0.01. We also provide an animation showing the
671
+ phase and amplitude for longer runs of 100 time units (Fig. S1). A similar pseudospectral approach was
672
+ used for the Oregonator and Swift-Hohenberg examples, but, in the Swift-Hohenberg case, with Nx = 256
673
+ discretization points, a domain of size L = 64π, an integration time of 5 time units, and random initial
674
+ condition given by the real part of u0 = �20
675
+ n=−20 αneiknx with kn = 2πn/L and αn complex random Gaussian
676
+ amplitudes with mean zero and variance 1.0/(1 +
677
+
678
+ |n|)2.
679
+ FIG. S1. Snapshot of the animation showing the phase φ and amplitude r of the trajectories, where A = reiφ.
680
+
681
+ 0
682
+ r/2
683
+ 3π/2
684
+ 2rl
685
+ L/2
686
+ 1.2
687
+ 0
688
+ -L/2
689
+ 0.9
690
+ -L
691
+ 0.6 r
692
+ L
693
+ L/2
694
+ 0.3
695
+ 0
696
+ 0.0
697
+ -L/2
698
+ -L
699
+ -L/2
700
+ 0
701
+ L/2
702
+ 7-7
703
+ -L/2
704
+ 0
705
+ L/2
706
+ 7
707
+ x
708
+ x2
709
+ S2.
710
+ DEMONSTRATIONS
711
+ Demonstrations of SINDyCP in discrete maps, ODEs and PDEs are shown in Fig. S2. The left panels
712
+ illustrate the logistic map,
713
+ xn+1 = rxn(1 − xn),
714
+ (S1)
715
+ which is a discrete-time system with a single dependent variable xn and a single parameter r. This equation
716
+ is the model for a universal period-doubling route to chaos as the parameter r increases past 3.56995. We
717
+ perform the SINDyCP fit using four sample trajectories of 1000 iterations, corresponding to parameter
718
+ values r = 3.6, 3.7, 3.8, 3.9 (red dotted lines in Fig. S2). We employ a library consisting of polynomials up to
719
+ third order in the dependent variable xn and linear functions of the control parameter r, and the SINDyCP
720
+ approach correctly identifies the parameterized equation. The middle panels illustrate the Lorenz system,
721
+ ˙x = σ(y − x), ˙y = x(ρ − z) − y, ˙z = xy − βz,
722
+ (S2)
723
+ which consists of three ordinary differential equations in three dependent variables x, y, and z and three
724
+ parameters σ, ρ and β. This equation exhibits the iconic butterfly-shaped Lorenz attractor for certain
725
+ parameter values.
726
+ We perform the SINDyCP fit using five sample trajectories that have converged to
727
+ their attractors, corresponding to the randomly selected parameter values σ = 10.0, 9.8, 9.9, 10.3, 9.5, ρ =
728
+ 27.6, 28.2, 28.3, 27.6, 28.1, and β = 3.1, 2.4, 2.4, 2.3, 2.4, respectively. We use feature and parameter libraries
729
+ consisting of polynomials up to fourth order in the dependent variables (x, y, z) and linear functions in
730
+ the parameters (σ, ρ, β), and the SINDyCP approach again correctly identifies the parameterized equation.
731
+ Finally, the right panels illustrate the CGLE described in the main text.
732
+ SINDyCP ft
733
+ Input data
734
+ Model
735
+ Logistic map
736
+ Lorenz system
737
+ Complex Ginzburg-Landau equation
738
+ FIG. S2.
739
+ Demonstrations of the SINDyCP approach for three models (top row) of nonlinear dynamics.
740
+ Several
741
+ trajectories produced from different parameter values (middle row) are supplied as input, and the SINDyCP fit
742
+ (bottom row) correctly identifies the governing equations in each case.
743
+
744
+ 3
745
+ S3.
746
+ OREGONATOR MODEL AND NORMAL FORM TRANSFORMATION
747
+ We mainly follow the analyses of the Oregonator model in Refs. [30,32], with realistic parameter values
748
+ shown in Table I. The fixed point (CX, CY , CZ) = (C0
749
+ X, C0
750
+ Y , C0
751
+ Z) undergoes a Hopf bifurcation as µ increases
752
+ from zero, leading to oscillatory chemical dynamics. For small µ, the weakly nonlinear theory follows from
753
+ a perturbative expansion of the model. Take x ≡ (CX, CY , CZ) − (C0
754
+ X, C0
755
+ Y , C0
756
+ Z) and express Eqs. (4)-(6)
757
+ as ˙x = F(x). Define the multilinear operators of partial derivatives Fxn(ei1, · · · , ein) = ∂nF/∂xi1 · · · ∂xin
758
+ with ei the ith component unit vector. Then the Taylor expansion for the system is
759
+ ˙x = (∂F/∂µ) µ + Fx1(x) + (∂F/∂µ)x1 (x)µ + 1
760
+ 2Fx2(x, x) + 1
761
+ 6Fx3(x, x, x) + D · ∇2x + · · · ,
762
+ (S3)
763
+ where D is a diagonal matrix with elements DX, DY and DZ. We develop a transformation x = y+h(y, µ)
764
+ perturbatively, where y ≡ Aeiω0tu + ¯Ae−iω0t¯u. Here u is one of the critical eigenvectors of the Jacobian
765
+ matrix Fx1 with eigenvalue λ = iω0 (with zero real part for µ = 0) and overbars represent complex
766
+ conjugates, and we also define the corresponding left eigenvector at u⊥. The near-identity transformation
767
+ function h(y, µ) is selected so as to eliminate the non-resonant terms in the evolution equation of A, which
768
+ can be accomplished under general conditions. This results in an amplitude equation ˙A = µσA + g|A|2A +
769
+ d∇2A, where
770
+ σ = u⊥ · (∂F/∂µ)x1 (u) − u⊥ · Fx2
771
+
772
+ u, (Fx1)−1 (∂F/∂µ)
773
+
774
+ ,
775
+ (S4)
776
+ g = 1
777
+ 2u⊥ · Fx3 (u, u, ¯u) − u⊥ · Fx2
778
+
779
+ u, [Fx1]−1 [Fx2 (u, ¯u)]
780
+
781
+ − 1
782
+ 2u⊥ · Fx2
783
+
784
+ ¯u,
785
+
786
+ Fx1 −
787
+
788
+ λ − ¯λ
789
+
790
+ I
791
+ �−1 [Fx2 (u, u)]
792
+
793
+ ,
794
+ (S5)
795
+ d = u⊥ · D · u.
796
+ (S6)
797
+ By rescaling the amplitude by a factor of µ1/2, time by a factor of 1/µ, and space by a factor of 1/µ1/2 and
798
+ employing additional rescalings to unitize the real components and eliminate the mean rotation, we can
799
+ arrive at the CGLE in Eq.(2), where b ≡ Im(d)/Re(d) = b0 = 0.173 and c ≡ −Im(g)/Re(g) = c0 = 2.379.
800
+ As expected, these parameter values correspond to the amplitude turbulence regime of the CGLE.
801
+ For our numerical simulations, we use a spatial domain of length L = 0.4/µ1/2 cm and an integration
802
+ time of T = 200/µ s, where we scaled by µ to ensure the trajectories have corresponding scales.
803
+ We
804
+ strobe the time in steps of 5.94804 s, which corresponds to the critical frequency of the instability. We
805
+ then interpolate the time series in steps of T/1000 to generate the trajectories. The first 200 time steps
806
+ are discarded as the trajectories relax to their attractors. The next 400 time steps are used to train the
807
+ SINDyCP model, while the remaining 400 steps are used as test trajectories to evaluate the R2 scores. We
808
+ finally employ the normal form transformation described above for the SINDyCP model to evaluate the
809
+ parameterized b(µ) and c(µ) shown in Fig. 2(b) of the main text. Consistently, the normal form parameters
810
+ very closely approximate the analytic results b(0) ≈ b0 and c(0) ≈ c0, but significant variations emerge for
811
+ larger µ.
812
+ k1 k2 k3
813
+ k4
814
+ k5 DX
815
+ DY
816
+ DZ
817
+ CH CA CB/(1 − µ) ν
818
+ 2 106 10 2 × 103 1 10−5 1.6 × 10−5 0.6 × 10−5 0.5
819
+ 1
820
+ 0.787
821
+ 1
822
+ TABLE I. Parameter values for the Oregonator model, in cgs units (suppressed for brevity).
823
+
824
+ 4
825
+ S4.
826
+ WEAK FORMULATION IMPLEMENTATION
827
+ We refer the reader to Refs. [25-28] for the theory of the weak formulation of SINDy. Here, we only
828
+ briefly describe our efficient numerical integration method for the weak formulation used in pysindy. We
829
+ suppose that the spatial grid is one-dimensional, for the moment, and the values of the coordinates on the
830
+ grid points are xi. The weak form requires us to calculate the integral of interpolated data f(x) weighted
831
+ by the dth derivatives of test function φ(x),
832
+ I(d) ≡
833
+ � xN
834
+ x0
835
+ f(x)φ(d)(x)dx.
836
+ (S7)
837
+ We choose to use test functions φ(x) = (x2 − 1)p in our implementation, and thus their dth derivatives are
838
+ φ(d)(x) =
839
+
840
+ ∂xd (x2 − 1)p =
841
+ p
842
+
843
+ k=0
844
+
845
+ p
846
+ k
847
+
848
+ (−1)k
849
+ (2(p − k))!
850
+ (2(p − k) − d)!x2(p−k)−d.
851
+ (S8)
852
+ We are provided with some feature values ui at the grid points, and we consider the value of a library
853
+ function f applied to that feature, fi ≡ f(ui).
854
+ We linearly interpolate the function as f(x) = fi +
855
+ x−xi
856
+ xi+1−xi (fi+1 − fi) where xi ≤ x ≤ xi+1. Expanding the interpolation, and integrating the xφ(d)(x) terms
857
+ by parts,
858
+ I(d) =
859
+ N−1
860
+
861
+ i=0
862
+ � xi+1
863
+ xi
864
+
865
+ fi +
866
+ x − xi
867
+ xi+1 − xi
868
+ (fi+1 − fi)
869
+
870
+ φ(d)(x)dx
871
+ =
872
+ N−1
873
+
874
+ i=0
875
+ fixi+1 − fi+1xi
876
+ xi+1 − xi
877
+
878
+ Φ(d)(xi+1) − Φ(d)(xi)
879
+
880
+ + fi+1 − fi
881
+ xi+1 − xi
882
+
883
+ Φ(d−1)(xi+1) − Φ(d−1)(xi)
884
+
885
+ ,
886
+ (S9)
887
+ where Φ(d)(x) are the antiderivatives of φ(d) [i.e. Φ(d)(x) = φ(d−1)(x) for d > 0].
888
+ By relabelling the dummy summation variables, we can recast Eq. (S9) as a dot product between the
889
+ input data fj and a weight wj
890
+ I(d) =
891
+ N−1
892
+
893
+ j=0
894
+ wj · fj,
895
+ (S10)
896
+ with
897
+ wj ≡ xj+1
898
+
899
+ Φ(d)(xj+1) − Φ(d)(xj)
900
+
901
+ xj+1 − xj
902
+ − xj−1
903
+
904
+ Φ(d)(xj) − Φ(d)(xj−1)
905
+
906
+ xj − xj−1
907
+ + Φ(d−1)(xj) − Φ(d−1)(xj−1)
908
+ xj − xj−1
909
+ − Φ(d−1)(xj+1) − Φ(d−1)(xj)
910
+ xj+1 − xj
911
+ ,
912
+ (S11)
913
+ where 0 < j < N − 1. At the left and right sides of the domain (for j = 0 and j = N − 1), we must adjust
914
+ the weights to correct for boundary effects,
915
+ w0 ≡ x1
916
+
917
+ Φ(d)(x1) − Φ(d)(x0)
918
+
919
+ x1 − x0
920
+ − Φ(d−1)(x1) − Φ(d−1)(x0)
921
+ x1 − x0
922
+ ,
923
+ (S12)
924
+ wN−1 ≡ −xN−2
925
+
926
+ Φ(d)(xN−1) − Φ(d)(xN−2)
927
+
928
+ xN−1 − xN−2
929
+ + Φ(d−1)(xN−1) − Φ(d−1)(xN−2)
930
+ xN−1 − xN−2
931
+ .
932
+ (S13)
933
+
934
+ 5
935
+ Expressing the integrals along each dimension as dot products [Eq. (S10)] enables efficient vectorization
936
+ with BLAS operations, and the integration weights [Eq. (S11)-(S13)] only need to be evaluated a single
937
+ time when the library is first initialized (in a vectorized fashion). We further vectorize the code by forming
938
+ tensor products over all integration dimensions to calculate multidimensional integrals using a single tensor
939
+ dot product.
940
+
GtE0T4oBgHgl3EQfzgJq/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
I9AzT4oBgHgl3EQfVPxY/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,156 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf,len=155
2
+ page_content='arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
3
+ page_content='01280v1 [math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
4
+ page_content='NA] 3 Jan 2023 An asymptotic formula for Aldaz-Kounchev-Render operators on the hypercube Ana-Maria Acua, Ioan Ra¸sab aLucian Blaga University of Sibiu, Department of Mathematics and Informatics, Romania, e-mail: anamaria.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
5
+ page_content='acu@ulbsibiu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
6
+ page_content='ro bTechnical University of Cluj-Napoca, Faculty of Automation and Computer Science, Department of Mathematics, Str.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
7
+ page_content=' Memorandumului nr.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
8
+ page_content=' 28, 400114 Cluj-Napoca, Romania e-mail: ioan.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
9
+ page_content='rasa@math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
10
+ page_content='utcluj.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
11
+ page_content='ro Abstract We prove a version of a conjecture concerning the asymptotic behavior of the Aldaz-Kounchev-Render operators on the hypercube.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
12
+ page_content=' Keywords: Aldaz-Kounchev-Render operators;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
13
+ page_content=' Bernstein operator;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
14
+ page_content=' Voronovskaja-type formula;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
15
+ page_content=' tensor product.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
16
+ page_content=' 2010 MSC: 41A36 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
17
+ page_content=' Introduction Let B[1] n : C[0, 1] → C[0, 1] be the classical Bernstein operator defined as B[1] n f(x) = n � i=0 f � i n � pn,i(x), where pn,i(x) = �n i � xi(1 − x)n−i, x ∈ [0, 1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
18
+ page_content=' For a fixed j ∈ N, j ≥ 2 and for n ≥ j, Aldaz, Kounchev and Render [2] introduced a polynomial operator B[1] n,j : C[0, 1] → C[0, 1] that fixes e0 and ej, investigated its approximation properties and gave applications to CAGD.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
19
+ page_content=' The operator is explicitly given by B[1] n,jf(x) = n � k=0 f � tj n,k � pn,k(x), where tj n,k = � k(k − 1) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
20
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
21
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
22
+ page_content=' (k − j + 1) n(n − 1) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
23
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
24
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
25
+ page_content=' (n − j + 1) �1/j .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
26
+ page_content=' The Voronovskaja-type formula for the sequence (B[1] n,j)n≥1 was conjectured in [4] and proved in [3], [5].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
27
+ page_content=' Preprint submitted to .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
28
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
29
+ page_content=' January 4, 2023 For f ∈ C([0, 1]2), the tensor product B[1] n ⊗ B[1] n is given by B[2] n f(x, y) := (B[1] n ⊗ B[1] n )f(x, y) = n � k=0 n � l=0 f �k n, l n � pn,k(x)pn,l(y).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
30
+ page_content=' (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
31
+ page_content='1) Let B[1] n,j : C[0, 1] → C[0, 1] be the AKR operator and (x, y) ∈ [0, 1]2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
32
+ page_content=' Then, for f ∈ C([0, 1]2), the tensor product B[1] n,j ⊗ B[1] n,j is given by B[2] n,jf(x, y) := (B[1] n,j ⊗ B[1] n,j)f(x, y) = n � k=0 n � l=0 f � tj n,k, tj n,l � pn,k(x)pn,l(y), (x, y) ∈ [0, 1]2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
33
+ page_content=' (1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
34
+ page_content='2) A conjecture concerning the Voronovskaja-type formula for the sequence (B[2] n,j) was formulated in [1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
35
+ page_content=' The aim of this paper is to prove a version of this conjecture.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
36
+ page_content=' 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
37
+ page_content=' Proof of Conjecture For the sake of conciseness we consider only the case j = 2, but obviously the proof can be extended to arbitrary j.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
38
+ page_content=' Let k and n be integers, n ≥ 2, 0 ≤ k ≤ n.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
39
+ page_content=' Define R(n, k) := k n − � k(k − 1) n(n − 1) − 1 2n + k 2n2 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
40
+ page_content=' It is elementary to prove that R(n, 0) = − 1 2n, (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
41
+ page_content='1) R(n, k) ≥ 0, k = 1, 2, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
42
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
43
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
44
+ page_content=' , n, (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
45
+ page_content='2) 0 ≤ k n − � k(k − 1) n(n − 1) ≤ 1 n.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
46
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
47
+ page_content='3) Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
48
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
49
+ page_content=' If 0 < x ≤ 1, then lim n→∞ n n � k=1 pn,k(x)R(n, k) = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
50
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
51
+ page_content='4) Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
52
+ page_content=' Let x ∈ (0, 1] and f ∈ C2[0, 1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
53
+ page_content=' It is known (see [3], [5]) that lim n→∞ n(B[1] n,2f(x) − f(x)) = x(1 − x) 2 f ′′(x) − 1 − x 2 f ′(x).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
54
+ page_content=' It is also well known that lim n→∞ n(B[1] n f(x) − f(x)) = x(1 − x) 2 f ′′(x).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
55
+ page_content=' 2 It follows that lim n→∞ n � B[1] n,2f(x) − B[1] n f(x) � = −1 − x 2 f ′(x).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
56
+ page_content=' In particular, for the function f(t) = t, we get lim n→∞ n n � k=1 pn,k(x) �� k(k − 1) n(n − 1) − k n � = −1 − x 2 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
57
+ page_content=' This can be written as lim n→∞ n n � k=1 pn,k(x) � 1 2n � 1 − k n � + R(n, k) � = 1 − x 2 , i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
58
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
59
+ page_content=', 1 2 lim n→∞ n � k=1 pn,k(x) � 1 − k n � + lim n→∞ n n � k=1 pn,k(x)R(n, k) = 1 − x 2 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
60
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
61
+ page_content='5) Let us remark that 1 2 lim n→∞ n � k=1 pn,k(x) � 1 − k n � = 1 2 lim n→∞ � B[1] n (1 − t;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
62
+ page_content=' x) − (1 − x)n� = 1 2(1 − x).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
63
+ page_content=' Combined with (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
64
+ page_content='5) this leads to (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
65
+ page_content='4), and the proof is finished.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
66
+ page_content=' Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
67
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
68
+ page_content=' Let 0 < x ≤ 1, 0 < y ≤ 1, f ∈ C2([0, 1]2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
69
+ page_content=' Then lim n→∞ n � B[2] n,2f(x, y) − f(x, y) � = x(1 − x) 2 f ′′ x2(x, y) + y(1 − y) 2 f ′′ y2(x, y) − 1 − x 2 f ′ x(x, y) − 1 − y 2 f ′ y(x, y).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
70
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
71
+ page_content='6) Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
72
+ page_content=' First we have n � B[2] n,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
73
+ page_content='2f(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
74
+ page_content=' y) − B[2] n f(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
75
+ page_content=' y) � = n n � k=0 n � l=0 pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
76
+ page_content='k(x)pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
77
+ page_content='l(y) � f �� k(k − 1) n(n − 1),' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
78
+ page_content=' � l(l − 1) n(n − 1) � − f �k n,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
79
+ page_content=' l n �� = Enf(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
80
+ page_content=' y) + Fnf(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
81
+ page_content=' y) + Gnf(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
82
+ page_content=' y),' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
83
+ page_content=' 3 where Enf(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
84
+ page_content=' y) := n n � k=0 n � l=0 pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
85
+ page_content='k(x)pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
86
+ page_content='l(y) �� k(k − 1) n(n − 1) − k n � f ′ x �k n,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
87
+ page_content=' l n � ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
88
+ page_content=' Fnf(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
89
+ page_content=' y) := n n � k=0 n � l=0 pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
90
+ page_content='k(x)pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
91
+ page_content='l(y) �� l(l − 1) n(n − 1) − l n � f ′ y �k n,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
92
+ page_content=' l n � ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
93
+ page_content=' Gnf(x,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
94
+ page_content=' y) := n 2 n � k=0 n � l=0 pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
95
+ page_content='k(x)pn,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
96
+ page_content='l(y) \uf8f1 \uf8f2 \uf8f3 �� k(k − 1) n(n − 1) − k n �2 f ′′ x2(ξ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
97
+ page_content=' η) + 2 �� k(k − 1) n(n − 1) − k n � �� l(l − 1) n(n − 1) − l n � f ′′ xy(ξ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
98
+ page_content=' η) + �� l(l − 1) n(n − 1) − l n �2 f ′′ y2(ξ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
99
+ page_content=' η) \uf8fc \uf8fd \uf8fe ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
100
+ page_content=' for suitable (ξ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
101
+ page_content=' η) furnished by Taylor’s formula.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
102
+ page_content=' Using (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
103
+ page_content='3) we see that lim n→∞ Gnf(x, y) = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
104
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
105
+ page_content='7) Moreover, lim n→∞ Enf(x, y) = − lim n→∞ n n � k=0 n � l=0 pn,k(x)pn,l(y) � 1 2n � 1 − k n � + R(n, k) � f ′ x �k n, l n � = −1 2 lim n→∞ n � k=0 n � l=0 pn,k(x)pn,l(y) � 1 − k n � f ′ x �k n, l n � − lim n→∞ n n � k=0 n � l=0 pn,k(x)pn,l(y)R(n, k)f ′ x �k n, l n � = −1 2 lim n→∞ B[2] n ((1 − s)f ′ x(s, t);' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
106
+ page_content=' (x, y)) − lim n→∞ n n � k=1 n � l=0 pn,k(x)pn,l(y)R(n, k)f ′ x �k n, l n � + lim n→∞ n n � l=0 (1 − x)npn,l(y) 1 2nf ′ x � 0, l n � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
107
+ page_content=' The first term equals −1 2(1 − x)f ′ x(x, y).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
108
+ page_content=' 4 Moreover, using (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
109
+ page_content='2) we have �����n n � k=1 n � l=0 pn,k(x)pn,l(y)R(n, k)f ′ x �k n, l n ������ ≤ n � l=0 � n n � k=1 pn,k(x)R(n, k)∥f ′ x∥∞ � pn,l(y) = n n � k=1 pn,k(x)R(n, k)∥f ′ x∥∞, and (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
110
+ page_content='4) shows that the second term is zero.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
111
+ page_content=' The third one is also zero, and so lim n→∞ Enf(x, y) = −1 − x 2 f ′ x(x, y).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
112
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
113
+ page_content='8) Similarly, lim n→∞ Fnf(x, y) = −1 − y 2 f ′ y(x, y).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
114
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
115
+ page_content='9) Now (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
116
+ page_content='7), (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
117
+ page_content='8), (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
118
+ page_content='9) yield lim n→∞ n � B[2] n,2f(x, y) − B[2] n f(x, y) � = −1 − x 2 f ′ x(x, y) − 1 − y 2 f ′ y(x, y).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
119
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
120
+ page_content='10) On the other hand, it is well known that lim n→∞ n(B[2] n f(x, y) − f(x, y)) = x(1 − x) 2 f ′′ x2(x, y) + y(1 − y) 2 f ′′ y2(x, y).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
121
+ page_content=' (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
122
+ page_content='11) From (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
123
+ page_content='10) and (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
124
+ page_content='11) we get (2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
125
+ page_content='6) and the theorem is proved.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
126
+ page_content=' References [1] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
127
+ page_content='M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
128
+ page_content=' Acu, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
129
+ page_content=' De Marchi, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
130
+ page_content=' Ra¸sa, Aldaz–Kounchev–Render Operators and Their Approximation Properties.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
131
+ page_content=' Results Math 78, 21 (2023).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
132
+ page_content=' [2] J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
133
+ page_content='M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
134
+ page_content=' Aldaz, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
135
+ page_content=' Kounchev, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
136
+ page_content=' Render, Shape preserving properties of gener- alized Bernstein operators on extended Chebyshev spaces, Numer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
137
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
138
+ page_content=', 2009, 114(1), 1–25.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
139
+ page_content=' [3] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
140
+ page_content=' Birou, A proof of a conjecture about the asymptotic formula of a Bern- stein type operator, Results Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
141
+ page_content=' 72 (2017), 1129–1138.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
142
+ page_content=' [4] D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
143
+ page_content=' C´ardenas-Morales, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
144
+ page_content=' Garrancho, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
145
+ page_content=' Rasa, Asymptotic Formulae via a Korovkin-Type Result, Abstr.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
146
+ page_content=' Appl.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
147
+ page_content=' Anal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
148
+ page_content=' Volume 2012, Article ID 217464, 12 pages.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
149
+ page_content=' [5] I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
150
+ page_content=' Gavrea, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
151
+ page_content=' Ivan, Complete asymptotic expansions related to conjecture on a Voronovskaja-type theorem, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
152
+ page_content=' Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
153
+ page_content=' Anal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
154
+ page_content=' Appl.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
155
+ page_content=' 458 (1) (2018), 452-463.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
156
+ page_content=' 5' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/I9AzT4oBgHgl3EQfVPxY/content/2301.01280v1.pdf'}
J9FJT4oBgHgl3EQfwy0B/content/2301.11631v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:33063760956e0d6f097653cc0ceae55dfce4878b76a9316bd95f2cb87d626715
3
+ size 7782181
J9FJT4oBgHgl3EQfwy0B/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:03a32503e123eede5eaa8e1a22456737fec5b2d83c4423cb952ad9a7245e73c5
3
+ size 129173
JtE2T4oBgHgl3EQfUwcR/content/2301.03815v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:227c1bd0009058a8d7f87502755bf85b828483be8f2ba8fe59ddf1983d85963c
3
+ size 688368
JtE2T4oBgHgl3EQfUwcR/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6365be6fc00836bc2e16cac69472ac88d9712188dcf644509c8d575731358034
3
+ size 4259885
JtE2T4oBgHgl3EQfUwcR/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:009ca451a59af2c8a8fde5e6dea031b36ce7f1c12d49a881e49877f8d3ba012a
3
+ size 170996
KNE3T4oBgHgl3EQfXwo4/content/tmp_files/2301.04481v1.pdf.txt ADDED
@@ -0,0 +1,1513 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.04481v1 [physics.optics] 11 Jan 2023
2
+ “Analytical Continuation” of Flattened Gaussian Beams
3
+ Riccardo Borghi
4
+ Dipartimento di Ingegneria Civile, Informatica e delle Tecnologie Aeronautiche,
5
+ Universit`a “Roma Tre”, Via Vito Volterra 62, I-00146 Rome, Italy
6
+ A purely analytical extension of the flattened Gaussian beams [Opt. Commun. 107, 335 (1994)]
7
+ to any values of the beam order, is here proposed. Thanks to it, the paraxial propagation problem of
8
+ axially symmetric, coherent flat-top beams through arbitrary ABCD optical systems can definitely
9
+ be closed in terms of a particular bivariate confluent hypergeometric function.
