jackkuo commited on
Commit
e61704c
·
verified ·
1 Parent(s): 4478090

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. -9A0T4oBgHgl3EQfPP_v/vector_store/index.faiss +3 -0
  2. -NE5T4oBgHgl3EQfRg5P/content/2301.05521v1.pdf +3 -0
  3. .gitattributes +34 -0
  4. 0NAyT4oBgHgl3EQf0_na/content/tmp_files/2301.00729v1.pdf.txt +1646 -0
  5. 0NAyT4oBgHgl3EQf0_na/content/tmp_files/load_file.txt +0 -0
  6. 19E3T4oBgHgl3EQfngq-/content/tmp_files/2301.04626v1.pdf.txt +1251 -0
  7. 19E3T4oBgHgl3EQfngq-/content/tmp_files/load_file.txt +0 -0
  8. 1dAzT4oBgHgl3EQfDfo-/vector_store/index.faiss +3 -0
  9. 29FST4oBgHgl3EQfYTgs/vector_store/index.faiss +3 -0
  10. 2NE4T4oBgHgl3EQfagwY/content/2301.05064v1.pdf +3 -0
  11. 2NE4T4oBgHgl3EQfagwY/vector_store/index.faiss +3 -0
  12. 2NE4T4oBgHgl3EQfagwY/vector_store/index.pkl +3 -0
  13. 4tE4T4oBgHgl3EQf1A0X/vector_store/index.faiss +3 -0
  14. 5NAyT4oBgHgl3EQfpPjF/content/tmp_files/2301.00523v1.pdf.txt +1189 -0
  15. 5NAyT4oBgHgl3EQfpPjF/content/tmp_files/load_file.txt +0 -0
  16. 6tE4T4oBgHgl3EQfCAsz/content/tmp_files/2301.04856v1.pdf.txt +0 -0
  17. 6tE4T4oBgHgl3EQfCAsz/content/tmp_files/load_file.txt +0 -0
  18. 79A0T4oBgHgl3EQfOf-b/vector_store/index.faiss +3 -0
  19. 7tE1T4oBgHgl3EQfBwI8/vector_store/index.pkl +3 -0
  20. 7tE2T4oBgHgl3EQfPgY3/content/tmp_files/2301.03759v1.pdf.txt +1407 -0
  21. 7tE2T4oBgHgl3EQfPgY3/content/tmp_files/load_file.txt +0 -0
  22. 8dAzT4oBgHgl3EQfgfzo/vector_store/index.pkl +3 -0
  23. 8tFRT4oBgHgl3EQfpzcC/content/2301.13614v1.pdf +3 -0
  24. 8tFRT4oBgHgl3EQfpzcC/vector_store/index.faiss +3 -0
  25. 9dA0T4oBgHgl3EQfO_9E/content/tmp_files/2301.02168v1.pdf.txt +0 -0
  26. 9dA0T4oBgHgl3EQfO_9E/content/tmp_files/load_file.txt +0 -0
  27. AdFQT4oBgHgl3EQfMTYr/content/tmp_files/2301.13267v1.pdf.txt +1613 -0
  28. AdFQT4oBgHgl3EQfMTYr/content/tmp_files/load_file.txt +0 -0
  29. CdFAT4oBgHgl3EQftB4M/vector_store/index.pkl +3 -0
  30. CtE1T4oBgHgl3EQf9wbT/vector_store/index.pkl +3 -0
  31. GNE1T4oBgHgl3EQf_AZD/content/tmp_files/2301.03575v1.pdf.txt +3147 -0
  32. GNE1T4oBgHgl3EQf_AZD/content/tmp_files/load_file.txt +0 -0
  33. HdAzT4oBgHgl3EQfUvz7/content/tmp_files/2301.01274v1.pdf.txt +801 -0
  34. HdAzT4oBgHgl3EQfUvz7/content/tmp_files/load_file.txt +387 -0
  35. JNFIT4oBgHgl3EQfZCuh/content/2301.11251v1.pdf +3 -0
  36. JdE2T4oBgHgl3EQfUgew/vector_store/index.faiss +3 -0
  37. KdFOT4oBgHgl3EQfzDRI/content/2301.12930v1.pdf +3 -0
  38. MNE4T4oBgHgl3EQfiw29/content/tmp_files/2301.05137v1.pdf.txt +2136 -0
  39. MNE4T4oBgHgl3EQfiw29/content/tmp_files/load_file.txt +0 -0
  40. MdE1T4oBgHgl3EQftQUl/content/2301.03374v1.pdf +3 -0
  41. OtAyT4oBgHgl3EQfUfen/vector_store/index.faiss +3 -0
  42. OtAyT4oBgHgl3EQfUfen/vector_store/index.pkl +3 -0
  43. PdE0T4oBgHgl3EQfkAEs/content/2301.02466v1.pdf +3 -0
  44. PdE0T4oBgHgl3EQfkAEs/vector_store/index.pkl +3 -0
  45. PtAyT4oBgHgl3EQf7fqt/vector_store/index.pkl +3 -0
  46. QNFJT4oBgHgl3EQf2i1I/vector_store/index.pkl +3 -0
  47. QNFRT4oBgHgl3EQfJje8/content/tmp_files/2301.13496v1.pdf.txt +1447 -0
  48. QNFRT4oBgHgl3EQfJje8/content/tmp_files/load_file.txt +0 -0
  49. QtFJT4oBgHgl3EQfJyy0/content/tmp_files/2301.11462v1.pdf.txt +2310 -0
  50. QtFJT4oBgHgl3EQfJyy0/content/tmp_files/load_file.txt +0 -0
