jackkuo commited on
Commit
88700d5
·
verified ·
1 Parent(s): a9d2e92

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. -9FQT4oBgHgl3EQf7Ta5/content/2301.13442v1.pdf +3 -0
  2. -9FQT4oBgHgl3EQf7Ta5/vector_store/index.pkl +3 -0
  3. -NFLT4oBgHgl3EQfuy_h/content/2301.12157v1.pdf +3 -0
  4. -NFLT4oBgHgl3EQfuy_h/vector_store/index.faiss +3 -0
  5. -NFLT4oBgHgl3EQfuy_h/vector_store/index.pkl +3 -0
  6. .gitattributes +61 -0
  7. 09E3T4oBgHgl3EQfnArs/content/2301.04622v1.pdf +3 -0
  8. 09E3T4oBgHgl3EQfnArs/vector_store/index.pkl +3 -0
  9. 09FAT4oBgHgl3EQfChxb/content/tmp_files/2301.08410v1.pdf.txt +0 -0
  10. 09FAT4oBgHgl3EQfChxb/content/tmp_files/load_file.txt +0 -0
  11. 0NAyT4oBgHgl3EQfbfc0/vector_store/index.pkl +3 -0
  12. 0NE2T4oBgHgl3EQf4gjD/content/2301.04182v1.pdf +3 -0
  13. 0NE2T4oBgHgl3EQf4gjD/vector_store/index.faiss +3 -0
  14. 29E0T4oBgHgl3EQfuwF7/content/2301.02609v1.pdf +3 -0
  15. 29E0T4oBgHgl3EQfuwF7/vector_store/index.faiss +3 -0
  16. 29E0T4oBgHgl3EQfuwF7/vector_store/index.pkl +3 -0
  17. 3NE2T4oBgHgl3EQfjQer/content/tmp_files/2301.03967v1.pdf.txt +1984 -0
  18. 3NE2T4oBgHgl3EQfjQer/content/tmp_files/load_file.txt +0 -0
  19. 3tE2T4oBgHgl3EQf6AiT/vector_store/index.faiss +3 -0
  20. 49E0T4oBgHgl3EQfvgE1/vector_store/index.pkl +3 -0
  21. 49E1T4oBgHgl3EQfAwK8/vector_store/index.faiss +3 -0
  22. 59E0T4oBgHgl3EQfewCK/content/2301.02395v1.pdf +3 -0
  23. 59E0T4oBgHgl3EQfewCK/vector_store/index.pkl +3 -0
  24. 69AyT4oBgHgl3EQfQfbA/content/tmp_files/2301.00048v1.pdf.txt +1233 -0
  25. 69AyT4oBgHgl3EQfQfbA/content/tmp_files/load_file.txt +0 -0
  26. 6NFJT4oBgHgl3EQflSzb/content/tmp_files/2301.11583v1.pdf.txt +1151 -0
  27. 6NFJT4oBgHgl3EQflSzb/content/tmp_files/load_file.txt +0 -0
  28. 79E0T4oBgHgl3EQfwQGR/content/tmp_files/2301.02630v1.pdf.txt +1079 -0
  29. 79E0T4oBgHgl3EQfwQGR/content/tmp_files/load_file.txt +0 -0
  30. 89E4T4oBgHgl3EQfDAsR/vector_store/index.faiss +3 -0
  31. 89E4T4oBgHgl3EQfDAsR/vector_store/index.pkl +3 -0
  32. 99AyT4oBgHgl3EQfRPYW/vector_store/index.pkl +3 -0
  33. 9NFRT4oBgHgl3EQfqTc6/content/2301.13616v1.pdf +3 -0
  34. 9NFRT4oBgHgl3EQfqTc6/vector_store/index.pkl +3 -0
  35. 9dE1T4oBgHgl3EQfUQOY/content/tmp_files/2301.03088v1.pdf.txt +0 -0
  36. 9dE1T4oBgHgl3EQfUQOY/content/tmp_files/load_file.txt +0 -0
  37. ANE1T4oBgHgl3EQf9AZG/content/tmp_files/2301.03551v1.pdf.txt +1727 -0
  38. ANE1T4oBgHgl3EQf9AZG/content/tmp_files/load_file.txt +0 -0
  39. AtE3T4oBgHgl3EQfTgrD/content/tmp_files/2301.04443v1.pdf.txt +897 -0
  40. AtE3T4oBgHgl3EQfTgrD/content/tmp_files/load_file.txt +0 -0
  41. BNAzT4oBgHgl3EQfv_6i/content/tmp_files/2301.01716v1.pdf.txt +3093 -0
  42. BNAzT4oBgHgl3EQfv_6i/content/tmp_files/load_file.txt +0 -0
  43. BdFIT4oBgHgl3EQf_ywO/content/2301.11416v1.pdf +3 -0
  44. CdAzT4oBgHgl3EQfGftE/vector_store/index.pkl +3 -0
  45. CdAzT4oBgHgl3EQfTvxb/content/tmp_files/2301.01254v1.pdf.txt +694 -0
  46. CdAzT4oBgHgl3EQfTvxb/content/tmp_files/load_file.txt +0 -0
  47. EdFRT4oBgHgl3EQfyTjG/content/2301.13645v1.pdf +3 -0
  48. EdFRT4oBgHgl3EQfyTjG/vector_store/index.faiss +3 -0
  49. EdFRT4oBgHgl3EQfyTjG/vector_store/index.pkl +3 -0
  50. FtA0T4oBgHgl3EQfBf8H/content/tmp_files/2301.01975v1.pdf.txt +2941 -0
-9FQT4oBgHgl3EQf7Ta5/content/2301.13442v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ee59292193eb91ec024c7846fd08204bda4ce6055ded9d893b249fcf4f23ea7
3
+ size 2143686
-9FQT4oBgHgl3EQf7Ta5/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7f1281296c04bee27a7f612a0b3110232133b3bcdb0ebd9e47ce5b9590d3ac05
3
+ size 259515
-NFLT4oBgHgl3EQfuy_h/content/2301.12157v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2175378c645913ab0fce1c17d65b3a00b094dacde36567d820148d90fafdfcf3
3
+ size 350948
-NFLT4oBgHgl3EQfuy_h/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b87f4082d8ece62454fcc23c4a3e6d59ee5374e203ea9a7f2b95b779a37929cb
3
+ size 1638445
-NFLT4oBgHgl3EQfuy_h/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:097cece77585a2cb3550a92aa7b6b3295a799e2afa85f788dd1a328b4ea5a966
3
+ size 63192
.gitattributes CHANGED
@@ -7379,3 +7379,64 @@ sdAyT4oBgHgl3EQf0Pk5/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -tex
7379
  hdE2T4oBgHgl3EQfHgaB/content/2301.03669v1.pdf filter=lfs diff=lfs merge=lfs -text
7380
  XdE0T4oBgHgl3EQf3gIm/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7381
  dNAzT4oBgHgl3EQfLvuL/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7379
  hdE2T4oBgHgl3EQfHgaB/content/2301.03669v1.pdf filter=lfs diff=lfs merge=lfs -text
7380
  XdE0T4oBgHgl3EQf3gIm/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7381
  dNAzT4oBgHgl3EQfLvuL/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7382
+ yNE4T4oBgHgl3EQfYQyw/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7383
+ wNAyT4oBgHgl3EQfavcI/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7384
+ R9E0T4oBgHgl3EQf1wJf/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7385
+ sdFJT4oBgHgl3EQfcSx2/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7386
+ dtFST4oBgHgl3EQfFDgM/content/2301.13716v1.pdf filter=lfs diff=lfs merge=lfs -text
7387
+ 0NE2T4oBgHgl3EQf4gjD/content/2301.04182v1.pdf filter=lfs diff=lfs merge=lfs -text
7388
+ P9E1T4oBgHgl3EQfagRh/content/2301.03162v1.pdf filter=lfs diff=lfs merge=lfs -text
7389
+ 0NE2T4oBgHgl3EQf4gjD/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7390
+ QNFJT4oBgHgl3EQf2i1I/content/2301.11656v1.pdf filter=lfs diff=lfs merge=lfs -text
7391
+ ZtFJT4oBgHgl3EQf7i35/content/2301.11679v1.pdf filter=lfs diff=lfs merge=lfs -text
7392
+ 3tE2T4oBgHgl3EQf6AiT/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7393
+ R9E0T4oBgHgl3EQf1wJf/content/2301.02703v1.pdf filter=lfs diff=lfs merge=lfs -text
7394
+ P9E1T4oBgHgl3EQfagRh/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7395
+ XdAyT4oBgHgl3EQfvPnb/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7396
+ wdFLT4oBgHgl3EQflC8K/content/2301.12117v1.pdf filter=lfs diff=lfs merge=lfs -text
7397
+ dNAyT4oBgHgl3EQf-fq3/content/2301.00894v1.pdf filter=lfs diff=lfs merge=lfs -text
7398
+ m9FQT4oBgHgl3EQfpTaJ/content/2301.13376v1.pdf filter=lfs diff=lfs merge=lfs -text
7399
+ ydFJT4oBgHgl3EQfhyyK/content/2301.11567v1.pdf filter=lfs diff=lfs merge=lfs -text
7400
+ QNFJT4oBgHgl3EQf2i1I/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7401
+ 49E1T4oBgHgl3EQfAwK8/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7402
+ htFPT4oBgHgl3EQfDzQQ/content/2301.12993v1.pdf filter=lfs diff=lfs merge=lfs -text
7403
+ EdFRT4oBgHgl3EQfyTjG/content/2301.13645v1.pdf filter=lfs diff=lfs merge=lfs -text
7404
+ s9E4T4oBgHgl3EQfVwyP/content/2301.05027v1.pdf filter=lfs diff=lfs merge=lfs -text
7405
+ bdFJT4oBgHgl3EQf9C1W/content/2301.11686v1.pdf filter=lfs diff=lfs merge=lfs -text
7406
+ m9FQT4oBgHgl3EQfpTaJ/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7407
+ _tFJT4oBgHgl3EQfqizo/content/2301.11605v1.pdf filter=lfs diff=lfs merge=lfs -text
7408
+ wNA0T4oBgHgl3EQfMP-b/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7409
+ HdE2T4oBgHgl3EQfTwd_/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7410
+ rNE3T4oBgHgl3EQf8gvv/content/2301.04809v1.pdf filter=lfs diff=lfs merge=lfs -text
7411
+ -9FQT4oBgHgl3EQf7Ta5/content/2301.13442v1.pdf filter=lfs diff=lfs merge=lfs -text
7412
+ dNAyT4oBgHgl3EQf-fq3/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7413
+ SNFKT4oBgHgl3EQfjS7T/content/2301.11845v1.pdf filter=lfs diff=lfs merge=lfs -text
7414
+ gtAzT4oBgHgl3EQfov1j/content/2301.01601v1.pdf filter=lfs diff=lfs merge=lfs -text
7415
+ EdFRT4oBgHgl3EQfyTjG/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7416
+ edFIT4oBgHgl3EQfpCt4/content/2301.11321v1.pdf filter=lfs diff=lfs merge=lfs -text
7417
+ xtAyT4oBgHgl3EQfa_c9/content/2301.00252v1.pdf filter=lfs diff=lfs merge=lfs -text
7418
+ h9AzT4oBgHgl3EQfo_2T/content/2301.01606v1.pdf filter=lfs diff=lfs merge=lfs -text
7419
+ BdFIT4oBgHgl3EQf_ywO/content/2301.11416v1.pdf filter=lfs diff=lfs merge=lfs -text
7420
+ -NFLT4oBgHgl3EQfuy_h/content/2301.12157v1.pdf filter=lfs diff=lfs merge=lfs -text
7421
+ Y9AzT4oBgHgl3EQfmv3V/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7422
+ vdAyT4oBgHgl3EQfnPg_/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7423
+ H9E3T4oBgHgl3EQfWwp_/content/2301.04472v1.pdf filter=lfs diff=lfs merge=lfs -text
7424
+ s9E3T4oBgHgl3EQf9gsi/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7425
+ 09E3T4oBgHgl3EQfnArs/content/2301.04622v1.pdf filter=lfs diff=lfs merge=lfs -text
7426
+ ktE2T4oBgHgl3EQfdwcD/content/2301.03908v1.pdf filter=lfs diff=lfs merge=lfs -text
7427
+ H9E3T4oBgHgl3EQfWwp_/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7428
+ -NFLT4oBgHgl3EQfuy_h/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7429
+ c9AzT4oBgHgl3EQf3P6k/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7430
+ u9E3T4oBgHgl3EQfkgqO/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7431
+ dNAzT4oBgHgl3EQfLvuL/content/2301.01120v1.pdf filter=lfs diff=lfs merge=lfs -text
7432
+ 29E0T4oBgHgl3EQfuwF7/content/2301.02609v1.pdf filter=lfs diff=lfs merge=lfs -text
7433
+ edFIT4oBgHgl3EQfpCt4/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7434
+ UdE3T4oBgHgl3EQfEgk6/content/2301.04296v1.pdf filter=lfs diff=lfs merge=lfs -text
7435
+ 89E4T4oBgHgl3EQfDAsR/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7436
+ 59E0T4oBgHgl3EQfewCK/content/2301.02395v1.pdf filter=lfs diff=lfs merge=lfs -text
7437
+ 29E0T4oBgHgl3EQfuwF7/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7438
+ bdFJT4oBgHgl3EQf9C1W/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7439
+ StE2T4oBgHgl3EQfCAbz/content/2301.03610v1.pdf filter=lfs diff=lfs merge=lfs -text
7440
+ OdFPT4oBgHgl3EQfmzXU/content/2301.13127v1.pdf filter=lfs diff=lfs merge=lfs -text
7441
+ sdAyT4oBgHgl3EQf0Pk5/content/2301.00713v1.pdf filter=lfs diff=lfs merge=lfs -text
7442
+ 9NFRT4oBgHgl3EQfqTc6/content/2301.13616v1.pdf filter=lfs diff=lfs merge=lfs -text
09E3T4oBgHgl3EQfnArs/content/2301.04622v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:09f6928868f17248a68392f32776b1d9772ac7013e40688d7fe5f9ed9bba801c
3
+ size 5413203
09E3T4oBgHgl3EQfnArs/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7958dea947bff8e701a3897fe0ad6e07dffc733f9e019f1cc46194b8b60ac99f
3
+ size 122818
09FAT4oBgHgl3EQfChxb/content/tmp_files/2301.08410v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
09FAT4oBgHgl3EQfChxb/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
0NAyT4oBgHgl3EQfbfc0/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:207f40c9b5aac4792de6086facfe723156eb5355e36e5cd6442a40e4b99f968a
3
+ size 196890
0NE2T4oBgHgl3EQf4gjD/content/2301.04182v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:964044bcd017ec31beb7ad9482d41c6604b945648e688c7332be9ebddf06cc12
3
+ size 903226
0NE2T4oBgHgl3EQf4gjD/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f6a6a62949d1b61db417000f27aa822619163c7fe05889df24c797f7e2eb615a
3
+ size 3145773
29E0T4oBgHgl3EQfuwF7/content/2301.02609v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05f4fa3387649f0491f1ec37ffb117e1c39d7b1ade26471746e47c3ab9ddc58d
3
+ size 676520
29E0T4oBgHgl3EQfuwF7/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fbbab9e5433198fc603d3de2e4970e86c5c973c224b753477b325750443b95b9
3
+ size 2424877
29E0T4oBgHgl3EQfuwF7/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:30ba7babb9aa922201ad3ddc8cac75a0c7b2c87033eba3cd134fb17cb716f79e
3
+ size 83982
3NE2T4oBgHgl3EQfjQer/content/tmp_files/2301.03967v1.pdf.txt ADDED
@@ -0,0 +1,1984 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Subgap states and quantum phase transitions in one-dimensional
2
+ superconductor-ferromagnetic insulator heterostructures
3
+ Javier Feijoo,1, 2 An´ıbal Iucci,1, 2 and Alejandro M. Lobos3, 4
4
+ 1Instituto de F´ısica La Plata - CONICET, Diag 113 y 64 (1900) La Plata, Argentina
5
+ 2Departamento de F´ısica, Universidad Nacional de La Plata, cc 67, 1900 La Plata, Argentina.
6
+ 3Instituto Interdisciplinario de Ciencias B´asicas (CONICET-UNCuyo)
7
+ 4Facultad de Ciencias Exactas y Naturales, Universidad Nacional de Cuyo, 5500 Mendoza, Argentina
8
+ We
9
+ theoretically
10
+ study
11
+ the
12
+ spectral
13
+ properties
14
+ of
15
+ a
16
+ one
17
+ dimensional
18
+ semiconductor-
19
+ superconductor-ferromagnetic insulator (SE-SU-FMI) hybrid nanostructure, motivated by recents
20
+ experiments where such devices have been fabricated using epitaxial growing techniques. We model
21
+ the hybrid structure as a one-dimensional single-channel semiconductor nanowire under the si-
22
+ multaneous effect of two proximity-induced interactions: superconducting pairing and a (spatially
23
+ inhomogeneous) Zeeman exchange field. The coexistence of these competing interactions generates
24
+ a rich quantum phase diagram and a complex subgap Andreev bound state (ABS) spectrum. By
25
+ exploiting the symmetries of the problem, we classify the solutions of the Bogoliubov-de Gennes
26
+ equations into even and odd ABS with respect to the spatial inversion symmetry x → −x. We
27
+ find the ABS spectrum of the device as a function of the different parameters of the model: the
28
+ length L of the coexisting SU-FMI region, the induced Zeeman exchange field h0, and the induced
29
+ superconducting coherence length ξ. In particular we analyze the evolution of the subgap spectrum
30
+ as a function of the length L. Interestingly, we have found that depending on the ratio h0/∆, the
31
+ emerging ABS can eventually cross below the Fermi energy at certain critical values Lc, and induce
32
+ spin-and fermion parity-changing quantum phase transitions. We argue that this type of device
33
+ constitute a promising highly-tunable platform to engineer subgap ABS.
34
+ I.
35
+ INTRODUCTION
36
+ The interplay of superconductivity and magnetism at
37
+ the microscopic scale has attracted a great deal of at-
38
+ tention in recent years [1–4].
39
+ For instance, the Yu-
40
+ Shiba-Rusinov (YSR) states [5–7] arising from the ex-
41
+ change interaction of an atomic magnetic moment in con-
42
+ tact with a superconductor, have been proposed as fun-
43
+ damental building blocks to engineer quantum devices
44
+ with topologically non-trivial ground states. In partic-
45
+ ular, the so-called “Shiba chains” (i.e., one-dimensional
46
+ arrays of magnetic atoms deposited on top of a clean
47
+ superconductor) are systems predicted to support Ma-
48
+ jorana zero-modes at the ends of the chain [8–10], and
49
+ could be used in topologically-protected quantum com-
50
+ putation schemes. Low-temperature scanning-tunneling
51
+ microscopy (STM) experiments have confirmed the pres-
52
+ ence of intruiguing zero-energy end-modes [11–17].
53
+ Other systems where the competition of superconduc-
54
+ tivity and magnetism at the nanoscale generates ex-
55
+ otic subgap states are superconductor (SU)- ferromag-
56
+ net (FM) heterostructures, such as SU-FM-SU Josephson
57
+ junctions and SU-FM proximity devices [18, 19]. Subgap
58
+ states generated in these structures are usually referred
59
+ to as Andreev bound states (ABS). More recently, a novel
60
+ class of hybrid device, i.e., semiconductor (SE) nanowire
61
+ systems combined with superconductors and ferromag-
62
+ netic insulator (FMI) materials have been fabricated us-
63
+ ing molecular-beam epitaxy techniques [20, 21]. These
64
+ SE-SU-FMI hybrid structures allow to build nanostruc-
65
+ tures with specific tailored properties which are impossi-
66
+ ble to obtain with the isolated individual components.
67
+ Despite the evident differences between the abovemen-
68
+ x
69
+ z
70
+ SC Bulk
71
+ FMI
72
+ Semiconductor
73
+ L
74
+ x
75
+ z
76
+ L
77
+ h0
78
+ Magnetic profile
79
+ FIG. 1. Schematic representation of the SC-FMI heterostruc-
80
+ ture.
81
+ tioned physical systems, from the theoretical perspective
82
+ they can be described within the same unified theoretical
83
+ model combining superconductivity and local exchange
84
+ fields at the microscopic scale.
85
+ The emerging subgap
86
+ states (which can be referred to as either YSR states or
87
+ ABS, depending on the context) appear symmetrically
88
+ around the Fermi level EF , and localize spatially around
89
+ the impurity or the FM region.
90
+ Their energy-position
91
+ within the gap depend on the value of the exchange
92
+ field and on other experimental parameters.
93
+ Interest-
94
+ ingly, whenever one of these states crosses EF , a spin-
95
+ and parity-changing quantum phase transition, usually
96
+ arXiv:2301.03967v1 [cond-mat.supr-con] 10 Jan 2023
97
+
98
+ 2
99
+ known as the “0 − π” phase transition, occurs [1, 22].
100
+ In the case of atomic “Shiba impurities” or ultra-short
101
+ SU-FM-SU junctions (i.e., junctions in which the length
102
+ L of the FM region is much smaller than λF , the Fermi
103
+ wavelength of the superconductor [23]), it is customary
104
+ to consider the magnetic scatterer as a point-like classical
105
+ spin S located at the point R0, interacting via a contact
106
+ s-d exchange interaction HZ = J(r) S·s(r) with the host
107
+ superconducting electrons [6]. Here J(r) = J0δ(r − R0)
108
+ is the local exchange potential and s(r) is the spin den-
109
+ sity vector of the electronic fluid. Subsequent theoretical
110
+ works considered atomic-sized systems with finite- (al-
111
+ beit short-ranged) exchange interactions with spherical
112
+ symmetry [7, 24–26]. In that case, theory predicts the
113
+ existence of multiple YSR states labelled by their orbital
114
+ momentum ℓ, a prediction that has been recently ob-
115
+ served in STM experiments [27–29].
116
+ The behavior of subgap states and the associated 0−π
117
+ quantum phase transitions has also been studied in the
118
+ opposite limit L ≫ λF in the context of ballistic SU-
119
+ FM-SU Josephson junctions with generic spin-dependent
120
+ fields in the sandwiched region [30–32]. In this case the
121
+ results differ from the well-known results of YSR states
122
+ due to the finite extension of the magnetic profile. In
123
+ particular, the subgap spectrum of long SU-FM-SU junc-
124
+ tions with zero phase difference is known to be double
125
+ degenerate [19, 31], showing the inherent complexity of
126
+ these hybrid heterostructures. On the experimental side,
127
+ the possibility to engineer and control the position of the
128
+ subgap states by a modification of the fabrication para-
129
+ maters (e.g., the length L or exchange field h0 via dif-
130
+ ferent FM materials) opens interesting perspectives for
131
+ potential electronic devices, where the precise knowledge
132
+ of the subgap spectrum is crucial to control their trans-
133
+ port properties.
134
+ Motivated by the experimental developments men-
135
+ tioned above, in this work we study the subgap states
136
+ emerging in one-dimensional (1D) SE-SU-FMI het-
137
+ erostructures where the SU and the FMI layers simul-
138
+ taneously generate coexisting proximity-induced pairing
139
+ and exchange interactions over a finite and arbitrary
140
+ length L in the SE nanowire, as schematically shown in
141
+ Fig. 1. This coexistence is a crucial aspect of this device,
142
+ which makes it unique and different from the abovemen-
143
+ tioned SU-FM-SU junctions, where such overlap occurs
144
+ only at the SU-FM interface. Our main goal in this work
145
+ is to study and understand the behavior of the subgap
146
+ ABS in this device as a function of the experimentally rel-
147
+ evant parameters of the model, i.e., the length L of the
148
+ FMI region and the magnitude of the induced exchange
149
+ field h0. As mentioned above, a device similar to that
150
+ shown in Fig. 1 has been recently experimentally real-
151
+ ized in SE nanowires with epitaxially-grown SU and FMI
152
+ layers [20, 21]. While the main interest of that work was
153
+ the fabrication of a device with non-trivial topological SU
154
+ ground state hosting Majorana zero modes, here we will
155
+ study the regime of parameters favoring a topologically-
156
+ trivial ground state.
157
+ As we will show below (see Sec.
158
+ II), this case is already very complex and rich as a result
159
+ of the antagonistic SU and FM interactions and, to the
160
+ best of our knowledge, the detailed behavior of subgap
161
+ states and the quantum phase diagram emerging in such
162
+ a system have not been explicitly studied before.
163
+ The article is organized as follows. In Section II, we
164
+ introduce the model representing a 1D SE-SU-FMI hy-
165
+ brid nanowire, discuss the solution to the Bogoliubov-
166
+ de Gennes equations for the subgap states, and derive
167
+ a generic equation for the subgap spectrum.
168
+ In Sec-
169
+ tion III, we analyze the results in two specific limits,
170
+ where we recover well-known results: a) the semiclassical
171
+ limit, where the superconducting coherence length ξ is
172
+ much larger than the Fermi wavelength λF , and b) the
173
+ atomic YSR limit, in which the exchange-field induced
174
+ by the FMI region becomes a delta-function potential:
175
+ i.e., infinitesimally narrow (L ≪ λF ), and infinitely deep
176
+ (h0 ≫ EF ), in such a way that the product h0.L = J
177
+ is kept constant.
178
+ In both cases, well-known analytical
179
+ solutions to the subgap spectrum can be recovered. In
180
+ addition, we numerically solve the characteristic equation
181
+ for the subgap states and provide a generic description
182
+ of the subgap spectrum, not restricted to any of these
183
+ limits.
184
+ We find a rich behaviour of the subgap ABS,
185
+ where the competing FM exchange and SU pairing inter-
186
+ actions give rise to parity- and spin-changing quantum
187
+ phase transitions. Finally, in Section IV, we present a
188
+ summary and our conclusions.
189
+ II.
190
+ THEORETICAL MODEL
191
+ We focus on the system schematically depicted in Fig.
192
+ 1, which represents a 1D SE-SU-FMI hybrid nanostruc-
193
+ ture of total length Lw, similar to those fabricated in
194
+ Refs. 20 and 21. We model this system with the Hamil-
195
+ tonian H = Hw + H∆ + HZ, where
196
+ Hw =
197
+
198
+ σ
199
+
200
+ Lw
201
+ 2
202
+ − Lw
203
+ 2
204
+ dx ψ†
205
+ σ(x)
206
+
207
+ −ℏ2∂2
208
+ x
209
+ 2m∗ − µ
210
+
211
+ ψ†
212
+ σ(x),
213
+ (1)
214
+ H ∆ = ∆
215
+
216
+ Lw
217
+ 2
218
+ − Lw
219
+ 2
220
+ dx
221
+
222
+ ψ†
223
+ ↑(x)ψ†
224
+ ↓(x) + H.c.
225
+
226
+ ,
227
+ (2)
228
+ HZ =
229
+
230
+ Lw
231
+ 2
232
+ − Lw
233
+ 2
234
+ dx h(x)
235
+
236
+ ψ†
237
+ ↑(x)ψ↑(x) − ψ†
238
+ ↓(x)ψ↓(x)
239
+
240
+ . (3)
241
+ Here Hw is the Hamiltonian of a single-channel SE
242
+ nanowire of length Lw, in which the fermionic operator
243
+ ψσ(x) creates an electron at position x with spin projec-
244
+ tion σ =↑, ↓ and effective mass m∗. The parameter µ is
245
+ the chemical potential, which can be experimentally var-
246
+ ied applying external gates beneath the nanostructure.
247
+ The terms H∆ and HZ represent, respectively, the
248
+ proximity-induced pairing interaction encoded by the pa-
249
+ rameter ∆, and the Zeeman exchange interaction intro-
250
+ duced by the FMI and described by a space-dependent
251
+ exchange field h(x), which we assume oriented along the
252
+
253
+ 3
254
+ z direction (see Fig. 1). Moreover, since these interac-
255
+ tions are externally induced into the semiconductor, we
256
+ make the additional assumption that ∆ is unaffected by
257
+ the presence of h(x) (a renormalized value of ∆ does not
258
+ change qualitatively our results). As mentioned before,
259
+ these two terms can be effectively induced by the pres-
260
+ ence of epitaxially-grown SU and FMI shells in contact
261
+ with the SE nanowire [20, 21]. It has been experimen-
262
+ tally confirmed [21] that the FMI shell (EuS in that case)
263
+ consists of a single magnetic monodomain, and there-
264
+ fore modelling this layer by the Hamiltonian HZ is a
265
+ reasonable approximation. In addition, the epitaxially-
266
+ generated interfaces are essentially disorder-free, a neces-
267
+ sary condition to produce a proximity-induced hard-gap
268
+ [33]. This feature allows to neglect the effects of disorder
269
+ and considerably simplifies the theoretical description.
270
+ The presence of both, a hard proximity-induced super-
271
+ conductor gap and an effectively induced Zeeman field,
272
+ in these nanowires have been reported in transport mea-
273
+ surements in Refs.
274
+ 20 and 21.
275
+ In addition, note that
276
+ in the above model we have neglected the effect of the
277
+ Rashba spin-orbit interaction. While this interaction is
278
+ crucial for the emergence of a topologically non-trivial
279
+ (i.e., D class) superconducting phase supporting Majo-
280
+ rana zero-modes [34], here we will focus strictly on the
281
+ topologically-trivial ground state. As we will show be-
282
+ low, the competition of SU and FM interactions make
283
+ this system already very complex and interesting in it-
284
+ self.
285
+ We note that since the total single-particle fermionic
286
+ spin along z
287
+ sz = 1
288
+ 2
289
+
290
+ Lw
291
+ 2
292
+ − Lw
293
+ 2
294
+ dx
295
+
296
+ ψ†
297
+ ↑(x)ψ↑(x) − ψ†
298
+ ↓(x)ψ↓(x)
299
+
300
+ ,
301
+ (4)
302
+ is a conserved quantity which verifies [sz, H] = 0, we
303
+ can label the electronic eigenstates of H with σ = {↑, ↓}.
304
+ Therefore, we introduce the following Nambu spinors
305
+ Ψ↑(x) =
306
+ � ψ↑(x)
307
+ ψ†
308
+ ↓(x)
309
+
310
+ ,
311
+ Ψ↓(x) =
312
+ � ψ↓(x)
313
+ ψ†
314
+ ↑(x)
315
+
316
+ ,
317
+ (5)
318
+ related to each other via the charge-conjugation transfor-
319
+ mation Ψ¯σ(x) = KτxΨσ(x), where τx is the 2 × 2 Pauli
320
+ matrix, and K is the complex conjugation operator. In
321
+ terms of these spinors the Hamiltonian writes
322
+ H = 1
323
+ 2
324
+
325
+ σ
326
+
327
+ Lw
328
+ 2
329
+ − Lw
330
+ 2
331
+ dx Ψ†
332
+ σ(x)HBdG,σ(x)Ψσ(x),
333
+ (6)
334
+ where the Bogoliubov-de Gennes (BdG) Hamiltonian is
335
+ defined as
336
+ HBdG,σ =
337
+
338
+ − ℏ2∂2
339
+ x
340
+ 2m − µ + σh(x)
341
+ σ∆
342
+ σ∆
343
+ ℏ2∂2
344
+ x
345
+ 2m + µ + σh(x)
346
+
347
+ .
348
+ (7)
349
+ In this expression, the spin projection σ =↑ (↓) on
350
+ the left-hand side corresponds to the + (−) sign in
351
+ the definition of the BdG matrix.
352
+ Using the above
353
+ charge-conjugation transformation, we note that the
354
+ BdG Hamiltonian Eq. (7) verifies the following symme-
355
+ try transformation
356
+ KτxHBdG,σ = −H∗
357
+ BdG,¯σKτx,
358
+ (8)
359
+ and therefore, provided χσ(x) is a solution of the BdG
360
+ eigenvalue equation
361
+ HBdG,σ(x)χσ(x) = Eσχσ(x),
362
+ (9)
363
+ with eigenenergy Eσ, the transformed spinor χ¯σ(x) =
364
+ Kτxχσ(x), is also a solution with eigenenergy E¯σ = −Eσ.
365
+ In what follows, we assume for simplicity the thermo-
366
+ dynamic limit Lw → ∞, and we focus on the features
367
+ introduced by the magnitude and spatial dependence of
368
+ h (x), which is crucial for the rest of this work. In addi-
369
+ tion, we assume the following step-like spatial profile for
370
+ the exchange field
371
+ h(x) =
372
+
373
+ −h0
374
+ if |x| < L
375
+ 2 ,
376
+ 0
377
+ if |x| ⩾ L
378
+ 2 ,
379
+ (10)
380
+ which models a uniform FMI shell of length L in contact
381
+ with the SE nanowire (see Fig. 1). This choice for h(x)
382
+ allows to split the problem into regions with either |x| <
383
+ L
384
+ 2 or |x| > L
385
+ 2 , with generic exponential solutions
386
+ χσ(x) ∼
387
+
388
+ ασ
389
+ βσ
390
+
391
+ eikx.
392
+ (11)
393
+ Linear combinations of Eq. (11), with appropriate coeffi-
394
+ cients and with allowed values of k for each region, must
395
+ be built so that continuity of the total wavefunction and
396
+ its derivative at the interfaces is satisfied. With this re-
397
+ quirement, the solution of Eq.(9) is finally obtained.
398
+ Note that the BdG Hamiltonian (7) is even under space
399
+ inversion x → −x, and therefore its eigenstates must be
400
+ even or odd under this transformation of coordinates.
401
+ This symmetry allows to reduce the number of unknowns
402
+ of the problem (i.e., coefficients of the linear combinta-
403
+ tion). Replacing the above ansatz Eq. (11) into the BdG
404
+ eigenvalue Eq.
405
+ (9), and looking for localized solutions
406
+ with energy within the gap |Eσ| < ∆, we obtain the fol-
407
+ lowing expressions for the eigenstates belonging to the
408
+ even-symmetry subspace:
409
+
410
+ 4
411
+ χe,σ
412
+
413
+ x > L
414
+ 2
415
+
416
+ = Ae
417
+
418
+
419
+ 1
420
+ σe−iϕσ
421
+
422
+ e−κσx + Ae
423
+
424
+
425
+ 1
426
+ σeiϕσ
427
+
428
+ e−κ∗
429
+ σx,
430
+ (12)
431
+ χe,σ
432
+
433
+ −L
434
+ 2 ≤ x ≤ L
435
+ 2
436
+
437
+ = Be
438
+
439
+
440
+ 1
441
+ σe−ησ
442
+
443
+ cos kσx + Be
444
+
445
+
446
+ 1
447
+ σeησ
448
+
449
+ cos ¯kσx,
450
+ (13)
451
+ and the following expressions for the odd-symmetry eigenfunctions
452
+ χo,σ
453
+
454
+ x > L
455
+ 2
456
+
457
+ = Ao
458
+
459
+
460
+ 1
461
+ σe−iϕσ
462
+
463
+ e−κσx + Ao
464
+
465
+
466
+ 1
467
+ σeiϕσ
468
+
469
+ e−κ∗
470
+ σx,
471
+ (14)
472
+ χo,σ
473
+
474
+ −L
475
+ 2 ≤ x ≤ L
476
+ 2
477
+
478
+ = Bo
479
+
480
+
481
+ 1
482
+ σe−ησ
483
+
484
+ sin kσx + Bo
485
+
486
+
487
+ 1
488
+ σeησ
489
+
490
+ sin ¯kσx,
491
+ (15)
492
+ where the coefficients {Aν
493
+ 1σ, Aν
494
+ 2σ, Bν
495
+ 1σ, Bν
496
+ 2σ}, with ν =
497
+ {e, o}, are unknowns to be fixed.
498
+ In addition, in the
499
+ above expressions we have introduced the parametriza-
500
+ tion
501
+ cos ϕσ = Eσ
502
+ ∆ ,
503
+ (16)
504
+ cosh ησ = Eσ + σh0
505
+
506
+ ,
507
+ (17)
508
+ where we fix the definition of ϕσ to the interval ϕσ ∈
509
+ (0, π]. The phase variable ϕσ is associated to the An-
510
+ dreev reflection taking place at the interface xb = L/2.
511
+ Note that the parametrization in Eq. (17) makes sense
512
+ whenever the right-hand side is positive. If this condi-
513
+ tion is not satisfied, one can always use the symmetry
514
+ Eq.(8) to send Eσ → −E¯σ and σ → ¯σ. In addition, note
515
+ that whenever 1 ≤ (Eσ + σh0) /∆ the parameter ησ is
516
+ purely real, while for 0 < (Eσ + σh0) /∆ < 1 it is purely
517
+ imaginary. Finally, we have introduced the quantities
518
+ κσ ≡ −ikF
519
+
520
+ 1 + 2i
521
+ kF ξ sin ϕσ,
522
+ (18)
523
+ kσ ≡ kF
524
+
525
+ 1 +
526
+ 2
527
+ kF ξ sinh ησ,
528
+ (19)
529
+ ¯kσ ≡ kF
530
+
531
+ 1 −
532
+ 2
533
+ kF ξ sinh ησ,
534
+ (20)
535
+ and the definition of the coherence length of the
536
+ (proximity-induced) 1D superconductor ξ = ℏvF /∆. No-
537
+ tice also that the spatial dependence of the wavefunc-
538
+ tions in the region x < −L/2 can be readily obtained by
539
+ symmetry from the relations χe,σ (x) = χe,σ (−x), and
540
+ χo,σ (x) = −χo,σ (−x).
541
+ We can intuitively understand the form of the scatter-
542
+ ing solutions in the regions x > L/2 and x < −L/2 in the
543
+ limit kF ξ ≫ 1 (i.e., the semiclassical limit, see Sec.III A),
544
+ where the momentum κσ in Eq. (18) can be expanded
545
+ as κσ ≃ −ikF + sin ϕσ/kF ξ, and the eigenfunctions Eqs.
546
+ (12) and (14) take the form
547
+ χν,σ
548
+
549
+ x > L
550
+ 2
551
+
552
+
553
+
554
+
555
+
556
+
557
+ 1
558
+ σe−iϕσ
559
+
560
+ eikF x+
561
+ +Aν
562
+
563
+
564
+ 1
565
+ σeiϕσ
566
+
567
+ e−ikF x
568
+
569
+ e− sin ϕσx
570
+ ξ
571
+ , (21)
572
+ with ν = {e, o}. In this way, it becomes evident that the
573
+ component proportional to Aν
574
+ 1σ corresponds to a right-
575
+ moving particle ∼ eikF x while Aν
576
+ 2σ corresponds to a left-
577
+ moving particle ∼ e−ikF x. In addition, the wavefunctions
578
+ exponentially decay into the superconductor within a lo-
579
+ calization length λloc = ξ/ sin ϕσ = ξ/
580
+
581
+ 1 − (Eσ/∆)2.
582
+ These results are in complete agreement with Ref. [32],
583
+ where the spectrum of SU-FM-SU Josephson junctions
584
+ has been recently studied as a function of the length L
585
+ of the FM region. However, in our case, the presence of
586
+ a finite pairing gap ∆ in the region −L/2 < x < L/2 (as
587
+ opposed to the assumption ∆ = 0 in the FM region in
588
+ that work), gives rise to important differences which we
589
+ analyze below in Sec. III.
590
+ A.
591
+ Continuity conditions at the interface
592
+ We now impose the continuity conditions on the wave-
593
+ function and its derivative at the boundary xb = L/2:
594
+ χν,σ
595
+
596
+ x−
597
+ b
598
+
599
+ = χν,σ
600
+
601
+ x+
602
+ b
603
+
604
+ (22)
605
+ ∂xχν,σ
606
+
607
+ x−
608
+ b
609
+
610
+ = ∂xχν,σ
611
+
612
+ x+
613
+ b
614
+
615
+ .
616
+ (23)
617
+ Note that the same equations are obtained by symme-
618
+ try at the other boundary −xb. Inserting the solutions
619
+ Eqs. (12)-(15), we can express the continuity equations
620
+ in matrix form as
621
+
622
+ 5
623
+
624
+ 1
625
+ σe−iϕσ
626
+ σe−iϕσ
627
+ 1
628
+ � �
629
+
630
+
631
+
632
+
633
+
634
+ =
635
+
636
+ 1
637
+ σe−ησ
638
+ σe−ησ
639
+ 1
640
+ � �
641
+
642
+ � kσL
643
+ 2
644
+
645
+ 0
646
+ 0
647
+
648
+ � ¯kσL
649
+ 2
650
+
651
+ � �
652
+
653
+
654
+
655
+
656
+
657
+ ,
658
+ (24)
659
+
660
+
661
+ 1
662
+ σe−iϕσ
663
+ σe−iϕσ
664
+ 1
665
+ � �
666
+ κσ
667
+ 0
668
+ 0
669
+ κ∗
670
+ σ
671
+ � �
672
+
673
+
674
+
675
+
676
+
677
+ = −s(ν)
678
+
679
+ 1
680
+ σe−ησ
681
+ σe−ησ
682
+ 1
683
+ � �
684
+ kσGν
685
+ � kσL
686
+ 2
687
+
688
+ 0
689
+ 0
690
+ ¯kσGν
691
+ � ¯kσL
692
+ 2
693
+
694
+ � �
695
+
696
+
697
+
698
+
699
+
700
+ ,
701
+ (25)
702
+ where we have conveniently redefined the unknown coef-
703
+ ficients as
704
+
705
+ 1σ → eκσL/2aν
706
+
707
+
708
+ 1σ → bν
709
+
710
+ (26)
711
+
712
+ 2σ → σeκ∗
713
+ σL/2e−iϕσaν
714
+
715
+
716
+ 2σ → σe−ησbν
717
+ 2σ,
718
+ (27)
719
+ in order to give these equations a more symmetric form.
720
+ In addition, we have used the notation s(ν) = +1(−1) for
721
+ ν = e(o), and Fe(x) = Go(x) ≡ cos(x), Ge(x) = Fo(x) ≡
722
+ sin(x) for compactness.
723
+ In each subspace (even or odd) we have four equa-
724
+ tions and four unknowns.
725
+ Eliminating the variables
726
+ (bν
727
+ 1σ, bν
728
+ 2σ)T , and writing the equation for (aν
729
+ 1σ, aν
730
+ 2σ)T , we
731
+ find from the nullification of the corresponding determi-
732
+ nant the following equations:
733
+ cosh ησ cos ϕσ − 1
734
+ sinh ησ sin ϕσ
735
+ =
736
+
737
+
738
+
739
+
740
+
741
+
742
+
743
+
744
+
745
+
746
+
747
+
748
+
749
+ |κσ|2 −
750
+
751
+ Kσ + ¯Kσ
752
+
753
+ Re κσ + Kσ ¯Kσ
754
+ � ¯Kσ − Kσ
755
+
756
+ Im κσ
757
+ (even-symmetry subspace),
758
+ |κσ|2 +
759
+
760
+ Qσ + ¯Qσ
761
+
762
+ Re κσ + Qσ ¯Qσ
763
+
764
+ Qσ − ¯Qσ
765
+
766
+ Im κσ
767
+ (odd-symmetry subspace),
768
+ (28)
769
+ where we have defined the quantities
770
+ Kσ = kσ tan
771
+ �kσL
772
+ 2
773
+
774
+ ,
775
+ (29)
776
+ ¯Kσ = ¯kσ tan
777
+ �¯kσL
778
+ 2
779
+
780
+ ,
781
+ (30)
782
+ Qσ = kσ cot
783
+ �kσL
784
+ 2
785
+
786
+ ,
787
+ (31)
788
+ ¯Qσ = ¯kσ cot
789
+ �¯kσL
790
+ 2
791
+
792
+ .
793
+ (32)
794
+ From Eq. (28), the eigenvalue Eσ for each subspace is
795
+ finally obtained. This equation summarizes our main the-
796
+ oretical results. In the next Sec. III we analyze the nu-
797
+ merical solution and different important limits.
798
+ B.
799
+ Spin-changing quantum phase transitions
800
+ We now focus on the quantum phase transitions which
801
+ occur whenever one of the subgap states crosses EF . To
802
+ that end, let us analyze the spinors defined in Eq. (5),
803
+ and consider the norm of the “up” spinor
804
+ q↑ =
805
+ � Lw/2
806
+ −Lw/2
807
+ dx
808
+
809
+ ψ†
810
+ ↑ (x) ψ↑ (x) + ψ↓ (x) ψ†
811
+ ↓ (x)
812
+
813
+ .
814
+ Recalling the definition of the single-particle sz operator
815
+ [see Eq. (4)], it is straightforward to associate these two
816
+ quantities through the relation q↑ = 2sz − 1. Since sz
817
+ is a conserved quantity, so is the norm q↑ of the “up”
818
+ Nambu spinors. This connection allows to interpret q↑ as
819
+ an effective “conserved charge”. Similar considerations
820
+ allow to write the relation q↓ = −2sz − 1. Due to the
821
+ particle-hole relation Eq.(8), the information about sz
822
+ can be obtained with either q↑ or q↓. A more symmetric
823
+ form involving both conserved charges is
824
+ sz = q↑ − q↓
825
+ 4
826
+ .
827
+ (33)
828
+ While redundant, this expression makes explicit that in
829
+ the spin-symmetric case q↑ = q↓, the net spin sz must
830
+ vanish (sz = 0).
831
+ We now return to Hamiltonian Eq.
832
+ (7), and let us
833
+ separate the effect of the proximity-induced Zeeman field,
834
+ by writing it as HBdG,σ = H0,σ + Vσ, where
835
+ H0,σ =
836
+
837
+ − ℏ2∂2
838
+ x
839
+ 2m − µ
840
+ σ∆
841
+ σ∆
842
+ ℏ2∂2
843
+ x
844
+ 2m + µ
845
+
846
+ ,
847
+ (34)
848
+ Vσ =
849
+
850
+ σh(x)
851
+ 0
852
+ 0
853
+ σh(x)
854
+
855
+ .
856
+ (35)
857
+ In this form, we can interpret the effect of the exchange
858
+ field as a “perturbation” on an otherwise homogeneous
859
+
860
+ 6
861
+ 1D superconductor represented by H0,σ. Therefore, the
862
+ full and the unperturbed single-particle Green’s functions
863
+ in this problem are respectively defined as
864
+ Gσ (z) = [z − H0,σ − Vσ]−1 ,
865
+ (36)
866
+ G0,σ (z) = [z − H0,σ]−1 ,
867
+ (37)
868
+ From here, the total number of effective “up” charges
869
+ Q↑ induced in the ground state due to the potential Vσ,
870
+ compared to the unperturbed homogeneous SU wire, can
871
+ be computed as
872
+ ∆Q↑ = − 1
873
+ π Im Tr
874
+ � ∞
875
+ −∞
876
+ dϵ nF (ϵ) ∆G↑ (ϵ + iδ) .
877
+ (38)
878
+ where ∆Gσ (z) ≡ Gσ (z) − G0,σ (z). At T = 0, Eq. (38)
879
+ can be easily computed from the well-known expression
880
+ of the Friedel sum rule [35]
881
+ ∆Q↑ = 1
882
+ π
883
+ � 0
884
+ −∞
885
+
886
+ �∂η↑ (ϵ)
887
+ ∂ϵ
888
+ − ∂η0,↑ (ϵ)
889
+ ∂ϵ
890
+
891
+ (39)
892
+ = η↑ (0) − η0,↑ (0)
893
+ π
894
+ (40)
895
+ where we have defined the phase shifts [32, 35]
896
+ ησ (ϵ) = Im ln det Gσ (ϵ + iδ) ,
897
+ (41)
898
+ η0,σ (ϵ) = Im ln det G0,σ (ϵ + iδ) ,
899
+ (42)
900
+ and where we have used that the phase shifts vanish in
901
+ the limit ϵ → ±∞.
902
+ Since the system is non-interacting, the Green’s func-
903
+ tion Eq. (36) can be written in terms of single-particle
904
+ eigenstates |α, σ⟩, with α a generic label, as
905
+ Gσ (z) =
906
+
907
+ α
908
+ |α, σ⟩ ⟨α, σ|
909
+ z − Eα,σ
910
+ .
911
+ (43)
912
+ Therefore, after simple algebra, and using the above re-
913
+ lations and the fact that in the absence of magnetic field
914
+ sz = 0 [see Eq. (33)], the total Sz of the ground state is
915
+ Sz = ∆Q↑
916
+ 2
917
+ = 1
918
+ 2
919
+ ��
920
+ α
921
+ Θ (−Eα,↑) −
922
+
923
+ α′
924
+ Θ
925
+
926
+ −E0
927
+ α′,↑
928
+
929
+
930
+ ,
931
+ (44)
932
+ where Θ(ϵ) is the unit-step function. The above expres-
933
+ sion allows to interpret the total Sz of the ground state
934
+ as a function of the “up” Nambu spinors with energy be-
935
+ low EF = 0, as compared to the (unperturbed) situation
936
+ h0 = 0. Since the effective charges are quantized in inte-
937
+ ger numbers, the total spin Sz can only change in discrete
938
+ “jumps” of 1/2 whenever a subgap state with projection
939
+ up crosses below EF (note that we have defined dimen-
940
+ sionless spin operators). This interpretation makes sense
941
+ since the ground state becomes spin-polarized when the
942
+ exchange field h0 becomes large enough [i.e., the Zeeman
943
+ energy of up-spin electron is decreased, see Eqs. (3) and
944
+ (10)]. While the result of Eq. (44) has been obtained re-
945
+ cently by the authors of Ref. [32], we note that here we
946
+ have rederived it in a different physical situation which
947
+ allows a more generic regime of parameters.
948
+ III.
949
+ RESULTS
950
+ We start this section by analyzing different limits of
951
+ the general result given in Eq. (28). In particular, in
952
+ Sec. III A we focus on the semiclassical limit, and in Sec.
953
+ III B we study the atomic limit, where we recover the
954
+ YSR results. In both cases, Eq. (28) reduces to well-
955
+ known analytical results. Finally in Sec. III C we show
956
+ results corresponding to intermediate regimes, obtained
957
+ by solving numerically Eq. (28).
958
+ A.
959
+ Semiclassical limit
960
+ Generally speaking, the semiclassical limit is verified
961
+ when EF is the largest scale of the problem [36]. In par-
962
+ ticular, the condition EF ≫ ∆ (which is very well satis-
963
+ fied in most experimental systems) can be expressed as
964
+ kF ξ ≫ 1, recalling that after linearization of the normal
965
+ quasiparticle dispersion, i.e., ϵk,σ ≃ ±ℏvF k, where the
966
+ +(−) sign corresponds to right-(left-)movers, the Fermi
967
+ energy can be approximated as EF ≃ ℏkF vF . In this
968
+ case, Eqs. (18)-(20) reduce to
969
+ rσ ≡ κσ
970
+ kF
971
+ ≃ −i + sin ϕσ
972
+ kF ξ ,
973
+ (45)
974
+ ζσ ≡ kσ
975
+ kF
976
+ ≃ 1 + sinh ησ
977
+ kF ξ
978
+ ,
979
+ (46)
980
+ ¯ζσ ≡
981
+ ¯kσ
982
+ kF
983
+ ≃ 1 − sinh ησ
984
+ kF ξ
985
+ ,
986
+ (47)
987
+ to leading order in O(kF ξ)−1, and Eq. (28) becomes
988
+ cosh ησ cos ϕσ − 1
989
+ sinh ησ sin ϕσ
990
+ ≃ s(ν)
991
+ 1 + tan
992
+
993
+ kF Lζσ
994
+ 2
995
+
996
+ tan
997
+
998
+ kF L¯ζσ
999
+ 2
1000
+
1001
+ tan
1002
+
1003
+ kF Lζσ
1004
+ 2
1005
+
1006
+ − tan
1007
+
1008
+ kF L¯ζσ
1009
+ 2
1010
+ � ,
1011
+ = s(ν) cot
1012
+ �L sinh ησ
1013
+ ξ
1014
+
1015
+ ,
1016
+ (48)
1017
+ where
1018
+ we
1019
+ have
1020
+ used
1021
+ the
1022
+ trigonometric
1023
+ identity
1024
+ tan (x + y) = (tan (x) + tan y)/(1 + tan x tan y). In gen-
1025
+ eral this transcendental equation cannot be solved ana-
1026
+ lytically. However, in the regime of parameters EF ≫
1027
+ h0 ≫ ∆, where the exchange field h0 is much larger
1028
+ than ∆, we can write cosh ησ ≈ sinh ησ ≈
1029
+ �� h0
1030
+
1031
+ �� ≫ 1
1032
+ [see Eqs.
1033
+ (16) and (17) ], and Eq.
1034
+ (48) reduces to
1035
+ cot ϕσ = s(ν) cot (Lh0/ℏvF ). Equivalently we can write
1036
+ this result as
1037
+ arccos
1038
+ �Eσ
1039
+
1040
+
1041
+ =
1042
+
1043
+
1044
+
1045
+
1046
+
1047
+
1048
+
1049
+
1050
+
1051
+ LEσ
1052
+ ℏvF
1053
+ + σ Lh0
1054
+ ℏvF
1055
+ + 2nπ,
1056
+ (even)
1057
+ LEσ
1058
+ ℏvF
1059
+ + σ Lh0
1060
+ ℏvF
1061
+ + (2n + 1) π.
1062
+ (odd)
1063
+ (49)
1064
+ This result can be interpreted as a semiclassical Bohr-
1065
+ Sommerfeld quantization condition for particles which
1066
+
1067
+ 7
1068
+ perform a complete a closed loop in the region −L/2 <
1069
+ x < L/2 [36]. In particular, it exactly coincides with the-
1070
+ oretical results obtained for SU-FM-SU Josephson junc-
1071
+ tions with a normal (i.e., ∆ = 0) FM region [30–32], the
1072
+ only difference being that within our theoretical treat-
1073
+ ment, we can distinguish the symmetry of the solutions.
1074
+ The similarity of these results can be rationalized noting
1075
+ that considering a normal sandwiched region in an SU-
1076
+ FM-SU junction corresponds to taking the limit h0 ≫ ∆
1077
+ in our Eq. (48) while keeping the ratio Eσ/∆ finite (since
1078
+ Eσ corresponds to a subgap state, it is always bounded
1079
+ by ∆), thus resulting in Eq. (49). This shows that our
1080
+ Eq. (28) is a generic relation describing different situa-
1081
+ tions regardless of the magnitude of the ratio h0/∆.
1082
+ B.
1083
+ YSR-impurity limit
1084
+ We now consider the atomic YSR (or simply Shiba)
1085
+ limit, in which the exchange profile becomes point-like,
1086
+ L → 0, while h0 → ∞, in such a way that the product
1087
+ Lh0 = J = const. Under these assumptions the magnetic
1088
+ barrier becomes a delta function and the Hamiltonian in
1089
+ Eq. (3) can be written as
1090
+ HZ ≈ −J
1091
+ � ∞
1092
+ −∞
1093
+ dx δ(x)
1094
+
1095
+ ψ†
1096
+ ↑(x)ψ↑(x) − ψ†
1097
+ ↓(x)ψ↓(x)
1098
+
1099
+ .
1100
+ (50)
1101
+ In this case, it is easy to see that the odd-symmetry solu-
1102
+ tions decouple from the above Hamiltonian (50), as they
1103
+ vanish at x = 0, and only even solutions can couple to
1104
+ the delta-potential.
1105
+ As in the previous section, note that the limit h0 → ∞
1106
+ implies cosh ησ ≈ sinh ησ ≈
1107
+ �� h0
1108
+
1109
+ �� ≫ 1. However, the limit
1110
+ h0 → ∞ is not compatible with the semiclassical ap-
1111
+ proach, as it violates the requirement h0 ≪ EF . There-
1112
+ fore we cannot use here our previous Eq. (49). Instead,
1113
+ we must first take the limit ησ ≫ 1 together with the
1114
+ limit L → 0, which applied to Eqs. (19) and (20) yield
1115
+ kσ → kF
1116
+
1117
+ 2h0
1118
+ ℏvF kF
1119
+ ,
1120
+ (51)
1121
+ ¯kσ → ikF
1122
+
1123
+ 2h0
1124
+ ℏvF kF
1125
+ .
1126
+ (52)
1127
+ In addition Eqs. (29)-(32) become
1128
+ Kσ → kF h0L
1129
+ ℏvF
1130
+ = kF ρ0J,
1131
+ (53)
1132
+ ¯Kσ → −kF h0L
1133
+ ℏvF
1134
+ = −kF ρ0J,
1135
+ (54)
1136
+ where the expressions for the density of states per spin
1137
+ of 1D quasiparticles at the Fermi energy ρ0 = 1/ℏvF ,
1138
+ and the exchange coupling J = h0L, have been used.
1139
+ Replacing these expressions into Eq. (28) for the even-
1140
+ symmetry solutions, we obtain
1141
+ σ
1142
+ Ee
1143
+ σ
1144
+
1145
+ ∆2 − (Ee)2
1146
+ σ
1147
+ = 1 − (ρ0J)2
1148
+ (2Jρ0)
1149
+ .
1150
+ (55)
1151
+ From this expression, we can easily solve for Ee
1152
+ σ
1153
+ Ee
1154
+ σ
1155
+ ∆ = σ 1 − (ρ0J)2
1156
+ 1 + (ρ0J)2 ,
1157
+ (56)
1158
+ which is the well-known expression for the energy of YSR-
1159
+ impurity subgap level [1]. This result indicates that any
1160
+ finite value of J produces a YSR in-gap state. This type
1161
+ of subgap YSR states has been observed in several STM
1162
+ experiments on atomic magnetic adsorbates on supercon-
1163
+ ducing substrates [27, 37–41].
1164
+ For completeness, and in order to illustrate the general
1165
+ scope of Eq. (28), here we also show the result for the
1166
+ YSR odd states for a small (but finite) L. Using similar
1167
+ approximations, we obtain the expression
1168
+ Eo
1169
+ σ
1170
+ ∆ = σ
1171
+ 1
1172
+
1173
+ 1 +
1174
+ �ρ0Jk2
1175
+ F L2
1176
+ 6
1177
+ �2 ,
1178
+ (57)
1179
+ where it becomes evident that in addition to a finite value
1180
+ of J, a finite value of kF L is needed to observe an odd-
1181
+ symmetry subgap YSR state.
1182
+ C.
1183
+ Subgap ABS spectrum in generic cases
1184
+ As stated in Section II, Eq. (28) implicitly defines the
1185
+ energy of the subgap states as a function of the param-
1186
+ eters h0/∆ , kF ξ, and kF L. These parameters can be
1187
+ directly or indirectly controlled in experiments, i.e., the
1188
+ parameter h0 can be controlled by modifying the FMI
1189
+ material, the length L of the FMI region can be modified
1190
+ varying the length Lw of the semiconductor via vapor-
1191
+ liquid-solid (VLS) method and subsequent evaporation of
1192
+ the FMI material [20], and the parameter kF in the semi-
1193
+ conductor can be varied by changing the SE material or
1194
+ by introducing external gates to modify the chemical po-
1195
+ tential µ. Therefore, due to this high degree of tunability,
1196
+ hybrid heterostructures might offer a unique platform to
1197
+ produce and control engineered subgap states. Probably
1198
+ the easiest way to experimentally control the subgap elec-
1199
+ tronic structure is by producing different devices with the
1200
+ same FMI material and different lengths L. Therefore, in
1201
+ this section we show the numerical solutions of Eq. (28)
1202
+ with fixed parameters h0/∆ and kF ξ (which control the
1203
+ “operation regime” of the device), and calculate both the
1204
+ energy dependence of the even- and odd-symmetry ABS,
1205
+ and the total spin Sz of the device as a function of L (i.e.,
1206
+ dimensionless variable kF L).
1207
+ Generally speaking, the overall evolution of the ABS
1208
+ spectrum from L = 0 to L → ∞ is quite complex and de-
1209
+ serves a detailed explanation. As shown in Fig. 2, as the
1210
+
1211
+ 8
1212
+ parameter kF L increases, more and more subgap states
1213
+ emerge from the gap edges. This behavior is reminiscent
1214
+ of a quantum particle in a square-well potential, tipically
1215
+ taught in introductory quantum mechanics courses [42],
1216
+ where increasing the width L of the well increases the
1217
+ number of allowed bound states. In our case, the emer-
1218
+ gence of new ABS as kF L increases can be intuitively
1219
+ understood in terms of a competition between supercon-
1220
+ ductivity and magnetic field: the magnetic field tends
1221
+ to break Cooper-pairs and to locally disrupt supercon-
1222
+ ductivity in the magnetic region by introducing subgap
1223
+ states that become macroscopic in number for large L,
1224
+ eventually populating the whole gap.
1225
+ We note that for any finite L, even- and odd-symmetry
1226
+ states are generically non-degenerate (except at isolated
1227
+ points). However, as it is clear from Figs. 2 and 3, their
1228
+ energy difference (evidenced as oscillations of the blue
1229
+ and red lines around the semiclassical value) decreases
1230
+ very rapidly and the solutions become degenerate in the
1231
+ limit L → ∞. This transition from non-degenerate YSR
1232
+ states in the limit L → 0, to double degenerate ABS
1233
+ states for L → ∞ has been discussed in previous works
1234
+ on ballistic SU-FM-SU junctions [19, 30–32], and in the
1235
+ case of extended Shiba impurities in 1D nanowires [43].
1236
+ It is also clearly visible in Fig. 2, and more dramatically
1237
+ in Fig. 3 below. In our 1D geometry, this degeneracy
1238
+ in the limit L → ∞ can be intuitively understood by
1239
+ linearizing the spectrum around the Fermi energy, and
1240
+ expressing the original fermionic operators in terms of
1241
+ right- and left-moving fields slowly varying in the scale
1242
+ of k−1
1243
+ F
1244
+ [44], i.e., ψσ (x) ≈ eikF xψR,σ (x) + e−ikF xψL,σ (x).
1245
+ The slowly-varying fields ψR,σ(x) and ψL,σ(x) are two
1246
+ independent chiral fermionic fields obeying the usual an-
1247
+ ticommutation relations, in terms of which the original
1248
+ Hamiltonian becomes [43]
1249
+ Hw ≈
1250
+
1251
+ σ
1252
+ � ∞
1253
+ −∞
1254
+ dx
1255
+
1256
+ −iℏvF ψ†
1257
+ R,σ(x)∂xψR,σ(x)
1258
+ + iℏvF ψ†
1259
+ L,σ(x)∂xψL,σ(x)
1260
+
1261
+ (58)
1262
+ H ∆ ≈ ∆
1263
+ � ∞
1264
+ −∞
1265
+ dx
1266
+
1267
+ ψ†
1268
+ R,↑(x)ψ†
1269
+ L,↓(x) + ψ†
1270
+ L,↑(x)ψ†
1271
+ R,↓(x) + H.c.
1272
+
1273
+ ,
1274
+ (59)
1275
+ HZ ≈ −
1276
+ � ∞
1277
+ −∞
1278
+ dx h0
1279
+
1280
+ ψ†
1281
+ R,↑(x)ψR,↑(x) − ψ†
1282
+ R,↓(x)ψR,↓(x)
1283
+ + ψ†
1284
+ L,↑(x)ψL,↑(x) − ψ†
1285
+ L,↓(x)ψL,↓(x)
1286
+
1287
+ ,
1288
+ (60)
1289
+ where oscillating terms proportional to e±2ikF x have been
1290
+ neglected as they cancel out in the limit L → ∞ due to
1291
+ destructive interference. Defining the new chiral Nambu
1292
+ spinors
1293
+ Ψ1,σ(x) =
1294
+ � ψR,σ(x)
1295
+ ψ†
1296
+ L,¯σ(x)
1297
+
1298
+ ,
1299
+ Ψ2,σ(x) =
1300
+ � ψL,σ(x)
1301
+ ψ†
1302
+ R,¯σ(x)
1303
+
1304
+ ,
1305
+ (61)
1306
+ the Hamiltonian of the system can be expressed in terms
1307
+ of two decoupled chiral sectors
1308
+ H = 1
1309
+ 2
1310
+
1311
+ σ=↑,↓
1312
+
1313
+ j=1,2
1314
+ � ∞
1315
+ −∞
1316
+ dx Ψ†
1317
+ j,σ(x)Hj,σ(x)Ψj,σ(x), (62)
1318
+ with the definitions of the chiral BdG Hamiltonians
1319
+ Hj,σ =
1320
+
1321
+ (−1)jivF ∂x − σh0
1322
+ σ∆
1323
+ σ∆
1324
+ (−1)j+1ivF ∂x − σh0
1325
+
1326
+ .
1327
+ (63)
1328
+ The Nambu spinors Eq. (61) define two independent chi-
1329
+ ral subspaces related by the inversion symmetry of the
1330
+ original Hamiltonian, i.e., under the space inversion op-
1331
+ eration x ↔ −x, the fermionic operators transform as
1332
+ ψL,σ(x) ↔ ψR,σ(x), and consequently we conclude that
1333
+ Ψ1,σ(x) ↔ Ψ2,σ(x), which must then be degenerate. In
1334
+ addition, the particle-hole symmetry Eq. (8) in this rep-
1335
+ resentation produces Ψ1,σ(x) → Ψ2,¯σ(x), and therefore
1336
+ H1,σ → −H2,¯σ, implying that the solutions verify the
1337
+ particle-hole symmetry property E1,σ = −E2,¯σ. More-
1338
+ over, notice that assuming periodic boundary conditions,
1339
+ the problem can be solved with the solutions ψR,σ(x) ∼
1340
+ eikx and ψL,σ(x) ∼ e−ikx, and the dispersion relation
1341
+ becomes E1,σ(k) = E2,σ(k) = ±
1342
+
1343
+ (ℏvF k)2 + ∆2 − σh0.
1344
+ From here, a renormalized quasiparticle gap 2∆ren =
1345
+ 2 |∆ − h0| is obtained, consistent with our previous re-
1346
+ sult.
1347
+ In terms of the chiral Nambu spinors, the most general
1348
+ solution is the linear combination
1349
+ Ψσ(x) = AeikF xΨ1,σ(x) + Be−ikF xΨ2,σ(x).
1350
+ (64)
1351
+ This is exactly the same form that can be obtained by
1352
+ combining the degenerate even and odd solutions in Eqs.
1353
+ (13) and (15) in the semiclassical limit where kF ξ ≫ 1.
1354
+ From the analysis of the linearized Hamiltonian, we
1355
+ conclude that the degeneracy in the limit L → ∞
1356
+ arises from the absence of chirality-breaking terms, i.e.,
1357
+ terms ∼ Ψ†
1358
+ 1,σ(x)Ψ2,σ(x) arising from, e.g., single par-
1359
+ ticle backscattering terms ψ†
1360
+ R,σ(x)ψL,σ(x) or Cooper-
1361
+ pairing channels ψ†
1362
+ R(L),↑(x)ψ†
1363
+ R(L),↓(x) carrying momen-
1364
+ tum ∓2kF .
1365
+ For this to occur, the magnetic FMI re-
1366
+ gion must be uniform and its length L must be much
1367
+ larger than k−1
1368
+ F
1369
+ in order to produce the required cancel-
1370
+ lation of the rapidly oscillating exponentials ∼ e±2ikF x.
1371
+ In other words, the product kF L must be kF L ≫ 1,
1372
+ consistent with our numerical results in Figs. 2 and 3.
1373
+ Only for small values of kF L, where this destructive in-
1374
+ terference is incomplete, residual couplings of the type
1375
+ ∼ Ψ†
1376
+ 1,σ(x)Ψ2,σ(x) remain, and the degeneracy is lifted.
1377
+ Finally, we stress that the degeneracy in the limit L → ∞
1378
+ is a robust property to the presence of interactions, as
1379
+ shown in previous works [43].
1380
+ On the other hand, in the limit L → 0 and for any
1381
+ finite value of the Zeeman field h0, both (even and odd)
1382
+ solutions converge to Eσ/∆ → ±1, indicating that the
1383
+ FMI region is no longer relevant (i.e., it physically drops
1384
+
1385
+ 9
1386
+ −1
1387
+ 1
1388
+ 0
1389
+ E/∆
1390
+ −1
1391
+ 1
1392
+ 0
1393
+ 0
1394
+ 10
1395
+ 20
1396
+ 30
1397
+ 40
1398
+ 50
1399
+ 0
1400
+ 1
1401
+ 2
1402
+ 3
1403
+ 4
1404
+ 5
1405
+ 6
1406
+ kF L
1407
+ Sz
1408
+ 0
1409
+ 2
1410
+ 4
1411
+ 6
1412
+ 8
1413
+ 10
1414
+ 12
1415
+ 14
1416
+ kF L
1417
+ FIG. 2. Energy of the Andreev bound states (upper panel) and total spin Sz(lower panel) as a function of kF L, for kF ξ = 7.8
1418
+ and h0/∆ = 3.0 (left panel) and kF ξ = 3.4 and h0/∆ = 2.1 (right panel). Blue and red colors correspond to even and odd states
1419
+ respectively. Lines starting from the top gap edge at positive energy E/∆ = 1 (bottom gap edge at negative energy E/∆ = −1)
1420
+ correspond to up (down) spin projections of the states. For smaller values of kF L (right panel), plateaus corresponding to
1421
+ regions of integer and half-integer spin are more separated and might become easier to observe in experiments.
1422
+ from the description). However, the behavior near L = 0
1423
+ is quite different for each case: while the even-symmetry
1424
+ solution tends to E/∆ → 1 as [see Eq. (56)]
1425
+ Ee
1426
+ σ
1427
+ ∆ ≈ σ
1428
+
1429
+ 1 − 2
1430
+ �h0L
1431
+ ℏvF
1432
+ �2
1433
+ . . .
1434
+
1435
+ ,
1436
+ (65)
1437
+ from Eq. (57) we conclude that the odd solution behaves
1438
+ as
1439
+ Eo
1440
+ σ
1441
+ ∆ ≈ σ
1442
+
1443
+ 1 − 1
1444
+ 2
1445
+ �h0k2
1446
+ F L3
1447
+ 6ℏvF
1448
+ �2
1449
+ . . .
1450
+
1451
+ ,
1452
+ (66)
1453
+ therefore approaching the gap edge much faster as L → 0.
1454
+ Besides the general features of the spectrum discussed
1455
+ up to this point, its evolution as L increases is strongly
1456
+ affected by the values of the parameters kF ξ and h0/∆.
1457
+ In what follows, we analyze their effects on Figs. 2 and
1458
+ Fig. 3 respectively.
1459
+ 1.
1460
+ Effect of varying the parameter kF ξ
1461
+ This parameter can be considered as a “knob” which
1462
+ tunes the device from the semiclassical behavior (kF ξ
1463
+ large, see left panel in Fig. 2) into a “quantum” regime
1464
+ (kF ξ small, see right panel) where the spectrum is dom-
1465
+ inated by quantum oscillations. The hybrid heterostruc-
1466
+ ture under study is promising in this sense since, due to
1467
+ the combination of materials (in particular, semiconduc-
1468
+ tors with a much smaller kF as compared to metals), it
1469
+ is in principle possible that kF ξ can be experimentally
1470
+ controlled. In addition, kF could be further modified by
1471
+ introducing external gating leads (through the modifica-
1472
+ tion of the chemical potential µ). To illustrate the dra-
1473
+ matic changes in the spectrum as kF ξ varies, in Fig 2 we
1474
+ show the numerically obtained subgap spectra as a func-
1475
+ tion of kF L for kF ξ = 7.8 and h0/∆ = 3.0 (left panel),
1476
+ for and kF ξ = 3.4 and h0/∆ = 2.1 (right panel). Solid
1477
+ blue (red) lines correspond to even(odd)-symmetry solu-
1478
+ tions. Moreover, since we always assume h0 > 0, solu-
1479
+ tions emerging from the top edge E/∆ = 1 (bottom edge
1480
+ E/∆ = −1) correspond to spin up (spin down) solutions.
1481
+ In addition, note the reflection symmetry of the solutions
1482
+ around the horizontal E = 0 axis, a consequence of the
1483
+ particle-hole symmetry of the BdG Hamiltonian, Eq. (8).
1484
+ Upon decreasing kF ξ, the subgap spectrum becomes
1485
+ much more intricate due to the enhanced even-odd
1486
+ energy-splitting, which results in an amplified oscillatory
1487
+ behavior of the ABS (we have reduced the range of kF L in
1488
+ the right panel for clarity in the figure). Unfortunately,
1489
+ in the regime kF ξ ∼ 1 no analytic expressions for the
1490
+ subgap ABS are possible, but qualitative considerations
1491
+
1492
+ 10
1493
+ can be provided. In fact, the amplified oscillations can
1494
+ be traced back to the larger energy dependence of the
1495
+ momenta Eq. (18)-(20) as kF ξ decreases. Then, whereas
1496
+ for large kF ξ all these quantities converge to a static (i.e.,
1497
+ energy-independent) value ∼ kF , the limit of small kF ξ
1498
+ produces a larger effect on the space-dependence of the
1499
+ wave functions through the exponential factors in Eqs.
1500
+ (12)-(15). This in turn produces larger interference ef-
1501
+ fects, and an enhanced lifting of the even-odd degener-
1502
+ acy.
1503
+ This phenomenological behavior enables interesting
1504
+ possibilities, such as the chance to observe half-integer
1505
+ spin (and fermion parity-switching) quantum phase tran-
1506
+ sitions in the ground state. To illustrate this effect, we
1507
+ show the ground-state Sz transitions in the bottom pan-
1508
+ els of Fig.
1509
+ 2 in each case.
1510
+ While for larger kF ξ, the
1511
+ half-integer Sz steps are very narrow due to the almost-
1512
+ degenerate even-odd solutions (i.e., the even and odd so-
1513
+ lutions cross zero energy almost at the same value of
1514
+ kF L), for smaller kF ξ the Sz transitions occur in well-
1515
+ defined half-integer steps. This behavior is well explained
1516
+ by the enhanced lifting of the even-odd degeneracy, which
1517
+ allows to observe one ABS crossing zero energy at a time.
1518
+ 2.
1519
+ Effect of varying the parameter h0/∆
1520
+ In Fig. 3 we show the evolution of the subgap spectrum
1521
+ as a function of kF L, for different values of the Zeeman
1522
+ field h0/∆ = 0.8, 1.54 and 2.2, and for a fixed relatively
1523
+ large value kF ξ = 8.2, allowing to interpret these results
1524
+ in terms of the semiclassical approximation. Here we can
1525
+ clearly distinguish three qualitatively different regimes:
1526
+ a) the “weak field” regime h0 < ∆ (top panel) where
1527
+ the ABS do not cross E = 0, b) the “intermediate field”
1528
+ regime ∆ < h0 < 2∆ (middle panel) where the ABS can
1529
+ evenually cross zero energy, and quantum phase transi-
1530
+ tions can be induced, and finally c) the “strong field”
1531
+ (2∆ < h0) regime (bottom panel), where the ABS can
1532
+ be found anywhere in the region −1 < Eσ/∆ < 1. In all
1533
+ cases, the value of h0 determines the asymptotic limit to
1534
+ which the ABS approach for large L (see dashed black
1535
+ lines in Fig. 3). Below we briefly discuss the main fea-
1536
+ tures of the spectrum in each regime.
1537
+ a.
1538
+ Weak-field regime 0 < h0 < ∆:
1539
+ This regime
1540
+ is characterized by a Zeeman field which is not strong
1541
+ enough to destroy the superconducting gap.
1542
+ In this
1543
+ case none of the ABS is able to cross E = 0 and in
1544
+ the limit L → ∞ they asymptotically approach the
1545
+ value Eσ/∆ → σ (1 − h0/∆) (see horizontal dashed black
1546
+ lines), and therefore a renormalized gap remains (see top
1547
+ panel in Fig. 3). More quantitatively, in the semiclassi-
1548
+ cal limit [Eq. (48)] they obey the asymptotic expression
1549
+ −1
1550
+ 1
1551
+ 0
1552
+ h0
1553
+ E/∆
1554
+ −1
1555
+ 1
1556
+ 0
1557
+ h0
1558
+ E/∆
1559
+ 0
1560
+ 20
1561
+ 40
1562
+ 60
1563
+ 80
1564
+ 100
1565
+ −1
1566
+ 1
1567
+ 0
1568
+ kF L
1569
+ E/∆
1570
+ FIG. 3. Energy of the Andreev bound states as a function of
1571
+ kF L for the three different values of h0 (h0/∆ = 0.8, 1.54, 2.2
1572
+ for the lower, middle and upper panels) and kF ξ = 8.2. Blue
1573
+ and red colors correspond to even and odd states respectively.
1574
+ Lines starting from negative (positive) energies correspond to
1575
+ down (up) spin projections of the state. Note that the value
1576
+ of h0/∆ sets the asymptotic limit for the Andreev states and
1577
+ is crucial to determine the overall subgap spectrum.
1578
+ valid for kF L → ∞
1579
+
1580
+ σ
1581
+ ∆ ≃ σ
1582
+
1583
+ �1 − h0
1584
+ ∆ + π2
1585
+ 2
1586
+ � ξ
1587
+ L
1588
+ �2 �
1589
+ 1 − s(ν)ξ
1590
+ L
1591
+
1592
+ 2∆
1593
+ h0
1594
+ − 1
1595
+ �2�
1596
+ � ,
1597
+ (67)
1598
+ with s(ν) = 1(−1) for ν = e(o).
1599
+ From here, we can
1600
+ clearly see that whereas the even-odd averaged quanti-
1601
+ ties (i.e., the semiclassical values) approach the asymp-
1602
+ totic limit as L−2, the energy difference between even
1603
+ and odd solutions (i.e., the amplitude of the oscillation
1604
+ around the semiclassical limit) decreases as L−3, and the
1605
+ solutions become degenerate in the limit L → ∞. On the
1606
+ other hand, the quasiparticle gap in the limit L → ∞ is
1607
+ renormalized to 2∆ren = 2 |∆ − h0|. Note that this gap
1608
+ renormalization is quite specific to this setup, and is not
1609
+ present, for instance, in the case of Ref. [32], where the
1610
+ magnetic region is normal and not superconducting, and
1611
+ in addition the system corresponds to a “short” SU-FM-
1612
+ SU junction with L < ξ, and therefore only few subgap
1613
+ states are allowed.
1614
+ Another feature of the weak-field regime is that the
1615
+ ABS require a minimal length Lmin to emerge in the sub-
1616
+
1617
+ 11
1618
+ gap region. This can be easily understood in terms of Eq.
1619
+ 49, where a minimal magnetic phase, represented by the
1620
+ product Lh0/ℏvF , must be accumulated in order to pro-
1621
+ duce an observable in-gap ABS. Finally, concerning the
1622
+ spin quantum number of the ground state, since none of
1623
+ the ABS cross EF , no quantum phase transitions are ex-
1624
+ pected according to the results of Sec. II B and the value
1625
+ of the ground state spin remains a spin-singlet Sz = 0.
1626
+ b. Intermediate field regime ∆ < h0 < 2∆: In this
1627
+ case the Zeeman field h0 is sufficiently strong to force the
1628
+ ABS to cross zero energy, eventually inducing quantum
1629
+ phase transitions (see middle panel in Fig. 3). The n-th
1630
+ critical value Lc,n can be obtained imposing the condition
1631
+ Eσ = 0 on the semiclassical approximation in Eq. (48),
1632
+
1633
+ c,n = ξ
1634
+ arctan
1635
+
1636
+ −s (ν) ∓
1637
+ �� h0
1638
+
1639
+ �2 − 1
1640
+
1641
+ + nπ
1642
+ �� h0
1643
+
1644
+ �2 − 1
1645
+ ,
1646
+ (68)
1647
+ with s(ν) = 1(−1) for ν = e(o).
1648
+ In this regime, the ABS follow the same asymp-
1649
+ totic behavior as in Eq.
1650
+ (67), approaching Eσ/∆ →
1651
+ σ (1 − h0/∆), although the overall subgap spectrum is
1652
+ completely different due to the closing of the gap, and
1653
+ due to the overlap of the E↑ and E↓ spectrum as L
1654
+ increases beyond the first critical Lc,0. In fact, in the
1655
+ regime L > Lc,0 the quasiparticle gap becomes com-
1656
+ pletely populated (and washed away) by subgap states.
1657
+ Moreover, we predict an accumulation of levels in the re-
1658
+ gion −∆ + h0 < E < ∆ − h0, which can eventually form
1659
+ a peak structure in the total density of states.
1660
+ c. Strong field regime 2∆ < h0: Finally, in this regime
1661
+ (see bottom panel in Fig. 3), the asymptotic dashed lines
1662
+ fall within the continuum and it is no longer possible to
1663
+ obtain an analytic expression for the ABS behavior in
1664
+ the limit L → ∞. As a result, the subgap ABS can be
1665
+ found anywhere in the subgap region −1 < Eσ/∆ < 1.
1666
+ In addition, we note that the minimal length required to
1667
+ observe in-gap ABS has reduced to Lmin ≈ 0.
1668
+ IV.
1669
+ SUMMARY AND CONCLUSIONS
1670
+ In this work we have analyzed the subgap electronic
1671
+ structure in the one dimensional SE-SU-FMI heterostruc-
1672
+ ture schematically depicted in Fig. 1, a novel physical
1673
+ system recently fabricated using molecular beam epitaxy
1674
+ techniques (MBE). The main motivation to study this
1675
+ type of hybrid systems is that, via a careful combina-
1676
+ tion of different materials, the emergent characteristics
1677
+ can be completely different from those of the individ-
1678
+ ual components, providing a way to build devices with
1679
+ tailored properties and specific functionalities. In partic-
1680
+ ular, much of the experimental effort has focused on the
1681
+ realization of topological superconducting phases host-
1682
+ ing Majorana zero modes, with possible applications in
1683
+ topological quantum computing [20, 21]. A distinguish-
1684
+ ing feature of these heterostructures is the coexistence
1685
+ of antagonistic superconductor and ferromagnetic insu-
1686
+ lating layers over a finite and arbitrary length L in a
1687
+ semiconductor wire, a combination that confers unique
1688
+ spectral properties which cannot be found in elemental
1689
+ materials in nature.
1690
+ In particular, we have modelled the hybrid struc-
1691
+ ture
1692
+ assuming
1693
+ non-interacting
1694
+ fermions
1695
+ in
1696
+ a
1697
+ one-
1698
+ dimensional single-channel nanowire under the effect of
1699
+ two proximity-induced interactions: a SU pairing and
1700
+ a space-dependent Zeeman exchange coupling [see Eqs.
1701
+ (1)-(3)].
1702
+ We have solved the associated Bogoliubov-de
1703
+ Gennes equations and, by imposing standard continuity
1704
+ conditions on the wave functions, we have obtained an
1705
+ equation [Eq. (28)] defining the subgap ABS spectrum
1706
+ of the device.
1707
+ This single equation encodes our main
1708
+ theoretical results. We stress that our approach is equiv-
1709
+ alent to other works using the scattering-matrix formal-
1710
+ ism. We have analytically solved Eq. (28) in two paradig-
1711
+ matic limits: the semiclassical limit (Sec. III A) and the
1712
+ Yu-Shiba-Rusinov limit, typical of atomic magnetic mo-
1713
+ ments interacting with a superconductor (Sec. III B). In
1714
+ both cases, we have been able to recover well-known ana-
1715
+ lytical results, providing important sanity checks for our
1716
+ theoretical results. As a consequence of the symmetries
1717
+ of the Hamiltonian (i.e., inversion x → −x and sz spin
1718
+ symmetries), it was possible to classify the solutions into
1719
+ even- and odd-symmetry, and with sz labels σ =↑, ↓. In
1720
+ particular, we note that the even-odd classification, aris-
1721
+ ing in the present case due to the inversion symmetry of
1722
+ the Hamiltonian, is nothing but the 1D analog of the clas-
1723
+ sification in angular momentum eigenstates ℓ occurring
1724
+ in 3D spherically-symmetric Hamiltonians [7, 25, 26].
1725
+ We have studied the subgap spectrum of ABS as a
1726
+ function of different parameters, namely: the length of
1727
+ the magnetic region (through the dimensionless parame-
1728
+ ter kF L), the strength of the Zeeman exchange induced
1729
+ by the FMI (parameter h0/∆), and the superconducting
1730
+ coherence length (parameter kF ξ). We stress that each
1731
+ one of these parameters could in principle (directly or in-
1732
+ directly) be controlled in experiments. However, due to
1733
+ its potential relevance for on-going experimental efforts,
1734
+ we have in particular focused our study on the evolution
1735
+ of the subgap spectrum as a function of the length L (i.e.,
1736
+ as it is probably the easiest parameter to vary in experi-
1737
+ ments), for fixed parameters kF ξ and h0/∆. The parame-
1738
+ ter L can be controlled by, e.g., changing the experimen-
1739
+ tal growing conditions of the semiconductor nanowires
1740
+ using the VLS growth method. In Figs. 2 and 3 we have
1741
+ analyzed the evolution of the subgap spectrum in terms
1742
+ of the parameter kF L for different values of h0/∆ and
1743
+ kF ξ. Roughly speaking, while kF ξ controls the “semi-
1744
+ classical vs quantum” operation regime of the device,
1745
+ and the magnitude of the even-odd energy separation,
1746
+ the parameter h0/∆ essentially controls the energy sep-
1747
+ aration of the E↑ and E↓ solutions, eventually enabling
1748
+ many interesting physical phenomena such as the possi-
1749
+
1750
+ 12
1751
+ bility to observe multiple ABS crossing zero-energy, the
1752
+ existence of multiple spin- and parity-changing quantum
1753
+ phase transitions in the device, quasiparticle gap renor-
1754
+ malization ∆ → ∆ren = |∆ − h0| in the limit of large
1755
+ kF L, etc.. An important conclusion here is that in order
1756
+ to experimentally observe a quantum phase transition,
1757
+ the condition h0 > ∆ must be fulfilled.
1758
+ Interpreting L as a “tunable” parameter has another
1759
+ theoretical advantage, as it enables to address the in-
1760
+ teresting fundamental question of how to connect two
1761
+ paradigmatic limits in SU-FM hybrid devices: the atomic
1762
+ limit (kF L → 0), where the physics is that of the well-
1763
+ known non-degenerate YSR states, and the ballistic limit
1764
+ (kF L ≫ 1) where the spectrum of the subgap ABS be-
1765
+ comes double degenerate. Until very recently, these lim-
1766
+ its were treated as disconnected from each other. In Ref.
1767
+ [32] this issue was addressed in the particular case of SU-
1768
+ FM-SU junctions in the limit L < ξ. Here we have revis-
1769
+ ited this intriguing question for a different setup where
1770
+ such constraint does not exist, and have studied the evo-
1771
+ lution of the subgap spectrum as a function of L. The
1772
+ abovementioned symmetry classification into even and
1773
+ odd solutions is critically important to allow the inter-
1774
+ pretation of the degeneracy in the limit kF L → ∞ as an
1775
+ “even-odd degeneracy”. At the same time, it enables to
1776
+ explain the degeneracy lifting in the limit L → 0, where
1777
+ only even states prevail in the subgap region of ener-
1778
+ gies.
1779
+ Using an approximate model of one-dimensional
1780
+ fermions with linearized dispersion, we have provided a
1781
+ simple picture where the even-odd degeneracy naturally
1782
+ emerges as a consequence of destructive interferences of
1783
+ terms e±i2kF x arising from single-particle backscattering
1784
+ mechanisms.
1785
+ The continuous evolution of the subgap spectrum as a
1786
+ function of kF L allows a better understanding of previ-
1787
+ ous experimental STM results on atomic magnetic adsor-
1788
+ bates on superconducting substrates, where the subgap
1789
+ YSR states are usually interpreted in terms of a point-like
1790
+ magnetic moment [27, 37–41]. While the delta-function
1791
+ limit is obviously a mathematical idealization, in terms
1792
+ of our model the observed YSR states can be rationalized
1793
+ assuming a finite value of kF L and a (more physically ap-
1794
+ pealing) finite value of the atomic local field h0. This is
1795
+ precisely the case if we note that for magnetic impurities
1796
+ (e.g. Fe, Co or Mn atoms) deposited on top of bulk metal-
1797
+ lic S surfaces (e.g., Pb or Al), the spatial extension of the
1798
+ short-ranged Zeeman field can be estimated as the size of
1799
+ the d-shell orbitals L ∼ 1 ˚A, while the Fermi wavevector
1800
+ of bulk superconductors (e.g., Pb) is kF ∼ 1−2×1010m−1
1801
+ (see Ref. [45]). This type of adsorbate/substrate combi-
1802
+ nation yields a parameter kF L ∼ 1, which is within the
1803
+ regime where we recover observable subgap states (see
1804
+ Figs. 2 and 3). On the other hand, in 1D semiconduc-
1805
+ tor heterostructures as those of Refs. 20 and 21, kF is
1806
+ usually much smaller than in metallic superconductors.
1807
+ Measurements of the number of carriers from the Hall
1808
+ conductance RH in 2D InGaAl quantum wells [46] yield
1809
+ the estimated value kF ∼ 2.2 × 107m−1, three orders of
1810
+ magnitude smaller as compared to bulk Pb. This much
1811
+ smaller value of kF allows for much larger, experimentally
1812
+ accessible values of L, while keeping values of h0 also
1813
+ within experimental reach. All together, this combina-
1814
+ tion makes these hybrid materials a much more versatile
1815
+ platform to control the spectrum of YSR/ABS subgap
1816
+ states.
1817
+ To characterize the quantum phase transitions occur-
1818
+ ring in the device, we have computed the value of the to-
1819
+ tal Sz using a spin version of the Friedel sum rule [see Eq.
1820
+ (44) and also Ref. [32]. We stress that these transitions
1821
+ are a generalization of the well-known “0-π” transition
1822
+ occurring in atomic Shiba impurities [22, 47] or quantum
1823
+ dots coupled to superconductors [48–50]. From this per-
1824
+ spective, the difference with respect to atomic systems
1825
+ is that instead of a single transition, actually multiple
1826
+ transitions can occur due to the finite extension L of
1827
+ the “impurity” and the many ABS states with different
1828
+ symmetry which can eventually cross below EF . Inter-
1829
+ estingly, we stress that the ocurrence of these quantum
1830
+ phase transitions can be tuned varying the length L.
1831
+ We now briefly address the effect of the Rashba spin-
1832
+ orbit interaction, which has been neglected in our work.
1833
+ As mentioned previously, this interaction was neglected
1834
+ to simplify the theoretical description of this (already
1835
+ quite complex and rich) problem. This interaction can
1836
+ drive the system into the topological superconductor
1837
+ class D [51, 52], hosting Majorana zero modes at the ends
1838
+ (see e.g., Ref. 34 for a related setup), and in that case
1839
+ we expect qualitative changes with respect to the results
1840
+ presented here. Consequently our results apply to exper-
1841
+ imental SE-SU-FMI systems where the spin-orbit energy
1842
+ term ESOC = α2
1843
+ Rm∗/2, with αR the Rashba parameter,
1844
+ is negligible compared to ∆ and h0.
1845
+ Finally, we consider the effect of disorder in this setup.
1846
+ This might be a relevant effect as a random disorder po-
1847
+ tential will eventually break the inversion symmetry of
1848
+ the model and might lift the predicted even-odd degen-
1849
+ eracy in the limit kF L ≫ 1. However, we believe the
1850
+ energy-lifting effect might be weak in epitaxially-grown
1851
+ samples, where disorder is a relatively small effect.
1852
+ ACKNOWLEDGMENTS
1853
+ This work was partially supported by CONICET un-
1854
+ der grant PIP 0792, UNLP under grant PID X497, and
1855
+ Agencia I+D+i under PICT 2017-2081, Argentina. AML
1856
+ is grateful to Liliana Arrachea for pointing out crucial
1857
+ bibliographic references.
1858
+
1859
+ 13
1860
+ [1] A. V. Balatsky, I. Vekhter, and J.-X. Zhu, Rev. Mod.
1861
+ Phys. 78, 373 (2006).
1862
+ [2] B. W. Heinrich, J. I. Pascual, and K. J. Franke, Prog.
1863
+ Surf. Sci. 93, 1 (2018).
1864
+ [3] R. Pawlak, S. Hoffman, J. Klinovaja, D. Loss, and
1865
+ E. Meyer, Prog. Part. Nucl. Phys. 107, 1 (2019).
1866
+ [4] D.-J. Choi, N. Lorente, J. Wiebe, K. von Bergmann,
1867
+ A. F. Otte, and A. J. Heinrich, Rev. Mod. Phys. 91,
1868
+ 041001 (2019).
1869
+ [5] L. Yu, Acta Phys. Sin. 21, 75 (1965).
1870
+ [6] H. Shiba, Prog. Theor. Phys. 40, 435 (1968).
1871
+ [7] A. I. Rusinov, Zh. Eksp. Teor. Fiz. Pisma. Red. 9, 146
1872
+ (1968), [JETP Lett. 9, 85 (1969)].
1873
+ [8] S. Nadj-Perge, I. K. Drozdov, B. A. Bernevig, and
1874
+ A. Yazdani, Phys. Rev. B 88, 020407(R) (2013).
1875
+ [9] J. Klinovaja, P. Stano, A. Yazdani, and D. Loss, Phys.
1876
+ Rev. Lett. 111, 186805 (2013).
1877
+ [10] J. Li, H. Chen, I. K. Drozdov, A. Yazdani, B. A.
1878
+ Bernevig, and A. H. MacDonald, Phys. Rev. B 90,
1879
+ 235433 (2014).
1880
+ [11] S. Nadj-Perge, I. K. Drozdov, J. Li, H. Chen, S. Jeon,
1881
+ J. Seo, A. H. MacDonald, B. A. Bernevig, and A. Yaz-
1882
+ dani, Science 346, 602 (2014).
1883
+ [12] M. Ruby, F. Pientka, Y. Peng, F. von Oppen, B. W.
1884
+ Heinrich, and K. J. Franke, Phys. Rev. Lett. 115, 197204
1885
+ (2015).
1886
+ [13] B. E. Feldman, M. T. Randeria, J. Li, S. Jeon, Y. Xie,
1887
+ Z. Wang, I. K. Drozdov, B. Andrei Bernevig, and A. Yaz-
1888
+ dani, Nat. Phys. 13, 286 EP (2016), article.
1889
+ [14] R. Pawlak, M. Kisiel, J. Klinovaja, T. Meier, S. Kawai,
1890
+ T. Glatzel, D. Loss, and E. Meyer, npj Quantum Inf. 2,
1891
+ 16035 (2016).
1892
+ [15] M. Ruby, B. W. Heinrich, Y. Peng, F. von Oppen, and
1893
+ K. J. Franke, Nano Lett. 17, 4473 (2017).
1894
+ [16] H. Kim,
1895
+ A. Palacio-Morales,
1896
+ T. Posske,
1897
+ L. R´ozsa,
1898
+ K. Palot´as, L. Szunyogh, M. Thorwart, and R. Wiesen-
1899
+ danger, Sci. Adv. 4, eaar5251 (2018).
1900
+ [17] B. J¨ack, Y. Xie, and A. Yazdani, Nat. Rev. Phys. 3, 541
1901
+ (2021).