10
+ I.
11
+ INTRODUCTION
12
+ Flat-top beams continue to attract a considerable at-
13
+ tention in optics: during the last five years more than
14
+ sixty papers have been published on the subject. In or-
15
+ der to model flat-top axially symmetric distributions, two
16
+ classes of different scenarios appeared: in the first one,
17
+ simple analytical profiles were employed, the most known
18
+ of them being the superGaussian (SG) [1, 2], which is for-
19
+ mally defined by
20
+ SGν(ξ) = exp(−ξ2ν) ,
21
+ (1)
22
+ where ν denotes a real parameter which controls the“flat-
23
+ ness” of the profile, with the particular case ν = 1 giving
24
+ the Gaussian profile. The symbol ξ denotes a normal-
25
+ ized radial transverse position.
26
+ Despite its mathemat-
27
+ ical simplicity, it is well known that Eq. (1) does not
28
+ allow the wavefield of paraxially propagated superGaus-
29
+ sian (i.e., for ν ̸= 1) beams to be analytically evaluated,
30
+ even within the simplest scenario, namely free space.
31
+ To overcome such a difficulty, which two or three
32
+ decades ago could represent a considerable computational
33
+ bottleneck in several practical situations, alternative ap-
34
+ proaches were proposed in 1994 and in 2002 by Gori and
35
+ Li, respectively, to conceive analytical models able to
36
+ solve the free space propagation problem. The former
37
+ was called flattened Gaussian (FG henceforth) [3], and,
38
+ differently from SG, is expressed through an explicit fi-
39
+ nite sum of terms, namely
40
+ FGN(ξ) = exp(−Nξ2)
41
+ N−1
42
+
43
+ m=0
44
+ (Nξ2)m
45
+ m!
46
+ ,
47
+ (2)
48
+ where the integer parameter N will be referred to as the
49
+ FG order. Scaling the ξ variable by the factor
50
+
51
+ N gives
52
+ the FG transverse profile a flat-topped shape which, for
53
+ N = 1, reduces to a Gaussian distribution, whereas for
54
+ N → ∞ tends to the characteristic function of the uni-
55
+ tary disk [4]. The model is computationally exact, since
56
+ the initial distribution (2) can be recast in terms of a su-
57
+ perposition of N standard Laguerre-Gauss (sLG hence-
58
+ forth) beams. Accordingly, in order to evaluate the field
59
+ propagated in free space, it was enough to sum up the
60
+ N propagated sLG, a job which can exactly be done, al-
61
+ ways [5]. In [6], a different superposition scheme of the
62
+ profile (2) was proposed, in which the sLG family was
63
+ replaced by the so- called elegant Laguerre-Gauss (eLG
64
+ henceforth) set. In this way, not only free-space propa-
65
+ gation, but also the interaction of FG beams with any
66
+ axially symmetric paraxial optical system can be dealt
67
+ with in exact terms, always through finite sums.
68
+ In 2002, Yaijun Li proposed an analytical model al-
69
+ ternative to the FG one. The idea was to impose a lo-
70
+ cal “flatness” condition, which required the first 2N ξ-
71
+ derivatives of the profile to be null at the origin ξ = 0 [7].
72
+ On using such condition, Li conceived the following ana-
73
+ lytical model:
74
+ LiGN(ξ) =
75
+ N
76
+
77
+ m=1
78
+ (−1)m−1
79
+ �N
80
+ m
81
+
82
+ exp(−mξ2) =
83
+ =
84
+
85
+ 1 −
86
+
87
+ 1 − exp
88
+
89
+ −ξ2���N
90
+ N
91
+ ,
92
+ (3)
93
+ which, differently from FG, is based on the superposi-
94
+ tion of N fundamental Gaussian beams having variable
95
+ widths.
96
+ Both Gori’s and Li’s models provide exact solutions
97
+ to the paraxial propagation problem of coherent, axially
98
+ symmetric flat-topped beams. From a merely mathemat-
99
+ ical perspective, their only own limit is represented by
100
+ the fact that, differently from SG, only positive integer
101
+ orders N can be dealt with to describe the initial flat-top
102
+ distribution. It is important to mention that, for 1D ge-
103
+ ometry (or rectangular 2D geometries), general analyt-
104
+ ical solutions were already provided, at least upon free
105
+ propagation, by modeling the flat-top profile via an error
106
+ function [8]. An attempt to extend the 2D circular FG
107
+ model to noninteger orders was also proposed in [9], but
108
+ only approximate estimates of the free space propagated
109
+ field were found within the asymptotic limit N ≫ 1.
110
+ The aim of the present paper is to solve exactly the
111
+ propagation problem of FG beams of any order (real or
112
+ even complex) through typical axially symmetric parax-
113
+ ial optical systems. To this end, the right side of Eq. (2)
114
+ will first be identified as an incomplete Gamma func-
115
+ tions, which is known to be defined onto the whole com-
116
+ plex plane, as far as both arguments are concerned. An
117
+ immediate byproduct of such identification will be the
118
+ closed form expression of the M 2 factor of FG beams
119
+ of any order, an interesting generalization of the result
120
+ found in [5]. This is shown in Sec. II of the present pa-
121
+ per. The most important results are indeed presented
122
+
123
+ 2
124
+ in Secs. III and IV. In the former, the free- space prop-
125
+ agation problem will be solved thanks to an important
126
+ class of integrals recently closed by Yuri Brychkov. Al-
127
+ though the more general propagation problem will be
128
+ solved in Sec. IV, the analysis presented in Sec. III should
129
+ be viewed as an important propaedeutical step. There, it
130
+ will be shown that a very important, but nevertheless not
131
+ so much known, class of special functions, called bivariate
132
+ hypergeometric functions, together with the correspond-
133
+ ing confluent versions, form the mathematical skeleton of
134
+ the paraxially diffracted wavefield. Bivariate hypergeo-
135
+ metric were first introduced in 1880 by Paul Appell [10],
136
+ their confluent version forty years later by Paul Hum-
137
+ bert [11]. The results we are going to present would also
138
+ give readers a partial answer about the lack, for more
139
+ than thirty years, of purely analytical solutions to the
140
+ problem of the paraxial propagation of coherent 2D flat-
141
+ topped beams.
142
+ The present work has a clear mathematical character:
143
+ for instance, dimensionless quantities will be used wher-
144
+ ever possible.
145
+ Moreover, the number of mathematical
146
+ appendices have been considerably limited, because we
147
+ strongly believe that following all most important math-
148
+ ematical steps could greatly help readers to fully grasp
149
+ the essence of our analysis, as well as the importance of
150
+ such still mysterious special functions, which will lead to
151
+ analytical, elegant, and exact solutions.
152
+ II.
153
+ PRELIMINARIES
154
+ A.
155
+ “Analytical continuation” of the FG model
156
+ Already in 1996, Sheppard & Saghafi [12] pointed out
157
+ that Eq. (2) can be given the closed form
158
+ FGN(ξ) = Γ(N, Nξ2)
159
+ Γ(N)
160
+ ,
161
+ (4)
162
+ where Γ(·) and Γ(·, ·) denote Gamma and incomplete
163
+ Gamma functions, respectively [13].
164
+ Differently from
165
+ Eq. (2), Eq. (4) is not limited to integer FG orders, but
166
+ rather it can be analytically continued to real and also
167
+ complex values of N.
168
+ As a preliminary result of the extended definition into
169
+ Eq. (4), an analytical check of Li’s“flatness condition”[7]
170
+ will now be carried out. To this end, it is sufficient to use
171
+ formulas 1.1.1.1 and 1.8.1.17 of [14] to prove, with long
172
+ but simple algebra, that
173
+ dn
174
+ dξn Γ(N, Nξ2) = −2nn!N N exp(−Nξ2) ξ2N−n
175
+ ×
176
+ [n/2]
177
+
178
+ k=0
179
+ (n − k − 1)!
180
+ 4kk!(n − 2k)! L(N−n+k)
181
+ n−k−1
182
+ (Nξ2) ,
183
+ (5)
184
+ which gives at once
185
+ � dn
186
+ dξn Γ(N, Nξ2)
187
+
188
+ ξ=0
189
+ = 0 ,
190
+ 0 ≤ n < 2 Re{N} ,
191
+ (6)
192
+ thus implying the real part of N to be chosen greater
193
+ than one.
194
+ B.
195
+ Spreading properties: closed form expression of the M 2
196
+ factor
197
+ An interesting byproduct of the extended Γ-based def-
198
+ inition into Eq. (4) is the evaluation of the M 2 factor of
199
+ FG beams, first established in [5] for N ∈ N, for nonin-
200
+ teger orders. To this end, consider an initial field distri-
201
+ bution across the plane z = 0 of a cylindrical reference
202
+ frame (r, z), say ψ0(r), given by
203
+ ψ0(r) = FGN
204
+ �r
205
+ a
206
+
207
+ =
208
+ Γ
209
+
210
+ N, N r2
211
+ a2
212
+
213
+ Γ(N)
214
+ ,
215
+ (7)
216
+ where an overall amplitude constant has been set to one
217
+ and the symbol a denotes the “width” of lat-top distri-
218
+ bution field distribution.
219
+ For simplicity, it will be set
220
+ a = 1.
221
+ The evaluation of the M 2 factor, which is defined as the
222
+ product of the normalized second order moments across
223
+ the z = 0 and the spatial frequency planes is detailed
224
+ in Appendix A, where it is proved the following closed-
225
+ form expression:
226
+ M 2 =
227
+
228
+ (N + 1) Γ(N + 1/2)
229
+ √π Γ(N + 1)
230
+
231
+ 1 −
232
+ Γ(N + 3/2)
233
+ √π Γ(N + 2)
234
+
235
+ 1 −
236
+ Γ(N + 1/2)
237
+ √π Γ(N + 1)
238
+ ,
239
+ (8)
240
+ which extends the 1996 analysis of [5] to N /∈ N. It is
241
+ worth comparing Eq. (8) with the corresponding expres-
242
+ sion of SG beam M 2 factor, namely [2]
243
+ M 2 =
244
+
245
+ Γ(2/ν)
246
+ Γ(1/ν)/ν ,
247
+ (9)
248
+ deceptively simpler. In the next two sections, our exten-
249
+ sion of the FG model will further reveal its powerfulness
250
+ and mathematical elegance.
251
+ III.
252
+ FREE-SPACE PARAXIAL PROPAGATION OF FG
253
+ BEAMS
254
+ A.
255
+ Preliminaries
256
+ Suppose the initial field distribution given by Eq. (7)
257
+ is allowed to propagate in free space. The corresponding
258
+
259
+ 3
260
+ field, say ψ(r; z), can be expressed, apart from an overall
261
+ phase factor exp(ikz), as follows:
262
+ ψ(r; z) = −i U
263
+
264
+
265
+ R2 d2ρ ψ0(ρ) exp
266
+ �iU
267
+ 2 |r − ρ|2
268
+
269
+ ,
270
+ (10)
271
+ where the Fresnel number U = ka2/z has been intro-
272
+ duced and the beam width a has been used as unit for
273
+ measuring all transverse sizes. This means that the quan-
274
+ tity r should be meant as the ratio between the trans-
275
+ verse position vector of the observation point and a. For
276
+ integer FG orders, the free space propagation problem
277
+ has already been solved in [3] by expanding the initial
278
+ field distribution ψ0 as the linear combination of a finite
279
+ number of sLG beams.
280
+ It is then sufficient to propa-
281
+ gate each sLG beam up to the observation plane and to
282
+ recombine all of them with the initial expanding coeffi-
283
+ cients for the correct value of ψ(r; z) to be retrieved. As
284
+ we are going to show in a moment, the Γ-based model
285
+ into Eq. (4) allows an exact evaluation of the propagated
286
+ wavefield (10) also for N /∈ N. It is worth recalling that,
287
+ from a mere practical perspective, the present section
288
+ could seem somewhat redundant, as in Sec. IV the more
289
+ general propagation problem within ABCD systems will
290
+ be solved.
291
+ Nevertheless, we believe what is contained
292
+ in the present section could help nonspecialist readers
293
+ to familiarize with the main notations and mathemati-
294
+ cal tools which will constitute the basis of the general
295
+ results presented into Sec. IV. In other words, it should
296
+ be considered as a useful, propaedeutical material.
297
+ We start on substituting from Eqs. (7) into Eq. (10),
298
+ which after simple algebra gives
299
+ ψ(r; z) = − i U
300
+ Γ(N) exp
301
+ �iU r2
302
+ 2
303
+
304
+ ×
305
+ � ∞
306
+ 0
307
+ dρ ρ exp
308
+ �iU
309
+ 2 ρ2
310
+
311
+ Γ
312
+
313
+ N, N ρ2�
314
+ J0(Ur ρ) ,
315
+ (11)
316
+ where J0 denotes the 0th-order Bessel function of the first
317
+ kind. It is worth recasting the incomplete Γ function as
318
+ Γ(N, Nξ)
319
+ Γ(N)
320
+ = 1 − γ(N, Nξ)
321
+ Γ(N)
322
+ ,
323
+ (12)
324
+ where γ(·, ·) denotes the “lower”incomplete gamma func-
325
+ tion. Then Eq. (11) takes on the form
326
+ ψ(r; z) =
327
+ = −i U exp
328
+ �iU r2
329
+ 2
330
+ � � ∞
331
+ 0
332
+ dρ ρ exp
333
+
334
+ −U
335
+ 2i ρ2
336
+
337
+ J0(Ur ρ)
338
+ +
339
+ i U exp
340
+ �iU r2
341
+ 2
342
+
343
+ Γ(N)
344
+ ×
345
+ � ∞
346
+ 0
347
+ dρ ρ exp
348
+
349
+ −U
350
+ 2i ρ2
351
+
352
+ γ
353
+
354
+ N, Nρ2�
355
+ J0(Ur ρ) .
356
+ (13)
357
+ The first term is identically equal to one (it is nothing
358
+ but a unitary plane wave propagating along the z-axis).
359
+ As far as the second is concerned, the following notable
360
+ formula has recently been published by Brychkov [15,
361
+ formula 9.2.20]:
362
+ � ∞
363
+ 0
364
+ dx xα−1 exp(−a x2) γ(µ, bx2) Jν(c x) =
365
+ =
366
+ 2−ν−1bµcνΓ
367
+
368
+ µ + α + ν
369
+ 2
370
+
371
+ µaµ+(α+ν)/2Γ(ν + 1)
372
+ Ψ1
373
+
374
+ µ + α + ν
375
+ 2
376
+ , µ
377
+ µ + 1, ν + 1
378
+ ����� − b
379
+ a, − c2
380
+ 4a
381
+
382
+ .
383
+ (14)
384
+ Then, on using Eqs. (13) and (14), long but straightfor-
385
+ ward algebra gives
386
+ ψ(r; z) = 1 − exp
387
+ �iU r2
388
+ 2
389
+ � �2iN
390
+ U
391
+ �N
392
+ × Ψ1
393
+
394
+ N + 1, N
395
+ N + 1, 1
396
+ ���� − 2iN
397
+ U , −iU r2
398
+ 2
399
+
400
+ .
401
+ (15)
402
+ B.
403
+ A short Tour on Bivariate Hypergeometric Functions
404
+ The symbol Ψ1 into Eq. (15) denotes a special func-
405
+ tion called bivariate confluent hypergeometric. It is worth
406
+ briefly describing the principal definitions and properties
407
+ which are important for our scopes. Function Ψ1 is for-
408
+ mally defined through the following double series power
409
+ expansion:
410
+ Ψ1
411
+
412
+ a, b
413
+ c, c′
414
+ ���� z, w
415
+
416
+ =
417
+
418
+
419
+ k=0
420
+
421
+
422
+ ℓ=0
423
+ (a)k+ℓ (b)k
424
+ (c)k(c′)ℓ
425
+ zk
426
+ k!
427
+ wℓ
428
+ ℓ! ,
429
+ (16)
430
+ valid for |z| ≤ 1.
431
+ The symbol (·)n denotes Pochham-
432
+ mer’s symbol. Another bivariate confluent hypergeomet-
433
+ ric function which will be meet in the present paper is
434
+ the function Φ1, defined by
435
+ Φ1
436
+
437
+ a, b
438
+ c
439
+ ���� z, w
440
+
441
+ =
442
+
443
+
444
+ k=0
445
+
446
+
447
+ ℓ=0
448
+ (a)k+ℓ (b)k
449
+ (c)k+ℓ
450
+ zk
451
+ k!
452
+ wℓ
453
+ ℓ! ,
454
+ (17)
455
+ valid for |z| ≤ 1.
456
+ Functions Ψ1 and Φ1 are members
457
+ of a family of functions that generalize Kummer’s con-
458
+ fluent hypergeometric function 1F1. In particular, Φ1 is
459
+ obtained from the so-called Appell function F1, defined
460
+ by
461
+ F1
462
+
463
+ a, b1, b2
464
+ c
465
+ ���� z, w
466
+
467
+ =
468
+
469
+
470
+ k=0
471
+
472
+
473
+ ℓ=0
474
+ (a)k+ℓ (b1)k (b2)ℓ
475
+ (c)k+ℓ
476
+ zk
477
+ k!
478
+ wℓ
479
+ ℓ! ,
480
+ (18)
481
+ (again valid for |z| ≤ 1), through the following limiting
482
+ definition:
483
+ Φ1
484
+
485
+ a, b
486
+ c
487
+ ���� z, w
488
+
489
+ = lim
490
+ ǫ→0 F1
491
+
492
+ a, b, 1
493
+ ǫ
494
+ c
495
+ ����� z, ǫw
496
+
497
+ ,
498
+ (19)
499
+
500
+ 4
501
+ which can be proved on first substituting the identity
502
+ lim
503
+ ǫ→0
504
+ �1
505
+ ǫ
506
+
507
+
508
+ ǫℓ = 1 ,
509
+ (20)
510
+ directly into Eq. (19), then on interchanging the limit
511
+ with the double series.
512
+ Multivariate hypergeometric and confluent hypergeo-
513
+ metric functions play a role of considerable importance
514
+ in theoretical physics and applied math. In optics, the
515
+ role of bivariate confluent hypergeometric functions in de-
516
+ scribing a large class of paraxial optical disturbances has
517
+ recently been pointed out [16, 17]. Moreover, it is worth
518
+ stressing that, from a practical viewpoint, Appell’s func-
519
+ tion F1 is nowadays part of the symbolic platform Math-
520
+ ematica, where it is computable with arbitrarily high ac-
521
+ curacies. Also the whole family of Appell functions, in-
522
+ cluding F1 as well as its three sisters F2, F3, and F4,
523
+ are currently implemented in the latest release of Maple.
524
+ It is then highly desirable that in a near future also the
525
+ set of bivariate confluent hypergeometric functions, in-
526
+ cluding Ψ1 and Φ1, could become part of such family of
527
+ “evaluable”special functions. In the meanwhile, someone
528
+ might rightly object to the practical usefulness of func-
529
+ tions that are defined through double infinite series like
530
+ those into Eqs. (16) - (18). To overcome such difficulties,
531
+ some tricks will be implemented in the rest of the paper,
532
+ tricks which are aimed at extending the validity domain
533
+ of Ψ1 and Φ1 beyond the series definitions, and then to
534
+ improve the practical usefulness of our analytical results.
535
+ C.
536
+ Free-space propagation formula
537
+ Function Ψ1 can be continued by using the following
538
+ transformation [18, formula 2.54]:
539
+ Ψ1
540
+
541
+ α, β
542
+ γ1, γ2
543
+ ���� z, w
544
+
545
+ =
546
+ =
547
+ 1
548
+ (1 − z)α Ψ1
549
+
550
+ α, γ1 − β
551
+ γ1, γ2
552
+ ����
553
+ z
554
+ z − 1,
555
+ w
556
+ 1 − z
557
+
558
+ ,
559
+ (21)
560
+ which, once substituted into Eq. (15), gives a new, closed-
561
+ form, expression of the paraxial propagated field
562
+ ψ(r; z) = 1 −
563
+ exp
564
+ �iU r2
565
+ 2
566
+
567
+ 1 + 2iN
568
+ U
569
+
570
+
571
+
572
+ 1
573
+ 1 +
574
+ U
575
+ 2iN
576
+
577
+
578
+
579
+ N
580
+ × Ψ1
581
+
582
+
583
+  N + 1, 1
584
+ N + 1, 1
585
+ ����
586
+ 1
587
+ 1 +
588
+ U
589
+ 2iN
590
+ , −
591
+ iU r2
592
+ 2
593
+ 1 + 2iN
594
+ U
595
+
596
+
597
+  ,
598
+ (22)
599
+ indubitably one of the main results of the present paper.
600
+ Waiting for Mathematica or Maple to develop their own
601
+ built-in version of Ψ1, it is worth working on the expres-
602
+ sion into Eq. (22) by using a notable integral represen-
603
+ tation found again in [18].
604
+ For the sake of clarity, all
605
+ mathematical steps are confined into Appendix B, where
606
+ it is proved that
607
+ Ψ1
608
+
609
+ N + 1, 1
610
+ N + 1, 1
611
+ ���� x, y
612
+
613
+ =
614
+ = N
615
+ � 1
616
+ 0
617
+ dξ (1 − ξ)N−1
618
+ (1 − xξ)N+1 1F1
619
+
620
+ N + 1; 1;
621
+ y
622
+ 1 − xξ
623
+
624
+ .
625
+ (23)
626
+ Equation (23) appears to be somewhat intriguing: the
627
+ wavefield of a free-space paraxially propagated FG beam
628
+ of any order can be represented via a 1D integral defined
629
+ over a finite integral. This could seem a somewhat pe-
630
+ culiar situation, due to the fact that the initial field dis-
631
+ tribution (7) has an infinite support, namely the whole
632
+ plane z = 0.
633
+ But what is, in our opinion, even more
634
+ important is that the integral representation (23) would
635
+ hardly be reachable starting from Fresnel’s integral (10),
636
+ without passing through the Ψ1 function and its trans-
637
+ formation rules. In the next section, a similar scenario
638
+ will also be found as far as the more general problem is
639
+ concerned.
640
+ IV.
641
+ PARAXIAL PROPAGATION THROUGH ABCD
642
+ SYSTEMS
643
+ A.