-9A0T4oBgHgl3EQfPP_v/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e305e359d81679c75fcb2d2df70990a3492235ccdc37b086d64fae53c0263d45
3
+ size 3604525
-NE5T4oBgHgl3EQfRg5P/content/2301.05521v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:08d21cc31de748753afdcc2d37b2ca9b3df2dd167d4fc0332295bd171df33756
3
+ size 753206
.gitattributes CHANGED
@@ -7130,3 +7130,37 @@ edE1T4oBgHgl3EQfLgP4/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -tex
7130
  ctE5T4oBgHgl3EQfEw7I/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7131
  b9FIT4oBgHgl3EQfmyvP/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7132
  FdAyT4oBgHgl3EQfe_hY/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7130
  ctE5T4oBgHgl3EQfEw7I/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7131
  b9FIT4oBgHgl3EQfmyvP/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7132
  FdAyT4oBgHgl3EQfe_hY/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7133
+ 8tFRT4oBgHgl3EQfpzcC/content/2301.13614v1.pdf filter=lfs diff=lfs merge=lfs -text
7134
+ 8tFRT4oBgHgl3EQfpzcC/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7135
+ MdE1T4oBgHgl3EQftQUl/content/2301.03374v1.pdf filter=lfs diff=lfs merge=lfs -text
7136
+ VNFKT4oBgHgl3EQfmS5u/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7137
+ 4tE4T4oBgHgl3EQf1A0X/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7138
+ 1dAzT4oBgHgl3EQfDfo-/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7139
+ edE1T4oBgHgl3EQfLgP4/content/2301.02979v1.pdf filter=lfs diff=lfs merge=lfs -text
7140
+ 2NE4T4oBgHgl3EQfagwY/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7141
+ 2NE4T4oBgHgl3EQfagwY/content/2301.05064v1.pdf filter=lfs diff=lfs merge=lfs -text
7142
+ aNE0T4oBgHgl3EQf4QKS/content/2301.02736v1.pdf filter=lfs diff=lfs merge=lfs -text
7143
+ dNE5T4oBgHgl3EQfgA8L/content/2301.05630v1.pdf filter=lfs diff=lfs merge=lfs -text
7144
+ dNE5T4oBgHgl3EQfgA8L/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7145
+ VNFKT4oBgHgl3EQfmS5u/content/2301.11857v1.pdf filter=lfs diff=lfs merge=lfs -text
7146
+ YtE1T4oBgHgl3EQfcQR-/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7147
+ JdE2T4oBgHgl3EQfUgew/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7148
+ OtAyT4oBgHgl3EQfUfen/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7149
+ KdFOT4oBgHgl3EQfzDRI/content/2301.12930v1.pdf filter=lfs diff=lfs merge=lfs -text
7150
+ ldAyT4oBgHgl3EQf_voI/content/2301.00912v1.pdf filter=lfs diff=lfs merge=lfs -text
7151
+ -9A0T4oBgHgl3EQfPP_v/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7152
+ rdFJT4oBgHgl3EQfbiw2/content/2301.11539v1.pdf filter=lfs diff=lfs merge=lfs -text
7153
+ rdFJT4oBgHgl3EQfbiw2/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7154
+ 29FST4oBgHgl3EQfYTgs/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7155
+ 79A0T4oBgHgl3EQfOf-b/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7156
+ RtAyT4oBgHgl3EQf7_pZ/content/2301.00848v1.pdf filter=lfs diff=lfs merge=lfs -text
7157
+ mdFPT4oBgHgl3EQf4TUa/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7158
+ etE2T4oBgHgl3EQfxghO/content/2301.04111v1.pdf filter=lfs diff=lfs merge=lfs -text
7159
+ ldAyT4oBgHgl3EQf_voI/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7160