1902
+ [18] F. S. Bergeret, A. F. Volkov, and K. B. Efetov, Rev. Mod.
1903
+ Phys. 77, 1321 (2005).
1904
+ [19] M. Eschrig, Phil. Trans. R. Soc. A 376, 20150149 (2018).
1905
+ [20] Y. Liu, S. Vaitiek˙enas, S. Mart´ı-S´anchez, C. Koch,
1906
+ S. Hart, Z. Cui, T. Kanne, S. A. Khan, R. Tanta,
1907
+ S. Upadhyay, M. E. Cachaza, C. M. Marcus, J. Ar-
1908
+ biol, K. A. Moler, and P. Krogstrup, Nano Lett. 20, 456
1909
+ (2020).
1910
+ [21] S. Vaitiek˙enas, Y. Liu, P. Krogstrup, and C. M. Marcus,
1911
+ Nat. Phys. 17, 43 (2021).
1912
+ [22] A. Sakurai, Prog. Theor. Phys. 44, 1472 (1970).
1913
+ [23] A. Costa, J. Fabian, and D. Kochan, Phys. Rev. B 98,
1914
+ 134511 (2018).
1915
+ [24] A. I. Rusinov, Zh. Eksp. Teor. Fiz. 56, 2047 (1969).
1916
+ [25] M. E. Flatt´e and J. M. Byers, Phys. Rev. B 56, 11213
1917
+ (1997).
1918
+ [26] L. Arrachea, Phys. Rev. B 104, 134515 (2021).
1919
+ [27] S.-H. Ji, T. Zhang, Y.-S. Fu, X. Chen, X.-C. Ma, J. Li,
1920
+ W.-H. Duan, J.-F. Jia, and Q.-K. Xue, Phys. Rev. Lett.
1921
+ 100, 226801 (2008).
1922
+ [28] M. Ruby, Y. Peng, F. von Oppen, B. W. Heinrich, and
1923
+ K. J. Franke, Phys. Rev. Lett. 117, 186801 (2016).
1924
+ [29] D.-J. Choi, C. Rubio-Verd´u, J. de Bruijckere, M. M.
1925
+ Ugeda, N. Lorente, and J. I. Pascual, Nat. Commun. 8,
1926
+ 15175 (2017), article.
1927
+ [30] F. Konschelle, F. S. Bergeret, and I. V. Tokatly, Phys.
1928
+ Rev. Lett. 116, 237002 (2016).
1929
+ [31] F. Konschelle, I. V. Tokatly, and F. S. Bergeret, Phys.
1930
+ Rev. B 94, 014515 (2016).
1931
+ [32] M. Rouco, I. V. Tokatly, and F. S. Bergeret, Phys. Rev.
1932
+ B 99, 094514 (2019).
1933
+ [33] S. Takei, B. M. Fregoso, H.-Y. Hui, A. M. Lobos, and
1934
+ S. Das Sarma, Phys. Rev. Lett. 110, 186803 (2013).
1935
+ [34] J. D. Sau and P. M. R. Brydon, Phys. Rev. Lett. 115,
1936
+ 127003 (2015).
1937
+ [35] A. C. Hewson, The Kondo Problem to Heavy Fermions
1938
+ (Cambridge University Press, Cambridge, 1993).
1939
+ [36] K. Duncan and B. Gy¨orffy, Ann. Phys. 298, 273 (2002).
1940
+ [37] A. Yazdani, B. A. Jones, C. P. Lutz, M. F. Crommie, and
1941
+ D. M. Eigler, Science 275, 1767 (1997).
1942
+ [38] M. Iavarone, G. Karapetrov, J. Fedor, D. Rosenmann,
1943
+ T. Nishizaki, and N. Kobayashi, J. Phys.: Condens. Mat-
1944
+ ter 22, 015501 (2010).
1945
+ [39] S.-H. Ji, T. Zhang, Y.-S. Fu, X. Chen, J.-F. Jia, Q.-K.
1946
+ Xue, and X.-C. Ma, App. Phys. Lett. 96, 073113 (2010).
1947
+ [40] J. Bauer, J. I. Pascual, and K. J. Franke, Phys. Rev. B
1948
+ 87, 075125 (2013).
1949
+ [41] N. Hatter, B. W. Heinrich, M. Ruby, J. I. Pascual, and
1950
+ K. J. Franke, Nat. Commun. 6, 8988 (2015).
1951
+ [42] L. D. Landau and E. M. Lifshitz, Quantum Mechanics:
1952
+ non-relativistic theory (Pergamon Press, Oxford, 1958).
1953
+ [43] T. Bortolin, A. Iucci, and A. M. Lobos, Phys. Rev. B
1954
+ 100, 155111 (2019).
1955
+ [44] T. Giamarchi, Quantum Physics in One Dimension (Ox-
1956
+ ford University Press, Oxford, 2003).
1957
+ [45] N. W. Ashcroft and N. D. Mermin, Solid state physics
1958
+ (Holt, Rinehart and Winston, New York, 1976).
1959
+ [46] M. Kjærgaard, Proximity Induced Superconducting Prop-
1960
+ erties in One and Two Dimensional Semiconductors: To-
1961
+ wards Topological States of Matter, Ph.D. thesis (2015).
1962
+ [47] K. J. Franke, G. Schulze, and J. I. Pascual, Science 332,
1963
+ 940 (2011).
1964
+ [48] J. Bauer, A. Oguri, and A. C. Hewson, J. Phys.: Con-
1965
+ dens. Matter 19, 486211 (2007).
1966
+ [49] R.
1967
+ S.
1968
+ Deacon,
1969
+ Y.
1970
+ Tanaka,
1971
+ A.
1972
+ Oiwa,
1973
+ R.
1974
+ Sakano,
1975
+ K. Yoshida, K. Shibata, K. Hirakawa, and S. Tarucha,
1976
+ Phys. Rev. Lett. 104, 076805 (2010).
1977
+ [50] E. J. H. Lee, X. Jiang, M. Houzet, R. Aguado, C. M.
1978
+ Lieber, and S. De Franceschi, Nat. Nanotechnol 9, 79
1979
+ (2014).
1980
+ [51] A. Altland and M. R. Zirnbauer, Phys. Rev. B 55, 1142
1981
+ (1997).
1982
+ [52] S. Ryu, A. P. Schnyder, A. Furusaki, and A. W. W. Lud-
1983
+ wig, New J. Phys. 12, 065010 (2010).
1984
+
3NE2T4oBgHgl3EQfjQer/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
3tE2T4oBgHgl3EQf6AiT/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e61989bde6a8de23d078be9e989299362ca0c0115f79bc139928f8216dec82a
3
+ size 2293805
49E0T4oBgHgl3EQfvgE1/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:695953e4fd98edb71020b54214bd2d1a6c191658430b4e3e357fb003fc089ad5
3
+ size 797643
49E1T4oBgHgl3EQfAwK8/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31a1dc2d0136f8e7227dde9f0aa6d0e6db67fd2973de1126e69f34ad6203c767
3
+ size 4980781
59E0T4oBgHgl3EQfewCK/content/2301.02395v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:759598288656ba3837be1ff4c853c5d457833be1ef600bf855486bf956aa6b6b
3
+ size 2858073
59E0T4oBgHgl3EQfewCK/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e5780af18254c405208ff382a6a2702f2aca827e809275f2145b858def9d1380
3
+ size 312751
69AyT4oBgHgl3EQfQfbA/content/tmp_files/2301.00048v1.pdf.txt ADDED
@@ -0,0 +1,1233 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ On the gate-error robustness of variational quantum algorithms
2
+ Daniil Rabinovich,1 Ernesto Campos,1 Soumik Adhikary,1 Ekaterina
3
+ Pankovets,1, 2 Dmitry Vinichenko,1, 3 and Jacob Biamonte4
4
+ 1Skolkovo Institute of Science and Technology, Moscow, Russian Federation
5
+ 2Moscow Institute of Physics and Technology, Moscow, Russian Federation
6
+ 3Moscow Engineering Physics Institute, Moscow, Russian Federation
7
+ 4Beijing Institute of Mathematical Sciences and Applications, Beijing, China
8
+ Variational algorithms are designed to work within the limitations of contemporary devices and
9
+ suffer from performance limiting errors.
10
+ Here we identify an experimentally relevant model for
11
+ gate errors, natural to variational quantum algorithms. We study how a quantum state prepared
12
+ variationally decoheres under this noise model, which manifests as a perturbation to the energy
13
+ approximation in the variational paradigm. A perturbative analysis of an optimized circuit allows
14
+ us to determine the noise threshold for which the acceptance criteria imposed by the stability
15
+ lemma remains satisfied. We benchmark the results against the variational quantum approximate
16
+ optimization algorithm for 3-SAT instances and unstructured search with up to 10 qubits and 30
17
+ layers. Finally, we observe that errors in certain gates have a significantly smaller impact on the
18
+ quality of the prepared state. Motivated by this, we show that it is possible to reduce the execution
19
+ time of the algorithm with minimal to no impact on the performance.
20
+ I.
21
+ INTRODUCTION
22
+ Noisy Intermediate Scale Quantum (NISQ) quantum
23
+ computing [1] suffers from limited coherence times and
24
+ opeartion precision [2–5]. In practice we are severely lim-
25
+ ited by the number of qubits and circuit depths that
26
+ one may implement with reasonable fidelity.
27
+ This has
28
+ piratical implications in that it limits contemporary ex-
29
+ perimental demonstrations. A host of theoretical results
30
+ are now emerging, leading to improved understanding
31
+ of the use of random circuit sampling as the basis of a
32
+ scalable experimental violation of the extended Church-
33
+ Turing thesis [6] and on the complexity analysis of NISQ
34
+ [7]. The variational model of quantum computation is
35
+ designed to work within these practical limitations [8–
36
+ 10]. More generally, the variational model is known to
37
+ be computationally universal, yet these results are highly
38
+ idealized and do not account for noise [11].
39
+ Reminiscent of machine learning, a variational algo-
40
+ rithm makes use of a short parameterized quantum cir-
41
+ cuit, known as ansatz, in which parameters are itera-
42
+ tively tuned to minimize a cost function in a quantum-to-
43
+ classical feedback loop [12]. The cost function is typically
44
+ given in the form of the expectation of a so called prob-
45
+ lem Hamiltonian; where the ground state of the problem
46
+ Hamiltonian encodes the solution of a given problem in-
47
+ stance. Thus, by the way of cost function (energy) min-
48
+ imization, a variational algorithm attempts to approx-
49
+ imate the ground state of a given Hamiltonian.
50
+ This
51
+ strategy, however, does not provide us with a guarantee
52
+ in regards to the quality of the approximate solution,
53
+ where the latter is typically quantified as the overlap
54
+ between the state prepared by the ansatz and the true
55
+ ground state. Nevertheless, the overlap can be bounded.
56
+ It has been shown using the stability lemma that the
57
+ bounds can be directly related to the energy, thus allow-
58
+ ing us to determine the energy threshold (upper bound)
59
+ required to guarantee a fixed minimum overlap. We call
60
+ this the acceptance threshold; a state with energy below
61
+ this threshold is said to be accepted by the algorithm
62
+ [11].
63
+ Variational algorithms by their design alleviate the ef-
64
+ fects of certain systematic limitations of NISQ devices.
65
+ Nevertheless, variational algorithms are not immune to
66
+ stochastic noise. While there exist some evidence that
67
+ variational algorithms can in fact benefit from certain
68
+ level of stochastic noise [13], in general, it is detrimental
69
+ to the performance; stochastic noise leads to decoherence
70
+ thus typically reducing solution quality.
71
+ In this paper we study the extent to which errors, in the
72
+ form of parameter alterations, affects the performance of
73
+ variational algorithms.
74
+ We analytically show that the
75
+ shift in energy varies quadratically with the strength of
76
+ noise (for small amounts of noise). We demonstrate this
77
+ numerically for variational quantum approximate opti-
78
+ misation in two common problems—3-SAT [14] and un-
79
+ structured search [15, 16]. Furthermore we also found the
80
+ performance to be more resilient to alterations in certain
81
+ parameters. With that in mind we propose avenues to
82
+ potentially improve performance and reduce the execu-
83
+ tion time of variational quantum algorithms.
84
+ II.
85
+ PRELIMINARIES
86
+ A.
87
+ Variational Quantum Approximate
88
+ Optimization
89
+ The quantum approximate optimization algorithm
90
+ (QAOA) [17], originally designed to approximately solve
91
+ combinatorial optimization problems [14, 17–28], consists
92
+ of ansatze circuits expressive enough to (in theory) emu-
93
+ late any quantum cirucuit [19, 20].
94
+ Consider a pseudo-Boolean function C : {0, 1}×n → R,
95
+ the objective of the algorithm is to approximate a bit
96
+ string that minimizes C. To accomplish this, C is first
97
+ arXiv:2301.00048v1 [quant-ph] 30 Dec 2022
98
+
99
+ 2
100
+ encoded as a problem Hamiltonian H, diagonal in the
101
+ computational basis. The ground state H encodes the
102
+ solution to the problem; in other words QAOA searches
103
+ for a solution |g⟩ such that ⟨g|H|g⟩ = min H.
104
+ The algorithm begins with an ansatz state |ψp(γ, β)⟩—
105
+ prepared by a circuit of depth p — parameterized as:
106
+ |ψp(γ, β)⟩ =
107
+ p
108
+
109
+ k=1
110
+ e−iβkHxe−iγkH |+⟩⊗n ,
111
+ (1)
112
+ with real parameters γk ∈ [0, 2π), βk ∈ [0, π).
113
+ Here
114
+ Hx = �n
115
+ j=1 Xj is the standard one-body mixer Hamil-
116
+ tonian with Pauli matrix Xj applied to the j-th qubit.
117
+ The cost function is given by the expectation of the prob-
118
+ lem Hamiltonian with respect to the ansatz state. The
119
+ algorith minimizes this cost function to output:
120
+ E∗ = minγ,β ⟨ψp(γ, β)| H |ψp(γ, β)⟩
121
+ (2)
122
+ γ∗, β∗ ∈ arg minγ,β ⟨ψp(γ, β)| H |ψp(γ, β)⟩
123
+ (3)
124
+ Here, |ψp(γ∗, β∗)⟩ is the approximate ground state of
125
+ H and hence the approximate solution to C. Indeed, the
126
+ quality of the approximation, quantified as the overlap
127
+ between the true solution and the approximate solution,
128
+ is not known a priori from (2).
129
+ Nevertheless one can
130
+ establish bounds on this quantity using the so called sta-
131
+ bility lemma.
132
+ B.
133
+ Stability lemma
134
+ The stability lemma states that if |g⟩ is the true ground
135
+ state of H with energy Eg and ∆ is the spectral gap
136
+ (the difference between the ground state energy and the
137
+ energy of the first excited state) the following relation
138
+ holds [11, 29]:
139
+ 1 − E∗ − Eg
140
+
141
+ ≤ |⟨ψp(γ∗, β∗)|g⟩|2 ≤ 1 − E∗ − Eg
142
+ Em − Eg
143
+ (4)
144
+ where Em is the maximum eigenvalue of H.
145
+ Thus to
146
+ guarantee a non-trivial overlap one must ensure that
147
+ E∗ ≤ Eg + ∆. We call the latter the acceptance con-
148
+ dition.
149
+ III.
150
+ VARIATIONAL QUANTUM ALGORITHMS
151
+ IN THE PRESENCE OF REALISTIC GATE
152
+ ERRORS
153
+ Implementation of unitary operations depends signif-
154
+ icantly on the considered hardware. However, typically
155
+ the implementation makes use of electromagnetic pulses,
156
+ such as in superconducting quantum computers [30, 31],
157
+ neutral atom based quantum computers [32, 33], and
158
+ trapped ion based quantum computers [34, 35].
159
+ Such
160
+ pulses can change the population of the energy levels
161
+ that constitute a qubit or introduce phases to the quan-
162
+ tum amplitudes, thus controlling the state of the qubits.
163
+ Consequently, the main contribution to gate errors comes
164
+ from variation in pulse shaping, meaning that amplitude
165
+ and timing of electromagnetic pulse can stochasticaly
166
+ vary.
167
+ In certain experimental setups, such as ground
168
+ state ion qubits, where entangling operations are per-
169
+ formed using the radial phonon modes [36], the variabil-
170
+ ity in pulse shaping is the main source of gate errors.
171
+ Angles of rotation in a typical gate operation depend
172
+ on time averaged intensity I(t) of the electromagnetic
173
+ pulse; θ ∝
174
+
175
+ I(t)dt. Thus, variations in the pulse shap-
176
+ ing lead to stochastic deviations of the angles of rota-
177
+ tions from the desired values. In other words, if a cir-
178
+ cuit is composed of the parameterised gates {Uk(θk)}k;
179
+ θ ∈ [0, 2π) and one tries to prepare a state |ψ(θ)⟩ =
180
+
181
+ k Uk(θk) |ψ0⟩, a different state
182
+ |ψ(θ + δθ)⟩ =
183
+
184
+ k
185
+ U(θk + δθk) |ψ0⟩ ,
186
+ (5)
187
+ is prepared instead due to the presence of errors. No-
188
+ tice here that the perturbation δθ to the parameters is
189
+ stochastic and is sampled with a certain probability den-
190
+ sity p(δθ). This implies that the prepared state can be
191
+ described by an ensemble {|ψ(θ + δθ)⟩ , p(δθ)}, which we
192
+ can equivalently view as a density matrix
193
+ ρ(θ) =
194
+
195
+ |ψ(θ + δθ)⟩⟨ψ(θ + δθ)|p(δθ)d(δθ).
196
+ (6)
197
+ Eq. (6) represents a noise model native to the vari-
198
+ ational paradigm of quantum computing. For the rest
199
+ of this paper we systematically study the effect of this
200
+ noise model on the performance of QAOA for instances
201
+ of 3-SAT and the unstructured search problem (see ap-
202
+ pendix A for more details on the considered problems).
203
+ In particular we study the energy perturbation around
204
+ E∗ in different scenarios subsequently recovering the
205
+ strength of noise under which the acceptance condition
206
+ continues to be satisfied.
207
+ IV.
208
+ RESULTS
209
+ A.
210
+ Perturbative analysis in presence of gate errors
211
+ Consider a problem Hamiltonian H and a variational
212
+ ansatz |ψ(θ)⟩ = U1(θ1) . . . Uq(θq) |ψ0⟩ used to mini-
213
+ mize H. Here the gates Uk(θk) have the form:
214
+ Uk(θk) = eiAkθk, A2
215
+ k = 1,
216
+ (7)
217
+ A typical example of such an ansatz is the checkerboard
218
+ ansatz, with Mølmer-Sørensen (MS) gates as the entan-
219
+ gling two qubit gates. Nevertheless, any quantum circuit
220
+ can admit a decomposition in terms of operations that
221
+ satisfy (7); this adds generality to this assumption.
222
+
223
+ 3
224
+ In the presence of gate errors the prepared quantum
225
+ state decoheres as |ψ(θ)��� → ρ(θ) as per (6). To obtain
226
+ the analytic form of ρ(θ) we first note that
227
+ Uk(θk + δθk) = Uk(θk)Uk(δθk)
228
+ = cos δθkUk(θk) + sin δθkUk
229
+
230
+ θk + π
231
+ 2
232
+
233
+ .
234
+ (8)
235
+ This follows directly from (7). Therefore we get:
236
+ |ψ(θ + δθ)⟩⟨ψ(θ + δθ)| =
237
+ 1
238
+
239
+ k1,...,kq,m1,...,mq=0
240
+ (cos2 δθ1 tank1+m1 δθ1) . . . (cos2 δθq tankq+mq δθq)|ψk1...kq⟩⟨ψm1...mq|, (9)
241
+ where
242
+ |ψk1...kq⟩ = U1(θ1 + k1
243
+ π
244
+ 2 ) . . . Uq(θq + kq
245
+ π
246
+ 2 ) |ψ0⟩ .
247
+ (10)
248
+ Here we make three realistic assumptions—(a) pertur-
249
+ bations to all the angles are independent, (b) average
250
+ perturbation ⟨δθk⟩ = 0 and (c) the distribution p(δθk)
251
+ vanishes quickly outside the range (−σk, σk); that is, the
252
+ error is localized on the scale σk ≪ 1. Note that if as-
253
+ sumption (b) does not hold, one can always shift the
254
+ parameters as θ → θ + ⟨δθ⟩.
255
+ Substituting (9) in (6) we arrive at the expression:
256
+ ρ(θ) = |ψ(θ)⟩⟨ψ(θ)| + δρ,
257
+ (11)
258
+ where
259
+ δρ ≈ −
260
+ q
261
+
262
+ k=1
263
+ ak|ψ(θ)⟩⟨ψ(θ)|+
264
+ q
265
+
266
+ k=1
267
+ ak|ψk⟩⟨ψk|+o(σ2
268
+ k). (12)
269
+ Here |ψk⟩ = |ψ00...1...00⟩ with 1 placed in the k-th posi-
270
+ tion, and
271
+ ak ≡ ⟨sin2 δθk⟩ =
272
+
273
+ sin2 δθkp(δθk)d(δθk) ∼ σ2
274
+ k.
275
+ (13)
276
+ Notice that the derivation above does not require θ to
277
+ be a minimum of the noiseless cost function.
278
+ Let us
279
+ now assume that θ∗ is a vector of parameters such that
280
+ |ψ(θ∗)⟩ approximates the ground state of H. The noise
281
+ induced energy perturbation around the optimal energy
282
+ E∗ is given as:
283
+ δE = Tr(ρ(θ∗)H) − ⟨ψ(θ∗)| H |ψ(θ∗)⟩
284
+ ≤ (Em − E∗)
285
+
286
+ k
287
+ ak.
288
+ (14)
289
+ For the simplest case where each parameter is sampled
290
+ from the same distribution (σk = σ) we can roughly es-
291
+ timate:
292
+ δE ≤ qσ2(Em − E∗).
293
+ (15)
294
+ Thus, requesting an energy threshold E ≤ Eg + ∆, we
295
+ conclude that for σ <∼
296
+
297
+ ∆ − (E∗ − Eg)
298
+ q(Em − E∗)
299
+ the acceptance
300
+ condition is still satisfied.
301
+ While our perturbative analysis holds for all varia-
302
+ tional algorithms, we substantiate our findings numer-
303
+ ically using QAOA. In particular we solve instances of
304
+ 3-SAT and unstructured search problems to study the
305
+ behaviour of energy perturbation around E∗ caused by
306
+ the presence of gate errors.
307
+ 1.
308
+ Constant perturbation
309
+ We begin with a simplified version of the noise model
310
+ proposed in (6). We ran QAOA for 100 uniformly gen-
311
+ erated 3-SAT instances of 6,8, and 10 variables with 26,
312
+ 34 and 42 clauses respectively.
313
+ All the instances were
314
+ selected to have a unique satisfying assignment. The in-
315
+ stances were minimized by QAOA sequences of 15, 25
316
+ and 30 layers respectively in order to obtain expected
317
+ values well below the energy gap. In order to numeri-
318
+ cally verify the behaviour of the energy perturbation, we
319
+ vary all optimal parameters by a constant angle δ. Fig-
320
+ ure 1 illustrates the shift in the energy for the minimized
321
+ instances, which can be seen to have a quadratic depen-
322
+ dence of the perturbed energy δE with respect to the
323
+ shift δ. This is natural to expect since the parameters
324
+ deviate from the local minimum, where linear contribu-
325
+ tion must have vanished (a rigorous expression showing
326
+ the quadratic behavior is derived in appendix B).
327
+ Similar to the case of 3-SAT, for the problem of un-
328
+ structured search we perturb optimal parameters of the
329
+ circuit by an angle δ and plot corresponding energy in
330
+ Fig. 2. Again, as expected, for small values of δ the en-
331
+ ergy perturbation is quadratic which comes from the fact
332
+ that the deviation happens around the minimum.
333
+
334
+ 4
335
+ 0.0000
336
+ 0.0025
337
+ 0.0050
338
+ 0.0075
339
+ 0.0100
340
+ 0.0125
341
+ 0.0150
342
+ 0.0175
343
+ 0.0200
344
+ δ
345
+ −0.02
346
+ 0.00
347
+ 0.02
348
+ 0.04
349
+ 0.06
350
+ 0.08
351
+ 0.10
352
+ 0.12
353
+ δE
354
+ 6.0 qubits
355
+ 71.8δ2
356
+ 8.0 qubits
357
+ 160.5δ2
358
+ 10.0 qubits
359
+ 250.3δ2
360
+ FIG. 1. Energy shift obtained by perturbing the ansatz state
361
+ as |ψp(γ∗ + δ, β∗ + δ)⟩. The curves illustrate averages over
362
+ 100 uniformly generated 3-SAT instances of 6, 8 and 10 qubits
363
+ with clause to variable ratio of 4.2 and unique satisfying as-
364
+ signment. The error bars depict standard error. Polynomial
365
+ fits of data indicates δ ∈ [0, 0.02] follow quadratic curves.
366
+ 0.00
367
+ 0.01
368
+ 0.02
369
+ 0.03
370
+ 0.04
371
+ 0.05
372
+ 0.06
373
+ 0.07
374
+ 0.08
375
+ δ
376
+ 0.0
377
+ 0.2
378
+ 0.4
379
+ 0.6
380
+ 0.8
381
+ 1.0
382
+ δE
383
+ 6 qubits
384
+ 208.9δ2
385
+ 8 qubits
386
+ 1228.4δ2
387
+ 10 qubits
388
+ 4664.2δ2
389
+ FIG. 2.
390
+ Energy shift for the problem of unstructured
391
+ search
392
+ obtained
393
+ by
394
+ perturbing
395
+ of
396
+ the
397
+ ansatz
398
+ state
399
+ as
400
+ |ψp(γ∗ + δ, β∗ + δ)⟩.
401
+ Polynomial fits for data points of 6,
402
+ 8 and 10 qubits follow quadratic curves in the ranges δ ∈
403
+ [0, 0.02], [0, 0.01], [0, 0.008] respectively.
404
+ 2.
405
+ Stochastic perturbation
406
+ We now consider the complete noise model in (6) and
407
+ verify our analytical prediction as shown in (15).
408
+ For
409
+ each 3-SAT instance, we randomly sample perturbations
410
+ δ to each of the gates from a uniform distribution on the
411
+ interval (−σ, σ) and average the obtained energy. Then
412
+ we average energies over instances of the same number
413
+ of qubits as depicted in Fig. 3. It is seen that for small
414
+ values of σ the behaviour is quadratic as per (15). It is
415
+ seen, that the value σ ∼ 0.075 could never violate the
416
+ acceptance criteria, as corresponding energy error never
417
+ exceeds the gap ∆ ≥ 1. For smaller number of qubits
418
+ and gates the threshold value of σ increases.
419
+ For unstructured search, we average the energy over
420
+ δ sampled for each gate from the uniform distribution
421
+ 0.00
422
+ 0.05
423
+ 0.10
424
+ 0.15
425
+ 0.20
426
+ 0.25
427
+ 0.30
428
+ 0.35
429
+ 0.40
430
+ σ
431
+ 0
432
+ 1
433
+ 2
434
+ 3
435
+ 4
436
+ 5
437
+ δE
438
+ 6.0 qubits
439
+ 61.0σ2
440
+ 8.0 qubits
441
+ 118.5σ2
442
+ 10.0 qubits
443
+ 172.8σ2
444
+ FIG. 3. Average energy shift of 100 uniformly generated 3-
445
+ SAT instances of 6, 8 and 10 qubits with clause to variable
446
+ ratio of 4.2 and unique satisfying assignment. The shifts are
447
+ obtained by the perturbation of γ∗, β∗ by δ uniformly sam-
448
+ pled from the range (−σ, σ). Error bars depict standard error.
449
+ Polynomial fits of data indicates σ ∈ [0, 0.1] follow quadratic
450
+ curves.
451
+ (−σ, σ). We again recover quadratic behaviour in σ, as
452
+ depicted in Fig. 4.
453
+ It is seen that the same threshold
454
+ σ ∼ 0.075 now increases energy by no more then 0.6,
455
+ which guaranties 40% overlap with the target state.
456
+ 0.00
457
+ 0.04
458
+ 0.08
459
+ 0.12
460
+ 0.16
461
+ 0.20
462
+ 0.24
463
+ 0.28
464
+ σ
465
+ 0.0
466
+ 0.2
467
+ 0.4
468
+ 0.6
469
+ 0.8
470
+ 1.0
471
+ δE
472
+ 6 qubits
473
+ 21.4σ2
474
+ 8 qubits
475
+ 60.3σ2
476
+ 10 qubits
477
+ 133.2σ2
478
+ FIG. 4.
479
+ Average energy for the problem of unstructured
480
+ search obtained by the perturbation of γ∗, β∗ by δ uni-
481
+ formly sampled from the range (−σ, σ).
482
+ Error bars de-
483
+ pict standard error.
484
+ Polynomial fits of data points of 6,
485
+ 8 and 10 qubits follow quadratic curves in the ranges σ ∈
486
+ [0, 0.1], [0, 0.07], [0, 0.05], respectively.
487
+ B.
488
+ Perturbation to individual parameters
489
+ Here we consider a modified version of (6), where pa-
490
+ rameters are perturbed one at a time while the rest are
491
+ kept intact. Effect of this model on the energy is illus-
492
+ trated in Figures 5 and 6. The results are numerical and
493
+ are yet to be explained analytically. We observe that per-
494
+ turbations to certain angles have a significantly smaller
495
+
496
+ 5
497
+ tbh
498
+ γ
499
+ β
500
+ n = 6
501
+ p = 8
502
+ 1
503
+ 2
504
+ 3
505
+ 4
506
+ 5
507
+ 6
508
+ 7
509
+ 8
510
+ k
511
+ 0.0105
512
+ 0.0110
513
+ 0.0115
514
+ 0.0120
515
+ 0.0125
516
+ ⟨H⟩
517
+ δ=0.0
518
+ δ=0.02
519
+ δ=0.05
520
+ δ=0.08
521
+ δ=0.1
522
+ 1
523
+ 2
524
+ 3
525
+ 4
526
+ 5
527
+ 6
528
+ 7
529
+ 8
530
+ k
531
+ 0.02
532
+ 0.04
533
+ 0.06
534
+ 0.08
535
+ 0.10
536
+ 0.12
537
+ ⟨H⟩
538
+ δ=0.0
539
+ δ=0.02
540
+ δ=0.05
541
+ δ=0.08
542
+ δ=0.1
543
+ n = 8
544
+ p = 15
545
+ 2
546
+ 4
547
+ 6
548
+ 8
549
+ 10
550
+ 12
551
+ 14
552
+ k
553
+ 0.0190
554
+ 0.0195
555
+ 0.0200
556
+ 0.0205
557
+ 0.0210
558
+ ⟨H⟩
559
+ δ=0.0
560
+ δ=0.02
561
+ δ=0.05
562
+ δ=0.08
563
+ δ=0.1
564
+ 2
565
+ 4
566
+ 6
567
+ 8
568
+ 10
569
+ 12
570
+ 14
571
+ k
572
+ 0.05
573
+ 0.10
574
+ 0.15
575
+ 0.20
576
+ ⟨H⟩
577
+ δ=0.0
578
+ δ=0.02
579
+ δ=0.05
580
+ δ=0.08
581
+ δ=0.1
582
+ n = 10
583
+ p = 25
584
+ 0
585
+ 3
586
+ 6
587
+ 9
588
+ 12
589
+ 15
590
+ 18
591
+ 21
592
+ 24
593
+ k
594
+ 0.0850
595
+ 0.0855
596
+ 0.0860
597
+ 0.0865
598
+ 0.0870
599
+ ⟨H⟩
600
+ δ=0.0
601
+ δ=0.02
602
+ δ=0.05
603
+ δ=0.08
604
+ δ=0.1
605
+ 0
606
+ 3
607
+ 6
608
+ 9
609
+ 12
610
+ 15
611
+ 18
612
+ 21
613
+ 24
614
+ k
615
+ 0.10
616
+ 0.15
617
+ 0.20
618
+ 0.25
619
+ 0.30
620
+ ⟨H⟩
621
+ δ=0.0
622
+ δ=0.02
623
+ δ=0.05
624
+ δ=0.08
625
+ δ=0.1
626
+ FIG. 5. Energy ⟨H⟩ = ⟨ψ(θ∗ + δθ)| H |ψ(θ∗ + δθ)⟩ from the unstructured search problem, where βk (right column) or γk (left
627
+ column), from the k-th layer, are perturbed.
628
+ effect on the energy.
629
+ Thus we can infer that reducing
630
+ the value of such angles would not have a significant ef-
631
+ fect on performance but will reduce the execution time of
632
+ the algorithm, that is texec = �p
633
+ k=1 βk + γk. Conversely,
634
+ we could limit the execution time as texec ≤ tmax and
635
+ increase the number of layers, since
636
+ min ⟨ψp| H |ψp⟩ ≥ min ⟨ψp+1| H |ψp+1⟩
637
+ (16)
638
+ for the same tmax.
639
+ Reducing the execution time is important to quantum
640
+ algorithms, since variational parameters are proportional
641
+ to the time required to execute a gate experimentally.
642
+ NISQ era devices suffers from limited coherence, thus
643
+ reducing execution times can lead to more efficient hard-
644
+ ware utilization [37, 38]. We test these ideas in the setting
645
+ of unstructured search, as depicted in Fig. 7. Here we
646
+ show the optimized QAOA energies for 6 qubits at mul-
647
+ tiple depths with execution time limited to tmax. The
648
+ highlighted green and orange rectangles depict the two
649
+ groups of optimal angles that minimize the energy at
650
+ each depth, as presented in [15]. Green rectangles also
651
+ indicate the depth and texec at which an ansatz will not
652
+ be able to decrease its energy by either increasing depth
653
+ or tmax. Following the observations of Fig. 5, by slightly
654
+ reducing tmax the optimizer will reduce the parameters
655
+ to which the energy is less sensitive. This results in a
656
+
657
+ 6
658
+ γ
659
+ β
660
+ n = 6
661
+ p = 15
662
+ 2
663
+ 4
664
+ 6
665
+ 8
666
+ 10
667
+ 12
668
+ 14
669
+ k
670
+ 0.08
671
+ 0.10
672
+ 0.12
673
+ 0.14
674
+ ⟨H⟩
675
+ δ = 0.0
676
+ δ = 0.02
677
+ δ = 0.05
678
+ δ = 0.08
679
+ δ = 0.1
680
+ 2
681
+ 4
682
+ 6
683
+ 8
684
+ 10
685
+ 12
686
+ 14
687
+ k
688
+ 0.075
689
+ 0.100
690
+ 0.125
691
+ 0.150
692
+ 0.175
693
+ 0.200
694
+ ⟨H⟩
695
+ δ = 0.0
696
+ δ = 0.02
697
+ δ = 0.05
698
+ δ = 0.08
699
+ δ = 0.1
700
+ n = 8
701
+ p = 25
702
+ 0
703
+ 3
704
+ 6
705
+ 9
706
+ 12
707
+ 15
708
+ 18
709
+ 21
710
+ 24
711
+ k
712
+ 0.08
713
+ 0.10
714
+ 0.12
715
+ 0.14
716
+ 0.16
717
+ ⟨H⟩
718
+ δ = 0.0
719
+ δ = 0.02
720
+ δ = 0.05
721
+ δ = 0.08
722
+ δ = 0.1
723
+ 0
724
+ 3
725
+ 6
726
+ 9
727
+ 12
728
+ 15
729
+ 18
730
+ 21
731
+ 24
732
+ k
733
+ 0.10
734
+ 0.15
735
+ 0.20
736
+ 0.25
737
+ ⟨H⟩
738
+ δ = 0.0
739
+ δ = 0.02
740
+ δ = 0.05
741
+ δ = 0.08
742
+ δ = 0.1
743
+ n = 10
744
+ p = 30
745
+ 0
746
+ 4
747
+ 8
748
+ 12
749
+ 16
750
+ 20
751
+ 24
752
+ 28
753
+ k
754
+ 0.10
755
+ 0.12
756
+ 0.14
757
+ 0.16
758
+ 0.18
759
+ 0.20
760
+ ⟨H⟩
761
+ δ = 0.0
762
+ δ = 0.02
763
+ δ = 0.05
764
+ δ = 0.08
765
+ δ = 0.1
766
+ 0
767
+ 4
768
+ 8
769
+ 12
770
+ 16
771
+ 20
772
+ 24
773
+ 28
774
+ k
775
+ 0.10
776
+ 0.15
777
+ 0.20
778
+ 0.25
779
+ 0.30
780
+ ⟨H⟩
781
+ δ = 0.0
782
+ δ = 0.02
783
+ δ = 0.05
784
+ δ = 0.08
785
+ δ = 0.1
786
+ FIG. 6.
787
+ Average energy ⟨H⟩ = ⟨ψ(θ∗ + δθ)| H |ψ(θ∗ + δθ)⟩ of 100 uniformly generated 3-SAT instances where βk (right
788
+ column) or γk (left column), from the k-th layer, are perturbed. The instances are of 6, 8 and 10 qubits with clause to variable
789
+ ratio of 4.2 and unique satisfying assignment.
790
+ slight energy increase as illustrated in Fig. 7 where to
791
+ the left of the green rectangles we can observe darkening
792
+ gradients.
793
+ By contrast, orange rectangles highlight longer execu-
794
+ tion times corresponding to different sets of angles that
795
+ also minimize the energy for a given number of layers.
796
+ Therefore if the optimization routine finds the a solu-
797
+ tion corresponding to the orange rectangle, setting tmax
798
+ to be slightly less than the texec of the orange rectangle
799
+ will lead the optimizer to find angles corresponding to
800
+ the green rectangle. This will amount to a considerable
801
+ reduction in execution time.
802
+ Alternatively increasing the number of layers while
803
+ keeping tmax will reduce the energy. In general, for an
804
+ arbitrary problem Hamiltonian we can not be sure if our
805
+ optimization has returned the ideal set of angles (green
806
+ ones in our example). For this reason, the best strategy
807
+ would be to reduce tmax or increase depth while fixing
808
+ tmax until performance stagnates.
809
+ V.
810
+ DISCUSSION
811
+ In this study we considered a realistic noise model—
812
+ one where the variational gate parameters are stochasti-
813
+ cally perturbed—and demonstrated its effect on the per-
814
+
815
+ 7
816
+ 4.1
817
+ 6.1
818
+ 8.1
819
+ 10.1
820
+ 12.1
821
+ 14.1
822
+ 16.1
823
+ 18.1
824
+ 20.1
825
+ 22.1
826
+ 24.1
827
+ 26.1
828
+ 28.1
829
+ 30.1
830
+ 32.1
831
+ 34.1
832
+ 36.1
833
+ 38.1
834
+ 40.0
835
+ 42.0
836
+ max execution time tmax
837
+ 8
838
+ 7
839
+ 6
840
+ 5
841
+ 4
842
+ 3
843
+ 2
844
+ 1
845
+ depth p
846
+ energy
847
+ 0.2
848
+ 0.4
849
+ 0.6
850
+ 0.8
851
+ FIG. 7. Expected value for multiple combinations of depth for maximum execution times. Green and orange rectangles depict
852
+ the two branches of angles that minimize expectation value for a given depth.
853
+ formance of variational algorithms.
854
+ Using a perturba-
855
+ tive analysis we showed that the change in energy δE
856
+ (from optimised energy E∗), caused due to the pres-
857
+ ence of the considered gate errors, behaves quadratically
858
+ with respect to the angle perturbations for small values
859
+ of the perturbations. This allows us to establish upper
860
+ bounds on the amount of perturbation such that the ac-
861
+ ceptance condition continues to be satisfied. This guar-
862
+ antees a fixed overlap between the target sate and the
863
+ state prepared by the noisy variational circuit. We con-
864
+ firm our analytical findings numerically in QAOA for two
865
+ common problems—3-SAT and unstructured search, us-
866
+ ing different modifications of the considered noise model.
867
+ Moreover we observed form our numerical results that
868
+ the algorithmic performance is more resilient to pertur-
869
+ bations of certain variational parameters. Motivated by
870
+ this observation we demonstrated that performance of
871
+ QAOA with a total execution time texec = �
872
+ k γk + βk
873
+ is stable if retrained with a maximum execution time
874
+ tmax = texec ± ϵ for ϵ ≪ texec. We also show that in
875
+ some cases (a) reduction in tmax can lead to dramatic
876
+ reductions in texec, and (b) increasing depth while fixing
877
+ texec can lead to an energy reduction.
878
+ While our study is primarily focused on energy pertur-
879
+ bations around the noiseless optimum θ∗, in practice one
880
+ has to train in the presence of noise. This would change
881
+ optimal angles θ∗ → θ∗ + δθ∗, where shift δθ∗ increases
882
+ with increase of the strength of the noise. Nevertheless,
883
+ using perturbation theory around the noiseless optimum
884
+ one can estimate δθ∗ = O(σ2), and the corresponding
885
+ change in the energy is Tr(ρ(θ∗+δθ∗)H)−Tr(ρ(θ∗)H) =
886
+ O(σ4). Therefore, working in the regime of weak noise
887
+ one can safely use noiseless optimum θ∗. See appendix
888
+ C for detailed calculation.
889
+ VI.
890
+ ACKNOWLEDGEMENT
891
+ * D.R., E.C., S.A., E.P., D.V. acknowledge support
892
+ from the research project, Leading Research Center on
893
+ Quantum Computing (agreement No. 014/20).
894
+ [1] John Preskill. Quantum computing in the nisq era and
895
+ beyond. Quantum, 2:79, 2018.
896
+ [2] Johannes Weidenfeller, Lucia C Valor, Julien Gacon,
897
+ Caroline Tornow, Luciano Bello, Stefan Woerner, and
898
+ Daniel J Egger.
899
+ Scaling of the quantum approximate
900
+ optimization algorithm on superconducting qubit based
901
+ hardware. arXiv preprint arXiv:2202.03459, 2022.
902
+ [3] Alexander K Ratcliffe, Richard L Taylor, Joseph J Hope,
903
+ and Andr´e RR Carvalho. Scaling trapped ion quantum
904
+ computers using fast gates and microtraps. Physical Re-
905
+ view Letters, 120(22):220501, 2018.
906
+ [4] Swathi S Hegde, Jingfu Zhang, and Dieter Suter. Toward
907
+ the speed limit of high-fidelity two-qubit gates. Physical
908
+ Review Letters, 128(23):230502, 2022.
909
+ [5] Adam R Mills, Charles R Guinn, Michael J Gullans, An-
910
+ thony J Sigillito, Mayer M Feldman, Erik Nielsen, and
911
+ Jason R Petta.
912
+ Two-qubit silicon quantum processor
913
+ with operation fidelity exceeding 99%. Science Advances,
914
+ 8(14):eabn5130, 2022.
915
+ [6] Dorit Aharonov, Xun Gao, Zeph Landau, Yunchao Liu,
916
+ and Umesh Vazirani. A polynomial-time classical algo-
917
+ rithm for noisy random circuit sampling. arXiv e-prints,
918
+ page arXiv:2211.03999, November 2022.
919
+ [7] Sitan Chen, Jordan Cotler, Hsin-Yuan Huang, and Jerry
920
+ Li.
921
+ The Complexity of NISQ.
922
+ arXiv e-prints, page
923
+ arXiv:2210.07234, October 2022.
924
+ [8] Abhinav Kandala, Antonio Mezzacapo, Kristan Temme,
925
+ Maika Takita,
926
+ Markus Brink,
927
+ Jerry M Chow,
928
+ and
929
+ Jay M Gambetta. Hardware-efficient variational quan-
930
+ tum eigensolver for small molecules and quantum mag-
931
+ nets. Nature, 549(7671):242–246, 2017.
932
+
933
+ 8
934
+ [9] Daniil Rabinovich, Soumik Adhikary, Ernesto Campos,
935
+ Vishwanathan Akshay, Evgeny Anikin, Richik Sengupta,
936
+ Olga Lakhmanskaya, Kirill Lakhmanskiy, and Jacob Bi-
937
+ amonte. Ion-native variational ansatz for quantum ap-
938
+ proximate optimization. Phys. Rev. A, 106:032418, Sep
939
+ 2022.
940
+ [10] Guido
941
+ Pagano,
942
+ Aniruddha
943
+ Bapat,
944
+ Patrick
945
+ Becker,
946
+ Katherine S Collins, Arinjoy De, Paul W Hess, Harvey B
947
+ Kaplan, Antonis Kyprianidis, Wen Lin Tan, Christopher
948
+ Baldwin, et al. Quantum approximate optimization of
949
+ the long-range ising model with a trapped-ion quantum
950
+ simulator. Proceedings of the National Academy of Sci-
951
+ ences, 117(41):25396–25401, 2020.
952
+ [11] Jacob Biamonte. Universal variational quantum compu-
953
+ tation. Physical Review A, 103(3):L030401, 2021.
954
+ [12] Marco Cerezo, Andrew Arrasmith, Ryan Babbush, Si-
955
+ mon C Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R
956
+ McClean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio,
957
+ et al. Variational quantum algorithms. Nature Reviews
958
+ Physics, 3(9):625–644, 2021.
959
+ [13] E Campos, D Rabinovich, V Akshay, and J Biamonte.
960
+ Training saturation in layerwise quantum approximate
961
+ optimisation. (Letter) Physical Review A, 104:L030401,
962
+ 2021.
963
+ [14] V. Akshay, H. Philathong, M. E.S. Morales, and J. D. Bi-
964
+ amonte. Reachability Deficits in Quantum Approximate
965
+ Optimization.
966
+ Physical Review Letters, 124(9):090504,
967
+ Mar 2020.
968
+ [15] Vishwanathan Akshay, Daniil Rabinovich, Ernesto Cam-
969
+ pos, and Jacob Biamonte. Parameter concentrations in
970
+ quantum approximate optimization. Physical Review A,
971
+ 104(1):L010401, 2021.
972
+ [16] Lov K Grover.
973
+ A fast quantum mechanical algorithm
974
+ for database search. In Proceedings of the twenty-eighth
975
+ annual ACM symposium on Theory of computing, pages
976
+ 212–219, 1996.
977
+ [17] Edward Farhi, Jeffrey Goldstone, and Sam Gutmann.
978
+ A quantum approximate optimization algorithm. arXiv
979
+ preprint arXiv:1411.4028, 2014.
980
+ [18] Murphy Yuezhen Niu, Sirui Lu, and Isaac L Chuang. Op-
981
+ timizing qaoa: Success probability and runtime depen-
982
+ dence on circuit depth. arXiv preprint arXiv:1905.12134,
983
+ May 2019.
984
+ [19] Seth Lloyd. Quantum approximate optimization is com-
985
+ putationally universal. arXiv preprint arXiv:1812.11075,
986
+ 2018.
987
+ [20] Mauro ES Morales, Jacob D Biamonte, and Zolt´an Zim-
988
+ bor´as. On the universality of the quantum approximate
989
+ optimization algorithm. Quantum Information Process-
990
+ ing, 19(9):1–26, 2020.
991
+ [21] Leo Zhou, Sheng-Tao Wang, Soonwon Choi, Hannes
992
+ Pichler, and Mikhail D. Lukin. Quantum approximate
993
+ optimization algorithm: Performance, mechanism, and
994
+ implementation on near-term devices.
995
+ Phys. Rev. X,
996
+ 10:021067, Jun 2020.
997
+ [22] Zhihui Wang, Nicholas C Rubin, Jason M Dominy, and
998
+ Eleanor G Rieffel. X y mixers: Analytical and numeri-
999
+ cal results for the quantum alternating operator ansatz.
1000
+ Physical Review A, 101(1):012320, 2020.
1001
+ [23] Lucas T. Brady, Christopher L. Baldwin, Aniruddha Ba-
1002
+ pat, Yaroslav Kharkov, and Alexey V. Gorshkov. Opti-
1003
+ mal Protocols in Quantum Annealing and Quantum Ap-
1004
+ proximate Optimization Algorithm Problems. Physical
1005
+ Review Letters, 126(7):070505, Feb 2021.
1006
+ [24] Edward
1007
+ Farhi
1008
+ and
1009
+ Aram
1010
+ W
1011
+ Harrow.
1012
+ Quantum
1013
+ supremacy through the quantum approximate optimiza-
1014
+ tion algorithm. arXiv preprint arXiv:1602.07674, 2016.
1015
+ [25] Edward Farhi, Jeffrey Goldstone, Sam Gutmann, and
1016
+ Leo Zhou. The quantum approximate optimization algo-
1017
+ rithm and the sherrington-kirkpatrick model at infinite
1018
+ size. arXiv preprint arXiv:1910.08187, Oct 2019.
1019
+ [26] Matteo M Wauters, Glen Bigan Mbeng, and Giuseppe E
1020
+ Santoro.
1021
+ Polynomial scaling of qaoa for ground-state
1022
+ preparation of the fully-connected p-spin ferromagnet.
1023
+ arXiv preprint arXiv:2003.07419, 2020.
1024
+ [27] Jahan Claes and Wim van Dam. Instance independence
1025
+ of single layer quantum approximate optimization al-
1026
+ gorithm on mixed-spin models at infinite size.
1027
+ arXiv
1028
+ preprint arXiv:2102.12043, 2021.
1029
+ [28] Leo Zhou, Sheng-Tao Wang, Soonwon Choi, Hannes
1030
+ Pichler, and Mikhail D. Lukin. Quantum approximate
1031
+ optimization algorithm: Performance, mechanism, and
1032
+ implementation on near-term devices.
1033
+ Phys. Rev. X,
1034
+ 10:021067, Jun 2020.
1035
+ [29] Vishwanathan Akshay, H Philathong, E Campos, Daniil
1036
+ Rabinovich, Igor Zacharov, Xiao-Ming Zhang, and J Bi-
1037
+ amonte. On circuit depth scaling for quantum approx-
1038
+ imate optimization.
1039
+ arXiv preprint arXiv:2205.01698,
1040
+ 2022.
1041
+ [30] John Clarke and Frank K Wilhelm.
1042
+ Superconducting
1043
+ quantum bits. Nature, 453(7198):1031–1042, 2008.
1044
+ [31] Jay M Gambetta, Jerry M Chow, and Matthias Steffen.
1045
+ Building logical qubits in a superconducting quantum
1046
+ computing system. npj quantum information, 3(1):1–7,
1047
+ 2017.
1048
+ [32] L. V. Gerasimov, R. R. Yusupov, A. D. Moiseevsky,
1049
+ I. Vybornyi, K. S. Tikhonov, S. P. Kulik, S. S. Straupe,
1050
+ C. I. Sukenik, and D. V. Kupriyanov. Coupled dynamics
1051
+ of spin qubits in optical dipole microtraps. 2022.
1052
+ [33] M. Morgado and S. Whitlock. Quantum simulation and
1053
+ computing with rydberg-interacting qubits. AVS Quan-
1054
+ tum Science, 3(2):023501, 2021.
1055
+ [34] Guido Pagano, A Bapat, P Becker, KS Collins, A De,
1056
+ PW Hess, HB Kaplan, A Kyprianidis, WL Tan, C Bald-
1057
+ win, et al.
1058
+ Quantum approximate optimization of the
1059
+ long-range ising model with a trapped-ion quantum sim-
1060
+ ulator. arXiv preprint arXiv:1906.02700, 2019.
1061
+ [35] J. Zhang, G. Pagano, P. W. Hess, A. Kyprianidis,
1062
+ P. Becker, H. Kaplan, A. V. Gorshkov, Z.-X. Gong,
1063
+ and C. Monroe. Observation of a many-body dynami-
1064
+ cal phase transition with a 53-qubit quantum simulator.
1065
+ Nature, 551(7682):601–604, 2017.
1066
+ [36] Laird Nicholas Egan. Scaling Quantum Computers with
1067
+ Long Chains of Trapped Ions. PhD thesis, University of
1068
+ Maryland, College Park, 2021.
1069
+ [37] Zhi-Cheng Yang,
1070
+ Armin Rahmani,
1071
+ Alireza Shabani,
1072
+ Hartmut Neven, and Claudio Chamon. Optimizing vari-
1073
+ ational quantum algorithms using pontryagin’s minimum
1074
+ principle. Physical Review X, 7(2):021027, 2017.
1075
+ [38] Mohannad
1076
+ Ibrahim,
1077
+ Hamed
1078
+ Mohammadbagherpoor,
1079
+ Cynthia Rios, Nicholas T Bronn, and Gregory T Byrd.
1080
+ Pulse-level optimization of parameterized quantum cir-
1081
+ cuits for variational quantum algorithms. arXiv preprint
1082
+ arXiv:2211.00350, 2022.
1083
+
1084
+ 9
1085
+ Appendix A: 3-SAT and unstructured search
1086
+ problems
1087
+ 1.
1088
+ 3-SAT
1089
+ Boolean satifyability, or SAT, is the problem of deter-
1090
+ mining weather a boolean formula written in conjunctive
1091
+ normal form (CNF) is satisfiable. It is possible to map
1092
+ any SAT instance via Karp reduction into 3-SAT, which
1093
+ are restricted to 3 literals per clause.
1094
+ In order to ap-
1095
+ proximate solutions to SAT we embed the instance into
1096
+ a Hamiltonian as
1097
+ HSAT =
1098
+
1099
+ j
1100
+ P(j),
1101
+ (A1)
1102
+ where j indexes clauses of an instance, and P(j) is the
1103
+ tensor product of projectors that penalizes bit string as-
1104
+ signments that do not satisfy the j-th clause.
1105
+ 2.
1106
+ Unstructured search
1107
+ Consider an unstructured database S indexed by j ∈
1108
+ {0, 1}×n. Let f : {0, 1}×n → {0, 1} be a Boolean function
1109
+ (a.k.a. black box) such that:
1110
+ f(j) =
1111
+
1112
+ 1
1113
+ iff j = t
1114
+ 0
1115
+ otherwise.
1116
+ (A2)
1117
+ The task is to find t ∈ {0, 1}×n. The corresponding prob-
1118
+ lem Hamiltonian for QAOA is
1119
+ Ht = 1 − |t⟩⟨t|,
1120
+ (A3)
1121
+ thus the expected value is given by
1122
+ ⟨H⟩ = 1 − |⟨t|ψp(γ, β)⟩|2.
1123
+ (A4)
1124
+ QAOA performance for unstructured search is not sen-
1125
+ sitive to the particular target state |t⟩ in the computa-
1126
+ tional basis. For any target state |t⟩ representing a binary
1127
+ string, there is a U = U † composed of X and 1 opera-
1128
+ tors such that U |0⟩⊗n = |t⟩. The overlap of an arbitrary
1129
+ state prepared by a QAOA sequence with |t⟩ is then:
1130
+ ⟨t|ψp(γ, β)⟩ = ⟨t|
1131
+ p
1132
+
1133
+ k=1
1134
+ e−iβkHxe−iγk|t⟩⟨t| |+⟩⊗n
1135
+ = ⟨0|⊗n U
1136
+ p
1137
+
1138
+ k=1
1139
+ e−iβkHxe−iγkU(|0⟩⟨0|)⊗nU |+⟩⊗n
1140
+ = ⟨0|⊗n U
1141
+ p
1142
+
1143
+ k=1
1144
+ e−iβkHxUe−iγk(|0⟩⟨0|)⊗nU |+⟩⊗n
1145
+ = ⟨0|⊗n
1146
+ p
1147
+
1148
+ k=1
1149
+ e−iβkHxe−iγk(|0⟩⟨0|)⊗n |+⟩⊗n ,
1150
+ which is independent on t.
1151
+ Appendix B: Energy variation in presence of
1152
+ constant perturbations to gate parameters
1153
+ Using (9) one can calculate perturbation to the energy
1154
+ caused by a shift of the optimal angles by a constant δθ
1155
+ as
1156
+ δE = ⟨ψ(θ∗ + δθ)| H |ψ(θ∗ + δθ)⟩ − ⟨ψ(θ∗)| H |ψ(θ∗)⟩
1157
+ = −
1158
+ q
1159
+
1160
+ k=1
1161
+ δθ2
1162
+ kE∗ +
1163
+ q
1164
+
1165
+ m̸=k
1166
+ δθkδθm(⟨ψ(θ∗)| H |ψkm⟩ + h.c.)
1167
+ +
1168
+ q
1169
+
1170
+ m,k
1171
+ δθkδθk ⟨ψm| H |ψk⟩ + o(δθkδθm)
1172
+ = 1
1173
+ 2(δθ)T Hδθ + o(δθkδθm),
1174
+ (B1)
1175
+ where |ψmk⟩ = |ψ0...1...1...0⟩ with 1 placed only at m-th
1176
+ and k-th positions. H is the Hessian of the energy at
1177
+ noiseless optimum, Hij =
1178
+ ∂2
1179
+ ∂θi∂θj
1180
+ ⟨ψ(θ)| H |ψ(θ)⟩ |θ=θ∗.
1181
+ Here we use the fact that at the optimal position linear
1182
+ contribution to the cost function necessarily vanishes. It
1183
+ is seen now that for the constant perturbation δθk = δ
1184
+ the energy changes as δE ∝ δ2.
1185
+ Appendix C: Optimal parameters variation in the
1186
+ presence of noise
1187
+ Let us use expressions (11) and (12) to estimate change
1188
+ in the energy if one accounts for shift of optimal param-
1189
+ eters θ∗ → θ∗ + δθ∗:
1190
+ Tr(ρ(θ∗ + δθ∗)H) = (1 −
1191
+ q
1192
+
1193
+ k=1
1194
+ ak) ⟨ψ(θ∗ + δθ∗)| H |ψ(θ∗ + δθ∗)⟩ +
1195
+ q
1196
+
1197
+ k=1
1198
+ ak ⟨ψk(θ∗ + δθ∗)| H |ψk(θ∗ + δθ∗)⟩ + o(σ2
1199
+ k)
1200
+ (C1)
1201
+ We introduce gradients of the noisy terms Bk
1202
+ =
1203
+
1204
+ ∂θ ⟨ψk(θ)| H |ψk(θ)⟩ |θ=θ∗. Notice that gradients of the
1205
+
1206
+ 10
1207
+ noiseless function ⟨ψ(θ)| H |ψ(θ)⟩ vanish at optimum.
1208
+ Then,
1209
+ Tr(ρ(θ∗ + δθ∗)H) ≈ (1 −
1210
+ q
1211
+
1212
+ k=1
1213
+ ak)E∗ + 1
1214
+ 2(δθ∗)T Hδθ∗
1215
+ +
1216
+ q
1217
+
1218
+ k=1
1219
+ ak[⟨ψk(θ∗)| H |ψk(θ∗)⟩ + (δθ∗)T Bk].
1220
+ (C2)
1221
+ Minimizing it with respect to δθ∗ one gets δθ∗ =
1222
+ �q
1223
+ k=1 akH−1Bk. Thus, if we account for the change of
1224
+ optimal parameters in the presence of noise, the energy
1225
+ shifts by
1226
+ Tr(ρ(θ∗ + δθ∗)H) − Tr(ρ(θ∗)H) ≈
1227
+ (δθ∗)T Hδθ∗ +
1228
+ q
1229
+
1230
+ k=1
1231
+ ak(δθ∗)T Bk = O(σ4).
1232
+ (C3)
1233
+
69AyT4oBgHgl3EQfQfbA/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
6NFJT4oBgHgl3EQflSzb/content/tmp_files/2301.11583v1.pdf.txt ADDED
@@ -0,0 +1,1151 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Tunable Strong Magnon-Magnon Coupling in Two-
2
+ Dimensional Array of Diamond Shaped Ferromagnetic
3
+ Nanodots
4
+ Sudip Majumder1, Samiran Choudhury1, Saswati Barman2, Yoshichika Otani3, 4,
5
+ Anjan Barman1,*
6
+ 1Department of Condensed Matter Physics and Material Sciences, S. N. Bose National Centre for
7
+ Basic Sciences, Block JD, Sector III, Salt Lake, 700106, Kolkata, India
8
+ 2Institute for Engineering and Management, Sector V, Salt Lake, 700091, Kolkata, India
9
+ 3 CEMS-RIKEN, 2-1 Hirosawa, Saitama, 3510198, Wako, Japan
10
+ 4Institute for Solid State Physics, University of Tokyo, 515 Kashiwanoha, Chiba, 277 8581,
11
+ Kashiwa, Japan
12
+ *Email: abarman@bose.res.in
13
+
14
+
15
+ Abstract
16
+ Hybrid magnonics involving coupling between magnons and different quantum particles have
17
+ been extensively studied during past few years for varied interests including quantum
18
+ electrodynamics. In such systems, magnons in magnetic materials with high spin density are
19
+ utilized where the “coupling strength” is collectively enhanced by the square root of the number
20
+ of spins to overcome the weaker coupling between individual spins and the microwave field.
21
+ However, achievement of strong magnon-magnon coupling in a confined nanomagnets would
22
+ be essential for on-chip integration of such hybrid systems. Here, through intensive study of
23
+ interaction between different magnon modes in a Ni80Fe20 (Py) nanodot array, we demonstrate
24
+ that the intermodal coupling can approach the strong coupling regime with coupling strength
25
+ up to 0.82 GHz and cooperativity of 2.51. Micromagnetic simulations reveal that the
26
+ intermodal coupling is mediated by the exchange field inside each nanodot. The coupling
27
+ strength could be continuously tuned by varying the bias field (Hext) strength and orientation
28
+ (), opening routes for external control over hybrid magnonic systems. These findings could
29
+ greatly enrich the rapidly evolving field of quantum magnonics.
30
+
31
+ 1. Introduction
32
+ Hybrid quantum systems [1] have recently attracted great attention due to their fundamental
33
+ importance and potential applications. It provides a new paradigm for the coherent transfer of
34
+
35
+ quantum states from one platform to another to execute quantum information processing [2,3].
36
+ This significantly facilitates the research on the fundamental physics of coupling between
37
+ different platforms which may lead to varied applications of quantum technologies, such as:
38
+ quantum computing [4,5], quantum communications [6,7], and quantum sensing [8]. The
39
+ introduction of magnons in hybrid systems was initiated from the exploration of spin ensembles
40
+ coupled to microwave photons [8-10]. The higher densities of spin in magnetic materials and
41
+ their collective dynamics as magnons, provide ultra-strong coupling with cooperativity up to
42
+ 103-104 [11,12]. During the last decade, extensive research has been done on magnon-magnon
43
+ coupling [13-19]. However, on-chip integration of hybrid systems requires downscaling the
44
+ dimensions of the systems to the nanometer range. The microwave cavity usually has the
45
+ dimension of millimeters. The coupling strength (g) is proportional to the square root of the
46
+ number of spins present in the magnetic material [20,21]. To increase the coupling strength the
47
+ number of spins in the magnetic material is usually required to be large enough (N  1013),
48
+ thereby restricting the size of the microwave cavity and magnet and the ensuing device
49
+ miniaturization towards CMOS integration.
50
+ To overcome this geometrical limitation of a microwave cavity, it becomes imperative to
51
+ search for different systems to act as nanometric resonators. In this context, the recent
52
+ development of interlayer magnon coupling or exchange-driven magnon-magnon coupling in
53
+ the magnetic systems has opened a new avenue for quantum magnonics [22-24]. In the last
54
+ decade, extensive studies have been done using both confined and propagating magnons in the
55
+ field of magnonics, which emerged as an exciting field of research. To this end single
56
+ nanomagnets have been studied extensively due to their geometrically confined rich volume
57
+ and localized magnetic modes [25-29] in nanometer dimension and their tunability with
58
+ different external parameters. Therefore, such systems possess great potential in quantum
59
+ magnonics with the possibility of developing magnon-based on-chip quantum information
60
+ processing systems in the GHz and THz frequency range with high energy efficiency. Recently
61
+ magnon-magnon coupling has been observed experimentally in ferromagnetic nanowire
62
+ array[15] and in single nanomagnet using micromagnetic simulation[30]. Furthermore,
63
+ moderate to strong magnon-magnon coupling have also been observed in Ni80Fe20 (Py)
64
+ nanocross array mediated by dynamic dipolar interaction [31] and anisotropic dipolar
65
+ interaction[32]. These studies have opened a new approach for executing and controlling this
66
+ phenomenon in a large variety of systems by tailoring the geometric and material parameters
67
+ of these artificially patterned systems and the external bias field. This leads the quest for
68
+
69
+ optimal solutions for applications in magnon-based quantum information technology.
70
+
71
+ Here, we have explored magon-magnon coupling in diamond-shaped Py nanodot array with
72
+ the aid of a broadband ferromagnetic resonance (FMR) spectrometer[33,34] and
73
+ micromagnetic simulations. Remarkably, we observe an avoided crossing (anticrossing) of
74
+ magnon modes [1] characteristic of the formation of hybrid system. Anticrossing gap of up to
75
+ 0.82 GHz and the ensuing cooperativity value as high as 2.51 are observed. Micromagnetic
76
+ simulations reveal that the coupling between two magnon modes is mediated by the exchange
77
+ field within each nanodot. Furthermore, the coupling strength is found to be highly dependent
78
+ on the orientation and strength of the bias magnetic field, leading towards the possibility of
79
+ externally controlled hybrid magnonic devices.
80
+
81
+
82
+ 2. Experimental Details
83
+ The 20-nm-thick diamond shaped Py nanodots, arranged in an array of dimensions 25 μm ×
84
+ 200 μm, were prepared on self-oxidized Si [100] substrate by using electron beam evaporation
85
+ (EBE), electron beam lithography (EBL), and Ar+ ion milling tools. A coplanar waveguide
86
+ (CPW) made of Au, having 150 nm thickness, 30 μm wide central conducting (signal) line and
87
+ 50 Ω characteristic impedance (Fig. 1(a)) was deposited on top of each array for broadband
88
+ FMR measurements. The CPW is separated from the nanodot array by a 60-nm-thick insulating
89
+ Al2O3 layer. The fabrication details are described in section S1 of the Supplementary Materials.
90
+ Fig. 1(b) exhibits the scanning electron microscope (SEM) image of the diamond nanodot array
91
+ arranged in a square lattice having width and height of the nanodots as 325 nm (dx) and 350
92
+ nm (dy) and lattice constant of 400 nm. The nanomagnet’s lateral dimensions and pitch are
93
+ shown in the SEM image of Fig. 1(b). The SEM image shows that the fabricated structures
94
+ suffer from slight edge deformations and rounded corners. All these deformations have been
95
+ incorporated in the micromagnetic simulations as described later. The applied bias magnetic
96
+
97
+ field orientation is shown in the inset of Fig. 1(b). The spin-wave (SW) spectra from the
98
+ samples were measured using a broadband FMR spectrometer, consisting of a high-frequency
99
+ Vector Network Analyzer (VNA, Agilent PNA-L, model no.: N5230C, frequency range: 10
100
+ MHz to 50 GHz) and a homemade high-frequency probe station equipped with nonmagnetic
101
+ ground-signal-ground (GSG)-type picoprobe (GGB Industries, model no.: 40A-GSG-150-
102
+ EDP) and a coaxial cable. One end of the CPW is shorted and the back-reflected signal is
103
+ collected and fed back to the VNA by the same GSG probe and the coaxial cable. From the
104
+ frequency dependent real part of the S-parameter in the reflection geometry (Re (S11)), different
105
+ SW frequencies are identified, which results in the characteristic SW spectrum of the sample.
106
+ Additional details of the experimental setup are given in section S2 of the Supplementary
107
+ Materials.
108
+
109
+
110
+
111
+ FIG. 1. (a) Schematic of the experimental geometry. The directions of the bias magnetic field
112
+ (Hext) and rf magnetic field (hrf) are shown in the schematic. (b) SEM image of diamond-shaped
113
+ Ni80Fe20 (Py) nanodots arranged in a square lattice having lattice constant a = 400 nm and nanodot
114
+ width dx = 325 nm, height dy = 350 nm. The inset again shows the orientation of Hext with respect
115
+ to hrf. (c) Real parts of the forward scattering parameter (S11) representing the FMR spectra at Hext
116
+ = 400 Oe applied at an azimuthal angle  = 0°. The observed spin-wave (SW) modes are marked
117
+ by down arrows. (d) Bias field (Hext) dependent SW absorption spectra of Py nanodots is shown
118
+ at  = 0°. The surface plots correspond to the experimental results, while the symbols represent
119
+ the simulated data. The color map for the surface plots and the schematic of Hext are given at the
120
+ bottom right corner of the figure.
121
+
122
+ 3
123
+ 6
124
+ 9
125
+ 0.0
126
+ 0.5
127
+ 1.0
128
+
129
+
130
+
131
+ 0.0
132
+ 0.3
133
+ 0.6
134
+ 0.9
135
+ 1.2
136
+ 3
137
+ 6
138
+ 9
139
+ 12
140
+ M1
141
+ M2
142
+ M3
143
+
144
+ Frequency (GHz)
145
+ Hext (kOe)
146
+ Frequency (GHz)
147
+ Re S11 (Normalized)
148
+ M1
149
+ M2
150
+ M3
151
+ Hext= 400 Oe
152
+ (a)
153
+ (b)
154
+ (c)
155
+ (d)
156
+ 500 nm
157
+ x
158
+ y
159
+ Hext
160
+
161
+ dx
162
+ a
163
+ dy
164
+ Re S11
165
+ Normalised
166
+ 1
167
+ 0
168
+ (b)
169
+
170
+ G
171
+ s
172
+ G
173
+ 3. Results and Discussion
174
+ 3.1. Experimental Result
175
+ 3.1.1. Field Dependence of SW
176
+ The SW absorption spectra (Re (S11)) are acquired from FMR measurements for a broad
177
+ range of bias magnetic field. Fig. 1(c) shows representative raw spectra at Hext = 400 Oe. At
178
+ first, the magnetization of the samples are saturated along the +x direction by applying Hext =
179
+ 1800 Oe, followed by gradual reduction of the field from 1600 Oe to 0 Oe at steps of 20 Oe in
180
+ a single trace. The surface plot in Fig. 1(d) displays the bias-field-dependent of SW absorption
181
+ spectra with their maximum power normalized to 1.0. These surface plots are generated from
182
+ the individual Re (S11) spectra acquired at a given applied magnetic field. Here, the bright
183
+ regions represent the experimental data while the symbols represent the micromagnetic
184
+ simulation results. The normalized surface plots help to identify three separate branches of SW,
185
+ among which the lowest frequency branch M1 shows maximum intensity in the entire field
186
+ regime. As we decrease the bias field M1 shows a dip (minimum) in f-Hext at Hext ≈ 300 Oe,
187
+ which indicates a mode softening due to transition in magnetization state of the nanomagnet
188
+ array. Other two SW modes M2 and M3 do not show any such transition and monotonically
189
+ decrease with the reduction in the bias field.
190
+
191
+ Fig. 2 shows the magnetic field dependences of the frequencies at different bias field angles.
192
+ The variation of magnetic field orientation creates some remarkable changes. First, the dip in
193
+ M1 occurring at ~300 Oe gradually disappears. Fig. 2(a) shows the f-Hext plot at  = 5, where
194
+ the dip shows an upward shift. At  = 15, the dip completely disappears and the M1 shows a
195
+ monotonic variation of frequency with the field, as shown in Fig. 2(b). Secondly, the relative
196
+ intensity of M2 and M3 shows a clear variation with the bias field orientation. For 5 ≤  ≤
197
+ 15, M2 gradually losses its intensity at the expense of gradual increment of intensity of M3,
198
+ which starts to dominate over M2 at  = 15. With further increment of angle, M2 further loses
199
+ its intensity and at  = 23 it completely disappears. Fig. 2(c) shows the f-Hext plot at  = 23
200
+ where a clear anticrossing between the branches representing modes M1 and M3 is observed
201
+ at Hext = 1060 Oe. The vertical dotted line represents the anticrossing field (Hac) in the f-Hext
202
+ plot. The value of Hac gradually shifts towards the lower field regime as we keep increasing .
203
+
204
+ Fig. 2(d) shows the magnetic field dispersion of SW frequencies at  = 30 where an
205
+ anticrossing is observed at Hext = 920 Oe in between the SW modes M1 and M3. Here, the mid
206
+ frequency SW mode M2* reappears, though the intensity of this mode is low. With further
207
+ increment of , this mode becomes more prominent and two different anticrossings are now
208
+ observed instead of one. One of those appears in between M1 and M2* and another one in
209
+ between M2* and M3. At  = 45, both of the anticrossings are observed at Hext = 475 Oe as
210
+ shown in Fig. 2(e). With further increment of , the first anticrossing shifts towards lower bias
211
+ magnetic field values, whereas the second one appears in higher bias field values. Fig. 2(f)
212
+ shows the magnetic field dispersion of SW frequencies at  = 60 where the first anticrossing
213
+ in between M1 and M2* appear at Hext = 410 Oe and second one at Hext = 600 Oe.
214
+
215
+ 3.1.2. Angular Dependence of SW
216
+
217
+ The variation of SW modes and their mutual interactions show high dependence on the in-
218
+ plane magnetic field orientation. For this reason, -dependence of SW spectra were acquired
219
+ at a constant bias field magnitude Hext in the range 0º ≤  ≤ 360º. In Fig. 3(a-d), we have
220
+
221
+
222
+
223
+ FIG. 2. Bias field (Hext) dependent SW absorption plots of Py diamond shaped nanodot array are shown
224
+ for the bias field orientation () of (a) 5°, (b) 15°, (c) 23°, (d) 30°, (e) 45° and (f) 60°. The surface plots
225
+ correspond to the experimental results, while the symbols represent the simulated data. The color map
226
+ for the surface plots and the schematic of the external applied field (Hext) are given at the bottom right
227
+ corner of the figure.
228
+
229
+ 0
230
+ 400
231
+ 800
232
+ 1200
233
+ 3
234
+ 6
235
+ 9
236
+ 12
237
+ M1
238
+ M2*
239
+ M3
240
+
241
+ 0
242
+ 500
243
+ 1000
244
+ 1500
245
+ M1
246
+ M3
247
+
248
+
249
+
250
+ Frequency (GHz)
251
+ Hext (kOe)
252
+ 0
253
+ 400
254
+ 800
255
+ 1200
256
+ M1
257
+ M2
258
+ M3
259
+
260
+  = 15
261
+  = 30
262
+ 0
263
+ 400
264
+ 800
265
+ 1200
266
+ M1
267
+ M2*
268
+ M3
269
+
270
+  = 45
271
+  = 23
272
+ 0
273
+ 400
274
+ 800
275
+ 1200
276
+ M1
277
+ M2*
278
+ M3
279
+
280
+  = 60
281
+ 0
282
+ 400
283
+ 800
284
+ 1200
285
+ 3
286
+ 6
287
+ 9
288
+ 12
289
+ M1
290
+ M2
291
+ M3
292
+
293
+  = 5
294
+ x
295
+ y
296
+ Hext
297
+
298
+ (a)
299
+ (b)
300
+ (c)
301
+ (d)
302
+ (e)
303
+ (f)
304
+ Re S11
305
+ Normalized
306
+ 1
307
+ 0
308
+
309
+ presented the -dependence at Hext = 200, 400, 600 and 800 Oe. To show the anticrossing points
310
+ we have magnified the relevant regions of the -dependent SW spectra. In the Supplementary
311
+ Information figure S4, we have shown the full range of -dependence. At a lower field value
312
+ like Hext = 200 Oe, only M1 shows angular dispersion as shown in Fig. 3(a). With an increment
313
+ in Hext, two more modes start to show angular dispersion. Here, mode M1 shows a sharp
314
+ variation of frequency with a minimum at  = 0, corresponding to the minimum observed in
315
+ Fig. 1(d). As we increase the field this sharp modulation gradually transforms into a continuous
316
+ angular variation. Fig. 3(b) shows the angular dispersion at Hext = 400 Oe. For  between 50
317
+ and 55, an anticrossing gap appears in between M1 and M2* which is shown by a white dotted
318
+ line. At a higher field of Hext = 600 Oe instead of one, two different anticrossings are observed.
319
+ The first one appears in between M1 and M3 at  = 40 while the 2nd one appears in between
320
+ M2* and M3 at  = 60. With an increment of magnetic field (e.g., 800 Oe) the first anticrossing
321
+ shifts towards lower angle (e.g. 35), while the second one gradually disappears as shown in
322
+ Fig. 3(d). Due to four fold symmetry[35] of diamond shaped nanodot array these anticrossing
323
+ also appear in other three quadrants of angular variation spectra of SW, which is shown in
324
+ section S4 of supplementary section.
325
+
326
+
327
+
328
+
329
+ 3.1.3. Anticrossing Strength
330
+ Fig. 4(a) shows the power spectrum measured at Hext = 1060 Oe, which is the anticrossing field
331
+ (Hac) for  = 23 configuration. The blue line represents the FMR spectra whereas the red line
332
+ represent the fitted spectra using an antisymmetric lorentzian function. Other FMR spectra for
333
+ varying anticrossing fields are presented in section S5 of Supplementary Information. The
334
+ magnon–magnon coupling strength g is defined as half of the peak-to-peak frequency spacing
335
+ at the anticrossing field, which is shown in Fig. 4(a). In order to estimate the strength of
336
+ interaction between these two modes, we have extracted the value of g13 and the corresponding
337
+ dissipation rates 1, 3 as shown in Fig. 4(a). Here, 1 and 3 are defined as half-width at half-
338
+ maximum of the FMR peak of SW mode M1 and M3, respectively.
339
+
340
+
341
+
342
+
343
+ FIG. 3. Variation of SW frequency as a function of the azimuthal angle () varying from 0° to 360° for
344
+ bias field value fixed at (a) Hext = 200 Oe, (b) 400 Oe, (c) 600 Oe and (d) 800 Oe. The surface plots
345
+ correspond to the experimental results, while the symbols represent the simulated data. The colour map
346
+ for the surface plots and the schematic of Hext are shown on the right side of the figure.
347
+
348
+
349
+
350
+ 6
351
+ 9
352
+ M2*
353
+ M3
354
+ M1
355
+ M2
356
+
357
+
358
+ 0
359
+ -60
360
+ 60
361
+ 3
362
+ 6
363
+ 9
364
+ M1
365
+ M2
366
+ M3
367
+ M1
368
+ M2
369
+ M3
370
+ M2*
371
+
372
+ 0
373
+ -60
374
+ 60
375
+ Frequency (GHz)
376
+ x
377
+ y
378
+ Hext
379
+
380
+ 6
381
+ 9
382
+ M2*
383
+ M3
384
+ M1
385
+ M2
386
+
387
+ 0
388
+ -60
389
+ 60
390
+ Azimuthal Angle,  (Degree)
391
+ 3
392
+ 6
393
+ 9
394
+ M2
395
+ *
396
+ M3
397
+ M1
398
+ M2
399
+
400
+ 0
401
+ -60
402
+ 60
403
+ (a)
404
+ (b)
405
+ (c)
406
+ (d)
407
+ Re S11
408
+ Normalized
409
+ 1
410
+ 0
411
+ 200 Oe
412
+ 400 Oe
413
+ 600 Oe
414
+ 800 Oe
415
+
416
+
417
+
418
+
419
+
420
+ At  = 23 the extracted value of g13 is 0.592 GHz, while the values of 1 and 3 are found to
421
+ be 0.60 GHz and 0.711 GHz, respectively. Since g13  1 and 3, therefore the interaction
422
+ between M1 and M3 can be considered as weak coupling. In the opposite case, i.e. when g13 >
423
+ 1 and 3 it will be considered as strong coupling between two SW branches. We have also
424
+ calculated magnon–magnon cooperativity (C), which is defined as C = g2/() (, = 1, 2,
425
+ 3) and obtained C13 = 0.821 for the coupling between M1 and M3. The extracted value of g,
426
+ k, k, and the estimated value of C for anticrossing points corresponds to different bias field
427
+
428
+
429
+ g13
430
+ (GHz)
431
+ g12
432
+ (GHz)
433
+ g23
434
+ (GHz)
435
+ 1(GHz)
436
+ 2(GHz)
437
+ 3(GHz)
438
+ C13
439
+ C12
440
+ C23
441
+ 23o
442
+ 0.592
443
+ -
444
+ -
445
+ 0.60
446
+ -
447
+ 0.711
448
+ 0.821
449
+
450
+
451
+ 30o
452
+ 0.82
453
+ -
454
+ -
455
+ 0.423
456
+ -
457
+ 0.660
458
+ 2.515
459
+
460
+
461
+ 45o
462
+ -
463
+ 0.745
464
+ 0.255
465
+ 0.426
466
+ 0.645
467
+ 0.645
468
+ -
469
+ 2.019
470
+ 0.113
471
+ 60o
472
+ -
473
+ 0.915
474
+ 0.205
475
+ 1.35
476
+ 0.69
477
+ 0.707
478
+ -
479
+ 0.878
480
+ 0.675
481
+
482
+ Table 1 The extracted values of coupling strength (g), FWHM (2k) and calculated cooperativity factor
483
+ (C) for different orientation of bias field at the anticrossing points. Values of g and k are extracted
484
+ from the FMR spectra).
485
+
486
+
487
+ FIG. 4. Real part of S11 parameter as a function of frequency to highlight the anticrossing field are
488
+ shown for  = (a) 23°. The frequency gap in the anticrossing mode reveals the coupling strength g. (b)
489
+ Variation of cooperativity factor with the orientation of bias field. It shows that coupling strength is
490
+ stronger at  = 30 and 45. The schematic of Hext are shown on the right side of the figure.
491
+
492
+
493
+
494
+
495
+
496
+ 8
497
+ 10
498
+ 0.0
499
+ 0.3
500
+ 0.6
501
+ 0.9
502
+
503
+
504
+  = 23
505
+ 1062 Oe
506
+ Frequency (GHz)
507
+ Re S11 (Normalized)
508
+ x
509
+ y
510
+ Hext
511
+
512
+ 2k1
513
+ 2k3
514
+ 2g
515
+ (a)
516
+ 20
517
+ 40
518
+ 60
519
+ 0
520
+ 1
521
+ 2
522
+ 3
523
+ C13
524
+ C23
525
+ C12
526
+
527
+
528
+  (Degree)
529
+ Cooperativity
530
+ (b)
531
+
532
+ angles are listed in Table 1. At  = 30 obtained value for g13, 1, 3 and C13 are estimated
533
+ 0.82, 0.423, 0.66, and 2.515, respectively and here this magnon-magnon coupling falls in the
534
+ strong coupling regime. From Table 1, we can see that first anticrossing at  = 45 also shows
535
+ strong magnon-magnon coupling with C = 2.019, while the second one shows weak interaction.
536
+ At  = 60 both the interactions are in the weak coupling regime. Fig. 4(b) shows the -
537
+ dependence of the C where it shows the tunability of coupling strength with the in-plane
538
+ magnetic field orientation. It also exhibits that the interaction between different SW branches
539
+ show strong coupling in-between 30 to 45 orientation.
540
+
541
+ 3.2. Micromagnetic Simulation
542
+ 3.2.1. Static Magnetic Configuration
543
+
544
+ In Fig. 1(d) at  = 0, a sharp minimum is observed which gradually vanishes for higher values
545
+ of . The answer to this lies in the nanodot structure and its rich and flexible spin configurations
546
+ which we have simulated using OOMMF software[36]. Details of the micromagnetic
547
+ simulations are given in section S3 of the Supplementary Materials. The simulations reproduce
548
+ important features of the experimental SW spectra with nearly identical frequencies and
549
+ number of modes besides their relative intensity variations. The simulated static spin textures
550
+ within the nanomagnet array for different bias field magnitudes Hext at  = 0 and 45 are shown
551
+ in Fig. 5. At  = 0, the nanodot structure shows drastic variation in spin configurations with
552
+ Hext. It shows the formation of an S-state at the lower field regime (Hext = 100 Oe) as shown in
553
+ Fig. 5. At larger bias fields (e.g., Hext = 800 Oe), the spins are nearly aligned along the bias-
554
+ field direction (x-axis) and switch to a leaf-state (Fig. 5). This transformation from S- to leaf-
555
+ state occurs for 250 Oe ≤ Hext ≤ 350 Oe, where the SW frequency shows a minimum as a
556
+ function of Hext. At  = 45 , this transformation is not observed. Here, for the entire field
557
+ range, the static magnetic configuration shows a leaf state.
558
+
559
+
560
+
561
+ 3.2.2. SW mode Characterization
562
+
563
+ To interpret the nature of the SW modes, we have further simulated the spatial profiles of power
564
+ and phase of each SW mode by using a home-built MATLAB based code Dotmag[37].
565
+ OOMMF simulation provides magnetization (M (r, t)) information of each rectangular prism-
566
+ like cell at different simulation times. By performing discrete Fourier transformation with
567
+ respect to time in each of these cells and subsequently extracting the power and phase of the
568
+ dynamic magnetization for a desired frequency gives rise to the spatial distribution of the power
569
+ phase profile for that particular mode. In Fig. 6, we have shown the power distribution profile
570
+ of SW mode at  = 45 orientation for five different fields, Hext = 200 Oe (Hext << Hac), 400
571
+ Oe (Hext < Hac), 475 Oe (Hac), 600 Oe (Hext  Hac) and 1000 Oe (Hext >> Hac), while the phase
572
+ profile for each case is shown in the inset. The power profile at Hext = 1000 Oe indicates that
573
+ at high bias field only existing mode is M3, which is boosted by all the available energy. With
574
+ a gradual decrement of bias field, two additional modes M1 and M2 appear and the power of
575
+
576
+
577
+
578
+ FIG. 5. Simulated static magnetic configurations for Py nanodot array at four different bias magnetic-
579
+ field magnitude (Hext) at  = 0 and  = 45. We have shown here a single nanodot from the center of
580
+ the array for clarity in spin configurations. The nanodot structure shows a drastic variation in spin
581
+ configurations with bias magnetic-field strength.
582
+
583
+
584
+
585
+
586
+
587
+
588
+
589
+
590
+
591
+
592
+
593
+
594
+
595
+
596
+
597
+
598
+
599
+
600
+
601
+
602
+
603
+ 100 Oe
604
+ 250 Oe
605
+ 800 Oe
606
+ 350 Oe
607
+
608
+ x
609
+ y
610
+ Hext
611
+
612
+ 0
613
+ 45
614
+ -Y
615
+ +Y
616
+ 0.0
617
+ 0.3
618
+ 0.6
619
+ 0.9
620
+ 1.2
621
+ M1
622
+ M2
623
+ M3
624
+ M4
625
+ M5
626
+ M'
627
+
628
+
629
+ 0.0
630
+ 0.3
631
+ 0.6
632
+ 0.9
633
+ 1.2
634
+ M1
635
+ M2
636
+ M3
637
+ M4
638
+ M5
639
+ M6
640
+ M7
641
+
642
+
643
+ 0.0
644
+ 0.3
645
+ 0.6
646
+ 0.9
647
+ 1.2
648
+ 3
649
+ 6
650
+ 9
651
+ 12
652
+ M1
653
+ M2
654
+ M3
655
+ M4
656
+ M5
657
+ M6
658
+
659
+
660
+ Frequency (GHz)
661
+ H1
662
+ H2
663
+ 0.0
664
+ 0.3
665
+ 0.6
666
+ 0.9
667
+ 1.2
668
+ M*
669
+ M1
670
+ M2
671
+ M3
672
+ M4
673
+ M5
674
+ M6
675
+ M7
676
+
677
+
678
+ H3
679
+ 0.0
680
+ 0.3
681
+ 0.6
682
+ 0.9
683
+ 1.2
684
+ M '
685
+ M1
686
+ M2
687
+ M3
688
+ M4
689
+ M5
690
+ M6
691
+
692
+
693
+ Applied Field Hext (kOe)
694
+ 0.0
695
+ 0.3
696
+ 0.6
697
+ 0.9
698
+ 1.2
699
+ 3
700
+ 6
701
+ 9
702
+ 12
703
+ M1
704
+ M2
705
+ M3
706
+ M4
707
+
708
+
709
+ O1
710
+ O2
711
+ O3
712
+ H2
713
+ Fig 2
714
+ Hext
715
+ x
716
+ y
717
+ 1
718
+ 0
719
+ Re (S11)
720
+ Normalised
721
+
722
+ M3 is gradually transferred to these two modes. At the anticrossing field, Hext = 475 Oe, M2
723
+ appears as the most intense mode although M1 and M3 have significant power at this field. At
724
+ lower fields, this power is gradually transferred to M1, and at 200 Oe, barring M1 other modes
725
+
726
+
727
+
728
+ FIG. 6. Simulated spatial distribution of power and phase (in the inset) profiles corresponding to
729
+ different SW modes at five different bias field values for  = 45 for the Py nanodot array. The
730
+ applied field direction is shown at the bottom left of the figure. Symbols with different colors
731
+ represent different SW modes. The color map is shown at the upper right side of the figure.
732
+
733
+
734
+
735
+
736
+
737
+
738
+
739
+
740
+
741
+
742
+
743
+
744
+
745
+
746
+
747
+ 200 Oe
748
+ 400 Oe
749
+ 475 Oe
750
+ 600 Oe
751
+ 1000 Oe
752
+ M1
753
+ M2*
754
+ M3
755
+ 20
756
+ 0
757
+ Power
758
+ (dB)
759
+ Phase
760
+ (rad)
761
+ +
762
+ -
763
+ x
764
+ y Hext
765
+
766
+  = 45
767
+
768
+ disappear. Similar to this energy exchange, the phase profiles also exhibit interchange of mode
769
+ behavior. At high bias fields (e.g., 1000 Oe), M3 shows quantized nature in BV-like geometry
770
+ with a quantization number n = 3. With a decrease in the field, this mode gradually transforms
771
+ into higher-order quantized mode and M2* is transformed into a quantized mode with n = 3.
772
+ At Hext = 475 Oe the quantization number of M1, and M3 are n = 5, and 7, respectively, while
773
+ for M2*, n = 3, which is identical to the quantization number of M3 at Hext = 1000 Oe. This
774
+ transformation of mode quantization number is also seen in-between M1 and M2* as we further
775
+ reduce the bias field and finally at Hext = 200 Oe, M1 shows a quantized behavior with n = 3.
776
+ This transformation of power as well as mode property from one branch of SW mode to another
777
+ at the anticrossing region indicates a strong interaction between these modes. For other
778
+ orientation like  = 23, 30 and 60, similar kind of behavior are observed, which are shown
779
+ in section S6 of the Supplementary Materials.
780
+
781
+ 3.2.3. Distribution of Exchange field
782
+ To understand the origin of the magnon-magnon coupling and its modulation with bias
783
+ magnetic field, we have simulated the spatial distribution of the dipole-exchange field
784
+ (Exchange field distribution of each dot, which is modulated by dipolar interaction of nanodot
785
+ array) lines at the equilibrium for different bias field orientations. Fig. 7 shows the exchange
786
+ field map of nanodots array at eight different fields for  = 45 orientation. Due to inter-dot
787
+ dipolar interactions, a dynamic variation of exchange field line with the bias field amplitude
788
+ (for better viewing purpose, we just present a single nanodot) is observed. The Supplementary
789
+ Movie A1 shows the dynamics of this exchange field in more detail. At lower bias fields (Hext
790
+ << Hac), due to dominating effect of demagnetizing field, spins take a configuration such that
791
+ at equilibrium condition the exchange field lines create three different regions within a single
792
+ dot. The field lines of center and edge regions are configured in opposite direction as denoted
793
+ with yellow and green arrows in Fig. 7(a). As we increase the bias field, the region around the
794
+ edge of the dot start to vanish and the center region gradually expands. At a very high bias field
795
+ (Hext  Hac), e.g., Hext = 1000 Oe, only the central region with unidirectional field lines are
796
+ observed inside a dot. This transformation from three mutually opposite (antiparallel) field-line
797
+ configuration to uniform (parallel) configuration occurs for 450 Oe ≤ Hext ≤ 500 Oe, which is
798
+ exactly the anticrossing field region for  = 45 orientation. This change in exchange field
799
+ profile can be observed much more clearly if we take a linescan along the bias field direction
800
+ (white dotted line in Fig. 7(a)) as shown in Fig. 7(b). In the inset, we have magnified the end
801
+
802
+ part of the linescan. Here, it is clearly visible that below the anticrossing field (Hext = Hac = 475
803
+ Oe) the linescan has two different local maxima
804
+
805
+ which transform into one maximum as we increase Hext. The exchange field profile for other
806
+ values of  are shown in section S7 of the Supplementary Materials, where similar
807
+ transformation is observed in the anticrossing field region. Our observation of correlation
808
+ between these two phenomena indicates that the anticrossing gap appears only when such a
809
+ variation of exchange field occurs due to the bias field strength as well as its orientation. The
810
+ internal field distribution in presence and absence of the exchange field leads to similar
811
+ conclusion, which we have described in section S8 of the Supplementary Materials.
812
+
813
+ 4. Conclusion
814
+ In summary, the interaction between magnons confined in a sole magnonic cavity has been
815
+ realized in the strong coupling regime. We have investigated a bias field strength and angle-
816
+ dependent magnetization dynamics in diamond-shaped Py nanodot arrays using the broadband
817
+ ferromagnetic resonance technique. Our study has demonstrated that the coupling between two
818
+ magnon modes is mediated by the exchange coupling inside individual nanodot. Furthermore,
819
+ the coupling strength is found to be highly dependent on the orientation and strength of
820
+ the bias magnetic field, leading towards the possibility of externally controlled hybrid
821
+
822
+
823
+
824
+
825
+ FIG. 7. Exchange field distributions for (a) single nanodot for eight different bias field values at 
826
+ = 45 . Yellow and green arrows represent the direction of exchange field at the center and edge
827
+ position of the nanodot. We have shown here a single nanodot from the center of the array for clarity
828
+ in spin configurations. The color bars are shown at the right side of the figure. (b) Linescan of the
829
+ simulated exchange field for nanodot array along the field direction. In the inset magnified portion
830
+ of simulated exchange field is shown.