644
+ Preliminaries
645
+ The free-space paraxial propagation formula derived in
646
+ the previous section will now be extended to the general
647
+ case of the paraxial propagation of FG beams of any or-
648
+ der through typical paraxial optical systems with axial
649
+ symmetry, characterized by the so-called ABCD optical
650
+ matrices. For FG beams of integer order, it was found
651
+ in [6] that the propagation problem can be dealt with
652
+ in exact terms by expanding the initial field distribution
653
+ given into Eqs. (7) and (2) as a finite superposition of
654
+ so-called elegant Laguerre (eLG henceforth) beams as fol-
655
+ lows:
656
+ ψ0(r) =
657
+ N−1
658
+
659
+ n=0
660
+ (−)n
661
+
662
+ N
663
+ n + 1
664
+
665
+ eLGn
666
+ �ikr2
667
+ 2qN
668
+
669
+ ,
670
+ (24)
671
+ where the symbol eLGn(x) = exp(x)Ln(−x) will be re-
672
+ ferred to as the elegant Laguerre function of order n and
673
+ the complex radius of curvature qN = ka2
674
+ 2iN has also been
675
+ introduced. The initial distribution ψ0 is then recast as
676
+ follows:
677
+ ψ0(r) = exp
678
+ �ikr2
679
+ 2qN
680
+
681
+ GN
682
+
683
+ 1, −ikr2
684
+ 2qN
685
+
686
+ ,
687
+ (25)
688
+ where the function GN (·, ·) is defined, for integer N, as
689
+ GN (t, s) =
690
+ N−1
691
+
692
+ n=0
693
+ (−t)n
694
+
695
+ N
696
+ n + 1
697
+
698
+ Ln(s) ,
699
+ (26)
700
+
701
+ 5
702
+ In [6] it was proved that, if the initial field distribution
703
+ given by Eq. (25) feeds an axially symmetric paraxial
704
+ optical system described by the optical matrix M
705
+ M =
706
+
707
+
708
+ A B
709
+ C D
710
+
711
+  ,
712
+ (27)
713
+ then the wavefield at the output plane of the system, say
714
+ ψ1(r), takes on the following form:
715
+ ψ1(r) =
716
+ =
717
+ exp
718
+ � ikr2
719
+ 2QN
720
+
721
+ A
722
+ 1
723
+ 1 +
724
+ B
725
+ A qN
726
+ GN
727
+
728
+
729
+
730
+
731
+ 1
732
+ 1 +
733
+ B
734
+ A qN
735
+ ,
736
+ kr2
737
+ 2iA2 qN
738
+ 1 +
739
+ B
740
+ A qN
741
+
742
+
743
+
744
+  ,
745
+ (28)
746
+ where an overall phase factor exp(ikℓ) (with ℓ being the
747
+ optical lenght) will be tacitly assumed and QN denotes
748
+ the complex quantity
749
+ QN = A qN + B
750
+ C qN + D .
751
+ (29)
752
+ The problem of extending the function GN (t, s) to N /∈ N
753
+ will now be addressed.
754
+ B.
755
+ Extension of the function GN (t, s) to N /∈ N
756
+ The starting point is the following Laplace transform
757
+ representation of GN(t, s) established in [9]:
758
+ GN(t, s) = exp(s)
759
+ � ∞
760
+ 0
761
+ dξ exp(−ξ) J0
762
+
763
+ 2
764
+
765
+ s ξ
766
+
767
+ L(1)
768
+ N−1(ξ t) .
769
+ (30)
770
+ For N ∈ N, the Laguerre polynomials L(1)
771
+ N−1 can be writ-
772
+ ten as
773
+ L(1)
774
+ N−1(ξ t) =
775
+ N−1
776
+
777
+ n=0
778
+ Ln(ξ t) ,
779
+ (31)
780
+ so that, on substituting from Eq. (31) into Eq. (30), it is
781
+ found
782
+ GN(t, s) =
783
+ = exp(s)
784
+ N−1
785
+
786
+ n=0
787
+ � ∞
788
+ 0
789
+ dξ exp(−ξ) J0
790
+
791
+ 2
792
+
793
+ s ξ
794
+
795
+ Ln(ξ t) =
796
+ =
797
+ N−1
798
+
799
+ n=0
800
+ (1 − t)n Ln
801
+ � st
802
+ t − 1
803
+
804
+ ,
805
+ (32)
806
+ where in the last passage, [22, formula 3.24.6.2] has been
807
+ used. Equation (32) is a valid alternative, for N ∈ N, to
808
+ the definition given into Eq. (26). For the scopes of the
809
+ present paper, its importance stems from the fact that
810
+ the quantity GN can also be thought of as function of
811
+ two new variables, namely
812
+
813
+
814
+
815
+
816
+
817
+
818
+
819
+
820
+
821
+
822
+
823
+
824
+
825
+
826
+
827
+
828
+
829
+ 1 − t =
830
+ 1
831
+ 1 + AqN
832
+ B
833
+ ,
834
+ st
835
+ t − 1 = ikr2
836
+ 2AB
837
+ 1
838
+ 1 +
839
+ B
840
+ AqN
841
+ ,
842
+ (33)
843
+ and this will reveal of a certain importance in the rest of
844
+ our analysis.
845
+ In order to extend the integral into Eq. (30) to N /∈
846
+ N, the following notable formula, again established by
847
+ Brychkov [15], will be employed:
848
+ � ∞
849
+ 0
850
+ xα−1 exp(−ax) Jν(b√x) L(λ)
851
+ n (cx) dx =
852
+ =
853
+ � b
854
+ 2
855
+ �ν Γ
856
+
857
+ α + ν
858
+ 2
859
+
860
+ (λ + 1)n
861
+ n! aα+ν/2Γ(ν + 1) Ψ1
862
+
863
+ α + ν
864
+ 2 , −n
865
+ λ + 1, ν + 1
866
+ �����
867
+ c
868
+ a, − b2
869
+ 4a
870
+
871
+ .
872
+ (34)
873
+ In particular, on letting α = 1, a = 1, ν = 0, b = 2√s,
874
+ t = c, n = N − 1, and λ = 1, Laplace’s transform into
875
+ Eq. (30) takes on the form
876
+ GN(t, s) = N exp(s) Ψ1
877
+
878
+ 1, 1 − N
879
+ 2, 1
880
+ ���� t, −s
881
+
882
+ .
883
+ (35)
884
+ Again, it can be appreciated how the confluent hyperge-
885
+ ometric function Ψ1 constitutes the mathematical skele-
886
+ ton of the propagated field. But there is more. In Ap-
887
+ pendix C, the following relationship has been established:
888
+ Ψ1
889
+
890
+ 1, 1 − N
891
+ 2, 1
892
+ ���� t, −s
893
+
894
+ =
895
+ =
896
+ exp(−s)
897
+ (1 − t)1−N Φ1
898
+
899
+ 1 − N, 1
900
+ 2
901
+ ����
902
+ t
903
+ t − 1,
904
+ st
905
+ t − 1
906
+
907
+ ,
908
+ (36)
909
+ where Φ1 is the confluent hypergeometric function de-
910
+ fined by Eq. (17). On substituting from Eq. (36) into
911
+ Eq. (35), we have
912
+ GN(t, s) = N (1 − t)N−1 Φ1
913
+
914
+ 1 − N, 1
915
+ 2
916
+ ����
917
+ t
918
+ t − 1,
919
+ st
920
+ t − 1
921
+
922
+ (37)
923
+ so that Eq. (28) eventually becomes
924
+ ψ1(r) = exp
925
+ � ikr2
926
+ 2QN
927
+ � qNN
928
+ B
929
+
930
+
931
+
932
+ 1
933
+ 1 + A qN
934
+ B
935
+
936
+
937
+
938
+ N
939
+ × Φ1
940
+
941
+
942
+
943
+
944
+ 1 − N, 1
945
+ 2
946
+ ���� − A qN
947
+ B
948
+ , ikr2
949
+ 2AB
950
+ 1
951
+ 1 +
952
+ B
953
+ AqN
954
+
955
+
956
+
957
+  .
958
+ (38)
959
+
960
+ 6
961
+ Equation (38) summarizes the main result of the present
962
+ paper: the general FG beam paraxial propagation prob-
963
+ lem is reduced to the evaluation of the bivariate confluent
964
+ hypergeometric Φ1.
965
+ Again, it is possible to give Eq. (38) a different dress on
966
+ using the following integral representation of Φ1, estab-
967
+ lished in 2012 by Brychkov and Saad [19, formula 3.4]:
968
+ Φ1
969
+
970
+ a, 1
971
+ 2
972
+ ���� w, z
973
+
974
+ =
975
+ = (1 − w)1−a
976
+ � 1
977
+ 0
978
+ dξ (1 − w ξ)a−2
979
+ 1F1(a; 1; zξ) ,
980
+ (39)
981
+ which eventually leads to
982
+ ψ1(r) = exp
983
+ � ikr2
984
+ 2QN
985
+ � qNN
986
+ B
987
+
988
+
989
+
990
+ 1
991
+ 1 + A qN
992
+ B
993
+
994
+
995
+
996
+ N
997
+ ×
998
+ � 1
999
+ 0
1000
+
1001
+
1002
+ 1 + A qN
1003
+ B
1004
+ ξ
1005
+ �N+1 1F1
1006
+
1007
+
1008
+
1009
+ 1 − N; 1; ikr2
1010
+ 2AB
1011
+ ξ
1012
+ 1 +
1013
+ B
1014
+ AqN
1015
+
1016
+
1017
+
1018
+  .
1019
+ (40)
1020
+ Similarly as it was found for the free-space propagation
1021
+ into Eq. (23), also the integral representation of ψ1 given
1022
+ by Eq. (40) turns out to be defined onto a finite interval
1023
+ [0, 1], despite the infinite support of both the initial field
1024
+ distribution ψ0, as well as its Fourier transform. In the
1025
+ present case, however, at least a qualitative explanation
1026
+ of such a mathematical counterintuitive behavior can be
1027
+ grasped by estimating the right side of Eq. (40) within the
1028
+ asymptotic limit N → ∞, which corresponds to replace
1029
+ the initial FG beam distribution ψ0 by that emerging
1030
+ from a circular hole of radius a.
1031
+ In particular, the asymptotics can be carried out in an
1032
+ elementary way, by first noting that QN → B/D and
1033
+ that
1034
+ lim
1035
+ N→∞
1036
+ 1
1037
+
1038
+ 1 + A qN
1039
+ B
1040
+ ξ
1041
+ �N+1 = exp
1042
+
1043
+ iA ka2
1044
+ 2B
1045
+ ξ
1046
+
1047
+ .
1048
+ (41)
1049
+ As far as Kummer’s function inside the integral is
1050
+ concerned, the following asymptotics holds [13, for-
1051
+ mula 13.8.13]:
1052
+ 1F1(1 − N; 1; z) ∼ exp(z/2) J0
1053
+
1054
+ 2
1055
+
1056
+ N z
1057
+
1058
+ ,
1059
+ N ≫ 1 ,
1060
+ (42)
1061
+ which, once substituted into Eq. (40) together with
1062
+ Eq. (41), leads to
1063
+ ψ1(r) ∼ U
1064
+ 2i exp
1065
+
1066
+ iUD
1067
+ 2
1068
+ �r
1069
+ a
1070
+ �2�
1071
+ ×
1072
+ � 1
1073
+ 0
1074
+ dξ exp
1075
+
1076
+ iA U
1077
+ 2
1078
+ ξ
1079
+
1080
+ J0
1081
+
1082
+ U r
1083
+ a
1084
+
1085
+ ξ
1086
+
1087
+ ,
1088
+ N ≫ 1 ,
1089
+ (43)
1090
+ where now U = ka2/B.
1091
+ Finally, it is not difficult to convince that Eq. (43)
1092
+ is nothing but von Lommel’s integral [23], namely, the
1093
+ result of Collins’ integral for an incident wavefield ψ0 =
1094
+ circ(r/a), as it should be expected.
1095
+ V.
1096
+ CONCLUSIONS
1097
+ Even today, the term“superGaussian beam”is synony-
1098
+ mous of flat-topped beam, despite the indisputable lim-
1099
+ its, both practical and theoretical, of the SG model and
1100
+ the availability of more efficient analytical approaches.
1101
+ For rectangular geometries, Sedukhin’s work should have
1102
+ contributed to identify flat-topped profiles with an error
1103
+ function. For two-dimensional, axially symmetric geome-
1104
+ tries, Gori’s and Li’s models, despite allowing to solve ex-
1105
+ actly the paraxial propagation problem, to date continue
1106
+ struggling to supplant the obsolete SG model.
1107
+ In the present paper, the FG model has been general-
1108
+ ized to any values, no longer necessarily integer, of the or-
1109
+ der N. In doing this, use has been made of the suggestion,
1110
+ dating back more than twenty-five years ago, by Shep-
1111
+ pard & Saghafi to mathematically identify the model FG
1112
+ through an incomplete Gamma function. From a merely
1113
+ technical viewpoint, our work rests on some beautiful re-
1114
+ sults recently established by Brychkov and co-workers. In
1115
+ this way, it has been possibile to analytically express the
1116
+ optical wavefield generated by the propagation of such
1117
+ flat-topped “Γ-beams”of any order through arbitrary ax-
1118
+ ially symmetric paraxial optical system (free space in-
1119
+ cluded) in terms of a single bivariate confluent hyperge-
1120
+ ometric function.
1121
+ Our model is purely analytical and provided purely an-
1122
+ alytical closed expressions of the paraxially propagated
1123
+ wavefield.
1124
+ It is a rare situation in physics in general
1125
+ and in optics in particular.
1126
+ The ubiquitous presence
1127
+ of less and less known special functions, such as bivari-
1128
+ ate hypergeometric ones certainly are, also constitutes
1129
+ in our opinion an added value of the present work. We
1130
+ strongly encourage our readers to go through an interest-
1131
+ ing paper written more than twenty years ago by Michael
1132
+ Berry [24], whose content seems nowadays more than ever
1133
+ more relevant. In particular, the current availability of
1134
+ powerful computational platforms, such as Mathematica
1135
+ and Maple, will allow in the future to increase the set of
1136
+ special functions whose evaluation could be implemented
1137
+ at arbitrarily high accuracies. We hope bivariate con-
1138
+ fluent hypergeometric functions, including of course Ψ1
1139
+ and Φ1, could soon become part of such a mathematical
1140
+ weaponry.
1141
+ Acknowledgements
1142
+ I wish to thank Turi Maria Spinozzi for his help during
1143
+ the preparation of the manuscript.
1144
+
1145
+ 7
1146
+ Appendix A: Proof of Eq. (8)
1147
+ The M 2 factor is defined by
1148
+ M 2 = 2π σr σp ,
1149
+ (A1)
1150
+ where σr and σp denote the widths across the plane z = 0
1151
+ and the plane of spatial frequencies, respectively, both of
1152
+ them normalized to the beam energy. Due to the axial
1153
+ symmetry, σr can then be expressed (in units of a) as
1154
+ follows:
1155
+ σ2
1156
+ r =
1157
+ � ∞
1158
+ 0
1159
+ dr r3 ψ2
1160
+ 0(r)
1161
+ � ∞
1162
+ 0
1163
+ dr r ψ2
1164
+ 0(r)
1165
+ .
1166
+ (A2)
1167
+ The denominator turns out to be
1168
+ � ∞
1169
+ 0
1170
+ dr r ψ2
1171
+ 0(r) = π
1172
+
1173
+ 1 −
1174
+ Γ
1175
+
1176
+ N + 1
1177
+ 2
1178
+
1179
+ √π Γ(N + 1)
1180
+
1181
+  ,
1182
+ (A3)
1183
+ while the numerator is
1184
+ � ∞
1185
+ 0
1186
+ dr r3 ψ2
1187
+ 0(r)π
1188
+ 2
1189
+
1190
+ 1 + 1
1191
+ N − (2N + 1)
1192
+ N
1193
+ Γ
1194
+
1195
+ N + 1
1196
+ 2
1197
+
1198
+ √π Γ(N + 1)
1199
+
1200
+  .
1201
+ (A4)
1202
+ The spectral width σp can also be expressed in terms
1203
+ of quantities defined across the plane z = 0, being (in
1204
+ units of 1/a)
1205
+ σ2
1206
+ p =
1207
+ 1
1208
+
1209
+ � ∞
1210
+ 0
1211
+ dr r
1212
+ �∂ψ0
1213
+ ∂r
1214
+ �2
1215
+ � ∞
1216
+ 0
1217
+ dr r ψ2
1218
+ 0(r)
1219
+ ,
1220
+ (A5)
1221
+ where the numerator turns out to be
1222
+ � ∞
1223
+ 0
1224
+ dr r
1225
+ �∂ψ0
1226
+ ∂r
1227
+ �2
1228
+ = 21−2N Γ(2N) ,
1229
+ (A6)
1230
+ so that, on using again Eq. (5),
1231
+ σ2
1232
+ p =
1233
+ 1
1234
+ π2 22N Γ(N)2
1235
+ √π Γ(N + 2) Γ(2N)
1236
+ √π Γ(N + 1) − Γ
1237
+
1238
+ N + 1
1239
+ 2
1240
+ � .
1241
+ (A7)
1242
+ Finally, on substituting from Eqs. (A2) and (A7) into
1243
+ Eq. (A1), Eq. (8) follows.
1244
+ Appendix B: Proof of Eq. (23)
1245
+ Thanks to the 2011 paper by Choi and Hasanov [18],
1246
+ the following integral representation of Ψ1 can be estab-
1247
+ lished:
1248
+ Ψ1
1249
+
1250
+ N + 1, 1
1251
+ N + 1, 1
1252
+ ���� x, y
1253
+
1254
+ =
1255
+ Γ(ǫ)
1256
+ Γ(N)Γ(ǫ − N − 1) ×
1257
+ � 1
1258
+ 0
1259
+ � 1
1260
+ 0
1261
+ dξ dη ηN(1 − ξ)N−1(1 − η)ǫ−N−2
1262
+ (1 − xξ)N+1
1263
+ × exp
1264
+
1265
+
1266
+
1267
+ xξ − 1
1268
+
1269
+ 1F1
1270
+
1271
+ 1 − ǫ; 1;
1272
+
1273
+ xξ − 1
1274
+
1275
+ (B1)
1276
+ where ǫ denotes an arbitrary complex parameters which
1277
+ must only satisfy the condition Re{ǫ} > Re{N} + 1. In
1278
+ particular, on letting ǫ = N + 2, Eq. (B1) yields
1279
+ Ψ1
1280
+
1281
+ N + 1, 1
1282
+ N + 1, 1
1283
+ ���� x, y
1284
+
1285
+ = Γ(N + 2)
1286
+ Γ(N)Γ(1) ×
1287
+ � 1
1288
+ 0
1289
+ dξ (1 − ξ)N−1
1290
+ (1 − xξ)N+1
1291
+ ×
1292
+ � 1
1293
+ 0
1294
+ dη ηN exp
1295
+
1296
+
1297
+
1298
+ xξ − 1
1299
+
1300
+ 1F1
1301
+
1302
+ −N − 1; 1;
1303
+
1304
+ xξ − 1
1305
+
1306
+ =
1307
+ = Γ(N + 2)
1308
+ Γ(N)
1309
+ ×
1310
+ � 1
1311
+ 0
1312
+ dξ (1 − ξ)N−1
1313
+ (1 − xξ)N+1
1314
+ � 1
1315
+ 0
1316
+ dη ηN 1F1
1317
+
1318
+ N + 2; 1;
1319
+
1320
+ 1 − xξ
1321
+
1322
+ ,
1323
+ (B2)
1324
+ where, in the last step, Kummer’s transformation has
1325
+ been employed. The inner η integral can be evaluated by
1326
+ using [21, formula 2.21.1.4], which yields
1327
+ � 1
1328
+ 0
1329
+ dη ηN
1330
+ 1F1
1331
+
1332
+ N + 2; 1;
1333
+
1334
+ 1 − xξ
1335
+
1336
+ =
1337
+ =
1338
+ 1
1339
+ N + 1 1F1
1340
+
1341
+ N + 1; 1;
1342
+ y
1343
+ 1 − xξ
1344
+
1345
+ .
1346
+ (B3)
1347
+ Finally, on substituting from Eq. (B3) into Eq. (B2), after
1348
+ simple algebra Eq. (23) follows.
1349
+
1350
+ 8
1351
+ Appendix C: Proof of Eq. (36)
1352
+ From the very definition into Eq. (16) we have
1353
+ Ψ1
1354
+
1355
+ 1, β
1356
+ 2, 1
1357
+ ���� t, −s
1358
+
1359
+ =
1360
+
1361
+
1362
+ k=0
1363
+
1364
+
1365
+ ℓ=0
1366
+ (1)k+ℓ (β)k
1367
+ (2)k(1)ℓ
1368
+ tk
1369
+ k!
1370
+ (−s)l
1371
+ ℓ!
1372
+ =
1373
+ =
1374
+
1375
+
1376
+ k=0
1377
+ (1)k (β)k
1378
+ (2)k
1379
+ tk
1380
+ k!
1381
+
1382
+
1383
+ ℓ=0
1384
+ (1 + k)ℓ
1385
+ (1)ℓ
1386
+ (−s)l
1387
+ ℓ!
1388
+ =
1389
+ =
1390
+
1391
+
1392
+ k=0
1393
+ (1)k (β)k
1394
+ (2)k
1395
+ tk
1396
+ k! 1F1(1 + k; 1; −s) =
1397
+ = exp(−s)
1398
+
1399
+
1400
+ k=0
1401
+ (β)k
1402
+ (2)k
1403
+ tkLk(s) .
1404
+ (C1)
1405
+ Last series can be expressed in closed form via [20,
1406
+ 5.11.2.7], i.e.,
1407
+
1408
+
1409
+ k=0
1410
+ (a)k tk
1411
+ (α + β)k
1412
+
1413
+ k(x) = (1 − t)−aΦ1
1414
+
1415
+ a, β − 1
1416
+ α + β
1417
+ ����
1418
+ t
1419
+ t − 1,
1420
+ tx
1421
+ t − 1
1422
+
1423
+ ,
1424
+ (C2)
1425
+ from which, on letting a = β, α = 0, β = 2, and x = s,
1426
+ after straightforward algebra Eq. (36) follows.
1427
+ [1] S. De Silvestri, P. Laporta, V. Magni, and 0. Svelto,
1428
+ “Solid-state laser unstable resonators with tapered reflec-
1429
+ tivity mirrors: the super-Gaussian approach,” IEEE J.
1430
+ Quant. El. 24, 1172 - 1177 (1988).
1431
+ [2] A. Parent, M. Morin, and P. Lavigne, “Propagation of
1432
+ super-Gaussian field distributions,” Opt. Quant. El. 24,
1433
+ S1071 - S1079 (1992).
1434
+ [3] F. Gori, “Flattened Gaussian beams,” Opt. Commun.
1435
+ 107, 335-341 (1994).
1436
+ [4] Equation (2) was originally derived starting from the
1437
+ identity 1 = exp(−ξ2) exp(ξ2) and on truncating the
1438
+ Taylor expansion of the second exponential up to N. In
1439
+ the present paper, however, we restrict the expansion to
1440
+ the first N terms. With such choice the case N = 1 cor-
1441
+ respond to the Gaussian beam. But in this
1442
+ [5] V. Bagini, R. Borghi, F. Gori, A. M. Pacileo, M. Santar-
1443
+ siero, D. Ambrosini, and G. Schirripa Spagnolo, “Prop-
1444
+ agation of axially symmetric flattened Gaussian beams,”
1445
+ J. Opt. Soc. Am. A 13, 1385-1394 (1996).
1446
+ [6] R. Borghi, “Elegant Laguerre-Gauss beams as a new tool
1447
+ for describing axisymmetric flattened Gaussian beams,”
1448
+ J. Opt. Soc. Am. A 18, 1627-1633 (2001).
1449
+ [7] Y. Li, “Light beams with flat-topped profiles,” Opt. Lett.
1450
+ 27, 1007-1009 (2002).
1451
+ [8] A.G. Sedukhin, “Rectangular symmetrical mesa beams
1452
+ and their comparison with flattened Gaussian and multi-
1453
+ Gaussian beams,” Optics Communications, 335, 284 - 292
1454
+ (2015).
1455
+ [9] R.
1456
+ Borghi,
1457
+ “Uniform
1458
+ approximation
1459
+ of
1460
+ flat-topped
1461
+ beams,” J. Opt. Soc. Am. A (2013)
1462
+ [10] P. Appell,“Sur les s´eries hyperg´eom´etriques de deux vari-
1463
+ ables et sur des ´equations diff´erentielles lin´eaires aux
1464
+ d´eriv´ees partielles,” Comptes rendus hebdomadaires des
1465
+ s´ances de l’Acad´emie des sciences 90, 296 - 298 (1880).
1466
+ [11] P. Humbert, “The Confluent Hypergeometric Functions
1467
+ of Two Variables,” Proceedings of the Royal Society of
1468
+ Edinburgh, IX, 73 - 96 1922.
1469
+ [12] C. J. R. Sheppard and S. Saghafi, “Flattened light
1470
+ beams,” Opt. Commun. 132, 144 -152 (1996).
1471
+ [13] Digital
1472
+ Library
1473
+ of
1474
+ Mathematical
1475
+ Functions,
1476
+ Na-
1477
+ tional
1478
+ Institute
1479
+ of
1480
+ Standards
1481
+ and
1482
+ Technology
1483
+ http://dlmf.nist.gov/.