+ hNAyT4oBgHgl3EQf-voB/content/2301.00895v1.pdf filter=lfs diff=lfs merge=lfs -text
7161
+ ntE5T4oBgHgl3EQfIA7J/content/2301.05446v1.pdf filter=lfs diff=lfs merge=lfs -text
7162
+ -NE5T4oBgHgl3EQfRg5P/content/2301.05521v1.pdf filter=lfs diff=lfs merge=lfs -text
7163
+ JNFIT4oBgHgl3EQfZCuh/content/2301.11251v1.pdf filter=lfs diff=lfs merge=lfs -text
7164
+ PdE0T4oBgHgl3EQfkAEs/content/2301.02466v1.pdf filter=lfs diff=lfs merge=lfs -text
7165
+ g9AzT4oBgHgl3EQf4P7n/content/2301.01843v1.pdf filter=lfs diff=lfs merge=lfs -text
7166
+ q9E3T4oBgHgl3EQf8gt0/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
0NAyT4oBgHgl3EQf0_na/content/tmp_files/2301.00729v1.pdf.txt ADDED
@@ -0,0 +1,1646 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ A Closed-Form EVSI Expression for a Multinomial
2
+ Data-Generating Process
3
+ Adam Fleischhacker∗, Pak-Wing Fok†, Mokshay Madiman‡, Nan Wu§
4
+ January 3, 2023
5
+ Abstract
6
+ This paper derives analytic expressions for the expected value of
7
+ sample information (EVSI), the expected value of distribution informa-
8
+ tion (EVDI), and the optimal sample size when data consists of inde-
9
+ pendent draws from a bounded sequence of integers. Due to challenges
10
+ of creating tractable EVSI expressions, most existing work valuing data
11
+ does so in one of three ways: 1) analytically through closed-form ex-
12
+ pressions on the upper bound of the value of data, 2) calculating the
13
+ expected value of data using numerical comparisons of decisions made
14
+ using simulated data to optimal decisions where the underlying data
15
+ distribution is known, or 3) using variance reduction as proxy for the
16
+ uncertainty reduction that accompanies more data. For the very flex-
17
+ ible case of modelling integer-valued observations using a multinomial
18
+ data-generating process with Dirichlet prior, this paper develops ex-
19
+ pressions that 1) generalize existing beta-Binomial computations, 2)
20
+ do not require prior knowledge of some underlying “true” distribution,
21
+ and 3) can be computed prior to the collection of any sample data.
22
+ 1
23
+ Introduction
24
+ The seminal work of [34] introduced preposterior analysis, a Bayesian recipe
25
+ for estimating the value of information (VOI) prior to knowing the informa-
26
+ ∗Department of Business Administration, University of Delaware, Newark, DE 19716,
27
+ email: ajf@udel.edu
28
+ †Department of Mathematical Sciences, University of Delaware, Newark, DE 19716,
29
+ email: pakwing@udel.edu
30
+ ‡Department of Mathematical Sciences, University of Delaware, Newark, DE 19716,
31
+ email: madiman@udel.edu
32
+ §Institute for Financial Services Analytics, University of Delaware, Newark, DE 19716,
33
+ email: nanw@udel.edu
34
+ 1
35
+ arXiv:2301.00729v1 [stat.ME] 2 Dec 2022
36
+
37
+ tion’s content. The expected value of sample information (EVSI), a particu-
38
+ larly valuable VOI computation, values the information contained in sample
39
+ observations prior to their collection. [34] include many closed-form and oft-
40
+ used expressions for calculating EVSI under the assumption of quadratic
41
+ loss. One such expression is for a Bernoulli data-generating process with
42
+ beta prior distribution (a.k.a.