831
+
832
+
833
+
834
+
835
+
836
+
837
+
838
+
839
+
840
+
841
+
842
+
843
+
844
+
845
+
846
+
847
+
848
+
849
+
850
+
851
+
852
+ 360
853
+ 420
854
+ 480
855
+ 540
856
+ 0
857
+ 250
858
+ 500
859
+ 1200 Oe
860
+ 550 Oe
861
+ 450 Oe
862
+ 200 Oe
863
+
864
+
865
+ Exchange Field (Oe)
866
+ Distance (nm)
867
+ x
868
+ y
869
+ Hext
870
+
871
+ (a)
872
+ (b)
873
+ 200 Oe
874
+ 300 Oe
875
+ 450 Oe
876
+ 500 Oe
877
+ 550 Oe
878
+ 700 Oe
879
+ 800 Oe
880
+ 1000 Oe
881
+ -8
882
+ -6
883
+ -4
884
+ -2
885
+ 1200 Oe
886
+ 550 Oe
887
+ 450 Oe
888
+ 200 Oe
889
+
890
+
891
+ Power(dB)
892
+ 0.0
893
+ 0.3
894
+ 0.6
895
+ 0.9
896
+ 1.2
897
+ M1
898
+ M2
899
+ M3
900
+ M4
901
+ M5
902
+ M'
903
+
904
+
905
+ 0.0
906
+ 0.3
907
+ 0.6
908
+ 0.9
909
+ 1.2
910
+ M1
911
+ M2
912
+ M3
913
+ M4
914
+ M5
915
+ M6
916
+ M7
917
+
918
+
919
+ 0.0
920
+ 0.3
921
+ 0.6
922
+ 0.9
923
+ 1.2
924
+ 3
925
+ 6
926
+ 9
927
+ 12
928
+ M1
929
+ M2
930
+ M3
931
+ M4
932
+ M5
933
+ M6
934
+
935
+
936
+ Frequency (GHz)
937
+ H1
938
+ H2
939
+ 0.0
940
+ 0.3
941
+ 0.6
942
+ 0.9
943
+ 1.2
944
+ M*
945
+ M1
946
+ M2
947
+ M3
948
+ M4
949
+ M5
950
+ M6
951
+ M7
952
+
953
+
954
+ H3
955
+ 0.0
956
+ 0.3
957
+ 0.6
958
+ 0.9
959
+ 1.2
960
+ M '
961
+ M1
962
+ M2
963
+ M3
964
+ M4
965
+ M5
966
+ M6
967
+
968
+
969
+ Applied Field Hext (kOe)
970
+ 0.0
971
+ 0.3
972
+ 0.6
973
+ 0.9
974
+ 1.2
975
+ 3
976
+ 6
977
+ 9
978
+ 12
979
+ M1
980
+ M2
981
+ M3
982
+ M4
983
+
984
+
985
+ O1
986
+ O2
987
+ O3
988
+ H2
989
+ Fig 2
990
+ Hext
991
+ x
992
+ y
993
+ 1
994
+ 0
995
+ Re (S11)
996
+ Normalised
997
+
998
+ 800 Oe7000e600 0e200 0e500 0emagnonic devices. The experimental results have been well reproduced by micromagnetic
999
+ simulation. The power and phase profiles of the resonant modes have been numerically
1000
+ calculated to gain insight into the spatial nature of the dynamics. The transformation of power
1001
+ as well as mode property from one branch of SW to another, apparently support the strong
1002
+ interaction in-between these modes. Numerical study shows that the anticrossing gap appears
1003
+ when the symmetry of exchange configuration inside each nanodot is broken due to the applied
1004
+ bias magnetic field. We have also observed mode softening phenomena when the static
1005
+ magnetic configuration switches from the S-state to the leaf state and with the variation of bias
1006
+ field angle it gradually disappears. Our findings offer a new approach toward tunable magnon-
1007
+ magnon coupling in ferromagnetic nanostructures for applications in quantum transduction
1008
+ using magnons.
1009
+
1010
+
1011
+
1012
+ 5. Acknowledgements
1013
+
1014
+ AB gratefully acknowledges the financial support from S. N. Bose National Centre for
1015
+ Basic Sciences, India (Grant No. SNB/AB/18-19/211). SB acknowledges Science and
1016
+ Engineering Research Board (SERB), India for funding (Grant no. CRG/2018/002080). SM
1017
+ and SC acknowledge S. N. Bose National Centre for Basic Sciences for senior research
1018
+ fellowship
1019
+
1020
+
1021
+ References
1022
+
1023
+ [1]
1024
+ A. A. Clerk, K. W. Lehnert, P. Bertet, J. R. Petta, and Y. Nakamura, Hybrid quantum systems
1025
+ with circuit quantum electrodynamics, Nature Physics 16, 257 (2020).
1026
+ [2]
1027
+ J. P. Home, D. Hanneke, J. D. Jost, J. M. Amini, D. Leibfried, and D. J. Wineland, Complete
1028
+ Methods Set for Scalable Ion Trap Quantum Information Processing, 325, 1227 (2009).
1029
+ [3]
1030
+ A. Blais, R.-S. Huang, A. Wallraff, S. M. Girvin, and R. J. Schoelkopf, Cavity quantum
1031
+ electrodynamics for superconducting electrical circuits: An architecture for quantum computation,
1032
+ Physical Review A 69, 062320 (2004).
1033
+ [4]
1034
+ R. P. Feynman, Quantum mechanical computers, Foundations of Physics 16, 507 (1986).
1035
+ [5]
1036
+ T. D. Ladd, F. Jelezko, R. Laflamme, Y. Nakamura, C. Monroe, and J. L. O'Brien, Quantum
1037
+ computers, Nature 464, 45 (2010).
1038
+ [6]
1039
+ H. J. Kimble, The quantum internet, Nature 453, 1023 (2008).
1040
+ [7]
1041
+ A. Reiserer and G. Rempe, Cavity-based quantum networks with single atoms and optical
1042
+ photons, Reviews of Modern Physics 87, 1379 (2015).
1043
+ [8]
1044
+ C. L. Degen, F. Reinhard, and P. Cappellaro, Quantum sensing, Reviews of modern physics
1045
+ 89, 035002 (2017).
1046
+ [9]
1047
+ Y. Kubo et al., Strong coupling of a spin ensemble to a superconducting resonator, Physical
1048
+ review letters 105, 140502 (2010).
1049
+
1050
+ [10]
1051
+ D. Schuster et al., High-cooperativity coupling of electron-spin ensembles to superconducting
1052
+ cavities, Physical review letters 105, 140501 (2010).
1053
+ [11]
1054
+ Ö. O. Soykal and M. Flatté, Strong field interactions between a nanomagnet and a photonic
1055
+ cavity, Physical review letters 104, 077202 (2010).
1056
+ [12]
1057
+ Y. Cao, P. Yan, H. Huebl, S. T. Goennenwein, and G. E. Bauer, Exchange magnon-polaritons
1058
+ in microwave cavities, Physical Review B 91, 094423 (2015).
1059
+ [13]
1060
+ S. Klingler et al., Spin-torque excitation of perpendicular standing spin waves in coupled
1061
+ YIG/Co heterostructures, Physical review letters 120, 127201 (2018).
1062
+ [14]
1063
+ H. Qin, S. J. Hämäläinen, and S. Van Dijken, Exchange-torque-induced excitation of
1064
+ perpendicular standing spin waves in nanometer-thick YIG films, Scientific reports 8, 1 (2018).
1065
+ [15]
1066
+ J. Chen, C. Liu, T. Liu, Y. Xiao, K. Xia, G. E. Bauer, M. Wu, and H. Yu, Strong interlayer
1067
+ magnon-magnon coupling in magnetic metal-insulator hybrid nanostructures, Physical review letters
1068
+ 120, 217202 (2018).
1069
+ [16]
1070
+ J. Chen et al., Excitation of unidirectional exchange spin waves by a nanoscale magnetic
1071
+ grating, Physical Review B 100, 104427 (2019).
1072
+ [17]
1073
+ J. Chen et al., Excitation of unidirectional exchange spin waves by a nanoscale magnetic
1074
+ grating, Physical Review B 100, 104427 (2019).
1075
+ [18]
1076
+ L. Liensberger et al., Exchange-enhanced ultrastrong magnon-magnon coupling in a
1077
+ compensated ferrimagnet, Physical review letters 123, 117204 (2019).
1078
+ [19]
1079
+ D. MacNeill, J. T. Hou, D. R. Klein, P. Zhang, P. Jarillo-Herrero, and L. Liu, Gigahertz
1080
+ frequency antiferromagnetic resonance and strong magnon-magnon coupling in the layered crystal
1081
+ crcl 3, Physical review letters 123, 047204 (2019).
1082
+ [20]
1083
+ Y. Tabuchi, S. Ishino, T. Ishikawa, R. Yamazaki, K. Usami, and Y. Nakamura, Hybridizing
1084
+ ferromagnetic magnons and microwave photons in the quantum limit, Physical review letters 113,
1085
+ 083603 (2014).
1086
+ [21]
1087
+ G. Agarwal, Vacuum-field Rabi splittings in microwave absorption by Rydberg atoms in a
1088
+ cavity, Physical review letters 53, 1732 (1984).
1089
+ [22]
1090
+ Y. Tabuchi, S. Ishino, A. Noguchi, T. Ishikawa, R. Yamazaki, K. Usami, and Y. Nakamura,
1091
+ Coherent coupling between a ferromagnetic magnon and a superconducting qubit, Science 349, 405
1092
+ (2015).
1093
+ [23]
1094
+ R. Hisatomi, A. Osada, Y. Tabuchi, T. Ishikawa, A. Noguchi, R. Yamazaki, K. Usami, and Y.
1095
+ Nakamura, Bidirectional conversion between microwave and light via ferromagnetic magnons,
1096
+ Physical Review B 93, 174427 (2016).
1097
+ [24]
1098
+ D. Lachance-Quirion, Y. Tabuchi, A. Gloppe, K. Usami, and Y. Nakamura, Hybrid quantum
1099
+ systems based on magnonics, Applied Physics Express 12, 070101 (2019).
1100
+ [25]
1101
+ J. Park, P. Eames, D. Engebretson, J. Berezovsky, and P. Crowell, Spatially resolved
1102
+ dynamics of localized spin-wave modes in ferromagnetic wires, Physical review letters 89, 277201
1103
+ (2002).
1104
+ [26]
1105
+ J. Jorzick, S. Demokritov, B. Hillebrands, M. Bailleul, C. Fermon, K. Y. Guslienko, A.
1106
+ Slavin, D. Berkov, and N. Gorn, Spin wave wells in nonellipsoidal micrometer size magnetic
1107
+ elements, Physical Review Letters 88, 047204 (2002).
1108
+ [27]
1109
+ M. Bailleul, R. Höllinger, and C. Fermon, Microwave spectrum of square Permalloy dots:
1110
+ Quasisaturated state, Physical Review B 73, 104424 (2006).
1111
+ [28]
1112
+ G. Carlotti, Pushing down the lateral dimension of single and coupled magnetic dots to the
1113
+ nanometric scale: Characteristics and evolution of the spin-wave eigenmodes, Applied Physics
1114
+ Reviews 6, 031304 (2019).
1115
+ [29]
1116
+ Z. Zhang, M. Vogel, M. B. Jungfleisch, A. Hoffmann, Y. Nie, and V. Novosad, Tuning edge-
1117
+ localized spin waves in magnetic microstripes by proximate magnetic structures, Physical Review B
1118
+ 100, 174434 (2019).
1119
+ [30]
1120
+ C. Dai, K. Xie, Z. Pan, and F. Ma, Strong coupling between magnons confined in a single
1121
+ magnonic cavity, Journal of Applied Physics 127, 203902 (2020).
1122
+ [31]
1123
+ K. Adhikari, S. Sahoo, A. K. Mondal, Y. Otani, and A. Barman, Large nonlinear
1124
+ ferromagnetic resonance shift and strong magnon-magnon coupling in N i 80 F e 20 nanocross array,
1125
+ Physical Review B 101, 054406 (2020).
1126
+ [32]
1127
+ K. Adhikari, S. Choudhury, S. Barman, Y. Otani, and A. Barman, Observation of magnon–
1128
+
1129
+ magnon coupling with high cooperativity in Ni80Fe20 cross-shaped nanoring array, Nanotechnology
1130
+ 32, 395706 (2021).
1131
+ [33]
1132
+ S. Choudhury, S. Majumder, S. Barman, Y. Otani, and A. Barman, Active control of mode
1133
+ crossover and mode hopping of spin waves in a ferromagnetic antidot lattice, Physical Review
1134
+ Applied 10, 064044 (2018).
1135
+ [34]
1136
+ S. Majumder, S. Choudhury, S. Barman, Y. Otani, and A. Barman, Reconfigurable spin-wave
1137
+ dynamics in two-dimensional quasiperiodic magnonic crystals, Physica E: Low-dimensional Systems
1138
+ Nanostructures 134, 114901 (2021).
1139
+ [35]
1140
+ B. Mahato, S. Choudhury, R. Mandal, S. Barman, Y. Otani, and A. Barman, Tunable
1141
+ configurational anisotropy in collective magnetization dynamics of Ni80Fe20 nanodot arrays with
1142
+ varying dot shapes, Journal of Applied Physics 117, 213909 (2015).
1143
+ [36]
1144
+ M. Donahue, D. Porter, J. Lau, and R. McMichael, Interagency report NISTIR 6376. National
1145
+ institute of standards and technology, Gaithersburg, NIST J. Res. 114, 57 (1999).
1146
+ [37]
1147
+ D. Kumar, O. Dmytriiev, S. Ponraj, and A. Barman, Numerical calculation of spin wave
1148
+ dispersions in magnetic nanostructures, Journal of Physics D: Applied Physics 45, 015001 (2011).
1149
+
1150
+
1151
+
6NFJT4oBgHgl3EQflSzb/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
79E0T4oBgHgl3EQfwQGR/content/tmp_files/2301.02630v1.pdf.txt ADDED
@@ -0,0 +1,1079 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.02630v1 [math.AG] 15 Aug 2022
2
+ HEIGHT PAIRING AND NEARBY CYCLES
3
+ A. Beilinson
4
+ To Yuri Ivanovich Manin with deepest gratitude
5
+ Abstract. We prove that, as was conjectured by Spencer Bloch, the Hodge period
6
+ of some limit Hodge structures equals the height pairing of algebraic cycles on the
7
+ resolution of singularities of the singular fiber.
8
+ §1. Introduction: the theorem and the idea of the proof
9
+ 1.1. The Hodge period.
10
+ Suppose we have a Q-Hodge structure E with weights
11
+ in [−2, 0] equiped with isomorphisms ι0 : grW
12
+ 0 E = Q(0), ι−2 : grW
13
+ −2E = Q(1).
14
+ One defines the Hodge period ⟨E⟩ = ⟨E, ι0, ι−2⟩ ∈ R as follows.
15
+ Consider the
16
+ R-Hodge structure E ⊗ R. Since the weight filtration on any R-Hodge structure
17
+ with two consequitive weights (canonically) splits one has E ⊗ R = G ⊕ grW
18
+ −1E ⊗ R
19
+ where G is an extension of R(0) by R(1). Our ⟨E⟩ is the class of this extension in
20
+ Ext1(R(0), R(1)) = R.
21
+ Remark. One computes ⟨E⟩ explicitly as follows. Let ER be E ⊗ R viewed a plain
22
+ R-vector space, EC be its complexification. Let 1F 0 ∈ F 0 ⊂ EC be any lifting of
23
+ ι−1
24
+ 0 (1). Then ⟨E⟩ is the image of 1F 0 in (ER + W−1EC)/(ER + (F 0 ∩ W−1EC))
25
+
26
+
27
+ W−2EC/W−2ER = C/2πiR
28
+
29
+ ← R.
30
+ 1.2. A geometric example. Let Y be a smooth proper equidimensional algebraic
31
+ variety over C. We denote by Hi(Y ) the homology of Hi(Y (C), Q) seen as an object
32
+ of the category of Q-Hodge structures; ditto for relative homology, etc. Let Zm(Y )
33
+ be the group of algebraic m-cycles on Y with Q-coefficients, Zm(Y )0 := Ker(cl :
34
+ Zm(Y ) → H2m(Y )(−m)) be the subgroup of cycles homologically equivalent to
35
+ zero. For a closed subset P ⊂ Y let Zm(P) ⊂ Zm(Y ) be the subgroup of cycles
36
+ supported on P, Zm(P)0 := Zm(P) ∩ Zm(Y )0. For an m-cycle A on Y we denote
37
+ by |A| its support (which is a closed subset of Y ).
38
+ Suppose m + m′ = dim Y − 1 and we have A ∈ Zm(Y )0, B ∈ Zm′(Y )0 such that
39
+ |A| ∩ |B| = ∅. Set E|A|,|B| := H2m+1(Y ∖ |B|, |A|)(−m). Notice that E|B|,|A| =
40
+ E∗
41
+ |A|,|B|(1) by the Poicar´e duality.
42
+ Lemma. E|A|,|B| has weights in [−2, 0]. One has grW
43
+ −2E|A|,|B| = Zm′(|B|)∗
44
+ 0(1),
45
+ grW
46
+ −1E|A|,|B| = H2m+1(Y )(−m), grW
47
+ 0 E|A|,|B| = Zm(|A|)0.
48
+ Proof. Notice that H2m(|A|)(−m) = Zm(|A|) and H>2m(|A|) = 0. By the Poincar´e
49
+ duality Hi(Y, Y ∖|B|)(− dim Y ) = H2 dim Y −i(|B|)∗, hence H2m+2(Y, Y ∖|B|)(−m) =
50
+ 1991 Mathematics Subject Classification. Primary 14C25; Secondary 14D07.
51
+ Key words and phrases. height pairing, nearby cycles, Hodge periods.
52
+ Typeset by AMS-TEX
53
+ 1
54
+
55
+ 2
56
+ A. BEILINSON
57
+ (H2m′(|B|)(−m′))∗(1) and H<2m+2(Y, Y ∖ |B|) = 0. Now use the long exact ho-
58
+ mology sequences for (Y ∖ |B|, |A|) and (Y, Y ∖ |B|).
59
+
60
+ Denote by EA,B the Hodge structure obtained from H|A|,|B| by pullback by A
61
+ and pushforward by B:
62
+ (1.2.1)
63
+ Zn(|B|)∗
64
+ 0(1)
65
+ ֒→
66
+ E|A|,|B|
67
+ ։
68
+ Zm(|A|)0
69
+ B ↓
70
+ ↑ A
71
+ Q(1)
72
+ ֒→
73
+ EA,B
74
+ ։
75
+ Q(0)
76
+ Our EA,B is as in 1.1, so we have ⟨EA,B⟩ ∈ R.
77
+ 1.3. The height pairing (cf. [B], [Bl1]). Let k be a subfield of C and suppose that
78
+ Y comes from a variety Yk over k, Y = Yk ⊗ C. Let Zm(Yk) ⊂ Zm(Y ) be the
79
+ group of algebraic cycles with Q-coefficients on Yk, Zm(Yk)0 := Zm(Yk) ∩ Zm(Y )0,
80
+ and let CHm(Yk)0 ⊂ CHm(Yk) be their quotients modulo the rational equivalence
81
+ relation. One checks (see §2) that if A, B as above are cycles on Yk then the class
82
+ of ⟨EA,B⟩ in R/Q log |k×| depends only on linear equivalence classes of A and B,
83
+ and so one has a bilinear height pairing
84
+ (1.3.1)
85
+ ⟨ , ⟩Yk : CHm(Yk)0 ⊗ CHm′(Yk)0 → R/Q log |k×|.
86
+ Namely ⟨a, b⟩Yk = ⟨EA,B⟩ where A, B are any cycles on Yk of classes a, b such that
87
+ |A| ∩ |B| = ∅.
88
+ Remark. If k = Q and we assume some motivic rationality conjectures (see (2.2.1),
89
+ (2.2.3) of [B]) then ⟨EA,B⟩ can be corrected (by adding a finite sum of corrections
90
+ log(p)⟨EA,B⟩p where p is a prime, ⟨EA,B⟩p is defined using the Gal(Qp)-action on
91
+ EA,B ⊗ Qℓ) so that the resulting real number depends only on rational equivalence
92
+ classes of A and B. In this manner (1.3.1) lifts naturally to an R-valued pairing.
93
+ 1.4. Finding elements of Chow groups that are homologically equivalent to zero
94
+ is an art. Spencer Bloch described one situation where they naturally arise, and
95
+ conjectured that the height pairing of his cycles can be computed in s different way,
96
+ namely, as Hodge periods of some nearby cycles. We start with preliminaries.
97
+ Let X be a smooth variety over C of pure dimension n ≥ 2, S be a smooth curve,
98
+ 0 ∈ S be a closed point, and f : X → S be a proper map which is smooth otside
99
+ a finite subset {xα} of the fiber X0 = f −1(0). Let Zα be the projectivized tangent
100
+ cone to X0 at xα; this is a hypersurface in the projectivization Pα := P(TxαX) of
101
+ the tangent space; denote by dα its degree. We assume the next condition:
102
+ (∗) All hypersurfaces Zα are smooth.
103
+ Let π : Y → X0 be the blowup of X0 at {xα}. Condition (∗) implies that Y is
104
+ a smooth variety, and Zα are pairwise disjoint divisors on Y . Set Z := ⊔Zα and
105
+ K := Ker(Hn−2(Z) → Hn−2(Y )) = Im(Hn−1(Y, Z) → Hn−2(Z)). If n = 2 then let
106
+ K0 ⊂ K be the subgroup of those elements A = ΣAα that deg Aα = 0 for every
107
+ α. One has a natural map Hn−1(X0) → Hn−1(Y, Z) defined as the composition
108
+ Hn−1(X0) → Hn−1(X0, {xα})
109
+
110
+ ← Hn−1(Y, Z).
111
+ Lemma. (i) The map Hn−1(X0) → Hn−1(Y, Z) is an isomorphism if n > 2. If
112
+ n = 2 it is injective and its image equals the preimage of K0 in Hn−1(Y, Z).
113
+ (ii) Hn−1(Y, Z) has weights 1 − n and 2 − n, and grW
114
+ 2−nHn−1(Y, Z) = K. The map
115
+
116
+ HEIGHT PAIRING AND NEARBY CYCLES
117
+ 3
118
+ Hn−1(Y ) → Hn−1(Y, Z) has image W1−nHn−1(Y, Z). If n is even then Hn−1(Y )
119
+
120
+
121
+ W1−nHn−1(Y, Z).
122
+ Proof. (i) Replace Hn−1(Y, Z) by Hn−1(X0, {xα}) and use the long exact homology
123
+ sequence. (ii) The first assertions follow from the exact homology sequence and
124
+ purity of weights on H·(Y ), H·(Z). The last one comes because Hn−1(Z) = 0 if n
125
+ is even (since Zα are hypersurfaces).
126
+
127
+ 1.5. Consider a variation of Q-Hodge structures V on S ∖ {0} with fibers Vs =
128
+ Hn−1(Xs). One has a nondegenerate intersection pairing ( , ) : V ⊗ V → Q(n − 1).
129
+ Choose a parameter t at 0 ∈ S and consider the limiting (a.k.a. nearby cycles)
130
+ Hodge structure ψtV. Let ψun
131
+ t V be its direct summand where the monodromy acts
132
+ unipotently. Since ψun
133
+ t
134
+ commutes with duality, ( , ) yields self-duality pairing on
135
+ it that we denote again by ( , ). One has the log of monodromy morphism N =
136
+ NV : ψun
137
+ t V(1) → ψun
138
+ t V and the specialization morphism sp : ψun
139
+ t V → Hn−1(X0).
140
+ Let (ψun
141
+ t V)N := Coker(NV) be the monodromy coinvariants. The next assertion
142
+ follows from the local invariant cycles theorem, see 3.5 for a detailed proof:
143
+ Proposition. sp factors through the isomorphism (ψun
144
+ t V)N
145
+
146
+ → Hn−1(X0).
147
+ Corollary. ψun
148
+ t V has weights in [−n, 2 − n]. One has grW
149
+ 2−nψun
150
+ t V = K if n > 2
151
+ and grW
152
+ 2−nψun
153
+ t V = K0 if n = 2. By self-duality, grW
154
+ −nψun
155
+ t V = (grW
156
+ 2−nψun
157
+ t V)∗(n − 1).
158
+ If n is even then grW
159
+ 1−nψun
160
+ t V = Hn−1(Y ).
161
+ Proof. Since ψun
162
+ t V is self-dual and N is nilpotent, the claim follows from the propo-
163
+ sition and the lemma in 1.4.
164
+
165
+ 1.6. Bloch cycles. We are in the setting of 1.4; suppose n is even, n = 2m + 2. Let
166
+ A = ΣAα be an m-cycle on Z. We say that A is a Bloch cycle if it is homologically
167
+ equivalent to zero on Y , i.e., cl(A) lies in K(−m) ⊂ Hn−2(Z)(−m). If m = 0 then
168
+ we demand, in addition, that cl(A) ∈ K0 ⊂ K.
169
+ Lemma. If A is a Bloch cycle then each cl(Aα) ∈ Hn−2(Zα)(−m) is primitive.
170
+ Proof. The composition Hn−2(Z)(−m) → Hn−2(Y )(−m) → Hn−4(Zα)(−m + 1),
171
+ where the second arrow is the pullback by Zα ֒→ Y , sends any class c = Σcα to
172
+ cα ∩ c1(O(−1)) (for O(−1) is the normal bundle to Zα in Y ). This composition
173
+ kills cl(Aα) since the first arrow does.
174
+
175
+ If A, B are two Bloch cycles then we denote by Eψ
176
+ A,B = Eψ
177
+ A,B,t the Hodge struc-
178
+ ture obtained from ψun
179
+ t V(−m) by pullback by cl(A) and pushforward by cl(B)∗:
180
+ (1.6.1)
181
+ K∗(m + 1)
182
+
183
+ ψun
184
+ t V(−m)
185
+
186
+ K(−m)
187
+ cl(B)∗ ↓
188
+ ↑ cl(A)
189
+ Q(1)
190
+ ֒→
191
+
192
+ A,B
193
+ ։
194
+ Q(0)
195
+ Our Eψ
196
+ A,B is as in 1.1 so we have ⟨Eψ
197
+ A,B⟩ ∈ R.
198
+ 1.7. Examples. Consider the case when we have single singular point x0 ∈ X0 of
199
+ f and the singularity at x0 is quadratic. Then the monodromy action on ψtV is
200
+ unipotent, the only possible Bloch cycle is the difference A of the rulings of the
201
+ quadric Z0, and it is actually a Bloch cycle if and only if the monodromy action on
202
+ ψtV is nontrivial or, equivalently, the Hodge structure on Hn−1(X0) is not pure.
203
+
204
+ 4
205
+ A. BEILINSON
206
+ Lemma. (i) If m = 0 then the curve X0 can have either 1 or 2 irreducible compo-
207
+ nents, and A is a Bloch cycle if and only if X0 is irreducible.
208
+ (ii) If X/S is a family of quadratic hypersurfaces in Pn then A is not a Bloch cycle.
209
+ (iii) If X/S is a family of hypersurfaces of degree d on a given smooth projective
210
+ variety P then A is a Bloch cycle if d is large enough.
211
+ Proof. (i) is clear. (ii) follows since the global monodromy for quadratic hypersur-
212
+ faces is ±1, and so it can’t contain non-trivial unipotent local monodromy.
213
+ (iii) Consider the corresponding map r : S → B := {hypersurfaces of degree
214
+ d on P}. Since X is smooth r is transversal to the locus D ⊂ B of degenerate
215
+ hypersurfaces. Replacing S by a germ of another transversal to D that intersects
216
+ D near r(0) would not change the topology of X over a small disc around 0. So we
217
+ can assume that S is a Zariski open subset of the base of a Lefschetz pencil on P.
218
+ Then, since local monodromies of a Lefschetz pencil are all conjugate, triviality of
219
+ one local monodromy amounts to triviality of the global monodromy. Thus A is a
220
+ Bloch cycle if and only if the global monodromy on V is not trivial. Let us check
221
+ that this happens for large enough d.
222
+ If R ⊂ P is the axis of our pencil then H·(X) = H·(P) ⊕ H·−2(R)(−1), and so
223
+ hn−1,0(P) = hn−1,0(X) which equals hn−1,0(Xs) if the global monodromy is trivial.
224
+ Thus the monodromy is not trivial when hn−1,0(Xs) > hn−1,0(P). To finish the
225
+ argument it remains to notice that hn−1,0(Xs) ≥ dim(H0(P, Ωn
226
+ P (d))/H0(P, Ωn
227
+ P )),
228
+ and so it tends to ∞ when d → ∞.
229
+
230
+ 1.8. Statement of the theorem. Now suppose we have a subfield k ⊂ C and our
231
+ datum is defined over k, i.e., there is Xk/Sk, a closed point 0 of Sk, a parameter t
232
+ on Sk at 0, and Bloch cycles A, B on Zk such that X/S, etc., come by base change
233
+ k → C. Let a ∈ CHm(Yk)0, b ∈ CHm(Yk)0 be the classes of A and B. The next
234
+ result was conjectured by Spencer Bloch:
235
+ Theorem. One has ⟨a, b⟩Yk = ⟨Eψ
236
+ A,B⟩ mod Q log |k×|.
237
+ In case n = 1 the theorem was proven in [BlJS].
238
+ Remark. Suppose we are in the situation of Remark in 1.3. If ⟨Eψ
239
+ A,B⟩ is corrected
240
+ in the same way as was discussed there, then the theorem lifts to an equality of real
241
+ numbers. The proof does not change; we will not discuss it below.
242
+ 1.9. Reformulation of the theorem that discards Hodge periods; the idea of the
243
+ proof. Let A′, B′ be cycles on Yk of classes a, b such that |A′| ∩ |B′| = ∅ (no-
244
+ tice that they are, most probably, not supported on Zk). We want to show that
245
+ ⟨EA′,B′⟩ = ⟨Eψ
246
+ A,B⟩ (see 1.2, 1.6). Let us compare the Hodge structures E = EA′,B′
247
+ and Eψ = Eψ
248
+ A,B themselves. Their weights lie in [−2, 0], and one has a canonical
249
+ identification grW
250
+ · E = grW
251
+ · Eψ. Indeed, grW
252
+ 0 E(ψ) = Q(0), grW
253
+ −2E(ψ) = Q(1) by the
254
+ constructions, and grW
255
+ −1E = H2m+1(Y )(−m) = grW
256
+ −1E(ψ) by the lemma in 1.2, and
257
+ the one in 1.4 combined with the corollary in 1.5. This identification lifts (uniquely)
258
+ to W−1E = W−1Eψ and E/W−2E = Eψ/W−2Eψ. Indeed, the classes of extensions
259
+ 0 → H2m+1(Y )(−m) → E(ψ)/W−2E(ψ) → Q(0) → 0 both equal Deligne cohomol-
260
+ ogy class clD(A) (a.k.a. Griffiths’ Abel-Jacobi periods) of A; by duality, the classes
261
+ of (the duals to) extensions 0 → Q(1) → W−1E(ψ) → H2m+1(Y )(−m) → 0 both
262
+ equal to clD(B) (see loc.cit.).
263
+
264
+ HEIGHT PAIRING AND NEARBY CYCLES
265
+ 5
266
+ Now suppose we have a Q-Hodge structure H of weight −1 and two classes
267
+ a ∈ Ext1(Q(0), H), b ∈ Ext1(H, Q(1)).
268
+ Consider the set EH
269
+ a,b = EH(H)a,b of
270
+ all Hodge structures E with weights in [−2, 0] and equipped with identifications
271
+ grW
272
+ 0 E = Q(0), grW
273
+ −1E = H, grW
274
+ −2E = Q(1) such that the extensions E/W−2E and
275
+ W−1E have classes a and b. The group Ext1(Q(0), Q(1)) = C× ⊗ Q acts on EH
276
+ a,b by
277
+ the Baer sum action, and EH
278
+ a,b is a C× ⊗ Q-torsor. Notice that for q ∈ C× one has
279
+ ⟨q · E⟩ = log |q| + ⟨E⟩. Applying this format to H = H2m+1(Y )(−m), a = clD(A),
280
+ b = clD(B) and EA′,B′, Eψ
281
+ A,B ∈ EH
282
+ a,b we get EA′,B′ − Eψ
283
+ A,B ∈ C× ⊗ Q. Now the
284
+ theorem in 1.8 follows immediately from the next result (notice that the Hodge
285
+ periods and the height pairing play no role here):
286
+ Theorem. One has EA′,B′ − Eψ
287
+ A,B ∈ k× ⊗ Q ⊂ C× ⊗ Q.
288
+ The theorem would be an immediate corollary of the motivic formalism if all
289
+ the above constructions could be spelled in motivic world: Indeed, we would have
290
+ then a motivic version EM of EH which is an Ext1
291
+ M(Q(0), Q(1)) = k× ⊗ Q-torsor
292
+ equipped with the Hodge realization embedding EM ֒→ EH; our EA′,B′, Eψ
293
+ A,B
294
+ would come from elements of EM, and so their difference lies in k× ⊗ Q. The only
295
+ problem is that in the present day formalism of motives, due to Voevodsky, Ayoub,
296
+ and Cisinski-D´eglise, the t-structure is not available, so we do not have the motivic
297
+ version of separate homology groups like Hi(Y ). The actual proof is an exercise in
298
+ spelling out the constructions in a way that makes the t-structure redundant.
299
+ I am very grateful to Spencer Bloch for explaining me his conjecture and stimu-
300
+ lating discussions (pity Spencer refused to coauthor the article), to Volodya Drinfeld
301
+ for valuable comments and discussions, and to Luc Illusie for calling my attention
302
+ to the construction of [I] which helped to clearify and simplify the argument.
303
+ §2. The height pairing and the construction of EM
304
+ a,b ∈ EM
305
+ a,b ⊂ EH
306
+ a,b
307
+ This section is a variation on the theme of [Bl2] and [G].
308
+ 2.1. Let C be a stable dg category. It yields two other dg categories C(1) and C(2)
309
+ constructed as follows:
310
+ An object of C(1) is a closed morphism α : M → N of degree 0 in C. One has
311
+ Hom((M, N, α), (M ′, N ′, α′))i = Hom(M, M ′)i ×Hom(N, N ′)i ×Hom(M, N ′)i+1 ⊂
312
+ Hom(Cone(α), Cone(α′))i, and the differential is defined so that the latter embed-
313
+ ding is a morphism of complexes; the composition of morphisms is defined in a sim-
314
+ ilar way. There are three dg functors C(1) → C which send (M, N, α) to M, N, and
315
+ Cone(α) respectively. We can view C(1) as the category of distinguished triangles,
316
+ and the rotation yields an autoequivalence ρ : C(1) → C(1) which sends α : M → N
317
+ to ρ(α) : N → Cone(α); the inverse autoequivalence is ρ−1(α) : Cone(��)[−1] → M.
318
+ An object of C(2) is a datum (P, M, Q, α, β, κ) where P, M, Q are objects of C,
319
+ α ∈ Hom(P, M)1, β ∈ Hom(M, Q)1 are closed maps, and κ ∈ Hom(P, Q)1 is such
320
+ that d(κ) = βα; we sometimes abbreviate it to (α, β, κ). One can assign to such a
321
+ datum an object E = E(α, β, κ) ∈ C which equals P ⊕ M ⊕ Q with α, β, and −κ
322
+ added as the components to the differential.1 There is a filtration Q ⊂ Cone(β :
323
+ M[−1], Q) ⊂ E, and morphisms in C(2) are the same as morphisms between the
324
+ 1Thus E = Cone((α, κ) : P [−1] → Cone(β : M[−1] → Q)) = Cone((κ, β) : Cone(α : P [−2] →
325
+ M[−1]) → Q).
326
+
327
+ 6
328
+ A. BEILINSON
329
+ corresponding objects E that preserve this filtration.
330
+ We have two dg functors
331
+ C(2) → C(1) which send to (α, β, κ) to α : P[−1] → M and β : M → Q[1], and
332
+ six dg functors C(2) → C which send (α, β, κ) to P, M, Q, Cone(α : P[−1] → M),
333
+ Cone(β : M[−1] → Q), and E(α, β, κ) respectively.
334
+ The dg category C(3) carries a natural involution σ which sends (P, M, Q, α, β, κ)
335
+ to the object (Q[−1], E(α, β, κ), P[1], ασ, βσ, 0) where ασ and βσ are the evident
336
+ embedding and projection.
337
+ Remark. One can view an object (α, β, κ) ∈ C(2) as an object of C equipped with
338
+ a 3-step filtration in two different ways. Namely, this could be E(α, β, κ) equipped
339
+ with an evident filtration with successive quotients Q, M, and P. Or this could be
340
+ M equipped with a filtration whose successive quotients are P[−1], E(α, β, κ), and
341
+ Q[1]. The involution σ exchanges the two perspectives.
342
+ 2.2. For C as above we denote by C× the ∞-groupoid of its homotopy equivalences,
343
+ by C×τ the corresponding 1-truncaded plain groupoid, and by HC the homotopy
344
+ category of C. For S, T ∈ C set Exti(S, T ) := HiHom(S, T ) = HomHC(S, T [i]).
345
+ Denote by Ext(S, T ) the plain Picard groupoid of extensions that corresponds to
346
+ the two-term complex τ [0,1]Hom(S, T ).
347
+ For M, N ∈ C let C(1)×
348
+ M,N be the ∞-groupoid of collections (α′ : M ′ → N ′, ιM, ιN)
349
+ where (α′ : M ′ → N ′) ∈ C(1) and ιM : M → M ′, ιN : N → N ′ are homotopy equiv-
350
+ alences. It is equivalent to the Picard ∞-groupoid that corresponds to the complex
351
+ τ ≤0Hom(M, N). The 1-truncated plain Picard groupoid C(1)×τ
352
+ M,N
353
+ corresponds to the
354
+ two-term complex τ [−1,0]Hom(M, N).
355
+ Similarly, for three objects P, M, Q ∈ C we have the ∞-groupoid C(2)×
356
+ P,M,Q whose
357
+ objects are data (P ′, M ′, Q′, α′, β′, κ′, ιP , ιM, ιQ) where (P ′, M ′, Q′, α′, β′, κ′) ∈ C(2)
358
+ and ιP : P → P ′, ιM : M → M ′, ιQ : Q → Q′ are homotopy equivalences.
359
+ The 1-truncated plain groupoid C(2)×τ
360
+ P,M,Q contains a normal subgroup Ext0(P, Q) =
361
+ HomHC(P, Q).
362
+ Let by E = E(M) = E(P, M, Q) be the quotient groupoid.
363
+ It
364
+ is equivalent to the groupoid of triples (α, β, κ) where α ∈ Hom(P, M)1, β ∈
365
+ Hom(M, Q)1 are closed maps, and κ ∈ Hom(P, Q)1/d(Hom(P, Q)0) is such that
366
+ d(κ) = βα; a morphism (α, β, κ) → (α′, β′, κ′) in E is a pair (φ, ψ) where φ ∈
367
+ Hom(P, M)0/d(Hom(P, M)−1), ψ ∈ Hom(M, Q)0/d(Hom(M, Q)−1) are such that
368
+ α′ − α = d(φ), β′ − β = d(ψ), κ′ − κ = βφ + ψα + ψd(φ).
369
+ The projection C(2)
370
+ P,M,Q → C(1)
371
+ P [−1],M × C(1)
372
+ M,Q[1] yields a map of plain groupoids
373
+ E(P, M, Q) → C(1)×τ
374
+ P [−1],M × C(1)×τ
375
+ M,Q[1] = Ext(P, M) × Ext(M, Q), (α, β, κ) �→ (α, β).
376
+ The group Ext1(P, Q) acts on E by translations of κ, and non-empty fibers Eα,β
377
+ are Ext1(P, Q)-torsors.
378
+ Remark. E(P, M, Q) is naturally functorial with respect to P and Q: every pair
379
+ of closed morphisms µ : P1 → P and ν : Q → Q1 yields a map E(P, M, Q) →
380
+ E(P1, M, Q1), (α, β, κ) �→ (αµ, νβ, νκµ); is compatible with the Ext1(P, Q)-action
381
+ via the map (µ∗, ν∗) : Ext1(P, Q) → Ext1(P1, Q1).
382
+ Suppose Ext2(P, Q) = 0.
383
+ Then Eα,β are non-empty, and the addition maps
384
+ Eα1,β × Eα2,β → Eα1+α2,β, Eα,β1 × Eα,β2 → Eα,β1+β2 define on E the structure of
385
+ an Ext1(P, Q)-biextension of (Ext(P, M), Ext(M, Q)).
386
+ 2.3. In our first example C is the dg category whose homotopy category is the
387
+ bounded derived category DH of the category H of Q-Hodge structures, and
388
+
389
+ HEIGHT PAIRING AND NEARBY CYCLES
390
+ 7
391
+ P = Q(0), Q = Q(1). We denote the corresponding E by EH = EH(M). Then
392
+ Ext̸=1
393
+ DH(P, Q) = 0 and Ext1
394
+ DH(P, Q) = C× ⊗ Q, so EH is a C× ⊗ Q-biextension of
395
+ (Ext(Q(0), M), Ext(M, Q(1))).
396
+ Let Ext1
397
+ 0(Q(0), M) ⊂ Ext1(Q(0), M), Ext1
398
+ 0(M, Q(1)) ⊂ Ext1(M, Q(1)) be the
399
+ subgroups of those elements a, b that the maps H0a : Q(0) → H1M, H−1b :
400
+ H−1M → Q(1) vanish. Let Ext0(Q(0), M) ⊂ Ext(Q(0), M), etc., be the Picard
401
+ groupoids of such extensions.
402
+ Lemma. Suppose that Hom(Q(0), H0M) = Hom(H0M, Q(1)) = 0.
403
+ (i) The restriction of EH to (Ext0(Q(0), M), Ext0(M, Q(1))) descends to the C×⊗Q-
404
+ biextension of (Ext1
405
+ 0(Q(0), M), Ext1
406
+ 0(M, Q(1))).
407
+ (ii) EH is naturally functorial with respect to M: if ϕ : M → M ′ is a morphism,
408
+ and we have a′ ∈ Ext1
409
+ 0(Q(0), M ′), b′ ∈ Ext1
410
+ 0(M ′, Q(1)) with ϕ∗(a) = a′, ϕ∗(b′) = b
411
+ then there is a canonical identification EH(M)a,b = EH(M ′)a′,b′.
412
+ (iii) The isomorphisms Ext1
413
+ 0(Q(0), M)
414
+
415
+ → Ext1(Q(0), H0M), Ext1
416
+ 0(M, Q(1))
417
+
418
+ → Ext1
419
+ (H0M, Q(1)) which assign to an extension its zero cohomology, lifts naturally to an
420
+ isomorphism of biextensions H0 : EH(M)
421
+
422
+ → EH(H0M). One has EH0 = H0E.
423
+ Proof. Let us prove (i); the rest is clear. We need to check that for every closed α ∈
424
+ Hom1
425
+ 0(Q(0), M), β ∈ Hom1
426
+ 0(M, Q(1)) the action of Aut(α)×Aut(β) = Hom(Q(0), M)
427
+ ×Hom(M, Q(1)) on EH
428
+ α,β is trivial.
429
+ Since H has homological dimension 1 our M is isomorphic to the direct sum of its
430
+ homologies and so Aut(α) = Ext1(Q(0), H−1M), Aut(β) = Ext1(H1(M), Q(1)) by
431
+ the condition on M. The action of (e, h) ∈ Ext1(Q(0), H−1M)×Ext1(H1(M), Q(1))
432
+ on EH
433
+ α,β is the translation by H−1(β)e + hH0(α) which is 0 since α, β ∈ Ext1
434
+ 0.
435
+
436
+ 2.4. Lemma. Suppose that H0M is pure of weight −1 (which implies the condition
437
+ of the lemma in 2.3). Then the function EH(M) → R, (α, β, κ) �→ ⟨E(α, β, κ)⟩ :=
438
+ ⟨H0E(α, β, κ)⟩, see 1.1, is a natural trivialization of the R-biextension log |EH(M)|.
439
+ Proof. Everything said in 2.3 works for the category HR of R-Hodge structures. The
440
+ extension of scalars functor H → HR, ? �→? ⊗ R, yields a morphism of our biex-
441
+ tensions EH(M) → EHR(M ⊗ R). The map Ext1(Q(0), Q(1)) → Ext1(R(0), R(1))
442
+ equals log | | after the standard identifications of the Ext groups with, respectively,
443
+ C× ⊗ Q and R.
444
+ Since Ext1(R(0), H0M ⊗ R) = Ext1(H0M ⊗ R, R(1)) = 0 by
445
+ the condition on M, one has EHR(M ⊗ R) = EHR(H0M ⊗ R) = R.
446
+ The map
447
+ EH(M) → EHR(M ⊗ R) = R is ⟨ ⟩ of 1.1.
448
+
449
+ 2.5. Let k ⊂ C be a subfield. Denote by DM(k) the dg category of geometric
450
+ Voevodsky Q-motives over k. We have the Hodge realization dg functor DM(k) →
451
+ DH, M �→ M H.
452
+ Consider the story of 2.2 for C = DM(k) with P = Q(0),
453
+ Q = Q(1). As before one has Ext̸=1
454
+ DM(k)(Q(0), Q(1)) = 0, and there is a canonical
455
+ identification Ext1(Q(0), Q(1)) = k× ⊗ Q such that the Hodge realization map
456
+ between the Ext1’s is the embedding k× ⊗ Q ֒→ C× ⊗ Q. So for any M ∈ DM(k)
457
+ we get a k× ⊗ Q-biextension of (Ext1(Q(0), M), Ext1(M, Q(1))) together with the
458
+ Hodge realization morphism EM(M) → EH(M) := EH(M H) of the biextensions.
459
+ Remark. Since the homomorphism k× ⊗ Q ֒→ C× ⊗ Q is injective, the maps of
460
+ torsors EM(M)α,β → EH(M)α,β := EH(M)αH,βH are injective too.
461
+ We define Ext1
462
+ 0(Q(0), M) ⊂ Ext1
463
+ 0(Q(0), M) and Ext1
464
+ 0(M, Q(1)) ⊂ Ext1(M, Q(1))
465
+ as preimages of the Ext1
466
+ 0 subgroups of the Hodge setting by the Hodge realiza-
467
+
468
+ 8
469
+ A. BEILINSON
470
+ tion maps.
471
+ Assume that H0M H is pure of weight −1.
472
+ Then (i) and (ii) of
473
+ the lemma in 2.3 remain true in the DM(k) setting (with C× replaced by k×):
474
+ this follows from loc.cit. by Remark above. Thus we have a k× ⊗ Q-biextension
475
+ EM(M) of (Ext1
476
+ 0(Q(0), M), Ext1
477
+ 0(M, Q(1))) together with a map of biextensions
478
+ EM(M) → EH(M), so the lemma in 2.4 provides a natural trivialization of the
479
+ R-biextension log |EM(M)|. The image of EM
480
+ a,b in R/Q log |k×| depends only on
481
+ a, b ∈ Ext1
482
+ 0(M, Q(1)) × Ext1
483
+ 0(Q(0), M), and we denote it by ⟨a, b⟩M. It is clearly
484
+ biadditive with respect to a, b.2 We have defined a canonical height pairing
485
+ (2.5.1)
486
+ ⟨ ⟩M : Ext1
487
+ 0(Q(0), M) × Ext1
488
+ 0(M, Q(1)) → R/Q log |k×|.
489
+ 2.6. We return to the situation of 1.3 and set M := M(Yk)(−m)[−1 − 2m] where
490
+ M(Yk) is the motive of Yk. One has Ext1(Q(0), M) = CHm(Yk), Ext1(M, Q(1)) =
491
+ CHm′(Yk) by the Poincar´e duality, and Ext1
492
+ 0 are the subgroups CH(Yk)0 of cycles
493
+ homologically equivalent to zero. Therefore we get a k× ⊗ Q-biextension EM of
494
+ (CHm(Yk)0, CHm′(Yk)0), the map of biextensions EM → EH, the trivialization of
495
+ log |EM|, and the height pairing ⟨ , ⟩M : CHm(Yk)0 × CHm′(Yk)0 → R/Q log |k×|.
496
+ By (iii) of the lemma in 2.3 one has H0 : EH(M)
497
+
498
+ → EH(H2m+1(Y )(−m)). For
499
+ a ∈ CHm(Yk)0, b ∈ CHm′(Yk)0 pick, as in 1.3, cycles A, B that represent them
500
+ such that |A| ∩ |B| = ∅.3 Let us construct (a, b, κA,B) ∈ EM
501
+ a,b such that the Hodge
502
+ realization EH
503
+ A,B of EM
504
+ A,B := E(a, b, κA,B) (see 2.1) has zero cohomology H0EH
505
+ A,B
506
+ equal to the Hodge structure EA,B from 1.3. This would imply that for our M the
507
+ height pairing (2.5.1) equals (1.3.1).
508
+ The composition of the maps M(|A|)
509
+ α→ M(Yk)
510
+ β→ M(Yk, Yk ∖ |B|) is naturally
511
+ homotopic to 0: indeed, M(Yk, Yk ∖ |B|) := Cone(M(Yk ∖ |B|) → M(Yk)), and the
512
+ homotopy κ|A|,|B| is M(|A|) → M(Yk ∖|B|) ⊂ Cone. Thus we have (α, β, κ|A|,|B|) ∈
513
+ DM(2) (see 2.1). Notice that E(α, β, κ|A|,|B|) = M(Yk ∖ |B|, |A|).
514
+ One has Ext−2m(Q(m), M(|A|)) = Zm(|A|) := the group of m-cycles supported
515
+ on |A| (recall that dim |A| = m), and Ext2m+2(M(Yk, Yk ∖ |B|), Q(m + 1)) =
516
+ Zm′(|B|) by the Poincar´e duality.
517
+ Therefore we have (αA, Bβ, Bκ|A|,|B|A) =
518
+ (Q(m)[2m+1], M(Yk), Q(m)[2m+2], αA, Bβ, Bκ|A|,|B|A) ∈ DM(2). The promised
519
+ (a, b, κA,B) ∈ EM
520
+ a,b is (αA, Bβ, Bκ|A|,|B|A)(−m)[−1 − 2m]. The fact that H0EH
521
+ A,B
522
+ equals the Hodge structure EA,B from 1.3 follows from the construction.
523
+ §3. The unipotent nearby cycles in the Hodge setting
524
+ 3.1. A nearby cycles reminder. In this section we play with algebraic varieties over
525
+ C. For an algebraic variety X we denote by H(X) the abelian category of perverse
526
+ Hodge Q-sheaves of M. Saito on X, by DH(X) its bounded derived category. It sat-
527
+ isfies the usual Grothendieck six functors formalism. Below ∗ is the Verdier duality.
528
+ Every object of H(X), hence of DH(X), carries a canonical weight filtration.
529
+ For F ∈ DH(X) let Γ(X, F), Γc(X, F) ∈ DH be the complex of chains, resp.
530
+ chains with compact support, with coefficients in F equipped with the natural
531
+ Hodge structure, H·
532
+ (c)(X, F) := H·Γ(c)(X, F) ∈ H; set Γ(c)(X) := Γ(c)(X, Q(0)X),
533
+
534
+ (c)(X) := H·
535
+ (c)(X, Q(0)), and denote by ( , ) the Poincar´e duality pairing. Simi-
536
+ larly for a closed subvariety A ⊂ X we set ΓA(X) := ΓA(X, Q(0)) ∈ DH, etc.
537
+ 2Indeed, a morphism from a biextension by a trivial group to a trivialized biextension amounts
538
+ to a biadditive pairing.
539
+ 3Recall that |A|, |B| ⊂ Yk are supports of the cycles.
540
+
541
+ HEIGHT PAIRING AND NEARBY CYCLES
542
+ 9
543
+ Let g : X → A1 be a function on X; set X0 := g−1(0), and let v : X ∖ X0 ֒→ X,
544
+ iX0 : X0 ֒→ X be the open and closed embeddings. One has the unipotent nearby
545
+ cycles functor ψun
546
+ g
547
+ : DH(X ∖ X0) → DH(X0) that carries a natural logarithm
548
+ of monodromy morphism N = Ng = NF : ψun
549
+ g (F)(1) → ψun
550
+ g (F) where F ∈
551
+ D(X ∖ X0). It has ´etale local origin with respect to X0. For sheaves on X there
552
+ is a natural morphism of functors ι : i∗
553
+ X0 → ψun
554
+ g v∗.
555
+ There are basic canonical
556
+ identifications:
557
+ (i) Compatibility with Verdier duality: One has ψun
558
+ g (F∗) = (ψun
559
+ g F)∗(1)[2].
560
+ (ii) Compatibility with proper direct images: Suppose h : X → T is a proper map
561
+ and t is a function on T such that g = th; then one has ψun
562
+ t h∗F = h∗ψun
563
+ g F.
564
+ (iii) One has Cone(NF) = i∗
565
+ X0v∗F(1)[1].
566
+ (iv) For every n > 0 one has ψun
567
+ gnF
568
+
569
+ → ψun
570
+ g F.
571
+ These identifications are mutually compatible; (i) and (ii) are compatible with
572
+ the action of N, and (iv) identifies Ngn with nNg. Finally, one has
573
+ (v) ψun[−1] is t-exact for the perverse t-structure.
574
+ Examples. Suppose that X is smooth of dimension n and F = Q(0)X∖X0. Then
575
+ F∗ = F(n)[2n] hence ψun
576
+ g (F)∗ = (ψun
577
+ g F)(n − 1)[2n − 2].
578
+ (a) If g is smooth then ιQ(0)X : Q(0)X0
579
+
580
+ → ψun
581
+ g F, NF = 0.
582
+ (b) Suppose g is semi-stable and X0 has two irreducible components Y and Y ′. By
583
+ (a) one has natural morphisms jY ′∖Y !QY ′∖Y → ψun
584
+ g F → jY ∖Y ′∗QY ∖Y ′ compatible
585
+ with the N-action (we take it that on the left and right object N acts trivially).
586
+ They form an exact triangle; its Verdier dual is the same triangle with Y and Y ′
587
+ interchanged.
588
+ 3.2. We are in the setting of 1.4 and follow the notation there.
589
+ Let j : U := X0 ∖ {xα} ֒→ X0 ←֓ {xα} : ⊔ixα be the complementary open
590
+ and closed embeddings. Let I be the intersection cohomology sheaf j!∗Q(0)U =
591
+ τ ≤n−2j∗Q(0)U 4 on X0; set I+ := π∗Q(0)Y . One has natural self-duality isomor-
592
+ phisms I∗ = I(n − 1)[2n − 2], I+∗ = I+(n − 1)[2n − 2] (recall that Y is smooth of
593
+ dimension n − 1 and π is proper).
594
+ The decomposition theorem for π is easy and explicit:
595
+ Proposition. There is a natural orthogonal direct sum decomposition
596
+ (3.2.1)
597
+ I+ = I ⊕ ⊕αixα∗τ [2,2n−4]Γ(Pα)
598
+ compatible with the self-dualities.
599
+ Proof. One has a natural orthogonal direct sum decomposition
600
+ (3.2.2)
601
+ Γ(Zα) = Hn−2
602
+ prim(Zα)[2 − n] ⊕ τ ≤2n−4Γ(Pα)
603
+ defined as follows. Consider the embedding Zα ֒→ Pα. The pullback and Gysin
604
+ maps Γ(Pα) → Γ(Zα) → Γ(Pα)(1)[2] are mutually dual for the Poincar´e duality
605
+ pairings, and their composition in either direction equals to the multiplication by
606
+ c1(O(dα)).5 Thus the composition of τ ≤2n−4Γ(Pα) → Γ(Zα) → τ ≥0(Γ(Pα)(1)[2]) is
607
+ an isomorphism. This yields a direct sum decomposition Γ(Zα) =?⊕τ ≤2n−4Γ(Pα).
608
+ 4Below τ is the usual truncation, pτ is the perverse one.
609
+ 5Since O(dα) is the normal bundle to Zα in Pα.
610
+
611
+ 10
612
+ A. BEILINSON
613
+ Since multiplication by c1(O(dα)) preserves the direct sum decomposition, the only
614
+ nonzero cohomology of ? is Hn−2
615
+ prim(Zα) ⊂ Hn−2(Zα), q.e.d.
616
+ Consider the embeddings of smooth divisors iZα : Zα ֒→ Y . One has i!
617
+ ZαQ(0)Y =
618
+ Q(−1)[−2]Zα, i∗
619
+ ZαQ(0)Y = Q(0)Zα, and the composition of the adjunction maps
620
+ iZα∗i!
621
+ ZαQ(0)Y → Q(0)Y → iZα∗i∗
622
+ ZαQ(0)Y equals the multiplication by c1(O(−1))
623
+ map Q(−1)[−2]Zα → Q(0)Zα.6 Apply π∗; then i!
624
+ xαI+ = Γ(Zα)(−1)[−2], i∗
625
+ xαI+ =
626
+ Γ(Zα) by base change, and the composition of the adjunctions ixα∗i!
627
+ xαI+ → I+ →
628
+ ixα∗i∗
629
+ xαI+ is multiplication by c1(O(−1)) map ixα∗Γ(Zα)(−1)[−2] → ixα∗Γ(Zα).
630
+ Composing the maps τ ≤2n−6Γ(Pα) ֒→ Γ(Zα) and Γ(Zα) ։ τ [2,2n−4]Γ(Pα) that
631
+ come from decomposition (3.2.2) from the left and from the right with the latter ad-
632
+ junctions, we get the maps ixα∗(τ ≤2n−6Γ(Pα))(−1)[−2] → I+ → ixα∗τ [2,2n−4]Γ(Pα).
633
+ Their composition is an isomorphism, which yields a decomposition I+ = I? ⊕
634
+ ixα∗τ [2,2n−4]Γ(Pα). Since the adjunctions are mutually dual, the decomposition is
635
+ orthogonal.
636
+ By (3.2.2) one has i!
637
+ xαI? = Hn−2
638
+ prim(Zα)(−1)[−n] ⊕ Q(n − 1)[2 − 2n], i∗
639
+ xαI? =
640
+ Hn−2
641
+ prim(Zα)[2 − n] ⊕ Q(0).
642
+ Thus I?[n − 1] is a perverse sheaf which equals Q(0)[n − 1]U on U and has no
643
+ subquotients supported on {xα}, and so I? = I. We are done.
644
+
645
+ Remarks. (i) The adjunction map Q(0)X0 → π∗Q(0)Y = I+ takes value in I ⊂ I+
646
+ since Hom(Q(0)X0, ixα∗τ [2,2n−4]Γ(Pα)) = 0.
647
+ (ii) Set B := ⊕ixα∗Hn−2
648
+ prim(Zα)[1 − n]. By the formula for i∗
649
+ xαI at the end of the
650
+ previous paragraph, one has an exact triangle Q(0)X0 → I → B[1].
651
+ 3.3. As in 1.5, t is a local coordinate at 0 ∈ S; shrinking S we can assume that
652
+ t is defined and invertible on S ∖ {0}, so X0 = (tf)−1(0). Consider the functor
653
+ ψun
654
+ tf : DH(X ∖ X0) → DH(X0) (see 3.1). Set R := ψun
655
+ tf Q(0)X∖X0. By 3.1(i) one
656
+ has a canonical self-duality identification R∗ = R(n − 1)[2n − 2] and the mutually
657
+ dual maps Q(0)X0
658
+ ι→ R
659
+ ι∗
660
+ → Q(0)∗
661
+ X0(1 − n)[2 − 2n] which are isomorphisms over U.
662
+ The next result is due to Illusie [Il]; we will need it in 4.5. The reader can skip
663
+ it at the moment and jump directly to section 3.4.
664
+ Proposition. For every critical point xα one has canonical isomorphisms
665
+ (3.3.1)
666
+ i!
667
+ xαR = Γc(Pα ∖ Zα),
668
+ i∗
669
+ xαR = Γ(Pα ∖ Zα)
670
+ interchanged by the duality. The N-action on i!
671
+ xαR, i∗
672
+ xαR is trivial.
673
+ Proof. (a) The claim is local at xα, so for the proof we remove from X the rest
674
+ of critical points, and still call it X by the abuse of notation. Let S♭ → S be the
675
+ covering of degree dα obtained by adding t♭ = t1/dα to the sheaf of functions; its
676
+ Galois group is µdα. Set X♭ := X×SS♭ and let f ♭ : X♭ → S♭ be the projection. Our
677
+ X♭ is a hypersurface {(x, t♭) : (tf)(x) − t♭dα = 0} in X × A1; its only singular point
678
+ is (xα, 0). The projectivized tangent cone Qα of X♭ at (xα, 0) is a hypersurface in
679
+ P +
680
+ α := P(T(xα,0)X × A1). The Galois group µdα acts on X♭ hence on Qα.
681
+ (b) Let us check that Qα is a µdα-covering of Pα completely ramified along Zα
682
+ and ´etale over its complement, and Qα is smooth. To see this, consider the leading
683
+ term [tf]dα(x) (of the Taylor expansion) of tf at xα; then the leading term of
684
+ 6Since O(−1) is the normal bundle to Zα in Y .
685
+
686
+ HEIGHT PAIRING AND NEARBY CYCLES
687
+ 11
688
+ (tf)(x) − t♭dα at (xα, 0) is [tf]dα(x) − t♭dα. The zeros of [tf]dα is Zα ⊂ Pα, of
689
+ [tf]dα(x) − t♭dα is Qα ⊂ P +
690
+ α , and so the projection Qα → Pα (x, t♭) �→ x, is as
691
+ claimed. The smoothness of Qα follows from that of Zα.
692
+ (c) Let π+ : X+ → X♭ be the blowup of X♭ at (xα, 0). By (b) X+ is smooth
693
+ and the map f + := f ♭π+ : X+ → S♭ has semistable reduction at 0 ∈ S♭. The
694
+ fiber X+
695
+ 0 has two irreducible components: one equals Y and the other Qα, and
696
+ their intersection equals Zα. The action of µdα on X♭ yields one on X+. The
697
+ µdα-action on X+
698
+ 0
699
+ fixes Y and acts on Qα as described in (b).
700
+ The projection
701
+ π+
702
+ 0 : X+
703
+ 0 → X♭
704
+ 0 = X0 contracts Qα to xα.
705
+ Set R+ := ψun
706
+ tf +Q(0)X+∖X+
707
+ 0 , R♭ := ψun
708
+ tf ♭Q(0)X♭∖X♭
709
+ 0. These are sheaves on X+
710
+ 0
711
+ and X♭
712
+ 0 = X0 respectively that are naturally µdα-equivariant.
713
+ By 3.1(ii) (with
714
+ h = π+) one has a natural identification π+
715
+ 0∗R+ = R♭ compatible with the µdα-
716
+ actions. Since the projection p : X♭ → X is a µdα-torsor over X ∖ X0 one has
717
+ Q(0)X∖X0 = (p∗Q(0)X♭∖X♭
718
+ 0)µdα , and so, by 3.1(ii) with h = p, one has R = R♭µdα .
719
+ Therefore R = (π+
720
+ 0∗R+)µdα .
721
+ (d) By 3.1(iv) with g = t♭f +, n = dα, one has ψun
722
+ tf + = ψun
723
+ t♭f +. Our t♭f + is semi-
724
+ stable, so we have the exact triangle jY ∖Zα!QY ∖Zα → R+ → jQα∖Zα∗QQα∖Zα
725
+ as in Example (b) in 3.1. Applying π+
726
+ 0∗ we get an exact triangle j!QU → R♭ →
727
+ ixα∗Γ(Qα∖Zα). Passing to µdα-invariants we get, by (b), an exact triangle j!QU →
728
+ R → ixα∗Γ(Pα ∖ Zα); here we use the identification Γ(Qα ∖ Zα)µdα
729
+
730
+ → Γ(Pα ∖ Zα)
731
+ defined as the composition Γ(Qα ∖ Zα)µdα ⊂ Γ(Qα ∖ Zα)
732
+ tr
733
+ → Γ(Pα ∖ Zα). Thus
734
+ we get the isomorphism i∗
735
+ xαR
736
+
737
+ → Γ(Pα ∖ Zα) in (3.3.1). The second isomorphism
738
+ there comes in the dual manner from the dual exact triangle jQα∖Zα!QQα∖Zα →
739
+ R+ → jY ∖Zα∗QY ∖Zα. Since π+
740
+ 0∗ commutes with duality, the two isomorphisms are
741
+ mutually dual, and we are done.
742
+
743
+ Let αR be the composition B
744
+ ∂→ Q(0)X0
745
+ ι→ R where ��� is the boundary map of the
746
+ triangle from Remark (ii) in 3.2, so I = Cone(∂). Let us compute the map i!
747
+ xα(αR).
748
+ Consider the standard triangle Hn−2
749
+ prim(Zα)[1−n]
750
+ δ→ Γc(Pα ∖Zα)
751
+ tr
752
+ → Q(1−n)[2−2n]
753
+ that comes from (3.2.2).
754
+ Lemma. −i!
755
+ xα(αR) equals the composition δR of the maps Hn−2
756
+ prim(Zα)[1 − n]
757
+ δ→
758
+ Γc(Pα ∖ Zα)
759
+ (3.3.1)
760
+ =
761
+ i!
762
+ xαR.
763
+ Proof. Consider the exact triangle
764
+ (3.3.2)
765
+ jQα∖Zα!Q(0)Qα∖Zα ⊕ jY ∖Zα!Q(0)Y ∖Zα → Q(0)X+
766
+ 0 → Q(0)Zα.
767
+ Let (δQ, δY ) : Q(0)Zα[−1] → jQα∖Zα!Q(0)Qα∖Zα ⊕ jY ∖Zα!Q(0)Y ∖Zα be the bound-
768
+ ary map. Its composition with the map to Q(0)X+
769
+ 0 , and hence with the further
770
+ composition with Q(0)X+
771
+ 0
772
+ ι→ R+, is 0.
773
+ Therefore the sum of the compositions
774
+ Q(0)Zα[−1]
775
+ δQ
776
+ −→ jQα∖Zα!
777
+ ι→ R+ and Q(0)Zα[−1]
778
+ δY
779
+ −→ jY ∖Zα!
780
+ ι→ R+ is 0.
781
+ Ap-
782
+ ply i!
783
+ xαπ+
784
+ ∗ and consider the restriction of our compositions to Hn−2
785
+ prim(Zα)[1 − n] ⊂
786
+ Γ(Zα)[−1]. For the first one it is δR, for the second one it is i!
787
+ xα(αR), and we are
788
+ done.
789
+
790
+
791
+ 12
792
+ A. BEILINSON
793
+ 3.4. Set P := R[n − 1] = ψun
794
+ tf Q(0)X∖X0[n − 1]; this is a perverse sheaf on X0; one
795
+ has a canonical self-duality identification P∗ = P(n − 1). Consider the perverse
796
+ sheaves PN := Ker(N : P → P(−1)), PN := Coker(N : P(1) → P).
797
+ Lemma. (i) Q(0)X0[n − 1] is a perverse sheaf of weights n − 1 and n − 2 with
798
+ grW
799
+ n−1 = I[n − 1], grW
800
+ n−2 = ⊕α ixα∗Hn−2
801
+ prim(Zα).
802
+ (ii) One has PN = Q(0)X0[n − 1], PN = (Q(0)X0[n − 1])∗(1 − n).
803
+ (iii) P has weights in [n − 2, n]. One has Wn−1P = Q(0)X0[n − 1], P/Wn−2P =
804
+ (Q(0)X0[n − 1])∗(1 − n), grW
805
+ n−2P = ⊕α ixα∗Hn−2
806
+ prim(Zα), grW
807
+ n−1P = I[n − 1], grW
808
+ n P =
809
+ (grW
810
+ n−2P)∗(1 − n).
811
+ Proof. (i) The exact triangle from Remark (ii) in 3.2 amounts to an exact triangle
812
+ ⊕ixα∗Hn−2
813
+ prim(Zα) → Q(0)X0[n − 1] → I[n − 1], and we are done since its left and
814
+ right terms are pure perverse sheaves of weights n − 2 and n − 1 respectively.
815
+ (ii) For any sheaf A on X one has a canonical exact triangle i∗
816
+ X0A → i∗
817
+ X0v∗v∗A →
818
+ i!
819
+ X0A[1]: Indeed, the map v!v∗A → v∗v∗A factors as composition v!v∗A → A →
820
+ v∗v∗A, and so one has an exact triangle Cone(v!v∗A → A) → Cone(v!v∗A →
821
+ v∗v∗A) → Cone(A → v∗v∗A) which is supported on X0.
822
+ The promised exact
823
+ triangle is its restriction to X0.
824
+ Now take for A the perverse sheaf Q(0)X[n]. The first term of the triangle is
825
+ Q(0)X0[n] which is perverse sheaf shifted by 1, its third term is (Q(0)X0[n−1])∗(−n)
826
+ which is a perverse sheaf. Therefore they equal, respectively, pH−1 and pH0 of
827
+ i∗
828
+ X0v∗v∗Q(0)X[2n], i.e., of Cone(N : P → P(−1)) by 3.1(iii), and we are done.
829
+ (iii) Since N is nilpotent, the weights of P are bounded from below by the
830
+ minimum of weights of PN, which is n − 2 by (ii) and (i). By self-duality of P they
831
+ are bounded then from above by n, and we have the first assertion. It implies that
832
+ Wn−2P ⊂ PN. The rest follows directly from (i), (ii), and self-duality of P.
833
+
834
+ 3.5. Proof of the proposition in 1.5. We use the notation in loc.cit. Injectivity of
835
+ sp : (ψun
836
+ t H)N → Hn−1(X0) follows from the local invariant cycles theorem. Let us
837
+ check the surjectivity. By 3.1(ii) applied to h = f (recall that f is proper) and 3.1(v)
838
+ applied to ψun
839
+ t , one has ψun
840
+ t H = H0(X0, P)(n−1). By 3.4 we have exact sequence of
841
+ perverse sheaves 0 → ⊕α ixα∗Hn−2
842
+ prim(Zα)(n−1) → P(n−1) → (Q(0)X0[n−1])∗ → 0.
843
+ Its left term has finite support, and so has no cohomology in degrees ̸= 0. Therefore
844
+ the map H0(X0, P)(n − 1) → H0(X0, (Q(0)X0[n − 1])∗) = Hn−1(X0) is surjective.
845
+ This map equals sp, and we are done.
846
+
847
+ §4. The motivic setting and the construction of EψM
848
+ a,b
849
+ ∈ EM
850
+ a,b
851
+ 4.1. We are in the setting of 1.8 so k ⊂ C is a subfield and we play with varieties
852
+ over k. Changing slightly the notation of 1.3 and 1.8, for a variety Z = Zk we set
853
+ ZC := Z ⊗k C. The notation of §3 is preserved except that we equip from now on
854
+ all Hodge sheaves and Hodge structures met previously with extra upper index H.
855
+ We play with motives (a.k.a. motivic sheaves) over varieties, see [A1] and [CD].
856
+ For a variety Z the category of constructible Q-motives over Z is denoted by
857
+ DM(Z).
858
+ We use Grothendieck’s six functors formalism for DM as developed
859
+ in [CD]. Recall that DM(Spec k) = DM(k) is the category of Voevodsky’s geo-
860
+ metric Q-motives over k.
861
+ For a variety Z one has M(Z) = πZ!π!
862
+ ZQ(0) where
863
+ πZ : Z → Spec k is the structure map. For a motivic sheaf F on Z set Γ(Z, F) :=
864
+ πZ∗F, Γc(Z, F) := πZ!F ∈ DM(k); we write Γ(c)(Z) := Γ(c)(Z, Q(0)Z). There is
865
+
866
+ HEIGHT PAIRING AND NEARBY CYCLES
867
+ 13
868
+ a Hodge realization functor DM(Z) → DH(ZC), F �→ FH, compatible with the
869
+ six functors and the Verdier duality ∗. For a smooth Z of dimension d one has
870
+ π!
871
+ ZQ(0) = Q(d)Z[2d], and so M(Z) = Γc(Z)(d)[2d].
872
+ The formalism of unipoteny nearby cycles in the setting of motivic sheaves was
873
+ developed in §§3.4, 3.6 of [A2]. The motivic version of everything said in 3.1 holds
874
+ except property (v) (for the t-structure is not available). The Hodge realization
875
+ functor commutes with the nearby cycles functors.
876
+ 4.2. Notation: Notice that Hom(Q(i)[2i], Q(j)[2j]) is 0 if i ̸= j and Q for i = j,7
877
+ and so every object M ∈ M(k) which is isomorphic to a direct sum of motives
878
+ Q(i)[2i], i ∈ Z, can be written in a unique manner as ⊕i Vi(i)[2i] where Vi is a
879
+ vector space (then Vi = Hom(Q(i)[2i], M)). Set τ ≤2aM := ⊕i≥−a Vi(i)[2i], etc.
880
+ We are in the situation of 3.2 in the setting of k-varieties. As in loc.cit., I+ :=
881
+ π∗Q(0)Y ∈ DM(X0) (so I+H is the corresponding Hodge sheaf from loc.cit.) Since
882
+ Y is smooth and π is proper one has a natural self-duality I+∗ = I+(n−1)[2n−2].
883
+ The t-structure in DM is not available, so we define the motivic intersection
884
+ cohomology sheaf I using a motivic version of decomposition (3.2.1):
885
+ Proposition. There is a natural orthogonal direct sum decomposition in DM(X0)
886
+ (4.2.1)
887
+ I+ = I ⊕ ⊕αixα∗τ [2,2n−4]Γ(Pα)
888
+ whose Hodge realization is (3.2.1)
889
+ Proof. It repeats the proof in 3.2 (minus its last paragraph). Namely, we first define
890
+ a natural orthogonal decomposition
891
+ (4.2.2)
892
+ Γ(Zα) = Hn−2
893
+ prim(Zα)[2 − n] ⊕ τ ≤2n−4Γ(Pα)
894
+ in DM(xα) = DM(kxα) whose Hodge realization is (3.2.2).8 The construction in
895
+ loc.cit. uses only basic six functors functoriality, so we can repeat it literally in the
896
+ motivic setting. Then we proceed to define (4.2.1) as in loc.cit.
897
+
898
+ Set B := ⊕α ixα∗Hn−2
899
+ prim(Zα)[1 − n] ∈ DM(X0). The self-dualities of Γ(Zα) and
900
+ of I+, and the above orthogonal decompositions yield natural self-dualities
901
+ (4.2.3)
902
+ B∗ ∼
903
+ → B(n − 2)[2n − 2],
904
+ I∗ ∼
905
+ → I(n − 1)[2n − 2].
906
+ 4.3. Lemma. (i) The adjunction χ : Q(0)X0 → π∗Q(0)Y = I+ takes values in
907
+ I ⊂ I+.
908
+ (ii) One has Cone(χ : Q(0)X0 → I) = B[1].
909
+ Proof. (i) Follows since Hom(Q(0)X0, ixα∗τ [2,2n−4]Γ(Pα)) = Hom(Q(0), τ [2,2n−4]Γ(Pα))
910
+ = 0.
911
+ (ii) Since χ|U = idQ(0)U the cone Cone(χ) is supported on {xα}. Now i∗
912
+ xαCone(χ) =
913
+ 7This follows since M(Pn) = ⊕i∈[0,n]Q(i)[2i] and End(M(Pn)) = CHn(Pn × Pn) = Q[0,n].
914
+ 8So Hn−2
915
+ prim(Zα) is a notation for a motive whose Hodge realization is the primitive cohomology
916
+ of Zα; its definition does not involve any cohomology. To construct it explicitly, choose a k-point
917
+ z in Pα ∖ Zα. Let πz : Zα → Pn−2 be the corresponding projection; this is a finite map of degree
918
+ dα. Then Hn−2
919
+ prim(Zα) is the kernel of the projector d−1
920
+ α πt
921
+ zπz acting on M(Zα)(2 − n)[4 − 2n].
922
+
923
+ 14
924
+ A. BEILINSON
925
+ Cone(i∗
926
+ xα(χ)) equals Hn−2
927
+ prim(Zα)[2 − n] by (4.2.2) and the construction of I, q.e.d.
928
+
929
+ Remark. Since Exti(Q(0)X0, Q(0)∗
930
+ X0(1−n)[2−2n]) = Exti(Q(0), M(X0)(1−n)[2−
931
+ 2n]) = CHn−1(X0, −i) we see that Ext0 = Zn−1(X0) and Ext̸=0 = 0, i.e., one has
932
+ Hom(Q(0)X0, Q(0)∗
933
+ X0(1 − n)[2 − 2n]) = Zn−1(X0) = Zn−1(U).
934
+ Example. One has χ∗χ = ǫ where ǫ : Q(0)X0 → Q(0)∗
935
+ X0(1 − n)[2 − 2n] is the map
936
+ that corresponds to the sum of irreducible components cycle (it is enough to check
937
+ the assertion on U where it is obvious).
938
+ 4.4. We are in the situation of 3.3 in the setting of k-varieties. Consider the functor
939
+ ψun
940
+ tf : DM(X ∖ X0) → DM(X0). There is a canonical morphism ι : i∗
941
+ X0 → ψun
942
+ tf v∗
943
+ of functors on DM(X) and its Verdier dual ι∗ : ψun
944
+ tf v∗ → i!
945
+ X0. Therefore we have
946
+ a motivic sheaf R := ψun
947
+ tf Q(0)X∖X0 equipped with a natural self-duality R∗
948
+
949
+
950
+ R(n − 1)[2n − 2] and mutually dual maps Q(0)X0
951
+ ι→ R
952
+ ι∗
953
+ → Q(0)∗
954
+ X0(1 − n)[2 − 2n]
955
+ that are isomorphisms over U.
956
+ Let ∂ : B → Q(0)X0 be the boundary map of the triangle from 4.3(ii).
957
+ Set
958
+ αR := ι∂ : B → R, and let βR be α∗
959
+ R combined with the self-duality identifications
960
+ for R and B, so we have
961
+ (4.4.1)
962
+ B
963
+ αR
964
+ −→ R
965
+ βR
966
+ −→ B(−1).
967
+ Lemma-construction. The composition βRαR is homotopic to zero.
968
+ In fact,
969
+ there is a canonical up to a homotopy κR such that d(κR) = βRαR.
970
+ Proof. By Remark and Example in 4.3 one has βRαR = ∂∗ι∗ι∂ = ∂∗ǫ∂ = ∂∗χ∗χ∂ =
971
+ (χ∂)∗χ∂. Notice that χ∂ is homotopic to 0; choose a homotopy λ, d(λ) = χ∂. Now
972
+ set κR := λ∗χ∂.
973
+ Independence of κR up to a homotopy from the choice of λ: if λ′ is another
974
+ homotopy as above, i.e., d(λ) = d(λ′), then κ′
975
+ R = λ′∗χ∂ = κR + (λ′ − λ)χ∂ =
976
+ κR + d((λ − λ′)λ).
977
+
978
+ Remark. Our κR is self-dual up to homotopy: Indeed, one has κ∗
979
+ R = (χ∂)∗λ =
980
+ κR + d(λ∗λ).
981
+ 4.5. Below we use the notation from 2.1, 2.2. We have defined (αR, βR, κR) ∈
982
+ DM(X0)(2). It yields the objects ER := E(αR, βR, κR) ∈ DM(X0) and (αI, βI, κI)
983
+ := σ(αR, βR, κR) ∈ DM(X0)(2). As follows from Remark in 4.4 and the defini-
984
+ tions, the above three objects are naturally self-dual.
985
+ Proposition. There is a homotopy equivalence θ : I
986
+
987
+ → ER such that the maps
988
+ βIθ : I → B[1], θ−1αI : B(−1)[−1] are a morphism of the triangle in 4.3(ii) and
989
+ its dual. Our θ is unitary, i.e., θ∗ = θ−1.
990
+ Proof. Recall that we have a natural homotopy equivalence (λ, χ) : Cone(∂ : B →
991
+ Q(0)X0)
992
+
993
+ → I (see 4.3(ii)), and ER is the direct sum B[1] ⊕ R ⊕ B(−1)[−1] with
994
+ (αR, −κR, βR) added to the differential (see 2.1). Our θ is the composition I
995
+
996
+
997
+ Cone(∂)
998
+ θ′
999
+ → ER where θ′ is the next morphism: its restriction to B[1] ⊂ Cone(∂)
1000
+ identifies it with the first summand in ER, and its restriction to Q(0)X0 ⊂ Cone(∂)
1001
+ is (0, ι, −λ∗χ).
1002
+
1003
+ HEIGHT PAIRING AND NEARBY CYCLES
1004
+ 15
1005
+ One has θ∗θ = idI: we need to check that θ′∗ρθ′ = (λ, χ)∗(λ, χ) : Cone(∂) →
1006
+ Cone(∂∗) where ρ : ER
1007
+
1008
+ → E∗
1009
+ R(1 − n)[2 − 2n] is the self-duality for ER. As follows
1010
+ from Remark in 4.4, ρ is the matrix with the self-dualities for R and B’s on the
1011
+ diagonal and the only non-zero off-diagonal entry being λ∗λ : B → B∗(1−n)[2−2n].
1012
+ The rest is an immediate calculation.
1013
+ The assertion that βIθ is the morphism of the triangle in 4.3(ii) means that
1014
+ βIθ′ is the projection Cone(∂) → B[1] which is evident from the construction. The
1015
+ assertion that αIθ is dual to βIθ follows from the unitarity of θ once we know that
1016
+ θ is a homotopy equivalence. Let us check it.
1017
+ Our θ′ is a morphism Cone(B → Q(0)X0) → Cone(B → Cone(βR)[−1]) com-
1018
+ patible with the projections to B, and so it is enough to check that the map
1019
+ (ι, −λ∗χ) : Q(0)X0 → Cone(βR)[−1] is a homotopy equivalence. Since ι is a homo-
1020
+ topy equivalence on U, it is enough to check our claim after applying i∗
1021
+ xα.
1022
+ The story of section 3.3 uses only the six functors formalism and basic facts
1023
+ from 3.1, so it remains literally true in the motivic setting. Consider the canonical
1024
+ homotopy equivalence a : i∗
1025
+ xαR
1026
+
1027
+ → Γ(Pα ∖ Zα) of (3.3.1). By the Verdier dual
1028
+ assertion to the lemma in 3.3, a identifies i∗
1029
+ xα(βR) with minus the residue map
1030
+ r : Γ(Pα ∖ Zα) → Hn−2
1031
+ prim(Zα)(−1)[1 − n] ⊂ Γ(Zα)(−1)[−1]. By (4.2.2) we have a
1032
+ split exact triangle Q(0) → Γ(Pα ∖ Zα)
1033
+ r→ Hn−2
1034
+ prim(Zα)(−1)[1 − n], so a identifies
1035
+ i∗
1036
+ xαCone(βR)[−1]) with Q(0) ⊂ Γ(Pα∖Zα). It follows directly from the construction
1037
+ of a that ai∗
1038
+ xα(ι) coincides with the latter embedding, and we are done.
1039
+
1040
+ 4.6. Proof of the theorem in 1.9.
1041
+ We have (αI, βI, κI) ∈ DM(X0)(2), hence
1042
+ Γ(αI, βI, κI) ∈ DM(2). For two Bloch cycles A, B of classes clA, clB ∈ Hom(Q(0),
1043
+ Hn−2
1044
+ prim(Zα)(m)) we have (cl∗
1045
+ A, clB∗)Γ(αI, βI, κI) ∈ EM(Γ(I)(m − 1)[1 − n]) =
1046
+ EM(Γ(I+)(m − 1)[1 − n]) = EM(M) where M := M(Y )(−m)[−1 − 2m]. By the
1047
+ construction the Hodge realization embedding EM(M) ֒→ EH(M) = EH(Hm(Y ))
1048
+ identifies it with Eψ
1049
+ A,B from 1.6, and we are done.
1050
+
1051
+ References
1052
+ [A1]
1053
+ J. Ayoub, Les six op´erations de Grothendieck et le formalisme des cycles ´evanescents
1054
+ dans le monde motivique (I), Ast´erisque 314, SMF, 2007.
1055
+ [A2]
1056
+ J. Ayoub, Les six op´erations de Grothendieck et le formalisme des cycles ´evanescents
1057
+ dans le monde motivique (II), Ast´erisque 315, SMF, 2007.
1058
+ [B]
1059
+ A. Beilinson, Height pairing between algebraic cycles, K-theory, Arithmetic and Geom-
1060
+ etry, Yu. I. Manin (Ed.), Lect. Notes in Math. 1289, Springer, 1987.
1061
+ [Bl1]
1062
+ S. Bloch, Height pairings for algebraic cycles, Journal of Pure and Applied Algebra 34
1063
+ (1984), 119–145.
1064
+ [Bl2]
1065
+ S. Bloch, Cycles and biextensions, Contemporary Mathematics 83 (1989), 19–30.
1066
+ [BlJS]
1067
+ S. Bloch, R. de Jong, E. Can Sert˜oz, Heights on curves and limits of Hodge structures,
1068
+ arXiv:2206.01220 (2022).
1069
+ [CD]
1070
+ D.-C. Cisinski, F. D´eglise, Triangulated categories of mixed motives, Springer Mono-
1071
+ graphs in Mathematics, Springer, 2019.
1072
+ [G]
1073
+ S. Gorchinskiy, Notes on the biextension of Chow groups, Motives and algebraic cycles,
1074
+ Fields Institute Commun., vol. 56, Amer. Math. Soc., 2009, pp. 111–148.
1075
+ [Il]
1076
+ L. Illusie, Sur la formule de Picard-Lefschetz, Algebraic geometry 2000, Azumino,
1077
+ Advanced Studies in Pure Math, vol. 36, Mathematical Society of Japan, 2002, pp. 249–
1078
+ 268.
1079
+
79E0T4oBgHgl3EQfwQGR/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
89E4T4oBgHgl3EQfDAsR/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8fe77e9914630554285c709198ce967381cbc785ee97e34cff81e5e2b22a9fdd
3
+ size 5767213
89E4T4oBgHgl3EQfDAsR/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a263164f26e513f80e0360149b5dd2aa7e4d98cdf4773bcf12c8760374089eb5
3
+ size 189042
99AyT4oBgHgl3EQfRPYW/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ee64f6c4c69e2857cd893d76e558709d617e4f464639cd9a3bd0ce05c417c099
3
+ size 118902
9NFRT4oBgHgl3EQfqTc6/content/2301.13616v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:49ec0b277626a017dc0caefb3d0f4ef26c8e00c4acd7281355bc831cb5020336
3
+ size 1355035
9NFRT4oBgHgl3EQfqTc6/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:17eb68d0731074ab64fa9aa10ba4244c5eecdf772fac8bc8555c2c62881e7966
3
+ size 203528
9dE1T4oBgHgl3EQfUQOY/content/tmp_files/2301.03088v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
9dE1T4oBgHgl3EQfUQOY/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
ANE1T4oBgHgl3EQf9AZG/content/tmp_files/2301.03551v1.pdf.txt ADDED
@@ -0,0 +1,1727 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ A Lightweight Blockchain and Fog-enabled Secure Remote Patient
2
+ Monitoring System
3
+ Omar Cheikhrouhoua,e,g,∗, Khaleel Mershadb, Faisal Jamilc, Redowan Mahmudd, Anis
4
+ Koubaae, Sanaz Rahimi Moosavif
5
+ aCES Laboratory, University of Sfax, Tunisia
6
+ bComputer Science and Mathematics Department, Lebanese American University, Beirut, Lebanon
7
+ cDepartment of Computer Engineering, Jeju National University, Korea
8
+ dSchool of Electrical Engineering, Computing and Mathematical Sciences, Curtin University, Perth,
9
+ Australia
10
+ eRobotics and Internet of Things Lab, Prince Sultan University, Riyadh, Saudi Arabia.
11
+ fCalifornia State University, Dominguez Hills (CSUDH)
12
+ gISIMA, Mahdia, University of Monastir, Tunisia
13
+ Abstract
14
+ IoT has enabled the rapid growth of smart remote healthcare applications. These IoT-based
15
+ remote healthcare applications deliver fast and preventive medical services to patients at risk
16
+ or with chronic diseases. However, ensuring data security and patient privacy while exchang-
17
+ ing sensitive medical data among medical IoT devices is still a significant concern in remote
18
+ healthcare applications. Altered or corrupted medical data may cause wrong treatment and
19
+ create grave health issues for patients. Moreover, current remote medical applications’ ef-
20
+ ficiency and response time need to be addressed and improved. Considering the need for
21
+ secure and efficient patient care, this paper proposes a lightweight Blockchain-based and Fog-
22
+ enabled remote patient monitoring system that provides a high level of security and efficient
23
+ response time. Simulation results and security analysis show that the proposed lightweight
24
+ blockchain architecture fits the resource-constrained IoT devices well and is secure against
25
+ attacks. Moreover, the augmentation of Fog computing improved the responsiveness of the
26
+ remote patient monitoring system by 40%.
27
+ Keywords:
28
+ IoT, Healthcare monitoring, Lightweight Blockchain, Fog computing,
29
+ consensus protocol.
30
+ 1. Introduction
31
+ Healthcare IoT networks are evolving from centralized to distributed systems to con-
32
+ nect with each other to provide patients with high-quality healthcare. According to pre-
33
+ ∗I am corresponding author
34
+ Email addresses: omar.cheikhrouhou@isetsf.rnu.tn (Omar Cheikhrouhou),
35
+ khaleel.mershad@lau.edu.lb (Khaleel Mershad), faisal@jejunu.ac.kr (Faisal Jamil),
36
+ mdredowan.mahmud@curtin.edu.au (Redowan Mahmud), akoubaa@psu.edu.sa (Anis Koubaa),
37
+ srahimimoosavi@csudh.edu (Sanaz Rahimi Moosavi)
38
+ Preprint submitted to Elsevier
39
+ January 10, 2023
40
+ arXiv:2301.03551v1 [cs.CR] 9 Jan 2023
41
+
42
+ dictions, the current hospital-centered healthcare monitoring systems will develop first to
43
+ hospital–home-balanced in 2025 and then ultimately to home-centered in 2030 [1]. New
44
+ system architectures, technologies, and computing paradigms are needed to realize such
45
+ evolution, specifically in the Healthcare Internet of Things (HIoT) [2].
46
+ Emerging tech-
47
+ nologies like IoT, blockchain, and artificial intelligence have made deploying smart remote
48
+ patient monitoring systems a fact. Indeed, IoT devices permit them to sense and moni-
49
+ tor patients’ physiological parameters, hence exempting them from a long waiting queue at
50
+ a doctor’s visit. All necessary physiological parameters needed by doctors can be sensed
51
+ by the biomedical IoT devices (also known as the Internet of Medical Things devices) and
52
+ sent remotely to the doctor, allowing the latter to decide the appropriate treatment for the
53
+ patient [3].
54
+ The evolution of sophisticated security attacks and the rising need for individualized
55
+ healthcare has made it essential for medical institutions to embrace blockchain technology.
56
+ The arrival of the blockchain provides solutions to several problems that the healthcare sys-
57
+ tem has been facing for a long time. The growing numbers of healthcare data breaches, pa-
58
+ tient privacy violations, counterfeit drugs, and many other issues are major reasons for steer-
59
+ ing the blockchain market’s growth in the healthcare industry. In general, the blockchain
60
+ brings a large number of opportunities to smart healthcare, which can be summarized as
61
+ follows:
62
+ • Secure access to personal health records: the decentralized blockchain system offers the
63
+ power of controlling data access to the owner of the data itself. Smart contracts register
64
+ and authorize users to access the patient’s data according to the patient consent policy.
65
+ • Patient Consent Management: the fundamental features of the blockchain, such as
66
+ transparency and immutability, enables healthcare applications to build trust among
67
+ patients and verify compliance with consent management policies.
68
+ • Traceability of remote treatment: the blockchain permits healthcare applications to
69
+ create immutable and coherent electronic records (EHRs) that can be viewed by all
70
+ stakeholders. The transparency and consistency of blockchain EHRs aid in tracing the
71
+ medical history of patients to offer the appropriate treatment.
72
+ • Traceability of in-home medical kits and devices: the blockchain provides immutable
73
+ and transparent record transactions to the ownership and performance of medical kits.
74
+ Reputation scores of medical devices and kits are saved in the blockchain using smart
75
+ contracts.
76
+ • Reputation-aware specialist referral services: during the treatment of a remote pa-
77
+ tient, medical referrals and expert suggestions are acquired through smart contracts.
78
+ Blockchain enables healthcare providers to store these referral documents on an Inter-
79
+ Planetary File System (IPFS) server, such that an IPFS hash of the document is stored
80
+ securely in the blockchain. The hash prevents the alteration of the stored document
81
+ and maintains its integrity.
82
+ 2
83
+
84
+ • Automated payments: blockchain provides digitally signed automatic payments to
85
+ guarantee non-repudiated secure transactions.
86
+ A complete discussion on the blockchain benefits to smart healthcare applications can
87
+ be found in [4].
88
+ Ensuring the security of the remote patient monitoring (RPM) system is a must. Since a
89
+ vulnerability in such a system could enable attackers to steal/modify sensitive information
90
+ and endanger the patient’s life. The blockchain has emerged as a promising technology that
91
+ can store and secure assets through a transparent and distributed ledger. In healthcare,
92
+ where patient data is a critical asset that needs to be securely managed, the blockchain could
93
+ become the right technology to address this challenge and provide a secure, transparent, and
94
+ tamper-proof management of patient healthcare data. However, the blockchain is a heavy
95
+ system requiring much processing and communication. Lightweight IoT devices would face
96
+ problems if they were to act as full blockchain nodes. Hence, a solution should be adopted to
97
+ enable IoT devices to participate in the blockchain network without affecting their limited
98
+ resources.
99
+ The lightweight blockchain [5, 6] has been proposed to achieve this purpose.
100
+ Here, the blockchain architecture and processes are modified to assign light roles to the IoT
101
+ devices while allowing them to benefit from the blockchain services.
102
+ In traditional RPM systems, patient healthcare data is stored in an Electronic Healthcare
103
+ Record (EHR) and saved in the cloud.
104
+ Cloud computing provides ubiquitous access to
105
+ patients’ data through a user-centric access control model, where the user chooses which
106
+ data and to whom he/she should give access. However, a cloud computing system presents
107
+ the disadvantages of high latency and, therefore, cannot fit critical healthcare application
108
+ requirements where immediate intervention is needed. More precisely, real-time detection
109
+ and notification of abnormal situations must be implemented in the context of a heart disease
110
+ use case. Otherwise, the patient’s life will be at risk.
111
+ To overcome the high latency limits of cloud computing and to fit the real-time require-
112
+ ments of most healthcare applications, we propose leveraging fog computing technology in
113
+ this paper. In our proposed architecture, fog computing will not replace cloud computing
114
+ but will cooperate via the lightweight blockchain to provide real-time and efficient service.
115
+ More precisely, we introduce the fog computing layer that will host a lightweight blockchain
116
+ application with low latency requirements. On the other hand, complex AI algorithms can
117
+ be executed at the cloud computing layer.
118
+ Currently, smart cities are moving towards adopting blockchain technology in many smart
119
+ city applications. In healthcare, and especially in remote patient monitoring, the blockchain
120
+ can change the methods in which the application is executed and managed. Integrating the
121
+ blockchain allows healthcare managers to guarantee the transparency of public healthcare
122
+ data and removes the need to apply trust-based mechanisms and systems to achieve this
123
+ target. In addition, the blockchain guarantees the privacy of patients’ personal data through
124
+ smart contracts. Moreover, the blockchain allows for fast and direct connectivity between
125
+ healthcare officials, providers, staff, and patients. Issuing blockchain transactions allows
126
+ these entities to communicate securely via the blockchain without intermediaries. Finally,
127
+ the blockchain allows healthcare and smart city officials to know the origin and destination
128
+ 3
129
+
130
+ of each medical resource. They can also find out how healthcare services are being used
131
+ without compromising people’s privacy.
132
+ To sum up, we propose a smart and secure remote patient monitoring system based
133
+ on three technology pillars: IoT, fog computing, and blockchain. More precisely, the key
134
+ contributions of this paper are as follows:
135
+ • We propose the architecture of a smart and secure remote patient monitoring system.