1484
+ [14] Y. A. Brychkov, Handbook of Special Functions (CRC
1485
+ Press, London, 2008).
1486
+ [15] Y. A. Brychkov, New Indefinite and Definite Integrals
1487
+ of Elementary and Special Functions (A. A. Dorodnicyn
1488
+ Computing Center of the Russian Academy of Sciences,
1489
+ Moscow, 2014).
1490
+ [16] E.M. El Halba, H. Nebdi, M. Boustimi, and A. Belafhal,
1491
+ “On the Humbert confluent hypergeometric function used
1492
+ in laser field,” Phys. Chem. News 73, 90 - 93 (2014).
1493
+ [17] A. Belafhal and F. Saad,“Conversion of circular beams by
1494
+ a spiral phase plate: Generation of Generalized Humbert
1495
+ beams,” Optik 138, 516 - 528 (2017).
1496
+ [18] J. Choi and A. Hasanov, “Applications of the operator
1497
+ H(α, β) to the Humbert double hypergeometric func-
1498
+ tions,” Computers and Mathematics with Applications
1499
+ 61, 663 - 671 (2011).
1500
+ [19] Y. A. Brychkov and N. Saad, “Some formulas for the
1501
+ Appell function F1(a, b, b′; c; w, z).” Integral Transforms
1502
+ and Special Functions 23, 793 - 802 (2012).
1503
+ [20] A. P. Prudnikov, Y. A. Brychkov, and O. I. Marichev,
1504
+ Integrals and Series (Gordon Breach, 1986), Vol. II.
1505
+ [21] A. P. Prudnikov, Y. A. Brychkov, and O. I. Marichev,
1506
+ Integrals and Series (Gordon Breach, 1986), Vol. III.
1507
+ [22] A. P. Prudnikov, Y. A. Brychkov, and O. I. Marichev,
1508
+ Integrals and Series (Gordon Breach, 1986), Vol. IV.
1509
+ [23] M. Born and E. Wolf, Principles of Optics (Cambridge
1510
+ University Press, Cambridge, 1999).
1511
+ [24] M. V. Berry, “Why are special functions special?,” Phys.
1512
+ Today, 11-12 (2001)
1513
+
KNE3T4oBgHgl3EQfXwo4/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,398 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf,len=397
2
+ page_content='arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
3
+ page_content='04481v1 [physics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
4
+ page_content='optics] 11 Jan 2023 “Analytical Continuation” of Flattened Gaussian Beams Riccardo Borghi Dipartimento di Ingegneria Civile, Informatica e delle Tecnologie Aeronautiche, Universit`a “Roma Tre”, Via Vito Volterra 62, I-00146 Rome, Italy A purely analytical extension of the flattened Gaussian beams [Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
5
+ page_content=' Commun.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
6
+ page_content=' 107, 335 (1994)] to any values of the beam order, is here proposed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
7
+ page_content=' Thanks to it, the paraxial propagation problem of axially symmetric, coherent flat-top beams through arbitrary ABCD optical systems can definitely be closed in terms of a particular bivariate confluent hypergeometric function.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
8
+ page_content=' I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
9
+ page_content=' INTRODUCTION Flat-top beams continue to attract a considerable at- tention in optics: during the last five years more than sixty papers have been published on the subject.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
10
+ page_content=' In or- der to model flat-top axially symmetric distributions, two classes of different scenarios appeared: in the first one, simple analytical profiles were employed, the most known of them being the superGaussian (SG) [1, 2], which is for- mally defined by SGν(ξ) = exp(−ξ2ν) , (1) where ν denotes a real parameter which controls the“flat- ness” of the profile, with the particular case ν = 1 giving the Gaussian profile.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
11
+ page_content=' The symbol ξ denotes a normal- ized radial transverse position.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
12
+ page_content=' Despite its mathemat- ical simplicity, it is well known that Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
13
+ page_content=' (1) does not allow the wavefield of paraxially propagated superGaus- sian (i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
14
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
15
+ page_content=', for ν ̸= 1) beams to be analytically evaluated, even within the simplest scenario, namely free space.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
16
+ page_content=' To overcome such a difficulty, which two or three decades ago could represent a considerable computational bottleneck in several practical situations, alternative ap- proaches were proposed in 1994 and in 2002 by Gori and Li, respectively, to conceive analytical models able to solve the free space propagation problem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
17
+ page_content=' The former was called flattened Gaussian (FG henceforth) [3], and, differently from SG, is expressed through an explicit fi- nite sum of terms, namely FGN(ξ) = exp(−Nξ2) N−1 � m=0 (Nξ2)m m!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
18
+ page_content=' , (2) where the integer parameter N will be referred to as the FG order.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
19
+ page_content=' Scaling the ξ variable by the factor √ N gives the FG transverse profile a flat-topped shape which, for N = 1, reduces to a Gaussian distribution, whereas for N → ∞ tends to the characteristic function of the uni- tary disk [4].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
20
+ page_content=' The model is computationally exact, since the initial distribution (2) can be recast in terms of a su- perposition of N standard Laguerre-Gauss (sLG hence- forth) beams.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
21
+ page_content=' Accordingly, in order to evaluate the field propagated in free space, it was enough to sum up the N propagated sLG, a job which can exactly be done, al- ways [5].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
22
+ page_content=' In [6], a different superposition scheme of the profile (2) was proposed, in which the sLG family was replaced by the so- called elegant Laguerre-Gauss (eLG henceforth) set.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
23
+ page_content=' In this way, not only free-space propa- gation, but also the interaction of FG beams with any axially symmetric paraxial optical system can be dealt with in exact terms, always through finite sums.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
24
+ page_content=' In 2002, Yaijun Li proposed an analytical model al- ternative to the FG one.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
25
+ page_content=' The idea was to impose a lo- cal “flatness” condition, which required the first 2N ξ- derivatives of the profile to be null at the origin ξ = 0 [7].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
26
+ page_content=' On using such condition, Li conceived the following ana- lytical model: LiGN(ξ) = N � m=1 (−1)m−1 �N m � exp(−mξ2) = = � 1 − � 1 − exp � −ξ2���N N , (3) which, differently from FG, is based on the superposi- tion of N fundamental Gaussian beams having variable widths.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
27
+ page_content=' Both Gori’s and Li’s models provide exact solutions to the paraxial propagation problem of coherent, axially symmetric flat-topped beams.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
28
+ page_content=' From a merely mathemat- ical perspective, their only own limit is represented by the fact that, differently from SG, only positive integer orders N can be dealt with to describe the initial flat-top distribution.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
29
+ page_content=' It is important to mention that, for 1D ge- ometry (or rectangular 2D geometries), general analyt- ical solutions were already provided, at least upon free propagation, by modeling the flat-top profile via an error function [8].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
30
+ page_content=' An attempt to extend the 2D circular FG model to noninteger orders was also proposed in [9], but only approximate estimates of the free space propagated field were found within the asymptotic limit N ≫ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
31
+ page_content=' The aim of the present paper is to solve exactly the propagation problem of FG beams of any order (real or even complex) through typical axially symmetric parax- ial optical systems.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
32
+ page_content=' To this end, the right side of Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
33
+ page_content=' (2) will first be identified as an incomplete Gamma func- tions, which is known to be defined onto the whole com- plex plane, as far as both arguments are concerned.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
34
+ page_content=' An immediate byproduct of such identification will be the closed form expression of the M 2 factor of FG beams of any order, an interesting generalization of the result found in [5].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
35
+ page_content=' This is shown in Sec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
36
+ page_content=' II of the present pa- per.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
37
+ page_content=' The most important results are indeed presented 2 in Secs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
38
+ page_content=' III and IV.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
39
+ page_content=' In the former, the free- space prop- agation problem will be solved thanks to an important class of integrals recently closed by Yuri Brychkov.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
40
+ page_content=' Al- though the more general propagation problem will be solved in Sec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
41
+ page_content=' IV, the analysis presented in Sec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
42
+ page_content=' III should be viewed as an important propaedeutical step.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
43
+ page_content=' There, it will be shown that a very important, but nevertheless not so much known, class of special functions, called bivariate hypergeometric functions, together with the correspond- ing confluent versions, form the mathematical skeleton of the paraxially diffracted wavefield.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
44
+ page_content=' Bivariate hypergeo- metric were first introduced in 1880 by Paul Appell [10], their confluent version forty years later by Paul Hum- bert [11].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
45
+ page_content=' The results we are going to present would also give readers a partial answer about the lack, for more than thirty years, of purely analytical solutions to the problem of the paraxial propagation of coherent 2D flat- topped beams.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
46
+ page_content=' The present work has a clear mathematical character: for instance, dimensionless quantities will be used wher- ever possible.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
47
+ page_content=' Moreover, the number of mathematical appendices have been considerably limited, because we strongly believe that following all most important math- ematical steps could greatly help readers to fully grasp the essence of our analysis, as well as the importance of such still mysterious special functions, which will lead to analytical, elegant, and exact solutions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
48
+ page_content=' II.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
49
+ page_content=' PRELIMINARIES A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
50
+ page_content=' “Analytical continuation” of the FG model Already in 1996, Sheppard & Saghafi [12] pointed out that Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
51
+ page_content=' (2) can be given the closed form FGN(ξ) = Γ(N, Nξ2) Γ(N) , (4) where Γ(·) and Γ(·, ·) denote Gamma and incomplete Gamma functions, respectively [13].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
52
+ page_content=' Differently from Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
53
+ page_content=' (2), Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
54
+ page_content=' (4) is not limited to integer FG orders, but rather it can be analytically continued to real and also complex values of N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
55
+ page_content=' As a preliminary result of the extended definition into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
56
+ page_content=' (4), an analytical check of Li’s“flatness condition”[7] will now be carried out.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
57
+ page_content=' To this end, it is sufficient to use formulas 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
58
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
59
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
60
+ page_content='1 and 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
61
+ page_content='8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
62
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
63
+ page_content='17 of [14] to prove, with long but simple algebra, that dn dξn Γ(N, Nξ2) = −2nn!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
64
+ page_content='N N exp(−Nξ2) ξ2N−n × [n/2] � k=0 (n − k − 1)!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
65
+ page_content=' 4kk!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
66
+ page_content=' (n − 2k)!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
67
+ page_content=' L(N−n+k) n−k−1 (Nξ2) , (5) which gives at once � dn dξn Γ(N, Nξ2) � ξ=0 = 0 , 0 ≤ n < 2 Re{N} , (6) thus implying the real part of N to be chosen greater than one.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
68
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
69
+ page_content=' Spreading properties: closed form expression of the M 2 factor An interesting byproduct of the extended Γ-based def- inition into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
70
+ page_content=' (4) is the evaluation of the M 2 factor of FG beams, first established in [5] for N ∈ N, for nonin- teger orders.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
71
+ page_content=' To this end, consider an initial field distri- bution across the plane z = 0 of a cylindrical reference frame (r, z), say ψ0(r), given by ψ0(r) = FGN �r a � = Γ � N, N r2 a2 � Γ(N) , (7) where an overall amplitude constant has been set to one and the symbol a denotes the “width” of lat-top distri- bution field distribution.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
72
+ page_content=' For simplicity, it will be set a = 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
73
+ page_content=' The evaluation of the M 2 factor, which is defined as the product of the normalized second order moments across the z = 0 and the spatial frequency planes is detailed in Appendix A, where it is proved the following closed- form expression: M 2 = � (N + 1) Γ(N + 1/2) √π Γ(N + 1) � 1 − Γ(N + 3/2) √π Γ(N + 2) � 1 − Γ(N + 1/2) √π Γ(N + 1) , (8) which extends the 1996 analysis of [5] to N /∈ N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
74
+ page_content=' It is worth comparing Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
75
+ page_content=' (8) with the corresponding expres- sion of SG beam M 2 factor, namely [2] M 2 = � Γ(2/ν) Γ(1/ν)/ν , (9) deceptively simpler.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
76
+ page_content=' In the next two sections, our exten- sion of the FG model will further reveal its powerfulness and mathematical elegance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
77
+ page_content=' III.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
78
+ page_content=' FREE-SPACE PARAXIAL PROPAGATION OF FG BEAMS A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
79
+ page_content=' Preliminaries Suppose the initial field distribution given by Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
80
+ page_content=' (7) is allowed to propagate in free space.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
81
+ page_content=' The corresponding 3 field, say ψ(r;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
82
+ page_content=' z), can be expressed, apart from an overall phase factor exp(ikz), as follows: ψ(r;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
83
+ page_content=' z) = −i U 2π � R2 d2ρ ψ0(ρ) exp �iU 2 |r − ρ|2 � , (10) where the Fresnel number U = ka2/z has been intro- duced and the beam width a has been used as unit for measuring all transverse sizes.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
84
+ page_content=' This means that the quan- tity r should be meant as the ratio between the trans- verse position vector of the observation point and a.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
85
+ page_content=' For integer FG orders, the free space propagation problem has already been solved in [3] by expanding the initial field distribution ψ0 as the linear combination of a finite number of sLG beams.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
86
+ page_content=' It is then sufficient to propa- gate each sLG beam up to the observation plane and to recombine all of them with the initial expanding coeffi- cients for the correct value of ψ(r;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
87
+ page_content=' z) to be retrieved.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
88
+ page_content=' As we are going to show in a moment, the Γ-based model into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
89
+ page_content=' (4) allows an exact evaluation of the propagated wavefield (10) also for N /∈ N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
90
+ page_content=' It is worth recalling that, from a mere practical perspective, the present section could seem somewhat redundant, as in Sec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
91
+ page_content=' IV the more general propagation problem within ABCD systems will be solved.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
92
+ page_content=' Nevertheless, we believe what is contained in the present section could help nonspecialist readers to familiarize with the main notations and mathemati- cal tools which will constitute the basis of the general results presented into Sec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
93
+ page_content=' IV.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
94
+ page_content=' In other words, it should be considered as a useful, propaedeutical material.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
95
+ page_content=' We start on substituting from Eqs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
96
+ page_content=' (7) into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
97
+ page_content=' (10), which after simple algebra gives ψ(r;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
98
+ page_content=' z) = − i U Γ(N) exp �iU r2 2 � × � ∞ 0 dρ ρ exp �iU 2 ρ2 � Γ � N, N ρ2� J0(Ur ρ) , (11) where J0 denotes the 0th-order Bessel function of the first kind.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
99
+ page_content=' It is worth recasting the incomplete Γ function as Γ(N, Nξ) Γ(N) = 1 − γ(N, Nξ) Γ(N) , (12) where γ(·, ·) denotes the “lower”incomplete gamma func- tion.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
100
+ page_content=' Then Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
101
+ page_content=' (11) takes on the form ψ(r;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
102
+ page_content=' z) = = −i U exp �iU r2 2 � � ∞ 0 dρ ρ exp � −U 2i ρ2 � J0(Ur ρ) + i U exp �iU r2 2 � Γ(N) × � ∞ 0 dρ ρ exp � −U 2i ρ2 � γ � N, Nρ2� J0(Ur ρ) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
103
+ page_content=' (13) The first term is identically equal to one (it is nothing but a unitary plane wave propagating along the z-axis).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
104
+ page_content=' As far as the second is concerned, the following notable formula has recently been published by Brychkov [15, formula 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
105
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
106
+ page_content='20]: � ∞ 0 dx xα−1 exp(−a x2) γ(µ, bx2) Jν(c x) = = 2−ν−1bµcνΓ � µ + α + ν 2 � µaµ+(α+ν)/2Γ(ν + 1) Ψ1 � µ + α + ν 2 , µ µ + 1, ν + 1 ����� − b a, − c2 4a � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
107
+ page_content=' (14) Then, on using Eqs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
108
+ page_content=' (13) and (14), long but straightfor- ward algebra gives ψ(r;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
109
+ page_content=' z) = 1 − exp �iU r2 2 � �2iN U �N × Ψ1 � N + 1, N N + 1, 1 ���� − 2iN U , −iU r2 2 � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
110
+ page_content=' (15) B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
111
+ page_content=' A short Tour on Bivariate Hypergeometric Functions The symbol Ψ1 into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
112
+ page_content=' (15) denotes a special func- tion called bivariate confluent hypergeometric.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
113
+ page_content=' It is worth briefly describing the principal definitions and properties which are important for our scopes.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
114
+ page_content=' Function Ψ1 is for- mally defined through the following double series power expansion: Ψ1 � a, b c, c′ ���� z, w � = ∞ � k=0 ∞ � ℓ=0 (a)k+ℓ (b)k (c)k(c′)ℓ zk k!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
115
+ page_content=' wℓ ℓ!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
116
+ page_content=' , (16) valid for |z| ≤ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
117
+ page_content=' The symbol (·)n denotes Pochham- mer’s symbol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
118
+ page_content=' Another bivariate confluent hypergeomet- ric function which will be meet in the present paper is the function Φ1, defined by Φ1 � a, b c ���� z, w � = ∞ � k=0 ∞ � ℓ=0 (a)k+ℓ (b)k (c)k+ℓ zk k!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
119
+ page_content=' wℓ ℓ!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
120
+ page_content=' , (17) valid for |z| ≤ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
121
+ page_content=' Functions Ψ1 and Φ1 are members of a family of functions that generalize Kummer’s con- fluent hypergeometric function 1F1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
122
+ page_content=' In particular, Φ1 is obtained from the so-called Appell function F1, defined by F1 � a, b1, b2 c ���� z, w � = ∞ � k=0 ∞ � ℓ=0 (a)k+ℓ (b1)k (b2)ℓ (c)k+ℓ zk k!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
123
+ page_content=' wℓ ℓ!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
124
+ page_content=' , (18) (again valid for |z| ≤ 1), through the following limiting definition: Φ1 � a, b c ���� z, w � = lim ǫ→0 F1 � a, b, 1 ǫ c ����� z, ǫw � , (19) 4 which can be proved on first substituting the identity lim ǫ→0 �1 ǫ � ℓ ǫℓ = 1 , (20) directly into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
125
+ page_content=' (19), then on interchanging the limit with the double series.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
126
+ page_content=' Multivariate hypergeometric and confluent hypergeo- metric functions play a role of considerable importance in theoretical physics and applied math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
127
+ page_content=' In optics, the role of bivariate confluent hypergeometric functions in de- scribing a large class of paraxial optical disturbances has recently been pointed out [16, 17].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
128
+ page_content=' Moreover, it is worth stressing that, from a practical viewpoint, Appell’s func- tion F1 is nowadays part of the symbolic platform Math- ematica, where it is computable with arbitrarily high ac- curacies.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
129
+ page_content=' Also the whole family of Appell functions, in- cluding F1 as well as its three sisters F2, F3, and F4, are currently implemented in the latest release of Maple.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
130
+ page_content=' It is then highly desirable that in a near future also the set of bivariate confluent hypergeometric functions, in- cluding Ψ1 and Φ1, could become part of such family of “evaluable”special functions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
131
+ page_content=' In the meanwhile, someone might rightly object to the practical usefulness of func- tions that are defined through double infinite series like those into Eqs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
132
+ page_content=' (16) - (18).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
133
+ page_content=' To overcome such difficulties, some tricks will be implemented in the rest of the paper, tricks which are aimed at extending the validity domain of Ψ1 and Φ1 beyond the series definitions, and then to improve the practical usefulness of our analytical results.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
134
+ page_content=' C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
135
+ page_content=' Free-space propagation formula Function Ψ1 can be continued by using the following transformation [18, formula 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
136
+ page_content='54]: Ψ1 � α, β γ1, γ2 ���� z, w � = = 1 (1 − z)α Ψ1 � α, γ1 − β γ1, γ2 ���� z z − 1, w 1 − z � , (21) which, once substituted into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
137
+ page_content=' (15), gives a new, closed- form, expression of the paraxial propagated field ψ(r;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
138
+ page_content=' z) = 1 − exp �iU r2 2 � 1 + 2iN U \uf8eb \uf8ec \uf8ed 1 1 + U 2iN \uf8f6 \uf8f7 \uf8f8 N × Ψ1 \uf8eb \uf8ec \uf8ed N + 1, 1 N + 1, 1 ���� 1 1 + U 2iN , − iU r2 2 1 + 2iN U \uf8f6 \uf8f7 \uf8f8 , (22) indubitably one of the main results of the present paper.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
139
+ page_content=' Waiting for Mathematica or Maple to develop their own built-in version of Ψ1, it is worth working on the expres- sion into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
140
+ page_content=' (22) by using a notable integral represen- tation found again in [18].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
141
+ page_content=' For the sake of clarity, all mathematical steps are confined into Appendix B, where it is proved that Ψ1 � N + 1, 1 N + 1, 1 ���� x, y � = = N � 1 0 dξ (1 − ξ)N−1 (1 − xξ)N+1 1F1 � N + 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
142
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
143
+ page_content=' y 1 − xξ � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
144
+ page_content=' (23) Equation (23) appears to be somewhat intriguing: the wavefield of a free-space paraxially propagated FG beam of any order can be represented via a 1D integral defined over a finite integral.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
145
+ page_content=' This could seem a somewhat pe- culiar situation, due to the fact that the initial field dis- tribution (7) has an infinite support, namely the whole plane z = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
146
+ page_content=' But what is, in our opinion, even more important is that the integral representation (23) would hardly be reachable starting from Fresnel’s integral (10), without passing through the Ψ1 function and its trans- formation rules.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
147
+ page_content=' In the next section, a similar scenario will also be found as far as the more general problem is concerned.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
148
+ page_content=' IV.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
149
+ page_content=' PARAXIAL PROPAGATION THROUGH ABCD SYSTEMS A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
150
+ page_content=' Preliminaries The free-space paraxial propagation formula derived in the previous section will now be extended to the general case of the paraxial propagation of FG beams of any or- der through typical paraxial optical systems with axial symmetry, characterized by the so-called ABCD optical matrices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
151
+ page_content=' For FG beams of integer order, it was found in [6] that the propagation problem can be dealt with in exact terms by expanding the initial field distribution given into Eqs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
152
+ page_content=' (7) and (2) as a finite superposition of so-called elegant Laguerre (eLG henceforth) beams as fol- lows: ψ0(r) = N−1 � n=0 (−)n � N n + 1 � eLGn �ikr2 2qN � , (24) where the symbol eLGn(x) = exp(x)Ln(−x) will be re- ferred to as the elegant Laguerre function of order n and the complex radius of curvature qN = ka2 2iN has also been introduced.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
153
+ page_content=' The initial distribution ψ0 is then recast as follows: ψ0(r) = exp �ikr2 2qN � GN � 1, −ikr2 2qN � , (25) where the function GN (·, ·) is defined, for integer N, as GN (t, s) = N−1 � n=0 (−t)n � N n + 1 � Ln(s) , (26) 5 In [6] it was proved that, if the initial field distribution given by Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
154
+ page_content=' (25) feeds an axially symmetric paraxial optical system described by the optical matrix M M = \uf8eb \uf8ed A B C D \uf8f6 \uf8f8 ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
155
+ page_content=' (27) then the wavefield at the output plane of the system,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
156
+ page_content=' say ψ1(r),' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
157
+ page_content=' takes on the following form: ψ1(r) = = exp � ikr2 2QN � A 1 1 + B A qN GN \uf8eb \uf8ec \uf8ec \uf8ed 1 1 + B A qN ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
158
+ page_content=' kr2 2iA2 qN 1 + B A qN \uf8f6 \uf8f7 \uf8f7 \uf8f8 ,' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
159
+ page_content=' (28) where an overall phase factor exp(ikℓ) (with ℓ being the optical lenght) will be tacitly assumed and QN denotes the complex quantity QN = A qN + B C qN + D .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
160
+ page_content=' (29) The problem of extending the function GN (t, s) to N /∈ N will now be addressed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
161
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
162
+ page_content=' Extension of the function GN (t, s) to N /∈ N The starting point is the following Laplace transform representation of GN(t, s) established in [9]: GN(t, s) = exp(s) � ∞ 0 dξ exp(−ξ) J0 � 2 � s ξ � L(1) N−1(ξ t) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
163
+ page_content=' (30) For N ∈ N, the Laguerre polynomials L(1) N−1 can be writ- ten as L(1) N−1(ξ t) = N−1 � n=0 Ln(ξ t) , (31) so that, on substituting from Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
164
+ page_content=' (31) into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
165
+ page_content=' (30), it is found GN(t, s) = = exp(s) N−1 � n=0 � ∞ 0 dξ exp(−ξ) J0 � 2 � s ξ � Ln(ξ t) = = N−1 � n=0 (1 − t)n Ln � st t − 1 � , (32) where in the last passage, [22, formula 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
166
+ page_content='24.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
167
+ page_content='6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
168
+ page_content='2] has been used.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
169
+ page_content=' Equation (32) is a valid alternative, for N ∈ N, to the definition given into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
170
+ page_content=' (26).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
171
+ page_content=' For the scopes of the present paper, its importance stems from the fact that the quantity GN can also be thought of as function of two new variables, namely \uf8f1 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f3 1 − t = 1 1 + AqN B , st t − 1 = ikr2 2AB 1 1 + B AqN , (33) and this will reveal of a certain importance in the rest of our analysis.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
172
+ page_content=' In order to extend the integral into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
173