43
+ a beta-Binomial model); each observation
44
+ being either zero or one [34, Table 6.2, p. 191]. In this paper, we gener-
45
+ alize the beta-binomial EVSI expression beyond binary-valued observations
46
+ to the case where each data point is drawn from a bounded sequence of
47
+ integers. These results expand the availability of tractable VOI expressions
48
+ to a useful scenario where previously value could only be approximated or
49
+ bounded when a closed-form expression was needed.
50
+ Depending on a modeler’s choices of actions, states of uncertainty, loss
51
+ (or utility) functions, and probability models, tractable calculations of VOI
52
+ may exist, but intractable formulations, especially for EVSI, are much more
53
+ common. In fact, reputed statistician Dennis Lindley has remarked that
54
+ the question of sample size “is embarrassingly difficult to answer” due to
55
+ difficulties calculating EVSI [26]. More generally, [14] shows that simply
56
+ characterizing the relationship between information and value is challeng-
57
+ ing; [14]’s work dispels the idea that information value will reliably exhibit
58
+ monotonic relationships with information value determinants such as action
59
+ flexibility, risk aversion, or a decision maker’s wealth.
60
+ While for some EVSI and VOI problems, closed-form solutions are at-
61
+ tainable [34, 5, 4], value of information solutions are often difficult to for-
62
+ mulate.
63
+ Hence, many papers are known for their ability to characterize
64
+ aspects of VOI expressions such as the distributional properties of the ex-
65
+ pected value of perfect information (EVPI) [28], the impact of an exogenous
66
+ variable on EVPI [20], and the additivity of information value when multi-
67
+ ple sources of uncertainty exist [21]. EVSI calculations, in particular, often
68
+ result in intractable expressions of multiple integrals where only numerical
69
+ methods can yield results [25]. Even then, many numerical methods still
70
+ require further simplifying assumptions (see, e.g., [36]). While it is possi-
71
+ ble to approximate VOI computations via normal approximations (see, e.g.,
72
+ [30, 19]) or using a computationally intense simulation-based methodology
73
+ (see, e.g., [10, 37]), closed-form expressions yield instantaneous and accurate
74
+ value computations with more interpretable insights regarding the effects of
75
+ prior beliefs and sample sizes.
76
+ In this paper, we provide a new EVSI calculation for a flexible (i.e. multi-
77
+ nomial) data-generating process that adheres to three desiderata outlined
78
+ in [34, p.44]:
79
+ 2
80
+
81
+ Tractable
82
+ EVSI is easily calculated using a closed-
83
+ form expression.
84
+ Rich
85
+ A decision maker’s prior beliefs and in-
86
+ formation are readily incorporated as part
87
+ of the calculation.
88
+ Interpretable
89
+ The expression for EVSI provides insight
90
+ as to the effects of prior beliefs and sam-
91
+ ple size choices on the expected value of
92
+ a sample.
93
+ Generating Process
94
+ Conjugate Prior
95
+ Source
96
+ Bernoulli(θ)
97
+ (θ) ∼Beta
98
+ [34]
99
+ [32]
100
+ Poisson(λ)
101
+ λ ∼gamma
102
+ [34]
103
+ Normal(µ, σ)
104
+ µ ∼ Normal, σ known
105
+ [34]
106
+ µ known, σ2 ∼ inv. Gamma
107
+ [34]
108
+ σ2 ∼ inv. Gamma, µ|σ2 ∼ Normal
109
+ [34]
110
+ Multinomial(t)1
111
+ t ∼ Dirichlet
112
+ This Paper
113
+ Table 1: Position of this paper in comparison to other tractable EVSI cal-
114
+ culations.
115
+ Shown in Table 1, our point of departure is generalizing the EVSI cal-
116
+ culation for a Bernoulli data-generating process with beta prior (a.k.a. a
117
+ beta-binomial model) to the case of a multinomial data generating process
118
+ with Dirichlet prior. Rich treatment and illustrative examples surround-
119
+ ing EVSI calculations for the beta-binomial conjugacy can be found in [15].