136
+ The proposed architecture uses IoT for patient vital signs collection and blockchain to
137
+ guarantee the privacy and security of the patient-collected data.
138
+ • The efficiency of the proposed architecture is achieved through the introduction of the
139
+ fog computing layer to provide real-time response and aggregate the patients’ collected
140
+ data.
141
+ • To reduce the heavy demands of traditional blockchain, we modify the blockchain
142
+ structure to include a local blockchain within the IoT ecosystems and a global chain
143
+ at the cloud layer. Each IoT ecosystem saves the block headers of all blockchain blocks,
144
+ the bodies of the blocks of interest to the local chain, and the smart contract functions
145
+ needed within the local chain. On the other hand, the global chain comprises whole
146
+ blocks and smart contracts.
147
+ • We propose a lightweight consensus model that enables the fog nodes to participate
148
+ in the consensus protocol without consuming a lot of processing and energy resources
149
+ and allows IoT nodes to store only the information they need to verify the legitimacy
150
+ and integrity of the blockchain data that they obtain from fog nodes and cloud servers.
151
+ The remainder of this paper is as follows.
152
+ Section 2 outlines the existing literature
153
+ on the remote patient monitoring system using blockchain and Fog Computing. Section 3
154
+ gives an overview of the proposed remote patient monitoring architecture with its different
155
+ components. Section 4 describes the details of the proposed lightweight blockchain model.
156
+ Section 5 describes the fog computing layer functions and properties.
157
+ The performance
158
+ evaluation of the system is discussed in Section 6. Section 7 analyses the security of the
159
+ proposed system. Finally, we conclude and give future directions in Section 8.
160
+ 2. Related Work
161
+ As our work is based on three technologies, namely the IoT, fog computing, and blockchain,
162
+ in this section, we present relevant work that uses one or more of these technologies to deploy
163
+ a healthcare solution. The discussed works are summarized in Table 1.
164
+ 4
165
+
166
+ Table 1: Summary of related work
167
+ Ref Contribution
168
+ Use case
169
+ Used Technologies
170
+ Pros(+)/Cons(-)
171
+ IoT BC
172
+ FC
173
+ [7]
174
+ Cloud based remote health
175
+ monitoring system with sig-
176
+ nal watermarking
177
+ ECG-based
178
+ health monitor-
179
+ ing
180
+
181
+
182
+ +Providing sig-
183
+ nal
184
+ authentica-
185
+ tion using water-
186
+ marking
187
+ [8]
188
+ A
189
+ hierarchical
190
+ fog-
191
+ computing-assisted
192
+ ar-
193
+ chitecture for IoT health
194
+ monitoring system
195
+ Arrhythmia de-
196
+ tection
197
+
198
+
199
+ +
200
+ Map
201
+ the
202
+ IBM’s
203
+ MAPE-
204
+ K
205
+ computing
206
+ model
207
+ to
208
+ the
209
+ healthcare
210
+ ap-
211
+ plication
212
+ [1]
213
+ They developed a smart e-
214
+ Health gateway localized at
215
+ the edge.
216
+ Heart disease
217
+
218
+
219
+ +Full-system
220
+ implementation
221
+ [9]
222
+ Improved the energy con-
223
+ sumption of sensor nodes
224
+ during
225
+ data
226
+ transmission
227
+ and processing.
228
+ Migraine disease
229
+
230
+
231
+ +Energy
232
+ con-
233
+ sumption reduc-
234
+ tion
235
+ [10]
236
+ An
237
+ Edge-Based
238
+ Architec-
239
+ ture for IoT-Healthcare ap-
240
+ plication.
241
+ Detect
242
+ high-
243
+ stress conditions
244
+ for workers and
245
+ athletes.
246
+
247
+
248
+ -security
249
+ is-
250
+ sues
251
+ are
252
+ not
253
+ addressed.
254
+ [11]
255
+ Used retraining of SDA in
256
+ the
257
+ testing
258
+ phase
259
+ of
260
+ ar-
261
+ rhythmia
262
+ classification
263
+ to
264
+ add or merge features in
265
+ the
266
+ anomaly
267
+ detector
268
+ +
269
+ Blockchain for access con-
270
+ trol
271
+ arrhythmia clas-
272
+ sification
273
+
274
+ +High accuracy
275
+ [12] Used blockchain to secure
276
+ remote patient monitoring
277
+ General
278
+
279
+
280
+ -Time issue
281
+ -Key
282
+ manage-
283
+ ment issue
284
+ Continued on next page
285
+ 5
286
+
287
+ Table 1: Summary of related work
288
+ Ref Contribution
289
+ Use case
290
+ Used Technologies
291
+ Pros(+)/Cons(-)
292
+ IoT BC
293
+ FC
294
+ [13]
295
+ Remote health monitoring
296
+ system using Tor to min-
297
+ imize the latency of the
298
+ blockchain network.
299
+ Cardiac
300
+ Pa-
301
+ tients.
302
+ Sleep
303
+ Apnoea
304
+ Pa-
305
+ tients. Epileptic
306
+ Patients
307
+
308
+
309
+ -The accuracy of
310
+ the system is not
311
+ tested.
312
+ [14]
313
+ HealthFog: A heart disease
314
+ analysis system based on
315
+ ensemble deep learning and
316
+ using integrated IoT and
317
+ Fog computing
318
+ Heart disease
319
+
320
+
321
+ -Security
322
+ is-
323
+ sue
324
+ are
325
+ not
326
+ addressed.
327
+ [15]
328
+ They developed an intelli-
329
+ gent e-Health architecture
330
+ integrating
331
+ AI,
332
+ IoT,
333
+ and
334
+ cloud computing.
335
+ ECG-based
336
+ arrhythmia
337
+ detection
338
+
339
+
340
+ +Hardware
341
+ im-
342
+ plementation of
343
+ AI algorithms
344
+ [16]
345
+ An IoT and fog comput-
346
+ ing architecture with par-
347
+ allelization and core allo-
348
+ cation capabilities to accel-
349
+ erate healthcare processor-
350
+ intensive services
351
+ ECG-based
352
+ arrhythmia
353
+ detection
354
+
355
+
356
+ +Response
357
+ Time
358
+ was
359
+ im-
360
+ proved.
361
+ [17]
362
+ Lightweight identity man-
363
+ agement and access control
364
+ scheme for IoT devices us-
365
+ ing IOTA.
366
+ General
367
+
368
+
369
+ -Does not sup-
370
+ port smart con-
371
+ tracts
372
+ [18]
373
+ Blockchain based architec-
374
+ ture to provide patient cen-
375
+ tric data access
376
+ Healthcare
377
+
378
+
379
+ +
380
+ Use
381
+ cluster-
382
+ ing techniques to
383
+ improve system
384
+ scalability
385
+ BC: Blockchain, FC: Fog Computing or any cloud computing related technologies
386
+ Hossain et al. [7] proposed a cloud-based architecture for ECG signal monitoring. To
387
+ authenticate the captured ECG signal, the authors add a watermark that will be checked on
388
+ the cloud side. Moreover, the authors proposed additional services, including ECG signal
389
+ enhancement, classification, and analysis. Azimi et al., [8] proposed a hierarchical computing
390
+ architecture leveraging fog and cloud computing technologies.
391
+ The authors proposed a
392
+ methodology to partition the existing machine learning methods for fog-enabled healthcare
393
+ IoT systems. The authors in [1] developed a smart e-Health gateway localized at the edge to
394
+ provide several functions, including local storage, real-time local data processing, embedded
395
+ data mining, etc. By releasing the small IoT devices from these functions, a considerable
396
+ 6
397
+
398
+ amount of energy can be saved by outsourcing some loads from sensor nodes to these smart
399
+ gateways.
400
+ In [9], the authors also proposed the integration of the IoT and cloud computing tech-
401
+ nologies to predict migraine disease.
402
+ The authors’ main contributions are the design of
403
+ low-power techniques in the radio and data processing for the sensor nodes. Moreover, the
404
+ authors proposed workload-balancing policies for cloud computing servers.
405
+ The authors in [10] proposed BodyEdge: an edge-based architecture for IoT-healthcare
406
+ applications. The system was implemented in two examples of edge gateway: Raspberry Pi3
407
+ and Zotac CI540 NANO Pc, and its performance was compared to cloud systems. Moreover,
408
+ as a validation example, the authors have implemented the system to detect high-stress
409
+ conditions for users in two different scenarios, namely Workers in a factory and Athletes
410
+ training in a fitness center. In [11], authors used blockchain to secure access control to patient
411
+ EHR. The authors proposed to store patient data in an off-chain database to overcome the
412
+ storage constraint of the blockchain.
413
+ The authors in [12] used a consortium blockchain based on the IBM Hyperledger platform
414
+ to secure remote patient monitoring.
415
+ In their proposed system, sensors interact with a
416
+ gateway (such as a mobile phone) that implements smart contracts for data analysis and
417
+ sends essential notifications to patients and healthcare providers. The blockchain was used
418
+ to securely log transactions (such as data reads and doctor’s commands). However, patient
419
+ data was stored on a local database.
420
+ They have proved that blockchain could be used
421
+ to resolve security concerns about the transfer and logging of data transactions in an IoT
422
+ healthcare system. The limitation rests in perfecting the time of the transmission of the
423
+ aggregated data sent by the gateway to the blockchain nodes.
424
+ The authors in [13] proposed a decentralized peer-to-peer remote health monitoring sys-
425
+ tem. The proposed architecture uses Tor hidden services for off-chain data delivery between
426
+ patients and doctors. The authors in [14] proposed HealthFog, a Fog-based healthcare sys-
427
+ tem that integrates Edge computing and IoT. Their work was motivated by latency-sensitive
428
+ healthcare applications, especially deep learning-based algorithms. The proposed system was
429
+ validated for a health disease use case.
430
+ In [15], an AI-driven e-health solution was proposed. The solution integrates IoT with
431
+ cloud computing. The key difference of this solution is the distribution of the AI intelligence
432
+ across the three architecture layers, namely: the Device layer, Fog layer, and Cloud layer.
433
+ Moreover, the authors proposed hardware implementation of the SVM, ANN, and CNN
434
+ algorithms using digital circuits. The authors in [16] proposed a framework for accelerating
435
+ the response to remote patients requiring the execution of smart eHealth services. Their
436
+ proposed framework supports distributed offloading to fog servers and multicore processors’
437
+ capacity to accelerate its execution.
438
+ The authors in [17] proposed a blockchain-based lightweight authentication and autho-
439
+ rization scheme for IoT devices. The proposed scheme uses distributed ledger technology
440
+ IOTA to design a lightweight and scalable mechanism for identity management and access
441
+ control of IoT devices. However, this solution did not support smart contacts. The authors
442
+ in [18] proposed a blockchain-based architecture that enables data owners to define their
443
+ desired access policies over their privacy-sensitive healthcare data. The architecture used
444
+ 7
445
+
446
+ two separate chains; one for storing data transactions and one for storing access policies.
447
+ Several lightweight blockchain architectures have been proposed in the literature [19, 20,
448
+ 21, 22, 23, 24]. For example, the ECLB protocol in [19] saves the full blockchain on edge
449
+ nodes, while the IoT nodes store what the authors call the fragmented ledger structure,
450
+ which contains the block headers and some of the transactions in each block that are needed
451
+ by the lightweight node. A multi-layer blockchain model is proposed in [20]. The blockchain
452
+ network is divided into three layers. At the first layer, ordinary IoT nodes are divided into
453
+ clusters. At the second layer, IoT cluster heads (CHs) store the local (i.e., cluster) copy of
454
+ the BC. The cellular base stations (BSs) store the full global BC at the third layer. Nodes
455
+ in Layers 2 and 3 collaborate to create new blocks and execute the consensus algorithm. On
456
+ the other hand, some IoT nodes at layer one can be peers that maintain a copy of the local
457
+ BC and act as transaction endorsers or committers, while CHs and BSs act as Hyperledger
458
+ orderers who order transactions and create blocks.
459
+ From the study of the existing work, we note that several works address only the per-
460
+ formance and energy aspect of their proposed RPMs by adding the fog layer that manages
461
+ the computing and data processing tasks [7, 8, 1, 9, 10, 14, 15, 16]. However, these schemes
462
+ did not address the security aspects, and therefore, they are vulnerable to attacks. Other
463
+ schemes such as [11, 12, 13, 17, 18], addressed the security aspect by adopting the blockchain
464
+ technology, however, they used classical blockchain platforms and architectures that cannot
465
+ fit the resource-constrained IoT devices. In this paper, we leverage the fog computing layer
466
+ not only to improve the performance of the RPM system but also to lighten the load of
467
+ blockchain technology. More precisely, the fog layer permits the proposal of a lightweight
468
+ blockchain architecture that provides security services adaptable to resources constrained
469
+ IoT devices. Moreover, the proposed consensus mechanism frees the architecture from the
470
+ burden of classical consensus algorithms such as PoW or PoS [25, 26].
471
+ 3. The Proposed RPM System Overview
472
+ This section gives an overview of the proposed RPM system. It highlights its three-layer
473
+ architecture and the different communication interaction between the components.
474
+ 3.1. RPM System Architecture
475
+ The proposed remote patient monitoring system is a three-layer architecture as shown
476
+ in Figure 1, and which are:
477
+ • The IoT devices layer: This layer, composed of biomedical sensor nodes, wearable
478
+ sensor nodes, and IoT medical devices, is responsible for collecting the vital signs of the
479
+ monitored patient. These sensor readouts are collected continuously; however, their
480
+ transmission to the gateway node located at the fog layer can be done periodically. The
481
+ transmission period depends on the nature of the vital sign and generally is determined
482
+ by the patient supervising doctors.
483
+ • The Fog Computing layer: This layer is responsible for the lightweight processing
484
+ of vital signs received from the IoT layer. For example, suppose the monitored vital
485
+ 8
486
+
487
+ Cloud Layer
488
+ DB
489
+ DB
490
+ DB
491
+ Gateway
492
+ Gateway
493
+ Gateway
494
+ Gateway
495
+ Edge/ Fog Layer
496
+ IoT Layer
497
+ Global Blockchain
498
+ Big data Analytics
499
+ Patient
500
+ Doctor
501
+ IoMT devices
502
+ GetVitalSign
503
+ Write/Order
504
+ Local Blockchain
505
+ GiveAccess/
506
+ Audit Data
507
+ GetVitalSign/
508
+ GetNotification
509
+ cloud
510
+ server
511
+ cloud
512
+ server
513
+ cloud
514
+ server
515
+ Blockchain
516
+ update
517
+ Figure 1: The three-layer architecture of the proposed remote patient monitoring system
518
+ sign exceeds a specific threshold. In that case, an alert message will be triggered and
519
+ sent to the patient and the supervising doctor to make the right decision. Moreover,
520
+ the fog layer first decides which data needs to be recorded in the blockchain network
521
+ and then interacts with this latter. The fog layer will also aggregate continuous sensed
522
+ data before sending it to the cloud server for permanent storage and data analytics.
523
+ Additionally, the fog layer contains IoT gateways that include the local blockchain
524
+ network, which is a subset of the global blockchain network (please refer to Section 5 for
525
+ full description). In the proposed system, the fog computing module consists of many
526
+ geographical intelligent gateways, that is, forming the fog. Each gateway supports
527
+ different protocols for communication and serves as a point of contact between the
528
+ sensor network and the cloud. It collects data from different sub-networks, translates
529
+ protocols, and offers other higher-level services, including filters, data aggregation,
530
+ analysis, and so on. The fog computing layer extends cloud computing to the edge of
531
+ the network and its facilities. From cloud to end users/devices, the fog recognizes real-
532
+ time interaction, dense geographical distribution, heterogeneity, accessibility support,
533
+ pre-processing interoperability along with cloud interaction [27]. This enables latency
534
+ to be decreased, particularly for real-time applications such as in-house IoT monitoring
535
+ of patients. The fog reduces contact with the cloud, particularly in the event of a loss
536
+ of cloud connectivity, where the data is stored locally on these gateways, and patients’
537
+ data is sent to the cloud when the connection is restored.
538
+ • The Cloud Computing layer: This layer is responsible for permanent data storage
539
+ and data analytics. Complex AI and deep learning algorithms can be implemented
540
+ at this layer for data classification, disease detection and prediction, and treatment
541
+ 9
542
+
543
+ +OAEplan decision.
544
+ Moreover, in the proposed architecture, cloud servers play the role
545
+ of full blockchain nodes that store the full copy of the blockchain and participate in
546
+ transaction validation, block generation, and consensus. The blockchain records pa-
547
+ tient data and actions of patients and caregivers and permits the patients to decide to
548
+ whom they give access to their data. Moreover, blockchain technology contains pieces
549
+ of code called smart contracts that can be automatically triggered when an event is
550
+ achieved. These smart contracts are a powerful tool for a remote patient monitoring
551
+ system as they can trigger an alarm and notify the doctor in an abnormal situation
552
+ (for example, when the vital sign value exceeds a specific threshold). In addition, the
553
+ blockchain is used to ensure patient data privacy and the system’s security. First,
554
+ thanks to blockchain technology, the patient will be given an anonymous identity.
555
+ This permits hiding the real patient’s identity; therefore, doctors can treat his/her
556
+ data privately. Moreover, in our system, we propose to use a private blockchain. This
557
+ type of blockchain has the advantage of restricting access to users’ data to only au-
558
+ thorized persons (such as patients, doctors, and caregivers). Furthermore, blockchain
559
+ architecture permits a patient-centric data management architecture. More precisely,
560
+ the patient will decide to whom he/she shares data access (please refer to Section 4
561
+ for more details).
562
+ 3.2. Communication Models
563
+ The fog layer enables us to control access to IoT devices for medical applications. Each
564
+ fog node manages and operates a group of medical IoT devices. This layer also interacts
565
+ with a network of fog nodes allowed by blockchain, which function together on the Internet.
566
+ All the related smart medical devices are connected with the closest blockchain-enabled
567
+ fog node, e.g., in an in-house monitoring scenario. These blockchain-enabled fog nodes are
568
+ communicated by IoT nodes and system users for authentication, authorization, and safe
569
+ communication synchronization. An intelligent contract with a collection of rules can also
570
+ be established on top of the fog nodes allowed by blockchain. Furthermore, the consensus
571
+ algorithm is performed to validate the transactions and blocks for those transactions after
572
+ they are created.
573
+ Transaction blocks can be exchanged between cloud servers and the
574
+ blockchain-enabled fog nodes or between the fog nodes to support robust authentication,
575
+ permission, and distributed secure communication. The proposed solution mainly includes
576
+ four forms of communication:
577
+ (1) Medical caregiver-to-fog communication: Where the end user (e.g., a healthcare
578
+ provider) is willing to use a particular IoT system, he first sends a request for au-
579
+ thentication with a query authentication function specifying the sensor details to the
580
+ blockchain-enabled fog node. The fog node with the blockchain feature will search for
581
+ that medical attendant in the available list of approved sensor equipment. A reject
582
+ message will be given when the user is not allowed to access the requested data. Oth-
583
+ erwise, if the user is approved, the blockchain-enabled fog node issues an access token
584
+ containing Unique Identification (UID) information, length, time of access, blockchain
585
+ address for the data, user blockchain address, and blockchain address of the fog node
586
+ 10
587
+
588
+ that stores the requested data. Notice that every fog node, sensor, and the user has a
589
+ unique blockchain address.
590
+ (2) Medical sensor-to-fog communication: The sensor-to-fog correspondence has two
591
+ principle objectives in our framework.
592
+ IoT medical services system mainly aims to
593
+ validate and authorize the clinical sensors. The following goal is to insert a blockchain-
594
+ enabled fog connected to sensor devices. It helps new sensors to enlist with the mist and
595
+ ensures that all sensors are recognizable by the blockchain network. In our context, each
596
+ IoT medical care system has at least one blockchain-enabled fog node close to the entire
597
+ blockchain network and is used for the enlistment, confirmation, and authorization of
598
+ IoT medical care gadgets with the same framework. Initially, the gadgets will enroll
599
+ with their associated blockchain fog node. As an exchange and blocks are made for
600
+ them, data concerning these gadgets are placed in the blockchain. These blocks would
601
+ then be transported between the wide range of different blockchain fog nodes. Should a
602
+ system with a collecting place need confirmation and consent, the associated blockchain-
603
+ enabled fog node should be given its certifications. The blockchain approves the provided
604
+ accreditation, and if there are significant requirements, the IoT gadgets for medical care
605
+ are effectively checked and authorized. If the certification is not valid, the gadget is
606
+ refused and will not obtain permission to access the blockchain data.
607
+ (3) Fog-to-fog communication: The main goal is to synchronize the information associ-
608
+ ated with IoT medical service confirmation and approval across all blockchain-enabled
609
+ fog nodes [28]. Several biological or physiological parameters are obtained by medical
610
+ sensors transmitted by patients. Medical IoT programs should be reliable and diligent
611
+ in supporting patients moving to a hospital or home. Typically, the mobility support
612
+ of the medical sensors from the upper layer (i.e., fog layer) should be given so that zero
613
+ reconfiguration in the sensor layer is essential. The strategic location and distribution
614
+ of smart gates in the fog layer can be used to provide smooth mobility for medical
615
+ sensors and relieve processing loads. Fog-to-fog contact helps patients wander around
616
+ the hospital wards, ensuring their health monitoring is not disrupted. The patient-free
617
+ movement provides a high level of medical services using a portable patient monitoring
618
+ system. Support of mobility for healthcare IoT systems is one of the most critical prob-
619
+ lems [29]. The improvements to patients’ quality of life in such programs are essential
620
+ [27]. It is important to encourage patients to walk into the hospital/medical facilities
621
+ knowing that monitoring their well-being is not disrupted. It is necessary to establish
622
+ self-configuration or transfer mechanisms to ensure safe and successful data transfer be-
623
+ tween different Medical Sensor Networks (MSNs) [30] to achieve ongoing monitoring of
624
+ patients considering mobility support. For example, when a patient is moving across the
625
+ clinics, a data transfer mechanism is described as the process of modifying or updating
626
+ the registration of mobile sensors on its MSN base. Data handover solutions should
627
+ allow ubiquity when they need to function independently without human interference.
628
+ (4) Medical sensor-to-medical sensor communication: When two clinical devices are
629
+ effectively tested and approved (both have a position with a similar system or another
630
+ one), they may create a safe link to each other and convey information.
631
+ In a case,
632
+ for example, where a patient is released from a clinic but still needs to be constantly
633
+ 11
634
+
635
+ monitored. The doctors bind the patient’s body before he/she leaves the medical cen-
636
+ ter to health tracking devices, including blood pressure monitors, pulse sensors, blood
637
+ glucose monitoring sensors, etc. These devices sense the patient’s blood pressure, heart
638
+ rate, and glucose level and transmit them through a safe channel to the health workers.
639
+ These devices can also interact with the patient’s intelligent home devices. For exam-
640
+ ple, if the patient’s condition becomes severe or a fall is detected, an immediate alarm
641
+ may automatically be activated. The hospital-related devices must interact to check the
642
+ availability of hospital beds in a smart city to ensure a correct count. The proposed
643
+ framework provides medical devices with access control in the IoT healthcare system.
644
+ Under this mechanism, devices can only communicate with recorded and successfully
645
+ authenticated and certified devices with blockchain-enabled fog nodes. A device not reg-
646
+ istered in the blockchain cannot authenticate itself or communicate with other devices
647
+ within the same healthcare ecosystem or external ones. The contact between malicious
648
+ devices and legitimate devices would also be alleviated.
649
+ In what follows, we detail the proposed lightweight blockchain model and the Fog layer
650
+ functions and properties.
651
+ 4. The Blockchain Module Description
652
+ Blockchain technology provides a decentralized, transparent, authenticated platform that
653
+ applies a consensus-driven approach to facilitate the interactions of multiple entities through
654
+ the use of a shared ledger. Beyond the financial sector, where much of the initial develop-
655
+ ment is taking place, blockchain has the potential to revolutionize the healthcare system.
656
+ By providing doctors, patients, researchers, and other healthcare professionals with a mech-
657
+ anism for the controlled exchange of sensitive, permissioned data, blockchain technology can
658
+ improve data sharing and transparency between clinical and research data systems. Any
659
+ healthcare organization participating in a blockchain consortium would be able to share
660
+ medical information, regardless of their native electronic health record system. Blockchain
661
+ provides significant opportunities for healthcare organizations to deliver more efficacious
662
+ treatments and diagnoses through increased provider data sharing and potentially safer and
663
+ more effective remote patient monitoring through advanced technologies such as AI.
664
+ 4.1. Blockchain Architecture of Proposed RPM System
665
+ We propose a lightweight blockchain architecture to manage the data storage and re-
666
+ trieval operations in the remote patient monitoring system. In our system, we implement a
667
+ lightweight blockchain architecture that aims at reducing the delay in accessing the cloud
668
+ by the end users while maintaining the security and immutability of data at all nodes. The
669
+ blockchain will store all healthcare-related data, such as the IoT sensor readings, lab test
670
+ results, physicians’ decisions, commands, etc.
671
+ In addition, the blockchain will comprise
672
+ transactions that contain management and security-related data, such as nodes’ and users’
673
+ registrations, access requests, smart contract results, etc.
674
+ In the proposed architecture, cloud servers play the role of full blockchain nodes that store
675
+ the full copy of the blockchain and participate in transaction validation, block generation,
676
+ 12
677
+
678
+ and consensus. On the other hand, IoT gateways play the role of light blockchain nodes
679
+ that store part of the blockchain. In our system, each gateway will be connected to a certain
680
+ number of IoT networks. For example, an IoT gateway at a patient’s home will connect to
681
+ a single IoT network that contains the IoT devices that are monitoring the patient. On the
682
+ other hand, an IoT gateway at a hospital could connect several IoT networks, such as IoT
683
+ devices, in several patients’ rooms. Here, the IoT devices in a certain room or Lab form a
684
+ separate IoT subnetwork since the data produced by these devices will be linked together
685
+ (for example, data related to a specific patient, doctor, lab, etc.).
686
+ The IoT gateways and sensors that exist in the same IoT ecosystem (for example, home,
687
+ hospital, health institution, etc.) form a cluster that store and manage a local blockchain.
688
+ Each local blockchain is created as part of the full blockchain that is related to the cor-
689
+ responding ecosystem. For example, in a certain patient’s home, a set of IoT devices are
690
+ connected to an IoT gateway. The devices and gateway form a cluster that store and man-
691
+ age a local blockchain that contains the blockchain blocks related to that home only. In a
692
+ hospital, several gateways and sets of IoT devices will form a cluster that store and manage
693
+ the blockchain of the hospital.
694
+ In each cluster, the sensor nodes store only the blocks headers of the full blockchain,
695
+ while the gateways store the block headers of the full blockchain in addition to the full blocks
696
+ of the local blockchain (as illustrated in Figure 2). In addition, to avoid overwhelming the
697
+ gateways with excessive storage as the blockchain grows, each transaction will have an expiry
698
+ time after which it becomes obsolete (for example, when the information in the transaction
699
+ becomes old and is no more relevant). Each gateway saves a data structure that contains,
700
+ for each transaction, the ID of the block in which the transaction is stored (BlockID) and
701
+ the transaction expiry date (Tex). The gateway continuously updates the data structure
702
+ when a transaction expires. In addition, the gateway searches the data structure to detect
703
+ any block in which all transactions have expired and deletes it. Using this approach allows
704
+ the gateway to remove old blocks and create room in its storage for new blocks in the local
705
+ blockchain.
706
+ In the proposed system, IoT nodes continuously generate data and send them to the
707
+ IoT gateway. In addition, healthcare providers (doctors, nurses, scientists, etc.) send their
708
+ data (such as prescriptions, sensors’ configurations, lab test results, commands to activate
709
+ actuators, data analytic results, etc.) to the nearest IoT gateway in their institution’s IoT
710
+ cluster. The IoT gateway stores the data it receives in a temporary cache. Each small period
711
+ (for example, every 100 ms), the IoT gateway aggregates and groups the received data into a
712
+ blockchain block and sends it to the cloud server. Note that each block can contain multiple
713
+ transactions. For example, the readings of a certain sensor can be aggregated into a single
714
+ transaction. Similarly, if the doctor is sending configuration commands to the IoT sensor,
715
+ the configuration settings of each sensor can be grouped into a transaction. Each transaction
716
+ will be signed by the owner that created the transaction.
717
+ 4.2. Consensus Protocol
718
+ We consider a network of cloud servers that are used by various healthcare providers
719
+ to manage the system. As mentioned, the cloud servers act as full blockchain nodes that
720
+ 13
721
+
722
+ Figure 2: The architecture of the proposed blockchain model: the cloud servers store the full blockchain,
723
+ the IoT gateways save the local blockchain, while the IoT nodes store the block headers.
724
+ store all the blockchain blocks. In addition, the cloud servers participate in the blockchain
725
+ consensus protocol. Each cloud server has a unique blockchain ID. The cloud servers create
726
+ the blockchain blocks successively based on their IDs. In other words, the server with the
727
+ smallest ID creates the first block, followed by the server that has the second smallest ID,
728
+ and so on. When the server that has the biggest ID creates a block, the turn goes back to
729
+ the first server. Note that the block generation time at the IoT gateway should be adjusted
730
+ to allow all the cloud servers to generate their blocks in order to avoid block accumulation
731
+ at the cloud servers.
732
+ When its turn to create the new block arrives, a cloud server CS 1 broadcasts the block
733
+ that it received from the gateway to all the cloud servers. Each cloud server CS i verifies that
734
+ all transactions in the block are legitimate by validating the signature of each transaction.
735
+ Next, CS i replies with a CONFIRM message to CS 1. The confirm message contains CS i’s
736
+ signature of the new block. However, If CS i discovers that one or more transactions in
737
+ the block are not valid, it replies with an ERROR message. In its turn, CS 1 waits until it
738
+ receives at least (N /2+1) CONFIRM messages before it adds the block to the blockchain
739
+ and broadcasts its ID in a “Block Add” message to all cloud servers. Here, N is the number
740
+ of the cloud servers.
741
+ This mechanism allows a cloud server to add the new block after
742
+ the majority of cloud servers confirm its validity. The ”Block Add” message contains the
743
+ signatures that CS 1 received in the CONFIRM messages. When a cloud server CS j receives
744
+ a ”Block Add” message, it checks the attached signatures to ensure that more than (N ÷ 2)
745
+ cloud servers have validated and confirmed the new block, before adding it to its blockchain.
746
+ After receiving the “Block Add” message, each cloud server adds the new block to its
747
+ blockchain and broadcasts it to its clusters. Note that each cloud server can serve multiple
748
+ 14
749
+
750
+ Block Header
751
+ Headers Blockchain
752
+ TX 1
753
+ Block Body!
754
+ TX 6
755
+ Block
756
+ Block
757
+ IoT Sensors
758
+ Full
759
+ Blockchain
760
+ Cloud
761
+ IoT Gateway
762
+ Server
763
+ Local Blockchain
764
+ +
765
+ Block 0
766
+ Block 1
767
+ Block 2
768
+ Block k-1
769
+ Block k
770
+ Header
771
+ Header
772
+ Header
773
+ Header
774
+ Header
775
+ Body
776
+ Body
777
+ Local BlocksFigure 3: A sample scenario of the proposed consensus algorithm.
778
+ institutions and organizations, with each institution/organization having its own IoT cluster.
779
+ Each gateway in a cluster examines the new block to determine if it contains transactions
780
+ that were generated by one of the IoT networks in the cluster. If yes, the gateway stores
781
+ the block in its local blockchain and sends it to the IoT devices that are connected to it.
782
+ Each IoT device validates the block (by hashing it and comparing the result to the hash
783
+ in the block header) and then stores the block header in the headers’ blockchain. Next,
784
+ the IoT device caches the block body for a small period of time before deleting it. On the
785
+ other hand, if the new block does not contain transactions that were generated by an IoT
786
+ network in the cluster, the gateway validates the block, sends it to the IoT devices that are
787
+ connected to it, extracts the block header, and adds it to the headers’ blockchain, and then
788
+ deletes the block. Each IoT device that receives the block performs the same operations
789
+ as the gateway. This allows the gateway and IoT devices to maintain the headers of all
790
+ blocks in the blockchain and use these headers to validate any block from outside their local
791
+ blockchain that they obtain from the cloud servers in the future. The proposed consensus
792
+ protocol is illustrated in Figure 3. In the figure, gateways G1 and G2 are connected to cloud
793
+ server CS1, while gateway G3 is connected to cloud server CS2. At a certain time, G2 creates
794
+ a new block B1 and sends it to CS1. When its turn to generate a new block arrives, CS1
795
+ broadcasts B1 to the cloud servers. Each cloud server confirms B1 by sending a CONFIRM
796
+ packet to CS1. Next, CS1 sends a “Block Add” packet to the cloud servers, and each cloud
797
+ server sends the new block to its gateways. G1 and G2 receive B1 from CS1 and add it to
798
+ 15
799
+
800
+ Block B1
801
+ Block B1
802
+ Block Bi
803
+ Block B1
804
+ Send Bi
805
+ Creation
806
+ Broadcast :
807
+ Commit
808
+ ppy
809
+ to Gateways
810
+ Gateway Gi
811
+ Gateway G2
812
+ B1
813
+ Gateway G3
814
+ Cloud Server CS
815
+ wait for turn
816
+ Cloud Server CS2
817
+ Cloud Server CS3
818
+ i
819
+ Cloud Server CS.
820
+ New Block Packet
821
+ Add B ock Packet
822
+ * : Add Block to Local Blockchain
823
+ Block Broadcast Packet
824
+ ^ : Add Block Header to Local Blockchain
825
+ Block ACK Packettheir copies of the local blockchain (since B1 was generated by a gateway in CS1’s cluster),
826
+ while G3 receives B1 from CS2 and adds its header only to the local blockchain.
827
+ 4.3. Smart Contracts and Data Management
828
+ When an IoT device or a user requires data from the blockchain, it sends a request to the
829
+ IoT gateway. The latter searches for the data in its local chain. If it finds it, the gateway
830
+ authenticates the sender and verifies that it has access to the requested data. If yes, the
831
+ gateway replies directly to the sender with the block that contains the data and the token
832
+ that enables the sender to access the data (more about this soon). If the gateway finds
833
+ that the required data doesn’t exist in the local blockchain, it forwards the request to the
834
+ cloud server. The latter performs the same operation, i.e., it authenticates the requesting
835
+ node and verifies that it has access to the requested data. If yes, the cloud server sends the
836
+ block that contains the data and the access token to the gateway, which forwards them to
837
+ the sender. When the latter receives the block, it validates it using the headers blockchain
838
+ before it retrieves the required transactions from the block and decrypts it using the access
839
+ token.
840
+ Note that in our system, all transactions that can be accessed together are assigned an
841
+ access token by the creator and saved into a smart contract. When the creator wants to
842
+ grant access to the transaction to a certain node/user, the creator executes a smart contract
843
+ function that adds the ID of the node/user to the access list of these transactions that is
844
+ saved in the smart contract. When the node/user wants to access the transactions, it should
845
+ authenticate itself and obtain the access token as described before. If the transactions belong
846
+ to the local chain, the smart contract is executed by the gateway within the local chain.
847
+ Else, the smart contract is executed by the cloud server within the full chain. The various
848
+ subsystems and interactions in the proposed RPM platform are presented in Figure 4.
849
+ 5. Specifications of Gateways at the Fog Layer
850
+ The fog layer is made up of IoT gateways which function primarily as a hub between
851
+ the cloud and IoT levels [31].
852
+ With an in-depth study of the role of the gateway in a
853
+ smart home/hospital, where the location and mobility of things and users are confined
854
+ to hospital premises or buildings, it can be recognized that the stationary nature of the
855
+ gateways empowers them with the property of being non-resource constrained in terms of
856
+ power consumption, processing power, and communication. These advantages can be used
857
+ by allowing gateways with ample intelligence, computing power, and structured networks.
858
+ An inter-device communication is the key task of a gateway and supports numerous
859
+ wireless protocols.
860
+ We broaden the function of such gateways into fog enablers by (1)
861
+ building a distributed gateway network and (2) implementing features such as the repository
862
+ (i.e., local data processing and storage using blockchain) to temporarily preserve data for
863
+ analysis by sensors and users. These are important to provide local pre-processing of sensor
864
+ information and, therefore, to be an intelligent gateway for medical services. In a smart
865
+ gateway, the main functions are:
866
+ 16
867
+
868
+ Figure 4: Interaction Model of the proposed blockchain-based remote patient monitoring system.
869
+ 5.1. Local data processing and storage
870
+ Local data processing is a key aspect of fog computing and is performed locally so
871
+ that intelligence is accessible at the doors.
872
+ Based on the device architecture, fog/edge
873
+ layers must continuously handle a large amount of information and respond to different
874
+ conditions in a short time. In the remote patient management system, this becomes more
875
+ important by allowing the system to respond to medical emergencies as quickly as possible.
876
+ Gateways should store inbound information in local storage to ensure that the remote patient
877
+ monitoring system can quickly recuperate patient medical data. In the proposed system, we
878
+ make use of the local blockchain to achieve this objective in a secure manner. The patient
879
+ data can be stored as blockchain transactions in an encrypted or compressed form depending
880
+ on their context and security requirements. The gateway stores all data related to the local
881
+ cluster in the local blockchain. In addition, when the gateway receives a blockchain block
882
+ that contains data related to other clusters, it caches the block for a small period of time to
883
+ allow users in the cluster who require data from the block to access it in a fast manner while
884
+ the data is hot. Moreover, since the network bandwidth is limited between the gateway and
885
+ the cloud, the locally cached blocks can be used to maintain a continuous data flow in the
886
+ event of a weak or unstable connection.
887
+ 5.2. Data filtering
888
+ Data from various medical sensors must be obtained before further processing, e.g., data
889
+ analysis, on the fog layer. The major sources of knowledge for the assessment of the health
890
+ status of a patient in the remote patient monitoring system [32, 33] are bio-signals, for exam-
891
+ ple, Electroencephalography (EEG), Electrocardiogram (ECG or EKG), and Electromyog-
892
+ raphy (EMG). They typically have complex types that have small amplitudes and varying
893
+ 17
894
+
895
+ Remote Patient Monitoring Healthcare System
896
+ Smart Contract
897
+ Call
898
+ Smart
899
+ Identity
900
+ Storage
901
+ Communication
902
+ Consensus
903
+ Smart Contract
904
+ Contract
905
+ Management
906
+ Management
907
+ Protocol
908
+ Reply
909
+ Interface
910
+ Patient
911
+ Data Access
912
+ Data
913
+ Local
914
+ All Block
915
+ All Blocks
916
+ Data
917
+ Blocks
918
+ Validation
919
+ Headers
920
+ Enrollment
921
+ loT Gateway
922
+ Cloud Server
923
+ Certificate
924
+ Smart Contract
925
+ Healthcare Blockchain Platform
926
+ Nurse
927
+ Result
928
+ Healthcare
929
+ Registration
930
+ Certificate
931
+ Data
932
+ Sensor1
933
+ Sensor2
934
+ Sensor3
935
+ Sensorn
936
+ Sensor4
937
+ Healthcare Sensors
938
+ Doctorfrequencies. It is important to remember that noise is often introduced to the signals during
939
+ a patient’s body sensing in a way that distorts the accuracy of the signals. These noises
940
+ are caused by different sources, including electromagnetic interference from other electrical
941
+ devices, shifts in current in the electricity grid, and inappropriate attachment of sensors
942
+ to the body of users. At the fog level, due to the proximity of the sensors, the gateway
943
+ addresses this issue. The fog layer is digitized via different contact protocols by sensors
944
+ (e.g., 6LoWPAN, Zigbee, etc.). While sensors are able to perform lightweight filtering to
945
+ eliminate certain noises during the data collection process, the fog layer offers more complex
946
+ and robust data filtering.
947
+ 5.3. Data analysis
948
+ With local data analysis in the fog layer, the sensitivity of the device can be corrected.
949
+ It helps the device to anticipate and diagnose situations of emergency. The developed deep
950
+ learning module for detecting irregular cardiac conditions is implemented in the fog layer in
951
+ our proposed RPM medical system. The deep learning module can categorize signals and
952
+ detect abnormal conditions on the basis of the sensed ECG signal. As a result, the device
953
+ responds more accurately, rapidly, and in real time to emergency situations. In addition,
954
+ local input and locally sensed data analysis change the quality and reliability of the device
955
+ in the event of the unavailability of the Internet link. Internet disconnection may occur
956
+ regularly for the long-term monitoring of patients with chronic diseases. Fog computing, in
957
+ this case, provides local maintenance of the system’s features. Thus, the sensed data and
958
+ processing results can be kept locally on the fog layer and later synchronized to the cloud via
959
+ the blockchain. Data analysis in the fog often helps the device minimize severe parameter
960
+ processing latencies.
961
+ 5.4. Improved latency
962
+ Agile responses and quick decision-making for acute diseases and emergencies, where
963
+ transmission time and data processing are to be reduced, are important for a continuous
964
+ remote control system. When raw medical data is transferred from medical sensor nodes to
965
+ the cloud, cloud computing can trigger response latencies indefinitely if the network condition
966
+ is not predictable. This becomes serious when streaming-based data processing, such as that
967
+ EEG or ECG signals that are obtained from patients, is needed. Hence, deploying high-
968
+ priority data analytics in distributed gateways in the fog later and making time-sensitive
969
+ and critical decisions inside the local network make the remote patient monitoring system
970
+ more predictable and robust. The processed data can then be transmitted for storage and
971
+ further processing to the cloud.
972
+ 5.5. Sensor nodes energy efficiency
973
+ There are various drawbacks to the processing of data at sensor nodes, as medical sensors
974
+ are resource-restricted devices. Complicated tasks can, in certain cases, be performed suc-
975
+ cessfully at sensor nodes but at significant energy costs. The transfer of heavy-weight tasks
976
+ from sensors to intelligent gateways in the fog layer can be an effective solution for solving
977
+ the above-mentioned problem, in particular when sensors do not have sufficient resources.
978
+ 18
979
+
980
+ Much energy can be saved with the aid of fog computing by outsourcing tasks from medical
981
+ sensors to intelligent gateways.
982
+ 6. Performance Evaluation
983
+ In this section, we present the performance evaluation of the proposed RPM system. We
984
+ have mainly evaluated the performance of the proposed blockchain module and demonstrated
985
+ the efficiency of Fog computing in dealing with critical healthcare applications.
986
+ 6.1. Blockchain Implementation and Performance Evaluation
987
+ The proposed blockchain model was implemented via the Hyperledger platform. Hy-
988
+ perledger is an open-source development platform for blockchain applications. It has been
989
+ widely used as an implementation platform by the research community and is considered
990
+ a benchmark tool to evaluate the performance of the proposed approach against state-of-
991
+ the-art approaches. For smart contracts, the Hyperledger tool provides easy-to-configure
992
+ and user APIs, thus making validation easy for our research work. Furthermore, the REST-
993
+ ful API is utilized to provide the functionality of interoperability and expose the back-end
994
+ blockchain services to the client application through which the patients or other medical
995
+ personnel interact with the system. The smart contacts are designed and aggregated in the
996
+ form of .bna files known as business network archive. Hyperledger Composer [34] is used
997
+ to implement and design the proposed medical blockchain, which aims to enhance system
998
+ operations in terms of throughput and latency. Hyperledger Composer is an open-source
999
+ tool used to design blockchain applications. The .bna in the designed platform consists
1000
+ of a model, query definition, transaction, and access control rules. The model file is the
1001
+ combination of participants, assets, and transactions. The participants are the user of the
1002
+ system who can interact with the system to commit transactions. Similarly, the assets are
1003
+ the medical services that are used by the system users (participants), which are stored in
1004
+ the blockchain. Likewise, transactions are operations that are used to communicate with
1005
+ assets. Moreover, transactions are also used to amend the values of assets and participants.
1006
+ Similarly, the access control rules are also defined to yield authentication and authorization
1007
+ to the users of the system. We also used the world state database to store the blockchain
1008
+ data. We specified the queries that are required to determine the interaction between the
1009
+ blockchain and the world state database. The queries are also used to fetch the user-based
1010
+ customized data from the database.
1011
+ Table 2 encapsulates the business network archive file with transactions, assets, and
1012
+ participants. The users are patients, doctors, and nurses. Similarly, the assets comprise
1013
+ patients’ medical records, sensors, vital sign readings, and other healthcare records. Lastly,
1014
+ transactions include getVitalSignReadings, AddHealthcareSensor, and DetectStatus.
1015
+ The business network archive is then used to construct a Representational state trans-
1016
+ fer (REST) Application Program Interface (API) in order to provide communication be-
1017
+ tween the client application and the back-end database. The RESTful API provides cross-
1018
+ accessibility, where the user of the system can access it from any platform with authentic
1019
+ credentials. Table 3, presents the RESTful API for the proposed medical blockchain, which
1020
+ 19
1021
+
1022
+ Table 2: Smart Contract Modeling for Proposed RPM System
1023
+ Type
1024
+ Components
1025
+ Description
1026
+ Asset
1027
+ Healthcare˙Sensor
1028
+ Healthcare sensors, such as ECG, or
1029
+ EMG etc.
1030
+ Vital Sign˙Sensing˙Data
1031
+ The vital signs of patients acquired
1032
+ from healthcare sensor.
1033
+ HealthRecord
1034
+ The patient medical information,
1035
+ such as current health condition, de-
1036
+ ployed sensors, etc.
1037
+ Participant
1038
+ Doctor
1039
+ System user.
1040
+ Patient
1041
+ System user.
1042
+ Nurse
1043
+ System user.
1044
+ Transaction
1045
+ getVitalSignReadings
1046
+ Get vital sign reading from health-
1047
+ care sensors.
1048
+ Add˙Healthcare˙Sensor
1049
+ Addition of new healthcare sensor in
1050
+ a medical blockchain platform.
1051
+ Modify˙Sensor
1052
+ Modify sensor composition.
1053
+ Detect˙Status
1054
+ Detect the patient vital sign status.
1055
+ is based on HTTP protocol. The generated RESTful API is used to expose the medical plat-
1056
+ form services to the client application. The services are related to patients, nurses, doctors,
1057
+ EMR, and other medical information. Figure 5 demonstrates how the major components of
1058
+ the proposed RPM system have interacted during the simulation study.
1059
+ Table 3: RESTful API for proposed Medical Blockchain
1060
+ Action
1061
+ Verb
1062
+ Media Type
1063
+ URI
1064
+ Patient Dashboard
1065
+ ALL
1066
+ Application/json
1067
+ /api/Patient
1068
+ Doctor Dashboard
1069
+ ALL
1070
+ Application/json
1071
+ /api/Doctor
1072
+ Nurse Dashboard
1073
+ ALL
1074
+ Application/json
1075
+ /api/Nurse
1076
+ Healthcare Sensor Dashboard
1077
+ ALL
1078
+ Application/json
1079
+ /api/Sensor
1080
+ Vital˙Sign
1081
+ Application/json
1082
+ /api/VitalSignReading
1083
+ EMR Dashboard
1084
+ ALL
1085
+ Application/json
1086
+ /api/PatientRecord
1087
+ Share patient record with healthcare personnel
1088
+ POST
1089
+ Application/json
1090
+ /api/ShareRecord
1091
+ Blockchain Network Text
1092
+ GET
1093
+ Application/json
1094
+ /api/system/ping
1095
+ Issue identity to system user
1096
+ POST
1097
+ Application/json
1098
+ /api/SystemIdentities/issue
1099
+ Get Identities
1100
+ GET
1101
+ Application/json
1102
+ /api/System/identities
1103
+ Retrieve historian records
1104
+ GET
1105
+ Application/json
1106
+ /api/System/historian
1107
+ Within our blockchain implementation, each piece of medical record has one user (owner)
1108
+ who can share the data they own with other users (doctors) at varying levels of access. Data
1109
+ sharing between users is modeled by a system where users can share data with other users
1110
+ in different groups, as well as receive data requests from other users at any access level.
1111
+ If a user responds to a request by granting data access, an access token is provided to the
1112
+ receiver in a way that allows that receiver to access the data at the specified access level only.
1113
+ Our system ensures that sensitive information is never exposed on the blockchain, including
1114
+ 20
1115
+
1116
+ Figure 5: Sequence of interactions conducted during simulation
1117
+ both private and document keys, which is necessary in order to maintain the privacy and
1118
+ security of user-controlled data.
1119
+ We evaluate the performance of the proposed blockchain model using Hyperledger Caliper
1120
+ [35]. For experimental analysis, we carried out several experiments in terms of the execution
1121
+ time when adding a new healthcare device and executing a healthcare data query. We also
1122
+ measure the average time of the proposed consensus algorithm. The execution time is the
1123
+ round-trip time which includes the total time of sending the request by the client and getting
1124
+ the response from the network. In order to evaluate the execution time, we utilized the Post-
1125
+ man tool, which is used to explore and test the RESTful APIs by simulating a customized
1126
+ user load within the network. In this study, we created three groups of devices: 150, 300,
1127
+ and 500, in order to investigate the execution time of registering a device in the proposed
1128
+ blockchain model. Furthermore, the execution time is analyzed using different statistical
1129
+ 21
1130
+
1131
+ Doctor/Nurse
1132
+ Gateway/Fog Node
1133
+ REST Server
1134
+ Patient
1135
+ (Sensor)
1136
+ (GUI)
1137
+ (Local Chain)
1138
+ (Global Chain)
1139
+ 1
1140
+ Device Registration
1141
+ Device Registation
1142
+ smart
1143
+ contract
1144
+ Device ldentity and Certificate
1145
+ call
1146
+ Vital Sign (ECG Signal)
1147
+ create
1148
+ block
1149
+ New Block
1150
+ block
1151
+ mining
1152
+ New Block Broadcast
1153
+ save block
1154
+ or block
1155
+ New Block Broadcast
1156
+ header
1157
+ save block
1158
+ or block
1159
+ header
1160
+ Data Request
1161
+ check
1162
+ data
1163
+ location
1164
+ alt
1165
+ certify
1166
+ sender's
1167
+ [data in local chain]
1168
+ Secure Health Data
1169
+ certificate
1170
+ [data in global chain]
1171
+ Data Request
1172
+ certify
1173
+ sender's
1174
+ certificate
1175
+ Secure Health Datameasures, such as the minimum, maximum, and average times. As shown in Figure 6, in
1176
+ the case of 150 users, the average, minimum, and maximum execution time to register the
1177
+ healthcare device is recorded as 2335 ms, 2257 ms, and 2795 ms, respectively. Likewise,
1178
+ the minimum, maximum, and average execution times for 300 healthcare device-group is
1179
+ are 1785 ms, 3204 ms, and 2454 ms, respectively. Finally, for 500 devices the minimum
1180
+ execution time is recorded as 2810 ms, whereas the maximum and average execution time
1181
+ is 3524 ms and 3015 ms respectively (Figure 6).
1182
+ Figure 6: Healthcare device registration execution time
1183
+ The execution time of the proposed system is also evaluated in the case of retrieving
1184
+ healthcare data from the blockchain network.
1185
+ Every healthcare device in the proposed
1186
+ platform has the HTTP client functionality which is used to send requests for vital sign
1187
+ sensing data through the IoT gateway. The request is initially processed by the IoT gateway.
1188
+ If the requested data is found in the local chain, the IoT gateway validates the device
1189
+ certificate via the local smart contract and then replies to the device with the encrypted
1190
+ data. Else, the IoT gateway forwards the request to the REST server, which performs a
1191
+ similar process. The execution time of reading the vital sign data is illustrated in Figure 7.
1192
+ The same set of device groups has been considered for the experimental evaluation, i.e., 150,
1193
+ 300, and 500 devices. It is observed from the graph that the increase in the device scale in
1194
+ the proposed healthcare system will also create an impact on the execution time. However,
1195
+ the overall execution time of the network remains stable until there is high congestion in the
1196
+ network. The average execution time of vital sign sensing data in the case of 150, 300, and
1197
+ 22
1198
+
1199
+ 4000
1200
+ ■150 Devices
1201
+ 300Devices
1202
+ 500 Devices
1203
+ 3500
1204
+ 3000
1205
+ Execution Time (ms)
1206
+ 2500
1207
+ 2000
1208
+ 1500
1209
+ 1000
1210
+ 500
1211
+ 0
1212
+ Minimum
1213
+ Average
1214
+ Maximum500 devices are 2552 ms, 2525 ms, and 2775 ms, respectively, which are comparable to the
1215
+ execution times of registering a device that is shown in Figure 6.
1216
+ Figure 7: Vital signs reading execution time
1217
+ In order to evaluate the effectiveness of the proposed consensus method, we tested several
1218
+ scenarios in which we deployed five REST servers and five IoT gateways. The IoT devices
1219
+ were distributed evenly among the gateways, and each gateway was connected to a REST
1220
+ server. The servers saved all the blocks that were confirmed by the consensus protocol,
1221
+ while the IoT gateways saved the blocks of the devices that connected to them only. In
1222
+ these scenarios, we measure the consensus time of each created block, then we calculate the
1223
+ minimum, maximum, and average values for all the created blocks. The results are shown
1224
+ in Figure 8. We notice that the consensus time generally increases as the number of devices
1225
+ increases, which is logical since, with more devices, the total number of transactions increase,
1226
+ which adds more time to validate the new blocks. However, the increase in the consensus
1227
+ time is only 12.5 ms (on average) as the number of devices increases from 150 to 500, which
1228
+ proves the efficiency of the proposed consensus approach. In addition, the average consensus
1229
+ time of the system is 140 ms. In case of Ethereum and Bitcoin, it requires 10 to 19 seconds
1230
+ and 10 minutes to an hour respectively to mine a new block. Hence, the proposed consensus
1231
+ algorithm outperforms those of other blockchain platforms in terms of consensus time.
1232
+ 23
1233
+
1234
+ 4000
1235
+ 150Devices
1236
+ 300Devices
1237
+ 500Devices
1238
+ 3500
1239
+ 3000
1240
+ Execution Time (ms)
1241
+ 2500
1242
+ 2000
1243
+ 1500
1244
+ 1000
1245
+ 500
1246
+ 0
1247
+ Minimum
1248
+ Average
1249
+ MaximumFigure 8: Block consensus time
1250
+ 6.2. Efficiency of the Fog computing infrastructure
1251
+ Figure 9 depicts the distributed data flow model for our proposed IoT-driven critical
1252
+ healthcare applications. According to this model, data signals generated by the IoT devices
1253
+ are pushed into the client module, an initial application interface for interacting with the IoT
1254
+ devices and actuators and receiving the user’s information, such as name, location, address,
1255
+ sex, and age of the patient.
1256
+ After pre-processing and filtering the data that is coming
1257
+ from the IoT devices, the client module forwards the data to the Data Processing module
1258
+ for further processing. Here, AI-enabled modules can execute data analytics processes for
1259
+ testing purposes. Based on the outcome of the data processing, a command is issued by
1260
+ the Data Processing module for the client module so that it can trigger physical emergency
1261
+ actions through the actuators. Next, the Data Processing module dispatches the processed
1262
+ data to the aggregator module, which simultaneously interacts with the blockchain module
1263
+ at the IoT gateway and cloud server to add the data to the blockchain and ensure data
1264
+ integrity and location-independent data access. The blockchain module interacts with the
1265
+ storage module in case the data is to be stored off-chain. Finally, The Data Processing
1266
+ module at the cloud server interacts with the blockchain module to consistently produce
1267
+ the results that are requested by the application users. Since the client module directly
1268
+ interacts with the IoT devices, it is preferable to be deployed at the IoT gateways (e.g.,
1269
+ ECG machines). For the deployment of other modules, there exist different approaches in
1270
+ the literature. For instance, cloud computation has been exploited in [36] [37] to execute the
1271
+ data analytics, aggregator, blockchain, storage, and training module. On the other hand,
1272
+ the proposed RPM system adopts Fog computing for executing these modules and utilizes
1273
+ the cloud to host the blockchain, storage, and processing modules.
1274
+ 24
1275
+
1276
+ 200
1277
+ 150 Devices
1278
+ 300Devices500Devices
1279
+ 175
1280
+ 150
1281
+ (ms)
1282
+ Time (
1283
+ 125
1284
+ Consensus
1285
+ 100
1286
+ 75
1287
+ 50
1288
+ 25
1289
+ 0
1290
+ Minimum
1291
+ Average
1292
+ MaximumFigure 9: Data flow model for the proposed RPM system
1293
+ In this phase of performance evaluation, we demonstrate how the augmentation of Fog
1294
+ computing in remote patient monitoring improves the service latency and the energy usage
1295
+ in comparison to harnessing cloud-based resources. The experiments are conducted in an
1296
+ iFogSim [38] simulated Fog-Cloud computing environment. The computing resources within
1297
+ the simulation environment are organized in a hierarchical order, as shown in Figure 10. At
1298
+ the lower level of the simulation environment, twenty-four ECG machines (EMs) equipped
1299
+ with ECG sensors and emergency alert systems are placed. Based on the simulation design,
1300
+ an EM can connect with any of the four Fog local servers (FLSs) at the upper level. All
1301
+ FLSs are also set connected with a Fog regional server (FRS) that helps the lower-level
1302
+ computing devices to maintain seamless communication with the Cloud datacenter. Table
1303
+ 4 presents the details of the simulation parameters used in the experiments. The numerical
1304
+ values have been extracted from real-world references as specified in [39] [40]. Additionally,
1305
+ Table 5 illustrates the configuration of different application modules for the simulations,
1306
+ Figure 10: Architecture of the simulated Fog-Cloud computing environment
1307
+ 25
1308
+
1309
+ Cloud
1310
+ Server
1311
+ FRS#1
1312
+ FLS#1
1313
+ FLS#6
1314
+ EM#1
1315
+ EM#2
1316
+ EM#3
1317
+ EM#4
1318
+ EM#21
1319
+ EM#22
1320
+ EM#23
1321
+ EM#24
1322
+ 883
1323
+ 89loT layer
1324
+ Fog layer
1325
+ Cloud layer
1326
+ Data
1327
+ 600
1328
+ Global
1329
+ Signal
1330
+ Processed
1331
+ Aggregator
1332
+ Records
1333
+ Blockchain
1334
+ Data
1335
+ Module
1336
+ ECG
1337
+ Raw Data
1338
+ Module
1339
+ Sensor
1340
+ Data
1341
+ Client
1342
+ Processing
1343
+ Data
1344
+ Retrieve
1345
+ Response
1346
+ Module
1347
+ Records
1348
+ Query
1349
+ Module
1350
+ Command
1351
+ Save
1352
+ Local
1353
+ Analytic
1354
+ Alert
1355
+ Storage
1356
+ Blockchain
1357
+ Training
1358
+ Module
1359
+ Block Metadata
1360
+ Module
1361
+ Emergency
1362
+ Module
1363
+ IndicatorTable 4: Parameters of simulated environment
1364
+ Device configuration
1365
+ Name
1366
+ Processing
1367
+ speed
1368
+ Downlink
1369
+ bandwidth
1370
+ Uplink
1371
+ bandwidth
1372
+ Memory
1373
+ capacity
1374
+ Busy
1375
+ power
1376
+ Idle
1377
+ power
1378
+ (in MIPS)
1379
+ (in MB)
1380
+ (in MB)
1381
+ (in GB)
1382
+ (in MWh)
1383
+ (in MWh)
1384
+ EM
1385
+ 1000
1386
+ 10
1387
+ 5
1388
+ 8
1389
+ 1.1
1390
+ 0.2
1391
+ FLS
1392
+ 7000
1393
+ 8
1394
+ 3
1395
+ 12
1396
+ 1.3
1397
+ 0.4
1398
+ FRS
1399
+ 15000
1400
+ 6
1401
+ 2
1402
+ 16
1403
+ 1.6
1404
+ 0.8
1405
+ Cloud
1406
+ 40000
1407
+ 3
1408
+ 4
1409
+ 32
1410
+ 3.2
1411
+ 1.4
1412
+ Sensing frequency of ECG sensors
1413
+ 5 signals per second
1414
+ Simulation time
1415
+ 500 seconds
1416
+ Table 5: Module configuration
1417
+ Name
1418
+ Program size
1419
+ Packet size
1420
+ RAM usage
1421
+ (in MB)
1422
+ (in KB)
1423
+ (in GB)
1424
+ Client module
1425
+ 2000
1426
+ 500
1427
+ 1
1428
+ Data analytic module
1429
+ 4000
1430
+ 1500
1431
+ 6
1432
+ Aggregator module
1433
+ 1500
1434
+ 1800
1435
+ 2
1436
+ Blockchain module (periodic)
1437
+ 1000
1438
+ 2000
1439
+ 4
1440
+ Storage module
1441
+ 1000
1442
+ 2000
1443
+ 2
1444
+ Analytic training module
1445
+ 8000
1446
+ 2000
1447
+ 12
1448
+ Figure 11: Performance in reducing sense-process-actuation delay
1449
+ which have been approximated based on the profiled run-time, resource utilization, and
1450
+ data communication delay of the proposed solutions in heterogeneous computing devices
1451
+ and networking context.
1452
+ The results of the simulation experiments conducted in the aforementioned computing
1453
+ 26
1454
+
1455
+ 350
1456
+ ms)
1457
+ 300 -
1458
+ Sense-Process-Actuation delay (in
1459
+ 250
1460
+ 200
1461
+ 150
1462
+ 100.
1463
+ S
1464
+ Proposed RPMS
1465
+ Cloud-based RPMSFigure 12: Performance in reducing energy consumption
1466
+ Figure 13: Performance in blockchain transaction retrieval
1467
+ setup demonstrate that our proposed Fog computing-based RPMS outperforms the Cloud
1468
+ computing-based RPMS both in terms of reducing sense-process-actuation delay (calculated
1469
+ using iFogSim AppLoop model on ECG sensors → client module → data analytic module →
1470
+ client module → emergency alert system data flow) and energy usage. Figure 11 indicates
1471
+ that the augmentation of Fog computing can improve the responsiveness of RPMS by 40%
1472
+ in initiating alert messages during emergency situations compared to its cloud counterpart.
1473
+ Such performance improvement happens mainly for executing the data analytics module
1474
+ closer to the sources, that consequently decreases the data transfer delay to remote cloud
1475
+ servers. Moreover, the computing devices in the Fog paradigm consume a reduced amount
1476
+ of energy than a cloud server because of their capacity constraints. Statistically, this feature
1477
+ 27
1478
+
1479
+ 0.40
1480
+ 0.35 -
1481
+ 0.30-
1482
+ 0.25
1483
+ 0.20
1484
+ 0.15
1485
+ 0.10.
1486
+ 0.05 -
1487
+ 0.00
1488
+ Proposed RPMS
1489
+ Cloud-based RPMS25
1490
+ ProposedRPMS
1491
+ Transaction Retrieval
1492
+ WCloud-basedRPMs
1493
+ 20
1494
+ ime (ms)
1495
+ 15
1496
+ 3lockchain
1497
+ B
1498
+ 0
1499
+ Transaction size = 500 KB
1500
+ Transactionsize=2000KBalso has an influence in lowering the idle energy consumption of Fog computing devices.
1501
+ Therefore, when the time-based energy consumption model (as programmed in the iFogSim
1502
+ simulator) is applied, the Fog computing-based RPMS promises to deliver its services by
1503
+ consuming around 36% less energy than its Cloud-based implementation (as shown in Figure
1504
+ 12).
1505
+ On the other hand, due to executing the blockchain module at the fog devices, the delay
1506
+ required to retrieve a random blockchain transaction decreases as compared to the cloud-
1507
+ based RPMS, as shown in Figure 13. The figure illustrates that when the transaction size
1508
+ is equal to 500 KB, the proposed system requires an average of 7.16 ms to retrieve the
1509
+ transaction from the blockchain, while cloud-based RPMS needs 16.54 ms. On the other
1510
+ hand, for a 2000 KB transaction, the proposed RPMS produces a delay equal to 8.9 ms,
1511
+ while the cloud-based RPMS needs 19.34 ms.
1512
+ Hence, the proposed RPMS reduces the
1513
+ transaction retrieval delay by an average of 55.1%. This is mainly due to the cases in which
1514
+ the transaction is fetched from the local chain, which require much less end-to-end delay
1515
+ than retrieving the transaction from the global chain, due to the deployment of fog nodes
1516
+ at locations that are much nearer to the sensor nodes than the cloud servers.
1517
+ 7. Security Analysis
1518
+ Having a robust architecture encryption scheme as part of a blockchain-based data-
1519
+ sharing system is particularly critical from a security perspective because most blockchain
1520
+ implementations replicate the entire transaction ledger onto each node, therefore, multiply-
1521
+ ing the potential attack surface by the number of nodes in the network. In the following,
1522
+ we discuss the security analysis which we performed on the proposed patient monitoring
1523
+ system.
1524
+ • Key attack: Elliptic curve encryption method is employed from a key pair, and an
1525
+ attacker can’t calculate the private key to address the elliptic curve logarithm problem;
1526
+ hence the security of the proposed model is ensured. Moreover, for each session, a
1527
+ temporary private key is generated for interaction among the nodes. In such a way, if
1528
+ a private key gets compromised in terms of leakage, then this will not have an impact
1529
+ on the session, as the attacker would not be able to calculate a session key for a session
1530
+ that is currently going on among the nodes; and (b) the leaked private key is of no use
1531
+ until the session is completed.
1532
+ • Replay attack: The proposed model uses an individual temporary private key that
1533
+ is different for each session agreement among the interacting nodes. It is improbable
1534
+ that a replay attack becomes successful since private keys hold a bounded lifetime.
1535
+ • Impersonation attack: This attack is executed only if the attacker has successfully
1536
+ obtained the private key. The proposed model employs an individual private key and
1537
+ elliptic curve encryption. Therefore, this attack cannot be executed.
1538
+ 28
1539
+
1540
+ • Sybil attack: there are different methods to remove the impact of Sybil attack on
1541
+ the proposed model, such as increasing the price to form a new identity. This method
1542
+ restricts attackers from obtaining fake identities, using a two-factor authentication
1543
+ mechanism and accumulating the MAC and IP addresses of the participants, which
1544
+ permits the detection of those participants who have varying identities.
1545
+ • False data injection attack: Prior to validating the records, the consensus algorithm
1546
+ is executed by the blockchain nodes. On arrival of the positive consensus, a node can
1547
+ confirm the legitimacy of the received record.
1548
+ • Tampering attack: For encryption and signing the transaction, a public key crypto-
1549
+ system is employed. This indicates that the tampering node cannot amend the transac-
1550
+ tion as it does not hold the private key of the signing node. Furthermore, the proposed
1551
+ model can handle the key attacks; therefore, the adversaries cannot exploit the private
1552
+ keys.
1553
+ • Modification attack: As explained above, this attack is impossible because the
1554
+ adversaries cannot exploit the private keys.
1555
+ • Hiding blocks attack: A record in the proposed vital sign monitoring platform
1556
+ holds a unique sequence number. It is a must for a blockchain node to provide its
1557
+ saved records if requested. If a node in the network does not offer its records, it is
1558
+ detached from the network and disallowed to interact with other nodes.
1559
+ • Man-in-the-middle attack: A mutual authentication is performed between the
1560
+ nodes in the proposed model, which employs private keys for each session agreement,
1561
+ therefore, man-in-the-middle attacks are prevented.
1562
+ • Compromisation attack: If an attacker compromises a cloud server and attempts
1563
+ to sabotage the consensus operation by sending a ”Block Add” message that contains
1564
+ an invalid block, the legitimate cloud servers will detect the attack from the invalid
1565
+ signatures in the ”Block Add” message, since the attacker will not be able to generate
1566
+ the valid signatures of the other cloud servers. If the attacker drops the block that
1567
+ it receives from the IoT gateway, the latter reports the attack to the IoT ecosystem
1568
+ administrator when it detects that its block was not added to the blockchain in due
1569
+ time. Finally, if the attacker sends a wrong reply message when it receives a new block
1570
+ from another cloud server, the attack will not have an effect as long as the number of
1571
+ legitimate cloud servers is greater than N /2.
1572
+ 8. Conclusion and Future Work
1573
+ In this work, we have presented a three-layer remote patient monitoring system that
1574
+ leverages blockchain technology for better security and Fog technology for providing low-
1575
+ latency services to IoT devices and healthcare users. The most important functions that
1576
+ encompass the system components are described and evaluated. In addition, a new consensus
1577
+ 29
1578
+
1579
+ protocol that is tailored to the RPM environment is discussed and analyzed. Moreover, the
1580
+ blockchain module was implemented and tested using Hyperledger Fabric Framework, and it
1581
+ achieved low execution and consensus delays. Moreover, the augmentation of Fog computing
1582
+ can improve the responsiveness of the remote patient monitoring system by 40%.
1583
+ Several future works are being studied to enhance the proposed system. For example,
1584
+ we are planning to perform the simulations using real healthcare datasets (such as that in
1585
+ [41]). In addition, we intend to add a prediction module at the cloud layer that can predict
1586
+ a heart disease problem before its occurrence. The module would analyze the patient’s data
1587
+ from the global blockchain over an extended period to enhance prediction accuracy. Another
1588
+ enhancement would be the integration of the proposed blockchain system with a body area
1589
+ network (BAN) framework that is used to collect patient medical data in an efficient manner.
1590
+ Such integration should be carefully designed in order to secure the BAN operations without
1591
+ adding significant overhead in terms of computation and energy consumption on the BAN
1592
+ nodes. A similar system was proposed in [42]. Hence, we aim to study the literature in order
1593
+ to adjust the proposed blockchain system to make it suitable for a BAN environment.
1594
+ Another important future work is to enhance the proposed fog layer by augmenting it
1595
+ with modern technological tools that will improve its performance. For example, federated
1596
+ learning can be used by fog nodes to filter and analyze the readings of IoT devices in order
1597
+ to provide more accurate results to healthcare providers. Another important aspect is to
1598
+ design the scheduling of IoT data on the fog layer using the blockchain. For this aspect, we
1599
+ intend to adopt a previous strategy that we proposed in [43] to guarantee that a fog node
1600
+ treats data from IoT nodes fairly and provides equal opportunities for IoT nodes to save
1601
+ their data in the blockchain.
1602
+ Finally, we will study the scalability of the proposed system and its ability to support a
1603
+ large number of IoT ecosystems. For this purpose, we will design a hierarchical clustering
1604
+ framework that distributes cloud servers, fog nodes, and IoT devices into clusters based on
1605
+ their geographic locations and the deployed healthcare application. Using clustering will
1606
+ allow us to reduce the delay overhead when the application contains a huge number of
1607
+ blockchain nodes. In such a system, it is possible to execute a blockchain query in parallel
1608
+ by distributing it over the cluster heads, which would result in a reduced end-to-end delay
1609
+ between the patient and the healthcare provider.
1610
+ References
1611
+ [1] A. M. Rahmani, T. N. Gia, B. Negash, A. Anzanpour, I. Azimi, M. Jiang, P. Liljeberg, Exploiting
1612
+ smart e-health gateways at the edge of healthcare internet-of-things: A fog computing approach, Future
1613
+ Generation Computer Systems 78 (2018) 641–658.
1614
+ [2] B. Zaabar, O. Cheikhrouhou, M. Ammi, A. I. Awad, M. Abid, Secure and privacy-aware blockchain-
1615
+ based remote patient monitoring system for internet of healthcare things, in: 2021 17th International
1616
+ Conference on Wireless and Mobile Computing, Networking and Communications (WiMob), IEEE,
1617
+ 2021, pp. 200–205.
1618
+ [3] O. Cheikhrouhou, R. Mahmud, R. Zouari, M. Ibrahim, A. Zaguia, T. N. Gia, One-dimensional cnn
1619
+ approach for ecg arrhythmia analysis in fog-cloud environments, IEEE Access 9 (2021) 103513–103523.
1620
+ [4] R. W. Ahmad, K. Salah, R. Jayaraman, I. Yaqoob, S. Ellahham, M. Omar, The role of blockchain tech-
1621
+ nology in telehealth and telemedicine, International journal of medical informatics 148 (2021) 104399.
1622
+ 30
1623
+
1624
+ [5] H.-N. Dai, Z. Zheng, Y. Zhang, Blockchain for internet of things: A survey, IEEE Internet of Things
1625
+ Journal 6 (5) (2019) 8076–8094.
1626
+ [6] K. Mershad, Proact: Parallel multi-miner proof of accumulated trust protocol for internet of drones,
1627
+ Vehicular Communications 36 (2022) 100495.
1628
+ [7] M. S. Hossain, G. Muhammad, Cloud-assisted industrial internet of things (iiot)–enabled framework
1629
+ for health monitoring, Computer Networks 101 (2016) 192–202.
1630
+ [8] I. Azimi, A. Anzanpour, A. M. Rahmani, T. Pahikkala, M. Levorato, P. Liljeberg, N. Dutt, HiCH:
1631
+ Hierarchical Fog-Assisted Computing Architecture for Healthcare IoT, ACM Transactions on Embedded
1632
+ Computing Systems (TECS) 16 (5s) (2017) 1–20.
1633
+ [9] J. Pag´an, M. Zapater, J. L. Ayala, Power transmission and workload balancing policies in ehealth
1634
+ mobile cloud computing scenarios, Future Generation Computer Systems 78 (2018) 587–601.
1635
+ [10] P. Pace, G. Aloi, R. Gravina, G. Caliciuri, G. Fortino, A. Liotta, An edge-based architecture to support
1636
+ efficient applications for healthcare industry 4.0, IEEE Transactions on Industrial Informatics 15 (1)
1637
+ (2018) 481–489.
1638
+ [11] A. Juneja, M. Marefat, Leveraging blockchain for retraining deep learning architecture in patient-
1639
+ specific arrhythmia classification, in: 2018 IEEE EMBS International Conference on Biomedical &
1640
+ Health Informatics (BHI), IEEE, 2018, pp. 393–397.
1641
+ [12] K. N. Griggs, O. Ossipova, C. P. Kohlios, A. N. Baccarini, E. A. Howson, T. Hayajneh, Healthcare
1642
+ blockchain system using smart contracts for secure automated remote patient monitoring, Journal of
1643
+ medical systems 42 (7) (2018) 130.
1644
+ [13] M. S. Ali, M. Vecchio, G. D. Putra, S. S. Kanhere, F. Antonelli, A decentralized peer-to-peer remote
1645
+ health monitoring system, Sensors 20 (6) (2020) 1656.
1646
+ [14] S. Tuli, N. Basumatary, S. S. Gill, M. Kahani, R. C. Arya, G. S. Wander, R. Buyya, Healthfog:
1647
+ An ensemble deep learning based smart healthcare system for automatic diagnosis of heart diseases
1648
+ in integrated iot and fog computing environments, Future Generation Computer Systems 104 (2020)
1649
+ 187–200.
1650
+ [15] B. Farahani, M. Barzegari, F. S. Aliee, K. A. Shaik, Towards collaborative intelligent iot ehealth: From
1651
+ device to fog, and cloud, Microprocessors and Microsystems 72 (2020) 102938.
1652
+ [16] M. Garc´ıa-Valls, C. Calva-Urrego, A. Garc´ıa-Fornes, Accelerating smart ehealth services execution at
1653
+ the fog computing infrastructure, Future Generation Computer Systems 108 (2020) 882–893.
1654
+ [17] S. Wang, H. Li, J. Chen, J. Wang, Y. Deng, Dag blockchain-based lightweight authentication and
1655
+ authorization scheme for iot devices, Journal of Information Security and Applications 66 (2022) 103134.
1656
+ [18] K. M. Hossein, M. E. Esmaeili, T. Dargahi, A. Khonsari, M. Conti, Bchealth: A novel blockchain-based
1657
+ privacy-preserving architecture for iot healthcare applications, Computer Communications 180 (2021)
1658
+ 31–47.
1659
+ [19] Q. Xie, F. Dong, X. Feng, Eclb: Edge-computing-based lightweight blockchain framework for mobile
1660
+ systems, Security and Communication Networks 2021.
1661
+ [20] H. Honar Pajooh, M. Rashid, F. Alam, S. Demidenko, Multi-layer blockchain-based security architec-
1662
+ ture for internet of things, Sensors 21 (3) (2021) 772.
1663
+ [21] Q. Yao, T. Li, C. Yan, Z. Deng, Accident responsibility identification model for internet of vehicles
1664
+ based on lightweight blockchain, Computational Intelligence.
1665
+ [22] K. Mershad, B. Said, A blockchain model for secure communications in internet of vehicles, in: 2020
1666
+ IEEE/ACS 17th International Conference on Computer Systems and Applications (AICCSA), IEEE,
1667
+ 2020, pp. 1–6.
1668
+ [23] A. R. Shahid, N. Pissinou, C. Staier, R. Kwan, Sensor-chain: a lightweight scalable blockchain frame-
1669
+ work for internet of things, in: 2019 International Conference on Internet of Things (iThings) and IEEE
1670
+ Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing
1671
+ (CPSCom) and IEEE Smart Data (SmartData), IEEE, 2019, pp. 1154–1161.
1672
+ [24] J. Sunny, S. Sankaran, V. Saraswat, Towards a lightweight blockchain platform for critical infrastructure
1673
+ protection, in: 2020 IEEE 40th International Conference on Distributed Computing Systems (ICDCS),
1674
+ IEEE, 2020, pp. 1287–1292.
1675
+ 31
1676
+
1677
+ [25] B. Zaabar, O. Cheikhrouhou, F. Jamil, M. Ammi, M. Abid, Healthblock: A secure blockchain-based
1678
+ healthcare data management system, Computer Networks 200 (2021) 108500.
1679
+ [26] K. Mershad, O. Cheikhrouhou, L. Ismail, Proof of accumulated trust: A new consensus protocol for
1680
+ the security of the iov, Vehicular Communications 32 (2021) 100392.
1681
+ [27] S. Rahimi Moosavi, T. Nguyen gia, E. Nigussie, A. M. Rahmani, S. Virtanen, H. Tenhunen, J. Isoaho,
1682
+ End-to-end security scheme for mobility enabled healthcare internet of things, Future Generation Com-
1683
+ puter Systems 64.
1684
+ [28] U. Khalid, M. Asim, T. Baker, P. Hung, M. A. Tariq, L. Rafferty, A decentralized lightweight blockchain-
1685
+ based authentication mechanism for iot systems, Cluster Computing 23.
1686
+ [29] Z. Shelby, C. Bormann, 6lowpan: The wireless embedded internet, 2009.
1687
+ [30] B. S. Negash, A. M. Rahmani, T. Westerlund, P. Liljeberg, H. Tenhunen, Lisa: Lightweight internet of
1688
+ things service bus architecture, Vol. 52, 2015. doi:10.1016/j.procs.2015.05.010.
1689
+ [31] S. Nunna, A. Kousaridas, M. Ibrahim, M. Dillinger, C. Thuemmler, H. Feussner, A. Schneider, En-
1690
+ abling real-time context-aware collaboration through 5g and mobile edge computing, in: 2015 12th
1691
+ International Conference on Information Technology - New Generations, 2015, pp. 601–605.
1692
+ [32] M. L. Hilton, Wavelet and wavelet packet compression of electrocardiograms, IEEE Transactions on
1693
+ Biomedical Engineering 44 (5) (1997) 394–402.
1694
+ [33] Zhitao Lu, Dong Youn Kim, W. A. Pearlman, Wavelet compression of ecg signals by the set partitioning
1695
+ in hierarchical trees algorithm, IEEE Transactions on Biomedical Engineering 47 (7) (2000) 849–856.
1696
+ [34] Composer,
1697
+ https://hyperledger.github.io/composer/latest/introduction/introduction.
1698
+ html, accessed: 2020-10-25.
1699
+ [35] Hyperledger caliper, https://www.hyperledger.org/use/caliper, accessed: 2020-10-25.
1700
+ [36] A. F. Hussein, M. Burbano-Fernandez, G. Ram´ırez-Gonz´alez, E. Abdulhay, V. H. C. De Albuquerque,
1701
+ et al., An automated remote cloud-based heart rate variability monitoring system, IEEE Access 6
1702
+ (2018) 77055–77064.
1703
+ [37] H. Kaur, M. A. Alam, R. Jameel, A. K. Mourya, V. Chang, A proposed solution and future direction
1704
+ for blockchain-based heterogeneous medicare data in cloud environment, Journal of medical systems
1705
+ 42 (8) (2018) 156.
1706
+ [38] R. Mahmud, R. Buyya, Modeling and Simulation of Fog and Edge Computing Environments Using
1707
+ iFogSim Toolkit, in: Fog and Edge Computing: Principles and Paradigms, John Wiley & Sons, Ltd,
1708
+ 2019, Ch. 17, pp. 433–465.
1709
+ [39] R. Mahmud, K. Ramamohanarao, R. Buyya, Edge affinity-based management of applications in fog
1710
+ computing environments, in: Proceedings of the 12th IEEE/ACM International Conference on Utility
1711
+ and Cloud Computing, UCC ’19, ACM, New York, NY, USA, 2019, pp. 1–10.
1712
+ [40] E. Ahvar, A. Orgerie, A. L´ebre, Estimating energy consumption of cloud, fog and edge computing
1713
+ infrastructures, IEEE Transactions on Sustainable Computing (2019) 1–1doi:10.1109/TSUSC.2019.
1714
+ 2905900.
1715
+ [41] S.
1716
+ Ulianova,
1717
+ Cardiovascular
1718
+ disease
1719
+ dataset,
1720
+ https://www.kaggle.com/datasets/sulianova/
1721
+ cardiovascular-disease-dataset, accessed: 2022-12-3.
1722
+ [42] Z. Shahbazi, Y.-C. Byun, Towards a secure thermal-energy aware routing protocol in wireless body
1723
+ area network based on blockchain technology, Sensors 20 (12) (2020) 3604.
1724
+ [43] K. Mershad, H. Artail, Score: Data scheduling at roadside units in vehicle ad hoc networks, in: 2012
1725
+ 19th International Conference on Telecommunications (ICT), IEEE, 2012, pp. 1–6.
1726
+ 32
1727
+
ANE1T4oBgHgl3EQf9AZG/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
AtE3T4oBgHgl3EQfTgrD/content/tmp_files/2301.04443v1.pdf.txt ADDED
@@ -0,0 +1,897 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Entangled States are Harder to Transfer than Product States
2
+ Tony J. G. Apollaro
3
+ ,1, ∗ Salvatore Lorenzo
4
+ ,2 Francesco
5
+ Plastina
6
+ ,3, 4 Mirko Consiglio
7
+ ,1 and Karol ˙Zyczkowski
8
+ 5, 6
9
+ 1Department of Physics, University of Malta, Msida MSD 2080, Malta
10
+ 2Universit`a degli Studi di Palermo, Dipartimento di Fisica e Chimica - Emilio Segr`e, via Archirafi 36, I-90123 Palermo, Italy
11
+ 3Dipartimento di Fisica, Universit`a della Calabria, 87036 Arcavacata di Rende (CS), Italy
12
+ 4INFN, gruppo collegato di Cosenza, 87036 Arcavacata di Rende (CS), Italy
13
+ 5Institute of Theoretical Physics, Jagiellonian University, ul. �Lojasiewicza 11, 30–348 Krak´ow, Poland
14
+ 6Center for Theoretical Physics, Polish Academy of Sciences, Al. Lotnik´ow 32/46, 02-668 Warszawa, Poland
15
+ The distribution of entangled states is a key task of utmost importance for many quantum infor-
16
+ mation processing protocols. A commonly adopted setup for distributing quantum states envisages
17
+ the creation of the state in one location, which is then sent to (possibly different) distant receivers
18
+ through some quantum channels. While it is undoubted and, perhaps, intuitively expected that the
19
+ distribution of entangled quantum states is less efficient than that of product states, a thorough
20
+ quantification of this inefficiency (namely, of the difference between the quantum-state transfer fi-
21
+ delity for entangled and factorized states) has not been performed. To this end, in this work, we
22
+ consider n-independent amplitude-damping channels, acting in parallel, i.e., each, locally, on one
23
+ part of an n-qubit state. We derive exact analytical results for the fidelity decrease, with respect to
24
+ the case of product states, in the presence of entanglement in the initial state, for up to four qubits.
25
+ Interestingly, we find that genuine multipartite entanglement has a more detrimental effect on the
26
+ fidelity than two-qubit entanglement. Our results hint at the fact that, for larger n-qubit states,
27
+ the difference in the average fidelity between product and entangled states increases with increasing
28
+ single-qubit fidelity, thus making the latter a less trustworthy figure of merit.
29
+ I.
30
+ INTRODUCTION
31
+ Distributing entangled states among several distant recipients is a task of paramount importance in a variety of
32
+ quantum-information processing protocols, ranging from n-party quantum key distribution [1] to distributed quantum
33
+ computing [2]. In many of these protocols, an n-partite entangled state is created at location S (the sender’s location),
34
+ and its parts are distributed among m ≤ n receivers, generally at different locations (which we dub R, the receivers’
35
+ location).
36
+ The special case of distributing a bipartite entangled state was already considered in the seminal paper by Bose on
37
+ quantum-state transfer (QST), where the transfer protocol is employed to send (the state of) one party of a two-qubit
38
+ Bell state to the opposite edge of a spin chain [3]. After this first instance, a considerable amount of research, both
39
+ theoretical and experimental, has been performed in order to improve the transfer performance and optimize the Bell
40
+ state distribution protocol [4–10]. Moreover, with the increasing exploration (and exploitation) of the fascinating
41
+ realm of quantum correlations by quantum technological applications, the distribution of n-partite entangled states,
42
+ with n > 2, has become a very active research topic [11–15].
43
+ At variance with the entanglement of two-qubits, which is the only system whose entanglement properties have been
44
+ fully characterized both for pure and mixed states, for n > 2 there are only a handful of closed, analytical results for
45
+ the quantification of entanglement [16], making the task of evaluating the efficiency of an entanglement distribution
46
+ protocol very difficult to assess.
47
+ Here, we address the distribution of an n-partite entangled state utilizing the fidelity between the sender’s and the
48
+ receivers’ state as a figure of merit for the quality of the protocol. Although the fidelity is not a bona fide tool to
49
+ characterize quantum resources [17, 18], it is nevertheless widely employed in constructing entanglement witnesses
50
+ following the idea that states close to an entangled state must be entangled as well [19]. Hence, building on recent
51
+ results [20, 21] reporting the fidelity of an n-qubits QST (n-QST) protocol for arbitrary quantum channels, we
52
+ investigate the effect of the presence of entanglement on the n-QST fidelity, which is evaluated when each qubit is
53
+ subject to an independent U(1)-symmetric quantum channel, e.g., an amplitude-damping channel. In particular, we
54
+ find that the presence of entanglement in the sender state is detrimental to the efficiency of the n-QST protocol. It
55
+ is not surprising that independent quantum channels acting on n qubits tend to destroy their quantum correlations,
56
+ thus lowering the transmission fidelity. We are able to provide a quantification of the fidelity reduction as a function
57
+ ∗ tony.apollaro@um.edu.mt
58
+ arXiv:2301.04443v1 [quant-ph] 11 Jan 2023
59
+
60
+ 2
61
+ of different entanglement monotones . In particular, we show that genuine multipartite entanglement, as quantified,
62
+ e.g., by the three-tangle, has a more pronounced effect on lowering the n-QST fidelity than bipartite entanglement
63
+ between two qubits, as quantified by the concurrence [22].
64
+ The paper is organised as follows: in Section II, we introduce our model and provide a brief recap of the n-QST
65
+ fidelity; in Section III, we apply the developed formalism to the case of n = 2, 3, 4 qubits; finally, in Section IV, we
66
+ draw our conclusions.
67
+ II.
68
+ N -QST FIDELITY FOR INDEPENDENT AMPLITUDE-DAMPING CHANNELS
69
+ Let us consider an n-qubit quantum-state transfer protocol as depicted in Figure 1. A sender, located at position
70
+ S, prepares an n-qubit arbitrary state and wants to transfer each party to different receivers to which the sender is
71
+ connected by different quantum channels. Without a loss of generality, let us assume the sender state to be a pure
72
+ state, ρS = |Ψ⟩⟨Ψ|n. The state at the receivers’ location reads
73
+ ρR(t) = [Φ1 ⊗ Φ2 ⊗ · · · ⊗ Φn] (t) (ρS) .
74
+ (1)
75
+ The fidelity between the sender and the receivers’ state is given by the Uhlmann–Jozsa fidelity [23]
76
+ F (|Ψ⟩ , ρ(t)) = ⟨Ψ| ρ |Ψ⟩ .
77
+ (2)
78
+ Expressing an arbitrary input state in the computational basis
79
+ |Ψ⟩ =
80
+ 2n
81
+
82
+ i=1
83
+ ai |i⟩ ,
84
+ (3)
85
+ the elements of the receivers’ density matrix read (sum over repeated indexes is assumed)
86
+ ρR
87
+ ij = Anm
88
+ ij ρS
89
+ nm
90
+ (4)
91
+ yielding the fidelity
92
+ F (|Ψ⟩ , ρ) =
93
+ d−1
94
+
95
+ ijnm=0
96
+ a∗
97
+ i ajana∗
98
+ mAnm
99
+ ij ,
100
+ (5)
101
+ where all of the amplitudes a refer to the initial state of the sender.
102
+ For the case represented in Figure 1, the total map is given by the tensor products of n independent maps as in
103
+ Equation (1). Hence, Equation (4) can be cast in the following form: [24]
104
+ ρR
105
+ i1i2...in;j1j2...jn = Ap1p2...pn;q1q2...qn
106
+ i1i2...in;j1j2...jn ρS
107
+ p1p2...pn;q1q2...qn,
108
+ (6)
109
+ where i, j, p, q = 0, 1 and the corresponding subscript refers to the i’s qubit, with
110
+ Ap1p2...pn;q1q2...qn
111
+ i1i2...in;j1j2...jn
112
+ = Ap1;q1
113
+ i1;j1 Ap2;q2
114
+ i2;j2 . . . Apn;qn
115
+ in;jn .
116
+ (7)
117
+ Each A in Equation (7) comes from a single qubit map connecting the sender qubit si and the receiver qubit ri,
118
+ which, for an U(1)-symmetric channel, can be expressed as
119
+
120
+
121
+
122
+ ρ00
123
+ ρ01
124
+ ρ10
125
+ ρ11
126
+
127
+
128
+
129
+ ri
130
+ =
131
+
132
+
133
+
134
+
135
+ 1
136
+ 0
137
+ 0
138
+ 1 −
139
+ ��f ri
140
+ si
141
+ ��2
142
+ 0 f ri
143
+ si
144
+ 0
145
+ 0
146
+ 0
147
+ 0
148
+
149
+ f ri
150
+ si
151
+ �∗
152
+ 0
153
+ 0
154
+ 0
155
+ 0
156
+ ��f ri
157
+ si
158
+ ��2
159
+
160
+
161
+
162
+
163
+
164
+
165
+
166
+ ρ00
167
+ ρ01
168
+ ρ10
169
+ ρ11
170
+
171
+
172
+
173
+ si
174
+ ,
175
+ (8)
176
+ where f ri
177
+ si is the transition amplitude for the excitation initially on si to reach ri. A widely used U(1)-symmetric
178
+ quantum channel is given by the so called XXZ spin- 1
179
+ 2 Hamiltonian,
180
+ H =
181
+
182
+ i,j
183
+ Jij
184
+
185
+ σx
186
+ i σx
187
+ j + σy
188
+ i σy
189
+ j
190
+
191
+ + ∆ijσz
192
+ i σz
193
+ j + hiσz
194
+ i
195
+ (9)
196
+
197
+ 3
198
+ FIG. 1. A quantum router. A dispatch center, encircled in red, creates an n-qubit entangled state (red spheres) with the aim
199
+ to send each party to a different receiver (green spheres) along independent quantum channels (blue spheres).
200
+ where σα
201
+ i (α = x, y, x) are Pauli matrices and i, j are the position indexes on an arbitrary d-dimensional lattice.
202
+ Assuming that each quantum channel is fully polarized, for the sender state , the map Φi reduces to an amplitude-
203
+ damping channel [25]. In particular, for f ri
204
+ si = 1, the map Φi entails a SWAP operation. Therefore, our formalism
205
+ also describes entanglement swapping protocols via imperfect operations [26]. Finally, to express Equation (6) in the
206
+ form of Equation (4), it is sufficient to express the bit strings in decimal notation.
207
+ By making use of Equations (4), (7), and (8), it is straightforward to evaluate the average fidelity [13] of an arbitrary
208
+ quantum state |Ψ⟩
209
+ ⟨F⟩ = 1
210
+
211
+
212
+
213
+ dΩ F (|Ψ⟩ , ρ(t)) ,
214
+ (10)
215
+ with Ω denoting the space of pure states and, with an abuse of notation, its volume and the measure of it .
216
+ An average with respect to an n qubit system will be denoted as ⟨Fn⟩. In the case of n independent channels,
217
+ making use of the transition amplitude f introduced in Equation (8), one arrives at the expression,
218
+ ⟨Fn⟩ =
219
+ 1
220
+ 2n + 1 +
221
+ 1
222
+ 2n (2n + 1) |1 + f|2n .
223
+ (11)
224
+ Notice that the average fidelity ⟨Fn⟩ ≤ �n
225
+ i=1 ⟨F1⟩, with equality holding only for f = 0, 1. While the left-hand side of
226
+ the latter inequality gives the average over all possible pure input states, its right-hand side, on the other hand, gives
227
+ the average restricted to fully factorized states only, i.e., to product states of the form |Ψ⟩n = �n
228
+ i=1 |ψ⟩i, thus not
229
+ including the entangled states. Hence, we conclude that, when n ≥ 2, in the set of all pure input states, entangled
230
+ states have a lower n-QST fidelity than the product state. In the next sections, we will provide a quantitative analysis
231
+ for this intuitive observation.
232
+ III.
233
+ N -QST FIDELITY AS A FUNCTION OF ENTANGLEMENT
234
+ This Section contains our main result, namely, that the presence of entanglement reduces the transfer fidelity.
235
+ Below, we illustrate this idea separately for two, three, and four qubit transmissions. In particular, we will show that,
236
+ in the presence of entanglement in the states to be sent, a reduction factor exists, which we dub Rn, such that the
237
+ average fidelity for the QST of n-qubits can be generically written
238
+ ⟨Fn⟩ = ⟨F1⟩n − En Rn.
239
+
240
+ 4
241
+ Here, ⟨F1⟩n gives the average fidelity for the transfer of factorized states (indeed, intuitively, qubits can be transferred
242
+ one by one, in this case), so that the difference ⟨Fn⟩ − ⟨F1⟩n is entirely due to the fact that entangled states are
243
+ possibly transferred. The coefficient En is an entanglement quantifier that changes with n. It is given by twice the
244
+ square of concurrence for n = 2, while for n = 3 it is proportional to a linear combination of the invariant polynomials
245
+ identifying the different classes of entangled states. Finally, the factor Rn (defined below for the various cases) gives
246
+ the weight of entanglement-induced fidelity decrease, and it also enters the fidelity averaged over specific entanglement
247
+ classes of states (for n = 3, 4).