+ page_content=' (30) to N /∈ N, the following notable formula, again established by Brychkov [15], will be employed: � ∞ 0 xα−1 exp(−ax) Jν(b√x) L(λ) n (cx) dx = = � b 2 �ν Γ � α + ν 2 � (λ + 1)n n!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
174
+ page_content=' aα+ν/2Γ(ν + 1) Ψ1 � α + ν 2 , −n λ + 1, ν + 1 ����� c a, − b2 4a � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
175
+ page_content=' (34) In particular, on letting α = 1, a = 1, ν = 0, b = 2√s, t = c, n = N − 1, and λ = 1, Laplace’s transform into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
176
+ page_content=' (30) takes on the form GN(t, s) = N exp(s) Ψ1 � 1, 1 − N 2, 1 ���� t, −s � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
177
+ page_content=' (35) Again, it can be appreciated how the confluent hyperge- ometric function Ψ1 constitutes the mathematical skele- ton of the propagated field.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
178
+ page_content=' But there is more.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
179
+ page_content=' In Ap- pendix C, the following relationship has been established: Ψ1 � 1, 1 − N 2, 1 ���� t, −s � = = exp(−s) (1 − t)1−N Φ1 � 1 − N, 1 2 ���� t t − 1, st t − 1 � , (36) where Φ1 is the confluent hypergeometric function de- fined by Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
180
+ page_content=' (17).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
181
+ page_content=' On substituting from Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
182
+ page_content=' (36) into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
183
+ page_content=' (35), we have GN(t, s) = N (1 − t)N−1 Φ1 � 1 − N, 1 2 ���� t t − 1, st t − 1 � (37) so that Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
184
+ page_content=' (28) eventually becomes ψ1(r) = exp � ikr2 2QN � qNN B \uf8eb \uf8ec \uf8ed 1 1 + A qN B \uf8f6 \uf8f7 \uf8f8 N × Φ1 \uf8eb \uf8ec \uf8ec \uf8ed 1 − N, 1 2 ���� − A qN B , ikr2 2AB 1 1 + B AqN \uf8f6 \uf8f7 \uf8f7 \uf8f8 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
185
+ page_content=' (38) 6 Equation (38) summarizes the main result of the present paper: the general FG beam paraxial propagation prob- lem is reduced to the evaluation of the bivariate confluent hypergeometric Φ1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
186
+ page_content=' Again, it is possible to give Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
187
+ page_content=' (38) a different dress on using the following integral representation of Φ1, estab- lished in 2012 by Brychkov and Saad [19, formula 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
188
+ page_content='4]: Φ1 � a, 1 2 ���� w, z � = = (1 − w)1−a � 1 0 dξ (1 − w ξ)a−2 1F1(a;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
189
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
190
+ page_content=' zξ) , (39) which eventually leads to ψ1(r) = exp � ikr2 2QN � qNN B \uf8eb \uf8ec \uf8ed 1 1 + A qN B \uf8f6 \uf8f7 \uf8f8 N × � 1 0 dξ � 1 + A qN B ξ �N+1 1F1 \uf8eb \uf8ec \uf8ec \uf8ed1 − N;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
191
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
192
+ page_content=' ikr2 2AB ξ 1 + B AqN \uf8f6 \uf8f7 \uf8f7 \uf8f8 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
193
+ page_content=' (40) Similarly as it was found for the free-space propagation into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
194
+ page_content=' (23), also the integral representation of ψ1 given by Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
195
+ page_content=' (40) turns out to be defined onto a finite interval [0, 1], despite the infinite support of both the initial field distribution ψ0, as well as its Fourier transform.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
196
+ page_content=' In the present case, however, at least a qualitative explanation of such a mathematical counterintuitive behavior can be grasped by estimating the right side of Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
197
+ page_content=' (40) within the asymptotic limit N → ∞, which corresponds to replace the initial FG beam distribution ψ0 by that emerging from a circular hole of radius a.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
198
+ page_content=' In particular, the asymptotics can be carried out in an elementary way, by first noting that QN → B/D and that lim N→∞ 1 � 1 + A qN B ξ �N+1 = exp � iA ka2 2B ξ � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
199
+ page_content=' (41) As far as Kummer’s function inside the integral is concerned, the following asymptotics holds [13, for- mula 13.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
200
+ page_content='8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
201
+ page_content='13]: 1F1(1 − N;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
202
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
203
+ page_content=' z) ∼ exp(z/2) J0 � 2 √ N z � , N ≫ 1 , (42) which, once substituted into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
204
+ page_content=' (40) together with Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
205
+ page_content=' (41), leads to ψ1(r) ∼ U 2i exp � iUD 2 �r a �2� × � 1 0 dξ exp � iA U 2 ξ � J0 � U r a � ξ � , N ≫ 1 , (43) where now U = ka2/B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
206
+ page_content=' Finally, it is not difficult to convince that Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
207
+ page_content=' (43) is nothing but von Lommel’s integral [23], namely, the result of Collins’ integral for an incident wavefield ψ0 = circ(r/a), as it should be expected.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
208
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
209
+ page_content=' CONCLUSIONS Even today, the term“superGaussian beam”is synony- mous of flat-topped beam, despite the indisputable lim- its, both practical and theoretical, of the SG model and the availability of more efficient analytical approaches.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
210
+ page_content=' For rectangular geometries, Sedukhin’s work should have contributed to identify flat-topped profiles with an error function.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
211
+ page_content=' For two-dimensional, axially symmetric geome- tries, Gori’s and Li’s models, despite allowing to solve ex- actly the paraxial propagation problem, to date continue struggling to supplant the obsolete SG model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
212
+ page_content=' In the present paper, the FG model has been general- ized to any values, no longer necessarily integer, of the or- der N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
213
+ page_content=' In doing this, use has been made of the suggestion, dating back more than twenty-five years ago, by Shep- pard & Saghafi to mathematically identify the model FG through an incomplete Gamma function.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
214
+ page_content=' From a merely technical viewpoint, our work rests on some beautiful re- sults recently established by Brychkov and co-workers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
215
+ page_content=' In this way, it has been possibile to analytically express the optical wavefield generated by the propagation of such flat-topped “Γ-beams”of any order through arbitrary ax- ially symmetric paraxial optical system (free space in- cluded) in terms of a single bivariate confluent hyperge- ometric function.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
216
+ page_content=' Our model is purely analytical and provided purely an- alytical closed expressions of the paraxially propagated wavefield.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
217
+ page_content=' It is a rare situation in physics in general and in optics in particular.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
218
+ page_content=' The ubiquitous presence of less and less known special functions, such as bivari- ate hypergeometric ones certainly are, also constitutes in our opinion an added value of the present work.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
219
+ page_content=' We strongly encourage our readers to go through an interest- ing paper written more than twenty years ago by Michael Berry [24], whose content seems nowadays more than ever more relevant.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
220
+ page_content=' In particular, the current availability of powerful computational platforms, such as Mathematica and Maple, will allow in the future to increase the set of special functions whose evaluation could be implemented at arbitrarily high accuracies.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
221
+ page_content=' We hope bivariate con- fluent hypergeometric functions, including of course Ψ1 and Φ1, could soon become part of such a mathematical weaponry.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
222
+ page_content=' Acknowledgements I wish to thank Turi Maria Spinozzi for his help during the preparation of the manuscript.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
223
+ page_content=' 7 Appendix A: Proof of Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
224
+ page_content=' (8) The M 2 factor is defined by M 2 = 2π σr σp , (A1) where σr and σp denote the widths across the plane z = 0 and the plane of spatial frequencies, respectively, both of them normalized to the beam energy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
225
+ page_content=' Due to the axial symmetry, σr can then be expressed (in units of a) as follows: σ2 r = � ∞ 0 dr r3 ψ2 0(r) � ∞ 0 dr r ψ2 0(r) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
226
+ page_content=' (A2) The denominator turns out to be � ∞ 0 dr r ψ2 0(r) = π \uf8ee \uf8ef\uf8ef\uf8f01 − Γ � N + 1 2 � √π Γ(N + 1) \uf8f9 \uf8fa\uf8fa\uf8fb , (A3) while the numerator is � ∞ 0 dr r3 ψ2 0(r)π 2 \uf8ee \uf8ef\uf8ef\uf8f01 + 1 N − (2N + 1) N Γ � N + 1 2 � √π Γ(N + 1) \uf8f9 \uf8fa\uf8fa\uf8fb .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
227
+ page_content=' (A4) The spectral width σp can also be expressed in terms of quantities defined across the plane z = 0, being (in units of 1/a) σ2 p = 1 2π � ∞ 0 dr r �∂ψ0 ∂r �2 � ∞ 0 dr r ψ2 0(r) , (A5) where the numerator turns out to be � ∞ 0 dr r �∂ψ0 ∂r �2 = 21−2N Γ(2N) , (A6) so that, on using again Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
228
+ page_content=' (5), σ2 p = 1 π2 22N Γ(N)2 √π Γ(N + 2) Γ(2N) √π Γ(N + 1) − Γ � N + 1 2 � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
229
+ page_content=' (A7) Finally, on substituting from Eqs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
230
+ page_content=' (A2) and (A7) into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
231
+ page_content=' (A1), Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
232
+ page_content=' (8) follows.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
233
+ page_content=' Appendix B: Proof of Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
234
+ page_content=' (23) Thanks to the 2011 paper by Choi and Hasanov [18], the following integral representation of Ψ1 can be estab- lished: Ψ1 � N + 1, 1 N + 1, 1 ���� x, y � = Γ(ǫ) Γ(N)Γ(ǫ − N − 1) × � 1 0 � 1 0 dξ dη ηN(1 − ξ)N−1(1 − η)ǫ−N−2 (1 − xξ)N+1 × exp � − yη xξ − 1 � 1F1 � 1 − ǫ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
235
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
236
+ page_content=' yη xξ − 1 � (B1) where ǫ denotes an arbitrary complex parameters which must only satisfy the condition Re{ǫ} > Re{N} + 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
237
+ page_content=' In particular, on letting ǫ = N + 2, Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
238
+ page_content=' (B1) yields Ψ1 � N + 1, 1 N + 1, 1 ���� x, y � = Γ(N + 2) Γ(N)Γ(1) × � 1 0 dξ (1 − ξ)N−1 (1 − xξ)N+1 × � 1 0 dη ηN exp � − yη xξ − 1 � 1F1 � −N − 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
239
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
240
+ page_content=' yη xξ − 1 � = = Γ(N + 2) Γ(N) × � 1 0 dξ (1 − ξ)N−1 (1 − xξ)N+1 � 1 0 dη ηN 1F1 � N + 2;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
241
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
242
+ page_content=' yη 1 − xξ � , (B2) where, in the last step, Kummer’s transformation has been employed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
243
+ page_content=' The inner η integral can be evaluated by using [21, formula 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
244
+ page_content='21.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
245
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
246
+ page_content='4], which yields � 1 0 dη ηN 1F1 � N + 2;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
247
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
248
+ page_content=' yη 1 − xξ � = = 1 N + 1 1F1 � N + 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
249
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
250
+ page_content=' y 1 − xξ � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
251
+ page_content=' (B3) Finally, on substituting from Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
252
+ page_content=' (B3) into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
253
+ page_content=' (B2), after simple algebra Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
254
+ page_content=' (23) follows.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
255
+ page_content=' 8 Appendix C: Proof of Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
256
+ page_content=' (36) From the very definition into Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
257
+ page_content=' (16) we have Ψ1 � 1, β 2, 1 ���� t, −s � = ∞ � k=0 ∞ � ℓ=0 (1)k+ℓ (β)k (2)k(1)ℓ tk k!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
258
+ page_content=' (−s)l ℓ!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
259
+ page_content=' = = ∞ � k=0 (1)k (β)k (2)k tk k!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
260
+ page_content=' ∞ � ℓ=0 (1 + k)ℓ (1)ℓ (−s)l ℓ!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
261
+ page_content=' = = ∞ � k=0 (1)k (β)k (2)k tk k!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
262
+ page_content=' 1F1(1 + k;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
263
+ page_content=' 1;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
264
+ page_content=' −s) = = exp(−s) ∞ � k=0 (β)k (2)k tkLk(s) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
265
+ page_content=' (C1) Last series can be expressed in closed form via [20, 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
266
+ page_content='11.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
267
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
268
+ page_content='7], i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
269
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
270
+ page_content=', ∞ � k=0 (a)k tk (α + β)k Lα k(x) = (1 − t)−aΦ1 � a, β − 1 α + β ���� t t − 1, tx t − 1 � , (C2) from which, on letting a = β, α = 0, β = 2, and x = s, after straightforward algebra Eq.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
271
+ page_content=' (36) follows.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
272
+ page_content=' [1] S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
273
+ page_content=' De Silvestri, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
274
+ page_content=' Laporta, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
275
+ page_content=' Magni, and 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
276
+ page_content=' Svelto, “Solid-state laser unstable resonators with tapered reflec- tivity mirrors: the super-Gaussian approach,” IEEE J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
277
+ page_content=' Quant.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
278
+ page_content=' El.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
279
+ page_content=' 24, 1172 - 1177 (1988).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
280
+ page_content=' [2] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
281
+ page_content=' Parent, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
282
+ page_content=' Morin, and P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
283
+ page_content=' Lavigne, “Propagation of super-Gaussian field distributions,” Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
284
+ page_content=' Quant.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
285
+ page_content=' El.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
286
+ page_content=' 24, S1071 - S1079 (1992).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
287
+ page_content=' [3] F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
288
+ page_content=' Gori, “Flattened Gaussian beams,” Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
289
+ page_content=' Commun.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
290
+ page_content=' 107, 335-341 (1994).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
291
+ page_content=' [4] Equation (2) was originally derived starting from the identity 1 = exp(−ξ2) exp(ξ2) and on truncating the Taylor expansion of the second exponential up to N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
292
+ page_content=' In the present paper, however, we restrict the expansion to the first N terms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
293
+ page_content=' With such choice the case N = 1 cor- respond to the Gaussian beam.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
294
+ page_content=' But in this [5] V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
295
+ page_content=' Bagini, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
296
+ page_content=' Borghi, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
297
+ page_content=' Gori, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
298
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
299
+ page_content=' Pacileo, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
300
+ page_content=' Santar- siero, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
301
+ page_content=' Ambrosini, and G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
302
+ page_content=' Schirripa Spagnolo, “Prop- agation of axially symmetric flattened Gaussian beams,” J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
303
+ page_content=' Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
304
+ page_content=' Soc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
305
+ page_content=' Am.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
306
+ page_content=' A 13, 1385-1394 (1996).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
307
+ page_content=' [6] R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
308
+ page_content=' Borghi, “Elegant Laguerre-Gauss beams as a new tool for describing axisymmetric flattened Gaussian beams,” J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
309
+ page_content=' Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
310
+ page_content=' Soc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
311
+ page_content=' Am.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
312
+ page_content=' A 18, 1627-1633 (2001).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
313
+ page_content=' [7] Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
314
+ page_content=' Li, “Light beams with flat-topped profiles,” Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
315
+ page_content=' Lett.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
316
+ page_content=' 27, 1007-1009 (2002).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
317
+ page_content=' [8] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
318
+ page_content='G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
319
+ page_content=' Sedukhin, “Rectangular symmetrical mesa beams and their comparison with flattened Gaussian and multi- Gaussian beams,” Optics Communications, 335, 284 - 292 (2015).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
320
+ page_content=' [9] R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
321
+ page_content=' Borghi, “Uniform approximation of flat-topped beams,” J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
322
+ page_content=' Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
323
+ page_content=' Soc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
324
+ page_content=' Am.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
325
+ page_content=' A (2013) [10] P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
326
+ page_content=' Appell,“Sur les s´eries hyperg´eom´etriques de deux vari- ables et sur des ´equations diff´erentielles lin´eaires aux d´eriv´ees partielles,” Comptes rendus hebdomadaires des s´ances de l’Acad´emie des sciences 90, 296 - 298 (1880).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
327
+ page_content=' [11] P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
328
+ page_content=' Humbert, “The Confluent Hypergeometric Functions of Two Variables,” Proceedings of the Royal Society of Edinburgh, IX, 73 - 96 1922.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
329
+ page_content=' [12] C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
330
+ page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
331
+ page_content=' R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
332
+ page_content=' Sheppard and S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
333
+ page_content=' Saghafi, “Flattened light beams,” Opt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
334
+ page_content=' Commun.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
335
+ page_content=' 132, 144 -152 (1996).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
336
+ page_content=' [13] Digital Library of Mathematical Functions, Na- tional Institute of Standards and Technology http://dlmf.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
337
+ page_content='nist.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
338
+ page_content='gov/.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
339
+ page_content=' [14] Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
340
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
341
+ page_content=' Brychkov, Handbook of Special Functions (CRC Press, London, 2008).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
342
+ page_content=' [15] Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
343
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
344
+ page_content=' Brychkov, New Indefinite and Definite Integrals of Elementary and Special Functions (A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
345
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
346
+ page_content=' Dorodnicyn Computing Center of the Russian Academy of Sciences, Moscow, 2014).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
347
+ page_content=' [16] E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
348
+ page_content='M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
349
+ page_content=' El Halba, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
350
+ page_content=' Nebdi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
351
+ page_content=' Boustimi, and A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
352
+ page_content=' Belafhal, “On the Humbert confluent hypergeometric function used in laser field,” Phys.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
353
+ page_content=' Chem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
354
+ page_content=' News 73, 90 - 93 (2014).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
355
+ page_content=' [17] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
356
+ page_content=' Belafhal and F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
357
+ page_content=' Saad,“Conversion of circular beams by a spiral phase plate: Generation of Generalized Humbert beams,” Optik 138, 516 - 528 (2017).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
358
+ page_content=' [18] J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
359
+ page_content=' Choi and A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
360
+ page_content=' Hasanov, “Applications of the operator H(α, β) to the Humbert double hypergeometric func- tions,” Computers and Mathematics with Applications 61, 663 - 671 (2011).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
361
+ page_content=' [19] Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
362
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
363
+ page_content=' Brychkov and N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
364
+ page_content=' Saad, “Some formulas for the Appell function F1(a, b, b′;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
365
+ page_content=' c;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
366
+ page_content=' w, z).” Integral Transforms and Special Functions 23, 793 - 802 (2012).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
367
+ page_content=' [20] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
368
+ page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
369
+ page_content=' Prudnikov, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
370
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
371
+ page_content=' Brychkov, and O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
372
+ page_content=' I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
373
+ page_content=' Marichev, Integrals and Series (Gordon Breach, 1986), Vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
374
+ page_content=' II.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
375
+ page_content=' [21] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
376
+ page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
377
+ page_content=' Prudnikov, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
378
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
379
+ page_content=' Brychkov, and O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
380
+ page_content=' I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
381
+ page_content=' Marichev, Integrals and Series (Gordon Breach, 1986), Vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
382
+ page_content=' III.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
383
+ page_content=' [22] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
384
+ page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
385
+ page_content=' Prudnikov, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
386
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
387
+ page_content=' Brychkov, and O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
388
+ page_content=' I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
389
+ page_content=' Marichev, Integrals and Series (Gordon Breach, 1986), Vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
390
+ page_content=' IV.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
391
+ page_content=' [23] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
392
+ page_content=' Born and E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
393
+ page_content=' Wolf, Principles of Optics (Cambridge University Press, Cambridge, 1999).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
394
+ page_content=' [24] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
395
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
396
+ page_content=' Berry, “Why are special functions special?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
397
+ page_content=',” Phys.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
398
+ page_content=' Today, 11-12 (2001)' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/KNE3T4oBgHgl3EQfXwo4/content/2301.04481v1.pdf'}
KdE1T4oBgHgl3EQfGgND/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:afb83cd29d32f2f57b81c8d96f8b65ea4b63020794f9b04930d1ee101e2aa3eb
3
+ size 5111853
LNE4T4oBgHgl3EQfiQ3g/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1ebb303a2c30c8d3e8135ecea94b44855ea104785f2e7adbb02013d1422ca8f4
3
+ size 2162733
MNAyT4oBgHgl3EQf6vpX/content/2301.00826v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4c9a39f3fc7b2cf1e6eb9f3f417dfc20798e78cbb94c3cf93536ea99d9f9eec8
3
+ size 387345
MNAyT4oBgHgl3EQf6vpX/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0054d0a64ba256c847bcf957268476cf5d3d8ffea739001b1064ddc8981d8713
3
+ size 63433
MtAzT4oBgHgl3EQfkv1s/content/tmp_files/2301.01536v1.pdf.txt ADDED
@@ -0,0 +1,671 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Condensed Matter Physics, 2022, Vol. 25, No. 4, 43708: 1–12
2
+ DOI: 10.5488/CMP.25.43708
3
+ http://www.icmp.lviv.ua/journal
4
+ Path integral Monte Carlo simulations of the
5
+ geometrical effects in KDP crystals
6
+ F. Torresi
7
+ , J. Lasave
8
+ , S. Koval
9
+
10
+ Instituto de Física Rosario, Universidad Nacional de Rosario and CONICET, 27 de Febrero 210 Bis, 2000 Rosario,
11
+ Argentina
12
+ Received July 10, 2022
13
+ Path integral Monte Carlo (PIMC) simulations with very simple models were used in order to unveil the physics
14
+ behind the isotope effects in H-bonded ferroelectrics. First, we studied geometrical effects in the H-bonds caused
15
+ by deuteration with a general three-site model based on a back-to-back double Morse potential plus a Morse
16
+ potential between oxygens, fitted to explain different general features for a wide set of H-bonded compounds.
17
+ Our model results show the Ubbelohde or geometrical effect (GE), i.e., the expansion of the H-bond with deute-
18
+ ration, in agreement to what is observed in H-bonded ferroelectrics with short H-bonds. Moreover, adjusting the
19
+ potential parameters to ab initio results, we have developed a 1D model which considers the bilinear proton-
20
+ proton interaction in mean-field to study nuclear quantum effects that give rise to the GE in KDP crystals. PIMC
21
+ simulations reveal that protons tunnel more efficiently than deuterons along the 1D chain, giving rise to a strong
22
+ attraction center that pulls the oxygens together. This mechanism, which is based on the correlation between
23
+ tunneling and geometrial modifications of the H-bonds, leads to a strong GE in the ordered phase of the chain
24
+ at low temperature which is in good agreement with the experimental data.
25
+ Key words: ferroelectric phase transition, H-bonded ferroelectrics, path integral Monte Carlo
26
+ 1. Introduction
27
+ KH2PO4 or KDP is the prototype of a wide family of H-bonded ferroelectric compounds which has
28
+ extensive applications as a key component in optoelectronic devices [1]. Besides the technological interest,
29
+ KDP has also attracted much attention due to its rich, complex and intriguing phenomenology, e.g., the
30
+ huge isotope effect that displays associated to its ferroelectric-paraelectric (FE-PE) phase transition. With
31
+ deuteration, the critical temperature 𝑇𝑐 changes from ≈ 122 K to ≈ 210 K. The saturated polarization 𝑃𝑠
32
+ at low 𝑇 also shows a large isotope effect, increasing from ≈ 5.0 µC/cm2 for KDP to ≈ 6.2 µC/cm2 for a
33
+ sample with 98% of deuteration [2].
34
+ The origin of these strong isotope effects is still controversial. The first explanation of the large
35
+ increase of 𝑇𝑐 upon deuteration was given by the quantum tunneling model [3], which focuses purely
36
+ on mass-dependent effects. However, increasing experimental evidence since the late nineteen eighties
37
+ showed that the large isotope effect is mainly driven by geometrical modifications of the H bonds [4, 5]
38
+ (Ubbelohde effect [6]). The recent observation of tunneling in the PE phase of KDP by neutron Compton
39
+ scattering experiments added even more controversy to the problem [7], although in deuterated KDP
40
+ (DKDP), tunneling could not be detected [8].
41
+ Ab initio calculations have recently shown that tunneling and geometric effects are complementary
42
+ aspects of the same phenomenon[9, 10]. With a simple selfconsistent model based on ab initio results, it
43
+ is demonstrated that the wave function solution of the nonlinear Schrödinger equation for deuteron/proton
44
+ clusters evolves from a double peak to a broad single peak located at the center of the H-bonds as the
45
+ cluster mass diminishes. This is explained by a strong nonlinear feedback between proton delocalization
46
+ (tunneling) and the effective proton potential barrier in the H-bonds, which changes concomitantly with
47
+ ∗Corresponding author: koval@ifir-conicet.gov.ar.
48
+ This work is licensed under a Creative Commons Attribution 4.0 International License. Further distribution
49
+ of this work must maintain attribution to the author(s) and the published article’s title, journal citation, and DOI.