120
+ Additionally, [32] provide explicit closed-form value of information compu-
121
+ tations for the beta-binomial case and is very close in spirit to this work,
122
+ but does not investigate the Dirichlet-multinomial setting. In relation to
123
+ the multinomial sampling process we explore in this paper, existing work
124
+ has focused on non-utility based approaches where data is valued based on
125
+ its ability to bound a parameter of interest within a certain level of preci-
126
+ sion [1, 6]. Our approach, in contrast, extends the utility-based valuation
127
+ of sampling to a multinomial sampling environment to yield closed-form
128
+ expressions for both EVSI and the expected value of distribution informa-
129
+ tion (EVDI). Publication of analytically tractable expressions will be able
130
+ 3
131
+
132
+ to supplant the still-present usage of Monte Carlo simulation in multinomial
133
+ settings (see, e.g., [38]).
134
+ When closed-form EVSI expressions are unavailable, quantification of
135
+ value created through uncertainty reduction typically relies on one of three
136
+ techniques: 1) closed-form expressions on the upper bound of the value of
137
+ data, 2) simulated comparisons between valuing decisions made by an or-
138
+ acle who knows the underlying data distribution to decisions made by a
139
+ less-informed decision maker, or 3) using variance-reduction as a proxy for
140
+ how data reduces underlying uncertainty in the data-generating process. For
141
+ examples of the first type, [27] bound EVPI for a risk-averse decision maker
142
+ and [40] place an upper bound on the value of knowing the true distribu-
143
+ tion when one already knows the mean and variance of that distribution.
144
+ Examples of the second type often compare a Bayesian updating procedure
145
+ to a known optimal solution [8, 29, 7, 35]. Lastly, computing the value of
146
+ variance reduction independent of the specific quantity of data is also seen
147
+ within the literature [11, 22].
148
+ 2
149
+ Problem Setup
150
+ Despite substantial efforts, notation for preposterior analysis has not been
151
+ standardized and is often a matter of personal taste [33]. To aid the reader
152
+ with this paper’s notation surrounding its random variables and their real-
153
+ izations, we present the following summary breaking the notation into three
154
+ levels of analysis:
155
+ 1.
156
+ Data/Sample.
157
+ Data is an integer-valued random variable with support
158
+ {0, 1, . . . , M}. Sample is a random vector referring to either a sequence of n data
159
+ observations or a vector of counts representing the number of occurrences of each
160
+ potential data value recorded in n observations.
161
+ D:
162
+ A random variable representing a single data observation.
163
+ d:
164
+ A single realization of D with integer valued support: d ∈ {0, 1, . . . , M}.
165
+ X ≡ (X1, . . . , Xn):
166
+ A random vector of n observations of D.
167
+ x ≡ (x1, . . . , xn):
168
+ A realization of data vector X.
169
+ Dn:
170
+ The support of X when n realizations are observed.
171
+ nk:
172
+ the number of times that k ∈ {0, 1, . . . , M} appears in x.
173
+ (n0, n1, . . . , nM):
174
+ A vector of counts of occurrences for each potential data value.
175
+ 2. Data/Sampling Distributions. Data and sampling distributions are iden-
176
+ tical terms referring to the probability distribution governing the data-generating
177
+ process. Data distribution refers to generating individual data points and sampling
178
+ 1With support interpreted as a sequence of integer values.
179
+ 4
180
+
181
+ distribution preferred when talking about a sequence of observations.
182
+ T ≡ (T0, T1, . . . , TM):
183
+ A random vector representing a data distribution.
184
+ Random elements Tk are data distribution parameters
185
+ representing the probability of data realization being k.
186
+ t ≡ (t0, t1, . . . , tM):
187
+ A realization of random vector T such that
188
+ tk = p(D = k) for k ∈ {0, 1, . . . , M}.
189
+ t∗:
190
+ The “true” data distribution or sampling distribution;
191
+ only knowable by an oracle.
192
+ T :
193
+ The space or set of all possible data distributions. T, t, t∗ ∈ T .
194
+ 3. Prior/Posterior Distributions. Continuous multivariate probability distri-
195
+ butions with domain of all possible data distributions.
196
+ π:
197
+ A prior from which data distributions are generated.