248
+ A.
249
+ Two Qubits
250
+ Adopting the (Schmidt-)parametrization of two-qubits pure states in terms of their entanglement [27], we write
251
+ |Ψ(s)⟩ =
252
+
253
+ 1 + s
254
+ 2
255
+ |00⟩ +
256
+
257
+ 1 − s
258
+ 2
259
+ |11⟩
260
+ (12)
261
+ where the parameter s ∈ [−1, 1] is related to the concurrence [22] via C =
262
+
263
+ 1 − s2, and every two-qubit pure state
264
+ can be obtained from Equation (12) via local, unitary operations |Φ(s)⟩ = U1U2 |Ψ(s)⟩, with Ui ∈ SU(2) acting on
265
+ qubit i = 1, 2. Below, we obtain the average fidelity for a two-qubit QST protocol with independent channels as a
266
+ function of the amount of entanglement of the sender state, which is invariant under local unitaries, and which we
267
+ denote as ⟨·⟩U ⊗2
268
+ This average fidelity (which, to say it shortly, is averaged at fixed values of entanglement) reads
269
+ ⟨F2⟩U ⊗2 = 1
270
+ 36
271
+
272
+ 3 + |f|2 + 2 |f| cos φ
273
+ �2
274
+ − 1
275
+ 18
276
+
277
+ |f|2 + 2 |f| cos φ
278
+ � �
279
+ 3 − |f|2 + 2 |f| cos φ
280
+
281
+ C2
282
+ (13)
283
+ where we expressed the complex transition amplitude f as f = |f| eiφ.
284
+ Following the procedure outlined by Bose [3], in order to maximise Equation (13), one sets cos φ = 1, which,
285
+ physically, can be obtained, e.g., by a uniform magnetic field applied over the spin chain. Hence, the average two-
286
+ qubit fidelity F2 can be cast in the form
287
+ ⟨F2⟩U ⊗2 = 1
288
+ 36
289
+
290
+ 3 + |f|2 + 2 |f|
291
+ �2
292
+ − 1
293
+ 18
294
+
295
+ |f|2 + 2 |f|
296
+ � �
297
+ 3 −
298
+
299
+ |f|2 + 2 |f|
300
+ ��
301
+ C2.
302
+ (14)
303
+ From Equation (14), since 0 ≤ |f| ≤ 1, one can readily appreciate that the more concurrence the sender’s pure state
304
+ contains, the lower the fidelity with the received state. Equation (14) can also be rewritten as a function of the
305
+ single-qubit QST average fidelity
306
+ ⟨F1⟩ = 1
307
+ 6
308
+
309
+ 3 + 2 |f| + |f|2�
310
+ ,
311
+ (15)
312
+ to read
313
+ ⟨F2⟩U ⊗2 = ⟨F1⟩2 − 2
314
+
315
+ ⟨F1⟩ − 1
316
+ 2
317
+
318
+ (1 − ⟨F1⟩) C2 = ⟨F1⟩2 − 2R2C2.
319
+ (16)
320
+ Again, as 1
321
+ 2 ≤ ⟨F1⟩ ≤ 1, the average fidelity ⟨F2⟩ decreases with the amount of concurrence of the sender state.
322
+ From Equation (16), we see that the average 2-QST fidelity is reduced in the presence of the squared concurrence
323
+ by a factor of
324
+ R2 =
325
+
326
+ ⟨F1⟩ − 1
327
+ 2
328
+
329
+ (1 − ⟨F1⟩) ,
330
+ (17)
331
+ which is reported in Figure 2 (left panel), together with the two-qubit average fidelity ⟨F2⟩, displayed for different
332
+ values of the squared concurrence as a function of the one-qubit fidelity ⟨F1⟩ (right panel).
333
+ B.
334
+ Three Qubits
335
+ Having obtained a quantitative expression giving the reduction in the transmission fidelity due to the presence
336
+ of entanglement for two qubits, we move to the more intricate three qubit case in order to try and obtain similar
337
+ relations.
338
+
339
+ 5
340
+ 0.6
341
+ 0.7
342
+ 0.8
343
+ 0.9
344
+ 1.0
345
+ 0.01
346
+ 0.02
347
+ 0.03
348
+ 0.04
349
+ 0.05
350
+ 0.06
351
+ 0.6
352
+ 0.7
353
+ 0.8
354
+ 0.9
355
+ 1.0
356
+ 0.4
357
+ 0.6
358
+ 0.8
359
+ 1.0
360
+ FIG. 2. (left) Reduction factor (17) for entangled states of the 2-QST average fidelity ⟨F2⟩ (16) as a function of the 1-QST
361
+ average fidelity ⟨F1⟩. The dotted, vertical line reports the maximum of R2 attained at ⟨F⟩1 = 0.75. (left)
362
+ 1.
363
+ Three-qubit pure-state entanglement
364
+ Let us now consider a system of three qubits A, B, and C. A three-qubit pure state can be written in canonical
365
+ form as [28]
366
+ |Ψ⟩ABC = λ0 |000⟩ + λ1eiφ |100⟩ + λ2 |101⟩ + λ3 |110⟩ + λ4 |111⟩ ,
367
+ (18)
368
+ where λi ≥ 0, 0 ≤ φ ≤ π, and the normalisation condition reads �
369
+ i λ2
370
+ i = 1.
371
+ In terms of the coefficients of the state in Equation (18), one can introduce five invariant polynomials, allowing to
372
+ identify different entanglement classes [29]:
373
+ J1 =
374
+ ��λ1λ4eiφ − λ2λ3
375
+ ��2 , J2 = λ2
376
+ 0λ2
377
+ 2 , J3 = λ2
378
+ 0λ2
379
+ 3
380
+ (19a)
381
+ J4 = λ2
382
+ 0λ2
383
+ 4 , J5 = λ2
384
+ 0
385
+
386
+ J1 + λ2
387
+ 2λ2
388
+ 3 − λ2
389
+ 1λ2
390
+ 4
391
+
392
+ .
393
+ (19b)
394
+ The relation between invariant polynomials and entanglement measures is given by
395
+ C2
396
+ jk = 4Ji,
397
+ (20)
398
+ for i ̸= j ̸= k = 1, 2, 3, and where now, (1, 2, 3) = (A, B, C) holds on the LHS of Equation (20) . At variance with the
399
+ two-qubit case, no single entanglement measure can capture genuine three-partite entanglement as three qubits can
400
+ be entangled in two inequivalent ways [30].
401
+ One type of entanglement is quantified by the three-tangle [31],
402
+ τ 2
403
+ 3 = 4J4 ,
404
+ (21)
405
+ while an inequivalent type of genuine multipartite entanglement (GME) is quantified by the so called GME concur-
406
+ rence, CGME [32], defined, in terms of invariant polynomials, for a three-qubit pure state as:
407
+ CGME = 4 (min {J2 + J3, J1 + J3, J1 + J2} + J4) .
408
+ (22)
409
+ Hence, three qubit states can be sorted in the following entanglement classes [30]:
410
+ • Product states. All Ji = 0, resulting into A − B − C (class 1). All entanglement measures vanish.
411
+ • Biseparable states. All Ji = 0 except i) J1 for A − BC, ii) J2 for B − AC, and iii) J3 for C − AB (class 2a).
412
+ Only the concurrence for one single pair of qubits is different from zero.
413
+ • W-states. CGME > 0 and τ3 = 0
414
+ 1. J4 = 0 and J1J2 + J1J3 + J2J3 = √J1J2J3 = J5
415
+ 2 (class 3a).
416
+ 2. J4 = 0 and √J1J2J3 = J5
417
+ 2 (class 4a).
418
+
419
+ 6
420
+ • GHZ-states. CGME > 0 and τ3 > 0, with 5 possible cases:
421
+ 1. All Ji = 0 except J4 (class 2b).
422
+ 2. J1 = J2 = J5 = 0, or J1 = J3 = J5 = 0 or J2 = J3 = J5 = 0 (class 3b).
423
+ 3. J2 = J5 = 0 or J3 = J5 = 0 (class 4b).
424
+ 4. J1J4 + J1J2 + J1J3 = √J1J2J3 = J5
425
+ 2 (class 4c).
426
+ 5. √J1J2J3 = |J5|
427
+ 2
428
+ and (J4 + J5)2 − 4 (J1 + J4) (J2 + J4) (J3 + J4) = 0 (class 4d) ,
429
+ where, with the notation X − Y , we indicate that subsystems X and Y do not share any type of entanglement.
430
+ Notably, for 3-qubit pure states, a monogamy relation exists between the amount of entanglement that can be
431
+ shared among the parties [31]
432
+ C2
433
+ A|BC = C2
434
+ AB + C2
435
+ AC + τ 2
436
+ 3 .
437
+ (23)
438
+ 2.
439
+ Fidelity of 3-QST
440
+ Here, we derive the fidelity of the QST of a tree-qubit pure state. First, we average the state in Equation (18) over
441
+ single-qubit, local operations U, and use the notation ⟨·⟩U ⊗3, in order to indicate the average fidelity at given values
442
+ of (and, thus, as a function of) the entanglement quantifiers. Subsequently, utilizing the averages of the invariant
443
+ polynomials obtained in Ref. [29], we derive the average fidelity of each three-qubit class and use the notation ⟨·⟩.
444
+ The 3-QST fidelity, expressed in terms of the single-qubit average, reads
445
+ ⟨F(|Ψ⟩ , ρ0)⟩U ⊗3 = ⟨F1⟩3 − 8 ⟨F1⟩
446
+
447
+ ⟨F1⟩ − 1
448
+ 2
449
+
450
+ (1 − ⟨F1⟩)
451
+
452
+ J1 + J2 + J3 + 3
453
+ 2J4
454
+
455
+ .
456
+ (24)
457
+ Since 1
458
+ 2 ≤ ⟨F1⟩ ≤ 1, and 0 ≤ Ji ≤ 1
459
+ 4, the second term on the right hand side of the latter equation is always
460
+ negative, so that one sees at once that entanglement reduces the 3-QST fidelity.
461
+ Introducing the reduction factor R3,
462
+ R3 = ⟨F1⟩
463
+
464
+ ⟨F1⟩ − 1
465
+ 2
466
+
467
+ (1 − ⟨F1⟩) ,
468
+ (25)
469
+ we can write
470
+ ⟨F(|Ψ⟩ , ρ0)⟩U ⊗3 = ⟨F1⟩3 − 8 R3
471
+
472
+ J1 + J2 + J3 + 3
473
+ 2J4
474
+
475
+ .
476
+ (26)
477
+ Now, using the averages of the invariant polynomials obtained in Ref. [29], ⟨J4⟩ =
478
+ 1
479
+ 12 and ⟨Jk⟩ =
480
+ 1
481
+ 24 (k = 1, 2, 3),
482
+ the average fidelity ⟨F3⟩ (with average taken over the full three qubit Hilbert space) is given by
483
+ ⟨F3⟩ = ⟨F1⟩3 − 2R3
484
+ (27)
485
+ Here, we see that, at fixed single-particle average fidelity, entanglement is responsible for a decrease in the 3-QST
486
+ average fidelity by twice the reduction factor R3.
487
+ In particular, the fidelities for the canonical states belonging to the different entanglement classes read
488
+ • class 1 (product state)
489
+ ⟨Fc1⟩U ⊗3 = ⟨F1⟩3
490
+ (28)
491
+ ⟨Fc1⟩ = ⟨F1⟩3
492
+ (29)
493
+ • class 2a (biseparable states)
494
+ ⟨Fc2a⟩U ⊗3 = ⟨F1⟩3 − 2 ⟨F1⟩
495
+
496
+ ⟨F1⟩ − 1
497
+ 2
498
+
499
+ (1 − ⟨F1⟩) C2
500
+ jk
501
+ (30)
502
+ ⟨Fc2a⟩ = ⟨F1⟩3 − R3
503
+ 3
504
+ (31)
505
+ where i ̸= j ̸= k = 1, 2, 3;
506
+
507
+ 7
508
+ • class 2b (GHZ-states)
509
+ ⟨Fc2b⟩U ⊗3 = ⟨F1⟩3 − 3 ⟨F1⟩
510
+
511
+ ⟨F1⟩ − 1
512
+ 2
513
+
514
+ (1 − ⟨F1⟩) τ 2
515
+ 3
516
+ (32)
517
+ ⟨Fc2b⟩ = ⟨F1⟩3 − R3
518
+ (33)
519
+ • class 3a (J4 = 0 and J1J2 + J1J3 + J2J3 = √J1J2J3 = J5
520
+ 2 )
521
+ ⟨Fc3a⟩U ⊗3 = ⟨F1⟩3 − 2 ⟨F1⟩
522
+
523
+ ⟨F1⟩ − 1
524
+ 2
525
+
526
+ (1 − ⟨F1⟩)
527
+
528
+ C2
529
+ BC + C2
530
+ AC + C2
531
+ AB
532
+
533
+ ⟨Fc3a⟩ = ⟨F1⟩3 − R3
534
+ (34)
535
+ • Class 3b (J1 = J2 = J5 = 0 or J1 = J3 = J5 = 0 or J2 = J3 = J5 = 0)
536
+ ⟨Fc3b⟩U ⊗3 = ⟨F1⟩3 − ⟨F1⟩
537
+
538
+ ⟨F1⟩ − 1
539
+ 2
540
+
541
+ (1 − ⟨F1⟩)
542
+
543
+ 2C2
544
+ BC + 3τ 2
545
+ 3
546
+
547
+ ⟨Fc3b⟩ = ⟨F1⟩3 − 4
548
+ 3R3
549
+ (35)
550
+ • Class 4a J4 = 0 and √J1J2J3 = J5
551
+ 2
552
+ ⟨Fc4a⟩U ⊗3 = ⟨F1⟩3 − 2 ⟨F1⟩
553
+
554
+ ⟨F1⟩ − 1
555
+ 2
556
+
557
+ (1 − ⟨F1⟩)
558
+
559
+ C2
560
+ BC + C2
561
+ AC + C2
562
+ AB
563
+
564
+ ⟨Fc4a⟩ = ⟨F1⟩3 − R3
565
+ (36)
566
+ • Class 4b (J2 = J5 = 0 or J3 = J5 = 0)
567
+ ⟨Fc4b⟩U ⊗3 = ⟨F1⟩3 − ⟨F1⟩
568
+
569
+ ⟨F1⟩ − 1
570
+ 2
571
+
572
+ (1 − ⟨F1⟩)
573
+
574
+ 2
575
+
576
+ C2
577
+ BC + C2
578
+ AC
579
+
580
+ + 3τ 2
581
+ 3
582
+
583
+ ⟨Fc4b⟩ = ⟨F1⟩3 − 5
584
+ 3R3
585
+ (37)
586
+ • Class 4c (J1J4 + J1J2 + J1J3 + J2J3 = √J1J2J3 = J5
587
+ 2 )
588
+ ⟨Fc4c⟩U ⊗3 = ⟨F1⟩3 − ⟨F1⟩
589
+
590
+ ⟨F1⟩ − 1
591
+ 2
592
+
593
+ (1 − ⟨F1⟩)
594
+
595
+ 2
596
+
597
+ C2
598
+ BC + C2
599
+ AC + C2
600
+ AB
601
+
602
+ + 3τ 2
603
+ 3
604
+
605
+ ⟨Fc4c⟩ = ⟨F1⟩3 − 2R3
606
+ (38)
607
+ • Class 4d (√J1J2J3 = |J5|
608
+ 2
609
+ and (J4 + J5)2 − 4 (J1 + J4) (J2 + J4) (J3 + J4) = 0)
610
+ ⟨Fc4c⟩U ⊗3 = ⟨F1⟩3 − ⟨F1⟩
611
+
612
+ ⟨F1⟩ − 1
613
+ 2
614
+
615
+ (1 − ⟨F1⟩)
616
+
617
+ 2
618
+
619
+ C2
620
+ BC + C2
621
+ AC + C2
622
+ AB
623
+
624
+ + 3τ 2
625
+ 3
626
+
627
+ ⟨Fc4d⟩ = ⟨F1⟩3 − 2R3
628
+ (39)
629
+ Comparing Equation (30) with Equation (32), it turns out that, for an equivalent amount of the entanglement
630
+ monotone C2
631
+ jk and τ 2
632
+ 3 , at fixed ⟨F1⟩ (or, equivalently, at a fixed transition amplitude f), the fidelity of the canonical
633
+ state in class 2a is greater than that in class 2b. This is in line with our intuition that the more entangled a state
634
+ is, the harder it is to achieve high fidelity in our parallel QST protocol, as shown in Figure 3 (right panel), where we
635
+ plot the average fidelity for different three qubit classes. In the left panel of Figure 3 we report the reduction factor
636
+ R3 of Equation (25) as a function of the single-particle average fidelity ⟨F1⟩.
637
+ From the above equations of the average fidelity of the three-qubit classes, we see that the average fidelity is
638
+ decreased, with respect to the product state class, whenever there is two-qubit concurrence or genuine multipartite
639
+ entanglement, both as CGME and as τ3. Moreover, per equal amount of squared two-qubit concurrence C2 and genuine
640
+ multipartite entanglement CGME, the reducing factor is respectively 2 and 3 times R3. As a consequence, we state
641
+ that, at fixed amount of entanglement, GME states are harder to transfer than biseparable states.
642
+
643
+ 8
644
+ 0.6
645
+ 0.7
646
+ 0.8
647
+ 0.9
648
+ 1.0
649
+ 0.01
650
+ 0.02
651
+ 0.03
652
+ 0.04
653
+ 0.05
654
+ 0.6
655
+ 0.7
656
+ 0.8
657
+ 0.9
658
+ 1.0
659
+ 0.2
660
+ 0.4
661
+ 0.6
662
+ 0.8
663
+ 1.0
664
+ FIG. 3. (left) Reduction factor for the average fidelity in the presence of entanglement as in Equation (25). (right) Average
665
+ fidelity for the three-qubit classes. The dotted, vertical line is at ⟨F⟩1 = 1
666
+ 2
667
+
668
+ 1 +
669
+ 1
670
+
671
+ 3
672
+
673
+ ≃ 0.789.
674
+ C.
675
+ Four Qubits
676
+ While for two and three qubits, the entanglement of pure states has been fully characterized, for four (or more)
677
+ qubits there are infinitely many inequivalent entanglement classes [30, 33] under SLOCC operations (stochastic local
678
+ operations and classical communication).
679
+ Here, we consider the fidelity of specific four-qubit states averaged over random local unitaries on each qubit.
680
+ Whereas this does not account for all entangled states within a given class, as the group of stocastic local operations
681
+ includes deterministic local operations, SU(2) ⊆ SL(2, C) (with equality holding for pure states), the results nev-
682
+ ertheless hint at the fact that the average fidelity decreases with the entanglement of the sender state and that the
683
+ reduction factor depends on the type of entanglement contained in the state.
684
+ We will consider the three irreducibly balanced states [34]: the 4-qubits GHZ-state, the cluster, and X4 :
685
+ |GHZ4⟩ =
686
+ 1
687
+
688
+ 2 (|0000⟩ + |1111⟩)
689
+ (40a)
690
+ |Cl4⟩ = 1
691
+ 2 (|0000⟩ + |0111⟩ + |1011⟩ + |1100⟩)
692
+ (40b)
693
+ |X4⟩ =
694
+ 1
695
+
696
+ 6
697
+ �√
698
+ 2 |1111⟩ + |0001⟩ + |0010⟩ + |0100⟩ + |1000⟩
699
+
700
+ (40c)
701
+ and two additional entangled states: the product of two-Bell states and the 4-qubit W-state:
702
+ |B2⟩ = |Φ⟩12 ⊗ |Φ⟩34
703
+ (41a)
704
+ |W4⟩ = 1
705
+ 2 (|0001⟩ + |0010⟩ + |0100⟩ + |1000⟩) .
706
+ (41b)
707
+ The average fidelities of the states reported in Equations (40) and (41), expressed in terms of 1-QST average fidelity,
708
+ read
709
+ ⟨FGHZ4⟩U ⊗4 = ⟨FB2⟩U ⊗4 = ⟨F1⟩4 − 2 ⟨F1⟩
710
+
711
+ ⟨F1⟩ − 1
712
+ 2
713
+
714
+ (1 − ⟨F1⟩)
715
+
716
+ 1 − 3 ⟨F⟩ + 4 ⟨F⟩2�
717
+ (42a)
718
+ ⟨FCl4⟩U ⊗4 = ⟨FX4⟩U ⊗4 = ⟨F1⟩4 − 4 ⟨F1⟩2
719
+
720
+ ⟨F1⟩ − 1
721
+ 2
722
+
723
+ (1 − ⟨F1⟩)
724
+ (42b)
725
+ ⟨FW4⟩U ⊗4 = ⟨F1⟩4 − 3 ⟨F1⟩2
726
+
727
+ ⟨F1⟩ − 1
728
+ 2
729
+
730
+ (1 − ⟨F1⟩) .
731
+ (42c)
732
+ Notice that the reduction factor for the states |Cl4⟩ , |X4⟩ , |W4⟩ is the same, although with different weights, and
733
+
734
+ 9
735
+ 0.6
736
+ 0.7
737
+ 0.8
738
+ 0.9
739
+ 1.0
740
+ 0.01
741
+ 0.02
742
+ 0.03
743
+ 0.04
744
+ 0.05
745
+ 0.06
746
+ 0.6
747
+ 0.7
748
+ 0.8
749
+ 0.9
750
+ 1.0
751
+ 0.2
752
+ 0.4
753
+ 0.6
754
+ 0.8
755
+ 1.0
756
+ FIG. 4. (left) Reduction factor for the average fidelity in the presence of entanglement as in Equations (43). (right) Average
757
+ fidelity for the entangled classes as reported in Equations (42). The blue and red dotted, vertical lines are, respectively, at
758
+ ⟨F⟩1 = 0.82 and ⟨F⟩1 = 0.85.
759
+ differs from the reduction factor for the states |GHZ4⟩ , |B2⟩, reading, respectively,
760
+ R4a = ⟨F1⟩
761
+
762
+ ⟨F1⟩ − 1
763
+ 2
764
+
765
+ (1 − ⟨F1⟩)
766
+
767
+ 1 − 3 ⟨F⟩ + 4 ⟨F⟩2�
768
+ (43a)
769
+ R4b = ⟨F1⟩2
770
+
771
+ ⟨F1⟩ − 1
772
+ 2
773
+
774
+ (1 − ⟨F1⟩) .
775
+ (43b)
776
+ A possible reason may be that, if one considers the four-tangle as an entanglement measure, although it is not a
777
+ measure of genuine multipartite entanglement, the first set of state has zero four-tangle, whereas for the second one
778
+ it is non-zero. In Figure 4 (left panel) we report the reduction factors in Equation (43), while in the right panel we
779
+ report the average fidelity of Equation (42).
780
+ IV.
781
+ DISCUSSION
782
+ We have shown that the QST of an entangled n ≥ 2 quantum state across parallel, independent U(1)-symmetric
783
+ quantum channels, as, e.g., embodied by an XXZ spin- 1
784
+ 2 Hamiltonian, leads to a lower average fidelity than that of
785
+ the QST of a product state at fixed one-qubit QST average fidelity, or, equivalently, at fixed transition amplitude.
786
+ For the case of n = 2, we have expressed the average fidelity reduction in terms of the squared concurrence times a
787
+ reduction factor. Similarly, for n = 3, we obtained that the presence of entanglement, both bipartite and multipartite,
788
+ has a detrimental effect on the average fidelity. In particular, we obtained that the reduction factor has a greater
789
+ weight in the presence of genuine three-partite entanglement, i.e., three-tangle and GME concurrence, than in the
790
+ presence of two-qubit squared concurrence for specific canonical classes of the three-qubit pure state. Finally, we have
791
+ considered specific cases of 4-qubit entangled states, which, again, result in an average fidelity reduction due to the
792
+ presence of entanglement in the initial state.
793
+ Our work clearly shows that for entanglement distribution in a routing configuration, where parties are sent over
794
+ independent quantum channels, the single-qubit average fidelity is not a reliable figure of merit. This calls for more
795
+ investigations into the properties of quantum channels able to faithfully distribute multipartite entangled states.
796
+ ACKNOWLEDGEMENTS
797
+ TJGA acknowledges funding through the IPAS+ (Internationalisation Partnership Awards Scheme +) QUEST
798
+ project by the MCST (The Malta Council for Science & Technology). MC acknowledges funding from the Tertiary
799
+ Education Scholarships Scheme and from the Project QVAQT financed by the Malta Council for Science & Technology,
800
+ for and on behalf of the Foundation for Science and Technology, through the FUSION: R&I Research Excellence
801
+ Programme REP-2022-003.
802
+ K ˙Z acknowledges support by Narodowe Centrum Nauki under the Quantera project
803
+
804
+ 10
805
+ number 2021/03/Y/ST2/00193 and by the Foundation for Polish Science under the Team-Net project POIR.04.04.00-
806
+ 00-17C1/18-00. SL acknowledges support by MUR under PRIN Project No. 2017 SRN-BRK QUSHIP
807
+ [1] Hillery, M.; Buˇzek, V.; Berthiaume, A. Quantum secret sharing. Phys. Rev. A 1999, 59, 1829–1834. https://doi.org/
808
+ 10.1103/PhysRevA.59.1829.
809
+ [2] Cuomo, D.; Caleffi, M.; Cacciapuoti, A.S. Towards a distributed quantum computing ecosystem. IET Quantum Commun.
810
+ 2020, 1, 3–8, [2002.11808]. https://doi.org/10.1049/iet-qtc.2020.0002.
811
+ [3] Bose, S.
812
+ Quantum Communication through an Unmodulated Spin Chain.
813
+ Phys. Rev. Lett. 2003, 91, 1–4.
814
+ https:
815
+ //doi.org/10.1103/PhysRevLett.91.207901.
816
+ [4] Aspelmeyer, M.; B¨ohm, H.R.; Gyatso, T.; Jennewein, T.; Kaltenbaek, R.; Lindenthal, M.; Molina-Terriza, G.; Poppe, A.;
817
+ Resch, K.; Taraba, M.; et al. Long-distance free-space distribution of quantum entanglement. Science (New York, N.Y.)
818
+ 2003, 301, 621–3. https://doi.org/10.1126/science.1085593.
819
+ [5] Di Franco, C.; Paternostro, M.; Kim, M. Nested entangled states for distributed quantum channels. Phys. Rev. A 2008,
820
+ 77, 1–4. https://doi.org/10.1103/PhysRevA.77.020303.
821
+ [6] Banchi, L.; Bayat, A.; Verrucchi, P.; Bose, S. Nonperturbative Entangling Gates between Distant Qubits Using Uniform
822
+ Cold Atom Chains. Phys. Rev. Lett. 2011, 106, 140501. https://doi.org/10.1103/PhysRevLett.106.140501.
823
+ [7] Apollaro, T.J.G.; Lorenzo, S.; Plastina, F. Transport of quantum correlations across a spin chain. Int. J. Mod. Phys. B
824
+ 2013, 27, 1345035. https://doi.org/10.1142/S0217979213450355.
825
+ [8] Almeida, G.M.A.; de Moura, F.A.B.F.; Apollaro, T.J.G.; Lyra, M.L. Disorder-assisted distribution of entanglement in
826
+ XY spin chains. Phys. Rev. A 2017, 96, 032315. https://doi.org/10.1103/PhysRevA.96.032315.
827
+ [9] Streltsov, A.; Kampermann, H.; Bruß, D.
828
+ Entanglement Distribution and Quantum Discord. In Lectures on General
829
+ Quantum Correlations and their Applications; Fanchini, F.F.; Soares Pinto, D.d.O.; Adesso, G., Eds.; Springer International
830
+ Publishing: Cham, Switzerland, 2017; pp. 217–230. https://doi.org/10.1007/978-3-319-53412-1_10.
831
+ [10] Wengerowsky, S.; Joshi, S.K.; Steinlechner, F.; Zichi, J.R.; Liu, B.; Scheidl, T.; Dobrovolskiy, S.M.; van der Molen, R.;
832
+ Los, J.W.; Zwiller, V.; et al. Passively stable distribution of polarisation entanglement over 192 km of deployed optical
833
+ fibre. npj Quantum Inf. 2020, 6, 1–5, [1907.04864]. https://doi.org/10.1038/s41534-019-0238-8.
834
+ [11] Vieira, R.; Rigolin, G. Almost perfect transmission of multipartite entanglement through disordered and noisy spin chains.
835
+ Phys. Lett. A 2020, 384, 126536. https://doi.org/10.1016/j.physleta.2020.126536.
836
+ [12] Apollaro, T.J.G.; Almeida, G.M.A.; Lorenzo, S.; Ferraro, A.; Paganelli, S. Spin chains for two-qubit teleportation. Phys.
837
+ Rev. A 2019, 100, 052308. https://doi.org/10.1103/PhysRevA.100.052308.
838
+ [13] Apollaro, T.J.; Sanavio, C.; Chetcuti, W.J.; Lorenzo, S. Multipartite entanglement transfer in spin chains. Phys. Lett. A
839
+ 2020, 384, 126306, [2001.03529]. https://doi.org/10.1016/j.physleta.2020.126306.
840
+ [14] Wang, M.; Xiang, Y.; Kang, H.; Han, D.; Liu, Y.; He, Q.; Gong, Q.; Su, X.; Peng, K. Deterministic Distribution of
841
+ Multipartite Entanglement and Steering in a Quantum Network by Separable States. Phys. Rev. Lett. 2020, 125, 6388–
842
+ 6396, [2101.01422]. https://doi.org/10.1103/PhysRevLett.125.260506.
843
+ [15] Mannalath, V.; Pathak, A. Multiparty Entanglement Routing in Quantum Networks, 2022. https://doi.org/10.48550/
844
+ ARXIV.2211.06690.
845
+ [16] Eltschka, C.; Siewert, J. Quantifying entanglement resources. J. Phys. Math. Theor. 2014, 47, 424005. https://doi.
846
+ org/10.1088/1751-8113/47/42/424005.
847
+ [17] Bina, M.; Mandarino, A.; Olivares, S.; Paris, M.G.A. Drawbacks of the use of fidelity to assess quantum resources. Phys.
848
+ Rev. A 2014, 89, 012305, [1309.5325]. https://doi.org/10.1103/PhysRevA.89.012305.
849
+ [18] Mandarino, A.; Bina, M.; Olivares, S.; Paris, M.G.
850
+ About the use of fidelity in continuous variable systems. Int. J.
851
+ Quantum Inf. 2014, 12, 1–9, [1402.0976]. https://doi.org/10.1142/S0219749914610152.
852
+ [19] G¨uhne, O.; Toth, G. Entanglement detection. Phys. Rep. 2008, 474, 1–75, [0811.2803]. https://doi.org/10.1016/j.
853
+ physrep.2009.02.004.
854
+ [20] Lorenzo, S.; Plastina, F.; Consiglio, M.; Apollaro, T.J.G.
855
+ Quantum Map Approach to Entanglement Transfer and
856
+ Generation in Spin Chains. In Entanglement in Spin Chains:
857
+ From Theory to Quantum Technology Applications;
858
+ Bayat, A.; Bose, S.; Johannesson, H., Eds.; Springer International Publishing: Cham, Switzerland, 2022; pp. 321–340.
859
+ https://doi.org/10.1007/978-3-031-03998-0_12.
860
+ [21] Apollaro, T.J.G.; Lorenzo, S.; Plastina, F.; Consiglio, M.; ˙Zyczkowski, K. Quantum transfer of interacting qubits. New J.
861
+ Phys. 2022, 24, 083025, [2205.01579]. https://doi.org/10.1088/1367-2630/ac86e7.
862
+ [22] Wootters, W.K. Entanglement of formation of an arbitrary state of two qubits. Phys. Rev. Lett. 1998, 80, 2245–2248,
863
+ [arXiv:quant-ph/9709029]. https://doi.org/10.1103/PhysRevLett.80.2245.
864
+ [23] Jozsa, R.
865
+ Fidelity for Mixed Quantum States.
866
+ J. Mod. Opt. 1994, 41, 2315–2323.
867
+ https://doi.org/10.1080/
868
+ 09500349414552171.
869
+ [24] Bellomo, B.; Lo Franco, R.; Compagno, G. Non-Markovian Effects on the Dynamics of Entanglement. Phys. Rev. Lett.
870
+ 2007, 99, 1–4. https://doi.org/10.1103/PhysRevLett.99.160502.
871
+ [25] Bose, S. Quantum communication through spin chain dynamics: an introductory overview. Contemp. Phys. 2007, 48, 13–
872
+ 30. https://doi.org/10.1080/00107510701342313.
873
+
874
+ 11
875
+ [26] Pan, J.W.; Bouwmeester, D.; Weinfurter, H.; Zeilinger, A. Experimental Entanglement Swapping: Entangling Photons
876
+ That Never Interacted. Phys. Rev. Lett. 1998, 80, 3891–3894. https://doi.org/10.1103/PhysRevLett.80.3891.
877
+ [27] Apollaro, T.J.G.; Lorenzo, S.; Sindona, A.; Paganelli, S.; Giorgi, G.L.; Plastina, F. Many-qubit quantum state transfer
878
+ via spin chains. Phys. Scr. 2015, T165, 014036. https://doi.org/10.1088/0031-8949/2015/T165/014036.
879
+ [28] Ac´ın, A.; Andrianov, A.; Costa, L.; Jan´e, E.; Latorre, J.I.; Tarrach, R.
880
+ Generalized Schmidt Decomposition and
881
+ Classification of Three-Quantum-Bit States. Phys. Rev. Lett. 2000, 85, 1560–1563, [arXiv:quant-ph/0003050]. https:
882
+ //doi.org/10.1103/PhysRevLett.85.1560.
883
+ [29] Enr´ıquez, M.; Delgado, F.; ˙Zyczkowski, K. Entanglement of Three-Qubit Random Pure States. Entropy 2018, 20, 745.
884
+ https://doi.org/10.3390/e20100745.
885
+ [30] D¨ur, W.; Vidal, G.; Cirac, J.I. Three qubits can be entangled in two inequivalent ways. Phys. Rev. A 2000, 62, 062314.
886
+ https://doi.org/10.1103/PhysRevA.62.062314.
887
+ [31] Coffman, V.; Kundu, J.; Wootters, W.K.
888
+ Distributed Entanglement.
889
+ Phys. Rev. A 1999, 61, 052306, [arXiv:quant-
890
+ ph/9907047]. https://doi.org/10.1103/PhysRevA.61.052306.
891
+ [32] Ma, Z.H.; Chen, Z.H.; Chen, J.L.; Spengler, C.; Gabriel, A.; Huber, M. Measure of genuine multipartite entanglement with
892
+ computable lower bounds. Phys. Rev. A 2011, 83, 062325, [1101.2001]. https://doi.org/10.1103/PhysRevA.83.062325.
893
+ [33] Verstraete, F.; Dehaene, J.; De Moor, B.; Verschelde, H. Four qubits can be entangled in nine different ways. Phys. Rev.
894
+ A 2002, 65, 052112, [arXiv:quant-ph/0109033]. https://doi.org/10.1103/PhysRevA.65.052112.
895
+ [34] Osterloh, A.; Siewert, J. The invariant-comb approach and its relation to the balancedness of multipartite entangled states.
896
+ New J. Phys. 2010, 12. https://doi.org/10.1088/1367-2630/12/7/075025.
897
+
AtE3T4oBgHgl3EQfTgrD/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
BNAzT4oBgHgl3EQfv_6i/content/tmp_files/2301.01716v1.pdf.txt ADDED
@@ -0,0 +1,3093 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.01716v1 [math.OC] 4 Jan 2023
2
+ First-order penalty methods for bilevel optimization
3
+ Zhaosong Lu ∗
4
+ Sanyou Mei ∗
5
+ January 4, 2023
6
+ Abstract
7
+ In this paper we study a class of unconstrained and constrained bilevel optimization problems in
8
+ which the lower-level part is a convex optimization problem, while the upper-level part is possibly
9
+ a nonconvex optimization problem. In particular, we propose penalty methods for solving them,
10
+ whose subproblems turn out to be a structured minimax problem and are suitably solved by a first-
11
+ order method developed in this paper. Under some suitable assumptions, an operation complexity
12
+ of O(ε−4 log ε−1) and O(ε−7 log ε−1), measured by their fundamental operations, is established for
13
+ the proposed penalty methods for finding an ε-KKT solution of the unconstrained and constrained
14
+ bilevel optimization problems, respectively.
15
+ To the best of our knowledge, the methodology and
16
+ results in this paper are new.
17
+ Keywords: bilevel optimization, minimax optimization, penalty methods, first-order methods, opera-
18
+ tion complexity
19
+ Mathematics Subject Classification: 90C26, 90C30, 90C47, 90C99, 65K05
20
+ 1
21
+ Introduction
22
+ Bilevel optimization is a two-level hierarchical optimization in which partial or full decision variables in
23
+ the upper level are also involved in the lower level. Generically, it can be written in the following form:
24
+ min
25
+ x,y
26
+ f(x, y)
27
+ s.t.
28
+ g(x, y) ≤ 0,
29
+ y ∈ Argmin
30
+ z
31
+ { ˜f(x, z)|˜g(x, z) ≤ 0}.1
32
+ (1)
33
+ Bilevel optimization has found a variety of important applications, including adversarial training [36,
34
+ 37, 46], continual learning [32], hyperparameter tuning [3, 17], image reconstruction [9], meta-learning
35
+ [4, 23, 42], neural architecture search [15, 30], reinforcement learning [20, 27], and Stackelberg games [48].
36
+ More applications about it can be found in [2, 8, 10, 11, 12, 44] and the references therein. Theoretical
37
+ properties including optimality conditions of (1) have been extensively studied in the literature (e.g., see
38
+ [12, 13, 34, 47, 50]).
39
+ Numerous methods have been developed for solving some special cases of (1). For example, constraint-
40
+ based methods [19, 43], deterministic gradient-based methods [16, 17, 21, 35, 41, 42], and stochastic
41
+ gradient-based methods [6, 18, 20, 24, 26] were proposed for solving (1) with g ≡ 0, ˜g ≡ 0, f, ˜f being
42
+ smooth, and ˜f being strongly convex with respect to y. Besides, when all the functions involved in
43
+ (1) are smooth and ˜f, ˜g are convex with respect to y, gradient-type methods were proposed by solving
44
+ the mathematical program with equilibrium constraints (MPEC) resulting from replacing the lower-level
45
+ optimization problem of (1) by its first-order optimality conditions (e.g., see [1, 33, 40]).
46
+ Recently,
47
+ difference-of-convex (DC) algorithms were developed in [51] for solving (1) with g ≡ 0, f being a DC
48
+ function, and ˜f, ˜g being convex functions. In addition, a double penalty method [22] was proposed for
49
+ (1), which solves a sequence of bilevel optimization problems of the form
50
+ min
51
+ x,y
52
+ f(x, y) + ρkΨ(x, y)
53
+ s.t.
54
+ y ∈ Argmin
55
+ z
56
+ ˜f(x, z) + ρk ˜Ψ(x, z),
57
+ (2)
58
+ ∗Department of Industrial and Systems Engineering, University of Minnesota, USA (email:
59
+ zhaosong@umn.edu,
60
+ mei00035@umn.edu). This work was partially supported by NSF Award IIS-2211491.
61
+ 1For ease of reading, throughout this paper the tilde symbol is particularly used for the functions related to the lower-level
62
+ optimization problem. Besides, “Argmin” denotes the set of optimal solutions of the associated problem.
63
+ 1
64
+
65
+ where {ρk} is a sequence of penalty parameters, and Ψ and ˜Ψ are a penalty function associated with the
66
+ sets {(x, y)|g(x, y) ≤ 0} and {(x, z)|˜g(x, z) ≤ 0}, respectively. Though problem (2) appears to be simpler
67
+ than (1), there is no method available for finding an approximate solution of (2) in general. Conse-
68
+ quently, the double penalty method [22] is typically not implementable. More discussion on algorithmic
69
+ development for bilevel optimization can be found in [2, 8, 12, 31, 45, 47]) and the references therein.
70
+ It has long been known that the notorious challenge of bilevel optimization (1) mainly comes from the
71
+ lower level part, which requires that the variable y be a solution of another optimization problem. Due
72
+ to this, for the sake of simplicity, we only consider a subclass of bilevel optimization with the constraint
73
+ g(x, y) ≤ 0 being excluded, namely,
74
+ min
75
+ x,y
76
+ f(x, y)
77
+ s.t.
78
+ y ∈ Argmin
79
+ z
80
+ { ˜f(x, z)|˜g(x, z) ≤ 0}.
81
+ (3)
82
+ Nevertheless, the results in this paper can be possibly extended to problem (1).
83
+ The main goal of this paper is to develop a first-order penalty method for solving problem (3). Our
84
+ key observations toward this development are: (i) problem (3) can be approximately solved as a penalty
85
+ problem (see (49)); (ii) such a penalty problem is equivalent to a structured minimax problem (see
86
+ (50)), which can be suitably solved by a first-order method proposed in Section 2. As a result, these
87
+ observations lead to development of a novel first-order penalty method for solving (3) (see Sections 3
88
+ and 4), which enjoys the following appealing features.
89
+ • It uses only the first-order information of the problem. Specifically, its fundamental operations
90
+ consist only of evaluations of the gradient of ˜g and the smooth component of f and ˜f and also
91
+ the proximal operator of the nonsmooth component of f and ˜f. Thus, it is suitable for solving
92
+ large-scale problems (see Sections 3 and 4).
93
+ • It has theoretical guarantees on operation complexity, which is measured by the aforementioned
94
+ fundamental operations, for finding an ε-KKT solution of (3).
95
+ In particular, when ˜g ≡ 0, it
96
+ enjoys an operation complexity of O(ε−4 log ε−1). Otherwise, it enjoys an operation complexity of
97
+ O(ε−7 log ε−1) (see Theorems 4 and 6).
98
+ • It is applicable to a broader class of problems than existing methods.
99
+ For example, it can be
100
+ applied to (3) with f, ˜f being nonsmooth and ˜f, ˜g being nonconvex with respect to x, which is
101
+ however not suitable for existing methods.
102
+ To the best of our knowledge, the methodology and results in this paper are new.
103
+ The rest of this paper is organized as follows. In Subsection 1.1 we introduce some notation and
104
+ terminology. In Section 2 we propose a first-order method for solving a nonconvex-concave minimax
105
+ problem and study its complexity.
106
+ In Sections 3 and 4, we propose first-order penalty methods for
107
+ unconstrained and constrained bilevel optimization and study their complexity, respectively. In Section
108
+ 5 we present the proofs of the main results. Finally, we make some concluding remarks in Section 6.
109
+ 1.1
110
+ Notation and terminology
111
+ The following notation will be used throughout this paper.
112
+ Let Rn denote the Euclidean space of
113
+ dimension n and Rn
114
+ + denote the nonnegative orthant in Rn. The standard inner product and Euclidean
115
+ norm are denoted by ⟨·, ·⟩ and ∥ · ∥, respectively. For any v ∈ Rn, let v+ denote the nonnegative part of
116
+ v, that is, (v+)i = max{vi, 0} for all i. For any two vectors u and v, (u; v) denotes the vector resulting
117
+ from stacking v under u. Given a point x and a closed set S in Rn, let dist(x, S) = minx′∈S ∥x′ − x∥ and
118
+ IS denote the indicator function associated with S.
119
+ A function or mapping φ is said to be Lφ-Lipschitz continuous on a set S if ∥φ(x)−φ(x′)∥ ≤ Lφ∥x−x′∥
120
+ for all x, x′ ∈ S. In addition, it is said to be L∇φ-smooth on S if ∥∇φ(x) − ∇φ(x′)∥ ≤ L∇φ∥x − x′∥ for
121
+ all x, x′ ∈ S. For a closed convex function p : Rn → R ∪ {∞},2 the proximal operator associated with p
122
+ is denoted by proxp, that is,
123
+ proxp(x) = arg min
124
+ x′∈Rn
125
+ �1
126
+ 2∥x′ − x∥2 + p(x′)
127
+
128
+ ∀x ∈ Rn.
129
+ (4)
130
+ 2For convenience, ∞ stands for +∞.
131
+ 2
132
+
133
+ Given that evaluation of proxγp(x) is often as cheap as proxp(x), we count the evaluation of proxγp(x)
134
+ as one evaluation of proximal operator of p for any γ > 0 and x ∈ Rn.
135
+ For a lower semicontinuous function φ : Rn → R∪{∞}, its domain is the set dom φ := {x|φ(x) < ∞}.
136
+ The upper subderivative of φ at x ∈ dom φ in a direction d ∈ Rn is defined by
137
+ φ′(x; d) = lim sup
138
+ x′ φ
139
+ →x, t↓0
140
+ inf
141
+ d′→d
142
+ φ(x′ + td′) − φ(x′)
143
+ t
144
+ ,
145
+ where t ↓ 0 means both t > 0 and t → 0, and x′
146
+ φ→ x means both x′ → x and φ(x′) → φ(x). The
147
+ subdifferential of φ at x ∈ dom φ is the set
148
+ ∂φ(x) = {s ∈ Rn��sT d ≤ φ′(x; d) ∀d ∈ Rn}.
149
+ We use ∂xiφ(x) to denote the subdifferential with respect to xi. In addition, for an upper semicontinuous
150
+ function φ, its subdifferential is defined as ∂φ = −∂(−φ). If φ is locally Lipschitz continuous, the above
151
+ definition of subdifferential coincides with the Clarke subdifferential. Besides, if φ is convex, it coincides
152
+ with the ordinary subdifferential for convex functions. Also, if φ is continuously differentiable at x , we
153
+ simply have ∂φ(x) = {∇φ(x)}, where ∇φ(x) is the gradient of φ at x. In addition, it is not hard to
154
+ verify that ∂(φ1 + φ2)(x) = ∇φ1(x) + ∂φ2(x) if φ1 is continuously differentiable at x and φ2 is lower or
155
+ upper semicontinuous at x. See [7, 49] for more details.
156
+ Finally, we introduce two types of approximate solutions for a general minimax problem
157
+ Ψ∗ = min
158
+ x max
159
+ y
160
+ Ψ(x, y),
161
+ (5)
162
+ where Ψ(·, y) : Rn → R ∪ {∞} is a lower semicontinuous function, Ψ(x, ·) : Rm → R ∪ {−∞} is an upper
163
+ semicontinuous function, and Ψ∗ is finite.
164
+ Definition 1. A point (xǫ, yǫ) is called an ǫ-optimal solution of the minimax problem (5) if
165
+ max
166
+ y
167
+ Ψ(xǫ, y) − Ψ(xǫ, yǫ) ≤ ǫ,
168
+ Ψ(xǫ, yǫ) − Ψ∗ ≤ ǫ.
169
+ Definition 2. A point (x, y) is called a stationary point of the minimax problem (5) if
170
+ 0 ∈ ∂xΨ(x, y),
171
+ 0 ∈ ∂yΨ(x, y).
172
+ In addition, for any ǫ > 0, a point (xǫ, yǫ) is called an ǫ-stationary point of the minimax problem (5) if
173
+ dist (0, ∂xΨ(xǫ, yǫ)) ≤ ǫ,
174
+ dist (0, ∂yΨ(xǫ, yǫ)) ≤ ǫ.
175
+ 2
176
+ A first-order method for nonconvex-concave minimax prob-
177
+ lem
178
+ In this section, we propose a first-order method for finding an approximate stationary point of a
179
+ nonconvex-concave minimax problem, which will be used as a subproblem solver for the penalty methods
180
+ proposed in Sections 3 and 4. In particular, we consider the minimax problem
181
+ H∗ = min
182
+ x max
183
+ y
184
+ {H(x, y) := h(x, y) + p(x) − q(y)} .
185
+ (6)
186
+ Assume that problem (6) has at least one optimal solution. In addition, h, p and q satisfy the following
187
+ assumptions.
188
+ Assumption 1.
189
+ (i) p : Rn → R ∪ {∞} and q : Rm → R ∪ {∞} are proper convex functions and
190
+ continuous on their domain, and moreover, their domain is compact.
191
+ (ii) The proximal operator associated with p and q can be exactly evaluated.
192
+ (iii) h is L∇h-smooth on dom p × dom q, and moreover, h(x, ·) is concave for any x ∈ dom p.
193
+ 3
194
+
195
+ Recently, an accelerated inexact proximal point smoothing (AIPP-S) scheme was proposed in [28]
196
+ for finding an approximate stationary point of a class of minimax composite nonconvex optimization
197
+ problems, which includes (6) as a special case. When applied to (6), AIPP-S requires the availability of
198
+ the oracle including exact evaluation of ∇xh(x, y) and
199
+ arg min
200
+ x
201
+
202
+ p(x) + 1
203
+ 2λ∥x − x′∥2
204
+
205
+ ,
206
+ arg max
207
+ y
208
+
209
+ h(x′, y) − q(y) − 1
210
+ 2λ∥y − y′∥2
211
+
212
+ (7)
213
+ for any λ > 0, x′ ∈ Rn and y′ ∈ Rm. Note that h is typically sophisticated and the exact solution of the
214
+ second problem in (7) usually cannot be found. As a result, AIPP-S is generally not implementable for
215
+ (6), though an oracle complexity of O(ǫ−5/2) was established in [28] for it to find an ǫ-stationary point
216
+ of (6).
217
+ In what follows, we first propose a modified optimal first-order method for solving a strongly-convex-
218
+ strongly-concave minimax problem in Subsection 2.1. Using this method as a subproblem solver for an
219
+ inexact proximal point scheme, we then propose a first-order method for (6) in Subsection 2.2, which
220
+ enjoys an operation complexity of O(ǫ−5/2 log ǫ−1), measured by the amount of evaluations of ∇h and
221
+ proximal operator of p and q, for finding an ǫ-stationary point of (6).
222
+ 2.1
223
+ A modified optimal first-order method for strongly-convex-strongly-concave
224
+ minimax problem
225
+ In this subsection, we consider the strongly-convex-strongly-concave minimax problem
226
+ ¯H∗ = min
227
+ x max
228
+ y
229
+ � ¯H(x, y) := ¯h(x, y) + p(x) − q(y)
230
+
231
+ ,
232
+ (8)
233
+ where p and q satisfy Assumption 1 and ¯h satisfies the following assumption.
234
+ Assumption 2. ¯h(x, y) is σx-strongly-convex-σy-strongly-concave and L∇¯h-smooth on dom p × dom q
235
+ for some σx, σy > 0.
236
+ The goal of this subsection is to propose a modified optimal first-order method for finding an approx-
237
+ imate stationary point of problem (8) and study its complexity. Before proceeding, we introduce some
238
+ more notation below, most of which is adopted from [29].
239
+ Let (x∗, y∗) denote the optimal solution of (8), z∗ = −σxx∗, and
240
+ Dp = max{∥u − v∥
241
+ ��u, v ∈ dom p},
242
+ Dq = max{∥u − v∥
243
+ ��u, v ∈ dom q},
244
+ (9)
245
+ ¯Hlow = min
246
+ � ¯H(x, y)|
247
+
248
+ x, y) ∈ dom p × dom q},
249
+ (10)
250
+ ˆh(x, y) = ¯h(x, y) − σx∥x∥2/2 + σy∥y∥2/2,
251
+ (11)
252
+ G(z, y) = sup
253
+ x {⟨x, z⟩ − p(x) − ˆh(x, y) + q(y)},
254
+ (12)
255
+ P(z, y) = σ−1
256
+ x ∥z∥2/2 + σy∥y∥2/2 + G(z, y),
257
+ (13)
258
+ ϑk = η−1
259
+ z ∥zk − z∗∥2 + η−1
260
+ y ∥yk − y∗∥2 + 2¯α−1(P(zk
261
+ f, yk
262
+ f) − P(z∗, y∗)),
263
+ (14)
264
+ ak
265
+ x(x, y) = ∇xˆh(x, y) + σx(x − σ−1
266
+ x zk
267
+ g)/2,
268
+ ak
269
+ y(x, y) = −∇yˆh(x, y) + σyy + σx(y − yk
270
+ g)/8,
271
+ where ¯α = min
272
+
273
+ 1,
274
+
275
+ 8σy/σx
276
+
277
+ , ηz = σx/2, ηy = min {1/(2σy), 4/(¯ασx)}, and yk, yk
278
+ f, yk
279
+ g, zk, zk
280
+ f and zk
281
+ g
282
+ are generated at iteration k of Algorithm 1 below. By Assumptions 1 and 2, one can observe that Dp,
283
+ Dq and ¯Hlow are finite.
284
+ We are now ready to present a modified optimal first-order method for solving (8) in Algorithm 1. It is
285
+ a slight modification of the novel optimal first-order method [29, Algorithm 4] by incorporating a forward-
286
+ backward splitting scheme and also a verifiable termination criterion (see steps 23-25 in Algorithm 1) in
287
+ order to find a τ-stationary point of (8) (see Definition 2) for any prescribed tolerance τ > 0.
288
+ 4
289
+
290
+ Algorithm 1 A modified optimal first-order method for (8)
291
+ Input: τ > 0, ¯z0 = z0
292
+ f ∈ −σxdom p,3 ¯y0 = y0
293
+ f ∈ dom q, (z0, y0) = (¯z0, ¯y0), ¯α = min
294
+
295
+ 1,
296
+
297
+ 8σy/σx
298
+
299
+ ,
300
+ ηz = σx/2, ηy = min {1/(2σy), 4/(¯ασx)}, βt = 2/(t + 3), ζ =
301
+
302
+ 2
303
+
304
+ 5(1 + 8L∇¯h/σx)
305
+ �−1, γx = γy =
306
+ 8σ−1
307
+ x , and ˆζ = min{σx, σy}/L2
308
+ ∇¯h.
309
+ 1: for k = 0, 1, 2, . . . do
310
+ 2:
311
+ (zk
312
+ g , yk
313
+ g) = ¯α(zk, yk) + (1 − ¯α)(zk
314
+ f, yk
315
+ f).
316
+ 3:
317
+ (xk,−1, yk,−1) = (−σ−1
318
+ x zk
319
+ g, yk
320
+ g).
321
+ 4:
322
+ xk,0 = proxζγxp(xk,−1 − ζγxak
323
+ x(xk,−1, yk,−1)).
324
+ 5:
325
+ yk,0 = proxζγyq(yk,−1 − ζγyak
326
+ y(xk,−1, yk,−1)).
327
+ 6:
328
+ bk,0
329
+ x
330
+ =
331
+ 1
332
+ ζγx (xk,−1 − ζγxak
333
+ x(xk,−1, yk,−1) − xk,0).
334
+ 7:
335
+ bk,0
336
+ y
337
+ =
338
+ 1
339
+ ζγy (yk,−1 − ζγyak
340
+ y(xk,−1, yk,−1) − yk,0).
341
+ 8:
342
+ t = 0.
343
+ 9:
344
+ while
345
+ γx∥ak
346
+ x(xk,t, yk,t) + bk,t
347
+ x ∥2 + γy∥ak
348
+ y(xk,t, yk,t) + bk,t
349
+ y ∥2 > γ−1
350
+ x ∥xk,t − xk,−1∥2 + γ−1
351
+ y ∥yk,t − yk,−1∥2
352
+ do
353
+ 10:
354
+ xk,t+1/2 = xk,t + βt(xk,0 − xk,t) − ζγx(ak
355
+ x(xk,t, yk,t) + bk,t
356
+ x ).
357
+ 11:
358
+ yk,t+1/2 = yk,t + βt(yk,0 − yk,t) − ζγy(ak
359
+ y(xk,t, yk,t) + bk,t
360
+ y ).
361
+ 12:
362
+ xk,t+1 = proxζγxp(xk,t + βt(xk,0 − xk,t) − ζγxak
363
+ x(xk,t+1/2, yk,t+1/2)).
364
+ 13:
365
+ yk,t+1 = proxζγyq(yk,t + βt(yk,0 − yk,t) − ζγyak
366
+ y(xk,t+1/2, yk,t+1/2)).
367
+ 14:
368
+ bk,t+1
369
+ x
370
+ =
371
+ 1
372
+ ζγx (xk,t + βt(xk,0 − xk,t) − ζγxak
373
+ x(xk,t+1/2, yk,t+1/2) − xk,t+1).
374
+ 15:
375
+ bk,t+1
376
+ y
377
+ =
378
+ 1
379
+ ζγy (yk,t + βt(yk,0 − yk,t) − ζγyak
380
+ y(xk,t+1/2, yk,t+1/2) − yk,t+1).
381
+ 16:
382
+ t ← t + 1.
383
+ 17:
384
+ end while
385
+ 18:
386
+ (xk+1
387
+ f
388
+ , yk+1
389
+ f
390
+ ) = (xk,t, yk,t).
391
+ 19:
392
+ (zk+1
393
+ f
394
+ , wk+1
395
+ f
396
+ ) = (∇xˆh(xk+1
397
+ f
398
+ , yk+1
399
+ f
400
+ ) + bk,t
401
+ x , −∇yˆh(xk+1
402
+ f
403
+ , yk+1
404
+ f
405
+ ) + bk,t
406
+ y ).
407
+ 20:
408
+ zk+1 = zk + ηzσ−1
409
+ x (zk+1
410
+ f
411
+ − zk) − ηz(xk+1
412
+ f
413
+ + σ−1
414
+ x zk+1
415
+ f
416
+ ).
417
+ 21:
418
+ yk+1 = yk + ηyσy(yk+1
419
+ f
420
+ − yk) − ηy(wk+1
421
+ f
422
+ + σyyk+1
423
+ f
424
+ ).
425
+ 22:
426
+ xk+1 = −σ−1
427
+ x zk+1.
428
+ 23:
429
+ ˆxk+1 = proxˆζp(xk+1 − ˆζ∇x¯h(xk+1, yk+1)).
430
+ 24:
431
+ ˆyk+1 = proxˆζq(yk+1 + ˆζ∇y¯h(xk+1, yk+1)).
432
+ 25:
433
+ Terminate the algorithm and output (ˆxk+1, ˆyk+1) if
434
+ ∥ˆζ−1(xk+1 − ˆxk+1, ˆyk+1 − yk+1) − (∇¯h(xk+1, yk+1) − ∇¯h(ˆxk+1, ˆyk+1))∥ ≤ τ.
435
+ (15)
436
+ 26: end for
437
+ The following theorem presents iteration and operation complexity of Algorithm 1 for finding a τ-
438
+ stationary point of problem (8), whose proof is deferred to Subsection 5.1.
439
+ Theorem 1 (Complexity of Algorithm 1). Suppose that Assumptions 1 and 2 hold. Let ¯H∗, Dp,
440
+ Dq, ¯Hlow, and ϑ0 be defined in (8), (9), (10) and (14), σx, σy and L∇¯h be given in Assumption 2, ¯α,
441
+ ηy, ηz, τ, ˆζ be given in Algorithm 1, and
442
+ ¯δ = (2 + ¯α−1)σxD2
443
+ p + max{2σy, ¯ασx/4}D2
444
+ q,
445
+ (16)
446
+ ¯K =
447
+
448
+ max
449
+ � 2
450
+ ¯α, ¯ασx
451
+ 4σy
452
+
453
+ log 4 max{ηzσ−2
454
+ x , ηy}ϑ0
455
+ (ˆζ−1 + L∇¯h)−2τ 2
456
+
457
+ +
458
+ ,
459
+ (17)
460
+ ¯N =
461
+
462
+ max
463
+
464
+ 2,
465
+ � σx
466
+ 2σy
467
+
468
+ log 4 max {1/(2σx), min {1/(2σy), 4/(¯ασx)}}
469
+ �¯δ + 2¯α−1 � ¯H∗ − ¯Hlow
470
+ ��
471
+ (L2
472
+ ∇¯h/ min{σx, σy} + L∇¯h)−2τ 2
473
+
474
+ +
475
+ ×
476
+ ��
477
+ 96
478
+
479
+ 2
480
+
481
+ 1 + 8L∇¯hσ−1
482
+ x
483
+ ��
484
+ + 2
485
+
486
+ .
487
+ (18)
488
+ Then Algorithm 1 outputs a τ-stationary point of (8) in at most ¯K iterations.
489
+ Moreover, the total
490
+ 3For convenience, −σxdom p stands for the set {−σxu|u ∈ dom p}.
491
+ 5
492
+
493
+ number of evaluations of ∇¯h and proximal operator of p and q performed in Algorithm 1 is no more than
494
+ ¯N, respectively.
495
+ Remark 1. It can be observed from Theorem 1 that Algorithm 1 enjoys an operation complexity of
496
+ O(log τ−1), measured by the amount of evaluations of ∇¯h and proximal operator of p and q, for finding
497
+ a τ-stationary point of the strongly-convex-strongly-concave minimax problem (8).
498
+ 2.2
499
+ A first-order method for problem (6)
500
+ In this subsection, we propose a first-order method for finding an ǫ-stationary point of problem (6) (see
501
+ Definition 2) for any prescribed tolerance ǫ > 0. In particular, we first add a perturbation to the max
502
+ part of (6) for obtaining an approximation of (6), which is given as follows:
503
+ min
504
+ x max
505
+ y
506
+
507
+ h(x, y) + p(x) − q(y) −
508
+ ǫ
509
+ 4Dq
510
+ ∥y − ˆy0∥2
511
+
512
+ (19)
513
+ for some ˆy0 ∈ dom q. We then apply an inexact proximal point method [25] to (19), which consists of
514
+ approximately solving a sequence of subproblems
515
+ min
516
+ x max
517
+ y
518
+ {Hk(x, y) := hk(x, y) + p(x) − q(y)} ,
519
+ (20)
520
+ where
521
+ hk(x, y) = h(x, y) − ǫ∥y − ˆy0∥2/(4Dq) + L∇h∥x − xk∥2.
522
+ (21)
523
+ By Assumption 1, one can observe that (i) hk is L∇h-strongly convex in x and ǫ/(2Dq)-strongly concave
524
+ in y on dom p × dom q; (ii) hk is (3L∇h + ǫ/(2Dq))-smooth on dom p × dom q. Consequently, problem
525
+ (20) is a special case of (8) and we can apply Algorithm 1 to solve (20). The resulting first-order method
526
+ for (6) is presented in Algorithm 2.
527
+ Algorithm 2 A first-order method for problem (6)
528
+ Input: ǫ > 0, ǫ0 ∈ (0, ǫ/2], (ˆx0, ˆy0) ∈ dom p × dom q, (x0, y0) = (ˆx0, ˆy0), and ǫk = ǫ0/(k + 1).
529
+ 1: for k = 0, 1, 2, . . . do
530
+ 2:
531
+ Call Algorithm 1 with ¯h ← hk, τ ← ǫk, σx ← L∇h, σy ← ǫ/(2Dq), L∇¯h ← 3L∇h + ǫ/(2Dq),
532
+ ¯z0 = z0
533
+ f ← −σxxk, ¯y0 = y0
534
+ f ← yk, and denote its output by (xk+1, yk+1), where hk is given in (21).
535
+ 3:
536
+ Terminate the algorithm and output (xǫ, yǫ) = (xk+1, yk+1) if
537
+ ∥xk+1 − xk∥ ≤ ǫ/(4L∇h).
538
+ (22)
539
+ 4: end for
540
+ Remark 2. It can be observed from step 2 of Algorithm 2 that (xk+1, yk+1) results from applying Algo-
541
+ rithm 1 to the subproblem (20). As will be shown in Lemma 2, (xk+1, yk+1) is an ǫk-stationary point of
542
+ (20).
543
+ We next study complexity of Algorithm 2 for finding an ǫ-stationary point of problem (6). Before
544
+ proceeding, we define
545
+ Hlow := min {H(x, y)|(x, y) ∈ dom p × dom q} .
546
+ (23)
547
+ By Assumption 1, one can observe that Hlow is finite.
548
+ The following theorem presents iteration and operation complexity of Algorithm 2 for finding an
549
+ ǫ-stationary point of problem (6), whose proof is deferred to Subsection 5.2.
550
+ Theorem 2 (Complexity of Algorithm 2). Suppose that Assumption 1 holds. Let H∗, H Dp, Dq,
551
+ and Hlow be defined in (6), (9) and (23), L∇h be given in Assumption 1, ǫ, ǫ0 and ˆx0 be given in
552
+ 6
553
+
554
+ Algorithm 2, and
555
+ α = min
556
+
557
+ 1,
558
+
559
+ 4ǫ/(DqL∇h)
560
+
561
+ ,
562
+ (24)
563
+ δ = (2 + α−1)L∇hD2
564
+ p + max {ǫ/Dq, αL∇h/4} D2
565
+ q,
566
+ (25)
567
+ K =
568
+
569
+ 16(max
570
+ y
571
+ H(ˆx0, y) − H∗ + ǫDq/4)L∇hǫ−2 + 32ǫ2
572
+ 0(1 + 4D2
573
+ qL2
574
+ ∇hǫ−2)ǫ−2 − 1
575
+
576
+ +
577
+ ,
578
+ (26)
579
+ N =
580
+ ��
581
+ 96
582
+
583
+ 2
584
+
585
+ 1 + (24L∇h + 4ǫ/Dq) L−1
586
+ ∇h
587
+ ��
588
+ + 2
589
+
590
+ max
591
+
592
+ 2,
593
+
594
+ DqL∇hǫ−1
595
+
596
+ ×
597
+
598
+ (K + 1)
599
+
600
+ log
601
+ 4 max
602
+
603
+ 1
604
+ 2L∇h , min
605
+
606
+ Dq
607
+ ǫ ,
608
+ 4
609
+ αL∇h
610
+ �� �
611
+ δ + 2α−1(H∗ − Hlow + ǫDq/4 + L∇hD2
612
+ p)
613
+
614
+ [(3L∇h + ǫ/(2Dq))2/ min{L∇h, ǫ/(2Dq)} + 3L∇h + ǫ/(2Dq)]−2 ǫ2
615
+ 0
616
+
617
+ +
618
+ + K + 1 + 2K log(K + 1)
619
+
620
+ .
621
+ (27)
622
+ Then Algorithm 2 terminates and outputs an ǫ-stationary point (xǫ, yǫ) of (6) in at most K + 1 outer
623
+ iterations that satisfies
624
+ max
625
+ y
626
+ H(xǫ, y) ≤ max
627
+ y
628
+ H(ˆx0, y) + ǫDq/4 + 2ǫ2
629
+ 0
630
+
631
+ L−1
632
+ ∇h + 4D2
633
+ qL∇hǫ−2�
634
+ .
635
+ (28)
636
+ Moreover, the total number of evaluations of ∇h and proximal operator of p and q performed in Algo-
637
+ rithm 2 is no more than N, respectively.
638
+ Remark 3. Since ǫ0 ∈ (0, ǫ/2], one can observe from Theorem 2 that α = O(ǫ1/2), δ = O(ǫ−1/2),
639
+ K = O(ǫ−2), and N = O(ǫ−5/2 log(ǫ−1
640
+ 0 ǫ−1). Consequently, Algorithm 2 enjoys an operation complexity
641
+ of O(ǫ−5/2 log(ǫ−1
642
+ 0 ǫ−1)), measured by the amount of evaluations of ∇h and proximal operator of p and
643
+ q, for finding an ǫ-stationary point of the nonconvex-concave minimax problem (6).
644
+ 3
645
+ Unconstrained bilevel optimization
646
+ In this section, we consider an unconstrained bilevel optimization problem4
647
+ f ∗ = min
648
+ f(x, y)
649
+ s.t.
650
+ y ∈ Argmin
651
+ z
652
+ ˜f(x, z).
653
+ (29)
654
+ Assume that problem (29) has at least one optimal solution. In addition, f and ˜f satisfy the following
655
+ assumptions.
656
+ Assumption 3.
657
+ (i) f(x, y) = f1(x, y)+f2(x) and ˜f(x, y) = ˜f1(x, y)+ ˜f2(y) are continuous on X ×Y,
658
+ where f2 : Rn → R ∪ {∞} and ˜f2 : Rm → R ∪ {∞} are proper closed convex functions, ˜f1(x, ·) is
659
+ convex for any given x ∈ X, and f1, ˜f1 are respectively L∇f1- and L∇ ˜f1-smooth on X × Y with
660
+ X := dom f2 and Y := dom ˜f2.
661
+ (ii) The proximal operator associated with f2 and ˜f2 can be exactly evaluated.
662
+ (iii) The sets X and Y (namely, dom f2 and dom ˜f2) are compact.
663
+ For notational convenience, we define
664
+ Dx := max{∥u − v∥
665
+ ��u, v ∈ X},
666
+ Dy := max{∥u − v∥
667
+ ��u, v ∈ Y},
668
+ (30)
669
+ ˜fhi := max{ ˜f(x, y)|(x, y) ∈ X × Y},
670
+ ˜flow := min{ ˜f(x, y)|(x, y) ∈ X × Y},
671
+ (31)
672
+ flow := min{f(x, y)|(x, y) ∈ X × Y}.
673
+ (32)
674
+ 4For convenience, problem (29) is referred to as an unconstrained bilevel optimization problem since its lower level part
675
+ does not have an explicit constraint. Strictly speaking, it can be a constrained bilevel optimization problem. For example,
676
+ when part of f and/or ˜f is the indicator function of a closed convex set, (29) is essentially a constrained bilevel optimization
677
+ problem.
678
+ 7
679
+
680
+ By Assumption 3, one can observe that Dx, Dy, ˜fhi, ˜flow and flow are finite.
681
+ The goal of this subsection is to propose penalty methods for solving problem for solving (29). To
682
+ this end, we observe that problem (29) can be viewed as
683
+ min
684
+ x,y {f(x, y)| ˜f(x, y) ≤ min
685
+ z
686
+ ˜f(x, z)}.
687
+ (33)
688
+ Notice that ˜f(x, y) − minz ˜f(x, z) ≥ 0 for all x, y. Consequently, a natural penalty problem associated
689
+ with (33) is
690
+ min
691
+ x,y f(x, y) + ρ( ˜f(x, y) − min
692
+ z
693
+ ˜f(x, z)),
694
+ (34)
695
+ where ρ > 0 is a penalty parameter. We further observe that (34) is equivalent to the minimax problem
696
+ min
697
+ x,y max
698
+ z
699
+ Pρ(x, y, z),
700
+ where
701
+ Pρ(x, y, z) := f(x, y) + ρ( ˜f(x, y) − ˜f(x, z)).
702
+ (35)
703
+ In view of Assumption 3(i), Pρ can be rewritten as
704
+ Pρ(x, y, z) =
705
+
706
+ f1(x, y) + ρ ˜f1(x, y) − ρ ˜f1(x, z)
707
+
708
+ +
709
+
710
+ f2(x) + ρ ˜f2(y) − ρ ˜f2(z)
711
+
712
+ .
713
+ (36)
714
+ By this and Assumption 3, one can observe that Pρ enjoys the following nice properties.
715
+ • Pρ is the sum of smooth function f1(x, y)+ ρ ˜f1(x, y)− ρ ˜f1(x, z) with Lipschitz continuous gradient
716
+ and possibly nonsmooth function f2(x)+ρ ˜f2(y)−ρ ˜f2(z) with exactly computable proximal operator.
717
+ • Pρ is nonconvex in (x, y) but concave in z.
718
+ Thanks to the nice structure of Pρ, an approximate stationary point of the minimax problem (35) can
719
+ be found by Algorithm 2 proposed in Subsection 2.2.
720
+ Based on the above observations, we are now ready to propose penalty methods for the unconstrained
721
+ bilevel optimization problem (29) by solving either a sequence of or a single minimax problem in the
722
+ form of (35). In particular, we first propose an ideal penalty method for (29) by solving a sequence of
723
+ minimax problems (see Algorithm 3). Then we propose a practical penalty method for (29) by finding
724
+ an approximate stationary point of a single minimax problem (see Algorithm 4).
725
+ Algorithm 3 An ideal penalty method for problem (29)
726
+ Input: positive sequences {ρk} and {ǫk} with limk→∞(ρk, ǫk) = (∞, 0).
727
+ 1: for k = 0, 1, 2, . . . do
728
+ 2:
729
+ Find an ǫk-optimal solution (xk, yk, zk) of problem (35) with ρ = ρk.
730
+ 3: end for
731
+ The following theorem states a convergence result of Algorithm 3, whose proof is deferred to Section
732
+ 5.3.
733
+ Theorem 3 (Convergence of Algorithm 3). Suppose that Assumption 3 holds and that {(xk, yk, zk)}
734
+ is generated by Algorithm 3. Then any accumulation point of {(xk, yk)} is an optimal solution of problem
735
+ (29).
736
+ Notice that (35) is a nonconvex-concave minimax problem. It is typically hard to find an ǫ-optimal
737
+ solution of (35) for an arbitrary ǫ > 0. Consequently, Algorithm 3 is not implementable in general. We
738
+ next propose a practical penalty method for problem (29) by finding an approximate stationary point of
739
+ a single minimax problem (35) with a suitable choice of ρ.
740
+ Algorithm 4 A practical penalty method for problem (29)
741
+ Input: ε ∈ (0, 1/4], ρ = ε−1, (x0, y0) ∈ X × Y with ˜f(x0, y0) ≤ miny ˜f(x0, y) + ε.
742
+ 1: Call Algorithm 2 with ǫ ← ε, ǫ0 ← ε3/2, ˆx0 ← (x0, y0), ˆy0 ← y0, and L∇h ← L∇f1 + 2ε−1L∇ ˜
743
+ f1 to
744
+ find an ǫ-stationary point (xǫ, yǫ, zǫ) of problem (35) with ρ = ε−1.
745
+ 2: Output: (xǫ, yǫ).
746
+ 8
747
+
748
+ Remark 4. (i) The initial point (x0, y0) of Algorithm 4 can be found by an additional procedure. Indeed,
749
+ one can first choose any x0 ∈ X and then apply accelerated proximal gradient method [38] to the problem
750
+ miny ˜f(x0, y) for finding y0 ∈ Y such that ˜f(x0, y0) ≤ miny ˜f(x0, y) + ε; (ii) As seen from Theorem 2,
751
+ an ǫ-stationary point of (35) can be successfully found in step 1 of Algorithm 4 by applying Algorithm 2
752
+ to (35); (iii) For the sake of simplicity, a single subproblem of the form (35) with static penalty and
753
+ tolerance parameters is solved in Algorithm 4. Nevertheless, Algorithm 4 can be modified into a perhaps
754
+ practically more efficient algorithm by solving a sequence of subproblems of the form (35) with dynamic
755
+ penalty and tolerance parameters instead.
756
+ In order to characterize the approximate solution found by Algorithm 4, we next introduce a termi-
757
+ nology called an ε-KKT solution of problem (29).
758
+ Recall that problem (29) can be viewed as problem (33). In the spirit of classical constrained opti-
759
+ mization, one would naturally be interested in a KKT solution (x, y) of (33) or equivalently (29), namely,
760
+ (x, y) satisfies ˜f(x, y) ≤ minz ˜f(x, z) and moreover (x, y) is a stationary point of the problem
761
+ min
762
+ x′,y′ f(x′, y′) + ρ
763
+ � ˜f(x′, y′) − min
764
+ z′
765
+ ˜f(x′, z′)
766
+
767
+ (37)
768
+ for some ρ ≥ 0. Yet, due to the sophisticated problem structure, characterizing a stationary point of (37)
769
+ is generally difficult. On another hand, notice that problem (37) is equivalent to the minimax problem
770
+ min
771
+ x′,y′ max
772
+ z′
773
+ f(x′, y′) + ρ( ˜f(x′, y′) − ˜f(x′, z′)),
774
+ whose stationary point (x, y, z) according to Definition 2 satisfies
775
+ 0 ∈ ∂f(x, y) + ρ∂ ˜f(x, y) − (ρ∇x ˜f(x, z); 0),
776
+ 0 ∈ ρ∂z ˜f(x, z).
777
+ (38)
778
+ Based on this observation, we are instead interested in a (weak) KKT solution of problem (29) and its
779
+ inexact counterpart that are defined below.
780
+ Definition 3. The pair (x, y) is said to be a KKT solution of problem (29) if there exists (z, ρ) ∈ Rm×R+
781
+ such that (38) and ˜f(x, y) ≤ minz′ ˜f(x, z′) hold. In addition, for any ε > 0, (x, y) is said to be an ε-KKT
782
+ solution of problem (29) if there exists (z, ρ) ∈ Rm × R+ such that
783
+ dist
784
+
785
+ 0, ∂f(x, y) + ρ∂ ˜f(x, y) − (ρ∇x ˜f(x, z); 0)
786
+
787
+ ≤ ε,
788
+ dist
789
+
790
+ 0, ρ∂z ˜f(x, z)
791
+
792
+ ≤ ε,
793
+ ˜f(x, y) − min
794
+ z′
795
+ ˜f(x, z′) ≤ ε.
796
+ We are now ready to present a theorem regarding operation complexity of Algorithm 4, measured by
797
+ the amount of evaluations of ∇f1, ∇ ˜f1 and proximal operator of f2 and ˜f2, for finding an O(ε)-KKT
798
+ solution of (29), whose proof is deferred to Subsection 5.3.
799
+ Theorem 4 (Complexity of Algorithm 4). Suppose that Assumption 3 holds. Let f ∗, f, ˜f, Dx, Dy,
800
+ ˜fhi, ˜flow and flow be defined in (29), (30), (31) and (32), L∇f1 and L∇ ˜
801
+ f1 be given in Assumption 3, ε,
802
+ ρ, x0, y0 and zǫ be given in Algorithm 4, and
803
+ �L = L∇f1 + 2ε−1L∇ ˜
804
+ f1, ˆα = min
805
+
806
+ 1,
807
+
808
+ 4ε/(Dy�L)
809
+
810
+ ,
811
+ (39)
812
+ ˆδ = (2 + ˆα−1)(D2
813
+ x + D2
814
+ y)�L + max
815
+
816
+ ε/Dy, ˆα�L/4
817
+
818
+ D2
819
+ y,
820
+ �C =
821
+ 4 max
822
+
823
+ 1
824
+ 2�L, min
825
+
826
+ Dy
827
+ ε ,
828
+ 4
829
+ ˆα�L
830
+ �� �
831
+ ˆδ + 2ˆα−1(f ∗ − flow + ε−1( ˜fhi − ˜flow) + εDy/4 + �L(D2
832
+ x + D2
833
+ y))
834
+
835
+
836
+ (3�L + ε/(2Dy))2/ min{�L, ε/(2Dy)} + 3�L + ε/(2Dy)
837
+ �−2
838
+ ε3
839
+ ,
840
+ �K =
841
+
842
+ 16(1 + f(x0, y0) − flow + εDy/4)�Lε−2 + 32(1 + 4D2
843
+ y�L2ε−2)ε − 1
844
+
845
+ + ,
846
+
847
+ N =
848
+ ��
849
+ 96
850
+
851
+ 2(1 + (24�L + 4ε/Dy)�L−1)
852
+
853
+ + 2
854
+
855
+ max
856
+
857
+ 2,
858
+
859
+ Dy�Lε−1
860
+
861
+ × (( �
862
+ K + 1)(log �C)+ + �K + 1 + 2 �K log( �K + 1)).
863
+ 9
864
+
865
+ Then Algorithm 4 outputs an approximate solution (xǫ, yǫ) of (29) satisfying
866
+ dist
867
+
868
+ 0, ∂f(xǫ, yǫ) + ρ∂ ˜f(xǫ, yǫ) − (ρ∇x ˜f(xǫ, zǫ); 0)
869
+
870
+ ≤ ε,
871
+ dist
872
+
873
+ 0, ρ∂ ˜f(xǫ, zǫ)
874
+
875
+ ≤ ε,
876
+ (40)
877
+ ˜f(xǫ, yǫ) ≤ min
878
+ z
879
+ ˜f(xǫ, z) + ε
880
+
881
+ 1 + f(x0, y0) − flow + 2ε3(�L−1 + 4D2
882
+ y�Lε−2) + Dyε/4
883
+
884
+ ,
885
+ (41)
886
+ after at most �
887
+ N evaluations of ∇f1, ∇ ˜f1 and proximal operator of f2 and ˜f2, respectively.
888
+ Remark 5. One can observe from Theorem 4 that �L = O(ε−1), ˆα = O(ε), ˆδ = O(ε−2), �C = O(ε−11),
889
+ �K = O(ε−3), and �
890
+ N = O(ε−4 log ε−1). Consequently, Algorithm 4 enjoys an operation complexity of
891
+ O(ε−4 log ε−1), measured by the amount of evaluations of ∇f1, ∇ ˜f1 and proximal operator of f2 and ˜f2,
892
+ for finding an O(ε)-KKT solution (xǫ, yǫ) of (29) satisfying
893
+ dist
894
+
895
+ 0, ∂f(xǫ, yǫ) + ρ∂ ˜f(xǫ, yǫ) − (ρ∇x ˜f(xǫ, zǫ); 0)
896
+
897
+ ≤ ε,
898
+ dist
899
+
900
+ 0, ρ∂ ˜f(xǫ, zǫ)
901
+
902
+ ≤ ε,
903
+ ˜f(xǫ, yǫ) − min
904
+ z
905
+ ˜f(xǫ, z) = O(ε),
906
+ where zǫ is given in Algorithm 4 and ρ = ε−1.
907
+ 4
908
+ Constrained bilevel optimization
909
+ In this section, we consider a constrained bilevel optimization problem5
910
+ f ∗ = min
911
+ f(x, y)
912
+ s.t.
913
+ y ∈ Argmin
914
+ z
915
+ { ˜f(x, z)|˜g(x, z) ≤ 0},
916
+ (42)
917
+ where f and ˜f satisfy Assumption 3. Recall from Assumption 3 that X = dom f2 and Y = dom ˜f2. We
918
+ now make some additional assumptions for problem (42).
919
+ Assumption 4.
920
+ (i) f and ˜f are Lf- and L ˜
921
+ f-Lipschitz continuous on X × Y, respectively.
922
+ (ii) ˜g : Rn × Rm → Rl is L∇˜g-smooth and L˜g-Lipschitz continuous on X × Y.
923
+ (iii) ˜gi(x, ·) is convex and there exists ˆzx ∈ Y for each x ∈ X such that ˜gi(x, ˆzx) < 0 for all i = 1, 2, . . ., l
924
+ and G := min{−˜gi(x, ˆzx)|x ∈ X, i = 1, . . . , l} > 0.6
925
+ For notational convenience, we define
926
+ ˜f ∗(x) := min
927
+ z { ˜f(x, z)|˜g(x, z) ≤ 0},
928
+ (43)
929
+ ˜f ∗
930
+ hi := sup{ ˜f ∗(x)|x ∈ X},
931
+ (44)
932
+ ˜ghi := max{∥˜g(x, y)∥
933
+ ��(x, y) ∈ X × Y},
934
+ (45)
935
+ It then follows from Assumption 4(ii) that
936
+ ∥∇˜g(x, y)∥ ≤ L˜g
937
+ ∀(x, y) ∈ X × Y.
938
+ (46)
939
+ In addition, by Assumptions 3 and 4 and the compactness of X and Y, one can observe that ˜ghi and G
940
+ are finite. Besides, as will be shown in Lemma 6(ii), ˜f ∗
941
+ hi is finite.
942
+ The goal of this subsection is to propose penalty methods for solving problem (42). To this end, let
943
+ us introduce a penalty function for the lower level optimization problem y ∈ Argmin
944
+ z
945
+ { ˜f(x, z)|˜g(x, z) ≤ 0}
946
+ of (42), which is given by
947
+ �Pµ(x, z) = ˜f(x, z) + µ ∥[˜g(x, z)]+∥2
948
+ (47)
949
+ 5For convenience, problem (42) is referred to as a constrained bilevel optimization problem since its lower level part has
950
+ at least one explicit constraint.
951
+ 6The latter part of this assumption can be weakened to the one that the pointwise Slater’s condition holds for the lower
952
+ level part of (42), that is, there exists ˆzx ∈ Y such that ˜g(x, ˆzx) < 0 for each x ∈ X. Indeed, if G > 0, Assumption 4(iii)
953
+ clearly holds. Otherwise, one can solve the perturbed counterpart of (42) with ˜g(x, z) being replaced by ˜g(x, z) − ǫ for
954
+ some suitable ǫ > 0 instead, which satisfies Assumption 4(iii).
955
+ 10
956
+
957
+ for a penalty parameter µ > 0. Observe that problem (42) can be approximately solved as the uncon-
958
+ strained bilevel optimization problem
959
+ f ∗
960
+ µ = min
961
+ x,y
962
+
963
+ f(x, y)|y ∈ Argmin
964
+ z
965
+ �Pµ(x, z)
966
+
967
+ .
968
+ (48)
969
+ Further, by the study in Section 3, problem (48) can be approximately solved as the penalty problem
970
+ min
971
+ x,y f(x, y) + ρ
972
+
973
+ �Pµ(x, y) − min
974
+ z
975
+ �Pµ(x, z)
976
+
977
+ (49)
978
+ for some suitable ρ > 0. One can also observe that problem (49) is equivalent to the minimax problem
979
+ min
980
+ x,y max
981
+ z
982
+ Pρ,µ(x, y, z),
983
+ where
984
+ Pρ,µ(x, y, z) := f(x, y) + ρ( �Pµ(x, y) − �Pµ(x, z)).
985
+ (50)
986
+ In view of (47), (50) and Assumption 3(i), Pρ,µ can be rewritten as
987
+ Pρ,µ(x, y, z) =
988
+
989
+ f1(x, y) + ρ ˜f1(x, y) + ρµ ∥[˜g(x, y)]+∥2 − ρ ˜f1(x, z) − ρµ ∥[˜g(x, z)]+∥2 �
990
+ +
991
+
992
+ f2(x) + ρ ˜f2(y) − ρ ˜f2(z)
993
+
994
+ .
995
+ (51)
996
+ By this and Assumptions 3 and 4, one can observe that Pρ,µ enjoys the following nice properties.
997
+ • Pρ,µ is the sum of smooth function f1(x, y)+ρ ˜f1(x, y)+ρµ ∥[˜g(x, y)]+∥2−ρ ˜f1(x, z)−ρµ ∥[˜g(x, z)]+∥2
998
+ with Lipschitz continuous gradient and possibly nonsmooth function f2(x) + ρ ˜f2(y) − ρ ˜f2(z) with
999
+ exactly computable proximal operator;
1000
+ • Pρ,µ is nonconvex in (x, y) but concave in z.
1001
+ Due to the nice structure of Pρ,µ, an approximate stationary point of the minimax problem (50) can be
1002
+ found by Algorithm 2 proposed in Subsection 2.2.
1003
+ Based on the above observations, we are now ready to propose penalty methods for the constrained
1004
+ bilevel optimization problem (42) by solving a sequence of or a single minimax problem of the form (50).
1005
+ In particular, we first propose an ideal penalty method for (42) by solving a sequence of minimax problems
1006
+ (see Algorithm 5). Then we propose a practical penalty method for (42) by finding an approximate
1007
+ stationary point of a single minimax problem (see Algorithm 6).
1008
+ Algorithm 5 An ideal penalty method for problem (42)
1009
+ Input: positive sequences {ρk}, {µk} and {ǫk} with limk→∞(ρk, µk, ǫk) = (∞, ∞, 0).
1010
+ 1: for k = 0, 1, 2, . . . do
1011
+ 2:
1012
+ Find an ǫk-optimal solution (xk, yk, zk) of problem (50) with (ρ, µ) = (ρk, µk).
1013
+ 3: end for
1014
+ To study convergence of Algorithm 5, we make the following error bound assumption on the solution
1015
+ set of the lower level optimization problem of (42). This type of error bounds has been considered in the
1016
+ context of set-value mappings in the literature (e.g., see [14]).
1017
+ Assumption 5. There exists a non-decreasing function ω : R+ → R+ with limθ↓0 ω(θ) = 0 and ¯θ > 0
1018
+ such that dist(z, Sθ(x)) ≤ ω(θ) for all x ∈ X, z ∈ S0(x) and θ ∈ [0, ¯θ], where
1019
+ Sθ(x) := Argmin
1020
+ z
1021
+ { ˜f(x, z) : ∥[˜g(x, z)]+∥ ≤ θ}.
1022
+ We are now ready to state a convergence result of Algorithm 5, whose proof is deferred to Section
1023
+ 5.4.
1024
+ Theorem 5 (Convergence of Algorithm 5). Suppose that Assumptions 3-5 hold and that {(xk, yk, zk)}
1025
+ is generated by Algorithm 5. Then any accumulation point of {(xk, yk)} is an optimal solution of problem
1026
+ (42).
1027
+ Notice that (50) is a nonconvex-concave minimax problem. It is generally hard to find an ǫ-optimal
1028
+ solution of (50) for an arbitrary ǫ > 0. As a result, Algorithm 5 is generally not implementable. We next
1029
+ propose a practical penalty method for problem (42) by finding an approximate stationary point of (50)
1030
+ with a suitable choice of ρ and µ.
1031
+ 11
1032
+
1033
+ Algorithm 6 A practical penalty method for problem (42)
1034
+ Input: ε ∈ (0, 1/4], ρ = ε−1, µ = ε−2, (x0, y0) ∈ X × Y with �Pµ(x0, y0) ≤ miny �Pµ(x0, y) + ε.
1035
+ 1: Call Algorithm 2 with ǫ ← ε, ǫ0 ← ε5/2, ˆx0 ← (x0, y0), ˆy0 ← y0, and L∇h ← L∇f1 + 2ρL∇ ˜f1 +
1036
+ 4ρµ(˜ghiL∇˜g +L2
1037
+ ˜g) to find an ǫ-stationary point (xǫ, yǫ, zǫ) of problem (50) with ρ = ε−1 and µ = ε−2.
1038
+ 2: Output: (xǫ, yǫ).
1039
+ Remark 6. (i) The initial point (x0, y0) of Algorithm 6 can be found by the similar procedure as described
1040
+ in Remark 4 with ˜f being replaced by �Pµ; (ii) As seen from Theorem 2, an ǫ-stationary point of (50)
1041
+ can be successfully found in step 1 of Algorithm 6 by applying Algorithm 2 to (50); (iii) For the sake of
1042
+ simplicity, a single subproblem of the form (50) with static penalty and tolerance parameters is solved in
1043
+ Algorithm 6. Nevertheless, Algorithm 6 can be modified into a perhaps practically more efficient algorithm
1044
+ by solving a sequence of subproblems of the form (50) with dynamic penalty and tolerance parameters
1045
+ instead.
1046
+ In order to characterize the approximate solution found by Algorithm 6, we next introduce a termi-
1047
+ nology called an ε-KKT solution of problem (42).
1048
+ By the definition of ˜f ∗ in (43), problem (42) can be viewed as
1049
+ min
1050
+ x,y {f(x, y)| ˜f(x, y) ≤ ˜f ∗(x), ˜g(x, y) ≤ 0}.
1051
+ (52)
1052
+ Its associated Lagrangian function is given by
1053
+ L(x, y, ρ, λ) = f(x, y) + ρ( ˜f(x, y) − ˜f ∗(x)) + ⟨λ, ˜g(x, y)⟩.
1054
+ (53)
1055
+ In the spirit of classical constrained optimization, one would naturally be interested in a KKT solution
1056
+ (x, y) of (52) or equivalently (42), namely, (x, y) satisfies
1057
+ ˜f(x, y) ≤ ˜f ∗(x),
1058
+ ˜g(x, y) ≤ 0,
1059
+ ρ( ˜f(x, y) − ˜f ∗(x)) = 0,
1060
+ ⟨λ, ˜g(x, y)⟩ = 0,
1061
+ (54)
1062
+ and moreover (x, y) is a stationary point of the problem
1063
+ min
1064
+ x′,y′ L(x′, y′, ρ, λ)
1065
+ (55)
1066
+ for some ρ ≥ 0 and λ ∈ Rl
1067
+ +. Yet, due to the sophisticated problem structure, characterizing a stationary
1068
+ point of (55) is generally difficult. On another hand, notice from Lemma 6 and (53) that problem (55)
1069
+ is equivalent to the minimax problem
1070
+ min
1071
+ x′,y′,˜λ′ max
1072
+ z′
1073
+
1074
+ f(x′, y′) + ρ
1075
+ � ˜f(x′, y′) − ˜f(x′, z′) − ⟨˜λ′, ˜g(x′, z′)⟩
1076
+
1077
+ + ⟨λ, ˜g(x′, y′)⟩ + IRl
1078
+ +(˜λ′)
1079
+
1080
+ ,
1081
+ whose stationary point (x, y, ˜λ, z) according to Definition 2 satisfies
1082
+ 0 ∈ ∂f(x, y) + ρ∂ ˜f(x, y) − ρ(∇x ˜f(x, z) + ∇x˜g(x, z)˜λ; 0) + ∇˜g(x, y)λ,
1083
+ (56)
1084
+ 0 ∈ ρ(∂z ˜f(x, z) + ∇z˜g(x, z)˜λ),
1085
+ (57)
1086
+ ˜λ ∈ Rl
1087
+ +,
1088
+ ˜g(x, z) ≤ 0,
1089
+ ⟨˜λ, ˜g(x, z)⟩ = 0.
1090
+ (58)
1091
+ Based on this observation and also the fact that (54) is equivalent to
1092
+ ˜f(x, y) = ˜f ∗(x),
1093
+ ˜g(x, y) ≤ 0,
1094
+ ⟨λ, ˜g(x, y)⟩ = 0,
1095
+ (59)
1096
+ we are instead interested in a (weak) KKT solution of problem (42) and its inexact counterpart that are
1097
+ defined below.
1098
+ Definition 4. The pair (x, y) is said to be a KKT solution of problem (42) if there exists (z, ρ, λ, ˜λ) ∈
1099
+ Rm × R+ × Rl
1100
+ + × Rl
1101
+ + such that (56)-(59) hold. In addition, for any ε > 0, (x, y) is said to be an ε-KKT
1102
+ solution of problem (42) if there exists (z, ρ, λ, ˜λ) ∈ Rm × R+ × Rl
1103
+ + × Rl
1104
+ + such that
1105
+ dist
1106
+
1107
+ 0, ∂f(x, y) + ρ∂ ˜f(x, y) − ρ(∇x ˜f(x, z) + ∇x˜g(x, z)˜λ; 0) + ∇˜g(x, y)λ
1108
+
1109
+ ≤ ε,
1110
+ dist
1111
+
1112
+ 0, ρ(∂z ˜f(x, z) + ∇z˜g(x, z)˜λ)
1113
+
1114
+ ≤ ε,
1115
+ ∥[˜g(x, z)]+∥ ≤ ε,
1116
+ |⟨˜λ, ˜g(x, z)⟩| ≤ ε,
1117
+ | ˜f(x, y) − ˜f ∗(x)| ≤ ε,
1118
+ ∥[˜g(x, y)]+∥ ≤ ε,
1119
+ |⟨λ, ˜g(x, y)⟩| ≤ ε,
1120
+ where ˜f ∗ is defined in (43).