50
+ 43708-1
51
+ arXiv:2301.01536v1 [cond-mat.mtrl-sci] 4 Jan 2023
52
+
53
+ F. Torresi, J. Lasave, S. Koval
54
+ the H-bond geometry. It is concluded that such a large mass dependence can explain the large isotope effect
55
+ found in KDP, via an amplified and selfconsistent geometric modification of the H bond in agreement
56
+ with experiments. On the other hand, these results are in striking contrast with the very weak dependence
57
+ obtained at fixed potential and geometry. Thus, the proton tunneling subunit and the host lattice are
58
+ strongly coupled and the host-and-tunneling system is not separable.
59
+ Many models were successfully developed in the past to shed light into the general phenomenology of
60
+ H-bonded ferroelectric materials [11–18]. In this paper, we address with very simple models the problem
61
+ of geometrical effects in KDP crystals by performing path integral Monte Carlo (PIMC) simulations.
62
+ First, we develop a three-site model for the H-bond to study local quantum geometric effects. This simple
63
+ model already serves us to gain knowledge about the interplay between proton tunneling and H-bond
64
+ geometric modifications such as the O–O distance variation. After this first insight, we develop a 1D
65
+ chain model of concatenated H-bonds to study in the ordered phase the geometrical effects caused by
66
+ deuteration. The model parameters are fitted using recent ab initio results [19]. We demonstrate that this
67
+ simple linear model can account for the geometrical effects observed in real H-bonded ferroelectrics,
68
+ which are at the root of the giant isotope effect in the critical temperature observed in the FE phase
69
+ transitions of these materials. The paper is organized as follows: in the next section we explain the models
70
+ used and describe details of the PIMC calculations. Section 3 describes and discusses the results obtained
71
+ for the three-site model and for the linear chain. Finally, we elaborate a summary and our conclussions
72
+ in section 4.
73
+ 2. Models and calculation details
74
+ 2.1. Three-site model
75
+ ���
76
+ ���
77
+ ���
78
+ ���
79
+
80
+
81
+ Figure 1. (Colour online) H-bond parameters in the three-site model. 𝑅 ≡ 𝑅OO is the distance between
82
+ oxygen nuclei. 𝑟OH is the proton-oxygen distance. The variable 𝛿 = 𝑅OO −2𝑟OH is defined as the distance
83
+ between the two possible equilibrium positions of the proton. Then, 𝑥 = 𝑅OO/2 − 𝑟OH is the proton
84
+ coordinate relative to the H-bond center. This parameter definition is also used in the linear chain model.
85
+ We developed a three-site (3S) model which represents a single O–H–O cluster embedded in the H-
86
+ bonded ferroelectric as it is sketched in figure 1. With the aim to model linear H-bonds, a Double Morse
87
+ (or back-to-back) potential (see e.g., [20–24]) is usually used, which is essentially the superposition of
88
+ two Morse potentials representing what the proton feels while interacting with both oxygens:
89
+ 𝑉OH (𝑥, 𝑅) = 𝑉𝑀
90
+
91
+ 𝑥 + 𝑅
92
+ 2
93
+
94
+ + 𝑉𝑀
95
+ � 𝑅
96
+ 2 − 𝑥
97
+
98
+ = 𝐷
99
+
100
+ 1 − exp
101
+
102
+ −𝑎
103
+ � 𝑅
104
+ 2 + 𝑥 − 𝑟0
105
+ ���2
106
+ + 𝐷
107
+
108
+ 1 − exp
109
+
110
+ −𝑎
111
+ � 𝑅
112
+ 2 − 𝑥 − 𝑟0
113
+ ���2
114
+ − 2𝐷,
115
+ (2.1)
116
+ where 𝑅 is the O–O distance, and 𝑥 represents the H position relative to the H-bridge center (see figure 1).
117
+ If we assume that 𝑅 is fixed, there is a critical value 𝑅𝑐 = 2(𝑎−1 ln 2+𝑟0) such that for 𝑅 < 𝑅𝑐 the potential
118
+ profile is a single well with a minimum at 𝑥 = 0. On the contrary, for 𝑅 > 𝑅𝑐 we have a symmetric double-
119
+ well potential, with a local maximum at 𝑥 = 0 and minima at 𝑥 = ±𝑎−1 cosh−1{1/2 exp[𝑎(𝑅/2 − 𝑟0)]}.
120
+ Notice that the energy barrier for the proton jump from one side to the other of the H-bond diminishes
121
+ concomitantly with the O–O distance 𝑅, vanishing for 𝑅 < 𝑅𝑐. Actually, we are interested in the
122
+ 43708-2
123
+
124
+ Quantum geometrical effects in KDP crystals
125
+ proton/deuteron tunneling regime, thus we would need that the equilibrium distance 𝑅 remains in the
126
+ region where the proton barrier exists, that is 𝑅 > 𝑅𝑐. However, simulations at low temperature with the
127
+ potential described in equation 2.1, relaxing both variables 𝑥 and 𝑅, yield to a collapse of the potential
128
+ barrier and the equilibrium energy profile displays one minimum only. Therefore, it is mandatory to
129
+ introduce a new interaction which preserves the system from the O–O distance collapse. This O–O
130
+ potential will represent the interaction between both oxygens and the lattice. The following Morse
131
+ potential between oxygens is chosen [19]:
132
+ 𝑉OO (𝑅) = 𝐷OO
133
+
134
+ 1 − e−𝑎OO(𝑅−𝑅0)�2
135
+ − 𝐷OO.
136
+ (2.2)
137
+ We adopted a Morse potential to describe the O–O interaction with the lattice because this kind of
138
+ anharmonic potential enables the system to explore with sufficient probability O–O distances larger
139
+ than 𝑅0, in such a way that the collapse tendency to a single well is drastically diminished. This is in
140
+ contrast to the case of a harmonic potential for the O–O interaction, where in this case the O–O collapse
141
+ is inevitable. The complete potential for the 3S model is as follows:
142
+ 𝑉3𝑆 (𝑥, 𝑅) = 𝑉OH (𝑥, 𝑅) + 𝑉OO (𝑅) = 𝐷
143
+
144
+ 1 − e−𝑎[(𝑅/2)+𝑥−𝑟0]�2
145
+ + 𝐷
146
+
147
+ 1 − e−𝑎[(𝑅/2)−𝑥−𝑟0]�2
148
+ − 2𝐷 + 𝐷OO
149
+
150
+ 1 − e−𝑎OO(𝑅−𝑅0)�2
151
+ − 𝐷OO.
152
+ (2.3)
153
+ The correlation between the H displacement 𝑥 and the O–O distance 𝑅 observed in experiments and ab
154
+ initio calculations is reflected by the anharmonic potential of equation (2.3): when the H approaches one
155
+ of the O’s in the covalent bond O–H (increasing 𝑥), the hydrogen-bond with the other O weakens and
156
+ the O–O distance (𝑅) increases. Moreover, 𝑅 diminishes with decreasing 𝑥, which is the inverse situation.
157
+ This correlation is precisely the important ingredient necessary for the existence of the Ubbelohde or the
158
+ geometrical effect observed in compounds with strong H-bonds.
159
+ 2.2. 1D model of concatenated H-bonds
160
+ Going a step beyond the simple three-site model, we have developed a one dimensional chain model
161
+ of concatenated H-bonds to study the GE in a more realistic way in the ordered phase. This 1D linear
162
+ model consists of a chain ...O–H...O–H...O–H...O–H..., which is built as a supercell containing 𝑁 = 200
163
+ unit cells of linear dimension 𝑅, the O–O distance, as shown schematically in figure 2. There are two
164
+ atoms, one oxygen and one hydrogen in each unit cell (O–H...). The supercell of dimension 𝐿 = 200𝑅 is
165
+ subjected to periodic boundary conditions. In the simulation, 𝐿 is allowed to relax at zero stress, as well
166
+ as each coordinate 𝑥𝑖 and 𝑅𝑖 of each unit cell 𝑖. For instance, this chain represents a model approximation
167
+ to the 1D H-bonded ferroelectric CsH2PO4 (CDP) if the model chain oxygen is interpreted as a PO4 unit
168
+ plus an ordered hydrogen covalently bonded to the phosphate at any temperature, and the model hydrogen
169
+ is the one that is disordered at high temperature in CDP [25]. Then, the global motion of hydrogens in our
170
+ linear model in the ordered phase, from one minimum to the other along the H-bonds of the chain, could
171
+ be related to the FE mode that accounts for the spontaneous polarization arising along the 𝑏 direction at
172
+ low 𝑇 in CDP [25]. Alternatively, the chain model may also represent an approximation to the study of
173
+ the GE in KH2PO4 (KDP) if the model effective oxygen now represents a KDP cluster of two phosphate
174
+ units including seven protons moving coordinately as a local FE mode [9, 10]. In all these cases, we must
175
+ adopt a convenient effective mass for the effective model hydrogen/deuteron considering that the real
176
+ displacements of H(D) are accompanied with the heavier atom motions [9, 10, 19].
177
+ The total potential energy for the linear chain (1D) model is defined as:
178
+ 𝑉1𝐷 (𝑅) =
179
+ ∑︁
180
+ 𝑖
181
+ 𝑉3𝑠 (𝑥𝑖, 𝑅𝑖) − 1
182
+ 2
183
+ ∑︁
184
+ ⟨𝑖 𝑗⟩
185
+ 𝐽𝑥𝑖𝑥 𝑗,
186
+ (2.4)
187
+ where 𝑉3𝑠 is the unit cell local potential defined exactly in the same way for the 3S model, as is shown
188
+ in equation (2.3), and the last term is the short-range interaction energy between protons/deuterons
189
+ 43708-3
190
+
191
+ F. Torresi, J. Lasave, S. Koval
192
+ Figure 2. (Colour online) Schematic representation of the 1D chain model in the ordered phase. Each
193
+ unit cell is formed with one oxygen (red sphere) and one hydrogen (white sphere). Our model consists of
194
+ a supercell subjected to periodic boundary conditions containing 200 unit cells (for better visualization
195
+ only 8 unit cells are shown).
196
+ stemming from the ice rules restrictions, i.e., in this 1D model, only one proton is attached to each
197
+ oxygen. The last sum in equation 2.4 is restricted to nearest neighbours for each index ⟨𝑖𝑗⟩. There is no
198
+ long-range part in this model, which precludes a phase transition in one dimension. However, the last
199
+ bilinear term is treated in mean-field, which enables the system to have a second order phase transition
200
+ at finite temperature [26]. Therefore, the 1D model total potential, is written in the following way [27]:
201
+ 𝑉1𝐷 (𝑅) =
202
+ ∑︁
203
+ 𝑖
204
+ 𝑉3𝑠 (𝑥𝑖, 𝑅𝑖) − 𝐽⟨𝑥⟩
205
+ ∑︁
206
+ 𝑖
207
+ 𝑥𝑖 + 1
208
+ 2 𝑁𝐽⟨𝑥⟩2,
209
+ (2.5)
210
+ where ⟨𝑥⟩ ≡ 1/𝑁 �
211
+ 𝑖 𝑥𝑖 is the time and lattice average of the 𝑥𝑖 positions for each unit cell 𝑖 taken at each
212
+ MC step in the simulation.
213
+ 2.3. Path integral Monte Carlo simulations
214
+ In the PIMC simulations [28], the effective short-time propagator for two adjacent points in the dis-
215
+ cretized imaginary-time path describing each quantum particle was evaluated to fourth-order accuracy
216
+ with the Takahashi-Imada approximation [28–30]. The effective action in this case allows us to signifi-
217
+ cantly reduce the Trotter number 𝑀 required for convergence. In all the simulations performed we have
218
+ used 𝑀 = 128 beads for the quantum polymer associated with each atom in the O–H...O bonds, which
219
+ yielded well-converged results [19, 25, 28]. Additionally, a normal-mode representation of the quantum
220
+ polymers was used in order to ensure ergodicity in the MC sampling [28, 30]. The PIMC simulations were
221
+ performed at low 𝑇 = 50 K such that the quantum nuclear effects were predominant compared to entropic
222
+ contributions in the 3S model and also with the aim to obtain GE in the ordered phase for the 1D model
223
+ (the classical version of this model has a transition to a disordered paraelectric phase at ≈ 350 K). The
224
+ simulations for the 3S model consisted of 1 × 106 MC steps preceded by 5 × 105 steps of thermalization.
225
+ In the 1D chain model simulations, we took 3 × 104 steps of thermalization plus 1 × 105 MC steps for
226
+ computing averages. In this case, each calculation performed was an average of 20 runs with different
227
+ random number generator seeds.
228
+ To characterize the degree of particle delocalization in the PIMC simulations, we studied the centroid
229
+ and radius of gyration (RG) distributions for the quantum polymers [31]. The centroid is defined as the
230
+ center of mass (CM) of the polymer and represents the average position of the quantum particle. The
231
+ radius of gyration represents the variance of the quantum path and is a quantitative measure of how
232
+ far away are the beads or monomers from the polymer center, and therefore, provides a measure of the
233
+ quantum delocalization of the particle [31].
234
+ 3. Results and discussion
235
+ 3.1. Geometrical effect study using the three-site model
236
+ The six potential parameters of equation (2.3) have been fitted in order to perform the GE study
237
+ with the 3S model. First, we fixed the values of 𝑎 = 2.89 Å
238
+ −1 [20, 21] and 𝐷 = 3.12 eV of the model
239
+ parameters for the proton potential defined in equation 2.1, such that the stretching frequency for the O–H
240
+ bond in the limit 𝑅 → ∞ coincides with the experimental average value 𝜔∞ ≈ 3750 cm−1 [20, 21, 32] for
241
+ 43708-4
242
+
243
+ Quantum geometrical effects in KDP crystals
244
+ (a) Proton
245
+ (b) Deuteron
246
+ Figure 3. (Colour online) Proton/Deuteron probability distribution contours for the three-site PIMC
247
+ simulations at 𝑇 = 50 K.
248
+ different H-bonded compounds. There is a strong correlation between the OH and OO distances for the
249
+ family of H-bonded compounds. The equilibrium distance 𝑟OH diminishes systematically with increasing
250
+ 𝑅 for 𝑅 > 𝑅𝑐 [33, 34], reaching a saturated value around 𝑟∞
251
+ OH ≈ 0.95 Å for very large 𝑅. Therefore, we
252
+ took the parameter value 𝑟0 = 0.93 Å so that the values 𝑥 that minimize 𝑉OH (𝑥, 𝑅) in equation (2.1) for
253
+ different values of 𝑅 give a curve 𝑟min
254
+ OH = 𝑅OO/2 − 𝑥min as a function of 𝑅 that is a lower bound for the
255
+ set of experimental points spread in the OH–OO correlation [20, 21, 33, 34]. With this choice, when the
256
+ nuclear quantum effects are included in the PIMC calculations, we observe a very good agreement with
257
+ the experimental correlation curve using the model of equation (2.1) with the OO distance 𝑅 fixed [35].
258
+ On the other hand, the parameter values for the OO interaction 𝑉OO (𝑅) [see equation (2.2)], were
259
+ initially taken from reference [23]. They were further adjusted, especially the value of 𝐷OO, due to the
260
+ important correlation between 𝑟OH and 𝑅OO, such that the classic potential profile has the minimum at
261
+ 𝑅cl
262
+ OO ≈ 2.55 Å. We considered this condition because the most important geometrical effects are observed
263
+ in H-bonded crystals with strong H-bonds which have distances 𝑅 in a range between 2.5 and 2.6 Å [36],
264
+ with 𝑅cl
265
+ OO lying precisely in the middle of that window. The final parameter values for the 3S model are
266
+ shown in table 1.
267
+ Table 1. Potential parameters used in the 3S model.
268
+ 𝐷 [eV]
269
+ 𝑎 [ Å−1]
270
+ 𝑟0 [ Å]
271
+ 𝐷OO [eV]
272
+ 𝑎OO [ Å−1]
273
+ 𝑅0 [ Å]
274
+ 3.12
275
+ 2.89
276
+ 0.93
277
+ 0.55
278
+ 2.28
279
+ 2.76
280
+ We have verified that the 3S-model PIMC simulations performed at 𝑇 = 50 K with 𝑀 = 128
281
+ beads for the quantum polymer representing each atom yielded probability distributions for the H-bond
282
+ parameters (𝑥 and 𝑅) and energies well converged. The low temperature of 50 K for the simulation was
283
+ chosen because we are interested in the nuclear quantum effects for the H-bonds and the geometrical
284
+ changes with deuteration without most of the influence of entropic contributions in the particle dynamics.
285
+ The 3S model results for the probability density contours to find the system in a given (𝑥, 𝑅) configuration
286
+ are shown in figure 3 for the proton and deuteron cases. The curves are qualitatively different but both
287
+ cases are found to have symmetric distributions around 𝑥 = 0 in the 𝑥 coordinate with two prominent
288
+ peaks with maximum probability, which are clearly shifted in the deuterated case. The OO distance for the
289
+ peak positions are in each case: 𝑅peak
290
+ OO (𝐻) = 2.527 Å and 𝑅peak
291
+ OO (𝐷) = 2.543 Å, which represents a distance
292
+ enlargement for the OO bond of Δ𝑅OO = 0.016 Å, evidencing the geometrical or Ubbelohde effect of
293
+ the H-bond expansion with deuteration. Moreover, the corresponding average values also increase with
294
+ 43708-5
295
+
296
+ F. Torresi, J. Lasave, S. Koval
297
+ -0,3
298
+ -0,2
299
+ -0,1
300
+ 0
301
+ 0,1
302
+ 0,2
303
+ 0,3
304
+ xCM [Å]
305
+ 0,05
306
+ 0,1
307
+ 0,15
308
+ 0,2
309
+ 0,25
310
+ 0,3
311
+ rG [Å]
312
+ (a) Proton
313
+ -0,3
314
+ -0,2
315
+ -0,1
316
+ 0
317
+ 0,1
318
+ 0,2
319
+ 0,3
320
+ xCM [Å]
321
+ 0,05
322
+ 0,1
323
+ 0,15
324
+ 0,2
325
+ 0,25
326
+ 0,3
327
+ rG [Å]
328
+ (b) Deuteron
329
+ Figure 4. (Colour online) Distribution of the radius of gyration 𝑟𝐺 vs. centroid coordinate 𝑥𝐶𝑀 for the
330
+ three-site simulations at 𝑇 = 50 K.
331
+ deuteration: ⟨𝑅OO(𝐻)⟩ = 2.525 Å and ⟨𝑅OO(𝐷)⟩ = 2.540 Å.
332
+ The PIMC simulations also show a change in the variable 𝛿 with deuteration for the peaks observed
333
+ in figure 3. The variation is: Δ𝛿 = 𝛿𝐷 − 𝛿𝐻 = 0.079 Å, where 𝛿𝐻 = 0.417 Å and 𝛿𝐷 = 0.496 Å. This
334
+ is also reflected in a shrinking of the O–H bonds: Δ𝑟 = 𝑟OH − 𝑟OD = 0.032 Å. The overall changes in
335
+ the variables 𝛿 and 𝑅 with deuteration in the simulations are in agreement with what is observed in the
336
+ experimental data for different H-bonded compounds with strong H-bonds [36, 37]. Thus, our simple 3S
337
+ model satisfactorily reproduces the isotopic geometrical effects for these systems.
338
+ It is worth to notice that if the OO distance is not allowed to relax, then the GE is smaller. For
339
+ instance, we have fixed the value 𝑅OO = 2.527 Å, which corresponds to the peak in the probability
340
+ distribution for the protonic system (see figure 3), and the simulations gave a change with deuteration in
341
+ the OH bond of only Δ𝑟 = 0.021 Å. Comparing this result with that considering the oxygen dynamics
342
+ (Δ𝑟 = 𝑟OH − 𝑟OD = 0.032 Å), we observe an increment of ≈ 50% in the isotopic geometrical effect in the
343
+ case where the oxygens are allowed to relax. This can be understood in the following way: first, when
344
+ the oxygens are fixed, protons, being more delocalized than deuterons, have more probability to stay
345
+ closer to the middle of the O–O bond. Second, when the oxygen dynamics is included, the protons act
346
+ as a strong attraction center that pulls the two bridge oxygens together, more effectively than deuterons
347
+ which are more localized near the oxygen. This proton-mediated O–O contraction lowers the potential
348
+ barrier, which delocalizes even more the proton, and so on, giving rise to a nonlinear selfconsistent
349
+ mechanism [9, 10]. For the deuteron, being less delocalized than the proton, the selfconsistent effect is
350
+ weaker. This mechanism leads to an isotopic geometrical effect which is stronger than that generated by
351
+ the proton/deuteron quantum delocalization at fixed potential (fixed oxygens) [9, 10].
352
+ To further illustrate the microscopic mechanism that rules the GE, we have analyzed the behavior
353
+ of the quantum polymers for the proton/deuteron in the simulation via an analysis of the center of mass
354
+ of the quantum polymer or centroid position 𝑥𝐶𝑀 and the radius of gyration 𝑟𝐺 representing a measure
355
+ of the quantum delocalization of the particle (i.e., the extension of the quantum polymer) [31]. We plot
356
+ in figure 4 the instantaneous values of 𝑟𝐺 as a function of the proton/deuteron centroids 𝑥𝐶𝑀, taken
357
+ every 100 MC steps in the PIMC simulation. As can be seen in the figure, the density of points reveals
358
+ that the deuteron prefers to be localized at both sides and far from the bond middle with small values
359
+ of 𝑟𝐺, indicating a more classical behavior in these cases. When the deuteron centroid takes the values
360
+ of 𝑥𝐶𝑀 closer to 0 (the bond middle), it is observed an increase of 𝑟𝐺 indicating that the quantum polymer
361
+ is delocalized and is spread through both sides of the potential barrier, signaling the presence of tunneling
362
+ in this case. Notice that the largest values of 𝑟𝐺 are found at 𝑥𝐶𝑀 ≈ 0 where delocalization is maximum.
363
+ On the other hand, in the proton case, tunneling is much more frequent because the region with larger
364
+ density of points appears near 𝑥𝐶𝑀 ≈ 0 with large values of 𝑟𝐺, as shown in figure 4. This is precisely
365
+ 43708-6
366
+
367
+ Quantum geometrical effects in KDP crystals
368
+ an important ingredient for the GE: the proton spends much more time delocalized with the quantum
369
+ polymer center of mass near the middle of the O–O bond, which finally produces a strong contraction
370
+ of the O–O distance. On the contrary, the deuteron is much more localized at both sides and far from
371
+ the bond middle which leads to a weakening of the O–O bond and to an increase of the O–O distance.
372
+ This yields the isotopic geometrical effect, which is observed in the calculated probability distribution of
373
+ figure 3.
374
+ 3.2. Isotope effects obtained with the 1D model simulations
375
+ The previous analysis of the 3S model results, which has clearly shown the isotopic GE, was carried
376
+ out based on the parametrization of the model which reproduces the universal OH–OO correlation
377
+ observed for a family of diverse H-bonded compounds. In this sense, this model is quite simple and
378
+ general, accounting for the geometrical effects with deuteration of a set of H-bonded ferroelectrics with
379
+ strong H-bonds. Now, we focus on the development of a 1D chain model, described in section 2.2 [see
380
+ equation (2.5)], which was specifically designed to explain the isotope effects in the phase transition of
381
+ KDP and was fitted to ab initio results [19]. This more realistic 1D model has, in the classical nuclei
382
+ version, a ferroelectric-paraelectric transition at 𝑇 ≈ 350 K [35]. In this paper, we have used it in the
383
+ ordered phase of KDP at 𝑇 = 50 K to analyze the isotopic GE which is at the root of the microscopic
384
+ mechanism that leads to the giant isotope effect in the critical temperature.
385
+ We start from equation 2.5 for the 1D model, which has seven parameters to be adjusted for the
386
+ KDP case. The six model parameters of the local proton potential 𝑉3𝑆 for each unit cell in the chain,
387
+ which is just the same that was used in the 3S model (see equation 2.3), have been adjusted to reproduce
388
+ six magnitudes obtained from ab initio calculations for KDP. These magnitudes were the global energy
389
+ barrier between the PE and FE states, the O–O and 𝛿 distances in the FE phase, the O–O distance in
390
+ the PE phase, the ab initio vibrational frequency of the PO4 rotation mode, which is equivalent to the
391
+ stretching mode in the 3S model, and the energy barrier between the energy minimum and the transition
392
+ state in the FE phase keeping the O–O distance fixed (see reference [19]). We adopted the model fit to the
393
+ ab initio calculations that includes dispersion corrections at the vdW-DF level, which exhibit, compared
394
+ to other methods, the best agreement with the experimental geometry for both KDP and deuterated KDP
395
+ (DKDP) [19].
396
+ Finally, we have fitted the remaining parameter 𝐽 that corresponds to the proton-proton interaction
397
+ term in equations (2.4) and (2.5). To this end, 𝐽 was adjusted to 0.55 eV/Å
398
+ 2 so that the critical temperature
399
+ 𝑇𝑐 for the FE-PE transition obtained by the 1D model simulation with classical nuclei reaches the value
400
+ of ≈ 350 K, similar to the value obtained by ab initio molecular dynamics calculations with dispersion
401
+ corrections at the vdW-DF level for DKDP [38].