1121
+ 12
1122
+
1123
+ We are now ready to present an operation complexity of Algorithm 6, measured by the amount of
1124
+ evaluations of ∇f1, ∇ ˜f1, ∇˜g and proximal operator of f2 and ˜f2, for finding an O(ε)-KKT solution of
1125
+ (42), whose proof is deferred to Subsection 5.4.
1126
+ Theorem 6 (Complexity of Algorithm 6). Suppose that Assumptions 3 and 4 hold. Let f ∗, f, ˜f,
1127
+ ˜g, Dx, Dy, ˜fhi, ˜flow, flow, ˜f ∗, ˜f ∗
1128
+ hi, and ˜ghi be defined in (29), (30), (31), (32), (43), (44) and (45),
1129
+ L∇f1, L∇ ˜
1130
+ f1, L ˜
1131
+ f, L∇˜g, L˜g and G be given in Assumptions 3 and 4, ε, ρ, µ, x0, y0 and zǫ be given in
1132
+ Algorithm 6, and
1133
+ ˜λ = 2ε−1[˜g(xǫ, zǫ)]+,
1134
+ ˆλ = 2ε−3[˜g(xǫ, yǫ)]+,
1135
+ (60)
1136
+ �L = L∇f1 + 2ε−1L∇ ˜
1137
+ f1 + 4ε−3(˜ghiL∇˜g + L2
1138
+ ˜g),
1139
+ (61)
1140
+ ˜α = min
1141
+
1142
+ 1,
1143
+
1144
+ 4ε/(Dy�L)
1145
+
1146
+ , ˜δ = (2 + ˜α−1)(D2
1147
+ x + D2
1148
+ y)�L + max
1149
+
1150
+ ε/Dy, ˜α�L/4
1151
+
1152
+ D2
1153
+ y,
1154
+ �C =
1155
+ 4 max{1/(2�L), min{Dyε−1, 4/(˜α�L)}}
1156
+ [(3�L + ε/(2Dy))2/ min{�L, ε/(2Dy)} + 3�L + ε/(2Dy)]−2ε5
1157
+ ×
1158
+
1159
+ ˜δ + 2˜α−1[f ∗ − flow + 2ε−1( ˜fhi − ˜flow) + ε−3˜g2
1160
+ hi + εDy/4 + �L(D2
1161
+ x + D2
1162
+ y)]
1163
+
1164
+ ,
1165
+ �K =
1166
+
1167
+ 32(1 + f(x0, y0) − flow + εDy/4)�Lε−2 + 32ε3 �
1168
+ 1 + 4D2
1169
+ y�L2ε−2�
1170
+ − 1
1171
+
1172
+ + ,
1173
+
1174
+ N =
1175
+ ��
1176
+ 96
1177
+
1178
+ 2
1179
+
1180
+ 1 + (24�L + 4ε/Dy)�L−1��
1181
+ + 2
1182
+
1183
+ max
1184
+
1185
+ 2,
1186
+
1187
+ Dy�Lε−1
1188
+
1189
+ × [( �K + 1)(log �C)+ + �K + 1 + 2 �K log( �K + 1)].
1190
+ Then Algorithm 6 outputs an approximate solution (xǫ, yǫ) of (42) satisfying
1191
+ dist
1192
+
1193
+ ∂f(xǫ, yǫ) + ρ∂ ˜f(xǫ, yǫ) − ρ(∇x ˜f(xǫ, zǫ) + ∇x˜g(xǫ, zǫ)˜λ; 0) + ∇˜g(xǫ, yǫ)ˆλ
1194
+
1195
+ ≤ ε,
1196
+ (62)
1197
+ dist
1198
+
1199
+ 0, ρ(∂z ˜f(xǫ, zǫ) + ∇z˜g(xǫ, zǫ)˜λ)
1200
+
1201
+ ≤ ε,
1202
+ (63)
1203
+ ∥[˜g(xǫ, zǫ)]+∥ ≤ ε2G−1Dy(ε2 + L ˜
1204
+ f)/2,
1205
+ (64)
1206
+ |⟨˜λ, ˜g(xǫ, zǫ)⟩| ≤ ε2G−2D2
1207
+ y(ρ−1ǫ + L ˜
1208
+ f)2/2,
1209
+ (65)
1210
+ | ˜f(xǫ, yǫ) − ˜f ∗(xǫ)| ≤ max
1211
+
1212
+ ε
1213
+
1214
+ 1 + f(x0, y0) − flow + 2ε5(�L−1 + 4D2
1215
+ y�Lε−2) + Dyε/4
1216
+
1217
+ ,
1218
+ ε2G−2D2
1219
+ yL ˜
1220
+ f(ε2 + εLf + L ˜
1221
+ f)/2
1222
+
1223
+ ,
1224
+ (66)
1225
+ ∥[˜g(xǫ, yǫ)]+∥ ≤ ε2G−1Dy(ε2 + εLf + L ˜
1226
+ f)/2,
1227
+ (67)
1228
+ |⟨ˆλ, ˜g(xǫ, yǫ)⟩| ≤ εG−2D2
1229
+ y(ε2 + εLf + L ˜
1230
+ f)2/2,
1231
+ (68)
1232
+ after at most �
1233
+ N evaluations of ∇f1, ∇ ˜f1, ∇˜g and proximal operator of f2 and ˜f2, respectively.
1234
+ Remark 7. One can observe from Theorem 6 that �L = O(ε−3), ˜α = O(ε2), ˜δ = O(ε−5), �C = O(ε−23),
1235
+ �K = O(ε−5), and �
1236
+ N = O(ε−7 log ε−1). Consequently, Algorithm 6 enjoys an operation complexity of
1237
+ O(ε−7 log ε−1), measured by the amount of evaluations of ∇f1, ∇ ˜f1, ∇˜g and proximal operator of f2
1238
+ and ˜f2, for finding an O(ε)-KKT solution (xǫ, yǫ) of (42) satisfying
1239
+ dist
1240
+
1241
+ 0, ∂f(xǫ, yǫ) + ρ∂ ˜f(xǫ, yǫ) − ρ(∇x ˜f(xǫ, zǫ) + ∇x˜g(xǫ, zǫ)˜λ; 0) + ∇˜g(xǫ, yǫ)ˆλ
1242
+
1243
+ ≤ ε,
1244
+ dist
1245
+
1246
+ 0, ρ(∂z ˜f(xǫ, zǫ) + ∇z˜g(xǫ, zǫ)˜λ)
1247
+
1248
+ ≤ ε,
1249
+ ∥[˜g(xǫ, zǫ)]+∥ = O(ε2),
1250
+ |⟨˜λ, ˜g(xǫ, zǫ)⟩| = O(ε2),
1251
+ | ˜f(xǫ, yǫ) − ˜f ∗(xǫ)| = O(ε),
1252
+ ∥[˜g(xǫ, yǫ)]+∥ = O(ε2),
1253
+ |⟨ˆλ, ˜g(xǫ, yǫ)⟩| = O(ε),
1254
+ where ˜f ∗ is defined in (43), ˆλ, ˜λ ∈ Rl
1255
+ + are defined in (60), zǫ is given in Algorithm 6 and ρ = ε−1.
1256
+ 5
1257
+ Proof of the main results
1258
+ In this section we provide a proof of our main results presented in Sections 2, 3 and 4, which are
1259
+ particularly Theorems 1-6.
1260
+ 13
1261
+
1262
+ 5.1
1263
+ Proof of the main results in Subsection 2.1
1264
+ In this subsection we prove Theorem 1. Before proceeding, we establish a lemma below.
1265
+ Lemma 1. Suppose that Assumptions 1 and 2 hold. Let ¯H∗, ¯Hlow, ϑ0 and ¯δ be defined in (8), (10),
1266
+ (14) and (16), and ¯α be given in Algorithm 1. Then we have
1267
+ ϑ0 ≤ ¯δ + 2¯α−1 � ¯H∗ − ¯Hlow
1268
+
1269
+ .
1270
+ (69)
1271
+ Proof. By (8), (10), (11) and (12), one has
1272
+ G(¯z0, ¯y0)
1273
+ (12)
1274
+ =
1275
+ sup
1276
+ x
1277
+
1278
+ ⟨x, ¯z0⟩ − p(x) − ˆh(x, ¯y0) + q(¯y0)
1279
+
1280
+ (11)
1281
+ =
1282
+ max
1283
+ x∈dom p
1284
+
1285
+ ⟨x, ¯z0⟩ − p(x) − ¯h(x, ¯y0) + σx
1286
+ 2 ∥x∥2 − σy
1287
+ 2 ∥¯y0∥2 + q(¯y0)
1288
+
1289
+ (8)(10)
1290
+
1291
+ max
1292
+ x∈dom p
1293
+
1294
+ ⟨x, ¯z0⟩ + σx
1295
+ 2 ∥x∥2�
1296
+ − σy
1297
+ 2 ∥¯y0∥2 − ¯Hlow
1298
+ =
1299
+ max
1300
+ x∈dom p
1301
+ σx
1302
+ 2 ∥x + σ−1
1303
+ x ¯z0∥2 − σ−1
1304
+ x
1305
+ 2 ∥¯z0∥2 − σy
1306
+ 2 ∥¯y0∥2 − ¯Hlow
1307
+ ≤ σxD2
1308
+ p
1309
+ 2
1310
+ − σ−1
1311
+ x
1312
+ 2 ∥¯z0∥2 − σy
1313
+ 2 ∥¯y0∥2 − ¯Hlow,
1314
+ (70)
1315
+ where the last inequality follows from (9) and the fact that z0 ∈ −σxdom p.
1316
+ Recall that (x∗, y∗) is the optimal solution of (8) and z∗ = −σxx∗. It follows from (8), (11) and (12)
1317
+ that
1318
+ G(z∗, y∗)
1319
+ (12)
1320
+ =
1321
+ sup
1322
+ x
1323
+
1324
+ ⟨x, z∗⟩ − p(x) − ˆh(x, y∗) + q(y∗)
1325
+
1326
+ ≥ ⟨x∗, z∗⟩ − p(x∗) − ˆh(x∗, y∗) + q(y∗)
1327
+ (11)
1328
+ = ⟨x∗, z∗⟩ + σx
1329
+ 2 ∥x∗∥2 − σy
1330
+ 2 ∥y∗∥2 − p(x∗) − ¯h(x∗, y∗) + q(y∗)
1331
+ =
1332
+ − σ−1
1333
+ x
1334
+ 2 ∥z∗∥2 − σy
1335
+ 2 ∥y∗∥2 − ¯H∗,
1336
+ where the last equality follows from (8), the definition of (x∗, y∗), and z∗ = −σxx∗. This together with
1337
+ (13) and (70) implies that
1338
+ P(¯z0, ¯y0) − P(z∗, y∗) = σ−1
1339
+ x
1340
+ 2 ∥¯z0∥2 + σy
1341
+ 2 ∥¯y0∥2 + G(¯z0, ¯y0) − σ−1
1342
+ x
1343
+ 2 ∥z∗∥2 − σy
1344
+ 2 ∥y∗∥2 − G(z∗, y∗)
1345
+ ≤ σxD2
1346
+ p/2 − ¯Hlow + ¯H∗.
1347
+ Notice from Algorithm 1 that z0 = z0
1348
+ f = ¯z0 ∈ −σxdom p and y0 = y0
1349
+ f = ¯y0 ∈ dom q.
1350
+ By these,
1351
+ z∗ = −σxx∗, (9), (14), and the above inequality, one has
1352
+ ϑ0
1353
+ (14)
1354
+ = η−1
1355
+ z ∥¯z0 − z∗∥2 + η−1
1356
+ y ∥¯y0 − y∗∥2 + 2¯α−1(P(¯z0, ¯y0) − P(z∗, y∗))
1357
+ ≤ η−1
1358
+ z σ2
1359
+ xD2
1360
+ p + η−1
1361
+ y D2
1362
+ q + 2¯α−1 �
1363
+ σxD2
1364
+ p/2 − ¯Hlow + ¯H∗�
1365
+ = η−1
1366
+ z σ2
1367
+ xD2
1368
+ p + ¯α−1σxD2
1369
+ p + η−1
1370
+ y D2
1371
+ q + 2¯α−1 � ¯H∗ − ¯Hlow
1372
+
1373
+ .
1374
+ Hence, the conclusion follows from this, (16), ηz = σx/2 and ηy = min {1/(2σy), 4/(¯ασx)}.
1375
+ We are now ready to prove Theorem 1.
1376
+ Proof of Theorem 1. Suppose for contradiction that Algorithm 1 runs for more than ¯K outer itera-
1377
+ tions, where ¯K is given in (17). By this and Algorithm 1, one can assert that (15) does not hold for
1378
+ k = ¯K − 1. On the other hand, by (17) and [29, Theorem 3], one has
1379
+ ∥(x
1380
+ ¯
1381
+ K, y
1382
+ ¯
1383
+ K) − (x∗, y∗)∥ ≤ (ˆζ−1 + L∇¯h)−1τ/2,
1384
+ (71)
1385
+ where (x∗, y∗) is the optimal solution of problem (8) and ˆζ is an input of Algorithm 1. Notice from
1386
+ Algorithm 1 that (ˆx ¯
1387
+ K, ˆy ¯
1388
+ K) results from the forward-backward splitting (FBS) step applied to the strongly
1389
+ monotone inclusion problem 0 ∈ (∇x¯h(x, y), −∇y¯h(x, y)) + (∂p(x), ∂q(y)) at the point (x ¯
1390
+ K, y ¯
1391
+ K). It then
1392
+ 14
1393
+
1394
+ follows from this, ˆζ = min{σx, σy}/L2
1395
+ ∇¯h (see Algorithm 1), and the contraction property of FBS [5,
1396
+ Corollary 2.5] that ∥(ˆx ¯
1397
+ K, ˆy ¯
1398
+ K) − (x∗, y∗)∥ ≤ ∥(x ¯
1399
+ K, y ¯
1400
+ K) − (x∗, y∗)∥. Using this and (71), we have
1401
+ ∥ˆζ−1(x
1402
+ ¯
1403
+ K − ˆx
1404
+ ¯
1405
+ K, ˆy
1406
+ ¯
1407
+ K − y
1408
+ ¯
1409
+ K) − (∇¯h(x
1410
+ ¯
1411
+ K, y
1412
+ ¯
1413
+ K) − ∇¯h(ˆx
1414
+ ¯
1415
+ K, ˆy
1416
+ ¯
1417
+ K))∥
1418
+ ≤ ˆζ−1∥(x
1419
+ ¯
1420
+ K, y
1421
+ ¯
1422
+ K) − (ˆx
1423
+ ¯
1424
+ K, ˆy
1425
+ ¯
1426
+ K)∥ + ∥∇¯h(x
1427
+ ¯
1428
+ K, y
1429
+ ¯
1430
+ K) − ∇¯h(ˆx
1431
+ ¯
1432
+ K, ˆy
1433
+ ¯
1434
+ K)∥
1435
+ ≤ (ˆζ−1 + L∇¯h)∥(x
1436
+ ¯
1437
+ K, y
1438
+ ¯
1439
+ K) − (ˆx
1440
+ ¯
1441
+ K, ˆy
1442
+ ¯
1443
+ K)∥
1444
+ ≤ (ˆζ−1 + L∇¯h)(∥(x
1445
+ ¯
1446
+ K, y
1447
+ ¯
1448
+ K) − (x∗, y∗)∥ + ∥(ˆx
1449
+ ¯
1450
+ K, ˆy
1451
+ ¯
1452
+ K) − (x∗, y∗)∥)
1453
+ ≤ 2(ˆζ−1 + L∇¯h)∥(x
1454
+ ¯
1455
+ K, y
1456
+ ¯
1457
+ K) − (x∗, y∗)∥
1458
+ (71)
1459
+ ≤ τ,
1460
+ where the second inequality uses the fact that ¯h is L∇¯h-smooth on dom p × dom q. It follows that (15)
1461
+ holds for k = ¯K − 1, which contradicts the above assertion. Hence, Algorithm 1 must terminate in at
1462
+ most ¯K outer iterations.
1463
+ We next show that the output of Algorithm 1 is a τ-stationary point of (8). To this end, suppose
1464
+ that Algorithm 1 terminates at some iteration k at which (15) is satisfied. Then by (4) and the definition
1465
+ of ˆxk+1 and ˆyk+1 (see steps 23 and 24 of Algorithm 1), one has
1466
+ 0 ∈ ˆζ∂p(ˆxk+1) + ˆxk+1 − xk+1 + ˆζ∇x¯h(xk+1, yk+1),
1467
+ 0 ∈ ˆζ∂q(ˆyk+1) + ˆyk+1 − yk+1 − ˆζ∇y¯h(xk+1, yk+1),
1468
+ which yield
1469
+ ˆζ−1(xk+1 − ˆxk+1) − ∇x¯h(xk+1, yk+1) ∈ ∂p(ˆxk+1), ˆζ−1(yk+1 − ˆyk+1) + ∇y¯h(xk+1, yk+1) ∈ ∂q(ˆyk+1).
1470
+ These together with the definition of ¯H in (8) imply that
1471
+ ∇x¯h(ˆxk+1, ˆyk+1) + ˆζ−1(xk+1 − ˆxk+1) − ∇x¯h(xk+1, yk+1) ∈ ∂x ¯H(ˆxk+1, ˆyk+1),
1472
+ ∇y¯h(ˆxk+1, ˆyk+1) − ˆζ−1(yk+1 − ˆyk+1) − ∇y¯h(xk+1, yk+1) ∈ ∂y ¯H(ˆxk+1, ˆyk+1).
1473
+ Using these and (15), we obtain
1474
+ dist(0, ∂x ¯H(ˆxk+1, ˆyk+1))2 + dist(0, ∂y ¯H(ˆxk+1, ˆyk+1))2
1475
+ ≤ ∥ˆζ−1(xk+1 − ˆxk+1) + ∇x¯h(ˆxk+1, ˆyk+1) − ∇x¯h(xk+1, yk+1)∥2
1476
+ + ∥ˆζ−1(ˆyk+1 − yk+1) + ∇y¯h(ˆxk+1, ˆyk+1) − ∇y¯h(xk+1, yk+1)∥2
1477
+ = ∥ˆζ−1(xk+1 − ˆxk+1, ˆyk+1 − yk+1) − (∇¯h(xk+1, yk+1) − ∇¯h(ˆxk+1, ˆyk+1))∥2 (15)
1478
+ ≤ τ2,
1479
+ which implies that dist(0, ∂x ¯H(ˆxk+1, ˆyk+1)) ≤ τ and dist(0, ∂y ¯H(ˆxk+1, ˆyk+1)) ≤ τ. It then follows from
1480
+ these and Definition 2 that the output (ˆxk+1, ˆyk+1) of Algorithm 1 is a τ-stationary point of (8).
1481
+ Finally, we show that the total number of evaluations of ∇¯h and proximal operator of p and q
1482
+ performed in Algorithm 1 is no more than ¯N, respectively.
1483
+ Indeed, notice from Algorithm 1 that
1484
+ ¯α = min
1485
+
1486
+ 1,
1487
+
1488
+ 8σy/σx
1489
+
1490
+ , which implies that 2/¯α = max{2,
1491
+
1492
+ σx/(2σy)} and ¯α ≤
1493
+
1494
+ 8σy/σx. By these,
1495
+ one has
1496
+ max
1497
+ � 2
1498
+ ¯α, ¯ασx
1499
+ 4σy
1500
+
1501
+ ≤ max
1502
+
1503
+ 2,
1504
+ � σx
1505
+ 2σy
1506
+ ,
1507
+
1508
+ 8σy
1509
+ σx
1510
+ σx
1511
+ 4σy
1512
+
1513
+ = max
1514
+
1515
+ 2,
1516
+ � σx
1517
+ 2σy
1518
+
1519
+ .
1520
+ (72)
1521
+ In addition, by [29, Lemma 4], the number of inner iterations performed in each outer iteration of
1522
+ Algorithm 1 is at most
1523
+ T =
1524
+
1525
+ 48
1526
+
1527
+ 2
1528
+
1529
+ 1 + 8L∇¯hσ−1
1530
+ x
1531
+ ��
1532
+ − 1.
1533
+ Then one can observe that the number of evaluations of ∇¯h and proximal operator of p and q performed
1534
+ 15
1535
+
1536
+ in Algorithm 1 is at most
1537
+ (2T + 3) ¯K ≤
1538
+ ��
1539
+ 96
1540
+
1541
+ 2
1542
+
1543
+ 1 + 8L∇¯hσ−1
1544
+ x
1545
+ ��
1546
+ + 2
1547
+ � �
1548
+ max
1549
+ � 2
1550
+ ¯α, ¯ασx
1551
+ 4σy
1552
+
1553
+ log 4 max{ηzσ−2
1554
+ x , ηy}ϑ0
1555
+ (ˆζ−1 + L∇¯h)−2τ 2
1556
+
1557
+ +
1558
+ (72)
1559
+
1560
+ ��
1561
+ 96
1562
+
1563
+ 2
1564
+
1565
+ 1 + 8L∇¯hσ−1
1566
+ x
1567
+ ��
1568
+ + 2
1569
+ � �
1570
+ max
1571
+
1572
+ 2,
1573
+ � σx
1574
+ 2σy
1575
+
1576
+ log 4 max{ηzσ−2
1577
+ x , ηy}ϑ0
1578
+ (ˆζ−1 + L∇¯h)−2τ 2
1579
+
1580
+ +
1581
+
1582
+ ��
1583
+ 96
1584
+
1585
+ 2
1586
+
1587
+ 1 + 8L∇¯hσ−1
1588
+ x
1589
+ ��
1590
+ + 2
1591
+
1592
+ ×
1593
+
1594
+ max
1595
+
1596
+ 2,
1597
+ � σx
1598
+ 2σy
1599
+
1600
+ log 4 max{1/(2σx), min {1/(2σy), 4/(¯ασx)}} ϑ0
1601
+ (L2
1602
+ ∇¯h/ min{σx, σy} + L∇¯h)−2τ 2
1603
+
1604
+ +
1605
+ (69)(18)
1606
+
1607
+ ¯N,
1608
+ where the second last inequality follows from the definition of ηy, ηz and ˆζ in Algorithm 1. Hence, the
1609
+ conclusion holds as desired.
1610
+ 5.2
1611
+ Proof of the main results in Subsection 2.2
1612
+ In this subsection we prove Theorem 2.
1613
+ Before proceeding, let {(xk, yk)}k∈K denote all the iterates
1614
+ generated by Algorithm 2, where K is a subset of consecutive nonnegative integers starting from 0. Also,
1615
+ we define K − 1 = {k − 1 : k ∈ K}. We first establish two lemmas and then use them to prove Theorem
1616
+ 2 subsequently.
1617
+ Lemma 2. Suppose that Assumption 1 holds. Let {(xk, yk)}k∈K be generated by Algorithm 2, H∗, Dp,
1618
+ Dq, Hlow, α, δ be defined in (6), (9), (23), (24) and (25), L∇h be given in Assumption 1, ǫ, ǫk be given
1619
+ in Algorithm 2, and
1620
+ Nk =
1621
+ ��
1622
+ 96
1623
+
1624
+ 2
1625
+
1626
+ 1 + (24L∇h + 4ǫ/Dq) L−1
1627
+ ∇h
1628
+ ��
1629
+ + 2
1630
+
1631
+ ×
1632
+
1633
+ max
1634
+
1635
+ 2,
1636
+
1637
+ DqL∇h
1638
+ ǫ
1639
+
1640
+ × log
1641
+ 4 max
1642
+
1643
+ 1
1644
+ 2L∇h , min
1645
+
1646
+ Dq
1647
+ ǫ ,
1648
+ 4
1649
+ αL∇h
1650
+ �� �
1651
+ δ + 2α−1(H∗ − Hlow + ǫDq/4 + L∇hD2
1652
+ p)
1653
+
1654
+ [(3L∇h + ǫ/(2Dq))2/ min{L∇h, ǫ/(2Dq)} + 3L∇h + ǫ/(2Dq)]−2 ǫ2
1655
+ k
1656
+
1657
+ +
1658
+ .
1659
+ (73)
1660
+ Then for all 0 ≤ k ∈ K−1, (xk+1, yk+1) is an ǫk-stationary point of (20). Moreover, the total number of
1661
+ evaluations of ∇h and proximal operator of p and q performed at iteration k of Algorithm 2 for generating
1662
+ (xk+1, yk+1) is no more than Nk, respectively.
1663
+ Proof. Let (x∗, y∗) be an optimal solution of (6). Recall that H, Hk and hk are given in (6), (20) and
1664
+ (21), respectively. Then we have
1665
+ Hk,∗ := min
1666
+ x max
1667
+ y
1668
+ Hk(x, y) = min
1669
+ x max
1670
+ y
1671
+
1672
+ H(x, y) −
1673
+ ǫ
1674
+ 4Dq
1675
+ ∥y − ˆy0∥2 + L∇h∥x − xk∥2
1676
+
1677
+ ≤ max
1678
+ y {H(x∗, y) + L∇h∥x∗ − xk∥2}
1679
+ (6)(9)
1680
+
1681
+ H∗ + L∇hD2
1682
+ p.
1683
+ (74)
1684
+ Moreover, by (9) and (23), one has
1685
+ Hk,low :=
1686
+ min
1687
+ (x,y)∈dom p×dom q Hk(x, y) =
1688
+ min
1689
+ (x,y)∈dom p×dom q
1690
+
1691
+ H(x, y) −
1692
+ ǫ
1693
+ 4Dq
1694
+ ∥y − ˆy0∥2 + L∇h∥x − xk∥2
1695
+
1696
+ (23)
1697
+ ≥ Hlow − max
1698
+ y∈dom q
1699
+ ǫ
1700
+ 4Dq
1701
+ ∥y − ˆy0∥2 (9)
1702
+ ≥ Hlow − ǫDq/4.
1703
+ (75)
1704
+ In addition, by Assumption 1 and the definition of hk in (21), it is not hard to verify that hk(x, y) is
1705
+ L∇h-strongly-convex in x, ǫ/(2Dq)-strongly-concave in y, and (3L∇h + ǫ/(2Dq))-smooth on its domain.
1706
+ Also, recall from Remark 2 that (xk+1, yk+1) results from applying Algorithm 1 to problem (20). The
1707
+ conclusion of this lemma then follows by using (74) and (75) and applying Theorem 1 to (20) with τ = ǫk,
1708
+ σx = L∇h, σy = ǫ/(2Dq), L∇¯h = 3L∇h + ǫ/(2Dq), ¯α = α, ¯δ = δ, ¯Hlow = Hk,low, and ¯H∗ = Hk,∗.
1709
+ 16
1710
+
1711
+ Lemma 3. Suppose that Assumption 1 holds. Let {xk}k∈K be generated by Algorithm 2, H, H∗ and
1712
+ Dq be defined in (6) and (9), L∇h be given in Assumption 1, and ǫ, ǫ0 and ˆx0 be given in Algorithm 2.
1713
+ Then for all 0 ≤ K ∈ K − 1, we have
1714
+ min
1715
+ 0≤k≤K ∥xk+1 − xk∥ ≤ maxy H(ˆx0, y) − H∗ + ǫDq/4
1716
+ L∇h(K + 1)
1717
+ + 2ǫ2
1718
+ 0(1 + 4D2
1719
+ qL2
1720
+ ∇hǫ−2)
1721
+ L2
1722
+ ∇h(K + 1)
1723
+ ,
1724
+ (76)
1725
+ max
1726
+ y
1727
+ H(xK+1, y) ≤ max
1728
+ y
1729
+ H(ˆx0, y) + ǫDq/4 + 2ǫ2
1730
+ 0
1731
+
1732
+ L−1
1733
+ ∇h + 4D2
1734
+ qL∇hǫ−2�
1735
+ .
1736
+ (77)
1737
+ Proof. For convenience of the proof, let
1738
+ H∗
1739
+ ǫ (x) = max
1740
+ y
1741
+
1742
+ H(x, y) − ǫ∥y − ˆy0∥2/(4Dq)
1743
+
1744
+ ,
1745
+ (78)
1746
+ H∗
1747
+ k(x) = max
1748
+ y
1749
+ Hk(x, y),
1750
+ yk+1
1751
+
1752
+ = arg max
1753
+ y
1754
+ Hk(xk+1, y).
1755
+ (79)
1756
+ One can observe from these, (20) and (21) that
1757
+ H∗
1758
+ k(x) = H∗
1759
+ ǫ (x) + L∇h∥x − xk∥2.
1760
+ (80)
1761
+ By this and Assumption 1, one can also see that H∗
1762
+ k is L∇h-strongly convex on dom p. In addition,
1763
+ recall from Lemma 2 that (xk+1, yk+1) is an ǫk-stationary point of problem (20) for all 0 ≤ k ∈ K − 1.
1764
+ It then follows from Definition 2 that there exist some u ∈ ∂xHk(xk+1, yk+1) and v ∈ ∂yHk(xk+1, yk+1)
1765
+ with ∥u∥ ≤ ǫk and ∥v∥ ≤ ǫk.
1766
+ Also, by (79), one has 0 ∈ ∂yHk(xk+1, yk+1
1767
+
1768
+ ), which together with
1769
+ v ∈ ∂yHk(xk+1, yk+1) and ǫ/(2Dq)-strong concavity of Hk(xk+1, ·), implies that ⟨−v, yk+1 − yk+1
1770
+
1771
+ ⟩ ≥
1772
+ ǫ∥yk+1 − yk+1
1773
+
1774
+ ∥2/(2Dq). This and ∥v∥ ≤ ǫk yield
1775
+ ∥yk+1 − yk+1
1776
+
1777
+ ∥ ≤ 2ǫkDq/ǫ.
1778
+ (81)
1779
+ In addition, by u ∈ ∂xHk(xk+1, yk+1), (20) and (21), one has
1780
+ u ∈ ∇xh(xk+1, yk+1) + ∂p(xk+1) + 2L∇h(xk+1 − xk).
1781
+ (82)
1782
+ Also, observe from (20), (21) and (79) that
1783
+ ∂H∗
1784
+ k(xk+1) = ∇xh(xk+1, yk+1
1785
+
1786
+ ) + ∂p(xk+1) + 2L∇h(xk+1 − xk),
1787
+ which together with (82) yields
1788
+ u + ∇xh(xk+1, yk+1
1789
+
1790
+ ) − ∇xh(xk+1, yk+1) ∈ ∂H∗
1791
+ k(xk+1).
1792
+ By this and L∇h-strong convexity of H∗
1793
+ k, one has
1794
+ H∗
1795
+ k(xk) ≥ H∗
1796
+ k(xk+1) + ⟨u + ∇xh(xk+1, yk+1
1797
+
1798
+ ) − ∇xh(xk+1, yk+1), xk − xk+1⟩ + L∇h∥xk − xk+1∥2/2. (83)
1799
+ Using this, (80), (81), (83), ∥u∥ ≤ ǫk, and the Lipschitz continuity of ∇h, we obtain
1800
+ H∗
1801
+ ǫ (xk) − H∗
1802
+ ǫ (xk+1)
1803
+ (80)
1804
+ = H∗
1805
+ k(xk) − H∗
1806
+ k(xk+1) + L∇h∥xk − xk+1∥2
1807
+ (83)
1808
+ ≥ ⟨u + ∇xh(xk+1, yk+1
1809
+
1810
+ ) − ∇xh(xk+1, yk+1), xk − xk+1⟩ + 3L∇h∥xk − xk+1∥2/2
1811
+
1812
+
1813
+ − ∥u + ∇xh(xk+1, yk+1
1814
+
1815
+ ) − ∇xh(xk+1, yk+1)∥∥xk − xk+1∥ + L∇h∥xk − xk+1∥2/2
1816
+
1817
+ + L∇h∥xk − xk+1∥2
1818
+ ≥ −(2L∇h)−1∥u + ∇xh(xk+1, yk+1
1819
+
1820
+ ) − ∇xh(xk+1, yk+1)∥2 + L∇h∥xk − xk+1∥2
1821
+ ≥ −L−1
1822
+ ∇h∥u∥2 − L−1
1823
+ ∇h∥∇xh(xk+1, yk+1
1824
+
1825
+ ) − ∇xh(xk+1, yk+1)∥2 + L∇h∥xk − xk+1∥2
1826
+ ≥ −L−1
1827
+ ∇hǫ2
1828
+ k − L∇h∥yk+1 − yk+1
1829
+
1830
+ ∥2 + L∇h∥xk − xk+1∥2
1831
+ (81)
1832
+ ≥ −(L−1
1833
+ ∇h + 4D2
1834
+ qL∇hǫ−2)ǫ2
1835
+ k + L∇h∥xk − xk+1∥2,
1836
+ where the second and fourth inequalities follow from Cauchy-Schwartz inequality, and the third inequal-
1837
+ ity is due to Young’s inequality, and the fifth inequality follows from L∇h-Lipschitz continuity of ∇h.
1838
+ Summing up the above inequality for k = 0, 1, . . ., K yields
1839
+ L∇h
1840
+ K
1841
+
1842
+ k=0
1843
+ ∥xk − xk+1∥2 ≤ H∗
1844
+ ǫ (x0) − H∗
1845
+ ǫ (xK+1) + (L−1
1846
+ ∇h + 4D2
1847
+ qL∇hǫ−2)
1848
+ K
1849
+
1850
+ k=0
1851
+ ǫ2
1852
+ k.
1853
+ (84)
1854
+ 17
1855
+
1856
+ In addition, it follows from (6), (9) and (78) that
1857
+ H∗
1858
+ ǫ (xK+1) = max
1859
+ y
1860
+
1861
+ H(xK+1, y) − ǫ∥y − ˆy0∥2/(4Dq)
1862
+
1863
+ ≥ min
1864
+ x max
1865
+ y
1866
+ H(x, y) − ǫDq/4 = H∗ − ǫDq/4,
1867
+ H∗
1868
+ ǫ (x0) = max
1869
+ y
1870
+
1871
+ H(x0, y) − ǫ∥y − ˆy0∥2/(4Dq)
1872
+
1873
+ ≤ max
1874
+ y
1875
+ H(x0, y).
1876
+ (85)
1877
+ These together with (84) yield
1878
+ L∇h(K + 1) min
1879
+ 0≤k≤K ∥xk+1 − xk∥2 ≤ L∇h
1880
+ K
1881
+
1882
+ k=0
1883
+ ∥xk − xk+1∥2
1884
+ ≤ max
1885
+ y
1886
+ H(x0, y) − H∗ + ǫDq/4 + (L−1
1887
+ ∇h + 4D2
1888
+ qL∇hǫ−2)
1889
+ K
1890
+
1891
+ k=0
1892
+ ǫ2
1893
+ k,
1894
+ which together with x0 = ˆx0, ǫk = ǫ0(k + 1)−1 and �K
1895
+ k=0(k + 1)−2 < 2 implies that (76) holds.
1896
+ Finally, we show that (77) holds. Indeed, it follows from (9), (78), (84), (85), ǫk = ǫ0(k + 1)−1, and
1897
+ �K
1898
+ k=0(k + 1)−2 < 2 that
1899
+ max
1900
+ y
1901
+ H(xK+1, y)
1902
+ (9)
1903
+
1904
+ max
1905
+ y
1906
+
1907
+ H(xK+1, y) − ǫ∥y − ˆy0∥2/(4Dq)
1908
+
1909
+ + ǫDq/4
1910
+ (78)
1911
+ = H∗
1912
+ ǫ (xK+1) + ǫDq/4
1913
+ (84)
1914
+ ≤ H∗
1915
+ ǫ (x0) + ǫDq/4 + (L−1
1916
+ ∇h + 4D2
1917
+ qL∇hǫ−2)
1918
+ K
1919
+
1920
+ k=0
1921
+ ǫ2
1922
+ k
1923
+ (85)
1924
+
1925
+ max
1926
+ y
1927
+ H(x0, y) + ǫDq/4 + 2ǫ2
1928
+ 0(L−1
1929
+ ∇h + 4D2
1930
+ qL∇hǫ−2).
1931
+ It then follows from this and x0 = ˆx0 that (77) holds.
1932
+ We are now ready to prove Theorem 2.
1933
+ Proof of Theorem 2. Suppose for contradiction that Algorithm 2 runs for more than K + 1 outer
1934
+ iterations, where K is given in (26). By this and Algorithm 2, one can then assert that (22) does not
1935
+ hold for all 0 ≤ k ≤ K. On the other hand, by (26) and (76), one has
1936
+ min
1937
+ 0≤k≤K ∥xk+1 − xk∥2
1938
+ (76)
1939
+
1940
+ maxy H(ˆx0, y) − H∗ + ǫDq/4
1941
+ L∇h(K + 1)
1942
+ + 2ǫ2
1943
+ 0(1 + 4D2
1944
+ qL2
1945
+ ∇hǫ−2)
1946
+ L2
1947
+ ∇h(K + 1)
1948
+ (26)
1949
+
1950
+ ǫ2
1951
+ 16L2
1952
+ ∇h
1953
+ ,
1954
+ which implies that there exists some 0 ≤ k ≤ K such that ∥xk+1 − xk∥ ≤ ǫ/(4L∇h), and thus (22) holds
1955
+ for such k, which contradicts the above assertion. Hence, Algorithm 2 must terminate in at most K + 1
1956
+ outer iterations.
1957
+ Suppose that Algorithm 2 terminates at some iteration 0 ≤ k ≤ K, namely, (22) holds for such k. We
1958
+ next show that its output (xǫ, yǫ) = (xk+1, yk+1) is an ǫ-stationary point of (6) and moreover it satisfies
1959
+ (28). Indeed, recall from Lemma 2 that (xk+1, yk+1) is an ǫk-stationary point of (20), namely, it satisfies
1960
+ dist(0, ∂xHk(xk+1, yk+1)) ≤ ǫk and dist(0, ∂yHk(xk+1, yk+1)) ≤ ǫk. By these, (6), (20) and (21), there
1961
+ exists (u, v) such that
1962
+ u ∈ ∂xH(xk+1, yk+1) + 2L∇h(xk+1 − xk),
1963
+ ∥u∥ ≤ ǫk,
1964
+ v ∈ ∂yH(xk+1, yk+1) − ǫ(yk+1 − ˆy0)/(2Dq),
1965
+ ∥v∥ ≤ ǫk.
1966
+ It then follows that u−2L∇h(xk+1−xk) ∈ ∂xH(xk+1, yk+1) and v+ǫ(yk+1−ˆy0)/(2Dq) ∈ ∂yH(xk+1, yk+1).
1967
+ These together with (9), (22), and ǫk ≤ ǫ0 ≤ ǫ/2 (see Algorithm 2) imply that
1968
+ dist
1969
+
1970
+ 0, ∂xH(xk+1, yk+1)
1971
+
1972
+ ≤ ∥u − 2L∇h(xk+1 − xk)∥ ≤ ∥u∥ + 2L∇h∥xk+1 − xk∥
1973
+ (22)
1974
+ ≤ ǫk + ǫ/2 ≤ ǫ,
1975
+ dist
1976
+
1977
+ 0, ∂yH(xk+1, yk+1)
1978
+
1979
+ ≤ ∥v + ǫ(yk+1 − ˆy0)/(2Dq)∥ ≤ ∥v∥ + ǫ∥yk+1 − ˆy0∥/(2Dq)
1980
+ (9)
1981
+ ≤ ǫk + ǫ/2 ≤ ǫ.
1982
+ Hence, the output (xk+1, yk+1) of Algorithm 2 is an ǫ-stationary point of (6). In addition, (28) holds
1983
+ due to Lemma 3.
1984
+ 18
1985
+
1986
+ Recall from Lemma 2 that the number of evaluations of ∇h and proximal operator of p and q
1987
+ performed at iteration k of Algorithm 2 is at most Nk, respectively, where Nk is defined in (73). Also,
1988
+ one can observe from the above proof and the definition of K that |K| ≤ K + 2. It then follows that the
1989
+ total number of evaluations of ∇h and proximal operator of p and q in Algorithm 2 is respectively no
1990
+ more than �|K|−2
1991
+ k=0
1992
+ Nk. Consequently, to complete the rest of the proof of Theorem 2, it suffices to show
1993
+ that �|K|−2
1994
+ k=0
1995
+ Nk ≤ N, where N is given in (27). Indeed, by (27), (73) and |K| ≤ K + 2, one has
1996
+ |K|−2
1997
+
1998
+ k=0
1999
+ Nk
2000
+ (73)
2001
+
2002
+ K
2003
+
2004
+ k=0
2005
+ ��
2006
+ 96
2007
+
2008
+ 2
2009
+
2010
+ 1 + (24L∇h + 4ǫ/Dq) L−1
2011
+ ∇h
2012
+ ��
2013
+ + 2
2014
+
2015
+ ×
2016
+
2017
+ max
2018
+
2019
+ 2,
2020
+
2021
+ DqL∇h
2022
+ ǫ
2023
+
2024
+ × log
2025
+ 4 max
2026
+
2027
+ 1
2028
+ 2L∇h , min
2029
+
2030
+ Dq
2031
+ ǫ ,
2032
+ 4
2033
+ αL∇h
2034
+ �� �
2035
+ δ + 2α−1(H∗ − Hlow + ǫDq/4 + L∇hD2
2036
+ p)
2037
+
2038
+ [(3L∇h + ǫ/(2Dq))2/ min{L∇h, ǫ/(2Dq)} + 3L∇h + ǫ/(2Dq)]−2 ǫ2
2039
+ k
2040
+
2041
+ +
2042
+
2043
+ ��
2044
+ 96
2045
+
2046
+ 2
2047
+
2048
+ 1 + (24L∇h + 4ǫ/Dq) L−1
2049
+ ∇h
2050
+ ��
2051
+ + 2
2052
+
2053
+ max
2054
+
2055
+ 2,
2056
+
2057
+ DqL∇h
2058
+ ǫ
2059
+
2060
+ ×
2061
+ K
2062
+
2063
+ k=0
2064
+
2065
+
2066
+
2067
+ log
2068
+ 4 max
2069
+
2070
+ 1
2071
+ 2L∇h , min
2072
+
2073
+ Dq
2074
+ ǫ ,
2075
+ 4
2076
+ αL∇h
2077
+ �� �
2078
+ δ + 2α−1(H∗ − hlow + ǫDq/4 + L∇hD2
2079
+ p)
2080
+
2081
+ [(3L∇h + ǫ/(2Dq))2/ min{L∇h, ǫ/(2Dq)} + 3L∇h + ǫ/(2Dq)]−2 ǫ2
2082
+ k
2083
+
2084
+
2085
+ +
2086
+ + 1
2087
+
2088
+
2089
+
2090
+ ��
2091
+ 96
2092
+
2093
+ 2
2094
+
2095
+ 1 + (24L∇h + 4ǫ/Dq) L−1
2096
+ ∇h
2097
+ ��
2098
+ + 2
2099
+
2100
+ max
2101
+
2102
+ 2,
2103
+
2104
+ DqL∇h
2105
+ ǫ
2106
+
2107
+ ×
2108
+
2109
+ (K + 1)
2110
+
2111
+ log
2112
+ 4 max
2113
+
2114
+ 1
2115
+ 2L∇h , min
2116
+
2117
+ Dq
2118
+ ǫ ,
2119
+ 4
2120
+ αL∇h
2121
+ �� �
2122
+ δ + 2α−1(H∗ − Hlow + ǫDq/4 + L∇hD2
2123
+ p)
2124
+
2125
+ [(3L∇h + ǫ/(2Dq))2/ min{L∇h, ǫ/(2Dq)} + 3L∇h + ǫ/(2Dq)]−2 ǫ2
2126
+ 0
2127
+
2128
+ +
2129
+ + K + 1 + 2
2130
+ K
2131
+
2132
+ k=0
2133
+ log(k + 1)
2134
+
2135
+ (27)
2136
+ ≤ N,
2137
+ where the last inequality is due to (27) and �K
2138
+ k=0 log(k + 1) ≤ K log(K + 1). This completes the proof
2139
+ of Theorem 2.
2140
+ 5.3
2141
+ Proof of the main results in Section 3
2142
+ In this subsection we prove Theorems 3 and 4. We first establish a lemma below, which will be used to
2143
+ prove Theorem 3 subsequently.
2144
+ Lemma 4. Suppose that Assumption 3 holds and (xǫ, yǫ, zǫ) is an ǫ-optimal solution of problem (35) for
2145
+ some ǫ > 0. Let f, ˜f, f ∗, flow and ρ be given in (29), (32) and (35), respectively. Then we have
2146
+ ˜f(xǫ, yǫ) ≤ min
2147
+ z
2148
+ ˜f(xǫ, z) + ρ−1(f ∗ − flow + 2ǫ),
2149
+ f(xǫ, yǫ) ≤ f ∗ + 2ǫ.
2150
+ Proof. Since (xǫ, yǫ, zǫ) is an ǫ-optimal solution of (35), it follows from Definition 1 that
2151
+ max
2152
+ z
2153
+ Pρ(xǫ, yǫ, z) ≤ Pρ(xǫ, yǫ, zǫ) + ǫ,
2154
+ Pρ(xǫ, yǫ, zǫ) ≤ min
2155
+ x,y max
2156
+ z
2157
+ Pρ(x, y, z) + ǫ.
2158
+ Summing up these inequalities yields
2159
+ max
2160
+ z
2161
+ Pρ(xǫ, yǫ, z) ≤ min
2162
+ x,y max
2163
+ z
2164
+ Pρ(x, y, z) + 2ǫ.
2165
+ (86)
2166
+ Let (x∗, y∗) be an optimal solution of (29).
2167
+ It then follows that f(x∗, y∗) = f ∗ and ˜f(x∗, y∗) =
2168
+ minz ˜f(x∗, z). By these and the definition of Pρ in (35), one has
2169
+ max
2170
+ z
2171
+ Pρ(x∗, y∗, z) = f(x∗, y∗) + ρ( ˜f(x∗, y∗) − min
2172
+ z
2173
+ ˜f(x∗, z)) = f(x∗, y∗) = f ∗,
2174
+ which implies that
2175
+ min
2176
+ x,y max
2177
+ z
2178
+ Pρ(x, y, z) ≤ max
2179
+ z
2180
+ Pρ(x∗, y∗, z) = f ∗.
2181
+ (87)
2182
+ 19
2183
+
2184
+ It then follows from (35), (86) and (87) that
2185
+ f(xǫ, yǫ) + ρ( ˜f(xǫ, yǫ) − min
2186
+ z
2187
+ ˜f(xǫ, z))
2188
+ (35)
2189
+ = max
2190
+ z
2191
+ Pρ(xǫ, yǫ, z)
2192
+ (86)(87)
2193
+
2194
+ f ∗ + 2ǫ,
2195
+ which together with ˜f(xǫ, yǫ) − minz ˜f(xǫ, z) ≥ 0 implies that
2196
+ f(xǫ, yǫ) ≤ f ∗ + 2ǫ,
2197
+ ˜f(xǫ, yǫ) ≤ min
2198
+ z
2199
+ ˜f(xǫ, z) + ρ−1 (f ∗ − f(xǫ, yǫ) + 2ǫ) .
2200
+ The conclusion of this lemma directly follows from these and (32).
2201
+ We are now ready to prove Theorem 3.
2202
+ Proof of Theorem 3. Let {(xk, yk, zk)} be generated by Algorithm 3 with limk→∞(ρk, ǫk) = (∞, 0).
2203
+ By considering a convergent subsequence if necessary, we assume without loss of generality that limk→∞(xk, yk) =
2204
+ (x∗, y∗). we now show that (x∗, y∗) is an optimal solution of problem (29). Indeed, since (xk, yk, zk)
2205
+ is an ǫk-optimal solution of (35) with ρ = ρk, it follows from Lemma 4 with (ρ, ǫ) = (ρk, ǫk) and
2206
+ (xǫ, yǫ) = (xk, yk) that
2207
+ ˜f(xk, yk) ≤ min
2208
+ z
2209
+ ˜f(xk, z) + ρ−1
2210
+ k (f ∗ − flow + 2ǫk),
2211
+ f(xk, yk) ≤ f ∗ + 2ǫk.
2212
+ By the continuity of f and ˜f, limk→∞(xk, yk) = (x∗, y∗), limk→∞(ρk, ǫk) = (∞, 0), and taking limits as
2213
+ k → ∞ on both sides of the above relations, we obtain that ˜f(x∗, y∗) ≤ minz ˜f(x∗, z) and f(x∗, y∗) ≤ f ∗,
2214
+ which clearly imply that y∗ ∈ Argminz ˜f(x∗, z) and f(x∗, y∗) = f ∗. Hence, (x∗, y∗) is an optimal solution
2215
+ of (29) as desired.
2216
+ We next prove Theorem 4. Before proceeding, we establish a lemma below, which will be used to
2217
+ prove Theorem 4 subsequently.
2218
+ Lemma 5. Suppose that Assumption 3 holds and (xǫ, yǫ, zǫ) is an ǫ-stationary point of (35). Let Dy,
2219
+ flow, ˜f, ρ, and Pρ be given in (30), (32) and (35), respectively. Then we have
2220
+ dist
2221
+
2222
+ 0, ∂f(xǫ, yǫ) + ρ∂ ˜f(xǫ, yǫ) − (ρ∇x ˜f(xǫ, zǫ); 0)
2223
+
2224
+ ≤ ǫ,
2225
+ dist
2226
+
2227
+ 0, ρ∂ ˜f(xǫ, zǫ)
2228
+
2229
+ ≤ ǫ,
2230
+ ˜f(xǫ, yǫ) ≤ min
2231
+ z
2232
+ ˜f(xǫ, z) + ρ−1(max
2233
+ z
2234
+ Pρ(xǫ, yǫ, z) − flow).
2235
+ Proof. Since (xǫ, yǫ, zǫ) is an ǫ-stationary point of (35), it follows from Definition 2 that
2236
+ dist
2237
+
2238
+ 0, ∂x,yPρ(xǫ, yǫ, zǫ)
2239
+
2240
+ ≤ ǫ,
2241
+ dist
2242
+
2243
+ 0, ∂zPρ(xǫ, yǫ, zǫ)
2244
+
2245
+ ≤ ǫ.
2246
+ Using these and the definition of Pρ in (35), we have
2247
+ dist
2248
+
2249
+ 0, ∂f(xǫ, yǫ) + ρ∂ ˜f(xǫ, yǫ) − (ρ∇x ˜f(xǫ, zǫ); 0)
2250
+
2251
+ ≤ ǫ,
2252
+ dist
2253
+
2254
+ 0, ρ∂ ˜f(xǫ, zǫ)
2255
+
2256
+ ≤ ε.
2257
+ In addition, by (35), we have
2258
+ f(xǫ, yǫ) + ρ( ˜f(xǫ, yǫ) − min
2259
+ z
2260
+ ˜f(xǫ, z)) = max
2261
+ z
2262
+ Pρ(xǫ, yǫ, z),
2263
+ which along with (32) implies that
2264
+ ˜f(xǫ, yǫ) − min
2265
+ z
2266
+ ˜f(xǫ, z) ≤ ρ−1(max
2267
+ z
2268
+ Pρ(xǫ, yǫ, z) − flow).
2269
+ This completes the proof of this lemma.
2270
+ We are now ready to prove Theorem 4.
2271
+ Proof of Theorem 4. Observe from (36) that problem (35) can be viewed as
2272
+ min
2273
+ x,y max
2274
+ z
2275
+ {Pρ(x, y, z) = h(x, y, z) + p(x, y) − q(z)} ,
2276
+ where h(x, y, z) = f1(x, y) + ρ ˜f1(x, y) − ρ ˜f1(x, z), p(x, y) = f2(x) + ρ ˜f2(y), and q(z) = ρ ˜f2(z). Hence,
2277
+ problem (35) is in the form of (6) with H = Pρ. By Assumption 3 and ρ = ε−1, one can see that h is
2278
+ 20
2279
+
2280
+ �L-smooth on its domain, where �L is given in (39). Also, notice from Algorithm 4 that ǫ0 = ε3/2 ≤ ε/2
2281
+ due to ε ∈ (0, 1/4]. Consequently, Algorithm 2 can be suitably applied to problem (35) with ρ = ε−1 for
2282
+ finding an ǫ-stationary point (xǫ, yǫ, zǫ) of it.
2283
+ In addition, notice from Algorithm 4 that ˜f(x0, y0) ≤ miny ˜f(x0, y)+ε. Using this, (35) and ρ = ε−1,
2284
+ we obtain
2285
+ max
2286
+ z
2287
+ Pρ(x0, y0, z) = f(x0, y0) + ρ( ˜f(x0, y0) − min
2288
+ z
2289
+ ˜f(x0, z)) ≤ f(x0, y0) + ρε = f(x0, y0) + 1.
2290
+ (88)
2291
+ By this and (28) with H = Pρ, ǫ = ε, ǫ0 = ε3/2, ˆx0 = (x0, y0), Dq = Dy, and L∇h = �L, one has
2292
+ Pρ(xǫ, yǫ, zǫ) ≤ max
2293
+ z
2294
+ Pρ(x0, y0, z) + εDy/4 + 2ε3(�L−1 + 4D2
2295
+ y�Lε−2)
2296
+ (88)
2297
+ ≤ 1 + f(x0, y0) + εDy/4 + 2ε3(�L−1 + 4D2
2298
+ y�Lε−2).
2299
+ It then follows from this and Lemma 5 with ǫ = ε and ρ = ε−1 that (xǫ, yǫ, zǫ) satisfies (40) and (41).
2300
+ We next show that at most �
2301
+ N evaluations of ∇f1, ∇ ˜f1, and proximal operator of f2 and ˜f2 are
2302
+ respectively performed in Algorithm 4. Indeed, by (31), (32) and (35), one has
2303
+ min
2304
+ x,y max
2305
+ z
2306
+ Pρ(x, y, z)
2307
+ (35)
2308
+ = min
2309
+ x,y {f(x, y) + ρ( ˜f(x, y) − min
2310
+ z
2311
+ ˜f(x, z))} ≥
2312
+ min
2313
+ (x,y)∈X ×Y f(x, y)
2314
+ (32)
2315
+ = flow,
2316
+ (89)
2317
+ min
2318
+ (x,y,z)∈X ×Y×Y Pρ(x, y, z)
2319
+ (35)
2320
+ =
2321
+ min
2322
+ (x,y,z)∈X ×Y×Y{f(x, y) + ρ( ˜f(x, y) − ˜f(x, z))}
2323
+ (31)(32)
2324
+
2325
+ flow + ρ( ˜flow − ˜fhi).
2326
+ (90)
2327
+ For convenience of the rest proof, let
2328
+ H = Pρ,
2329
+ H∗ = min
2330
+ x,y max
2331
+ z
2332
+ Pρ(x, y, z),
2333
+ Hlow = min{Pρ(x, y, z)|(x, y, z) ∈ X × Y × Y}.
2334
+ (91)
2335
+ In view of these, (87), (88), (89), (90), and ρ = ε−1, we obtain that
2336
+ max
2337
+ z
2338
+ H(x0, y0, z)
2339
+ (88)
2340
+ ≤ f(x0, y0) + 1,
2341
+ flow
2342
+ (89)
2343
+ ≤ H∗ (87)
2344
+ ≤ f ∗,
2345
+ Hlow
2346
+ (90)
2347
+ ≥ flow + ρ( ˜flow − ˜fhi) = flow + ε−1( ˜flow − ˜fhi).
2348
+ Using these and Theorem 2 with ǫ = ε, ˆx0 = (x0, y0), Dp =
2349
+
2350
+ D2x + D2y, Dq = Dy, ǫ0 = ε3/2, L∇h = �L,
2351
+ α = ˆα, δ = ˆδ, and H, H∗, Hlow given in (91), we can conclude that Algorithm 4 performs at most
2352
+
2353
+ N evaluations of ∇f1, ∇ ˜f1 and proximal operator of f2 and ˜f2 respectively for finding an approximate
2354
+ solution (xǫ, yǫ) of problem (29) satisfying (40) and (41).
2355
+ 5.4
2356
+ Proof of the main results in Section 4
2357
+ In this subsection we prove Theorems 5 and 6. Before proceeding, we define
2358
+ r = G−1Dy(ρ−1ǫ + L ˜
2359
+ f),
2360
+ B+
2361
+ r = {λ ∈ Rl
2362
+ + : ∥λ∥ ≤ r},
2363
+ (92)
2364
+ where Dy is defined in (30), G is given in Assumption 4(iii), and ǫ and ρ are given in Algorithm 6. In
2365
+ addition, one can observe from (43) and (47) that
2366
+ min
2367
+ z
2368
+ �Pµ(x, z) ≤ ˜f ∗(x)
2369
+ ∀x ∈ X,
2370
+ (93)
2371
+ which will be frequently used later.
2372
+ We next establish several technical lemmas that will be used to prove Theorem 5 subsequently.
2373
+ Lemma 6. Suppose that Assumptions 3 and 4 hold. Let Dy, L ˜
2374
+ f, G, ˜f ∗, ˜f ∗
2375
+ hi and B+
2376
+ r be given in (30),
2377
+ (43), (44), (92) and Assumption 4, respectively. Then the following statements hold.
2378
+ (i) ∥λ∗∥ ≤ G−1L ˜
2379
+ fDy and λ∗ ∈ B+
2380
+ r for all λ∗ ∈ Λ∗(x) and x ∈ X, where Λ∗(x) denotes the set of
2381
+ optimal Lagrangian multipliers of problem (43) for any x ∈ X.
2382
+ 21
2383
+
2384
+ (ii) The function ˜f ∗ is Lipschitz continuous on X and ˜f ∗
2385
+ hi is finite.
2386
+ (iii) It holds that
2387
+ ˜f ∗(x) = max
2388
+ λ
2389
+ min
2390
+ z
2391
+ ˜f(x, z) + ⟨λ, ˜g(x, z)⟩ − IRl
2392
+ +(λ)
2393
+ ∀x ∈ X,
2394
+ (94)
2395
+ where IRl
2396
+ +(·) is the indicator function associated with Rl
2397
+ +.
2398
+ Proof. (i) Let x ∈ X and λ∗ ∈ Λ∗(x) be arbitrarily chosen, and let z∗ ∈ Y be such that (z∗, λ∗) is a pair
2399
+ of primal-dual optimal solutions of (43). It then follows that
2400
+ z∗ ∈ Argmin
2401
+ z
2402
+ ˜f(x, z) + ⟨λ∗, ˜g(x, z)⟩,
2403
+ ⟨λ∗, ˜g(x, z∗)⟩ = 0,
2404
+ ˜g(x, z∗) ≤ 0,
2405
+ λ∗ ≥ 0.
2406
+ The first relation above yields
2407
+ ˜f(x, z∗) + ⟨λ∗, ˜g(x, z∗)⟩ ≤ ˜f(x, ˆzx) + ⟨λ∗, ˜g(x, ˆzx)⟩,
2408
+ where ˆzx is given in Assumption 4(iii). By this and ⟨λ∗, ˜g(x, z∗)⟩ = 0, one has
2409
+ ⟨λ∗, −˜g(x, ˆzx)⟩ ≤ ˜f(x, ˆzx) − ˜f(x, z∗),
2410
+ which together with λ∗ ≥ 0, (30) and Assumption 4 implies that
2411
+ G
2412
+ l
2413
+
2414
+ i=1
2415
+ λ∗
2416
+ i ≤ ⟨λ∗, −˜g(x, ˆzx)⟩ ≤ ˜f(x, ˆzx) − ˜f(x, z∗) ≤ L ˜
2417
+ f∥ˆzx − z∗∥ ≤ L ˜
2418
+ fDy,
2419
+ (95)
2420
+ where the first inequality is due to Assumption 4(iii), and the third inequality follows from (30) and L ˜
2421
+ f-
2422
+ Lipschitz continuity of ˜f (see Assumption 4(i)). By (92), (95) and λ∗ ≥ 0, we have ∥λ∗∥ ≤ �l
2423
+ i=1 λ∗
2424
+ i ≤
2425
+ G−1L ˜
2426
+ fDy and λ∗ ∈ B+
2427
+ r .
2428
+ (ii) Recall from Assumptions 3(i) and 4(iii) that ˜f(x, ·) and ˜gi(x, ·), i = 1, . . . , l, are convex for any
2429
+ given x ∈ X. Using this, (43) and the first statement of this lemma, we observe that
2430
+ ˜f ∗(x) = min
2431
+ z
2432
+ max
2433
+ λ∈B+
2434
+ r
2435
+ ˜f(x, z) + ⟨λ, ˜g(x, z)⟩
2436
+ ∀x ∈ X.
2437
+ (96)
2438
+ Notice from Assumption 4 that ˜f and ˜g are Lipschitz continuous on their domain. Then it is not hard to
2439
+ observe that max{ ˜f(x, z)+⟨λ, ˜g(x, z)⟩|λ ∈ B+
2440
+ r } is a Lipschitz continuous function of (x, z) on its domain.
2441
+ By this and (96), one can easily verify that ˜f ∗ is Lipschitz continuous on X. In addition, the finiteness
2442
+ of ˜f ∗
2443
+ hi follows from (44), the continuity of ˜f ∗, and the compactness of X.
2444
+ (iii) One can observe from (43) that for all x ∈ X,
2445
+ ˜f ∗(x) = min
2446
+ z
2447
+ max
2448
+ λ
2449
+ ˜f(x, z) + ⟨λ, ˜g(x, z)⟩ − IRl
2450
+ +(λ) ≥ max
2451
+ λ
2452
+ min
2453
+ z
2454
+ ˜f(x, z) + ⟨λ, ˜g(x, z)⟩ − IRl
2455
+ +(λ)
2456
+ where the inequality follows from the weak duality. In addition, it follows from Assumption 3 that the
2457
+ domain of ˜f(x, ·) is compact for all x ∈ X. By this, (96) and the strong duality, one has
2458
+ ˜f ∗(x) = max
2459
+ λ∈B+
2460
+ r
2461
+ min
2462
+ z
2463
+ ˜f(x, z) + ⟨λ, ˜g(x, z)⟩ − IRl
2464
+ +(λ)
2465
+ ∀x ∈ X,
2466
+ which together with the above inequality implies that (94) holds.
2467
+ Lemma 7. Suppose that Assumptions 3 and 4 hold and that (xǫ, yǫ, zǫ) is an ǫ-optimal solution of
2468
+ problem (50) for some ǫ > 0. Let flow, f, �Pµ, f ∗
2469
+ µ, ρ and µ be given in (32), (42), (47), (48) and (50),
2470
+ respectively. Then we have
2471
+ �Pµ(xǫ, yǫ) ≤ min
2472
+ z
2473
+ �Pµ(xǫ, z) + ρ−1(f ∗
2474
+ µ − flow + 2ǫ),
2475
+ f(xǫ, yǫ) ≤ f ∗
2476
+ µ + 2ǫ.
2477
+ (97)
2478
+ Proof. The proof follows from the same argument as the one for Lemma 4 with f ∗ and ˜f being replaced
2479
+ by f ∗
2480
+ µ and �Pµ, respectively.
2481
+ Lemma 8. Suppose that Assumptions 3-5 hold. Let ˜flow, f ∗, ˜f ∗
2482
+ hi, f ∗
2483
+ µ be defined in (31), (42), (44) and
2484
+ (48), and Lf, ω and ¯θ be given in Assumptions 4 and 5. Suppose that µ ≥ ( ˜f ∗
2485
+ hi − ˜flow)/¯θ2. Then we have
2486
+ f ∗
2487
+ µ ≤ f ∗ + Lfω
2488
+ ��
2489
+ µ−1( ˜f ∗
2490
+ hi − ˜flow)
2491
+
2492
+ .
2493
+ (98)
2494
+ 22
2495
+
2496
+ Proof. Let x ∈ X, y ∈ Argminz{ ˜f(x, z)|˜g(x, z) ≤ 0} and z∗ ∈ Argminz �Pµ(x, z) be arbitrarily chosen.
2497
+ One can easily see from (47) and (93) that ˜f(x, z∗) + µ ∥[˜g(x, z∗)]+∥2 ≤ ˜f ∗(x), which together with (31)
2498
+ and (44) implies that
2499
+ ∥[˜g(x, z∗)]+∥2 ≤ µ−1( ˜f ∗
2500
+ hi − ˜flow).
2501
+ (99)
2502
+ Since µ ≥ ( ˜f ∗
2503
+ hi− ˜flow)/¯θ2, it follows from (99) that ∥[˜g(x, z∗)]+∥ ≤ ¯θ. By this relation, y ∈ Argmin
2504
+ z
2505
+ { ˜f(x, z)|˜g(x, z) ≤
2506
+ 0} and Assumption 5, there exists some ˆz∗ such that
2507
+ ∥y − ˆz∗∥ ≤ ω(∥[˜g(x, z∗)]+∥),
2508
+ ˆz∗ ∈ Argmin
2509
+ z
2510
+
2511
+ ˜f(x, z)
2512
+ �� ∥[˜g(x, z)]+∥ ≤ ∥[˜g(x, z∗)]+∥
2513
+
2514
+ .
2515
+ (100)
2516
+ In view of (47), z∗ ∈ Argminz �Pµ(x, z) and the second relation in (100), one can observe that ˆz∗ ∈
2517
+ Argminz �Pµ(x, z), which along with (48) yields f(x, ˆz∗) ≥ f ∗
2518
+ µ. Also, using (100) and Lf-Lipschitz conti-
2519
+ nuity of f (see Assumption 4), we have
2520
+ f(x, y) − f(x, ˆz∗) ≥ −Lf∥y − ˆz∗∥
2521
+ (100)
2522
+
2523
+ −Lfω(∥[˜g(x, z∗)]+∥).
2524
+ Taking minimum over x ∈ X and y ∈ Argminz{ ˜f(x, z)|˜g(x, z) ≤ 0} on both sides of this relation, and
2525
+ using (42), (99), f(x, ˆz∗) ≥ f ∗
2526
+ µ and the monotonicity of ω, we can conclude that (98) holds.
2527
+ Lemma 9. Suppose that Assumptions 3-5 hold. Let ˜flow, flow, f, ˜f, f ∗, ˜f ∗, ˜f ∗
2528
+ hi, ρ and µ be given in
2529
+ (31), (32), (42), (43), (44) and (50), and Lf, ω and ¯θ be given in Assumptions 4 and 5, respectively.
2530
+ Suppose that µ ≥ ( ˜f ∗
2531
+ hi − ˜flow)/¯θ2 and (xǫ, yǫ, zǫ) is an ǫ-optimal solution of problem (50) for some ǫ > 0.
2532
+ Then we have
2533
+ f(xǫ, yǫ) ≤ f ∗ + Lfω
2534
+ ��
2535
+ µ−1( ˜f ∗
2536
+ hi − ˜flow)
2537
+
2538
+ + 2ǫ,
2539
+ ˜f(xǫ, yǫ) ≤ ˜f ∗(xǫ) + ρ−1�
2540
+ f ∗ − flow + Lfω
2541
+ ��
2542
+ µ−1( ˜f ∗
2543
+ hi − ˜flow)
2544
+
2545
+ + 2ǫ
2546
+
2547
+ ,
2548
+ ∥[˜g(xǫ, yǫ)]+∥2 ≤ µ−1�
2549
+ ˜f ∗(xǫ) − ˜flow + ρ−1�
2550
+ f ∗ − flow + Lfω
2551
+ ��
2552
+ µ−1( ˜f ∗
2553
+ hi − ˜flow)
2554
+
2555
+ + 2ǫ
2556
+ ��
2557
+ .
2558
+ Proof. By (47), (93), and the first relation in (97), one has
2559
+ ˜f(xǫ, yǫ) + µ ∥[˜g(xǫ, yǫ)]+∥2 (47)
2560
+ =
2561
+ �Pµ(xǫ, yǫ)
2562
+ (93)(97)
2563
+
2564
+ ˜f ∗(xǫ) + ρ−1(f ∗
2565
+ µ − flow + 2ǫ).
2566
+ It then follows from this and (31) that
2567
+ ˜f(xǫ, yǫ) ≤ ˜f ∗(xǫ) + ρ−1(f ∗
2568
+ µ − flow + 2ǫ),
2569
+ ∥[˜g(xǫ, yǫ)]+∥2 ≤ µ−1( ˜f ∗(xǫ) − ˜flow + ρ−1(f ∗
2570
+ µ − flow + 2ǫ)).
2571
+ In addition, recall from (97) that f(xǫ, yǫ) ≤ f ∗
2572
+ µ + 2ǫ. The conclusion of this lemma then follows from
2573
+ these three relations and (98).
2574
+ We are now ready to prove Theorem 5.
2575
+ Proof of Theorem 5. Let {(xk, yk, zk)} be generated by Algorithm 5 with limk→∞(ρk, µk, ǫk) = (∞, ∞, 0).
2576
+ By considering a convergent subsequence if necessary, we assume without loss of generality that limk→∞(xk, yk) =
2577
+ (x∗, y∗). We now show that (x∗, y∗) is an optimal solution of problem (42). Indeed, since (xk, yk, zk) is
2578
+ an ǫk-optimal solution of (50) with (ρ, µ) = (ρk, µk) and limk→∞ µk = ∞, it follows from Lemma 9 with
2579
+ (ρ, µ, ǫ) = (ρk, µk, ǫk) and (xǫ, yǫ) = (xk, yk) that for all sufficiently large k, one has
2580
+ f(xk, yk) ≤ f ∗ + Lfω
2581
+ ��
2582
+ µ−1
2583
+ k ( ˜f ∗
2584
+ hi − ˜flow)
2585
+
2586
+ + 2ǫk,
2587
+ ˜f(xk, yk) ≤ ˜f ∗(xk) + ρ−1
2588
+ k
2589
+
2590
+ f ∗ − flow + Lfω
2591
+ ��
2592
+ µ−1
2593
+ k ( ˜f ∗
2594
+ hi − ˜flow)
2595
+
2596
+ + 2ǫk
2597
+
2598
+ ,
2599
+ ��[˜g(xk, yk)]+
2600
+ ��2 ≤ µ−1
2601
+ k
2602
+
2603
+ ˜f ∗(xk) − ˜flow + ρ−1
2604
+ k
2605
+
2606
+ f ∗ − flow + Lfω
2607
+ ��
2608
+ µ−1
2609
+ k ( ˜f ∗
2610
+ hi − ˜flow)
2611
+
2612
+ + 2ǫk
2613
+ ��
2614
+ .
2615
+ By the continuity of f, ˜f and ˜f ∗ (see Assumption 3(i) and Lemma 6(ii)), limk→∞(xk, yk) = (x∗, y∗),
2616
+ limk→∞(ρk, µk, ǫk) = (∞, ∞, 0), limθ↓0 ω(θ) = 0, and taking limits as k → ∞ on both sides of the above
2617
+ relations, we obtain that f(x∗, y∗) ≤ f ∗, ˜f(x∗, y∗) ≤ ˜f ∗(x∗) and [˜g(x∗, y∗)]+ = 0, which along with
2618
+ (42) and (43) imply that f(x∗, y∗) = f ∗ and y∗ ∈ Argminz{ ˜f(x∗, z)|˜g(x∗, z) ≤ 0}. Hence, (x∗, y∗) is an
2619
+ optimal solution of (42) as desired.
2620
+ 23
2621
+
2622
+ We next prove Theorem 6. Before proceeding, we establish several technical lemmas below, which
2623
+ will be used to prove Theorem 6 subsequently.
2624
+ Lemma 10. Suppose that Assumptions 3 and 4 hold and that (xǫ, yǫ, zǫ) is an ǫ-stationary point of
2625
+ problem (50) for some ǫ > 0. Let Dy, ˜g, ρ, µ, Lf, L ˜
2626
+ f and G be given in (30), (42), (50) and Assumption
2627
+ 4, respectively. Then we have
2628
+ ∥[˜g(xǫ, zǫ)]+∥ ≤ (2µG)−1Dy(ρ−1ǫ + L ˜
2629
+ f),
2630
+ (101)
2631
+ ∥[˜g(xǫ, yǫ)]+∥ ≤ (2µG)−1Dy(ρ−1ǫ + ρ−1Lf + L ˜
2632
+ f).
2633
+ (102)
2634
+ Proof. We first prove (101). Since (xǫ, yǫ, zǫ) is an ǫ-stationary point of (50), it follows from Definition
2635
+ 2 that dist(0, ∂zPρ,µ(xǫ, yǫ, zǫ)) ≤ ǫ. Also, by (47) and (50), one has
2636
+ Pρ,µ(x, y, z) = f(x, y) + ρ( ˜f(x, y) + µ ∥[˜g(x, y)]+∥2) − ρ( ˜f(x, z) + µ ∥[˜g(x, z)]+∥2).
2637
+ (103)
2638
+ Using these relations, we have
2639
+ dist
2640
+
2641
+ 0, ∂z ˜f(xǫ, zǫ) + 2µ
2642
+ l
2643
+
2644
+ i=1
2645
+ [˜gi(xǫ, zǫ)]+∇z˜gi(xǫ, zǫ)
2646
+
2647
+ ≤ ρ−1ǫ.
2648
+ Hence, there exists s ∈ ∂z ˜f(xǫ, zǫ) such that
2649
+ ���s + 2µ
2650
+ l
2651
+
2652
+ i=1
2653
+ [˜gi(xǫ, zǫ)]+∇z˜gi(xǫ, zǫ)
2654
+ ��� ≤ ρ−1ǫ.
2655
+ (104)
2656
+ Let ˆzxǫ and G be given in Assumption 4(iii). It then follows that ˆzxǫ ∈ Y and −˜gi(xǫ, ˆzxǫ) ≥ G > 0 for
2657
+ all i. Notice that [˜gi(xǫ, zǫ)]+˜gi(xǫ, zǫ) ≥ 0 for all i and ∥zǫ − ˆzxǫ∥ ≤ Dy due to (30). Using these, (104),
2658
+ and the convexity of ˜f(xǫ, ·) and ˜gi(xǫ, ·) for all i, we have
2659
+ ˜f(xǫ, zǫ) − ˜f(xǫ, ˆzxǫ) + 2µG
2660
+ l
2661
+
2662
+ i=1
2663
+ [˜gi(xǫ, zǫ)]+ ≤ ˜f(xǫ, zǫ) − ˜f(xǫ, ˆzxǫ) − 2µ
2664
+ l
2665
+
2666
+ i=1
2667
+ [˜gi(xǫ, zǫ)]+˜gi(xǫ, ˆzxǫ)
2668
+ ≤ ˜f(xǫ, zǫ) − ˜f(xǫ, ˆzxǫ) + 2µ
2669
+ l
2670
+
2671
+ i=1
2672
+ [˜gi(xǫ, zǫ)]+(˜gi(xǫ, zǫ) − ˜gi(xǫ, ˆzxǫ))
2673
+ ≤ ⟨s, zǫ − ˆzxǫ⟩ + 2µ
2674
+ l
2675
+
2676
+ i=1
2677
+ [˜gi(xǫ, zǫ)]+⟨∇z˜gi(xǫ, zǫ), zǫ − ˆzxǫ⟩
2678
+ = ⟨s + 2µ
2679
+ l
2680
+
2681
+ i=1
2682
+ [˜g(xǫ, zǫ)]+∇z˜gi(xǫ, zǫ), zǫ − ˆzxǫ⟩ ≤ ρ−1Dyǫ,
2683
+ (105)
2684
+ where the first inequality is due to −˜gi(xǫ, ˆzxǫ) ≥ G for all i, the second inequality follows from
2685
+ [˜gi(xǫ, zǫ)]+˜gi(xǫ, zǫ) ≥ 0 for all i, the third inequality is due to s ∈ ∂z ˜f(xǫ, zǫ) and the convexity of
2686
+ ˜f(xǫ, ·) and ˜gi(xǫ, ·) for all i, and the last inequality follows from (30) and (104). In view of (30), (105),
2687
+ and L ˜
2688
+ f-Lipschitz continuity of ˜f(x, y) (see Assumption 4), one has
2689
+ ∥[˜g(xǫ, zǫ)]+∥ ≤
2690
+ l
2691
+
2692
+ i=1
2693
+ [˜gi(xǫ, zǫ)]+
2694
+ (105)
2695
+
2696
+ (2µG)−1(ρ−1Dyǫ + ˜f(xǫ, ˆzxǫ) − ˜f(xǫ, zǫ))
2697
+ ≤ (2µG)−1(ρ−1Dyǫ + L ˜
2698
+ f∥ˆzxǫ − zǫ∥)
2699
+ (30)
2700
+ ≤ (2µG)−1Dy(ρ−1ǫ + L ˜
2701
+ f).
2702
+ Hence, (101) holds.
2703
+ We next prove (102). Since (xǫ, yǫ, zǫ) is an ǫ-stationary point of (50), it follows from Definition 2
2704
+ that dist(0, ∂yPρ,µ(xǫ, yǫ, zǫ)) ≤ ǫ. This together with (103) implies that
2705
+ dist
2706
+
2707
+ 0, ∂yf(xǫ, yǫ) + ρ∂y ˜f(xǫ, yǫ) + 2ρµ∇y˜g(xǫ, yǫ)[˜g(xǫ, yǫ)]+
2708
+
2709
+ ≤ ǫ.
2710
+ Hence, there exists s ∈ ∂yf(xǫ, yǫ) and ˜s ∈ ∂y ˜f(xǫ, yǫ) such that
2711
+ ∥s + ρ˜s + 2ρµ∇y˜g(xǫ, yǫ)[˜g(xǫ, yǫ)]+∥ ≤ ǫ.
2712
+ (106)
2713
+ 24
2714
+
2715
+ Let ¯
2716
+ A(xǫ, yǫ) = {i|˜gi(xǫ, yǫ) > 0, 1 ≤ i ≤ l}, ˆzxǫ and G be given in Assumption 4(iii). It then follows
2717
+ that ˆzxǫ ∈ Y and −˜gi(xǫ, ˆzxǫ) ≥ G > 0 for all i. Using these and the convexity of ˜gi(xǫ, ·) for all i, we
2718
+ have
2719
+ ⟨∇y˜g(xǫ, yǫ)[˜g(xǫ, yǫ)]+, yǫ − ˆzxǫ⟩ =
2720
+
2721
+ i∈ ¯
2722
+ A(xǫ,yǫ)
2723
+ ⟨∇y˜gi(xǫ, yǫ), yǫ − ˆzxǫ⟩[gi(xǫ, yǫ)]+
2724
+
2725
+
2726
+ i∈ ¯
2727
+ A(xǫ,yǫ)
2728
+ (˜gi(xǫ, yǫ) − ˜gi(xǫ, ˆzxǫ))[˜gi(xǫ, yǫ)]+
2729
+
2730
+
2731
+ i∈ ¯
2732
+ A(xǫ,yǫ)
2733
+ G[˜gi(xǫ, yǫ)]+ = G
2734
+ l
2735
+
2736
+ i=1
2737
+ [˜gi(xǫ, yǫ)]+ ≥ G ∥[˜g(xǫ, yǫ)]+∥ ,
2738
+ (107)
2739
+ where the first inequality follows from the convexity of ˜g(xǫ, ·) and the second inequality is due to
2740
+ −˜gi(xǫ, ˆzxǫ) ≥ G. It then follows from this, (106) and (107) that
2741
+ Dyǫ ≥ ∥s + ρ˜s + 2ρµ∇y˜g(xǫ, yǫ)[˜g(xǫ, yǫ)]+∥ · ∥yǫ − ˆzxǫ∥
2742
+ ≥ ⟨s + ρ˜s + 2ρµ∇y˜g(xǫ, yǫ)[˜g(xǫ, yǫ)]+, yǫ − ˆzxǫ⟩
2743
+ = ��s + ρ˜s, yǫ − ˆzxǫ⟩ + 2ρµ⟨∇y˜g(xǫ, yǫ)[˜g(xǫ, yǫ)]+, yǫ − ˆzxǫ⟩
2744
+ (107)
2745
+ ≥ − (∥s∥ + ρ∥˜s∥) ∥yǫ − ˆzxǫ∥ + 2ρµG ∥[˜g(xǫ, yǫ)]+∥
2746
+ ≥ −(Lf + ρL ˜
2747
+ f)Dy + 2ρµG ∥[˜g(xǫ, yǫ)]+∥ ,
2748
+ (108)
2749
+ where the last inequality follows from ∥yǫ − ˆzxǫ∥ ≤ Dy and the fact that ∥s∥ ≤ Lf and ∥˜s∥ ≤ L ˜
2750
+ f, which
2751
+ are due to (30), s ∈ ∂yf(xǫ, yǫ), ˜s ∈ ∂y ˜f(xǫ, yǫ) and Assumption 4(i). By (108), one can immediately see
2752
+ that (102) holds.
2753
+ Lemma 11. Suppose that Assumptions 3 and 4 hold. Let f, ˜f, ˜g, Dy, flow, ˜f ∗ and Pρ,µ be given in (29),
2754
+ (30), (32), (43) and (50), Lf, L ˜
2755
+ f and G be given in Assumptions 3 and 4, (xǫ, yǫ, zǫ) be an ǫ-stationary
2756
+ point of (50) for some ǫ > 0, and
2757
+ ˜λ = 2µ[˜g(xǫ, zǫ)]+,
2758
+ ˆλ = 2ρµ[˜g(xǫ, yǫ)]+.
2759
+ (109)
2760
+ Then we have
2761
+ dist
2762
+
2763
+ ∂f(xǫ, yǫ) + ρ∂ ˜f(xǫ, yǫ) − ρ(∇x ˜f(xǫ, zǫ) + ∇x˜g(xǫ, zǫ)˜λ; 0) + ∇˜g(xǫ, yǫ)ˆλ
2764
+
2765
+ ≤ ǫ,
2766
+ (110)
2767
+ dist
2768
+
2769
+ 0, ρ(∂z ˜f(xǫ, zǫ) + ∇z˜g(xǫ, zǫ)˜λ)
2770
+
2771
+ ≤ ǫ,
2772
+ (111)
2773
+ ∥[˜g(xǫ, zǫ)]+∥ ≤ (2µG)−1Dy(ρ−1ǫ + L ˜
2774
+ f),
2775
+ (112)
2776
+ |⟨˜λ, ˜g(xǫ, zǫ)⟩| ≤ (2µ)−1G−2D2
2777
+ y(ρ−1ǫ + L ˜
2778
+ f)2,
2779
+ (113)
2780
+ | ˜f(xǫ, yǫ) − ˜f ∗(xǫ)| ≤ max
2781
+
2782
+ ρ−1(max
2783
+ z
2784
+ Pρ,µ(xǫ, yǫ, z) − flow), (2µ)−1G−2D2
2785
+ yL ˜
2786
+ f(ρ−1ǫ + ρ−1Lf + L ˜
2787
+ f)
2788
+
2789
+ ,
2790
+ (114)
2791
+ ∥[˜g(xǫ, yǫ)]+∥ ≤ (2µG)−1Dy(ρ−1ǫ + ρ−1Lf + L ˜
2792
+ f),
2793
+ (115)
2794
+ |⟨ˆλ, ˜g(xǫ, yǫ)⟩| ≤ (2µ)−1ρG−2D2
2795
+ y(ρ−1ǫ + ρ−1Lf + L ˜
2796
+ f)2.
2797
+ (116)
2798
+ Proof. Since (xǫ, yǫ, zǫ) is an ǫ-stationary point of (50), it easily follows from (103), (109) and Definition
2799
+ 2 that (110) and (111) hold. Also, it follows from (101) and (102) that (112) and (115) hold. In addition,
2800
+ in view of (109), (112) and (115), one has
2801
+ |⟨˜λ, ˜g(xǫ, zǫ)⟩|
2802
+ (109)
2803
+ =
2804
+ 2µ ∥[˜g(xǫ, zǫ)]+∥2 (112)
2805
+
2806
+ (2µ)−1G−2D2
2807
+ y(ρ−1ǫ + L ˜
2808
+ f)2,
2809
+ |⟨ˆλ, ˜g(xǫ, yǫ)⟩|
2810
+ (109)
2811
+ =
2812
+ 2ρµ ∥[˜g(xǫ, yǫ)]∥+∥2 (115)
2813
+
2814
+ (2µ)−1ρG−2D2
2815
+ y(ρ−1ǫ + ρ−1Lf + L ˜
2816
+ f)2,
2817
+ and hence (113) and (116) hold. Also, observe from the definition of Pρ,µ in (50) that
2818
+ �Pµ(xǫ, yǫ) − min
2819
+ z
2820
+ �Pµ(xǫ, z) = ρ−1(max
2821
+ z
2822
+ Pρ,µ(xǫ, yǫ, z) − f(xǫ, yǫ)).
2823
+ 25
2824
+
2825
+ Using this, (32), (47) and (93), we obtain that
2826
+ ˜f(xǫ, yǫ) + µ ∥[˜g(xǫ, yǫ)]+∥2 (47)
2827
+ =
2828
+ �Pµ(xǫ, yǫ) =
2829
+ min
2830
+ z
2831
+ �Pµ(xǫ, z) + ρ−1(max
2832
+ z
2833
+ Pρ,µ(xǫ, yǫ, z) − f(xǫ, yǫ))
2834
+ (32)(93)
2835
+
2836
+ ˜f ∗(xǫ) + ρ−1(max
2837
+ z
2838
+ Pρ,µ(xǫ, yǫ, z) − flow).
2839
+ (117)
2840
+ On the other hand, let λ∗ ∈ Rl
2841
+ + be an optimal Lagrangian multiplier of problem (43) with x = xǫ. It
2842
+ then follows from Lemma 6(i) that ∥λ∗∥ ≤ G−1L ˜
2843
+ fDy. Using these and (115), we have
2844
+ ˜f ∗(xǫ) = min
2845
+ y
2846
+
2847
+ ˜f(xǫ, y) + ⟨λ∗, ˜g(xǫ, y)⟩
2848
+
2849
+ ≤ ˜f(xǫ, yǫ) + ⟨λ∗, ˜g(xǫ, yǫ)⟩
2850
+ ≤ ˜f(xǫ, yǫ) + ∥λ∗∥∥[˜g(xǫ, yǫ)]+∥ ≤ ˜f(xǫ, yǫ) + (2µ)−1G−2D2
2851
+ yL ˜
2852
+ f(ρ−1ǫ + ρ−1Lf + L ˜
2853
+ f).
2854
+ By this and (117), one can see that (114) holds.
2855
+ We are now ready to prove Theorem 6.
2856
+ Proof of Theorem 6. Observe from (51) that problem (50) can be viewed as
2857
+ min
2858
+ x,y max
2859
+ z
2860
+ {Pρ,µ(x, y, z) = h(x, y, z) + p(x, y) − q(z)} ,
2861
+ where h(x, y, z) = f1(x, y) + ρ ˜f1(x, y) + ρµ ∥[˜g(x, y)]+∥2 − ρ ˜f1(x, z) − ρµ ∥[˜g(x, z)]+∥2, p(x, y) = f2(x) +
2862
+ ρ ˜f2(y) and q(z) = ρ ˜f2(z). Hence, problem (50) is in the form of (6) with H = Pρ,µ. By Assumption 3,
2863
+ (45), (46), ρ = ε−1 and µ = ε−2, one can see that h is �L-smooth on its domain, where �L is given in (61).
2864
+ Also, notice from Algorithm 6 that ǫ0 = ε5/2 ≤ ε/2 = ǫ/2 due to ε ∈ (0, 1/4]. Consequently, Algorithm 2
2865
+ can be suitably applied to problem (50) with ρ = ε−1 and µ = ε−2 for finding an ǫ-stationary point
2866
+ (xǫ, yǫ, zǫ) of it.
2867
+ In addition, notice from Algorithm 6 that �Pµ(x0, y0) ≤ miny �Pµ(x0, y) + ε. Using this, (50) and
2868
+ ρ = ε−1, we obtain
2869
+ max
2870
+ z
2871
+ Pρ,µ(x0, y0, z)
2872
+ (50)
2873
+ = f(x0, y0) + ρ( �Pµ(x0, y0) − min
2874
+ z
2875
+ �Pµ(x0, z)) ≤ f(x0, y0) + ρε = f(x0, y0) + 1. (118)
2876
+ By this and (28) with H = Pρ,µ, ǫ = ε, ǫ0 = ε5/2, ˆx0 = (x0, y0), Dq = Dy and L∇h = �L, one has
2877
+ Pρ,µ(xǫ, yǫ, zǫ) ≤
2878
+ max
2879
+ z
2880
+ Pρ,µ(x0, y0, z) + εDy/4 + 2ε5(�L−1 + 4D2
2881
+ y�Lε−2)
2882
+ (118)
2883
+
2884
+ 1 + f(x0, y0) + εDy/4 + 2ε5(�L−1 + 4D2
2885
+ y�Lε−2).
2886
+ It then follows from this and Lemma 11 with ǫ = ε, ρ = ε−1 and µ = ε−2 that (xǫ, yǫ, zǫ) satisfies the
2887
+ relations (62)-(68).
2888
+ We next show that at most �
2889
+ N evaluations of ∇f1, ∇ ˜f1, ∇˜g and proximal operator of f2 and ˜f2 are
2890
+ respectively performed in Algorithm 6. Indeed, by (31), (32), (45), (47) and (50), one has
2891
+ min
2892
+ x,y max
2893
+ z
2894
+ Pρ,µ(x, y, z)
2895
+ (50)
2896
+ = min
2897
+ x,y {f(x, y) + ρ( �Pµ(x, y) − min
2898
+ z
2899
+ �Pµ(x, z))} ≥
2900
+ min
2901
+ (x,y)∈X ×Y f(x, y)
2902
+ (32)
2903
+ = flow, (119)
2904
+ min{Pρ,µ(x, y, z)|(x, y, z) ∈ X × Y × Y}
2905
+ (50)
2906
+ = min{f(x, y) + ρ( �Pµ(x, y) − �Pµ(x, z))|(x, y, z) ∈ X × Y × Y}
2907
+ (47)
2908
+ = min{f(x, y) + ρ( ˜f(x, y) + µ∥[˜g(x, y)]+∥2 − ˜f(x, z) − µ∥[˜g(x, z)]+∥2)|(x, y, z) ∈ X × Y × Y}
2909
+ ≥ flow + ρ( ˜flow − ˜fhi) − ρµ˜g2
2910
+ hi,
2911
+ (120)
2912
+ where the last inequality follows from (31), (32) and (45). In addition, let (x∗, y∗) be an optimal solution
2913
+ of (42). It then follows that f(x∗, y∗) = f ∗ and [˜g(x∗, y∗)]+ = 0. By these, (31), (47) and (50), one has
2914
+ min
2915
+ x,y max
2916
+ z
2917
+ Pρ,µ(x, y, z) ≤ max
2918
+ z
2919
+ Pρ,µ(x∗, y∗, z)
2920
+ (50)
2921
+ = f(x∗, y∗) + ρ
2922
+
2923
+ �Pµ(x∗, y∗) − min
2924
+ z
2925
+ �Pµ(x∗, z)
2926
+
2927
+ (47)
2928
+ = f(x∗, y∗) + ρ( ˜f(x∗, y∗) + µ∥[˜g(x∗, y∗)]+∥2 − min
2929
+ z { ˜f(x∗, z) + µ∥[˜g(x∗, z)]+∥2})
2930
+ (31)
2931
+ ≤ f ∗ + ρ( ˜fhi − ˜flow).
2932
+ (121)
2933
+ 26
2934
+
2935
+ For convenience of the rest proof, let
2936
+ H = Pρ,µ,
2937
+ H∗ = min
2938
+ x,y max
2939
+ z
2940
+ Pρ,µ(x, y, z),
2941
+ Hlow = min{Pρ,µ(x, y, z)|(x, y, z) ∈ X × Y × Y}.
2942
+ (122)
2943
+ In view of these, (118), (119), (120), (121), ρ = ε−1 and µ = ε−2, we obtain that
2944
+ max
2945
+ z
2946
+ H(x0, y0, z)
2947
+ (118)
2948
+ ≤ f(x0, y0) + 1,
2949
+ flow
2950
+ (119)
2951
+
2952
+ H∗ (121)
2953
+
2954
+ f ∗ + ρ( ˜fhi − ˜flow) = f ∗ + ε−1( ˜fhi − ˜flow),
2955
+ Hlow
2956
+ (120)
2957
+
2958
+ flow + ρ( ˜flow − ˜fhi) − ρµ˜g2
2959
+ hi = flow + ε−1( ˜flow − ˜fhi) − ε−3˜g2
2960
+ hi.
2961
+ Using these and Theorem 2 with ǫ = ε, ˆx0 = (x0, y0), Dp =
2962
+
2963
+ D2x + D2y, Dq = Dy, ǫ0 = ε5/2, L∇h = �L,
2964
+ α = ˜α, δ = ˜δ, and H, H∗, Hlow given in (122), we can conclude that Algorithm 6 performs at most �
2965
+ N
2966
+ evaluations of ∇f1, ∇ ˜f1, ∇˜g and proximal operator of f2 and ˜f2 for finding an approximate solution
2967
+ (xǫ, yǫ) of problem (42) satisfying (62)-(68).
2968
+ 6
2969
+ Concluding remarks
2970
+ For the sake of simplicity, first-order penalty methods are proposed only for problem (3) in this paper.
2971
+ It would be interesting to extend them to problem (1) by using a standard technique (e.g., see [39]) for
2972
+ handling the constraint g(x, y) ≤ 0. In addition, a single subproblem with static penalty and tolerance
2973
+ parameters is solved in our methods (Algorithms 4 and 6), which may be conservative in practice. To
2974
+ make the methods possibly practically more efficient, it would be natural to modify them by solving
2975
+ a sequence of subproblems with dynamic penalty and tolerance parameters instead. These along with
2976
+ numerical experiments will be left for the future research.
2977
+ References
2978
+ [1] G. B. Allende and G. Still. Solving bilevel programs with the KKT-approach. Mathematical pro-
2979
+ gramming, 138(1):309–332, 2013.
2980
+ [2] J. F. Bard. Practical bilevel optimization: algorithms and applications, volume 30. Springer Science
2981
+ & Business Media, 2013.
2982
+ [3] K. P. Bennett, G. Kunapuli, J. Hu, and J.-S. Pang. Bilevel optimization and machine learning. In
2983
+ IEEE World Congress on Computational Intelligence, pages 25–47. Springer, 2008.
2984
+ [4] L. Bertinetto, J. F. Henriques, P. Torr, and A. Vedaldi. Meta-learning with differentiable closed-form
2985
+ solvers. In International Conference on Learning Representations, 2018.
2986
+ [5] G. H. Chen and R. T. Rockafellar. Convergence rates in forward–backward splitting. SIAM Journal
2987
+ on Optimization, 7(2):421–444, 1997.
2988
+ [6] T. Chen, Y. Sun, and W. Yin. A single-timescale stochastic bilevel optimization method. arXiv
2989
+ preprint arXiv:2102.04671, 2021.
2990
+ [7] F. H. Clarke. Optimization and nonsmooth analysis. SIAM, 1990.
2991
+ [8] B. Colson, P. Marcotte, and G. Savard. An overview of bilevel optimization. Annals of operations
2992
+ research, 153(1):235–256, 2007.
2993
+ [9] C. Crockett, J. A. Fessler, et al. Bilevel methods for image reconstruction. Foundations and Trends®
2994
+ in Signal Processing, 15(2-3):121–289, 2022.
2995
+ [10] S. Dempe. Foundations of bilevel programming. Springer Science & Business Media, 2002.
2996
+ [11] S. Dempe, V. Kalashnikov, G. A. P´erez-Vald´es, and N. Kalashnykova. Bilevel programming prob-
2997
+ lems. Energy Systems. Springer, Berlin, 10:978–3, 2015.
2998
+ [12] S. Dempe and A. Zemkoho. Bilevel optimization. In Springer optimization and its applications. Vol.
2999
+ 161. Springer, 2020.
3000
+ 27
3001
+
3002
+ [13] S. Dempe and A. B. Zemkoho. The bilevel programming problem: reformulations, constraint qual-
3003
+ ifications and optimality conditions. Mathematical Programming, 138(1):447–473, 2013.
3004
+ [14] A. L. Dontchev and R. T. Rockafellar.
3005
+ Implicit functions and solution mappings, volume 543.
3006
+ Springer, 2009.
3007
+ [15] M. Feurer and F. Hutter. Hyperparameter optimization. In Automated machine learning, pages
3008
+ 3–33. Springer, Cham, 2019.
3009
+ [16] L. Franceschi, M. Donini, P. Frasconi, and M. Pontil.
3010
+ Forward and reverse gradient-based hy-
3011
+ perparameter optimization. In International Conference on Machine Learning, pages 1165–1173,
3012
+ 2017.
3013
+ [17] L. Franceschi, P. Frasconi, S. Salzo, R. Grazzi, and M. Pontil. Bilevel programming for hyperpa-
3014
+ rameter optimization and meta-learning. In International Conference on Machine Learning, pages
3015
+ 1568–1577, 2018.
3016
+ [18] Z. Guo and T. Yang. Randomized stochastic variance-reduced methods for stochastic bilevel opti-
3017
+ mization. arXiv e-prints, pages arXiv–2105, 2021.
3018
+ [19] P. Hansen, B. Jaumard, and G. Savard. New branch-and-bound rules for linear bilevel programming.
3019
+ SIAM Journal on scientific and Statistical Computing, 13(5):1194–1217, 1992.
3020
+ [20] M. Hong, H.-T. Wai, Z. Wang, and Z. Yang. A two-timescale framework for bilevel optimization:
3021
+ Complexity analysis and application to actor-critic. arXiv preprint arXiv:2007.05170, 2020.
3022
+ [21] X. Hu, N. Xiao, X. Liu, and K.-C. Toh. An improved unconstrained approach for bilevel optimiza-
3023
+ tion. arXiv preprint arXiv:2208.00732, 2022.
3024
+ [22] Y. Ishizuka and E. Aiyoshi. Double penalty method for bilevel optimization problems. Annals of
3025
+ Operations Research, 34(1):73–88, 1992.
3026
+ [23] K. Ji, J. D. Lee, Y. Liang, and H. V. Poor. Convergence of meta-learning with task-specific adapta-
3027
+ tion over partial parameters. Advances in Neural Information Processing Systems, 33:11490–11500,
3028
+ 2020.
3029
+ [24] K. Ji, J. Yang, and Y. Liang. Bilevel optimization: Nonasymptotic analysis and faster algorithms.
3030
+ arXiv preprint arXiv:2010.07962, 2020.
3031
+ [25] A. Kaplan and R. Tichatschke. Proximal point methods and nonconvex optimization. Journal of
3032
+ global Optimization, 13(4):389–406, 1998.
3033
+ [26] P. Khanduri, S. Zeng, M. Hong, H.-T. Wai, Z. Wang, and Z. Yang. A near-optimal algorithm for
3034
+ stochastic bilevel optimization via double-momentum. Advances in Neural Information Processing
3035
+ Systems, 34, 2021.
3036
+ [27] V. Konda and J. Tsitsiklis. Actor-critic algorithms. Advances in neural information processing
3037
+ systems, 12, 1999.
3038
+ [28] W. Kong and R. D. Monteiro. An accelerated inexact proximal point method for solving nonconvex-
3039
+ concave min-max problems. SIAM Journal on Optimization, 31(4):2558–2585, 2021.
3040
+ [29] D. Kovalev and A. Gasnikov. The first optimal algorithm for smooth and strongly-convex-strongly-
3041
+ concave minimax optimization. arXiv preprint arXiv:2205.05653, 2022.
3042
+ [30] H. Liu, K. Simonyan, and Y. Yang. Darts: Differentiable architecture search. In International
3043
+ Conference on Learning Representations, 2018.
3044
+ [31] R. Liu, J. Gao, J. Zhang, D. Meng, and Z. Lin. Investigating bi-level optimization for learning and
3045
+ vision from a unified perspective: A survey and beyond. IEEE Transactions on Pattern Analysis
3046
+ and Machine Intelligence, 2021.
3047
+ [32] D. Lopez-Paz and M. Ranzato. Gradient episodic memory for continual learning. Advances in neural
3048
+ information processing systems, 30, 2017.
3049
+ 28
3050
+
3051
+ [33] Z.-Q. Luo, J.-S. Pang, and D. Ralph. Mathematical programs with equilibrium constraints. Cam-
3052
+ bridge University Press, 1996.
3053
+ [34] X. Ma, W. Yao, J. J. Ye, and J. Zhang. Combined approach with second-order optimality conditions
3054
+ for bilevel programming problems. arXiv preprint arXiv:2108.00179, 2021.
3055
+ [35] D. Maclaurin, D. Duvenaud, and R. Adams. Gradient-based hyperparameter optimization through
3056
+ reversible learning. In International conference on machine learning, pages 2113–2122, 2015.
3057
+ [36] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu.
3058
+ Towards deep learning models
3059
+ resistant to adversarial attacks. In International Conference on Learning Representations, 2018.
3060
+ [37] J. A. Mirrlees. The theory of moral hazard and unobservable behaviour: Part I. The Review of
3061
+ Economic Studies, 66(1):3–21, 1999.
3062
+ [38] Y. Nesterov. Gradient methods for minimizing composite functions. Mathematical programming,
3063
+ 140(1):125–161, 2013.
3064
+ [39] J. Nocedal and S. J. Wright. Numerical optimization. Springer, 1999.
3065
+ [40] J. Outrata, M. Kocvara, and J. Zowe. Nonsmooth approach to optimization problems with equilibrium
3066
+ constraints: theory, applications and numerical results, volume 28. Springer Science & Business
3067
+ Media, 2013.
3068
+ [41] F. Pedregosa. Hyperparameter optimization with approximate gradient. In International conference
3069
+ on machine learning, pages 737–746, 2016.
3070
+ [42] A. Rajeswaran, C. Finn, S. M. Kakade, and S. Levine.
3071
+ Meta-learning with implicit gradients.
3072
+ Advances in neural information processing systems, 32, 2019.
3073
+ [43] C. Shi, J. Lu, and G. Zhang. An extended Kuhn–Tucker approach for linear bilevel programming.
3074
+ Applied Mathematics and Computation, 162(1):51–63, 2005.
3075
+ [44] K. Shimizu, Y. Ishizuka, and J. F. Bard. Nondifferentiable and two-level mathematical programming.
3076
+ Springer Science & Business Media, 2012.
3077
+ [45] A. Sinha, P. Malo, and K. Deb. A review on bilevel optimization: from classical to evolutionary
3078
+ approaches and applications.
3079
+ IEEE Transactions on Evolutionary Computation, 22(2):276–295,
3080
+ 2017.
3081
+ [46] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus. Intriguing
3082
+ properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.
3083
+ [47] L. N. Vicente and P. H. Calamai. Bilevel and multilevel programming: A bibliography review.
3084
+ Journal of Global optimization, 5(3):291–306, 1994.
3085
+ [48] H. Von Stackelberg. Market structure and equilibrium. Springer Science & Business Media, 2010.
3086
+ [49] D. Ward and J. M. Borwein. Nonsmooth calculus in finite dimensions. SIAM Journal on control
3087
+ and optimization, 25(5):1312–1340, 1987.
3088
+ [50] J. J. Ye. Constraint qualifications and optimality conditions in bilevel optimization. In Bilevel
3089
+ Optimization, pages 227–251. Springer, 2020.
3090
+ [51] J. J. Ye, X. Yuan, S. Zeng, and J. Zhang. Difference of convex algorithms for bilevel programs with
3091
+ applications in hyperparameter selection. Mathematical Programming, pages 1–34, 2022.
3092
+ 29
3093
+
BNAzT4oBgHgl3EQfv_6i/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
BdFIT4oBgHgl3EQf_ywO/content/2301.11416v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f8ca3c04bb6dbd55aaa31c7a0960e7a0bf2233f487e36f4631ce12b17b876368
3
+ size 4829854
CdAzT4oBgHgl3EQfGftE/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e092b4b481c3e5ab13902a006cf18c414f3e86f675ba564fbdc0f71eef13ee1b
3
+ size 4223863
CdAzT4oBgHgl3EQfTvxb/content/tmp_files/2301.01254v1.pdf.txt ADDED
@@ -0,0 +1,694 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ Machine learning prediction of the MJO extends beyond one month
3
+ Tamaki Suematsu1*, Kengo Nakai2*, Tsuyoshi Yoneda3, Daisuke Takasuka4,5, Takuya Jinno6,3,
4
+ Yoshitaka Saiki7, Hiroaki Miura6
5
+
6
+ 1RIKEN Center for Computational Science; Kobe, Hyogo, 650-0047, Japan.
7
+ *Tamaki Suematsu. E-mail: tamaki.suematsu@riken.jp
8
+ 2Faculty of Marine Technology, Tokyo University of Marine Science and Technology; Tokyo 135-8533,
9
+ Japan.
10
+ *Kengo Nakai. E-mail: knakai0@kaiyodai.ac.jp
11
+ 3Graduate School of Economics, Hitotsubashi University; Kunitachi, Tokyo, 186-8601; Japan.
12
+ 4Atmosphere and Ocean Research Institute, The University of Tokyo; Kashiwa, Chiba, 277-0882, Japan.
13
+ 5Japan Agency for Marine-Earth Science and Technology; Yokohama, Kanagawa, 236-0001, Japan.
14
+ 6Graduate School of Science, The University of Tokyo; Bunkyo-ku, Tokyo, 113-0033, Japan.
15
+ 7Graduate School of Business Administration, Hitotsubashi University; Kunitachi, Tokyo, 186-8601,
16
+ Japan.
17
+ *These authors contributed equally to this work
18
+ Abstract
19
+ The prediction of the Madden-Julian Oscillation (MJO), a massive tropical weather
20
+ event with vast global socio-economic impacts1,2, has been infamously difficult with
21
+ physics-based weather prediction models3–5. Here we construct a machine learning
22
+ model using reservoir computing technique that forecasts the real-time multivariate
23
+ MJO index (RMM)6, a macroscopic variable that represents the state of the MJO.
24
+ The training data was refined by developing a novel filter that extracts the
25
+ recurrency of MJO signals from the raw atmospheric data and selecting a suitable
26
+
27
+ 2
28
+ time-delay coordinate of the RMM. The model demonstrated the skill to forecast
29
+ the state of MJO events for a month from the pre-developmental stages. Best-
30
+ performing cases predicted the RMM sequence over two months, which exceeds the
31
+ expected inherent predictability limit of the MJO.
32
+
33
+ Main text
34
+ The Madden–Julian Oscillation (MJO) 7 is a massive cluster of convective activities in the tropics that
35
+ spans thousands of kilometers traveling slowly eastward from the Indian Ocean to the central Pacific in
36
+ approximately 20 to 60 days. It has far reaching influence on the global weather 1,8 and is recognized to
37
+ be one of the most important sources of predictability in extended-range weather forecast longer than
38
+ weeks 9,10. However, simulation of the MJO by physics-based dynamical numerical models (hereafter
39
+ dynamical models) has been shown to be notoriously difficult 3,11,12. It has only been since the mid 2000s
40
+ that the predictability of the MJO by dynamical models 13,14 exceeded that of empirical statistical models,
41
+ such as atmospheric-only linear inverse models, at two to three weeks 15. The current forecast skills of
42
+ dynamical models for MJO prediction lie between two to five weeks 16, which falls short of the expected
43
+ inherent predictability of the MJO estimated from multi-model ensemble studies at six to seven weeks 17.
44
+ Weather forecasts by dynamical models require physical parameterizations that incorporate the mean
45
+ effects of the sub-grid scale processes on the evolution of the grid-scale flows. However, parameterizations
46
+ are prone to empirical tuning and are inevitable sources of uncertainty of the dynamical models 18–20
47
+ because we have yet to determine the correct theoretical formulations and parameters of the statistical mean
48
+ states of microscopic processes. The difficulties of reducing the ambiguities in parameterizations continue
49
+ to be a major limiting factor for improving dynamical models and for identifying processes essential for
50
+ successful weather predictions. In contrast, machine learning models with relatively small number of neural
51
+ networks trained only by the time series of macroscopic variables has the potential to implicitly incorporate
52
+
53
+ 3
54
+ the influence of microscopic variables on the macroscopic variables and to eliminate the parameterizations
55
+ that unsatisfactorily replicate the multiscale interactions between the unresolved and the resolved processes.
56
+ The effectiveness of the machine learning methods has been demonstrated in the fields of atmospheric
57
+ and climate science. Considerable progress has been achieved in areas for forecasting phenomena with
58
+ large socio-economic impacts such as the El Niño Southern Oscillation 21,22, Asian summer monsoons 23,
59
+ and hurricanes 24 at incomparably small computational costs compared to the dynamical models of the
60
+ atmosphere and the ocean. However, phenomena at the intraseasonal time scale have been difficult to
61
+ predict with the use of machine learning methods because complex interactions between processes with
62
+ various spatio-temporal scales that range from the convective to seasonal scales play important role in
63
+ determining its time evolution. Particularly with regards to the MJO forecasts, machine learning models
64
+ have been outperformed by dynamical models 25,26.
65
+ Here, we employ the reservoir computing method to advance the machine learning prediction of the MJO.
66
+ The reservoir computing method is a brain-inspired machine-learning technique that constructs a data-
67
+ driven dynamical model (hereafter reservoir computing models) 27–31. By training on a time series of
68
+ macroscopic variables of high-dimensional dynamics, the method can efficiently predict time series and
69
+ frequency spectra of its chaotic behaviors 32,33. For example, it is useful for predicting the statistical
70
+ quantities of a chaotic fluid flow, which cannot be calculated directly from a numerical simulation of the
71
+ Navier–Stokes equation due to its high computational cost 34. In this study, we construct a reservoir
72
+ computing MJO prediction model, trained only by the time series of a macroscopic variable, with a
73
+ performance competitive with the state-of-the-art physics based dynamical models. Our results
74
+ demonstrate that the inherent predictability of some MJO cases is longer than have been expected from
75
+ studies by dynamical models 17.
76
+
77
+ 4
78
+
79
+ Fig. 1. Schematic picture of reservoir computing. (A) In the training phase, the input data 𝒖(𝑡) for time
80
+ 𝑡 is fed to the reservoir state vector 𝒓(𝑡) through input weight matrix 𝑾!" and the output weight matrix
81
+ 𝑾#$% is determined by reservoir computing. (B) In the prediction phase, the time evolution of 𝒖(𝑡) for
82
+ time 𝑡 + Δ 𝑡 is predicted as 𝒖*(𝑡 + Δ 𝑡) from the 𝑾∗
83
+ #$% determined in the training phase.
84
+ A reservoir is a recurrent neural network whose internal parameters are adjusted to fit the data in the
85
+ training process 27,28. It is trained by feeding an input time series and fitting a linear function of the high
86
+ dimensional reservoir state vector to the desired output time series (Fig. 1). The construction of a reservoir
87
+ computing model simply assumes recurrent and deterministic property of the input time series and does
88
+ not involve any physical knowledge of the input data. The reservoir computing model of this study is
89
+ described by:
90
+
91
+ ATraining phase
92
+ Reservoir
93
+ statevector
94
+ r(t)
95
+ Training
96
+ Input data
97
+ Outputdata
98
+ ()n
99
+ (+)n
100
+ Win
101
+ r(t+△t)
102
+ W
103
+ out
104
+ B Prediction phase
105
+ Reservoir
106
+ statevector
107
+ r(t
108
+ Predicted
109
+ Input data
110
+ outputdata
111
+ u(t)
112
+ (+)
113
+ Determined
114
+ Win
115
+ r(t+△t)
116
+ W
117
+ Nout5
118
+ +𝒓(𝑡 + Δ 𝑡) = (1 − 𝛼) 𝒓(𝑡) + 𝛼 tanh4𝑨 𝒓 (𝑡) + 𝑾!" 𝒖(𝑡)6
119
+ 𝒖*(𝑡 + Δ 𝑡) = 𝑾∗
120
+ #$% 𝒓(𝑡 + Δ 𝑡)
121
+
122
+ (1)
123
+ where 𝒖(𝑡) ∈ ℝ( is both the input variable vector, 𝒓(𝑡) ∈ ℝ) (𝑁 ≫ 𝑀) is the reservoir state vector,
124
+ 𝑨 ∈ ℝ)×) , 𝑾!" ∈ ℝ)×( , and 𝑾∗
125
+ #$% ∈ ℝ(×) are reservoir, input, and output weight matrices,
126
+ respectively, 𝛼 (0 < 𝛼 ≤ 1) is the coefficient that adjusts the nonlinearity of the dynamics of 𝒓, and Δ 𝑡
127
+ is the time step. We define tanh(𝐪) = (tanh(q+) , tanh(q,) , … tanh(q)))- , for a vector 𝐪 =
128
+ (q+, q,, … q))- , where 𝑇 represents the transpose of a vector. 𝑾∗
129
+ #$% is determined to satisfy 𝒖(𝑡) ≈
130
+ 𝑾#$% 𝒓(𝑡) using the training data 𝒖(𝑡), where 𝑾#$% is the output weight matrix in the training phase.
131
+ Further details on the construction of the reservoir computing model are provided in the supplementary
132
+ materials. In the prediction phase, the predicted variable 𝒖*(𝑡 + Δ 𝑡) is obtained from 𝒖(𝑡) and 𝒓 (𝑡), using
133
+ eqn. (1) with fixed 𝑨, 𝑾!" , and 𝑾∗#$%. A successful training will give 𝒖*(𝑡 + Δ 𝑡) that approximates the
134
+ desired unmeasured quantity 𝒖(𝑡 + Δ 𝑡).
135
+ The objective of our reservoir computing model is to predict the sequence of the Realtime Multi-variate
136
+ MJO (RMM) index 6, which is widely accepted as the standard proxy for diagnosing the state of an MJO1.
137
+ It captures the signals of the MJO as an envelope of convective activities coupled to planetary-scale
138
+ circulation from the leading pair of principal components (RMM1, RMM2) of the equatorial outgoing
139
+ longwave radiation and zonal winds at 850 hPa and 200 hPa. The RMM calculated from data without
140
+ smoothing in time 6 has been applied to machine learning prediction of the MJO 25,26; however, their
141
+ machine learning predictions were susceptible to degradation ascribed to noises in unsmoothed data from
142
+ atmospheric variabilities outside of the MJO timescale. Moreover, signals at time scales longer than the
143
+ MJO needs to be removed from the training data for the machine learning to identify recurring patterns.
144
+ Thus, to refine the RMM time series to train our reservoir computing model for MJO prediction, signals
145
+ outside of the MJO frequency range were removed from the raw data by an application of a filter that
146
+ approximately retains signals only between 20 days and 120 days frequency range 35.
147
+
148
+ 6
149
+ The Lanczos filter 36, which is conventionally used to filter MJO signals, cannot be employed as the filter
150
+ to generate the training data for the machine learning. This is because the Lanczos filter, whose weights
151
+ are symmetric in time, requires data from both the past and the future to calculate a filtered value at a
152
+ certain point in time (Fig. 2 A). To resolve this problem, we design a novel filter, applicable for real-time
153
+ use, that does not require data from the future. The filter Ψ.!,.",0 is defined as:
154
+ Ψ.!,.",0 (𝑡) = F.!,." (𝑡)
155
+ sin(𝑡
156
+ 𝑐 − 𝜋)
157
+ (𝑡
158
+ 𝑐 − 𝜋)
159
+ ,
160
+ where
161
+ F.!,." (𝑡) = K
162
+ sin L 𝑡
163
+ 𝑟1N
164
+ 𝑡
165
+
166
+ sin L 𝑡
167
+ 𝑟2N
168
+ 𝑡
169
+ (𝑡 ≤ 0)
170
+ 0 (𝑡 > 0)
171
+ ,
172
+ and 𝑐 is a parameter that adjusts the center of the weights. We set the parameters as (𝑟1, 𝑟2, 𝑐) = (
173
+ ,3
174
+ 4 ,
175
+ +,3
176
+ 4 , 14) to remove the signals at frequencies lower than 120 days and higher than 20 days. The shape of
177
+ the filter function in real-space and in Fourier space is compared against that of the Lanczos filter in Fig. 2
178
+ A, B. In contrast to the Lanczos filter, the weights of the new filter vanish at 𝑡 = 0 and require only the
179
+ data from the past. The asymmetric weights of the new filter make it suitable for its application to real-
180
+ time use such as filtering the input variable data for machine learning predictions. Due to the asymmetry,
181
+ the center of the weight of the new filter shifts backward only by approximately eight days. Hereafter, this
182
+ filter will be referred to as the real-time band-pass filter (RB filter). The RMM time series is calculated
183
+ from data filtered by the RB filter in this study (see methods for details).
184
+
185
+ 7
186
+
187
+ Fig. 2. Comparison of Real-time Band-pass filter and Lanczos filter. The shape of the Lanczos (red)
188
+ and RB (blue) filters are shown in (A) real space and in (B) Fourier space. (C) Sample trajectories, from
189
+ 1st December 2018 (indicated by circles) to 9th January 2019, for the original Wheeler and Hendon 2004
190
+ RMM index (WH04, grey), and RMM index filtered by Lanczos (red) and RB (blue) filters and RMM2
191
+ replaced by 12-day time-delay coordinate of RMM1. The axis for both RMM2 and 12-day time-delay
192
+ coordinate of RMM1 is labeled as RMM2 for brevity.
193
+ Furthermore, the MJO prediction is refined by employing the RMM phase space spanned by RMM1 and
194
+ its delay coordinate to diagnose the state of the MJO. That is, we replace RMM2 with the delay coordinate
195
+ of RMM1 and eliminate the model prediction of RMM2. This enhances the recurrency of the input data
196
+ and contributes to the robustness of the trained model. The modification is founded on the expectation that
197
+ RMM2 can be reconstructed from the delay coordinate of RMM1, since RMM1 and RMM2 are orthogonal
198
+ by definition and the trajectories of the projections of MJO events on the RMM phase space are near
199
+ circular. The delay time of the delay coordinate is chosen at 12 days, when the auto-correlation of RMM1
200
+ crosses zero for the first time. The correlation coefficient of RMM2 and 12-day delay coordinate variable
201
+ of RMM1 is 0.75. The trajectories of the RB-filtered and Lanczos filtered RMM sequences with RMM2
202
+ replaced by the 12-day delay coordinate of RMM1 is compared with the original Wheeler and Hendon
203
+ RMM (WH04) 6 in Fig. 2C. We confirm that the RB filter removes signals from slow variabilities and
204
+ noises as effectively as the Lanczos filter and that the trajectory of the RMM sequence on the phase space
205
+ with RMM2 replaced by the 12-day delay coordinate of RMM1 is similar to that of the WH04 RMM on
206
+ phase space spanned by RMM1 and RMM2. Thus, we focus on the RMM phase space spanned by RMM1
207
+
208
+ 8
209
+ and its 12-day delay coordinate, which we will refer to as the machine learning RMM (ML-RMM) phase
210
+ space. The relevance of ML-RMM phase space to the conventional one spanned by RMM1 and RMM2 is
211
+ further discussed in the supplementary materials (Fig. S1). We will denote RMM1 and its time-delay
212
+ coordinate at time 𝑡 as 𝑎(𝑡) ≔ RMM1(𝑡) and 𝑏(𝑡): = 𝑎(𝑡 − 12).
213
+ It is known that a chaotic attractor can be reconstructed by some observable variables and its delay
214
+ coordinates 37,38. For the construction of a reservoir computing model, it is efficient to take the delay
215
+ coordinate variable with an appropriate delay time as the input when the number of observable variable is
216
+ smaller than the effective dimension of the attractor 33. A suitable delay time and dimension of the delay
217
+ coordinate of RMM1 is inferred by computing its auto-correlation function. Thus, an M-dimensional
218
+ delay coordinate vector of RMM1 is introduced as the input and output variable vector 𝒖 in Eq. (1):
219
+ 𝒖(𝑡) = 4RMM1(𝑡), RMM1(𝑡 − 1Δ 𝜏), … , RMM1(𝑡 − (𝑀 − 1)Δ 𝜏)6,
220
+ where ∆τ is the delay time, and (Δ 𝜏, 𝑀) = (6, 7). The pair of parameters are chosen so that the behavior
221
+ of 𝒖(𝑡) is deterministic and has recurrency, which are essential properties for successful modelling. The
222
+ reservoir model (Eq. (1)) of the RMM1 time sequence is constructed by determining 𝑾∗
223
+ #$%. The time series
224
+ of the RMM1 data from 30th December 1986 to 29th December 2011 was used as the training data. The
225
+ optimal reservoir computing model was selected from evaluation of test cases of RMM1 forecasts
226
+ initialized from every day between 8th April 2014 and 6th July 2014. The selected model is used throughout
227
+ this study for all predictions.
228
+ The predicted variables at time t initialized from time p are denoted as 𝑎\(𝑡, 𝑝) and 𝑏^(𝑡, 𝑝). We note that
229
+ 𝑏^(𝑡, 𝑝) is predicted by the reservoir computing model simultaneously with 𝑎\(𝑡, 𝑝). The relationship
230
+ 𝑏^(𝑡, 𝑝) = 𝑎\(𝑡 − 12, 𝑝) would hold only in an ideal case in which the model learns the delay coordinate of
231
+ the RMM1 perfectly. The reference time series in this case are 𝑎(𝑡) and 𝑏(𝑡). We compare the time series
232
+ of predicted variables against the reference time series using the bivariate correlation coefficient (COR)
233
+ 16,39, defined by the equation:
234
+
235
+ 9
236
+ COR(𝑞) ≔
237
+
238
+ L𝑎(𝑝 + 𝑞)𝑎\(𝑝 + 𝑞, 𝑝) + 𝑏(𝑝 + 𝑞)𝑏^(𝑝 + 𝑞, 𝑝)N
239
+ )
240
+ 56+
241
+ c∑
242
+ ((𝑎(𝑝 + 𝑞))𝟐 + 4𝑏(𝑝 + 𝑞)6
243
+ 𝟐)
244
+ )
245
+ 56+
246
+ + c∑
247
+ ((𝑎\(𝑝 + 𝑞, 𝑝))𝟐 + (𝑏^(𝑝 + 𝑞, 𝑝))𝟐)
248
+ )
249
+ 56+
250
+ ,
251
+ where 𝑞 is the forecast lead time. The COR corresponds to a covariance between the actual vector
252
+ (𝑎(𝑡), 𝑏(𝑡)) and the predicted vector (𝑎\(𝑡, 𝑝), 𝑏^(𝑡, 𝑝)), and is conventionally used to evaluate the MJO
253
+ prediction skills of dynamical and statistical models 40,41. Here, the 𝑁 = 2010 is the number of
254
+ predictions initialized for all days between 28th July 2014 and 28th January 2020. In Fig. 3, we show the
255
+ time series of COR(𝑞) for all predictions and three cases, the details of which will be described next. The
256
+ COR(𝑞) stays above 0.5 for 28 days for all predictions. The threshold value 0.5 is customarily adopted
257
+ for MJO prediction skill score 16. This signifies that the expectancy of the skill score of the model is at
258
+ three weeks for all days, including periods devoid of MJO activity. The forecast skill was evaluated as
259
+ three weeks in consideration of the approximate 8-day shift by the RB filter as discussed above.
260
+ It is customary to evaluate the skill of MJO predictions from the forecasts of periods when MJO events
261
+ are identified 14. We reevaluate the forecast skill of the reservoir model following the custom. Here, the
262
+ MJO events were identified as continuous sequences from phase 2 to phase 7 on the RMM phase space
263
+ spanned by RMM1 and its delay coordinate of 12 days 8,35 (See methods for details). For the predictions
264
+ initialized on three, five, and seven days before the onsets of MJO events, the COR remains above 0.5 for
265
+ 38 days for all three cases. Considering the 8-day shift by the RB filter, this signifies that the constructed
266
+ model has the potential to skillfully forecast the time evolutions of the MJO events for 30 days.
267
+ Counterintuitively, the COR decays below 0.5 faster for the forecasts initialized three days before the
268
+ MJO onsets than those for five and seven days before the MJO onsets. We note however, that this is
269
+ consistent with the fact that the predictions reach the terminations of MJO events faster for predictions
270
+ that are initiated closer to the onsets.
271
+ The performance of the MJO prediction on individual cases are examined to illuminate the similarity
272
+ between the predicted and the actual trajectories of the RMM. Figure 4 compares the actual and predicted
273
+
274
+ 10
275
+ trajectories on the ML-RMM phase, prediction errors, and the phase difference for the 10th (A, B, C), 26th
276
+ (D, E, F), and 50th (G, H, I) best performing cases in terms of mean error over the first 60 days of the
277
+ prediction. The three samples are chosen so that there are no overlaps in the forecast lead times. The errors
278
+ are measured by the distance between the actual (𝑎(𝑡), 𝑏(𝑡)) and the predicted (𝑎\(𝑡, 𝑝), 𝑏^(𝑡, 𝑝)) vectors.
279
+ The phase difference is evaluated from the cosine of the angle between (𝑎(𝑡), 𝑏(𝑡)) and (𝑎\(𝑡, 𝑝), 𝑏^(𝑡, 𝑝))
280
+ (cos(𝜃5(𝑡))). In all three cases, the error remains below 1.4, the threshold of the root mean square errors
281
+ of the predicted RMM adopted to evaluate the skills of climate simulations 42, well beyond two months (>
282
+ 75 days). The prediction also stays in phase (cos(𝜃5(𝑡)) > 0.7) for nearly two months (58, 83, and 76
283
+ days for the 10th, 26th, and 50th best case, respectively). We note that the rapid increases in phase differences
284
+ occur when the amplitude of the RMM1 decreases. This is reasonable considering that the RMM phases
285
+ become physically meaningless with diminishment of its amplitude. These results indicate that our
286
+ reservoir computing model can predict the state of some MJO events well beyond the estimated inherent
287
+ predictability limit of 7 weeks from dynamical model studies 17. This inference is supported by cases of
288
+ RMM1 predictions that skillfully forecast RMM1 phases for longer than 120 days (see Fig. S2).
289
+
290
+ 11
291
+
292
+
293
+ Fig. 3. Bivariate correlation coefficient. The mean bivariate correlation coefficient (COR) as a function
294
+ of forecast lead time (days) for all 2010 predictions (red), and for predictions initialized on 3 (navy), 5
295
+ (blue), and 7 (light blue) days before onsets of MJO events. The dash-dot and the dotted lines indicate 28th
296
+ and the 38th day in the forecast lead time, respectively.
297
+ We constructed a computationally inexpensive machine learning model, using the reservoir computing
298
+ technique, that is capable of month-long forecasts of the state of the MJO. This prediction skill is superior
299
+ to that of preceding machine learning methods and is matched only by physics-based dynamical models
300
+ that inevitably demand the state-of-the-art supercomputers 13,14,40,43. It is remarkable that our model was
301
+ trained only by the macroscopic time series of the RMM1. This signifies that intricate information of the
302
+ atmospheric and oceanic states that influences the MJO 44,45 were implicitly incorporated into the reservoir
303
+ state variables of the neural network. The extended prediction skill of our reservoir model is attributed to
304
+ the refinement of the training data. The signals from slow variability and high frequency noise were filtered
305
+ out from the input data with the RB filter to restrict the degrees of freedom of the training. This was
306
+
307
+ 12
308
+ essential because it was necessary for the model to efficiently learn from merely 26 years of RMM1 data
309
+ with a limited number (< 100) of MJO events. To further enhance the efficacy of the reservoir computing,
310
+ we introduced the delay coordinate variable of RMM1 to employ suitably correlated variables as our
311
+ training data 33. It is of interest how the extension of the training data with accumulation of observational
312
+ data in the future will enhance the performance of the reservoir model.
313
+ The best performing forecasts by our reservoir model predicted the RMM time series for more than two
314
+ months. These results indicate that some MJO events are inherently predictable beyond the potential
315
+ predictability estimates made from dynamical model studies at seven weeks 17. This implies a possibility
316
+ for significant improvements in dynamical models to extend their lead time in MJO prediction, which is
317
+ crucial for reliable global weather forecasts. However, observations suggest that global warming alters the
318
+ characteristics of the MJO 8, meaning that the applicability of machine learning models trained on historical
319
+ data for MJO predictions could be undermined by climate change in the future. Furthermore, the reservoir
320
+ model of this study can only forecast the RMM sequence and cannot directly assess the impact of the MJO
321
+ on the midlatitude weather. Thus, dynamical models are expected to continue to be an imperative tool for
322
+ predicting and understanding the behaviors of our atmosphere and it is important to make the efforts to
323
+ exploit machine learning weather predictions to advance the dynamical models.
324
+
325
+
326
+ 13
327
+
328
+ Fig. 4. Samples of best performing cases of RMM1 predictions and their errors. The (A, B, C) 10th,
329
+ (D, E, F) 26th, and (G, H, I) 50th best performing cases of RMM1 predictions initialized from 19th June
330
+ 2019, 14th April 2018, and 9th December 2015 (indicated by the red dots), respectively. (A, D, G) The
331
+ trajectories of the actual (red) and the predicted (blue) RMM1 (𝑎(𝑡) and 𝑎\(𝑡, 𝑝)) and its time-delay
332
+ coordinate (𝑏(𝑡, 𝑝) and 𝑏^(𝑡, 𝑝)) are shown on the RMM phase space and (B, E, H) as a function of the
333
+
334
+ 14
335
+ forecast lead time with the errors shown as the width of the gray shade. (C, F, I) The time series of
336
+ prediction errors measured by the cosines of the angles between (𝑎(𝑡), 𝑏(𝑡)) and (𝑎\(𝑡, 𝑝), 𝑏^(𝑡, 𝑝))
337
+ (cos(𝜃5(𝑡)) ). The gray lines at ±0.7 in panels B, E, H indicate the threshold for the error = 1.4 and
338
+ cos(𝜃5(𝑡)) = 0.7 in panels C, F, I.
339
+ Acknowledgments
340
+ Funding:
341
+ Japan Society for the Promotion of Science KAKENHI Grants, 21K13991 and 20H05730 (TS)
342
+ Japan Society for the Promotion of Science KAKENHI Grants, 22K17965 (KN)
343
+ Japan Science and Technology Agency PRESTO, 22724051(KN)
344
+ Japan Society for the Promotion of Science KAKENHI Grants, 20H01819 (TY)
345
+ Japan Society for the Promotion of Science KAKENHI Grants, 20H05728 and 22H01297 (DT)
346
+ MEXT Program for Promoting Researches on the Supercomputer Fugaku, hp210166 and hp220167 (DT)
347
+ Japan Society for the Promotion of Science KAKENHI Grants, 20J11246 (TJ)
348
+ Japan Society for the Promotion of Science KAKENHI Grants, 19KK0067 and 21K18584 (YS)
349
+ "Joint Usage/Research Center for Interdisciplinary Large-scale Information Infrastructures" in Japan,
350
+ jh210027and jh220007, (YS)
351
+ HPCI System Research Project, hp210072 (YS)
352
+ Japan Society for the Promotion of Science KAKENHI Grants, 16H04048 and 20H05729 (HM)
353
+
354
+ Author contributions:
355
+ Conceptualization: KN, TS, DT, TY, HM, YS
356
+ Methodology: KN, TY, TS, HM, YS
357
+ Investigation: KN, TS, DT, HM, YS
358
+ Visualization: TS, KN
359
+
360
+ 15
361
+ Funding acquisition: TS, HM, KN, DT, TJ, TY, YS
362
+ Supervision: HM, YS
363
+ Writing – original draft: TS, KN, HM, YS
364
+ Writing – review and editing: TS, HM, KN, DT, TJ, TY, YS
365
+
366
+ Competing interests: Authors declare that they have no competing interests.
367
+
368
+ Data availability
369
+ NOAA-OLR data are available at https://www.psl.noaa.gov/data/gridded/data.olrcdr.interp.html .
370
+ NCEP-DOE reanalysis data for zonal wind data are available at
371
+ https://psl.noaa.gov/data/gridded/data.ncep.reanalysis2.html .
372
+ The original Wheeler and Hendon 2004 RMM time series are available at
373
+ http://www.bom.gov.au/climate/mjo/ .
374
+
375
+ Code availability
376
+ All source codes of our reservoir model, filter-function of the RB filter, input and output data of the
377
+ reservoir computing, and the list of MJO events will be provided via zenodo before publication of this
378
+ work.
379
+
380
+ References
381
+ 1.
382
+ Zhang, C. Madden–Julian Oscillation: Bridging Weather and Climate. Bull. Am. Meteorol. Soc. 94,
383
+ 1849–1870 (2013).
384
+ 2.
385
+ Roxy, M. K. et al. Twofold expansion of the Indo-Pacific warm pool warps the MJO life cycle.
386
+ Nature 575, (2019).
387
+
388
+ 16
389
+ 3.
390
+ Ahn, M. S. et al. MJO Propagation Across the Maritime Continent: Are CMIP6 Models Better Than
391
+ CMIP5 Models? Geophys. Res. Lett. 47, e2020GL087250 (2020).
392
+ 4.
393
+ Ahn, M. S. et al. MJO simulation in CMIP5 climate models: MJO skill metrics and process-oriented
394
+ diagnosis. Clim. Dyn. 49, 4023–4045 (2017).
395
+ 5.
396
+ Jiang, X. et al. Vertical structure and physical processes of the Madden-Julian oscillation: Exploring
397
+ key model physics in climate simulations. J. Geophys. Res. Atmos. 120, 4718–4748 (2015).
398
+ 6.
399
+ Wheeler, M. C. & Hendon, H. H. An All-Season Real-Time Multivariate MJO Index: Development
400
+ of an Index for Monitoring and Prediction. Mon. Weather Rev. 132, 1917–1932 (2004).
401
+ 7.
402
+ Madden, R. A. & Julian, P. R. Detection of a 40–50 Day Oscillation in the Zonal Wind in the
403
+ Tropical Pacific. J. Atmos. Sci. 28, 702–708 (1971).
404
+ 8.
405
+ Roxy, M. K. et al. Twofold expansion of the Indo-Pacific warm pool warps the MJO life cycle.
406
+ Nature 575, 647–651 (2019).
407
+ 9.
408
+ Waliser, D. Predictability and forecasting. Intraseasonal Var. Atmos. Clim. Syst. 389–423 (2005)
409
+ doi:10.1007/3-540-27250-X_12.
410
+ 10.
411
+ Tseng, K. C., Maloney, E. & Barnes, E. A. The Consistency of MJO Teleconnection Patterns on
412
+ Interannual Time Scales. J. Clim. 33, 3471–3486 (2020).
413
+ 11.
414
+ Ahn, M. S. et al. MJO simulation in CMIP5 climate models: MJO skill metrics and process-oriented
415
+ diagnosis. Clim. Dyn. 49, 4023–4045 (2017).
416
+ 12.
417
+ Jiang, X. et al. Vertical structure and physical processes of the Madden-Julian oscillation: Exploring
418
+ key model physics in climate simulations. J. Geophys. Res. Atmos. 120, 4718–4748 (2015).
419
+ 13.
420
+ Miura, H., Satoh, M., Nasuno, T., Noda, A. T. & Oouchi, K. A Madden-Julian Oscillation Event
421
+ Realistically Simulated by a Global Cloud-Resolving Model. Science (80-. ). 318, 1763–1765 (2007).
422
+
423
+ 17
424
+ 14.
425
+ Miyakawa, T. et al. Madden–Julian Oscillation prediction skill of a new-generation global model
426
+ demonstrated using a supercomputer. Nat. Commun. 5, (2014).
427
+ 15.
428
+ Cavanaugh, N. R. et al. The skill of atmospheric linear inverse models in hindcasting the Madden–
429
+ Julian Oscillation. Clim. Dyn. 44, 897–906 (2015).
430
+ 16.
431
+ Kim, H., Vitart, F. & Waliser, D. E. Prediction of the Madden–Julian Oscillation: A Review. J. Clim.
432
+ 31, 9425–9443 (2018).
433
+ 17.
434
+ Neena, J. M., Lee, J. Y., Waliser, D., Wang, B. & Jiang, X. Predictability of the Madden–Julian
435
+ Oscillation in the Intraseasonal Variability Hindcast Experiment (ISVHE). J. Clim. 27, 4531–4543
436
+ (2014).
437
+ 18.
438
+ Randall, D. A. & Wielicki, B. A. Measurements, Models, and Hypotheses in the Atmospheric
439
+ Sciences. Bull. Am. Meteorol. Soc. 78, 399–406 (1997).
440
+ 19.
441
+ Mauritsen, T. et al. Tuning the climate of a global model. J. Adv. Model. Earth Syst. 4, 0–01 (2012).
442
+ 20.
443
+ Mauritsen, T. & Roeckner, E. Tuning the MPI-ESM1.2 Global Climate Model to Improve the Match
444
+ With Instrumental Record Warming by Lowering Its Climate Sensitivity. J. Adv. Model. Earth Syst.
445
+ 12, e2019MS002037 (2020).
446
+ 21.
447
+ Chen, N., Gilani, F. & Harlim, J. A Bayesian Machine Learning Algorithm for Predicting ENSO
448
+ Using Short Observational Time Series. Geophys. Res. Lett. 48, e2021GL093704 (2021).
449
+ 22.
450
+ Ham, Y. G., Kim, J. H. & Luo, J. J. Deep learning for multi-year ENSO forecasts. Nature 573, 568–
451
+ 572 (2019).
452
+ 23.
453
+ Mitsui, T. & Boers, N. Seasonal prediction of Indian summer monsoon onset with echo state
454
+ networks. Environ. Res. Lett. 16, 074024 (2021).
455
+ 24.
456
+ Weyn, J. A., Durran, D. R., Caruana, R. & Cresswell-Clay, N. Sub-Seasonal Forecasting With a
457
+
458
+ 18
459
+ Large Ensemble of Deep-Learning Weather Prediction Models. J. Adv. Model. Earth Syst. 13,
460
+ e2021MS002502 (2021).
461
+ 25.
462
+ Silini, R., Barreiro, M. & Masoller, C. Machine learning prediction of the Madden-Julian oscillation.
463
+ npj Clim. Atmos. Sci. 4, 1–7 (2021).
464
+ 26.
465
+ Hagos, S., Leung, L. R., Zhang, C. & Balaguru, K. An Observationally Trained Markov Model for
466
+ MJO Propagation. Geophys. Res. Lett. 49, e2021GL095663 (2022).
467
+ 27.
468
+ Jaeger, H. The “echo state” approach to analysing and training recurrent neural networks-with an
469
+ erratum note. Bonn, Ger. Ger. Natl. Res. Cent. Inf. Technol. GMD Tech. Rep. 148, 13 (2001).
470
+ 28.
471
+ Jaeger, H. & Haas, H. Harnessing Nonlinearity: Predicting Chaotic Systems and Saving Energy in
472
+ Wireless Communication. Science (80-. ). 304, 78–80 (2004).
473
+ 29.
474
+ Lu, Z. et al. Reservoir observers: Model-free inference of unmeasured variables in chaotic systems.
475
+ Chaos An Interdiscip. J. Nonlinear Sci. 27, 041102 (2017).
476
+ 30.
477
+ Lu, Z., Hunt, B. R. & Ott, E. Attractor reconstruction by machine learning. Chaos An Interdiscip. J.
478
+ Nonlinear Sci. 28, 061104 (2018).
479
+ 31.
480
+ Arcomano, T. et al. A Machine Learning-Based Global Atmospheric Forecast Model. Geophys. Res.
481
+ Lett. 47, e2020GL087776 (2020).
482
+ 32.
483
+ Nakai, K. & Saiki, Y. Machine-learning inference of fluid variables from data using reservoir
484
+ computing. Phys. Rev. E 98, 023111 (2018).
485
+ 33.
486
+ Nakai, K. & Saiki, Y. Machine-learning construction of a model for a macroscopic fluid variable
487
+ using the delay-coordinate of a scalar observable. Discret. Contin. Dyn. Syst. - Ser. S 14, 1079 (2020).
488
+ 34.
489
+ Kobayashi, M. U., Nakai, K., Saiki, Y. & Tsutsumi, N. Dynamical system analysis of a data-driven
490
+ model constructed by reservoir computing. Phys. Rev. E 104, 044215 (2021).
491
+
492
+ 19
493
+ 35.
494
+ Suematsu, T. & Miura, H. Zonal SST Difference as a Potential Environmental Factor Supporting
495
+ the Longevity of the Madden–Julian Oscillation. J. Clim. 31, 7549–7564 (2018).
496
+ 36.
497
+ Duchon, C. E. Lanczos Filtering in One and Two Dimensions. J. Appl. Meteorol. 18, 1016–1022
498
+ (1979).
499
+ 37.
500
+ Sauer, T. & Yorke, J. A. Rigorous verification of trajectories for the computer simulation of
501
+ dynamical systems. Nonlinearity 4, 961 (1991).
502
+ 38.
503
+ Takens, F. Detecting strange attractors in turbulence. in Dynamical Systems and Turbulence,
504
+ Warwick 1980 (eds. Rand, D. & Young, L.-S.) 366–381 (Springer Berlin Heidelberg, 1981).
505
+ 39.
506
+ Lin, H., Brunet, G. & Derome, J. Forecast Skill of the Madden–Julian Oscillation in Two Canadian
507
+ Atmospheric Models. Mon. Weather Rev. 136, 4130–4149 (2008).
508
+ 40.
509
+ Vitart, F. Evolution of ECMWF sub-seasonal forecast skill scores. Q. J. R. Meteorol. Soc. 140,
510
+ 1889–1899 (2014).
511
+ 41.
512
+ Kim, H., Vitart, F. & Waliser, D. E. Prediction of the Madden–Julian Oscillation: A Review. J. Clim.
513
+ 31, 9425–9443 (2018).
514
+ 42.
515
+ Rashid, H. A., Hendon, H. H., Wheeler, M. C. & Alves, O. Prediction of the Madden-Julian
516
+ oscillation with the POAMA dynamical prediction system. Clim. Dyn. 36, 649–661 (2011).
517
+ 43.
518
+ Slingo, J. et al. Ambitious partnership needed for reliable climate prediction. Nat. Clim. Chang.
519
+ 2022 126 12, 499–503 (2022).
520
+ 44.
521
+ Pohl, B. & Matthews, A. J. Observed Changes in the Lifetime and Amplitude of the Madden–Julian
522
+ Oscillation Associated with Interannual ENSO Sea Surface Temperature Anomalies. J. Clim. 20,
523
+ 2659–2674 (2007).
524
+ 45.
525
+ Suematsu, T. & Miura, H. Changes in the Eastward Movement Speed of the Madden–Julian
526
+
527
+ 20
528
+ Oscillation with Fluctuation in the Walker Circulation. J. Clim. 35, 211–225 (2022).
529
+ 46.
530
+ Lukoševičius, M. & Jaeger, H. Reservoir computing approaches to recurrent neural network training.
531
+ Comput. Sci. Rev. 3, 127–149 (2009).
532
+ 47.
533
+ Willoughby, R. A. Solutions of Ill-Posed Problems (A. N. Tikhonov and V. Y. Arsenin). SIAM Rev.
534
+ 21, 266–267 (1979).
535
+ 48.
536
+ Pathak, J., Lu, Z., Hunt, B. R., Girvan, M. & Ott, E. Using machine learning to replicate chaotic
537
+ attractors and calculate Lyapunov exponents from data. Chaos An Interdiscip. J. Nonlinear Sci. 27,
538
+ 121102 (2017).
539
+ 49.
540
+ Liebmann, B. & Smith, C. Description of a complete (interpolated) outgoing longwave radiation
541
+ dataset. Bull. Amer. Met. Soc. 77, 1275–1277 (1996).
542
+ 50.
543
+ Kanamitsu, M. et al. NCEP–DOE AMIP-II Reanalysis (R-2). Bull. Am. Meteorol. Soc. 83, 1631–
544
+ 1644 (2002).
545
+
546
+
547
+
548
+
549
+ 21
550
+ Methods
551
+
552
+ Reservoir computing technique
553
+ The method of determining the output weight matrix 𝑾∗
554
+ #$%
555
+ of our reservoir computing machine learning
556
+ model is described. The time development of the reservoir state vector 𝒓(𝑙 Δ 𝑡) is determined by:
557
+ 𝒓(𝑡 + Δ 𝑡) = (1 − 𝛼) 𝒓(𝑡) + 𝛼 tanh4𝑨 𝒓 (𝑡) + 𝑾!" 𝒖(𝑡)6 ,
558
+ (1 − M)
559
+ where {𝒖(𝑙 Δ 𝑡)} (−𝐿3 ≤ 𝑙 ≤ 𝐿) is the training time series data, 𝐿3 is the transient time, and 𝐿 is the time
560
+ length to determine 𝑾∗
561
+ #$%
562
+ . For given random matrices 𝑨 and 𝑾!" , we determine 𝑾#$% so that the
563
+ following quadratic form takes the minimum:
564
+ lm𝑾#$% 𝒓 (𝑙Δ𝑡) − 𝒖4(𝑙 + 1)Δ𝑡6m
565
+ , + 𝛽[𝑇𝑟 (𝑾#$% 𝑾#$%
566
+ -
567
+ )],
568
+ 8
569
+ 963
570
+
571
+ (2 − M)
572
+ where ‖𝒒‖, = 𝒒𝑻𝒒 for a vector 𝒒. The minimizer is
573
+ 𝑾∗
574
+ #$% = 𝛿𝑼𝛿𝑹-(𝛿𝑹𝛿𝑹- + 𝛽 𝑰)2+
575
+ (3 − M)
576
+ where 𝑰 is the 𝑁 × 𝑁 identity matrix, 𝛿𝑹 and 𝛿𝑼 are the matrices whose 𝑙-th column is 𝒓 (𝑙Δ𝑡) and
577
+ 𝒖 (𝑙Δ𝑡), respectively. (see Lukosevivcius and Jaeger (2009)46 P.140 and Tikhonov and Arsenin (1977)47
578
+ Chapter 1 for details).
579
+ Note that 𝑨 is chosen to have a maximum eigenvalue 𝜌 (|𝜌| < 1) in order for eqn. (2-M) to satisfy the
580
+ so called echo state property. It is known that addition of noises to the training time series data is potentially
581
+ useful in the construction of a data-driven model 22. Further details on the reservoir computing can be found
582
+ in preceding literatures 32–34,48.
583
+ The set of parameter values used to construct the reservoir computing model is shown in Table 1. We
584
+ determine 𝑾#$% using the training time series data 𝒖, which in this case is the RMM1 data from 30th
585
+ December 1986 (𝑡 = 0) to 29th December 2011 (𝑡 = 9131). The optimal reservoir computing model was
586
+
587
+ 22
588
+ selected based on predictions of the RMM1 initialized every day between 8th April 2014 and 16th July 2014
589
+ by using 𝑾#$% for a given 𝑨 and 𝑾!" . We selected a model with the smallest prediction error
590
+ max
591
+ ;∈[+,>]|𝑢+(𝑡) − 𝑢\+(𝑡)| and max
592
+ ;∈[+,@3]|𝑢+(𝑡) − 𝑢\+(𝑡)|, where 𝑢+(𝑡) is the first component of 𝒖 and 𝑢\+(𝑡) is the
593
+ predicted variables of 𝑢+(𝑡).
594
+
595
+ parameter
596
+ value
597
+ 𝑀
598
+ Dimension of input and output variables
599
+ 7
600
+ 𝑁
601
+ Dimension of reservoir state vector
602
+ 1000
603
+ Δt
604
+ Time step of the model
605
+ 1 (day)
606
+ ρ
607
+ Maximal eigenvalue of 𝑨
608
+ 0.8
609
+ α
610
+ Nonlinearity degree in the model
611
+ 0.7
612
+ β
613
+ Regularization parameter
614
+ 0.01
615
+ Δτ
616
+ Delay-time for input and output variables
617
+ 6 (day)
618
+
619
+ Table 1. The list of parameters and their values for the selected reservoir computing model
620
+
621
+
622
+ MJO detection method
623
+
624
+ The RMM is calculated from the combined empirical orthogonal functions of the outgoing
625
+ longwave radiation data from National Oceanic and Atmospheric Administration 49, and zonal wind data
626
+ from National Centers for Environmental Prediction-Department of Energy reanalysis 50. With the
627
+ exception of replacing RMM2 with the 12-day time delay coordinate of RMM1, the orientation of RMM1
628
+ and definitions of the RMM phases follow the convention introduced by Wheeler and Hendon 6. The MJO
629
+ events were identified from time sequences that were projected on to the RMM index from phase 2 to phase
630
+ 7, while satisfying the following four conditions employed by Suematsu and Miura (2018) 35: (1) Phases
631
+ do not skip forward nor recede backward by more than one phase. (2) The average amplitude is greater
632
+ than the critical value of 0.8. (3) Period of consecutive days with amplitude below 0.8 is less than 15. (4)
633
+ Transition from phase 2 to phase 7 takes 20 to 90 days.
634
+
635
+
636
+ 23
637
+ Supplementary materials
638
+
639
+ Validity of the Machine Learning-RMM Phase Space
640
+
641
+
642
+ The relevance of employing the RMM phase space spanned by RMM1 and its delay coordinate, the
643
+ machine learning RMM (ML-RMM) phase space, to describe the MJO instead of that spanned by RMM1
644
+ and RMM2 is discussed. Conventionally, MJO events are identified using RMM phase space spanned by
645
+ the first two orthogonal functions, RMM1 and RMM2, of 20 - 120 day Lanczos bandpass filtered 36
646
+ outgoing longwave radiation and zonal winds at 850hPa and 200 hPa. Figure S1 compares the composites
647
+ of 1979 – 2020 December to February outgoing longwave radiation for each of the RMM phases spanned
648
+ by the conventional RMM1 and RMM2 with ML-RMM phase space.
649
+ The composites indicate that the definition of the ML-RMM phases (Fig. S1 B) can capture the
650
+ characteristic of the MJO convection to shift eastward from the Indian Ocean to the Western Pacific as
651
+ the conventional method (Fig. S1 A). We note however, that compared to the conventional method, the
652
+ convective signals over the Indian Ocean in the ML-RMM phase 2 is weaker. This may be a caveat to our
653
+ method that arises from replacing the RMM2 with the delay coordinate of RMM1, since the structure of
654
+ the eigenvector of RMM2 reflects the state of the atmosphere with deep convection over the Indian
655
+ Ocean (see Fig. 1 in 30). Despite the abovementioned concern, the method employed in this study is
656
+ capable of adequately tracking MJO events on the RMM phase space (Fig. 2C).
657
+
658
+
659
+ 24
660
+
661
+
662
+
663
+
664
+ Fig. S1. Composites of 1979 – 2020 December to February outgoing longwave radiation for each of the
665
+ RMM phases on the (A) conventional RMM1 and RMM2 phase space calculated from 20-120 days
666
+ Lanczos bandpass filtered data and (B) on the ML-RMM calculated from the 20-120 days RB filtered
667
+ data.
668
+
669
+
670
+
671
+ 25
672
+ Examples of best performing cases in terms of phase prediction
673
+ The best performing prediction cases in terms of ML-RMM phase predictions are examined. Figure S2
674
+ shows the best three cases evaluated by the first day the cosine of the phase difference between the actual
675
+ (𝑎(𝑡), 𝑏(𝑡)) and the predicted (𝑎\(𝑡, 𝑝), 𝑏^(𝑡, 𝑝)) vector, cos(𝜃5(𝑡)), becomes less than 0.7. The first (Fig.
676
+ S2. A, D), second (Fig. S2. B, E) and third (Fig. S2.C, F) best performing cases are the predictions of ML-
677
+ RMM initiated on 10th October 2017, 26th March 2019, and 18th April 2018, respectively. In all three
678
+ cases, the predictions stay in phase (cos(𝜃5(𝑡)) > 0.7) for longer than 120 days. However, there is a
679
+ tendency for the amplitudes to be underestimated in these cases, which leads to growth in error as measured
680
+ by the distance between (𝑎(𝑡), 𝑏(𝑡)) and (𝑎\(𝑡, 𝑝), 𝑏^(𝑡, 𝑝)) from early stages of the predictions. Thus,
681
+ while the long predictability of the RMM phases over 120 days suggest the possibility of predicting the
682
+ MJO over a season (three months), overcoming the difficulty of accurately predicting the RMM phase and
683
+ amplitude simultaneously remains a challenge.
684
+
685
+ Fig. S2. The three best performing prediction of ML-RMM time series in terms of phase
686
+ prediction. (A, B, C) Predictions of ML-RMM initialized from (A, D) 2nd October 2017, (B, E) 18th
687
+ March 2019, and (C, F) 10th April 2018, which are the three best prediction cases measured by the
688
+ cosine of the phase difference between the actual (𝑎(𝑡), 𝑏(𝑡)) and the predicted vectors (cos(𝜃5(𝑡))).
689
+
690
+ 26
691
+ (D, E, F) show the time evolution of the cos(𝜃5(𝑡)). The width of the grey shades in A, B, C indicates
692
+ the error measured by the distance between (𝑎(𝑡), 𝑏(𝑡)) and (𝑎\(𝑡, 𝑝), 𝑏^(𝑡, 𝑝)).
693
+
694
+
CdAzT4oBgHgl3EQfTvxb/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
EdFRT4oBgHgl3EQfyTjG/content/2301.13645v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc6f577740e4027ff6f8f4d4d3efbeab9bbabf149d89c5ce4365ad11b3e40417
3
+ size 198426
EdFRT4oBgHgl3EQfyTjG/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c0865a6f243a107c6dab9b0b5753affcb10a8b9f3511588d4f6e03b935ddef4e
3
+ size 1376301
EdFRT4oBgHgl3EQfyTjG/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e66ed1e8693c08df70df037316da5973ea590e427864d113ffe15ad47d69cd2
3
+ size 60458
FtA0T4oBgHgl3EQfBf8H/content/tmp_files/2301.01975v1.pdf.txt ADDED
@@ -0,0 +1,2941 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ STABILIZED WEIGHTED REDUCED ORDER METHODS FOR
2
+ PARAMETRIZED ADVECTION-DOMINATED OPTIMAL CONTROL
3
+ PROBLEMS GOVERNED BY PARTIAL DIFFERENTIAL EQUATIONS WITH
4
+ RANDOM INPUTS
5
+ FABIO ZOCCOLAN1, MARIA STRAZZULLO2, AND GIANLUIGI ROZZA3
6
+ Abstract. In this work, we analyze Parametrized Advection-Dominated distributed Optimal
7
+ Control Problems with random inputs in a Reduced Order Model (ROM) context. All the simula-
8
+ tions are initially based on a finite element method (FEM) discretization; moreover, a space-time
9
+ approach is considered when dealing with unsteady cases. To overcome numerical instabilities that
10
+ can occur in the optimality system for high values of the P´eclet number, we consider a Streamline
11
+ Upwind Petrov–Galerkin technique applied in an optimize-then-discretize approach.
12
+ We com-
13
+ bine this method with the ROM framework in order to consider two possibilities of stabilization:
14
+ Offline-Only stabilization and Offline-Online stabilization. Moreover we consider random parame-
15
+ ters and we use a weighted Proper Orthogonal Decomposition algorithm in a partitioned approach
16
+ to deal with the issue of uncertainty quantification. Several quadrature techniques are used to
17
+ derive weighted ROMs: tensor rules, isotropic sparse grids, Monte-Carlo and quasi Monte-Carlo
18
+ methods. We compare all the approaches analyzing relative errors between the FEM and ROM
19
+ solutions and the computational efficiency based on the speedup-index.
20
+ 1. Introduction
21
+ Here we present a numerical study concerning stabilized Parametrized Advection-Dominated Op-
22
+ timal Control Problems (OCP(µ)s) with random inputs in a Reduced Order Methods (ROMs)
23
+ framework. As a matter of fact, engineering and scientific applications often need very fast evalu-
24
+ ations of the numerical solutions for many parameters that characterize the problem, for instance
25
+ in real-time scenarios. A solution to these many-query situations can be to exploit the parameter
26
+ dependence of the OCP(µ)s using ROMs [6, 24, 41, 40, 39]. This process is divided in two stages.
27
+ The former is the offline phase, when many numerical solutions for different values of parameters
28
+ are collected considering a first discretization of the OCP(µ), such a finite element method (FEM)
29
+ one, called the high-fidelity or truth approximation. Then all parameter-independent components
30
+ are calculated and stored, and reduced spaces are built. The latter is the online phase, when all
31
+ parameter-dependent parts and, then, the reduced solutions are computed. More precisely, to deal
32
+ with the randomness which is hidden in the parameters, we consider a modification of the Proper
33
+ Orthogonal Decomposition (POD) that takes into account the probability distribution of the random
34
+ inputs: the weighted POD (wPOD) [61, 60]. We apply this procedure in a partitioned approach,
35
+ following good results shown in literature [30, 34, 53, 63]. As this algorithm aims to minimize the
36
+ expectation of the square error between the truth and the ROM solutions, we can identify different
37
+ types of weighted ROMs (wROMs) [11, 15, 13, 16, 17, 18, 49, 59, 61, 60] based on the chosen quad-
38
+ rature rules. In this work, we will exploit Monte-Carlo and Quasi Monte-Carlo procedures, tensor
39
+ rules based on Gauss-Jacobi and Clenshaw-Curtis quadrature techniques, and Smolyak isotropic
40
+ sparse grids.
41
+ 1 Section de Math´ematiques, ´Ecole Polytechnique F´ed´erale de Lausanne, 1015 Lausanne, Switzerland,
42
+ email: fabio.zoccolan@epfl.ch
43
+ 2 DISMA, Politecnico di Torino, Corso Duca degli Abruzzi 24, Turin, Italy.
44
+ email: maria.strazzullo@polito.it
45
+ 3 mathLab, Mathematics Area, SISSA, via Bonomea 265, I-34136 Trieste, Italy.
46
+ email: gianluigi.rozza@sissa.it
47
+ 1
48
+ arXiv:2301.01975v1 [math.NA] 5 Jan 2023
49
+
50
+ 2
51
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
52
+ The optimization problem will always concern a linear-quadratic cost functional. We use FEM
53
+ as the truth approximation, both for steady and unsteady problems. At a first level, FEM approx-
54
+ imations of stochastic steady OCP(µ)s have been already presented, for example, in [26] consid-
55
+ ering stochastic PDEs. In the parabolic case, we discretize time-dependency via a space-time ap-
56
+ proach [25, 50, 51]. Concerning stabilization, we considered the Streamline Upwind Petrov–Galerkin
57
+ (SUPG) [9, 28, 38] suitably combined with the ROM setting in an optimize-then-discretize approach.
58
+ We exploit two possibilities: when stabilization only occurs in the offline phase, Offline-Only stabi-
59
+ lization or when SUPG is applied in both phases, Offline-Online stabilization.
60
+ Stabilized Advection-Dominated problems in a ROM framework without control are studied,
61
+ for instance, in [37, 59], both for steady and unsteady cases. Instead, in [11] wROMs for generic
62
+ OCP(µ)s are applied to experiments concerning environmental sciences. Instead, SUPG Advection-
63
+ Dominated distributed OCP(µ)s are analyzed in a deterministic context in [63], both for elliptic and
64
+ parabolic experiments. To the best of our knowledge, this is the first time that stabilized Advection-
65
+ Dominated OCP(µ)s with random inputs are analyzed in a ROM context, both for elliptic and
66
+ parabolic problems.
67
+ This work is arranged as follows. In Section 2, we introduce linear-quadratic optimal control
68
+ theory for PDEs and its FEM discretization. Section 3 firstly concerns the basic theory of SUPG
69
+ stabilization for Advection-Dominated PDEs in an optimize-then-discretize approach [19], then the
70
+ space-time procedure that will be used is presented. wROMs features will be illustrated in Sec-
71
+ tion 4. Section 5 will concern numerical simulations. Two Advection-Dominated problems under
72
+ distributed control and random inputs will be analyzed: the Graetz-Poiseuille Problem under ge-
73
+ ometrical parametrization and the Propagating Front in a Square Problem. We will compare the
74
+ wPOD procedures through relative errors between the FEM and the ROM solutions and computa-
75
+ tional time considering the speedup-index. Finally, conclusions follow in Section 6.
76
+ 2. Problem Formulation for Random Input Optimal Control Problems
77
+ 2.1. Mathematical Setting. Let Ω be an open and bounded regular domain in R2, where ΓN and
78
+ ΓD will indicate the portions of the boundary ∂Ω where Neumann and Dirichlet boundary conditions
79
+ are imposed, respectively. With the symbol Ωobs ⊆ Ω the observation domain will be indicated, i.e.
80
+ the subset of the domain where we seek the state variable to be similar to a desired solution profile
81
+ yd ∈ Y , with Y Hilbert space, in a sense that will be specified later. For time-dependent problems we
82
+ will also take into account the time interval (0, T) ⊂ R+. Let us consider a compact set P ⊆ Rp, for
83
+ natural number p. We will call P and as the parameter space and a p-vector µ ∈ P is the parameter
84
+ of our Parametric OCP(µ)s. As the setting is completely general, for instance µ can characterize
85
+ our yd or geometrical and physical properties of the problem. Furthermore, we denote with B(Q, R)
86
+ the space of linear continuous operators between Banach spaces Q and R.
87
+ The triplet (A, F, P) will denote a complete probability space, composed by A, which is the set
88
+ of outcomes ω ∈ A, F, that is a σ-algebra of events, and P : F → [0, 1] with P(A) = 1, which is
89
+ the chosen probability measure. As dealing with random input OCP(µ)s, the parameter µ will be
90
+ a real-valued random vector. In detail, µ : (A, F) → (Rp, B) is a measurable function, where B is
91
+ the Borel σ-algebra on Rp. The distribution function of µ : A → P ⊂ Rp, being P the image of µ,
92
+ is defined as Pµ : P → [0, 1] such that
93
+ (1)
94
+ ∀µ ∈ P,
95
+ Pµ(µ) = P(ω ∈ A : µ(ω) ≤ µ).
96
+ Let dPµ(µ) denote the distribution measure of µ, i.e.,
97
+ (2)
98
+ ∀H ⊂ P,
99
+ P(µ ∈ H) =
100
+
101
+ H
102
+ dPµ(µ).
103
+ We assume that µ admits a Lebesgue density, i.e. dPµ(µ) is absolutely continuous with respect
104
+ to the Lebesgue measure dµ. This practically means that there exists a probability density func-
105
+ tion ρµ : P → R+ such that ρµ(µ)dµ = dPµ(µ). It is worth to notice that the measure space
106
+ (P, B(P), ρµ(µ)dµ) is isometric to (A, F, P) under the random vector µ.
107
+ The aim of this work is to analyze random input OCP(µ)s from the numerical point of view.
108
+
109
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
110
+ 3
111
+ Problem 2.1.1 (Random Input Parametric Optimal Control Problem). Consider the state equation
112
+ E : Y × U → Q, with Y, U, and Q real Banach spaces, satisfying a set of boundary and/or initial
113
+ conditions, and a real functional J : Y ×U → R. Then for Pµ-a.e. find the pair
114
+
115
+ y(µ), u(µ)
116
+
117
+ ∈ X :=
118
+ Y × U that minimizes cost functional J (y(µ), u(µ); µ) under the constraint E(y(µ), u(µ); µ) = 0.
119
+ Let Xad be the set of all couples (y, u) solutions of E: we will only consider the case of full
120
+ admissibility, i.e. when Xad = Y × U. Problem 2.1.1 looks for minimizers among all state-control
121
+ pairs such that:
122
+ min
123
+ (y(µ),u(µ))∈Y ×U J (y(µ), u(µ); µ) s.t. E(y(µ), u(µ); µ) = 0.
124
+ This can be achieved through the research of the critical points of the Lagrangian operator
125
+ L : Y × U × Q∗ → R defined as:
126
+ (3)
127
+ L(y(µ), u(µ), p(µ); µ) = J (y(µ), u(µ); µ) + ⟨p(µ), E(y(µ), u(µ); µ)⟩Q∗Q,
128
+ where p(µ) is a Lagrange multiplier belonging to the adjoint space Q∗, the dual space of Q. For
129
+ the sake of notation we write y := y(µ), u := u(µ) and p := p(µ). In case that Pµ is the uniform
130
+ distribution with support in P, then Problem 2.1.1 is called to be deterministic problem. In this
131
+ work linear-quadratic problems will be involved.
132
+ Definition 2.1.2 (Linear-Quadratic OCP(µ). Let us consider the bilinear forms m : Z × Z → R
133
+ and n : U × U → R, which are symmetric and continuous, where Z is a Banach space called the the
134
+ observation space. Fix α > 0, a constant called the penalization parameter and consider a quadratic
135
+ objective functional J of the form
136
+ (4)
137
+ J (y, u; µ) = 1
138
+ 2m (Oy(µ) − zd(µ), Oy(µ) − zd(µ)) + α
139
+ 2 n(u(µ), u(µ)),
140
+ where O : Y → Z is a linear and bounded operator called the observation map and zd(µ) ∈ Z is
141
+ the observed desired solution profile. Consider an affine map E defined as
142
+ (5)
143
+ E(y, u; µ) = A(µ)y + C(µ)u − f(µ),
144
+
145
+
146
+ y(µ), u(µ)
147
+
148
+ ∈ Y × U,
149
+ where A(µ) ∈ B(Y, Q), C(µ) ∈ B(U, Q) and f(µ) ∈ Q.
150
+ Then an OCP(µ)s with the above properties is said a Linear-Quadratic Optimal Control Problem.
151
+ For Linear-Quadratic OCP(µ)s well-posedness of Problem 2.1.2 yields [7, 8]. More precisely, the
152
+ reader can refer to [10] to a comparison between the Lagrangian approach for the full-admissibility
153
+ case and the adjoint one. Via the functional derivative of L, we obtain a optimality system to be
154
+ solved to find the unique solution. In this case, this reads as finding (y, u, p) ∈ Y × U × Q∗ that
155
+ satisfies [10],
156
+ (6)
157
+
158
+
159
+
160
+
161
+
162
+ DyL(y, u, p; µ)(¯y) = 0 =⇒ m(Oy, O¯y; µ) + ⟨A∗(µ)p, ¯y⟩Y ∗Y = m (O¯y, zd; µ) ,
163
+ ∀¯y ∈ Y,
164
+ DuL(y, u, p; µ)(¯u) = 0 =⇒ αn(u, ¯u; µ) + ⟨C∗(µ)p, ¯u⟩U ∗U = 0,
165
+ ∀¯u ∈ U,
166
+ DpL(y, u, p; µ)(¯p) = 0 =⇒ ⟨¯p, A(µ)y + C(µ)u⟩Q∗Q = ⟨¯p, f(µ)⟩Q∗Q,
167
+ ∀¯p ∈ Q∗.
168
+ In system (6), the first equation is called the adjoint equation, the second one is the gradient
169
+ equation and the last one is state equation.
170
+ Remark 2.1.3 (Notation). For the sake of notation, when Hilbert spaces will be taken into account,
171
+ bilinear forms A, B and their adjoint counterparts will be indicate uniquely as:
172
+ ⟨A(µ)y, p⟩QQ∗ := a(y, p; µ)
173
+ ⟨C(µ)u, p⟩QQ∗ := c(u, p; µ).
174
+ Remark 2.1.4 (Parabolic Problems). Concerning unsteady problems, one must add more hypotheses
175
+ to the mathematical setting of Linear-Quadratic OCP(µ)ss to reach well-posedness. We will consider
176
+ the following Hilbert spaces Y = L2(0, T; Y ), Y∗ = L2 (0, T; Y ∗), Z = L2 (0, T; Z), U = L2(0, T; U)
177
+ with respective norms given by
178
+ (7) ∥y∥2
179
+ Y :=
180
+ T
181
+
182
+ 0
183
+ ∥y∥2
184
+ Y dt,
185
+ ∥y∥2
186
+ Y∗ :=
187
+ T
188
+
189
+ 0
190
+ ∥y∥2
191
+ Y ∗dt,
192
+ ∥z∥2
193
+ Z :=
194
+ T
195
+
196
+ 0
197
+ ∥z∥2
198
+ Zdt,
199
+ and
200
+ ∥u∥2
201
+ U :=
202
+ T
203
+
204
+ 0
205
+ ∥u∥2
206
+ Udt.
207
+
208
+ 4
209
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
210
+ Then we define the Hilbert space Yt := {y ∈ Y
211
+ s.t.
212
+ ∂ty ∈ Y∗}. For parabolic problems we will
213
+ also consider the case of full-admissibility as Xad = Yt × U [5, 54, 55].
214
+ 2.2. High-Fidelity Discretization. In this work, the discretization of the optimality sistem (6)
215
+ follows an one shot or all-at-once approach [25, 50, 51]. When we will consider Advection-Dominated
216
+ OCP(µ)s, a stabilization technique will be also needed. Therefore, a SUPG method will be applied
217
+ in a optimize-then-discretize approach, as we will see in Section 3.
218
+ This means that firstly the
219
+ optimality conditions are computed obtaining system (6) and then we discretize and stabilize it.
220
+ Concerning numerical implementation, we use a FEM discretization for all three variables, where
221
+ Th is a regular triangularization on Ω. Its elements K are triangles and the parameter h denotes the
222
+ mesh size, i.e. the maximum diameter of an element of the chosen grid. In addition, we define
223
+ Ωh := int
224
+ � �
225
+ K∈Th
226
+ K
227
+
228
+ ,
229
+ as a quasi-uniform mesh for Ω. Considering Pr(K) as the space of polynomials of degree at most
230
+ equal to r defined on K and defining
231
+ XN ,r =
232
+
233
+ qN ∈ C(¯Ω) : qN
234
+ |K ∈ Pr(K), ∀K ∈ Th
235
+
236
+ we set Y N = Y ∩ XN ,r, U N = U ∩ XN ,r and
237
+
238
+ QN �∗ = Q∗ ∩ XN ,r. In this work, the numerical
239
+ implementation will always made by a P1-FEM approximation and the same triangulation Th for
240
+ Y N , U N , and
241
+
242
+ QN �∗. A similar discussion can be made for time-dependent problem, as we will
243
+ see in Section 3.2. This first discretization procedure will be indicated as the truth or high-fidelity
244
+ approximation.
245
+ From now on, Y, U, Q will be always Hilbert spaces and we will consider the Identity restricted to
246
+ our observation domain Ωobs as the Observation function O for both steady and unsteady problems.
247
+ Therefore, Z = Y for steady problems and Z = Y for unsteady ones are assumed. Our desired state
248
+ will be denoted by yd and with the same symbol will also indicate its FEM discretization.
249
+ 3. SUPG stabilization for Advection-Dominated OCP(µ)s
250
+ In this work we only deal with Advection-Diffusion equations.
251
+ Definition 3.0.1 (Advection-Diffusion Equations). Let us take into account the following problem:
252
+ (8)
253
+ T(µ)y := −γ(µ)∆y + η(µ) · ∇y = f(µ) in Ω ⊂ R2,
254
+ where suitable boundary conditions are applied on ∂Ω. In addition, we require that:
255
+ • the diffusion coefficient γ : Ω → R is uniformly bounded, i.e. there exists γmax, γmin > 0 such
256
+ that
257
+ (9)
258
+ P
259
+
260
+ ω ∈ A : γmin < γ(x; µ) < γmax ∀x ∈ Ω
261
+
262
+ = 1.
263
+ • the advection field η : Ω → R2 belongs to (L∞(Ω))2 for a.e. µ ∈ P. More precisely, for a.e.
264
+ µ the following inequality holds: 0 ≥ div η(x) ≥ −ϑ, ∀x ∈ Ω, with ϑ ∈ R+
265
+ 0 ;
266
+ • f : Ω → R is an L2(Ω)-function for a.e. µ; in addition, f has bounded second moments with
267
+ respect to the integral along A and Ω.
268
+ With these hypotheses, Problem (8) is called Advection-Diffusion problem and the operator T(µ)y :=
269
+ −γ(µ)∆y + η(µ) · ∇y is said the Advection-Diffusion operator.
270
+ For more details regarding the well-posedness and theoretical results of Stochastic Advection-
271
+ Diffusion OCP(µ)s, we refer to [14, 13].
272
+ Definition 3.0.2 (P´eclet number and Advection-Dominated problem). Let us consider the FEM
273
+ discretization related to an Advection-Diffusion problem and its regular triangulation Th. For any
274
+ element K ∈ Th, the local P´eclet number is defined as [42, 38]:
275
+ (10)
276
+ PeK(x) := |η(x)|hK
277
+ 2γ(x)
278
+ ∀x ∈ K,
279
+
280
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
281
+ 5
282
+ where hK is the diameter of K. If PeK(x) > 1 ∀x ∈ K, ∀K ∈ Th, we say to study an Advection-
283
+ Dominated problem.
284
+ 3.1. Setting for Stabilized Advection-Dominated OCP(µ)s - Steady case. Numerical in-
285
+ stabilities can appear, when dealing with Advection-Dominated OCP(µ)s, i.e. when |η(µ)| ≫ γ. In
286
+ order to adjust this unpleasant behaviour without modifying the mesh size, we use the Streamline
287
+ upwind/Petrov Galerkin (SUPG) method [9, 27, 28, 42] in a optimize-then-discretize approach [19].
288
+ This assures the strongly consistency of the optimality system [19]. For the sake of simplicity, we
289
+ define our Advection-Dominated problem on H1
290
+ 0(Ω) and we do not indicate the parameter depen-
291
+ dence. We denote with T ∗ the adjoint operator related to T. This last operator can be split into its
292
+ symmetric and skew-symmetric parts as T = TS + TSS [42], where:
293
+ (11)
294
+ symmetric part: TSy = −γ∆y − 1
295
+ 2(div η)y,
296
+ ∀y ∈ H1
297
+ 0(Ω),
298
+ skew-symmetric part: TSSy = η · ∇y + 1
299
+ 2(div η)y,
300
+ ∀y ∈ H1
301
+ 0(Ω).
302
+ This two parts can be immediately recovered using the formulae:
303
+ (12)
304
+ TS = T + T ∗
305
+ 2
306
+ ,
307
+ TSS = T − T ∗
308
+ 2
309
+ .
310
+ After having considered FEM spaces, the stabilization occurs in the bilinear and linear terms involved
311
+ in the state and the adjoint equations. Instead, the gradient equation is left unstabilized [19]. We
312
+ recall that we use distributed control.
313
+ We defined the stabilized bilinear form as and cs, and the stabilized forcing term Fs as
314
+ (13)
315
+ as
316
+
317
+ yN , qN ; µ
318
+
319
+ := a
320
+
321
+ yN , qN ; µ
322
+
323
+ +
324
+
325
+ K∈Th
326
+ δK
327
+
328
+ TyN , hK
329
+ |η| TSSqN
330
+
331
+ K
332
+ ,
333
+ yN , qN ∈ Y N ,
334
+ cs
335
+
336
+ uN , qN ; µ
337
+
338
+ := −
339
+
340
+
341
+ uN qN −
342
+
343
+ K∈Th
344
+ δK
345
+
346
+ uN , hK
347
+ |η| TSSqN
348
+
349
+ K
350
+ ,
351
+ uN ∈ U N , qN ∈ Y N ,
352
+ Fs
353
+
354
+ qN ; µ
355
+
356
+ := F
357
+
358
+ qN ; µ
359
+
360
+ +
361
+
362
+ K∈Th
363
+ δK
364
+
365
+ f(µ), hK
366
+ |η| TSSqN
367
+
368
+ K
369
+ ,
370
+ ∀qN ∈ Y N .
371
+ where δK is a local positive dimensionless parameter related to the element K ∈ Th, consequently
372
+ it can be different for each triangle, and (·, ·)K is the inner scalar product in L2(K).
373
+ In (13)
374
+ a
375
+
376
+ yN , qN ; µ
377
+
378
+ =
379
+
380
+ TyN , qN �
381
+ L2(Ω) and F
382
+
383
+ qN ; µ
384
+
385
+ =
386
+
387
+ f, qN �
388
+ L2(Ω), where f collects all forcing and
389
+ lifting terms of the problem.
390
+ For the remaining conditions of the optimality system, we will always consider m and n form as
391
+ the L2(Ωobs) and the L2(Ω) products for steady problems. The adjoint equation is an Advection-
392
+ Dominated equation, too, where the advective term has opposite sign with respect to the state one:
393
+ indeed, T ∗ = TS − TSS from (12). We use the next SUPG forms for zN ∈ Y N :
394
+ (14)
395
+ a∗
396
+ s
397
+
398
+ zN , pN ; µ
399
+
400
+ := a∗ �
401
+ zN , pN ; µ
402
+
403
+ +
404
+
405
+ K∈Th
406
+ δa
407
+ K
408
+
409
+ (TS − TSS)pN , hK
410
+ |η| (−TSS) zN
411
+
412
+ K
413
+ ,
414
+
415
+ yN − yd, zN ; µ
416
+
417
+ s :=
418
+
419
+ Ωobs
420
+ (yN − yd)zN dx +
421
+
422
+ K∈Th|Ωobs
423
+ δa
424
+ K
425
+
426
+ yN − yd, hK
427
+ |η| (−TSS) zN
428
+
429
+ K
430
+ ,
431
+ where a∗ is the adjoint form of a, δa
432
+ K is the positive stabilization parameter of the stabilized adjoint
433
+ equation.
434
+ In our numerical experiments, we will always consider δK = δa
435
+ K.
436
+ Finally, the SUPG
437
+ optimality system for a steady OCP(µ) reads as:
438
+ (15)
439
+ discretized adjoint equation:
440
+ a∗
441
+ s
442
+
443
+ zN , pN ; µ
444
+
445
+ +
446
+
447
+ yN − yd, zN ; µ
448
+
449
+ s = 0, ∀zN ∈ Y N ,
450
+ discretized gradient equation:
451
+ c∗�
452
+ vN , pN ; µ
453
+
454
+ + αn
455
+
456
+ uN , vN ; µ
457
+
458
+ = 0, ∀vN ∈ U N ,
459
+ discretized state equation:
460
+ as
461
+
462
+ yN , qN ; µ
463
+
464
+ + cs
465
+
466
+ uN , qN ; µ
467
+
468
+ = Fs(qN ; µ), ∀qN ∈
469
+
470
+ QN �∗ .
471
+
472
+ 6
473
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
474
+ We denote with Ks and KT
475
+ s the stiffness matrices related to the stabilized forms as and a∗
476
+ s,
477
+ respectively, M is the not-stabilized mass matrix related to n, instead, Ms is the stabilized mass
478
+ matrix related to m after stabilization, Cs is the matrix linked to stable form cs, the block CT refers
479
+ to c, and fs is the vector that contains the coefficients of the stabilized force term as components.
480
+ Moreover, we consider with the symbol y, u and p as the vectors of coefficients of yN , uN and pN ,
481
+ expressed in terms of the nodal basis of Y N , U N , (QN )∗, respectively. Finally, the discretized block
482
+ system related to (15) is:
483
+ (16)
484
+
485
+
486
+ Ms
487
+ 0
488
+ KT
489
+ s
490
+ 0
491
+ αM
492
+ CT
493
+ Ks
494
+ Cs
495
+ 0
496
+
497
+
498
+
499
+
500
+ y
501
+ u
502
+ p
503
+
504
+ � =
505
+
506
+
507
+ Msyd
508
+ 0
509
+ fs
510
+
511
+ � .
512
+ 3.2. Setting for Stabilized Advection-Dominated OCP(µ)s - Unsteady case. We show the
513
+ SUPG approach for time-dependent OCP(µ)s proposed in [63]. A classical implicit Euler discretiza-
514
+ tion is applied to all forms including time-derivatives [3, 25, 50, 54, 55, 56]. We divide the time
515
+ interval (0, T) in Nt sub-intervals of equal length ∆t := tj − tj−1, j ∈ {1, . . . , Nt}. Starting from
516
+ this framework, a discretization along time is done, where each discrete instant of time is considered
517
+ as a steady-state Advection-Dominated equation in a space-time approach [25, 50, 51, 54, 55, 56].
518
+ In addition, the SUPG stabilization occurs for time-dependent forms, too. The general scheme is
519
+ described as follows.
520
+ Let us firstly define the discrete vectors y =
521
+
522
+ yT
523
+ 1 , . . . , yT
524
+ Nt
525
+ �T , u =
526
+
527
+ uT
528
+ 1 , . . . , uT
529
+ Nt
530
+ �T and p =
531
+
532
+ pT
533
+ 1 , . . . , pT
534
+ Nt
535
+ �T , where yi ∈ Y N , ui ∈ U N and pi ∈ (QN )∗ for 1 ≤ j ≤ Nt.
536
+ Also here, yj, uj
537
+ and pj indicate the column vectors containing the coefficients of the FEM discretization for state,
538
+ control and adjoint, respectively (unlike the steady case, there are not denoted in bold style). This
539
+ implies Ntot = 3 × Nt × N as the global dimension of the block system.
540
+ We express all other
541
+ terms in based of the respective nodal basis. The vector representing the initial condition for the
542
+ state variable is y0 =
543
+
544
+ yT
545
+ 0 , 0T , . . . , 0T �T , where 0 is the zero vector in RN , yd =
546
+
547
+ yT
548
+ d1, . . . , yT
549
+ dNt
550
+ �T
551
+ is the vector including discrete time components of the discretized desired solution profile; instead,
552
+ f s =
553
+
554
+ f T
555
+ s1, . . . , f T
556
+ sNt
557
+ �T
558
+ corresponds to the stabilized forcing term. We recall that Y, U, Q are Hilbert
559
+ Spaces and, for the sake of simplicity, we assume Y N ≡ (QN )∗. So now we can see locally the time
560
+ block discretization.
561
+ • Adjoint equation: this equation is discretized backward in time using the forward Euler
562
+ method, which is equal to a backward Euler with respect to time T − t, for t ∈ (0, T) [21].
563
+ Firstly, we add a stabilized term to the form related to ∂tp and a∗ defined as:
564
+ s∗ �
565
+ zN , pN (t); µ
566
+
567
+ =
568
+
569
+ K∈Th
570
+ δK
571
+
572
+ −∂tpN (t) + T ∗pN (t), −hK
573
+ |η| TSSzN
574
+
575
+ K
576
+ ,
577
+ where we define the form
578
+ (17)
579
+ m∗
580
+ s
581
+
582
+ pN , zN ; µ
583
+
584
+ =
585
+
586
+ pN , zN �
587
+ L2(Ω) −
588
+
589
+ K∈Th
590
+ δK
591
+
592
+ pN , hK
593
+ |η| TSSzN
594
+
595
+ K
596
+ .
597
+ Then, the time discretization is: for each j ∈ {Nt − 1, Nt − 2, ..., 1}, find pN
598
+ j ∈ Y N
599
+ s.t.
600
+ (18)
601
+ 1
602
+ ∆tm∗
603
+ s
604
+
605
+ pN
606
+ j (µ) − pN
607
+ j+1(µ), zN ; µ
608
+
609
+ + a∗
610
+ s
611
+
612
+ zN , pN
613
+ j (µ); µ
614
+
615
+ = −
616
+
617
+ yN
618
+ j − ydj, zN ; µ
619
+
620
+ s
621
+ ∀zN ∈ Y N ,
622
+ Considering M T
623
+ s as the matrix inherent to m∗
624
+ s, the block subsystem reads
625
+ M T
626
+ s pj = M T
627
+ s pj+1 + ∆t
628
+
629
+ −M T
630
+ s yj − KT
631
+ s pj + M T
632
+ s ydj
633
+
634
+ for j ∈ {Nt − 1, Nt − 2, . . . , 1} .
635
+
636
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
637
+ 7
638
+ Finally, we derive the following block system:
639
+
640
+ ����
641
+ M T
642
+ s + ∆tKT
643
+ s
644
+ −M T
645
+ s
646
+ ...
647
+ ...
648
+ M T
649
+ s + ∆tKT
650
+ s
651
+ −M T
652
+ s
653
+ M T
654
+ s + ∆tKT
655
+ s
656
+
657
+ ����
658
+
659
+ ��
660
+
661
+ AT
662
+ s
663
+ p +
664
+
665
+ �����
666
+ ∆tM T
667
+ s y1
668
+ ...
669
+ ...
670
+ ∆tM T
671
+ s yNt
672
+
673
+ �����
674
+ =
675
+
676
+ �����
677
+ ∆tM T
678
+ s yd1
679
+ ...
680
+ ...
681
+ ∆tM T
682
+ s ydNt
683
+
684
+ �����
685
+ .
686
+ Setting the diagonal block matrix MT
687
+ s ∈ RN ·Nt×RN ·Nt with diagonal entries [M T
688
+ s , . . . , M T
689
+ s ],
690
+ the adjoint system to be solved is: ∆tMT
691
+ s y + AT
692
+ s p = ∆tMT
693
+ s yd.
694
+ • Gradient equation. We seek the solution of α∆tMuj+∆tCT pj = 0, ∀j ∈ {1, 2, . . . , Nt} ,
695
+ which is equal to the following block system:
696
+ (19)
697
+ α∆t
698
+
699
+ ����
700
+ M
701
+ M
702
+ ...
703
+ ...
704
+ M
705
+
706
+ ����
707
+
708
+ ��
709
+
710
+ M
711
+
712
+ ����
713
+ u1
714
+ u2
715
+ ...
716
+ uNt
717
+
718
+ ���� +∆t
719
+
720
+ ����
721
+ CT
722
+ 0
723
+ · · ·
724
+ CT
725
+ ...
726
+ CT
727
+
728
+ ����
729
+
730
+ ��
731
+
732
+ CT
733
+
734
+ ����
735
+ p1
736
+ p2
737
+ ...
738
+ pNt
739
+
740
+ ���� =
741
+
742
+ ����
743
+ 0
744
+ 0
745
+ ...
746
+ 0
747
+
748
+ ���� .
749
+ More compactly, we solve α∆tMu+∆tCT p = 0.
750
+ • State equation. A backward Euler method is used for a discretization forward in time. The
751
+ stabilized term related to ∂ty and the bilinear form a is [29, 37, 58]:
752
+ s
753
+
754
+ yN (t), qN ; µ
755
+
756
+ =
757
+
758
+ K∈Th
759
+ δK
760
+
761
+ ∂tyN (t) + TyN (t), hK
762
+ |η| TSSqN
763
+
764
+ K
765
+ ,
766
+ where yN (t) ∈ Y N for each t ∈ (0, T) and qN ∈ Y N . Defining the stabilized term ms as
767
+ (20)
768
+ ms
769
+
770
+ yN , qN ; µ
771
+
772
+ =
773
+
774
+ yN , qN �
775
+ L2(Ω) +
776
+
777
+ K∈Th
778
+ δK
779
+
780
+ yN , hK
781
+ |η| TSSqN
782
+
783
+ K
784
+ ,
785
+ then the backward Euler approach reads as: for each j ∈ {1, 2, · · · , Nt}, find yN
786
+ j ∈ Y N s.t.
787
+ (21)
788
+ 1
789
+ ∆tms
790
+
791
+ yN
792
+ j (µ) − yN
793
+ j−1(µ), qN ; µ
794
+
795
+ + as
796
+
797
+ yN
798
+ j (µ), qN ; µ
799
+
800
+ + cs
801
+
802
+ uN
803
+ j , qN ; µ
804
+
805
+ = Fs
806
+
807
+ qN ; µ
808
+
809
+ ,
810
+ given the initial condition yN
811
+ 0
812
+ which satisfies
813
+
814
+ yN
815
+ 0 , qN �
816
+ L2(Ω) =
817
+
818
+ y0, qN �
819
+ L2(Ω) , ∀qN ∈ Y N .
820
+ The matrix state equation to be solved becomes
821
+ (22)
822
+ Msyj + ∆tKsyj + ∆tCsuj = Msyj−1 + fsj∆t
823
+ for j ∈ {1, 2, . . . , Nt} ,
824
+ where the stabilized mass matrix of ms is Ms. Thus, we have
825
+
826
+ ����
827
+ Ms + ∆tKs
828
+ 0
829
+ −Ms
830
+ Ms + ∆tKs
831
+ 0
832
+ ...
833
+ ...
834
+ 0
835
+ 0
836
+ −Ms
837
+ Ms + ∆tKs
838
+
839
+ ����
840
+
841
+ ��
842
+
843
+ As
844
+ y+∆t
845
+
846
+ ��
847
+ Cs
848
+ 0
849
+ 0
850
+ ...
851
+ 0
852
+ 0
853
+ Cs
854
+
855
+ ��
856
+
857
+ ��
858
+
859
+ Cs
860
+ u = Msy0 + ∆tf s,
861
+ where Ms ∈ RN ·Nt×RN ·Nt is a block diagonal matrix which diagonal entries are [Ms, . . . , Ms].
862
+ In a more compact notation, we have Asy+∆tCsu = Msy0 + ∆tf s.
863
+ The final system considered and solved through an one shot approach is the following:
864
+ (23)
865
+
866
+
867
+ ∆tMT
868
+ s
869
+ 0
870
+ AT
871
+ s
872
+ 0
873
+ α∆tM
874
+ ∆tCT
875
+ As
876
+ ∆tCs
877
+ 0
878
+
879
+
880
+
881
+
882
+ y
883
+ u
884
+ p
885
+
886
+ � =
887
+
888
+
889
+ ∆tMT
890
+ s yd
891
+ 0
892
+ Msy0 + ∆tf s
893
+
894
+ � .
895
+
896
+ 8
897
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
898
+ 4. Weighted ROMs for random inputs advection-dominated OCP(µ)s
899
+ Numerical simulations for OCP(µ)s can be very expensive in relation to computational time and
900
+ storage. To overcome this problem, in this work we will consider ROMs [6, 24, 41, 40, 39]. We will
901
+ study the case when the parameter µ can be affected by randomness, i.e. it can follow a particular
902
+ probability distribution. That is the case of random inputs OCP(µ)s. In this scenario, a suitable
903
+ modification of the ROMs, the wROMs [11, 15, 13, 16, 17, 18, 49, 59, 61, 60], takes into account
904
+ the uncertainty quantification (UQ) of the problems and shows efficient results concerning errors
905
+ and computational time. For the sake of notation, from now on we denote a generic probability
906
+ distribution with the symbol ρ. ROM procedure is divided in two stages: an offline phase and an
907
+ online phase.