402
+ The final values for the parameters used in the 1D model are listed in table 2.
403
+ Table 2. Potential parameters used in the 1D model.
404
+ 𝐷 [eV]
405
+ 𝑎 [Å−1]
406
+ 𝑟0 [Å]
407
+ 𝐷OO [eV]
408
+ 𝑎OO [Å−1]
409
+ 𝑅0 [Å]
410
+ 𝐽 [ eV/Å
411
+ 2]
412
+ 8.838
413
+ 3.027
414
+ 0.966
415
+ 10.542
416
+ 0.831
417
+ 2.917
418
+ 0.55
419
+ The motion of the proton/deuteron is strongly correlated with that of the heavy ions, and its mass is
420
+ dressed accordingly as discussed in reference [10]. Therefore, instead of using the bare proton (deuteron)
421
+ masses 𝑚 𝑝 (2𝑚 𝑝), we have used in the PIMC simulations the effective masses for H and D: 𝜇𝐻 = 2.3𝑚 𝑝
422
+ and 𝜇𝐷 = 3𝑚 𝑝, respectively, with 𝑚 𝑝 the proton mass [9, 10, 19].
423
+ We plot in figure 5 the probability distribution contours for the PIMC simulation with the 1D model,
424
+ obtained in the ordered phase at 𝑇 = 50 K. Due to the ordered phase, only one peak is observed in the
425
+ proton and deuteron distributions, which is in contrast to the symmetrical double peaks around 𝑥 = 0 found
426
+ in the 3S model distribution results (see figure 3). The calculated distribution for the chain of protons
427
+ in figure 5 is asymmetric around the peak position due to the potential anharmonicity and quantum
428
+ delocalization, which is in qualitative agreement with the experimental diffraction pattern measured near
429
+ 43708-7
430
+
431
+ F. Torresi, J. Lasave, S. Koval
432
+ (a) Proton
433
+ (b) Deuteron
434
+ Figure 5. (Colour online) Proton/Deuteron probability distribution contours in the H-bonds for the linear-
435
+ chain PIMC simulation at 𝑇 = 50 K.
436
+ 𝑇𝑐 in the FE phase of KDP [39]. The asymmetry around the peak is less pronounced in the deuterated
437
+ case as shown in figure 5, because the deuteron is less delocalized than the proton.
438
+ The prominent single peak found in the distribution results for the 1D simulation is clearly shifted
439
+ in the deuterated case towards larger 𝑥 and 𝑅, revealing the existence of the isotopic geometrical effect,
440
+ i.e., the expansion of the H-bonds in the chain with deuteration. The O–O distance for the peak positions
441
+ are in each case: 𝑅peak
442
+ OO (𝐻) = 2.515 Å and 𝑅peak
443
+ OO (𝐷) = 2.542 Å, which represents a distance enlargement
444
+ for the O–O bond of Δ𝑅OO ≡ 𝑅OO(𝐷) − 𝑅OO(𝐻) = 0.027 Å. The 𝑥 coordinate of the peak position also
445
+ expands with deuteration, from 𝑥peak
446
+ 𝐻
447
+ = 0.188 Å to 𝑥peak
448
+ 𝐷
449
+ = 0.218 Å, with a net increase of Δ𝑥 = 0.030 Å
450
+ or similarly Δ𝛿 ≡ 𝛿𝐷 − 𝛿𝐻 = 0.060 Å. These results are summarized in table 3 and compared with the
451
+ available experimental data for KDP and DKDP [40]. We observe a good agreement between theory
452
+ and experiment, although the GE is a little bit underestimated, with difference values under deuteration
453
+ ≈ 25% lower than the experimental data.
454
+ Table 3. Nuclear quantum calculations of the H-bond geometries for KDP and DKDP using the 1D
455
+ linear model. The results, which correspond to the peak positions of figure 5, are contrasted with the
456
+ experimental data of reference [40]. Distances are in Å.
457
+ PIMC
458
+ KDP (𝜇𝐻 = 2.3 𝑚 𝑝)
459
+ DKDP (𝜇𝐷 = 3.0 𝑚 𝑝)
460
+ Δ𝑅OO
461
+ Δ𝛿
462
+ results
463
+ 𝑅OO
464
+ 𝛿
465
+ 𝑅OO
466
+ 𝛿
467
+ 1D model
468
+ 2.515
469
+ 0.376
470
+ 2.542
471
+ 0.436
472
+ 0.027
473
+ 0.060
474
+ Expt. [40]
475
+ 2.497
476
+ 0.385
477
+ 2.533
478
+ 0.472
479
+ 0.036
480
+ 0.087
481
+ To get a deeper insight into the microscopic mechanism of the geometrical effect in the linear chain
482
+ model, we plot in figure 6 the distribution of the instantaneous radius of gyration 𝑟𝐺 as a function of the
483
+ centroid positions 𝑥𝐶𝑀 for all H-bonds in the chain, where the points are taken every 100 MC steps along
484
+ the PIMC simulation. The region with largest density of points in figure 6 coincides with the position of
485
+ the peaks in both proton and deuteron cases (see figure 5). We again observe an asymmetric distribution
486
+ centered in one of the sides of the H-bond consistent with the (𝑥, 𝑅) distribution pattern of figure 5. The
487
+ asymmetry observed in figure 6 is more pronounced in the proton case, indicating that protons jump more
488
+ often than deuterons to the other side of the O–H–O bond. The mechanism to pass through the potential
489
+ barrier is to increase the radius of gyration near 𝑥𝐶𝑀 ≈ 0 which means that the particle tunnels through
490
+ the barrier. This is helped by a strong contraction of the 𝑅 distance, which diminishes concomitantly with
491
+ 43708-8
492
+
493
+ Quantum geometrical effects in KDP crystals
494
+ (a) Proton
495
+ (b) Deuteron
496
+ Figure 6. Distribution of the radius of gyration 𝑟𝐺 vs. centroid coordinate 𝑥𝐶𝑀 of the quantum polymer
497
+ representing the protons (a) and deuterons (b) relative to the center of the H-bonds, for the linear-chain
498
+ PIMC simulation at 𝑇 = 50 K.
499
+ the potential barrier, to a lower bound of 𝑅min ≈ 2.3 Å near 𝑥 = 0 as shown in figure 5. Thus, we conclude
500
+ that tunneling is assisted by the 𝑅 distance modulation. However, in this ordered phase at 𝑇 = 50 K,
501
+ the proton spends more time in one of the sides of the O–H–O bond where the behavior is more classic
502
+ (low value of 𝑟𝐺). On the other hand, in the deuteron case, the particle remains localized practically all
503
+ the time, with a general classical behavior with low values of 𝑟𝐺. In other words, the tunneling for the
504
+ deuteron is very scarce. These results are consistent with the general assumption in the tunneling model:
505
+ protons are capable of tunelling while deuterons are not [3]. However, there is an essential difference:
506
+ protons tunnel being assisted by the strong correlation with the O–O distance, which is the behavior that
507
+ originates the geometrical effect [9, 10]. Therefore, the proton has a larger probability than the deuteron
508
+ to spend more time tunneling through the barrier near the middle of the O–H–O bond, and this generates
509
+ a strong attraction center that pulls the two oxygens together, much more efficiently than deuterons. This
510
+ “tunneling – geometrical effect” interrelation gives rise to the final geometrical effect observed in KDP
511
+ crystals, that is, the H-bond expansion with deuteration, which is crucial for the isotope effects in the
512
+ FE-PE phase transitions [9, 35].
513
+ 4. Summary and conclusions
514
+ We have carried out PIMC simulations with simple models to account for the geometrical effects (GE)
515
+ with deuteration in H-bonded ferroelectrics such as KDP crystals. Firstly, we have developed a general
516
+ three-site (3S) model consisting in a back-to-back double Morse potential for the O–H interaction and
517
+ a Morse potential which represents the interaction between the oxygens and the lattice. The model was
518
+ fitted to reproduce general features for a large set of different H-bonded compounds. The computed
519
+ probability distribution contours in the (𝑅, 𝑥) configuration space, with 𝑅 the O–O distance and 𝑥 the
520
+ proton/deuteron distance to the middle of the O–O bond, reveal a symmetric distribution around 𝑥 = 0
521
+ with two peaks on either side, for both proton and deuteron cases. The results show an increase with
522
+ deuteration of 𝑅 and 𝑥 for the observed peaks, i.e., a GE, which is in agreement with that observed in
523
+ H-bonded compounds with strong H-bonds. Moreover, if the oxygens are not allowed to relax during the
524
+ simulation, the GE in the 𝑥 coordinate is much smaller, which means that there is a strong correlation
525
+ between 𝑅 and 𝑥 that is important for the GE. During the PIMC simulations we have also plotted the
526
+ instantaneous radius of gyration 𝑟𝐺 vs. the centroid position 𝑥𝐶𝑀 of the quantum polymer representing
527
+ the proton/deuteron. The results show that the proton tunnels more frequently than the deuteron (that is,
528
+ it spends more time with the center of mass near 𝑥𝐶𝑀 = 0 with large values of 𝑟𝐺), while the deuteron is
529
+ 43708-9
530
+
531
+ 0,25
532
+ 0,2
533
+ 0,15
534
+ rG
535
+ 0,1
536
+ 0,05
537
+ -0,4
538
+ -0,2
539
+ 0
540
+ 0,2
541
+ 0,4
542
+ XcM [A]0,25
543
+ 0,2
544
+ 0,15
545
+ rG
546
+ 0,1
547
+ 0,05
548
+ -0,4
549
+ -0,2
550
+ 0
551
+ 0,2
552
+ 0,4
553
+ XcM [A]F. Torresi, J. Lasave, S. Koval
554
+ more localized in both sides and far from the O–H–O bond center, with small values of 𝑟𝐺 (i.e., a more
555
+ classsical behavior). These features yield a more effective contraction of the O–O bond in the proton
556
+ case, explaining the GE observed.
557
+ Secondly, we have developed a more realistic 1D model, with the same local potential for the H-bonds
558
+ as that used in the 3S model, but adding also a bilinear proton-proton interaction treated in mean-field.
559
+ The parameters of the 1D model were fitted to ab initio results for KDP. The bilinear interaction parameter
560
+ of the model was adjusted such that the classical nuclei version of the model has a second order FE-PE
561
+ phase transition at 𝑇 = 350 K in agreement with ab initio molecular dynamics simulations for DKDP.
562
+ In this paper, by means of PIMC simulations of the 1D model, we have studied the GE caused by
563
+ deuteration in the ordered phase at 𝑇 = 50 K. The calculated probability distribution contours show
564
+ only one peak in the (𝑅, 𝑥) configuration space for both proton/deuteron cases. The distribution is more
565
+ asymmetric in the proton case due to the anharmonicity of the potential and the quantum delocalization.
566
+ The distribution pattern is in qualitative agreement with the experimental distribution determined by high-
567
+ resolution neutron diffraction studies [39]. The probability distribution contours show a peak which shifts
568
+ substantially with deuteration. The changes in H-bond geomentry caused by the GE observed in the 1D
569
+ model simulations are in good agreement with the corresponding experimental data. The distribution of
570
+ the radius of gyration vs. the quantum path centroids shows that the protons tunnel through the potential
571
+ barrier frequently while the deuterons are much more localized in one of the sides of the O–H–O bond
572
+ and practically do not tunnel, in agreement with the well-known tunneling model [3], and also with
573
+ recent neutron Compton scattering experiments [7, 8]. We have shown that proton tunneling is assisted
574
+ by a strong contraction of the O–O distance in the 1D model. Thus, there is a strong correlation between
575
+ instantaneous tunneling and geometrical effects of the H-bond that is much more efficient in the proton
576
+ case than in the deuterated system, which gives in average a strong GE for the whole simulation. This
577
+ mechanism is expected to be at the root of the huge isotope effect observed in H-bonded ferroelectrics of
578
+ the KDP type [9, 10].
579
+ Acknowledgements
580
+ We acknowledge support from Consejo Nacional de Investigaciones Científicas y Técnicas (CON-
581
+ ICET), Argentina.
582
+ References
583
+ 1. Lines M. E., Glass A. M., Principles and Applications of Ferroelectric and Related Materials, Clarendon, Oxford,
584
+ 1977.
585
+ 2. Samara G. A., Ferroelectrics, 1973, 5, 25, doi:10.1080/00150197308235776.
586
+ 3. Blinc R., J. Phys. Chem. Solids, 1960, 13, 204, doi:10.1016/0022-3697(60)90003-2.
587
+ 4. McMahon M. I., Nelmes R. J., Kuhst W. F., Dorwarth R., Piltz R. O., Tun Z., Nature, 1990, 348, 317,
588
+ doi:10.1038/348317a0.
589
+ 5. Nelmes R. J., McMahon M. I., Piltz R. O., Wright N. G., Ferroelectrics, 1991, 124, 355,
590
+ doi:10.1080/00150199108209465.
591
+ 6. Robertson J. M., Ubbelohde A. R. J. P., Proc. R. Soc. London, Ser. A, 1939, 170, 222,
592
+ doi:10.1098/rspa.1939.0028.
593
+ 7. Reiter G. F., Mayers J., Platzman P., Phys. Rev. Lett., 2002, 89, 135505, doi:10.1103/PhysRevLett.89.135505.
594
+ 8. Reiter G., Shukla A., Platzman P. M., Mayers J., New J. Phys., 2008, 10, 013016, doi:10.1088/1367-
595
+ 2630/10/1/013016.
596
+ 9. Koval S., Kohanoff J., Migoni R. L., Tosatti E., Phys. Rev. Lett., 2002, 89, 187602,
597
+ doi:10.1103/PhysRevLett.89.187602.
598
+ 10. Koval S., Kohanoff J., Lasave J., Colizzi G., Migoni R. L., Phys. Rev. B, 2005, 71, 184102,
599
+ doi:10.1103/PhysRevB.71.184102.
600
+ 11. Blinc R., Svetina S., Phys. Rev., 1966, 147, 430, doi:10.1103/PhysRev.147.430.
601
+ 12. Kojyo M., Onodera Y., J. Phys. Soc. Jpn., 1988, 57, 4391, doi:10.1143/JPSJ.57.4391.
602
+ 13. Sugimoto H., Ikeda S., Phys. Rev. Lett., 1991, 67, 1306, doi:10.1103/PhysRevLett.67.1306.
603
+ 43708-10
604
+
605
+ Quantum geometrical effects in KDP crystals
606
+ 14. Shchur Ya. I., Levitskii R. R., Vlokh O. G., Kityk A. V., Vysochansky Y. M., Grabar A. A., Condens. Matter
607
+ Phys., 1999, 2, 93, doi:10.5488/CMP.2.1.93.
608
+ 15. Merunka D., Rakvin B., Phys. Rev. B, 2002, 66, 174101, doi:10.1103/PhysRevB.66.174101.
609
+ 16. Shchur Ya., Phys. Rev. B, 2006, 74, 054301, doi:10.1103/PhysRevB.74.054301.
610
+ 17. Levitskii R. R., Zachek I. R., Vdovych A. S., Sorokov S. I., Condens. Matter Phys., 2009, 12, 75,
611
+ doi:10.5488/CMP.12.1.75.
612
+ 18. Lasave J., Kohanoff J., Migoni R. L., Koval S., Physica B, 2009, 404, 2736, doi:10.1016/j.physb.2009.06.143.
613
+ 19. Menchón R., Colizzi G., Johnston C., Torresi F., Lasave J., Koval S., Kohanoff J., Migoni R., Phys. Rev. B, 2018
614
+ 98, 104108, doi:10.1103/PhysRevB.98.104108.
615
+ 20. Matsushita E., Matsubara T., Prog. Theor. Phys., 1982, 67, 1, doi:10.1143/PTP.67.1.
616
+ 21. Matsubara T., Matsushita E., Prog. Theor. Phys., 1984, 71, 209, doi:10.1143/PTP.71.209.
617
+ 22. Tanaka S., Phys. Rev. B, 1990, 42, 10488, doi:10.1103/PhysRevB.42.10488.
618
+ 23. Yanovitskii O., Vlastou-Tsinganos G., Flytzanis N., Phys. Rev. B, 1993, 48, 12645,
619
+ doi:10.1103/PhysRevB.48.12645.
620
+ 24. Scivetti I., Gidopoulos N., Kohanoff J., Phys. Rev. B, 2008, 78, 224108, doi:10.1103/PhysRevB.78.224108.
621
+ 25. Lasave J., Abufager P., Koval S., Phys. Rev. B, 2016, 93, 134112, doi:10.1103/PhysRevB.93.134112.
622
+ 26. Koehler T. R., Gillis N. S., Phys. Rev. B, 1976, 13, 4183, doi:10.1103/PhysRevB.13.4183.
623
+ 27. Tuckerman M. E., Statistical Mechanics: Theory and Molecular Simulation, Oxford University Press Inc., New
624
+ York, 2010.
625
+ 28. Chakravarty C., Int. Rev. Phys. Chem., 1997, 16, 421, doi:10.1080/014423597230190.
626
+ 29. Takahashi M., Imada M., J. Phys. Soc. Jpn., 1984, 53, 3765, doi:10.1143/JPSJ.53.3765.
627
+ 30. Weht R. O., Kohanoff J., Estrín D. A., Chakravarty C., J. Chem. Phys., 1998, 108, 8848, doi:10.1063/1.476331.
628
+ 31. Morrone J. A., Lin L., Car R., J. Chem. Phys., 2009, 130, 204511, doi:10.1063/1.3142828.
629
+ 32. McKenzie R. H., Bekker C., Athokpam B., Ramesh S. G., J. Chem. Phys., 2014, 140, 174508,
630
+ doi:10.1063/1.4873352.
631
+ 33. Ichikawa M., Acta Crystallogr., Sect. B: Struct. Crystallogr. Cryst. Chem., 1978, 34, 2074–2080,
632
+ doi:10.1107/S0567740878007475.
633
+ 34. Joswig W., Fuess H., Ferraris G., Acta Crystallogr., Sect. B: Struct. Crystallogr. Cryst. Chem., 1982, 38,
634
+ 2798–2801, doi:10.1107/S0567740882009984.
635
+ 35. Torresi F., Lasave J., Koval S., (unpublished).
636
+ 36. Ichikawa M., J. Mol. Struct., 2000, 552, 63–70, doi:10.1016/S0022-2860(00)00465-8.
637
+ 37. Sokolov N. D., Verner M. V., Savel’ev V. A., J. Mol. Struct., 1988, 177, 93–110, doi:10.1016/0022-
638
+ 2860(88)80081-4.
639
+ 38. Menchón R. E., Ph.D. Thesis, Universidad Nacional de Rosario (UNR), Argentina, 2019.
640
+ 39. Nelmes R. J., Kuhs W. F., Howard C. J., Tibballs J. E., Ryan T. W., J. Phys. C: Solid State Phys., 1985, 18, L711,
641
+ doi:10.1088/0022-3719/18/24/001.
642
+ 40. Nelmes R. J., Tun Z., Kuhs W. F., Ferroelectrics, 1987, 71, 125, doi:10.1080/00150198708224833.
643
+ 43708-11
644
+
645
+ F. Torresi, J. Lasave, S. Koval
646
+ Метод iнтегралiв за траєкторiями у моделюваннi
647
+ Монте-Карло геометричних ефектiв у кристалах KDP
648
+ Ф. Торрезi, Х. Ласаве, С. Коваль
649
+ Iнститут фiзики Росарiо, Нацiональний унiверситет Росарiо та Нацiональна рада з науково-технiчних
650
+ дослiджень, вул. 27 лютого, 210 Bis, 2000 Росарiо, Аргентина
651
+ Метод iнтегралiв за траєкторiями у моделюваннi Монте-Карло (IТМК) для дуже простих моделей застосо-
652
+ вано для з’ясування фiзичних механiзмiв, що лежать в основi iзотопiчного ефекту в сегнетоелектриках з
653
+ водневими зв’язками. Зумовленi дейтеруванням геометричнi ефекти у водневих зв’язках було дослiдже-
654
+ но за допомогою загальної тривузлової моделi, в якiй використовуються подвiйний потенцiал Морзе та
655
+ потенцiал Морзе мiж киснями; параметри моделi вибрано так, щоб пояснити рiзноманiтнi загальнi влас-
656
+ тивостi низки сполук з водневими зв’язками. З розрахункiв у рамках цiєї моделi випливає виникнення
657
+ геометричного ефекту (ефекту Уббелоде): видовження водневого зв’язка при дейтеруваннi, i це узгоджу-
658
+ ється з тим, що спостерiгається в сегнетоелектриках з короткими водневими зв’язками. Використовуючи
659
+ для параметрiв потенцiалiв результати першопринципних розрахункiв, розвинено одновимiрну модель,
660
+ в якiй бiлiнiйнi протон-протоннi взаємодiї розглядаються в наближеннi середнього поля. Ця модель вико-
661
+ ристовується для дослiдження квантових ефектiв у ядрах, якi призводять до виникнення геометричного
662
+ ефекту в кристалах KDP. Пiдхiд IТМК дає змогу виявити, що протони тунелюють бiльш ефективно вздовж
663
+ одновимiрного ланцюжка, нiж дейтрони; це спричиняє появу сильного притягувального центра, який
664
+ зменшує вiдстань мiж атомами киснiв. Цей механiзм, який ґрунтується на кореляцiї мiж тунелюванням i
665
+ геометричними змiнами водневих зв’язкiв, призводить до виникнення сильного геометричного ефекту
666
+ в ланцюжку у впорядкованiй фазi при низьких температурах, що добре узгоджується з експерименталь-
667
+ ними даними.
668
+ Ключовi слова: сегнетоелектричний фазовий перехiд, сегнетоелектрики з водневими зв’язками, метод
669
+ iнтегралiв за траєкторiями у моделюваннi Монте-Карло
670
+ 43708-12
671
+
MtAzT4oBgHgl3EQfkv1s/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf ADDED
Binary file (59 kB). View file
 
MtE2T4oBgHgl3EQfqgg6/content/tmp_files/2301.04039v1.pdf.txt ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.04039v1 [physics.hist-ph] 10 Jan 2023
2
+ A Theory of Theories
3
+ Mich`ele Levi
4
+ Mathematical Institute, University of Oxford,
5
+ Oxford OX2 6GG, United Kingdom
6
+ levi@maths.ox.ac.uk
7
+ Abstract
8
+ We take a tour through the past, present and future of Effective
9
+ Field Theory, with applications ranging from LHC physics to cosmol-
10
+ ogy.
11
+ 1
12
+
13
+ High-energy physics spans a wide range of energies, from a few MeV to
14
+ TeV, that are all relevant. It is therefore often difficult to take all phenomena
15
+ into account at the same time. Effective field theories (EFTs) are designed
16
+ to break down this range of scales into smaller segments so that physicists
17
+ can work in the relevant range. Theorists “cut” their theory’s energy scale
18
+ at the order of the mass of the lightest particle omitted from the theory,
19
+ such as the proton mass.
20
+ Thus, multi-scale problems reduce to separate
21
+ and single-scale problems.EFTs are today also understood to be “bottom-
22
+ up” theories. Built only out of the general field content and symmetries at
23
+ the relevant scales, they allow us to test hypotheses efficiently and to select
24
+ the most promising ones without needing to know the underlying theories in
25
+ full detail. Thanks to their applicability to all generic classical and quantum
26
+ field theories, the sheer variety of EFT applications is striking.
27
+ In hindsight, particle physicists were working with EFTs from as early
28
+ as Fermi’s phenomenological picture of beta decay in which a four-fermion
29
+ vertex replaces the W-boson propagator because the momentum is much
30
+ smaller compared to the mass of the W boson.
31
+ Like so many profound
32
+ concepts in theoretical physics, EFT was first considered in a narrow phe-
33
+ nomenological context. One of the earliest instances was in the 1960s, when
34
+ ad-hoc methods of current algebras were utilised to study weak interactions
35
+ of hadrons.
36
+ This required detailed calculations, and a simpler approach
37
+ was needed to derive useful results. The heuristic idea of describing hadron
38
+ dynamics with the most general Lagrangian density based on symmetries,
39
+ the relevant energy scale and the relevant particles, which can be written in
40
+ terms of operators multiplied by Wilson coefficients, was yet to be known.
41
+ With this approach, it was possible to encode local symmetries in terms of
42
+ the current algebra due to their association with conserved currents.
43
+ For strong interactions, physicists described the interaction between pi-
44
+ ons with chiral perturbation theory, an effective Lagrangian, which sim-
45
+ plified current algebra calculations and enabled the low-energy theory to
46
+ be investigated systematically. This “mother” of modern EFTs describes
47
+ the physics of hadrons and remains valid to an energy scale of the proton
48
+ mass. Heavy-quark effective theory (HQET), introduced by Howard Georgi
49
+ in 1990, complements chiral perturbation theory by describing the interac-
50
+ tions of charm and bottom quarks. HQET allowed us to make predictions on
51
+ B-meson decay rates, since the corrections could now be classified. The more
52
+ powers of energy are allowed, the more infinities appear. These infinities are
53
+ cancelled by available counter-terms.