908
+ To exploit the potential of the ROMs setting, we assume an affine decomposition of the forms in
909
+ (15) [24]. Therefore, Assumption 4.0.1 is required here.
910
+ Assumption 4.0.1. We request that all the forms in (15) are affine in µ = (µ1, . . . , µp) ∈ P. More
911
+ precisely, we request that [15, 13]:
912
+ (1) the random diffusivity γ : Ω × P → R is of the form
913
+ (24)
914
+ γ(µ, x) = γ0(x) +
915
+ p
916
+
917
+ k=1
918
+ θγ
919
+ k(µk)γk(x),
920
+ with γk ∈ L∞(Ω), for k = 0, . . . , p and θγ
921
+ k depending only on µk;
922
+ (2) the random advection field η : Ω × P → R2 is of the form
923
+ (25)
924
+ η(µ, x) = η0(x) +
925
+ p
926
+
927
+ k=1
928
+ θη
929
+ k(µk)ηk(x),
930
+ with ηk ∈ (L∞(Ω))2, for k = 0, . . . , p and θη
931
+ k depending only on µk;
932
+ (3) the random forcing term f : Ω × P → R is of the form
933
+ (26)
934
+ f(µ, x) = f0(x) +
935
+ p
936
+
937
+ k=1
938
+ θf
939
+ k(µk)fk(x),
940
+ with fk ∈ L2(Ω), for k = 0, . . . , p and θf
941
+ k depending only on µk.
942
+ For example, Assumption 4.0.1 can be satisfied by truncating a Karhunen–Lo`eve expansion [47].
943
+ 4.1. Offline phase. The offline phase is the most expensive stage of the wROMs, which usually
944
+ depends on N. However, this should be done only once. The aim of this procedure is to build reduced
945
+ spaces Y N, U N and (QN)∗ that are good approximations of the high-fidelity ones and to compute
946
+ all block matrix components that are µ-independent. Then, everything is memorized in order to be
947
+ ready to be used in the online phase. The construction of the reduced basis is achieved through a
948
+ modified version of the POD algorithm: the wPOD [11, 60, 61], described in Section 4.1.1. Here,
949
+ we firstly compute high-fidelity evaluation of optimal solutions
950
+
951
+ yN (µ), uN (µ), pN (µ)
952
+
953
+ for different
954
+ parameters µ, the so-called snapshots, to build the bases. Because of good performance presented in
955
+ literature [30, 34, 53], this process will go through a partitioned approach, i.e. the wPOD is executed
956
+ separately for all three variables. After this step, the three reduced spaces for state, control and
957
+ adjoint are constructed as, respectively,
958
+ (27)
959
+ Y N = span {ξy
960
+ n, n = 1, . . . , N},
961
+ U N = span {ξu
962
+ n, n = 1, . . . , N} ,
963
+ (QN)∗ = span {ξp
964
+ n, n = 1, . . . , N} .
965
+ In order to ensure well-posedness for the reduced space approximation, we need to implement an
966
+ enriched space for state and adjoint variables. This means to impose GN ≡ Y N ≡ (QN)∗, where
967
+ GN = span {σn, n = 1, . . . , 2N} and {σn}2N
968
+ n=1 = {ξy
969
+ n}N
970
+ n=1 ∪ {ξp
971
+ n}N
972
+ n=1 [20, 23, 30, 31, 35, 34]. This
973
+ whole discussion holds true for parabolic problems in a space-time context, too. As a matter of fact,
974
+
975
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
976
+ 9
977
+ when dealing with time-dependent OCP(µ)s in a space-time approach, the time instances are not
978
+ separated in the wPOD algorithm. Therefore, each snapshot carries all the time instances.
979
+ 4.1.1. Weighted Proper Orthogonal Decomposition. The peculiarity of wPOD is to take into account
980
+ the probability distribution that characterizes µ to create reduced spaces with less number of basis
981
+ with respect to the deterministic case without losing in accuracy [11, 60, 61]. We will notice that
982
+ there will be different ways to consider randomness in the wPOD: the general idea is to suitably
983
+ attribute a larger weight to those samples that are more significant according to the distribution of µ.
984
+ From now we will refer to the POD algorithm based on the Monte-Carlo procedure in a deterministic
985
+ context, i.e. when the distribution ρ is the uniform one, as Standard POD to distinguish it from
986
+ the wPOD. As we will consider a partitioned approach, we show the procedure for the state space:
987
+ adjoint and control variables will follow the same process.
988
+ To consider stochasticity, wPOD needs to find the N-dimensional space Y N, with N ≪ N, such
989
+ that it minimizes the following estimate:
990
+ (28)
991
+ E =
992
+ ��
993
+ P
994
+ inf
995
+ ζy∈Y N ∥yN (µ) − ζy∥2
996
+ Y ρ(µ)dµ.
997
+ Let us consider a set of Ntrain ordered parameters µ1, . . . , µNtrain ∈ PNtrain, where PNtrain ∈ P is a
998
+ discretization of P called the training set and its cardinality is |PNtrain| = Ntrain. One can choose
999
+ Ntrain so that PNtrain is a good approximation of P. We can relate µ1, . . . , µNtrain to the ordered
1000
+ snapshots yN (µ1) , . . . , yN �
1001
+ µNtrain
1002
+
1003
+ . Considering w : P → R+ a weight function, a discretization
1004
+ of problem (28) is meant to find the N-dimensional space Y N which minimize the quantity
1005
+ (29)
1006
+ 1
1007
+ Ntrain
1008
+ Ntrain
1009
+
1010
+ k=1
1011
+ w (µk)
1012
+ ��yN (µk) − yN (µk)
1013
+ ��2
1014
+ Y .
1015
+ One could think that the natural choice can be w(µ) = ρ(µ) and in an UQ context this means
1016
+ to just discretize the expectation of the square error
1017
+ (30)
1018
+ E
1019
+ ���yN − yN��2
1020
+ Y
1021
+
1022
+ :=
1023
+
1024
+ P
1025
+ ��yN (µ) − yN(µ)
1026
+ ��2 ρ(µ)dµ,
1027
+ which is the argument of the square root in (28). However, this is not the unique choice in this
1028
+ scenario: therefore it will be interesting to understand which method is better to approximate (30).
1029
+ Here we illustrate different techniques that we use in the numerical tests in Section 5 to approximate
1030
+ (30). Considering the training set PNtrain =
1031
+
1032
+ µ1, . . . , µNtrain
1033
+
1034
+ , which can be composed by the nodes
1035
+ of the chosen quadrature formula that approximates (30), we indicate with ω = (ω1, . . . , ωNtrain) the
1036
+ standard weights of a chosen quadrature rule, with ρ1, . . . , ρNtrain the values of the density ρ in the
1037
+ nodes in PNtrain, and with w = (w1, . . . , wNtrain) the definitive weights used in wPOD algorithm. For
1038
+ a node µk, we have the correspondent quantities ωk, ρk, and wk. As a final result of this first step,
1039
+ the wPOD furnished the following sum to minimize
1040
+ (31)
1041
+ 1
1042
+ Ntrain
1043
+ Ntrain
1044
+
1045
+ k=1
1046
+ wk
1047
+ ��yN (µk) − yN (µk)
1048
+ ��2
1049
+ Y ,
1050
+ which is achieved here through the following algorithms:
1051
+ • Weighted Monte-Carlo method, where µ1, . . . , µNtrain are Ntrain parameters extracted from
1052
+ the random variable µ according to its distribution ρ and ρi are the values of the density ρ in
1053
+ these points. For this approximation, we have PNtrain =
1054
+
1055
+ µ1, . . . , µNtrain
1056
+
1057
+ and wk = ρ(µk),
1058
+ for all k = 1, . . . , Ntrain;
1059
+ • Pseudo-Random method based on a Halton Sequence, where µ1, . . . , µNtrain are the nodes
1060
+ extracted by a sampling completely based on the Halton sequence [57] and ρk = ρ(µk). Also
1061
+ in this case, PNtrain =
1062
+
1063
+ µ1, . . . , µNtrain
1064
+
1065
+ and wk = ρk, for all k = 1, . . . , Ntrain;
1066
+
1067
+ 10
1068
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1069
+ • Tensor product Gauss-Jacobi rule, where µ1, . . . , µNtrain are the nodes of the tensor product
1070
+ Gauss-Jacobi quadrature rule and ω1, . . . , ωNtrain are the correspondent quadrature weights.
1071
+ We can use this formula when the distribution is a Beta(αk, βk), as suitable Jacobi polyno-
1072
+ mials are orthogonal to this distribution [43]. As a matter of fact, simulations in Section 5
1073
+ will consider different Beta distributions for all components of µ. Therefore, we implement
1074
+ a Gauss-Jacobi formula using (αk, βk) as its parameters in each dimension [43], accordingly
1075
+ to the distribution of µ. For this approximation, we have PNtrain =
1076
+
1077
+ µ1, . . . , µNtrain
1078
+
1079
+ and
1080
+ wk = ωk, for all k = 1, . . . , Ntrain;
1081
+ • Tensor product Clenshaw-Curtis rule, where µ1, . . . , µNtrain are the nodes of the tensor prod-
1082
+ uct Clenshaw-Curtis quadrature rule and ω1, . . . , ωNtrain are the correspondent quadrature
1083
+ weights [57]. In this case we obtain PNtrain =
1084
+
1085
+ µ1, . . . , µNtrain
1086
+
1087
+ and wk = ρkωk, for all
1088
+ k = 1, . . . , Ntrain.
1089
+ In numerical tests of Section 5, we will respectively call as Weighted Monte-Carlo, Pseudo-
1090
+ Random, Gauss-Jacobi, and Clenshaw-Curtis wPOD algorithms the rules just specified. As it is
1091
+ know, tensor rule can be efficient, but their structure implies huge computational costs for elevate
1092
+ cardinality of the training set Ptrain or high-dimensional parameter space P. For this purpose, when
1093
+ we will use Clenshaw-Curtis or Gauss-Jacobi methods, we will consider sparse grid techniques based
1094
+ on a Smolyak algorithm, too [48, 62]: we will implement isotropic ones [36].
1095
+ Once chosen the rule (29) to approximate (30), the procedure to minimize (29) is described as
1096
+ follows. Let us define the deterministic correlation matrix of the snapshots of the state variable
1097
+ Dy ∈ RNtrain×Ntrain in the following way:
1098
+ (32)
1099
+ Dy
1100
+ kl :=
1101
+ 1
1102
+ Ntrain
1103
+
1104
+ yN (µk) , yN (µl)
1105
+
1106
+ Y ,
1107
+ 1 ≤ k, l ≤ Ntrain.
1108
+ Firstly, we define the weighted correlation matrix as
1109
+ (33)
1110
+ W y := W · Dy,
1111
+ where W := diag(w1, · · · , wNtrain) is the diagonal matrix whose elements are the weights of (29).
1112
+ The matrix W y is not symmetric in the usual matrix sense, but with respect to the scalar product
1113
+ induced by W y, hence W y is diagonalizable anyway [60]. Therefore, we seek the solution of the
1114
+ eigenvalue problem W ygy
1115
+ n = λy
1116
+ ngy
1117
+ n, 1 ≤ n ≤ Ntrain, where ∥gy
1118
+ n∥Y = 1, i.e. we pursue to find an
1119
+ eigenvalue λy
1120
+ n with the relative eigenvector of norm equal to one. We will indicate with (gy
1121
+ n)t the t-
1122
+ th component of the eigenvector gy
1123
+ n ∈ RNtrain. For the sake of simplicity, we rearrange the eigenvalues
1124
+ λy
1125
+ 1, . . . , λy
1126
+ Ntrain in a decreasing layout. Then, let us look at the first N eigenvalue-eigenvector pairs
1127
+ (λy
1128
+ 1, gy
1129
+ 1), . . . , (gy
1130
+ N, λy
1131
+ N). The basis functions χy
1132
+ n for the state equation are constructed through the
1133
+ following relation:
1134
+ (34)
1135
+ ζy
1136
+ n =
1137
+ 1
1138
+
1139
+ λy
1140
+ n
1141
+ Ntrain
1142
+
1143
+ t=1
1144
+ (gy
1145
+ n)t yN (µk) ,
1146
+ 1 ≤ n ≤ N.
1147
+ In order to choose N, one can refer to same study of eigenvalues of W y [24, 39, 61]. At the end, our
1148
+ reduced spaces are built as (27) and, then, enriched spaces are constructed.
1149
+ We summarise all the wPOD procedure for OCP(µ)s in Algorithm 1.
1150
+ 4.2. Online phase. In this stage, all operations have usually a N-independent cost. This process
1151
+ reflects to be computationally cheap and, therefore, it can be recalled multiple times using small
1152
+ machine resources. Firstly, we choose a parameter µ. We get all the µ-independent quantities and
1153
+ reduced spaces back from the storage. Immediately, we combined parameter independent part with
1154
+ the µ-dependent ones, that are rapidly calculated here. Then a Galerkin projector onto Y N, U N
1155
+ and (QN)∗ is performed, computing the reduced solution yN, uN and pN through a reduced block
1156
+ matrix system. As previously seen in Section 3, a stabilization is needed in the truth approximation.
1157
+ However, it could also not be the case for the online stage. This scenario lead to two possibilities:
1158
+ we do not use SUPG in the online phase, Offline-Only stabilization, or, on the contrary, stabilization
1159
+ occurs also here Offline-Online stabilization.
1160
+
1161
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1162
+ 11
1163
+ Algorithm 1 Weighted POD algorithm for OCP(µ) problems through a partitioned approach
1164
+ Input: FEM spaces Y N , U N , and (QN )∗ parameter domain P, and Ntrain.
1165
+ Output: reduced spaces Y N, U N and (QN)∗.
1166
+ Considering the high-fidelity spaces Y N , U N and (QN )∗:
1167
+ 1: Choose a quadrature rule (29) to approximate (30). This step defines a sample Ptrain ⊂ P and
1168
+ the respective weights w1, . . . , wNtrain. Define the matrix W := diag(w1, · · · , wNtrain) ;
1169
+ 2: for all µ ∈ Ptrain do
1170
+ 3:
1171
+ Solve the high-fidelity SUPG OCP(µ) system (15);
1172
+ 4: end for
1173
+ 5: Calculate the matrices Dy
1174
+ kl :=
1175
+ 1
1176
+ Ntrain
1177
+
1178
+ yN (µk) , yN (µl)
1179
+
1180
+ Y , 1 ≤ k, l ≤ Ntrain and W y := W · Dy.
1181
+ Do the same for the control u and the adjoint p;
1182
+ 6: Compute eigenvalues λy
1183
+ 1, . . . , λy
1184
+ Ntrain
1185
+ and the corresponding orthonormalized eigenvectors
1186
+ gy
1187
+ 1, . . . , gy
1188
+ Ntrain of W y. Do the same procedure for u and p variables;
1189
+ 7: Fix N according to a certain criterion and construct Y N = span {ξy
1190
+ n, n = 1, . . . , N}, where
1191
+ ξy
1192
+ n =
1193
+ 1
1194
+
1195
+ λy
1196
+ n
1197
+ �Ntrain
1198
+ t=1
1199
+ (gy
1200
+ n)t yN (µk). Do the same for u and p variables.
1201
+ 8: Build the aggregated space GN = span
1202
+
1203
+ {ξy
1204
+ n}N
1205
+ n=1 ∪ {ξp
1206
+ n}N
1207
+ n=1
1208
+
1209
+ and set GN ≡ Y N ≡ (QN)∗.
1210
+ 5. Numerical Results
1211
+ In this last part we illustrate numerical simulations concerning two Advection-Dominated OCP(µ)s
1212
+ under random inputs: the Graetz-Poiseuille Problem and the Propagating Front in a Square Prob-
1213
+ lem. In both experiments, the parameter µ will be a random vector and it will follow a prescribed
1214
+ probability density function that will be specified. The deterministic version of both experiments
1215
+ can be founded in [63].
1216
+ The Offline approximation will be always based on a P1−FEM, which means to consider a finite
1217
+ element method characterized by polynomials of degree less or equal than 1. In steady and unsteady
1218
+ simulations, the same stabilization parameter δK will be employed for both stabilization in the high-
1219
+ fidelity approximation and in the Online phase: namely in Offline-Online stabilization, δK is the
1220
+ same for both phases.
1221
+ For each simulation, relative errors between the FEM and the reduced solutions, i.e.
1222
+ (35)
1223
+ ey,N(µ) :=
1224
+ ��yN (µ) − yN(µ)
1225
+ ��
1226
+ Y
1227
+ ∥yN (µ)∥Y
1228
+ , eu,N(µ) :=
1229
+ ��uN (µ) − uN(µ)
1230
+ ��
1231
+ U
1232
+ ∥uN (µ)∥U
1233
+ , ep,N(µ) :=
1234
+ ��pN (µ) − pN(µ)
1235
+ ��
1236
+ Q∗
1237
+ ∥pN (µ)∥Q∗
1238
+ ,
1239
+ for the state, the control and the adjoint, respectively, will be shown. Due to the parametric nature
1240
+ of the problems, for each quantity in (35) a simple average is computed for µ distributed according
1241
+ to its probability density in a testing set Ptest ⊆ P of size Ntest, for every dimension N = 1, . . . , Nmax
1242
+ of the reduced space built through a chosen wPOD procedure. In every graph, the base-10 logarithm
1243
+ of these averages will be shown. When we will specify to use a POD procedure based on a Monte-
1244
+ Carlo sampling [57] of a uniform density distribution, we will talk about Standard POD. In order
1245
+ to compare the different wPOD possibilities, we use the same testing set for all of them: it will be
1246
+ taken using a Monte-Carlo method according to the distribution of µ. Obviously, the performance
1247
+ of the Standard POD will be based on a testing set of uniform density. The sum of the errors with
1248
+ respect to each discretized instant of time t will be taken into account in the unsteady versions.
1249
+ In order to compare the computational cost between the FEM solution with that of the reduced
1250
+ one for any possible dimension N, we use the speedup-index, i.e.
1251
+ (36)
1252
+ speedup-index = computational time of the high-fidelity solution
1253
+ computational time of the reduced solution
1254
+ ,
1255
+ which will be calculated for any µ in the testing set. Again, we will shown its sample average
1256
+ for any dimension N. For each test case, we will use the same Ptest to compute relative errors and
1257
+
1258
+ 12
1259
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1260
+ the speedup-index. The steady experiments are run using a machine with 16GB of RAM and Intel
1261
+ Core i7-7500U Dual Core, 2.7GHz for the CPU; whereas all parabolic simulations are computed
1262
+ considering 16GB of RAM and Intel Core i7 − 7700 Quad Core, 3.60GHz for the CPU.
1263
+ The code concerning steady experiments is implemented using the RBniCS library [2]; instead,
1264
+ the unsteady ones are provided using both RBniCS and multiphenics [1] libraries. These are python-
1265
+ based libraries, built on FEniCS [32].
1266
+ 5.1. Numerical Tests for the Graetz-Poiseuille Problem. The Graetz-Poiseuille problem is
1267
+ an Advection-Diffusion problem that represents the heat conduction in a rectilinear pipe. Here the
1268
+ transfer of heat can be regulated through the walls of the duct, which can be held at certain fixed
1269
+ temperature [22, 37, 46, 59].
1270
+ Firstly, we present simulation concerning the stationary case, where a distributed control is em-
1271
+ ployed all over the whole domain. The parameter µ = (µ1, µ2) is composed by the diffusion compo-
1272
+ nent µ1 and the geometrical one µ2, which characterizes the length of the plate.
1273
+ Ωobs
1274
+ Ωobs
1275
+ Ωo
1276
+ Γo,1
1277
+ Γo,2
1278
+ Γo,3
1279
+ Γo,4
1280
+ Γo,5
1281
+ Γo,6
1282
+ (0,0)
1283
+ (1,0)
1284
+ (1+µ2,0)
1285
+ (1+µ2,0.2)
1286
+ (1+µ2,0.8)
1287
+ (1+µ2,1)
1288
+ (1,1)
1289
+ (0,1)
1290
+ Figure 1. Geometry of the Graetz-Poiseuille Problem.
1291
+ The problem is studied using (x0, x1) as spatial coordinates. Ωo is the domain observed for a
1292
+ certain value µ2 with boundary Γo.
1293
+ We deal with homogeneous Neumann boundary conditions
1294
+ (BC) on Γo,3 := {1 + µ2} × [0, 1] considering Figure 1. Instead, Dirichlet conditions are set on sides
1295
+ Γo,1 := [0, 1]×{0}, Γo,5 := [0, 1]×{1}, Γo,6 := {0}×[0, 1] by imposing y = 0 and Γo,2 := [1, 1+µ2]×{0}
1296
+ and Γo,4 := [1, 1 + µ2] × {1} by imposing y = 1.
1297
+ The formulation of the problem is the following: given µ ∈ P, find (y, u) ∈ ˜Y × U which solves
1298
+ min
1299
+ (y,u)
1300
+ 1
1301
+ 2
1302
+
1303
+ Ωobs(µ)
1304
+ (y(µ) − yd)2 dΩo(µ) + α
1305
+ 2
1306
+
1307
+ Ωo(µ)
1308
+ u(µ)2 dΩo(µ),
1309
+ such that
1310
+ (37)
1311
+
1312
+
1313
+
1314
+
1315
+
1316
+
1317
+
1318
+
1319
+
1320
+
1321
+
1322
+
1323
+
1324
+
1325
+
1326
+ − 1
1327
+ µ1
1328
+ ∆y(µ) + 4x1(1 − x1)∂x0y(µ) = u(µ),
1329
+ in Ωo(µ),
1330
+ y(µ) = 0,
1331
+ on Γo,1(µ) ∪ Γo,5(µ) ∪ Γo,6(µ),
1332
+ y(µ) = 1,
1333
+ on Γo,2(µ) ∪ Γo,4(µ),
1334
+ ∂y(µ)
1335
+ ∂ν
1336
+ = 0,
1337
+ on Γo,3(µ),
1338
+ where ˜Y :=
1339
+
1340
+ v ∈ H1�
1341
+ Ωo
1342
+
1343
+ s.t. it satisfies the BC in (37)
1344
+
1345
+ and U = L2(Ωo). For the sake of clarity,
1346
+ a lifting function Ry ∈ H1(Ω) that fulfills the BC in (37) is used.
1347
+ Consequently, the variable
1348
+ ¯y := y − Ry, with ¯y ∈ Y , is used, where
1349
+ Y :=
1350
+
1351
+ v ∈ H1
1352
+ 0
1353
+
1354
+
1355
+
1356
+ s.t. ∂¯y
1357
+ ∂ν = 0, on Γ3 and ¯y = 0 on Γ \ Γ3
1358
+
1359
+ .
1360
+ Furthermore, we settle Q := Y ∗ without any loss of generality. Therefore, the adjoint variable p
1361
+ is null on Γ \ Γ3. The observation domain is Ωobs := [1, 1 + µ2] × [0.8, 1] ∪ [1, 1 + µ2] × [0, 0.2] as
1362
+ illustrated in Figure 1. The value µ2 can change the domain under study. Having that the domain
1363
+
1364
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1365
+ 13
1366
+ Ωo is µ-dependent itself, in the Offline Phase snapshots are based on different domains due to the
1367
+ sampling of the geometrical parameter components [24, 41, 45, 44]. To deal with the geometrical
1368
+ parametrization of the problem, we set a reference domain Ω and we build affine maps that transform
1369
+ Ω in Ωo for a defined µ. This procedure implies an automatic modification of some bilinear and
1370
+ linear forms involved in the weak formulation of Problem (37).
1371
+ We choose Ω = (0, 2) × (0, 1) as reference domain, that is the original one Ωo(µ) corresponding
1372
+ to µ2 = 1. We assume that µ2 is positive for the sake of simplicity. Considering Figure 1, we divide
1373
+ this into 2 subdomains, which are defined as Ω1 = (0, 1) × (0, 1) and Ω2 = (1, 2) × (0, 1). Then, we
1374
+ build two affine transformations:
1375
+ (38)
1376
+ V1(µ) : Ω1 → Ωo,1(µ) ⊂ R2,
1377
+ such that V1
1378
+ ��� x
1379
+ y
1380
+
1381
+ ; µ
1382
+
1383
+ :=
1384
+ � x
1385
+ y
1386
+
1387
+ ,
1388
+ which is the identity map defined on the first subdomain Ω1 and V2(µ) : Ω2 → Ωo,2(µ) ⊂ R2 as
1389
+ (39)
1390
+ V2
1391
+ �� x
1392
+ y
1393
+
1394
+ ; µ
1395
+
1396
+ =
1397
+ � µ2x
1398
+ y
1399
+
1400
+ +
1401
+ � 1 − µ2
1402
+ 0
1403
+
1404
+ = R2
1405
+ � x
1406
+ y
1407
+
1408
+ +
1409
+ � 1 − µ2
1410
+ 0
1411
+
1412
+ ,
1413
+ where we have
1414
+ (40)
1415
+ R2 :=
1416
+
1417
+ µ2
1418
+ 0
1419
+ 0
1420
+ 1
1421
+
1422
+ .
1423
+ Glueing together V1 and V2 for each µ ∈ P, we manage to build a one-to-one transformation
1424
+ V (µ) defined all over Ω. We denote the restrictions of Th to Ω1 and Ω2 with T 1
1425
+ h and T 2
1426
+ h , respec-
1427
+ tively. Therefore, we can express all the forms of the weak formulation under the effect of this
1428
+ transformation. For instance, after possible lifting, we have as = a + s and a∗
1429
+ s = a∗ + s∗ as
1430
+ (41)
1431
+ a
1432
+
1433
+ yN , qN ; µ
1434
+
1435
+ : =
1436
+
1437
+ Ω1
1438
+ 1
1439
+ µ1
1440
+ ∇yN · ∇qN + 4x1(1 − x1)∂x0yN qN
1441
+ +
1442
+
1443
+ Ω2
1444
+ 1
1445
+ µ1µ2
1446
+ ∂x0yN ∂x0qN + µ2
1447
+ µ1
1448
+ ∂x1yN ∂x1qN + 4x1(1 − x1)∂x0yN qN ,
1449
+ s
1450
+
1451
+ yN , qN ; µ
1452
+
1453
+ : =
1454
+
1455
+ K∈T 1
1456
+ h
1457
+ δKhK
1458
+
1459
+ K
1460
+
1461
+ 4x1(1 − x1)∂x0yN �
1462
+ ∂x0qN
1463
+ +
1464
+
1465
+ K∈T 2
1466
+ h
1467
+ δK
1468
+ hK
1469
+ õ2
1470
+
1471
+ K
1472
+
1473
+ 4x1(1 − x1)∂x0yN �
1474
+ ∂x0qN ,
1475
+ a∗ �
1476
+ zN , pN ; µ
1477
+
1478
+ : =
1479
+
1480
+ Ω1
1481
+ 1
1482
+ µ1
1483
+ ∇pN · ∇zN − 4x1(1 − x1)∂x0pN zN
1484
+
1485
+
1486
+ Ω2
1487
+ 1
1488
+ µ1µ2
1489
+ ∂x0pN ∂x0zN − µ2
1490
+ µ1
1491
+ ∂x1pN ∂x1zN − 4x1(1 − x1)∂x0pN zN ,
1492
+ s∗ �
1493
+ zN , pN ; µ
1494
+
1495
+ : =
1496
+
1497
+ K∈T 1
1498
+ h
1499
+ δKhK
1500
+
1501
+ K
1502
+
1503
+ 4x1(1 − x1)∂x0pN �
1504
+ ∂x0zN
1505
+ +
1506
+
1507
+ K∈T 2
1508
+ h
1509
+ δK
1510
+ hK
1511
+ õ2
1512
+
1513
+ K
1514
+
1515
+ 4x1(1 − x1)∂x0pN �
1516
+ ∂x0zN ,
1517
+ for all yN , qN , zN , pN , ∈ Y N . In order to take into account the possible bad effect on stabilized
1518
+ forms due to a extension or shortening of our domain Ωo, we choose the stabilization parameter for
1519
+ K ∈ T 2
1520
+ h as δK
1521
+ hK
1522
+ õ2 , where õ2 =
1523
+
1524
+ | det(R2)| [35, 37, 58].
1525
+ For the FEM discretization, a quite coarse mesh of size h = 0.034 is used and the total dimension
1526
+ of the numerical problem is 13146. We take δK = 1.0 for all K ∈ Th. The parameter space is set
1527
+ as P :=
1528
+
1529
+ 1, 105�
1530
+ ×
1531
+
1532
+ 0.5, 1.5
1533
+
1534
+ , from which we want to extract a training set Ptrain with cardinality
1535
+ Ntrain = 100. For the n bilinear form, we consider a penalization α = 0.01. Our aim is to minimize
1536
+ the L2-error between the state and the desired solution profile yd(x) = 1.0, function defined on Ωobs
1537
+
1538
+ 14
1539
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1540
+ Figure 2. (top) FEM not stabilized and (bottom) FEM stabilized solution, y (right) and
1541
+ u (left), µ = (105, 1.5), h = 0.034, α = 0.01, δK = 1.0.
1542
+ of Figure 1. Each wPOD procedure is computed until a Nmax = 20 in a partitioned approach and
1543
+ then all algorithms are compared using a testing set Ptest of 100 elements in P.
1544
+ We suppose that µ follows a Beta(5, 3) distribution for both parameter µ1 and µ2, i.e.
1545
+ (42)
1546
+ µ1 ∼ 1 +
1547
+
1548
+ 105 − 1
1549
+
1550
+ X1, where X1 ∼ Beta(5, 3),
1551
+ µ2 ∼ 0.5 +
1552
+
1553
+ 1.5 − 0.5
1554
+
1555
+ X2, where X2 ∼ Beta(5, 3),
1556
+ where µ1 and µ2 are independent random variables. This implies that we consider more probable
1557
+ the parameters for which the Graetz-Poiseuille Problem has high values of the P´eclet number. In
1558
+ Figure 2, we highlight how the FEM solutions of the state and the control are for µ = (105, 1.5).
1559
+ The adjoint solution is not shown here because it is proportional to the control due to the gradient
1560
+ equation [19].
1561
+ From Figure 2, one can see that a stabilization is necessarily needed.
1562
+ We firstly exploit the
1563
+ Offline-Only stabilization procedure, which results regarding errors are shown in Figure 3. The
1564
+ performance is not good for any kind of wPOD. Moreover, the Standard POD does not perform
1565
+ good, either. Relative errors never drop under 10−2 for any variables, hence more stabilization is
1566
+ necessary in this case.
1567
+ In Figure 4 relative errors of the Offline-Online stabilization procedure are presented.
1568
+ Here
1569
+ the trend seems better than the Offline-Only one, because these quantities decay faster along the
1570
+ value of N. The wPOD Monte-Carlo is the best performer for all y, u, p variables, as a matter of
1571
+ fact, it reaches ey,16 = 2.13 · 10−7 for the state, for the adjoint ep,16 = 3.95 · 10−7 and the control
1572
+ eu,16 = 3.80·10−7. This procedure has a better performance of the Standard POD, which its accuracy
1573
+ is at least 100 times inferior of the wPOD Monte-Carlo after N > 11. Concerning other rules, it can
1574
+ be noticed that Smolyak grid techniques perform better than their tensor-rule counterparts, despite
1575
+ having a training set whose cardinality is similar, but less of 100: 93 and 91 for the Clenshaw-Curtis
1576
+ and Gauss-Jacobi sparse grids, respectively.
1577
+ In Figure 5 we visually compare the two possibilities of stabilization for the geometrical parametriza-
1578
+ tion of the Graetz-Pouiseuille problem for the wPOD Monte-Carlo.
1579
+ In Table 1 we compare the speedup-index for all wPOD algorithms. We see that computational
1580
+ values are all of the same order of magnitude. For the wPOD Monte-Carlo we calculate 87 reduced
1581
+ solutions in the time of a FEM one.
1582
+ Now we want to present the parabolic version of Problem (37). This unsteady problem has been
1583
+ studied without optimization in [37, 59] in a deterministic context and in [59] in a UQ one. Instead,
1584
+ the deterministic OCP(µ) Graetz Problem under boundary control without Advection-dominancy
1585
+ is studied in [54, 52] and the deterministicdistributed control configuration is analyzed in [63].
1586
+
1587
+ 1.2e+00
1588
+ 0.8
1589
+ 0.6
1590
+ 0.4
1591
+ 0.2
1592
+ -1.5e-021.0e+00
1593
+ 0.5
1594
+ 0
1595
+ -0.5
1596
+ -9.3e-011.2e+00
1597
+ 0.8
1598
+ 0.6
1599
+ 0.4
1600
+ 0.2
1601
+ -1.5e-021.0e+00
1602
+ -0.5
1603
+ - 0
1604
+ -0.5
1605
+ -9.3e-01STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1606
+ 15
1607
+ Figure 3. Relative Errors for the Graetz-Poiseuille Problem - Offline-Only Stabiliza-
1608
+ tion; State (top-left), Control (top-right), Adjoint (bottom); Standard POD (blue), wPOD
1609
+ Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi Smolyak grid (red),
1610
+ Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid (dark green), Pseudo-
1611
+ Random based on Halton numbers (pink).
1612
+ Speedup-index Graetz-Poiseuille Problem: Offline-Online Stabilization - µ1, µ2 ∼ Beta(5,3)
1613
+ N
1614
+ POD
1615
+ wPOD
1616
+ Gauss tensor
1617
+ Gauss Smolyak
1618
+ CC tensor
1619
+ CC Smolyak
1620
+ Ps. Random
1621
+ 4
1622
+ 113.0
1623
+ 108.9
1624
+ 110.1
1625
+ 110.1
1626
+ 106.5
1627
+ 109.4
1628
+ 112.0
1629
+ 8
1630
+ 108.4
1631
+ 105.1
1632
+ 104.9
1633
+ 106.1
1634
+ 102.1
1635
+ 105.9
1636
+ 107.4
1637
+ 12
1638
+ 103.3
1639
+ 100.2
1640
+ 99.9
1641
+ 99.8
1642
+ 99.1
1643
+ 96.9
1644
+ 101.7
1645
+ 16
1646
+ 97.2
1647
+ 92.5
1648
+ 95.1
1649
+ 94.5
1650
+ 92.6
1651
+ 94.2
1652
+ 96.9
1653
+ 20
1654
+ 90.5
1655
+ 87.3
1656
+ 87.0
1657
+ 88.0
1658
+ 85.8
1659
+ 86.3
1660
+ 89.7
1661
+ Table 1. Average Speedup-index of Offline-Online Stabilization for the Graetz-Poiseuille
1662
+ Problem under geometrical parametrization. From left to right: Standard POD, wPOD
1663
+ Monte-Carlo, Gauss-Jacobi tensor, Gauss-Jacobi Smolyak grid, Clenshaw-Curtis tensor,
1664
+ Clenshaw-Curtis Smolyak grid, Pseudo-Random based on Halton numbers.
1665
+ Recalling Figure 1, for a fixed T > 0 the unsteady Graetz-Poiseuille Problem is posed as follows:
1666
+ find (y, u) ∈ ˜Y × U which solves
1667
+ min
1668
+ (y,u)
1669
+ 1
1670
+ 2
1671
+
1672
+ Ωobs(µ)×(0,T )
1673
+ (y(µ) − yd)2 dΩ + α
1674
+ 2
1675
+
1676
+ Ω(µ)×(0,T )
1677
+ u(µ)2 dΩ,
1678
+ such that
1679
+
1680
+ FEM vs ROM averaged relative error - y (state)
1681
+ 101
1682
+ Log-Error
1683
+ 100
1684
+ Relative L
1685
+ StandardPOD
1686
+ WeightedPODMonte-Carlo
1687
+ Gaussjacobi-tensor
1688
+ Gausslacobi-Smolyak
1689
+ ClenshawCurtis tensor
1690
+ 10-1
1691
+ ClenshawCurtis+Smolyak
1692
+ PseudoRandom-Halton
1693
+ 2.5
1694
+ 5.0
1695
+ 7.5
1696
+ 10.0
1697
+ 12.5
1698
+ 15.0
1699
+ 17.5
1700
+ 20.0
1701
+ NFEM vs ROM averaged relative error - u (control)
1702
+ 101
1703
+ 100
1704
+ Standard POD
1705
+ Weighted POD Monte-Carlo
1706
+ Gaussjacobi-tensor
1707
+ 10-1
1708
+ Gaussjacobi- Smolyak
1709
+ ClenshawCurtis-tensor
1710
+ ClenshawCurtis Smolyak
1711
+ PseudoRandom Halton
1712
+ 2.5
1713
+ 5.0
1714
+ 7.5
1715
+ 10.0
1716
+ 12.5
1717
+ 15.0
1718
+ 17.5
1719
+ 20.0
1720
+ NFEM vs ROM averaged relative error - p (adjoint)
1721
+ 102
1722
+ Log-Error
1723
+ 101
1724
+ Relative
1725
+ StandardPOD
1726
+ WeightedPoDMonte-Carlo
1727
+ Gaussjacobi-tensor
1728
+ 100
1729
+ Gaussjacobi--Smolyak
1730
+ ClenshawCurtis-tensor
1731
+ ClenshawCurtis Smolyak
1732
+ PseudoRandom Halton
1733
+ 2.5
1734
+ 5.0
1735
+ 7.5
1736
+ 10.0
1737
+ 12.5
1738
+ 15.0
1739
+ 17.5
1740
+ 20.0
1741
+ N16
1742
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1743
+ Figure 4. Relative Errors for the Graetz-Poiseuille Problem - Offline-Online Stabiliza-
1744
+ tion; State (top-left), Control (top-right), Adjoint (bottom); Standard POD (blue), wPOD
1745
+ Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi Smolyak grid (red),
1746
+ Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid (dark green), Pseudo-
1747
+ Random based on Halton numbers (pink).
1748
+ Figure 5. (top) wPOD Monte-Carlo Offline-Only stabilized and (bottom) Offline-Online
1749
+ stabilized solution, y (right) and u (left), µ = (105, 1.5), h = 0.034, α = 0.01, Ntrain = 100,
1750
+ δK = 1.0, N = 20.
1751
+ (43)
1752
+
1753
+
1754
+
1755
+
1756
+
1757
+
1758
+
1759
+
1760
+
1761
+
1762
+
1763
+
1764
+
1765
+
1766
+
1767
+
1768
+
1769
+
1770
+
1771
+ ∂ty(µ) − 1
1772
+ µ1
1773
+ ∆y(µ) + 4x1(1 − x1)∂x0y(µ) = u(µ),
1774
+ in Ω(µ) × (0, T),
1775
+ y(µ) = 0,
1776
+ on Γ1 ∪ Γ5 ∪ Γ6 × (0, T),
1777
+ y(µ) = 1,
1778
+ on Γ2(µ) ∪ Γ4(µ) × (0, T),
1779
+ ∂y(µ)
1780
+ ∂ν
1781
+ = 0,
1782
+ on Γ3(µ) × (0, T),
1783
+ y(µ)(0) = y0(x),
1784
+ in Ω(µ).
1785
+
1786
+ FEM vs ROM averaged relative error - p (adjoint)
1787
+ 101.4
1788
+ 100
1789
+ 10-1
1790
+ 10-2
1791
+ 10-3
1792
+ 10
1793
+ Standard POD
1794
+ 10-5
1795
+ Weighted POD Monte-Carlo
1796
+ Gaussjacobi+tensor
1797
+ GaussjacobiSmolyak
1798
+ ClenshawCurtis-tensor
1799
+ 10-6
1800
+ ClenshawCurtis - Smolyak
1801
+ PseudoRandom - Halton
1802
+ 2.5
1803
+ 5.0
1804
+ 7.5
1805
+ 10.0
1806
+ 12.5
1807
+ 15.0
1808
+ 17.5
1809
+ 20.0
1810
+ N1.2e+00
1811
+ 1
1812
+ 0.8
1813
+ 0.6
1814
+ 0.4
1815
+ 0.2
1816
+ -1.5e-021.0e+00
1817
+ 0.5
1818
+
1819
+ 0
1820
+ -0.5
1821
+ -9.3e-011.2e+00
1822
+ 0.8
1823
+ 0.6
1824
+ 0.4
1825
+ 0.2
1826
+ -1.5e-021.0e+00
1827
+ 0.5
1828
+ 0
1829
+ -0.5
1830
+ -9.3e-01FEM vs ROM averaged relative error - y (state)
1831
+ 10-1
1832
+ 10-2
1833
+ 10-3
1834
+ 10
1835
+ 10-5
1836
+ Standard POD
1837
+ Weighted PODMonte-Carlo
1838
+ Gaussjacobi +tensor
1839
+ 10-6
1840
+ Gaussjacobi↓. Smolyak
1841
+ ClenshawCurtis -tensor
1842
+ ClenshawCurtis-Smolyak
1843
+ PseudoRandom-Halton
1844
+ 2.5
1845
+ 5.0
1846
+ 7.5
1847
+ 10.0
1848
+ 12.5
1849
+ 15.0
1850
+ 17.5
1851
+ 20.0
1852
+ NFEM vs ROM averaged relative error - u (control)
1853
+ 10-1
1854
+ 10-3
1855
+ 10
1856
+ StandardPOD
1857
+ 10-5
1858
+ Weighted PODMonte-Carlo
1859
+ Gaussjacobi+tensor
1860
+ Gaussjacobi + Smolyak
1861
+ 10-6
1862
+ ClenshawCurtis.-.tensor
1863
+ ClenshawCurtis - Smolyak
1864
+ PseudoRandom-Halton
1865
+ 2.5
1866
+ 5.0
1867
+ 7.5
1868
+ 10.0
1869
+ 12.5
1870
+ 15.0
1871
+ 17.5
1872
+ 20.0
1873
+ NSTABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1874
+ 17
1875
+ Figure 6. Relative Errors for the Parabolic Graetz-Poiseuille Problem - Offline-Only Sta-
1876
+ bilization; State (top-left), Control (top-right), Adjoint (bottom); Standard POD (blue),
1877
+ wPOD Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi Smolyak
1878
+ grid (red), Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid (dark green),
1879
+ Pseudo-Random based on Halton numbers (pink).
1880
+ As made in the steady version, we firstly consider a lifting procedure. Simulations are run following
1881
+ the space-time setting proposed in Section 3.2 for a prearranged number of time-steps Nt.
1882
+ The initial condition is y0(x) = 0 for all x ∈ Ω referring to Figure 1 and we set T = 3.0. The
1883
+ penalization parameter is α = 0.01 and we want the state solution to be similar in the L2-norm to
1884
+ a desired solution profile yd(x, t) = 1.0, function defined for all t ∈ (0, 3.0) and for all x in Ωobs in
1885
+ Figure 1. Choosing Nt = 30, the time step is ∆t = 0.1. For the spatial discretization a quite coarse
1886
+ mesh of h = 0.038 is implemented: consequently the total high-fidelity dimension is Ntot = 314820
1887
+ and a single FEM space is characterized by N = 3498 for a fixed instant t . Again, δK = 1.0 for all
1888
+ K ∈ Th. We take P :=
1889
+
1890
+ 1, 105�
1891
+ ×
1892
+
1893
+ 1, 3.0
1894
+
1895
+ and µ is determined by the probability distribution
1896
+ (44)
1897
+ µ1 ∼ 1 +
1898
+
1899
+ 105 − 1
1900
+
1901
+ X1, where X1 ∼ Beta(5, 3),
1902
+ µ2 ∼ 1.0 +
1903
+
1904
+ 3.0 − 1.0
1905
+
1906
+ X2, where X2 ∼ Beta(5, 3).
1907
+ We choose a training set Ptrain of cardinality Ntrain = 100 (with exception of sparse grids, which
1908
+ have similar cardinality) and we performed the wPOD algorithms with Nmax = 15.
1909
+ In Figure 6 we present relative errors related to Offline-Only stabilization. Also in the parabolic
1910
+ case this procedure does not perform well. Therefore an online stabilization is needed.
1911
+ As a matter of fact, one can see in Figure 7 that the trends for Offline-Online stabilization seems
1912
+ a lot better than the previous strategy. Besides the Clenshaw-Curtis quadrature rule, errors decrease
1913
+ along the dimension N. Again, the best performance is given by the wPOD Monte-Carlo, where the
1914
+ following values are reached for N = 14: ey,14 = 9.71·10−7,ep,14 = 9.21·10−7, and eu,14 = 2.64·10−7.
1915
+
1916
+ FEM vs ROM averaged relative error - y (state)
1917
+ Standard POD
1918
+ WeightedPODMonte-Carlo
1919
+ Gaussjacobi-tensor
1920
+ GaussJacobi-Smolyak
1921
+ ClenshawCurtis -tensor
1922
+ 101
1923
+ ClenshawCurtis--Smolyak
1924
+ PseudoRandom-Halton
1925
+ Relative Log-Error
1926
+ 100
1927
+ 10-1
1928
+ 2
1929
+ 4
1930
+ 6
1931
+ 8
1932
+ 10
1933
+ 12
1934
+ 14
1935
+ NFEM vs ROM averaged relative error - u (control)
1936
+ StandardPOD
1937
+ WeightedPODMonte-Carlo
1938
+ Gaussjacobi-tensor
1939
+ 101
1940
+ Gaussjacobi- Smolyak
1941
+ ClenshawCurtis-tensor
1942
+ ClenshawCurtis -Smolyak
1943
+ Relative Log-Error
1944
+ PseudoRandom - Halton
1945
+ 100
1946
+ 10-
1947
+ 10-2
1948
+ 2
1949
+ 4
1950
+ 6
1951
+ 8
1952
+ 10
1953
+ 12
1954
+ 14
1955
+ NFEM vs ROM averaged relative error - p (adjoint)
1956
+ Standard POD
1957
+ WeightedPODMonte-Carlo
1958
+ Gaussjacobi-tensor
1959
+ Gaussjacobi-Smolyak
1960
+ ClenshawCurtis -tensor
1961
+ ClenshawCurtis - Smolyak
1962
+ PseudoRandom-Halton
1963
+ Relative Log-Error
1964
+ 101
1965
+ 100
1966
+ 2
1967
+ 4
1968
+ 6
1969
+ 8
1970
+ 10
1971
+ 12
1972
+ 14
1973
+ N18
1974
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
1975
+ Figure 7. Relative Errors for the Parabolic Graetz-Poiseuille Problem - Offline-Online
1976
+ Stabilization; State (top-left), Control (top-right), Adjoint (bottom); Standard POD
1977
+ (blue), wPOD Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi
1978
+ Smolyak grid (red), Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid
1979
+ (dark green), Pseudo-Random based on Halton numbers (pink).
1980
+ Finally, in Table 2 we illustrate the performance of the speedup-index. All weighted algorithms
1981
+ performs similar: we compute an order of magnitude of 104 reduced solution in the time of a FEM
1982
+ one. This efficiency is given by the nature of the space-time procedure, where each snapshot carries
1983
+ all the time instances, and the reduction is very effective.
1984
+ Speedup-index Parabolic Graetz-Poiseuille Problem: Offline-Online Stab. - µ1, µ2 ∼ Beta(5,3)
1985
+ N
1986
+ POD
1987
+ wPOD
1988
+ Gauss tensor
1989
+ Gauss Smolyak
1990
+ CC tensor
1991
+ CC Smolyak
1992
+ Ps. Random
1993
+ 3
1994
+ 14299.7
1995
+ 14571.0
1996
+ 13970.7
1997
+ 14013.4
1998
+ 14524.6
1999
+ 14578.1
2000
+ 14106.2
2001
+ 6
2002
+ 14666.3
2003
+ 15393.5
2004
+ 14621.8
2005
+ 14952.8
2006
+ 15302.6
2007
+ 15117.8
2008
+ 14482.0
2009
+ 9
2010
+ 14245.6
2011
+ 14803.1
2012
+ 14125.6
2013
+ 14546.5
2014
+ 14756.7
2015
+ 14608.5
2016
+ 13986.9
2017
+ 12
2018
+ 13693.6
2019
+ 14206.2
2020
+ 13554.3
2021
+ 13935.7
2022
+ 14050.5
2023
+ 14075.4
2024
+ 13453.0
2025
+ 15
2026
+ 13090.9
2027
+ 13606.4
2028
+ 13055.8
2029
+ 13455.1
2030
+ 13548.6
2031
+ 13544.0
2032
+ 12875.2
2033
+ Table 2. Average Speedup-index of Offline-Online Stabilization for the Parabolic Graetz-
2034
+ Poiseuille Problem under geometrical parametrization. From left to right: Standard POD,
2035
+ wPOD Monte-Carlo, Gauss-Jacobi tensor, Gauss-Jacobi Smolyak grid, Clenshaw-Curtis
2036
+ tensor, Clenshaw-Curtis Smolyak grid, Pseudo-Random based on Halton numbers.
2037
+ 5.2. Numerical Tests for Propagating Front in a Square Problem. Here we analyze an
2038
+ Advection-Dominated PDE problem illustrated without control in a deterministic and in a stochastic
2039
+ context in [37, 59] and in [59], respectively. A distributed control is applied all over the domain Ω,
2040
+
2041
+ FEM vs ROM averaged relative error - y (state)
2042
+ Standard POD
2043
+ Weighted POD Monte-Carlo
2044
+ 101
2045
+ Gaussjacobi- tensor
2046
+ GaussJacobi-Smolyak
2047
+ ClenshawCurtis-tensor
2048
+ 100
2049
+ ClenshawCurtis-Smolyak
2050
+ PseudoRandom-Halton
2051
+ 10-
2052
+ 10-2
2053
+ 10-3
2054
+ 10
2055
+ 10-5
2056
+ 10-6
2057
+ 2
2058
+ 4
2059
+ 6
2060
+ 8
2061
+ 10
2062
+ 12
2063
+ 14
2064
+ NFEM vs ROM averaged relative error - u (control)
2065
+ 101
2066
+ 100
2067
+ 10
2068
+ 10-2
2069
+ 10-3
2070
+ 10-
2071
+ .4
2072
+ Standard POD
2073
+ 10-5
2074
+ Weighted POD Monte-Carlo
2075
+ Gaussjacobi-tensor
2076
+ Gaussjacobi-Smolyak
2077
+ 10-6
2078
+ ClenshawCurtis.-.tensor
2079
+ ClenshawCurtis - Smolyak
2080
+ PseudoRandom -Halton
2081
+ 2
2082
+ 4
2083
+ 6
2084
+ 8
2085
+ 10
2086
+ 12
2087
+ 14
2088
+ NFEM vs ROM averaged relative error - p (adioint)
2089
+ 101
2090
+ 10
2091
+ 10-3
2092
+ Standard POD
2093
+ WeightedPODMonte-Cairlo
2094
+ Gaussjacobi-tensor
2095
+ 10-5
2096
+ Gaussjacobi- Smolyak
2097
+ ClenshawCurtis - tensor
2098
+ ClenshawCurtis - Smolyak
2099
+ PseudoRandom-Halton
2100
+ 2
2101
+ 4
2102
+ 6
2103
+ 8
2104
+ 10
2105
+ 12
2106
+ 14
2107
+ NSTABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2108
+ 19
2109
+ which is the square (0, 1) × (0, 1), as shown under Cartesian coordinates (x0, x1) in Figure 8. The
2110
+ boundary is composed as follows: Γ1 := {0} × [0, 0.25], Γ2 := [0, 1] × {0}, Γ3 := {1} × [0, 1],
2111
+ Γ4 := [0, 1] × {1}, Γ5 := {0} × [0.25, 1]; instead Ωobs := [0.25, 1] × [0.75, 1].
2112
+ Γ1
2113
+ Γ2
2114
+ Γ3
2115
+ Γ4
2116
+ Γ5
2117
+
2118
+ Ωobs
2119
+ (0,0.25)
2120
+ (0,1)
2121
+ (1,0.75)
2122
+ (1,1)
2123
+ (0.25,1)
2124
+ (1,0)
2125
+ (0,0)
2126
+ Figure 8. Geometry of the Propagating Front in a Square Problem
2127
+ Given µ = (µ1, µ2), our aim is to solve the following OCP(µ) problem: find (y, u) ∈ ˜Y × U which
2128
+ solves
2129
+ min
2130
+ (y,u)
2131
+ 1
2132
+ 2
2133
+
2134
+ Ωobs
2135
+ (y(µ) − yd)2 dΩ + α
2136
+ 2
2137
+
2138
+
2139
+ u(µ)2 dΩ,
2140
+ such that
2141
+ (45)
2142
+
2143
+
2144
+
2145
+
2146
+
2147
+
2148
+
2149
+ − 1
2150
+ µ1
2151
+ ∆y(µ) + [cos µ2, sin µ2] · ∇y(µ) = u(µ),
2152
+ in Ω,
2153
+ y(µ) = 1,
2154
+ on Γ1 ∪ Γ2,
2155
+ y(µ) = 0,
2156
+ on Γ3 ∪ Γ4 ∪ Γ5.
2157
+ In this case, we have that the domain of definition of our state y is
2158
+ ˜Y :=
2159
+
2160
+ v ∈ H1�
2161
+
2162
+
2163
+ s.t. BC in (45)
2164
+
2165
+ .
2166
+ Again, we define a lifting function Ry ∈ H1�
2167
+
2168
+
2169
+ such that satisfies BC in (45), applying a lifting
2170
+ procedure before the Lagrangian approach. We define ¯y := y − Ry, with ¯y ∈ Y and Y := H1
2171
+ 0(Ω),
2172
+ U = L2(Ω) and Q := Y ∗, with p = 0 on ∂Ω.
2173
+ The mesh size h is equal to 0.025, which entails an overall dimension of the truth approximation
2174
+ of 12087.
2175
+ Consequently, we have N = 4029 for state, control and adjoint spaces.
2176
+ Concerning
2177
+ stabilization, δK = 1.0 for all K ∈ Th. The penalization parameter is α = 0.01 and we pursue the
2178
+ state solution to be similar in the L2-norm to yd(x) = 0.5, defined for all x in Ωobs of Figure 8. In
2179
+ our test cases, P :=
2180
+
2181
+ 1, 4 · 104�
2182
+ ×
2183
+
2184
+ 0.9, 1.5
2185
+
2186
+ and µ follow the subsequent probability distribution:
2187
+ (46)
2188
+ µ1 ∼ 1 +
2189
+
2190
+ 4 · 104 − 1
2191
+
2192
+ X1, where X1 ∼ Beta(10, 10),
2193
+ µ2 ∼ 0.9 +
2194
+
2195
+ 1.5 − 0.9
2196
+
2197
+ X2, where X2 ∼ Beta(10, 10),
2198
+ where µ1 and µ2 are independent random variables. The training set Ptrain and the testing set
2199
+ Ptest have both cardinality equal to ntrain = 100, with exception of sparse grid samplings, whose
2200
+ cardinality is similar to 100. We apply a wPOD procedure for a Nmax = 50 dimension. In Figure 9,
2201
+ we show the performance of relative errors for the Offline-Only stabilization procedure. As in the
2202
+ Graetz-Poiseuille Problem, these trends are not acceptable, as no quantity drops under 10−1 for all
2203
+ state, control and adjoint variables. Therefore, a stabilization applied in the Online Phase is needed,
2204
+ too.
2205
+ In Figure 10 relative errors for Offline-Online Stabilization procedure are shown. Again, wPOD
2206
+ Monte-Carlo presents the best behaviour: in this case it reaches ey,50 = 5.03 · 10−7 for the state, for
2207
+ the adjoint ep,50 = 1.07·10−6, and the control eu,50 = 4.21·10−6. Moreover, the wPOD Monte-Carlo
2208
+ has an accuracy of nearly a factor of 100 better than a Standard POD in a deterministic context for
2209
+ N > 20. Also here, Smolyak grids perform better than their tensor counterpart: for instance, we
2210
+ obtain in this case it reaches ey,50 = 2.77 · 10−6 for the state, for the adjoint ep,50 = 5.80 · 10−6, and
2211
+
2212
+ 20
2213
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2214
+ Figure 9. Relative Errors for the Propagating Front in a Problem - Offline-Only Stabiliza-
2215
+ tion; State (top-left), Control (top-right), Adjoint (bottom); Standard POD (blue), wPOD
2216
+ Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi Smolyak grid (red),
2217
+ Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid (dark green), Pseudo-
2218
+ Random based on Halton numbers (pink).
2219
+ the control eu,50 = 1.02 · 10−5 for Gauss-Jacobi. Concerning the training set, we have Ntrain = 89
2220
+ and Ntrain = 93 for the Gauss-Jacobi and the Clenshaw-Curtis ones, respectively. In Figure 11
2221
+ we see a comparison between the FEM solution for the state and the adjoint without stabilization
2222
+ and the Offline-Online Stabilized wPOD Monte-Carlo reduced solution for these variables with
2223
+ µ = (2 · 104, 1.2).
2224
+ The values of the speedup-index for the Offline-Online stabilization for each type of wPOD are
2225
+ reported in Table 3. For N = 50 the wPOD Monte-Carlo is the best choice again with a computation
2226
+ of 50 reduced solutions in the time of a FEM one. All the other possibilities perform a little bit lower
2227
+ for N = 50; however, all weighted algorithms have similar performances concerning the speedup-
2228
+ index: an order of magnitude of 102 for the first 50 reduced basis.
2229
+ Numerical tests of the parabolic version of the Propagating Front in a Square Problem are here
2230
+ illustrated. For a fix T > 0 and a given µ ∈ P we have to find the pair (y, u) ∈ ˜Y × U which solves
2231
+ min
2232
+ (y,u)
2233
+ 1
2234
+ 2
2235
+
2236
+ Ωobs×(0,T )
2237
+ (y(µ) − yd)2 dΩ + α
2238
+ 2
2239
+
2240
+ Ω×(0,T )
2241
+ u(µ)2 dΩ,
2242
+ such that
2243
+ (47)
2244
+
2245
+
2246
+
2247
+
2248
+
2249
+
2250
+
2251
+
2252
+
2253
+
2254
+
2255
+
2256
+
2257
+ ∂ty(µ) − 1
2258
+ µ1
2259
+ ∆y(µ) + [cos µ2, sin µ2] · ∇y(µ) = u(µ),
2260
+ in Ω × (0, T),
2261
+ y(µ) = 1,
2262
+ on Γ1 ∪ Γ2 × (0, T),
2263
+ y(µ) = 0,
2264
+ on Γ3 ∪ Γ4 ∪ Γ5 × (0, T),
2265
+ y(µ)(0) = y0(x),
2266
+ in Ω,
2267
+
2268
+ FEM vs ROM averaged relative error - y (state)
2269
+ Relative Log-Error
2270
+ StandardPOD
2271
+ 100
2272
+ WeightedPoDMonte-Carlo
2273
+ GaussJacobi-tensor
2274
+ Gaussjacobi - Smolyak
2275
+ ClenshawCurtis-tensor
2276
+ ClenshawCurtis-Smolyak
2277
+ PseudoRandom-Halton
2278
+ 10
2279
+ 20
2280
+ 30
2281
+ 40
2282
+ 50
2283
+ NFEM vs ROM averaged relative error - u (control)
2284
+ Relative Log-Error
2285
+ 100
2286
+ Standard POD
2287
+ Weighted PODMonte-Carlo
2288
+ Gaussjacobi-tensor
2289
+ Gausslacobi-Smolyak
2290
+ ClenshawCurtis-tensor
2291
+ ClenshawCurtis-Smolyak
2292
+ PseudoRandom -Halton
2293
+ 10
2294
+ 20
2295
+ 30
2296
+ 40
2297
+ 50
2298
+ NFEM vs ROM averaged relative error - p (adioint)
2299
+ Relative Log-Error
2300
+ 101
2301
+ StandardPOD
2302
+ Weighted POD Monte-Carlo
2303
+ Gausslacobi-tensor
2304
+ Gaussjacobi- Smolyak
2305
+ ClenshawCurtis-tensor
2306
+ ClenshawCurtis-Smolyak
2307
+ PseudoRandom-Halton
2308
+ 10
2309
+ 20
2310
+ 30
2311
+ 40
2312
+ 50STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2313
+ 21
2314
+ Figure 10. Relative Errors for the Propagating Front in a Problem - Offline-Online Sta-
2315
+ bilization; State (top-left), Control (top-right), Adjoint (bottom); Standard POD (blue),
2316
+ wPOD Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi Smolyak
2317
+ grid (red), Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid (dark green),
2318
+ Pseudo-Random based on Halton numbers (pink).
2319
+ Figure 11. FEM not stabilized and wPOD Monte-Carlo Offline-Online stabilized solu-
2320
+ tion for y (left) and for p (right), µ = (2 · 104, 1.2), h = 0.025 α = 0.01, Ntrain = 100,
2321
+ δK = 1.0, N = 50.
2322
+ where y0(x) = 0 for all x ∈ Ω in Figure 8. A final time T = 3.0 is set. Considering the time
2323
+ discretization, we chose a number of time steps equal to Nt = 30, then we have ∆t = 0.1. Instead,
2324
+ for the spatial approximation, the mesh size is set to h = 0.036, that implies an overall dimension
2325
+ of the space-time setting equal to Ntot = 174780.
2326
+ For a fixed instant t, a single FEM space is
2327
+ characterized by N = 1942.
2328
+ For the SUPG procedure, we impose δK = 1.0 for all K ∈ Th.
2329
+ Setting a penalization parameter α = 0.01, we try to achieve in a L2-mean a desired solution profile
2330
+ yd(x, t) = 0.5, defined for all t ∈ (0, 3) and x in Ωobs of Figure 8.
2331
+ P :=
2332
+
2333
+ 1, 4 · 104�
2334
+ ×
2335
+
2336
+ 0.9, 1.5
2337
+
2338
+ , as in the steady version. We suppose that µ follows the probability
2339
+ distribution (46). Our training set has cardinality Ntrain = 100, with exception for Gauss-Jacobi and
2340
+ Clenshaw-Curtis Smolyak grids with Ntrain = 89 and Ntrain = 93, respectively, which are the number
2341
+ of nodes nearest to 100 for this kind of procedure. In Figure 12 and 13, we show a representative
2342
+
2343
+ FEM vs ROM averaged relative error - y (state)
2344
+ 10-1
2345
+ 10-2
2346
+ 10-3
2347
+ 10
2348
+ 4
2349
+ StandardPOD
2350
+ 10-5
2351
+ WeightedPODMonte-Carlo
2352
+ GaussJacobi-tensor
2353
+ Gaussjaciobi-Smolyak
2354
+ ClenshawCurtis-tensor
2355
+ 10-6
2356
+ ClenshawCurtis-Smolyak
2357
+ PseudoRandom-Halton
2358
+ 10
2359
+ 20
2360
+ 30
2361
+ 40
2362
+ 50
2363
+ NFEM vs ROM averaged relative error - u (control)
2364
+ 100
2365
+ Standard.POD
2366
+ Weighted POD Monte-Carlo
2367
+ GaussJacobi-tensor
2368
+ GaussJacobi - Smolyak
2369
+ ClenshawCurtis-tensor
2370
+ 10-1
2371
+ ClenshawCurtis-Smolyak
2372
+ PseudoRandom -Halton
2373
+ 10-2
2374
+ 10-3
2375
+ 10-5
2376
+ 10
2377
+ 20
2378
+ 30
2379
+ 40
2380
+ 50
2381
+ NFEM ys ROM averaged relative error - p (adioint)
2382
+ 100
2383
+ 10-1
2384
+ 10-2
2385
+ 10-3
2386
+ StandardPOD
2387
+ WeightedPODMonte-Carlo
2388
+ 10-5
2389
+ Gaussjacobi--tensor
2390
+ Gaussjacobi - Smolyak
2391
+ ClenshawCurtis-tensor
2392
+ ClenshawCurtis - Smolyak
2393
+ 10-6
2394
+ PseudoRandom---Halton
2395
+ 10
2396
+ 20
2397
+ 30
2398
+ 40
2399
+ 50
2400
+ N1.2e+00
2401
+ 1.1
2402
+ 0.9
2403
+ 0.8
2404
+ 0.7
2405
+ 0.6
2406
+ 0.5
2407
+ 0.4
2408
+ 0.3
2409
+ 0.2
2410
+ 0.1
2411
+ -1.8e-029.7e-03
2412
+ 0.008
2413
+ 0.006
2414
+ 0.004
2415
+ 0.002
2416
+ 0
2417
+ -0.002
2418
+ -0.004
2419
+ -0.006
2420
+ -8.4e-0322
2421
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2422
+ Speedup-index Propagating front in a Square Problem: Offline-Online Stab. - µ1, µ2 ∼ Beta(10,10)
2423
+ N
2424
+ POD
2425
+ wPOD
2426
+ Gauss tensor
2427
+ Gauss Smolyak
2428
+ CC tensor
2429
+ CC Smolyak
2430
+ Ps. Random
2431
+ 10
2432
+ 151.3
2433
+ 179.2
2434
+ 175.0
2435
+ 178.7
2436
+ 181.5
2437
+ 176.4
2438
+ 173.9
2439
+ 20
2440
+ 123.3
2441
+ 140.4
2442
+ 139.8
2443
+ 141.0
2444
+ 140.9
2445
+ 140.5
2446
+ 143.6
2447
+ 30
2448
+ 88.5
2449
+ 103.3
2450
+ 102.6
2451
+ 102.8
2452
+ 100.6
2453
+ 102.6
2454
+ 104.3
2455
+ 40
2456
+ 61.6
2457
+ 73.7
2458
+ 73.2
2459
+ 69.9
2460
+ 68.6
2461
+ 70.4
2462
+ 70.2
2463
+ 50
2464
+ 43.4
2465
+ 50.2
2466
+ 49.0
2467
+ 47.6
2468
+ 46.8
2469
+ 49.2
2470
+ 48.2
2471
+ Table 3. Average Speedup-index of Offline-Online Stabilization for the Propagating
2472
+ Front in a Square Problem.
2473
+ From left to right: Standard POD, wPOD Monte-Carlo,
2474
+ Gauss-Jacobi tensor, Gauss-Jacobi Smolyak grid, Clenshaw-Curtis tensor, Clenshaw-
2475
+ Curtis Smolyak grid, Pseudo-Random based on Halton numbers.
2476
+ Figure 12. wPOD Monte-Carlo Offline-Online stabilized reduced solution of y, for t =
2477
+ 0.1, t = 1.5, t = 3.0, µ = (2 · 104, 1.2), h = 0.036, α = 0.01, Ntrain = 100, δK = 1.0,
2478
+ N = 30.
2479
+ Figure 13. wPOD Monte-Carlo Offline-Online stabilized reduced solution of p, for t =
2480
+ 0.1, t = 1.5, t = 3.0, µ = (2 · 104, 1.2), h = 0.036, α = 0.01, Ntrain = 100, δK = 1.0,
2481
+ N = 30.
2482
+ stabilized FEM solution for µ = (2 · 104, 1.2) for some instants of time of the state y and the adjoint
2483
+ p, respectively. We choose to perform all wPOD procedure with Nmax = 30.
2484
+ Let us move to the error analysis. In Figure 14 we illustrate the relative errors for the Offline-Only
2485
+ stabilization. The performance are not satisfactory here, too, where no quantity drops below the
2486
+ accuracy of 10−1 for all N.
2487
+ Instead, Offline-Online stabilization procedure performs well, as one can notice from Figure 15.
2488
+ Again, wPOD Monte-Carlo has the best behaviour, it reaches ey,30 = 1.12 · 10−7 for the state, for
2489
+ the adjoint ep,30 = 4.55 · 10−7 and the control eu,30 = 1.36 · 10−7. Also in this case, isotropic sparse
2490
+ grid techniques is a better choice than tensor rules, both for Gauss-Jacobi and Clenshaw-Curtis
2491
+ approximations.
2492
+ In Table 4 we compare the speedup-index for all the weighted algorithms: performance are similar
2493
+ for all N, for N = 30 we computed nearly 4000 Offline-Online stabilized reduced solutions in the
2494
+ time of a FEM one.
2495
+
2496
+ 1.2e+00
2497
+ 1.1
2498
+ 0.9
2499
+ 0.8
2500
+ 0.7
2501
+ 0.6
2502
+ 0.5
2503
+ 0.4
2504
+ 0.3
2505
+ 0.2
2506
+ 0.1
2507
+ -1.8e-021.2e+00
2508
+ 1.1
2509
+ 0.9
2510
+ 0.8
2511
+ 0.7
2512
+ 0.6
2513
+ 0.5
2514
+ 0.4
2515
+ 0.3
2516
+ 0.2
2517
+ 0.1
2518
+ -1.8e-021.2e+00
2519
+ 0.9
2520
+ 0.8
2521
+ 0.7
2522
+ 0.6
2523
+ 0.5
2524
+ 0.4
2525
+ 0.3
2526
+ 0.2
2527
+ 0.1
2528
+ -1.8e-029.7e-03
2529
+ 0.008
2530
+ 0.006
2531
+ 0.004
2532
+ 0.002
2533
+ 0
2534
+ -0.002
2535
+ -0.004
2536
+ -0.006
2537
+ -8.4e-039.7e-03
2538
+ 0.008
2539
+ 0.006
2540
+ 0.004
2541
+ 0.002
2542
+ 0
2543
+ -0.002
2544
+ -0.004
2545
+ -0.006
2546
+ -8.4e-039.7e-03
2547
+ 0.008
2548
+ 0.006
2549
+ 0.004
2550
+ 0.002
2551
+ 0
2552
+ -0.002
2553
+ -0.004
2554
+ -0.006
2555
+ -8.4e-03STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2556
+ 23
2557
+ Figure 14. Relative Errors for the Parabolic Propagating Front in a Problem - Offline-
2558
+ Only Stabilization; State (top-left), Control (top-right), Adjoint (bottom); Standard POD
2559
+ (blue), wPOD Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi
2560
+ Smolyak grid (red), Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid
2561
+ (dark green), Pseudo-Random based on Halton numbers (pink).
2562
+ Speedup-index Parabolic Propagating front in a Square Problem: Offline-Online Stabilization
2563
+ N
2564
+ POD
2565
+ wPOD
2566
+ Gauss tensor
2567
+ Gauss Smolyak
2568
+ CC tensor
2569
+ CC Smolyak
2570
+ Ps. Random
2571
+ 5
2572
+ 6601.9
2573
+ 6503.5
2574
+ 6702.6
2575
+ 6629.1
2576
+ 6566.6
2577
+ 6605.6
2578
+ 6575.8
2579
+ 10
2580
+ 6275.9
2581
+ 6208.0
2582
+ 6336.0
2583
+ 6277.9
2584
+ 6204.4
2585
+ 6293.4
2586
+ 6212.4
2587
+ 15
2588
+ 5814.3
2589
+ 5702.4
2590
+ 5838.7
2591
+ 5794.3
2592
+ 5699.6
2593
+ 5752.1
2594
+ 5723.7
2595
+ 20
2596
+ 5327.9
2597
+ 5190.4
2598
+ 5329.8
2599
+ 5270.3
2600
+ 5277.2
2601
+ 5235.9
2602
+ 5197.6
2603
+ 25
2604
+ 4465.2
2605
+ 4303.3
2606
+ 4562.6
2607
+ 4422.2
2608
+ 4541.3
2609
+ 4433.1
2610
+ 4479.9
2611
+ 30
2612
+ 4061.5
2613
+ 3959.5
2614
+ 4140.3
2615
+ 4026.3
2616
+ 4100.3
2617
+ 4035.6
2618
+ 4043.8
2619
+ Table 4. Average Speedup-index of Offline-Online Stabilization for the Parabolic Propa-
2620
+ gating Front in a Square Problem. From left to right: Standard POD, wPOD Monte-Carlo,
2621
+ Gauss-Jacobi tensor, Gauss-Jacobi Smolyak grid, Clenshaw-Curtis tensor, Clenshaw-
2622
+ Curtis Smolyak grid, Pseudo-Random based on Halton numbers. µ1, µ2 ∼ Beta(10,10)
2623
+ 6. Conclusions and Perspectives
2624
+ In this work, we illustrated some numerical tests concerning stabilized Parametrized Advection-
2625
+ Dominated OCPs with random parametric inputs in a ROM context. We deal with both steady and
2626
+ unsteady cases and we took advantage of the SUPG stabilization to overcome numerical issues due
2627
+ to high values of the P´eclet number. Two possibilities of stabilization were analyzed: when SUPG
2628
+ is only used occurs in the offline phase, Offline-Only stabilization, or when it is provided in both
2629
+ online and offline phases, Offline-Online stabilization.
2630
+
2631
+ FEM vs ROM averaged relative error - y (state)
2632
+ StandardPOD
2633
+ Weighted POD Monte-Carlo
2634
+ 100
2635
+ Gaussjacobi- tensor
2636
+ Gaussjacobi -Smolyak
2637
+ ClenshawCurtis-tensor
2638
+ ClenshawCurtis-Smolyak
2639
+ PseudoRandom-Halton
2640
+ 5
2641
+ 10
2642
+ 15
2643
+ 20
2644
+ 25
2645
+ 30
2646
+ NFEM vs ROM averaged relative error - u (control)
2647
+ Standard POD
2648
+ WeightedPODMonte-Carlo
2649
+ 2 × 100
2650
+ Gaussjacobi-tensor
2651
+ GaussJacobi-Smolyak
2652
+ ClenshawCurtis-tensor
2653
+ ClenshawCurtis -Smolyak
2654
+ PseudoRandom-Halton
2655
+ 100
2656
+ 6×10-1
2657
+ 5
2658
+ 10
2659
+ 15
2660
+ 20
2661
+ 25
2662
+ 30
2663
+ NFEM vs ROM averaged relative error - p (adjoint)
2664
+ Standard POD
2665
+ Weighted PODMonte-Carlo
2666
+ Gaussjacobi-tensor
2667
+ GaussJacobi-Smolyak
2668
+ ClenshawCurtis-tensor
2669
+ ClenshawCurtis - Smolyak
2670
+ PseudoRandom-Halton
2671
+ Log-Error
2672
+ 101
2673
+ Relative
2674
+ 100
2675
+ 5
2676
+ 10
2677
+ 15
2678
+ 20
2679
+ 25
2680
+ 30
2681
+ N24
2682
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2683
+ Figure 15. Relative Errors for the Parabolic Propagating Front in a Problem - Offline-
2684
+ Online Stabilization; State (top-left), Control (top-right), Adjoint (bottom); Standard
2685
+ POD (blue), wPOD Monte-Carlo (orange), Gauss-Jacobi tensor rule (green), Gauss-Jacobi
2686
+ Smolyak grid (red), Clenshaw-Curtis tensor rule (cyan), Clenshaw-Curtis Smolyak grid
2687
+ (dark green), Pseudo-Random based on Halton numbers (pink).
2688
+ In order to deal with the uncertainty quantification caused by random inputs, we consider wROM.
2689
+ More precisely, we built our reduced bases using a wPOD in a partitioned approach, using different
2690
+ quadrature rules. We implemented wPOD Monte-Carlo, Gaussian quadrature formulae based on
2691
+ Jacobi polynomials in a tensor rule, approximation related to Clenshaw-Curtis tensor rule, Smolyak
2692
+ isotropic sparse grid approximation of the last two methods, quasi Monte-Carlo method as a Pseudo-
2693
+ Random rule defined on Halton numbers.
2694
+ We analyzed relative errors between the reduced and the high fidelity solutions and the speedup-
2695
+ index concerning the Graetz-Poiseuille and Propagating Front in a Square Problems, always under a
2696
+ distributed control. For the state, control and adjoint spaces we implemented a P1-FEM approxima-
2697
+ tion in a optimize-then-discretize framework as the truth solution. Concerning parabolic problems,
2698
+ a space-time approach is followed applying SUPG in a suitable way. In order to established which
2699
+ wPOD performs better, we compare them through the same testing set sampled by a Monte-Carlo
2700
+ method according to the probability distribution of the parameter.
2701
+ Offline-Only stabilization technique performed very poorly in terms of errors, this happened for
2702
+ all wROMs considered. Instead, in all the steady and unsteady experiments, the wROM technique
2703
+ performed excellently in an Offline-Online stabilization framework. For parabolic problems, the
2704
+ speedup-index features large values thanks to the space-time formulation. More precisely, wPOD
2705
+ Monte-Carlo technique was always the best performer for relative errors, instead, concerning compu-
2706
+ tational efficiency all methods seem equivalent. In addition, the efficiency of the wPOD Monte-Carlo
2707
+
2708
+ FEM vs ROM averaged relative error - y (state)
2709
+ 10-1
2710
+ 10-2
2711
+ 10-3
2712
+ 10-5
2713
+ Standard POD
2714
+ WeightedPODMonte-Carlo
2715
+ 10-6
2716
+ GaussJacobi-tensor
2717
+ Gaussjacobi- Smolyak
2718
+ ClenshawCurtis -tensor
2719
+ ClenshawCurtis - Smolyak
2720
+ 10-7
2721
+ PseudoRandom-Halton
2722
+ 5
2723
+ 10
2724
+ 15
2725
+ 20
2726
+ 25
2727
+ 30
2728
+ NFEM vs ROM averaged relative error - u (control)
2729
+ 10-1
2730
+ 10-2
2731
+ 10-3
2732
+ 10-
2733
+ 4
2734
+ 10-5
2735
+ Standard POD
2736
+ WeightedPODMonte-Carlo
2737
+ Gaussjacobi-tensor
2738
+ 10-6
2739
+ GaussJacobi-Smolyak
2740
+ ClenshawCurtis-tensor
2741
+ ClenshawCurtis-Smolyak
2742
+ 10-7
2743
+ PseudoRandom-Halton
2744
+ 5
2745
+ 10
2746
+ 15
2747
+ 20
2748
+ 25
2749
+ 30
2750
+ NFEM vs ROM averaged relative error - p (adjoint)
2751
+ 100
2752
+ 10-
2753
+ 10
2754
+ 10-3
2755
+ 10
2756
+ Standard POD
2757
+ 10-5
2758
+ WeightedPODMonte-Carlo
2759
+ Gaussjacobi-tensor
2760
+ Gaussjacobi- Smolyak
2761
+ ClenshawCurtis -tensor
2762
+ 10-6
2763
+ ClenshawCurtis -Smolyak
2764
+ PseudoRandom-Halton
2765
+ 5
2766
+ 10
2767
+ 15
2768
+ 20
2769
+ 25
2770
+ 30
2771
+ NSTABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2772
+ 25
2773
+ is supported by the fact that after a small number of reduced basis it is nearly 100 times more accu-
2774
+ rate than a Standard POD in a deterministic context. Moreover, we notice that sparse grids perform
2775
+ better than relative tensor ones, although having a bit less number of quadrature nodes.
2776
+ Furthermore, in the Graetz-Poiseuille Problem we illustrate that under geometrical parametriza-
2777
+ tion affected by randomness, wROMs still have good performance, despite small fluctuations in the
2778
+ graph of relative errors.
2779
+ As a first perspective, it might be interesting to create a strongly-consistent stabilization technique
2780
+ that flattens all the fluctuations of geometrical parametrization in a UQ context. Moreover, we want
2781
+ to extend the study to boundary control. Finally, it might be interesting to study the performance
2782
+ of other stabilization techniques for the online phases, for instance, of the Online Vanishing Viscosity
2783
+ and the Online Rectification methods [4, 12, 33] combined with the SUPG technique in the offline
2784
+ phase or with the stabilization strategy used in [59].
2785
+ Acknowledgements
2786
+ We acknowledge the support by European Union Funding for Research and Innovation – Horizon
2787
+ 2020 Program – in the framework of European Research Council Executive Agency: Consolidator
2788
+ Grant H2020 ERC CoG 2015 AROMA-CFD project 681447 “Advanced Reduced Order Methods
2789
+ with Applications in Computational Fluid Dynamics”. We also acknowledge the PRIN 2017 “Nu-
2790
+ merical Analysis for Full and Reduced Order Methods for the efficient and accurate solution of
2791
+ complex systems governed by Partial Differential Equations” (NA-FROM-PDEs) and the INDAM-
2792
+ GNCS project “Tecniche Numeriche Avanzate per Applicazioni Industriali”. The computations in
2793
+ this work have been performed with RBniCS [2] library, developed at SISSA mathLab, which is
2794
+ an implementation in FEniCS [32] of several reduced order modelling techniques; we acknowledge
2795
+ developers and contributors to both libraries.
2796
+ References
2797
+ [1] multiphenics - easy prototyping of multiphysics problems in FEniCS. https://mathlab.sissa.it/multiphenics.
2798
+ [2] RBniCS – reduced order modelling in FEniCS. https://www.rbnicsproject.org/.
2799
+ [3] Tu˘gba Akman, B¨ulent Karas¨ozen, and Zahire Kanar-Seymen. Streamline Upwind/Petrov-Galerkin solution of
2800
+ optimal control problems governed by time-dependent diffusion-convection-reaction equations. TWMS Journal
2801
+ of Applied and Engineering Mathematics, 7(2):221–235, 2017.
2802
+ [4] Shafqat Ali. Stabilized reduced basis methods for the approximation of parametrized viscous flows. PhD. Thesis,
2803
+ SISSA, 2018.
2804
+ [5] Francesco Ballarin, Gianluigi Rozza, and Maria Strazzullo. Chapter 9 - Space-time POD-Galerkin approach for
2805
+ parametric flow control. In Emmanuel Tr´elat and Enrique Zuazua, editors, Numerical Control: Part A, volume 23
2806
+ of Handbook of Numerical Analysis, pages 307–338. Elsevier, 2022.
2807
+ [6] Peter Benner, Mario Ohlberger, Anthony Patera, Gianluigi Rozza, and Karsten Urban. Model reduction of
2808
+ parametrized systems. Springer, 2017.
2809
+ [7] Franco Brezzi. On the existence, uniqueness and approximation of saddle-point problems arising from Lagrangian
2810
+ multipliers. Publications math´ematiques et informatique de Rennes, (S4):1–26, 1974.
2811
+ [8] Franco Brezzi and Michel Fortin. Mixed and hybrid finite element methods, volume 15. Springer Science &
2812
+ Business Media, 2012.
2813
+ [9] Alexander N. Brooks and Thomas J.R. Hughes. Streamline Upwind/Petrov-Galerkin formulations for convection
2814
+ dominated flows with particular emphasis on the incompressible Navier-Stokes equations. Computer methods in
2815
+ applied mechanics and engineering, 32(1-3):199–259, 1982.
2816
+ [10] Giuseppe Carere. Reduced Order Methods for Optimal Control Problems constrained by PDEs with random
2817
+ inputs and applications. Master’s thesis, University of Amsterdam and SISSA, 2019.
2818
+ [11] Giuseppe Carere, Maria Strazzullo, Francesco Ballarin, Gianluigi Rozza, and Rob Stevenson. A weighted POD-
2819
+ reduction approach for parametrized PDE-constrained Optimal Control Problems with random inputs and ap-
2820
+ plications to environmental sciences. Computers & Mathematics with Applications, 102:261–276, 2021.
2821
+ [12] Rachida Chakir, Yvon Maday, and Philippe Parnaudeau. A non-intrusive reduced basis approach for parametrized
2822
+ heat transfer problems. Journal of Computational Physics, 376:617–633, 2019.
2823
+ [13] Peng Chen and Alfio Quarteroni. Weighted reduced basis method for stochastic optimal control problems with
2824
+ elliptic PDE constraint. SIAM/ASA Journal on Uncertainty Quantification, 2(1):364–396, 2014.
2825
+ [14] Peng Chen, Alfio Quarteroni, and Gianluigi Rozza. Stochastic optimal Robin boundary control problems of
2826
+ advection-dominated elliptic equations. SIAM Journal on Numerical Analysis, 51(5):2700–2722, 2013.
2827
+
2828
+ 26
2829
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2830
+ [15] Peng Chen, Alfio Quarteroni, and Gianluigi Rozza. A weighted reduced basis method for elliptic partial differential
2831
+ equations with random input data. SIAM Journal on Numerical Analysis, 51(6):3163–3185, 2013.
2832
+ [16] Peng Chen, Alfio Quarteroni, and Gianluigi Rozza. Comparison between reduced basis and stochastic collocation
2833
+ methods for elliptic problems. Journal of Scientific Computing, 59(1):187–216, 2014.
2834
+ [17] Peng Chen, Alfio Quarteroni, and Gianluigi Rozza. Multilevel and weighted reduced basis method for stochastic
2835
+ optimal control problems constrained by Stokes equations. Numerische Mathematik, 133(1):67–102, 2016.
2836
+ [18] Peng Chen, Alfio Quarteroni, and Gianluigi Rozza. Reduced basis methods for uncertainty quantification.
2837
+ SIAM/ASA Journal on Uncertainty Quantification, 5(1):813–869, 2017.
2838
+ [19] S. Scott Collis and Matthias Heinkenschloss. Analysis of the Streamline Upwind/Petrov Galerkin method applied
2839
+ to the solution of optimal control problems. CAAM TR02-01, 108, 2002.
2840
+ [20] Luca Ded`e. Reduced Basis Method and A Posteriori Error Estimation for Parametrized Linear-Quadratic Optimal
2841
+ Control Problems. SIAM Journal on Scientific Computing, 32(2):997–1019, 2010.
2842
+ [21] Kenneth Eriksson and Claes Johnson. Error estimates and automatic time step control for nonlinear parabolic
2843
+ problems, I. SIAM journal on numerical analysis, 24(1):12–23, 1987.
2844
+ [22] Fabrizio Gelsomino and Gianluigi Rozza. Comparison and combination of reduced-order modelling techniques
2845
+ in 3D parametrized heat transfer problems. Mathematical and Computer Modelling of Dynamical Systems,
2846
+ 17(4):371–394, 2011.
2847
+ [23] Anna-Lena Gerner and Karen Veroy. Certified Reduced Basis Methods for Parametrized Saddle Point Problems.
2848
+ SIAM Journal on Scientific Computing, 34(5):A2812–A2836, 2012.
2849
+ [24] Jan S. Hesthaven, Gianluigi Rozza, and Benjamin Stamm. Certified reduced basis methods for parametrized
2850
+ partial differential equations, volume 590. Springer, 2016.
2851
+ [25] Michael Hinze, Michael K¨oster, and Stefan Turek. A hierarchical space-time solver for distributed control of the
2852
+ Stokes equation. Technical Report,SPP1253-16-01, 2008.
2853
+ [26] L.S. Hou, J. Lee, and H. Manouzi. Finite element approximations of stochastic optimal control problems con-
2854
+ strained by stochastic elliptic PDEs. Journal of Mathematical Analysis and Applications, 384(1):87–103, 2011.
2855
+ [27] Thomas J.R. Hughes. A multidimentional upwind scheme with no crosswind diffusion. Finite Element Methods
2856
+ for Convection Dominated Flows, AMD 34, 1979.
2857
+ [28] Thomas J.R. Hughes. Recent progress in the development and understanding of SUPG methods with special
2858
+ reference to the compressible Euler and Navier-Stokes equations. International journal for numerical methods in
2859
+ fluids, 7(11):1261–1275, 1987.
2860
+ [29] Volker John and Julia Novo. Error analysis of the SUPG finite element discretization of evolutionary convection-
2861
+ diffusion-reaction equations. SIAM journal on numerical analysis, 49(3):1149–1176, 2011.
2862
+ [30] Mark K¨archer, Zoi Tokoutsi, Martin A. Grepl, and Karen Veroy. Certified reduced basis methods for parametrized
2863
+ elliptic optimal control problems with distributed controls. Journal of Scientific Computing, 75(1):276–307, 2018.
2864
+ [31] Karl Kunisch and Stefan Volkwein. Proper orthogonal decomposition for optimality systems. ESAIM: Mathe-
2865
+ matical Modelling and Numerical Analysis, 42(1):1–23, 2008.
2866
+ [32] Anders Logg, Kent-Andre Mardal, and Garth Wells. Automated solution of differential equations by the finite
2867
+ element method: The FEniCS book, volume 84. Springer Science & Business Media, 2012.
2868
+ [33] Yvon Maday and Eitan Tadmor. Analysis of the spectral vanishing viscosity method for periodic conservation
2869
+ laws. SIAM Journal on Numerical Analysis, 26(4):854–870, 1989.
2870
+ [34] Federico Negri, Andrea Manzoni, and Gianluigi Rozza. Reduced basis approximation of parametrized optimal
2871
+ flow control problems for the Stokes equations. Computers & Mathematics with Applications, 69(4):319–336,
2872
+ 2015.
2873
+ [35] Federico Negri, Gianluigi Rozza, Andrea Manzoni, and Alfio Quarteroni. Reduced basis method for parametrized
2874
+ elliptic optimal control problems. SIAM Journal on Scientific Computing, 35(5):A2316–A2340, 2013.
2875
+ [36] Fabio Nobile, Ra´ul Tempone, and Clayton G. Webster. A sparse grid stochastic collocation method for partial
2876
+ differential equations with random input data. SIAM Journal on Numerical Analysis, 46(5):2309–2345, 2008.
2877
+ [37] Paolo Pacciarini and Gianluigi Rozza. Stabilized reduced basis method for parametrized advection–diffusion
2878
+ PDEs. Computer Methods in Applied Mechanics and Engineering, 274:1–18, 2014.
2879
+ [38] Alfio Quarteroni. Numerical models for differential problems, volume 2. Springer, 2009.
2880
+ [39] Alfio Quarteroni, Andrea Manzoni, and Federico Negri. Reduced basis methods for partial differential equations:
2881
+ an introduction, volume 92. Springer, 2015.
2882
+ [40] Alfio Quarteroni, Gianluigi Rozza, et al. Reduced order methods for modeling and computational reduction,
2883
+ volume 9. Springer, 2014.
2884
+ [41] Alfio Quarteroni, Gianluigi Rozza, and Andrea Manzoni. Certified reduced basis approximation for parametrized
2885
+ partial differential equations and applications. Journal of Mathematics in Industry, 1(1):1–49, 2011.
2886
+ [42] Alfio Quarteroni and Alberto Valli. Numerical approximation of partial differential equations, volume 23. Springer
2887
+ Science & Business Media, 2008.
2888
+ [43] Anthony Ralston and Philip Rabinowitz. A first course in numerical analysis. Courier Corporation, 2001.
2889
+ [44] Gianluigi Rozza. Reduced basis approximation and error bounds for potential flows in parametrized geometries.
2890
+ Communications in Computational Physics, 9(1):1–48, 2011.
2891
+
2892
+ STABILIZED WEIGHTED ROMS FOR ADVECTION-DOMINATED OCPS WITH RANDOM INPUTS
2893
+ 27
2894
+ [45] Gianluigi Rozza, Dinh Bao Phuong Huynh, and Anthony T. Patera. Reduced basis approximation and a posteriori
2895
+ error estimation for affinely parametrized elliptic coercive partial differential equations. Archives of Computational
2896
+ Methods in Engineering, 15(3):229–275, 2008.
2897
+ [46] Gianluigi Rozza, Ngoc-Cuong Nguyen, Anthony T. Patera, and Simone Deparis. Reduced basis methods and
2898
+ a posteriori error estimators for heat transfer problems. In Heat Transfer Summer Conference, volume 43574,
2899
+ pages 753–762, 2009.
2900
+ [47] Christoph Schwab and Radu Alexandru Todor. Karhunen-Lo`eve approximation of random fields by generalized
2901
+ fast multipole methods. Journal of Computational Physics, 217(1):100–122, 2006.
2902
+ [48] Sergei Abramovich Smolyak. Quadrature and interpolation formulas for tensor products of certain classes of
2903
+ functions. In Doklady Akademii Nauk, volume 148, pages 1042–1045. Russian Academy of Sciences, 1963.
2904
+ [49] Christopher Spannring, Sebastian Ullmann, and Jens Lang. A weighted reduced basis method for parabolic PDEs
2905
+ with random data. In International Conference on Computational Engineering, pages 145–161. Springer, 2017.
2906
+ [50] Martin Stoll and Andrew J. Wathen. All-at-once solution of time-dependent PDE-constrained optimization prob-
2907
+ lems. Unspecified, Tech. Rep, 2010.
2908
+ [51] Martin Stoll and Andy Wathen. All-at-once solution of time-dependent Stokes control. Journal of Computational
2909
+ Physics, 232(1):498–515, 2013.
2910
+ [52] Maria Strazzullo. Model Order Reduction for Nonlinear and Time-Dependent Parametric Optimal Flow Control
2911
+ Problems. PhD. Thesis, SISSA, 2021.
2912
+ [53] Maria Strazzullo, Francesco Ballarin, Renzo Mosetti, and Gianluigi Rozza. Model reduction for parametrized op-
2913
+ timal control problems in environmental marine sciences and engineering. SIAM Journal on Scientific Computing,
2914
+ 40(4):B1055–B1079, 2018.
2915
+ [54] Maria Strazzullo, Francesco Ballarin, and Gianluigi Rozza. POD-Galerkin model order reduction for parametrized
2916
+ time dependent linear quadratic optimal control problems in saddle point formulation. Journal of Scientific
2917
+ Computing, 83:1–35, 2020.
2918
+ [55] Maria Strazzullo, Francesco Ballarin, and Gianluigi Rozza. A certified reduced basis method for linear
2919
+ parametrized parabolic optimal control problems in space-time formulation. arXiv preprint arXiv:2103.00460,
2920
+ 2021.
2921
+ [56] Maria Strazzullo, Francesco Ballarin, and Gianluigi Rozza. POD-Galerkin model order reduction for parametrized
2922
+ nonlinear time-dependent optimal flow control: an application to shallow water equations. Journal of Numerical
2923
+ Mathematics, 30(1):63–84, 2022.
2924
+ [57] Timothy John Sullivan. Introduction to Uncertainty Quantification, volume 63. Springer, 2015.
2925
+ [58] Davide Torlo. Stabilized reduced basis method for transport PDEs with random inputs. Master’s thesis, University
2926
+ of Trieste and SISSA, 2016.
2927
+ [59] Davide Torlo,
2928
+ Francesco Ballarin,
2929
+ and Gianluigi Rozza. Stabilized weighted reduced basis methods for
2930
+ parametrized advection dominated problems with random inputs. SIAM/ASA Journal on Uncertainty Quantifi-
2931
+ cation, 6(4):1475–1502, 2018.
2932
+ [60] Luca Venturi, Francesco Ballarin, and Gianluigi Rozza. A weighted POD method for elliptic PDEs with random
2933
+ inputs. Journal of Scientific Computing, 81(1):136–153, 2019.
2934
+ [61] Luca Venturi, Davide Torlo, Francesco Ballarin, and Gianluigi Rozza. Weighted reduced order methods for
2935
+ parametrized partial differential equations with random inputs. In Uncertainty Modeling for Engineering Appli-
2936
+ cations, pages 27–40. Springer, 2019.
2937
+ [62] Dongbin Xiu and Jan S. Hesthaven. High-order collocation methods for differential equations with random inputs.
2938
+ SIAM Journal on Scientific Computing, 27(3):1118–1139, 2005.
2939
+ [63] Fabio Zoccolan, Maria Strazzullo, and Gianluigi Rozza. A Streamline Upwind Petrov-Galerkin Reduced Order
2940
+ Method for Advection-Dominated Partial Differential Equations under Optimal Control. preprint, 2022.
2941
+