54
+ Similarly, it is possible to regard the Standard Model as the truncation
55
+ of a much more general theory including non-renormalizable interactions,
56
+ which yield corrections of higher order in energy. This perception of the
57
+ whole Standard Model as an effective field theory started to be formed in
58
+ the late 1970s by Weinberg and others. Among the known corrections to
59
+ the Standard Model that do not satisfy its approximate symmetries are
60
+ 2
61
+
62
+ neutrino masses, postulated in the 1960s and discovered via the observation
63
+ of neutrino oscillations in the late 1990s.
64
+ While the scope of EFTs was
65
+ unclear initially, today we understand that all successful field theories, with
66
+ which we have been working in many areas of theoretical physics, are nothing
67
+ but effective field theories. EFTs provide the theoretical framework to probe
68
+ new physics and to establish precision programmes at experiments.
69
+ The
70
+ former is crucial for making accurate theoretical predictions, while the latter
71
+ is central to the physics programme of CERN in general.
72
+ 1
73
+ EFTs in Particle Physics
74
+ More than a decade has passed since the first run of the LHC, in which the
75
+ Higgs boson and the mechanism for electroweak symmetry breaking were
76
+ discovered. So far, there are no signals of new physics beyond the SM. EFTs
77
+ are well suited to explore LHC physics in depth. A typical example for an
78
+ event involving two scales is Higgs-boson production because there is a factor
79
+ 10−100 between its mass and transverse momentum. The calculation of each
80
+ Higgs-boson production process leads to large logarithms that can invalidate
81
+ perturbation theory due to the large-scale separation. This is just one of
82
+ many examples of the two-scale problem that arises when the full quantum
83
+ field theory approach for high-energy colliders is applied. Traditionally, such
84
+ two-scale problems have been treated in the framework of QCD factorisation
85
+ and resummation.
86
+ Over the past two decades, it has been possible to recast two-scale prob-
87
+ lems at high-energy colliders with the advent of soft-collinear effective theory
88
+ (SCET). SCET is nowadays a popular framework that is used to describe
89
+ Higgs physics, jets and their substructure, as well as more formal problems,
90
+ such as power corrections to reconstruct full amplitudes eventually.
91
+ The
92
+ difference between HQET and SCET is that SCET considers long-distance
93
+ interactions between quarks and both soft and collinear particles, whereas
94
+ HQET takes into account only soft interactions between a heavy quark and
95
+ a parton. SCET is just one example where the EFT methodology has been
96
+ indispensable, even though the underlying theory at much higher energies
97
+ is known. Other examples of EFT applications include precision measure-
98
+ ments of rare decays that can be described by QCD with its approximate
99
+ chiral symmetry, or heavy quarks at finite temperature and density. EFT
100
+ is also central to a deeper understanding of the so-called flavour anomalies,
101
+ enabling comparisons between theory and experiment in terms of particular
102
+ Wilson coefficients.
103
+ Moreover, precision measurements of Higgs and electroweak observables
104
+ at the LHC and future colliders will provide opportunities to detect new
105
+ physics signals, such as resonances in invariant mass plots, or small devia-
106
+ tions from the SM, seen in tails of distributions for instance at the HL-LHC
107
+ 3
108
+
109
+ – testing the perception of the SM as a low-energy incarnation of a more fun-
110
+ damental theory being probed at the electroweak scale. This is dubbed the
111
+ SMEFT (SM EFT) or HEFT (Higgs EFT), depending on whether the Higgs
112
+ fields are expressed in terms of the Higgs doublet or the physical Higgs bo-
113
+ son. This particular EFT framework has recently been implemented in the
114
+ data-analysis tools at the LHC, enabling the analyses across different chan-
115
+ nels and even different experiments.At the same time, the study of SMEFT
116
+ and HEFT has sparked a plethora of theoretical investigations that have
117
+ uncovered its remarkable underlying features, for example allowing EFT to
118
+ be extended or placing constraints on the EFT coefficients due to Lorentz
119
+ invariance, causality and analyticity.
120
+ 2
121
+ EFTs in Gravity
122
+ Since the inception of EFT, it was believed that the framework is applicable
123
+ only to the description of quantum field theories for capturing the physics
124
+ of elementary particles at high-energy scales, or alternatively at very small
125
+ length scales. Thus, EFT seemed mostly irrelevant regarding gravitation,
126
+ for which we are still lacking a full theory valid at quantum scales. The only
127
+ way in which EFT seemed to be pertinent for gravitation was to think of
128
+ general relativity as a first approximation to an EFT description of quantum
129
+ gravity, which indeed provided a new EFT perspective at the time. However,
130
+ in the past decade it has become widely acknowledged that EFT provides a
131
+ powerful framework to capture gravitation occurring completely across large
132
+ length scales, as long as these scales display a clear hierarchy.
133
+ The most notable application to such classical gravitational systems
134
+ came when it was realised that the EFT framework would be ideal to handle
135
+ gravitational radiation emitted at the inspiral phase of a binary of compact
136
+ objects, such as black holes. At this phase in the evolution of the binary, the
137
+ compact objects are moving at non-relativistic velocities. Using the small
138
+ velocity as the expansion parameter exhibits the separation between the
139
+ various characteristic length scales of the system. Thus, the physics can be
140
+ treated perturbatively. For example, it was found that even couplings man-
141
+ ifestly change in classical systems across their characteristic scales, which
142
+ was previously believed to be unique to quantum field theories. The appli-
143
+ cation of EFT to the binary inspiral problem has been so successful that
144
+ the precision frontier has been pushed beyond the state of the art, quickly
145
+ surpassing the reach of work that has been focused on the two-body problem
146
+ for decades via traditional methods in general relativity.
147
+ This theoretical progress has made an even broader impact since the
148
+ breakthrough direct discovery of gravitational waves (GWs) was announced
149
+ in 2016. An inspiraling binary of black holes merged into a single black hole
150
+ in less than a split second, releasing an enormous amount of energy in the
151
+ 4
152
+
153
+ form of GWs, which instigated even greater, more intense use of EFTs for
154
+ the generation of theoretical GW data. In the coming years and decades,
155
+ a continuous increase in the quantity and quality of real-world GW data is
156
+ expected from the rapidly growing worldwide network of ground-based GW
157
+ detectors, and future space-based interferometers, covering a wide range of
158
+ target frequencies.
159
+ 3
160
+ EFTs in Cosmology
161
+ Cosmology is inherently a cross-cutting domain, spanning scales over about
162
+ 1060 orders of magnitude, from the Planck scale to the size of the observable
163
+ universe. As such, cosmology generally cannot be expected to be tackled
164
+ directly by each of the fundamental theories that capture particle physics
165
+ or gravity. The correct description of cosmology relies heavily on the work
166
+ in many disparate areas of research in theoretical and experimental physics,
167
+ including particle physics and general relativity among many more.
168
+ The development of EFT applications in cosmology – including EFTs of
169
+ inflation, dark matter, dark energy and even EFTs of large-scale structure
170
+ – has become essential to make observable predictions in cosmology. The
171
+ discovery of the accelerated expansion of the universe in 1998 shows our diffi-
172
+ culty in understanding gravity both at the quantum regime and the classical
173
+ one. The cosmological constant problem and dark-matter paradigm might
174
+ be a hint for alternative theories of gravity at very large scales. Indeed, the
175
+ problems with gravity in the very-high and very-low energy range may well
176
+ be tied together. The science programme of next-generation large surveys,
177
+ such as ESA’s Euclid satellite, rely heavily on all these EFT applications
178
+ for the exploitation of the enormous data that is going to be collected to
179
+ constrain unknown cosmological parameters, thus helping to pinpoint viable
180
+ theories.
181
+ 4
182
+ The Future of EFTs in Physics
183
+ The EFT framework plays a key role at the exciting and rich interface be-
184
+ tween theory and experiment in particle physics, gravity and cosmology as
185
+ well as in other domains, such as condensed-matter physics, which were
186
+ not covered here. The technology for precision measurements in these do-
187
+ mains is constantly being upgraded, and in the coming years and decades
188
+ we are heading towards a growing influx of real-world data of higher qual-
189
+ ity. Future particle-collider projects, such as the Future Circular Collider
190
+ at CERN, or China’s Circular Electron Positron Collider, are being planned
191
+ and developed.
192
+ Precision cosmology is also thriving, with an upcoming
193
+ next-generation of very large surveys, such as the ground-based LSST, or
194
+ space-based Euclid.
195
+ GW detectors keep improving and multiplying, and
196
+ 5
197
+
198
+ besides those that are currently operating many more are planned, aimed
199
+ at measuring various frequency ranges, which will enable a richer array of
200
+ sources and events to be found.
201
+ Half a century after the concept has formally emerged, effective field
202
+ theory is still full of surprises. Recently, the physical space of EFTs has been
203
+ studied as a fundamental entity in its own right. These studies, by numerous
204
+ groups worldwide, have exposed a new hidden “totally positive” geometric
205
+ structure dubbed the EFT-hedron that constrains the EFT expansion in any
206
+ quantum field theory, and even string theory, from first principles, including
207
+ causality, unitarity and analyticity, to be satisfied by any amplitudes of
208
+ these theories. This recent formal progress reflects the ultimate leap in the
209
+ perception of EFT nowadays as the most fundamental and most generic
210
+ theory concept to capture the physics of nature at all scales. Clearly, in
211
+ the vast array of formidable open questions in physics that still lie ahead,
212
+ effective field theory is here to stay – for good.
213
+ Acknowledgements
214
+ We dedicate this article to the memory of Steven Weinberg, who so gen-
215
+ erously graced us with a spectacular inaugural lecture to the international
216
+ online series hosted at CERN “All Things EFT”, which turned out to be
217
+ his final published lecture.
218
+ We thank Cliff Burgess and HuaXing Zhu for comments and input on a
219
+ preliminary draft. ML has been supported by the Science and Technology
220
+ Facilities Council (STFC) Rutherford Grant ST/V003895 “Harnessing QFT
221
+ for Gravity”, and by the Mathematical Institute University of Oxford.
222
+ References
223
+ [1] S. Weinberg, Eur. Phys. J. H 46 (2021), 6 [arXiv:2101.04241 [hep-th]].
224
+ [2] S. Weinberg, Physica A 96 (1979), 327-340.
225
+ [3] C. P. Burgess, Les Houches100 (2015),148-197 [arXiv:1309.4133 [hep-th]].
226
+ [4] M. Levi, Rept. Prog. Phys. 83 (2020),075901 [arXiv:1807.01699 [hep-th]].
227
+ [5] I. Brivio and M. Trott, Phys. Rept. 793 (2019), 1-98 [arXiv:1706.08945
228
+ [hep-ph]].
229
+ [6] M. Tanabashi et al. [Particle Data Group], Phys. Rev. D 98 (2018),
230
+ 030001.
231
+ 6
232
+
MtE2T4oBgHgl3EQfqgg6/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf,len=113
2
+ page_content='arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
3
+ page_content='04039v1 [physics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
4
+ page_content='hist-ph] 10 Jan 2023 A Theory of Theories Mich`ele Levi Mathematical Institute, University of Oxford, Oxford OX2 6GG, United Kingdom levi@maths.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
5
+ page_content='ox.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
6
+ page_content='ac.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
7
+ page_content='uk Abstract We take a tour through the past, present and future of Effective Field Theory, with applications ranging from LHC physics to cosmol- ogy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
8
+ page_content=' 1 High-energy physics spans a wide range of energies, from a few MeV to TeV, that are all relevant.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
9
+ page_content=' It is therefore often difficult to take all phenomena into account at the same time.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
10
+ page_content=' Effective field theories (EFTs) are designed to break down this range of scales into smaller segments so that physicists can work in the relevant range.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
11
+ page_content=' Theorists “cut” their theory’s energy scale at the order of the mass of the lightest particle omitted from the theory, such as the proton mass.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
12
+ page_content=' Thus, multi-scale problems reduce to separate and single-scale problems.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
13
+ page_content='EFTs are today also understood to be “bottom- up” theories.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
14
+ page_content=' Built only out of the general field content and symmetries at the relevant scales, they allow us to test hypotheses efficiently and to select the most promising ones without needing to know the underlying theories in full detail.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
15
+ page_content=' Thanks to their applicability to all generic classical and quantum field theories, the sheer variety of EFT applications is striking.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
16
+ page_content=' In hindsight, particle physicists were working with EFTs from as early as Fermi’s phenomenological picture of beta decay in which a four-fermion vertex replaces the W-boson propagator because the momentum is much smaller compared to the mass of the W boson.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
17
+ page_content=' Like so many profound concepts in theoretical physics, EFT was first considered in a narrow phe- nomenological context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
18
+ page_content=' One of the earliest instances was in the 1960s, when ad-hoc methods of current algebras were utilised to study weak interactions of hadrons.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
19
+ page_content=' This required detailed calculations, and a simpler approach was needed to derive useful results.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
20
+ page_content=' The heuristic idea of describing hadron dynamics with the most general Lagrangian density based on symmetries, the relevant energy scale and the relevant particles, which can be written in terms of operators multiplied by Wilson coefficients, was yet to be known.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
21
+ page_content=' With this approach, it was possible to encode local symmetries in terms of the current algebra due to their association with conserved currents.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
22
+ page_content=' For strong interactions, physicists described the interaction between pi- ons with chiral perturbation theory, an effective Lagrangian, which sim- plified current algebra calculations and enabled the low-energy theory to be investigated systematically.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
23
+ page_content=' This “mother” of modern EFTs describes the physics of hadrons and remains valid to an energy scale of the proton mass.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
24
+ page_content=' Heavy-quark effective theory (HQET), introduced by Howard Georgi in 1990, complements chiral perturbation theory by describing the interac- tions of charm and bottom quarks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
25
+ page_content=' HQET allowed us to make predictions on B-meson decay rates, since the corrections could now be classified.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
26
+ page_content=' The more powers of energy are allowed, the more infinities appear.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
27
+ page_content=' These infinities are cancelled by available counter-terms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
28
+ page_content=' Similarly, it is possible to regard the Standard Model as the truncation of a much more general theory including non-renormalizable interactions, which yield corrections of higher order in energy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
29
+ page_content=' This perception of the whole Standard Model as an effective field theory started to be formed in the late 1970s by Weinberg and others.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
30
+ page_content=' Among the known corrections to the Standard Model that do not satisfy its approximate symmetries are 2 neutrino masses, postulated in the 1960s and discovered via the observation of neutrino oscillations in the late 1990s.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
31
+ page_content=' While the scope of EFTs was unclear initially, today we understand that all successful field theories, with which we have been working in many areas of theoretical physics, are nothing but effective field theories.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
32
+ page_content=' EFTs provide the theoretical framework to probe new physics and to establish precision programmes at experiments.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
33
+ page_content=' The former is crucial for making accurate theoretical predictions, while the latter is central to the physics programme of CERN in general.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
34
+ page_content=' 1 EFTs in Particle Physics More than a decade has passed since the first run of the LHC, in which the Higgs boson and the mechanism for electroweak symmetry breaking were discovered.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
35
+ page_content=' So far, there are no signals of new physics beyond the SM.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
36
+ page_content=' EFTs are well suited to explore LHC physics in depth.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
37
+ page_content=' A typical example for an event involving two scales is Higgs-boson production because there is a factor 10−100 between its mass and transverse momentum.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
38
+ page_content=' The calculation of each Higgs-boson production process leads to large logarithms that can invalidate perturbation theory due to the large-scale separation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
39
+ page_content=' This is just one of many examples of the two-scale problem that arises when the full quantum field theory approach for high-energy colliders is applied.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
40
+ page_content=' Traditionally, such two-scale problems have been treated in the framework of QCD factorisation and resummation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
41
+ page_content=' Over the past two decades, it has been possible to recast two-scale prob- lems at high-energy colliders with the advent of soft-collinear effective theory (SCET).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
42
+ page_content=' SCET is nowadays a popular framework that is used to describe Higgs physics, jets and their substructure, as well as more formal problems, such as power corrections to reconstruct full amplitudes eventually.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
43
+ page_content=' The difference between HQET and SCET is that SCET considers long-distance interactions between quarks and both soft and collinear particles, whereas HQET takes into account only soft interactions between a heavy quark and a parton.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
44
+ page_content=' SCET is just one example where the EFT methodology has been indispensable, even though the underlying theory at much higher energies is known.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
45
+ page_content=' Other examples of EFT applications include precision measure- ments of rare decays that can be described by QCD with its approximate chiral symmetry, or heavy quarks at finite temperature and density.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
46
+ page_content=' EFT is also central to a deeper understanding of the so-called flavour anomalies, enabling comparisons between theory and experiment in terms of particular Wilson coefficients.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
47
+ page_content=' Moreover, precision measurements of Higgs and electroweak observables at the LHC and future colliders will provide opportunities to detect new physics signals, such as resonances in invariant mass plots, or small devia- tions from the SM, seen in tails of distributions for instance at the HL-LHC 3 – testing the perception of the SM as a low-energy incarnation of a more fun- damental theory being probed at the electroweak scale.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
48
+ page_content=' This is dubbed the SMEFT (SM EFT) or HEFT (Higgs EFT), depending on whether the Higgs fields are expressed in terms of the Higgs doublet or the physical Higgs bo- son.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
49
+ page_content=' This particular EFT framework has recently been implemented in the data-analysis tools at the LHC, enabling the analyses across different chan- nels and even different experiments.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
50
+ page_content='At the same time, the study of SMEFT and HEFT has sparked a plethora of theoretical investigations that have uncovered its remarkable underlying features, for example allowing EFT to be extended or placing constraints on the EFT coefficients due to Lorentz invariance, causality and analyticity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
51
+ page_content=' 2 EFTs in Gravity Since the inception of EFT, it was believed that the framework is applicable only to the description of quantum field theories for capturing the physics of elementary particles at high-energy scales, or alternatively at very small length scales.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
52
+ page_content=' Thus, EFT seemed mostly irrelevant regarding gravitation, for which we are still lacking a full theory valid at quantum scales.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
53
+ page_content=' The only way in which EFT seemed to be pertinent for gravitation was to think of general relativity as a first approximation to an EFT description of quantum gravity, which indeed provided a new EFT perspective at the time.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
54
+ page_content=' However, in the past decade it has become widely acknowledged that EFT provides a powerful framework to capture gravitation occurring completely across large length scales, as long as these scales display a clear hierarchy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
55
+ page_content=' The most notable application to such classical gravitational systems came when it was realised that the EFT framework would be ideal to handle gravitational radiation emitted at the inspiral phase of a binary of compact objects, such as black holes.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
56
+ page_content=' At this phase in the evolution of the binary, the compact objects are moving at non-relativistic velocities.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
57
+ page_content=' Using the small velocity as the expansion parameter exhibits the separation between the various characteristic length scales of the system.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
58
+ page_content=' Thus, the physics can be treated perturbatively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
59
+ page_content=' For example, it was found that even couplings man- ifestly change in classical systems across their characteristic scales, which was previously believed to be unique to quantum field theories.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
60
+ page_content=' The appli- cation of EFT to the binary inspiral problem has been so successful that the precision frontier has been pushed beyond the state of the art, quickly surpassing the reach of work that has been focused on the two-body problem for decades via traditional methods in general relativity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
61
+ page_content=' This theoretical progress has made an even broader impact since the breakthrough direct discovery of gravitational waves (GWs) was announced in 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
62
+ page_content=' An inspiraling binary of black holes merged into a single black hole in less than a split second, releasing an enormous amount of energy in the 4 form of GWs, which instigated even greater, more intense use of EFTs for the generation of theoretical GW data.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
63
+ page_content=' In the coming years and decades, a continuous increase in the quantity and quality of real-world GW data is expected from the rapidly growing worldwide network of ground-based GW detectors, and future space-based interferometers, covering a wide range of target frequencies.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
64
+ page_content=' 3 EFTs in Cosmology Cosmology is inherently a cross-cutting domain, spanning scales over about 1060 orders of magnitude, from the Planck scale to the size of the observable universe.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
65
+ page_content=' As such, cosmology generally cannot be expected to be tackled directly by each of the fundamental theories that capture particle physics or gravity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
66
+ page_content=' The correct description of cosmology relies heavily on the work in many disparate areas of research in theoretical and experimental physics, including particle physics and general relativity among many more.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
67
+ page_content=' The development of EFT applications in cosmology – including EFTs of inflation, dark matter, dark energy and even EFTs of large-scale structure – has become essential to make observable predictions in cosmology.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
68
+ page_content=' The discovery of the accelerated expansion of the universe in 1998 shows our diffi- culty in understanding gravity both at the quantum regime and the classical one.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
69
+ page_content=' The cosmological constant problem and dark-matter paradigm might be a hint for alternative theories of gravity at very large scales.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
70
+ page_content=' Indeed, the problems with gravity in the very-high and very-low energy range may well be tied together.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
71
+ page_content=' The science programme of next-generation large surveys, such as ESA’s Euclid satellite, rely heavily on all these EFT applications for the exploitation of the enormous data that is going to be collected to constrain unknown cosmological parameters, thus helping to pinpoint viable theories.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
72
+ page_content=' 4 The Future of EFTs in Physics The EFT framework plays a key role at the exciting and rich interface be- tween theory and experiment in particle physics, gravity and cosmology as well as in other domains, such as condensed-matter physics, which were not covered here.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
73
+ page_content=' The technology for precision measurements in these do- mains is constantly being upgraded, and in the coming years and decades we are heading towards a growing influx of real-world data of higher qual- ity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
74
+ page_content=' Future particle-collider projects, such as the Future Circular Collider at CERN, or China’s Circular Electron Positron Collider, are being planned and developed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
75
+ page_content=' Precision cosmology is also thriving, with an upcoming next-generation of very large surveys, such as the ground-based LSST, or space-based Euclid.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
76
+ page_content=' GW detectors keep improving and multiplying, and 5 besides those that are currently operating many more are planned, aimed at measuring various frequency ranges, which will enable a richer array of sources and events to be found.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
77
+ page_content=' Half a century after the concept has formally emerged, effective field theory is still full of surprises.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
78
+ page_content=' Recently, the physical space of EFTs has been studied as a fundamental entity in its own right.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
79
+ page_content=' These studies, by numerous groups worldwide, have exposed a new hidden “totally positive” geometric structure dubbed the EFT-hedron that constrains the EFT expansion in any quantum field theory, and even string theory, from first principles, including causality, unitarity and analyticity, to be satisfied by any amplitudes of these theories.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
80
+ page_content=' This recent formal progress reflects the ultimate leap in the perception of EFT nowadays as the most fundamental and most generic theory concept to capture the physics of nature at all scales.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
81
+ page_content=' Clearly, in the vast array of formidable open questions in physics that still lie ahead, effective field theory is here to stay – for good.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
82
+ page_content=' Acknowledgements We dedicate this article to the memory of Steven Weinberg, who so gen- erously graced us with a spectacular inaugural lecture to the international online series hosted at CERN “All Things EFT”, which turned out to be his final published lecture.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
83
+ page_content=' We thank Cliff Burgess and HuaXing Zhu for comments and input on a preliminary draft.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
84
+ page_content=' ML has been supported by the Science and Technology Facilities Council (STFC) Rutherford Grant ST/V003895 “Harnessing QFT for Gravity”, and by the Mathematical Institute University of Oxford.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
85
+ page_content=' References [1] S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
86
+ page_content=' Weinberg, Eur.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
87
+ page_content=' Phys.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
88
+ page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
89
+ page_content=' H 46 (2021), 6 [arXiv:2101.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
90
+ page_content='04241 [hep-th]].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
91
+ page_content=' [2] S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
92
+ page_content=' Weinberg, Physica A 96 (1979), 327-340.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
93
+ page_content=' [3] C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
94
+ page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
95
+ page_content=' Burgess, Les Houches100 (2015),148-197 [arXiv:1309.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
96
+ page_content='4133 [hep-th]].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
97
+ page_content=' [4] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
98
+ page_content=' Levi, Rept.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
99
+ page_content=' Prog.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
100
+ page_content=' Phys.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
101
+ page_content=' 83 (2020),075901 [arXiv:1807.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
102
+ page_content='01699 [hep-th]].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
103
+ page_content=' [5] I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
104
+ page_content=' Brivio and M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
105
+ page_content=' Trott, Phys.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
106
+ page_content=' Rept.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
107
+ page_content=' 793 (2019), 1-98 [arXiv:1706.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
108
+ page_content='08945 [hep-ph]].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
109
+ page_content=' [6] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
110
+ page_content=' Tanabashi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
111
+ page_content=' [Particle Data Group], Phys.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
112
+ page_content=' Rev.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
113
+ page_content=' D 98 (2018), 030001.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
114
+ page_content=' 6' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/MtE2T4oBgHgl3EQfqgg6/content/2301.04039v1.pdf'}
N9AzT4oBgHgl3EQfk_2S/content/2301.01541v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c48978a2502d9bb140a2cf3e233c29feea59428f02aaa56923f4b0ba49df19a3
3
+ size 1211473