198
+ πX:
199
+ A posterior that updates π in light of data X.
200
+ 2.1
201
+ Modelling Data and Loss
202
+ Consider a data-generating process that generates independent and identi-
203
+ cally distributed samples from a bounded sequence of M + 1 integers. For
204
+ notational simplicity, we rescale the sequence to be [M] ≡ {0, 1, . . . , M}. For
205
+ practical motivation, the data could represent product demand and the goal
206
+ is to make accurate predictions for inventory control [39]. For the specific
207
+ case of demand uncertainty, we note that there are asymmetric and other
208
+ loss functions that would be preferred to the quadratic loss function used
209
+ here, but closed-form expressions are not forthcoming for those cases.
210
+ The data-generating process is governed by an unknown data distribu-
211
+ tion, t, with discrete-finite support [M]. Thus the statistical model for the
212
+ data-generating process is parameterized by the standard M-dimensional
213
+ simplex of probabilities
214
+ T = {t = (t0, . . . , tM) ∈ RM+1
215
+ +
216
+ : t0 + . . . + tM = 1};
217
+ this infinite (but finite-dimensional) parameter space describes how we are
218
+ labeling the potential data distributions. If the sample size of the data is n,
219
+ we have n values x1, . . . , xn ∈ [M] being generated by the data-generating
220
+ process.
221
+ For a given t ∈ T , the associated data-generating process p(n)
222
+ t
223
+ assigns probability
224
+ p(n)
225
+ t
226
+ (x1, . . . , xn) =
227
+ n
228
+
229
+ i=1
230
+ txi
231
+ (1)
232
+ 5
233
+
234
+ to this particular sequence of data values. In particular, if the sample size
235
+ is 1, the data-generating process is simply given by
236
+ pt(d) ≡ p(1)
237
+ t (d) = td,
238
+ d ∈ [M].
239
+ It is clear that the number of occurrences of particular data values in the
240
+ sample is a sufficient statistic for the model described, and that the sam-
241
+ pling distribution for this sufficient statistic is just the multinomial model.
242
+ Specifically, if nd = |{1 ≤ i ≤ n : xi = d}|, then (n0, . . . , nM) is a sufficient
243
+ statistic, and we have, with obvious abuse of notation,
244
+ pt(n0, . . . , nM) =
245
+
246
+ n
247
+ n0 · · · nM
248
+ � M
249
+
250
+ d=0
251
+ tnd
252
+ d .
253
+ (2)
254
+ Note that n0+. . .+nM = n by definition; so we do not write the superscript
255
+ (n) when using the sufficient statistic to represent the data.
256
+ When making predictions for future data, ideally the action (or predic-
257
+ tion) is close to the actual data realization. For tractability, we consider a
258
+ quadratic terminal opportunity loss function for a single prediction to be of
259
+ the following form:
260
+ ℓ(d, a) = k(d − a)2
261
+ (3)
262
+ where k > 0 is a known constant, a is the action/prediction, and d ∈ [M] is
263
+ the actual data realization.
264
+ To briefly make the above notation more concrete, let’s imagine fore-
265
+ casting product demand for a product that will sell between 0 and 5 units
266
+ (M = 5). Each period’s i.i.d demand, d ∈ {0, 1, . . . , 5}, has an associated
267
+ probability of occurrence, pt(0), pt(1), . . . , pt(5), which is represented more
268
+ compactly as t0, t1, . . . , t5. The effectiveness of any action will be measured
269
+ using quadratic loss scaled by a factor k such that if k = 5, d = 4, and
270
+ a = 1, then ℓ(4, 1) = 45. The decision maker is contemplating the value of
271
+ n = 3 observations where generated data, (x1, x2, x3), might be something
272
+ like (0, 5, 0) and the associated sufficient statistic of counts, (n0, . . . , n5),
273
+ would be (2, 0, 0, 0, 0, 1). Note that t ≡ t0, t1, . . . , t5 parameterizes both the
274
+ data-generating process of eq. (1) yielding (x1, x2, x3) and the equivalent
275
+ sampling process of eq. (2) yielding (n0, . . . , n5). As a result, we refer to t
276
+ as both data distribution and sampling distribution depending on context.
277
+ 6
278
+
279
+ 2.2
280
+ Preposterior Analysis
281
+ For any data distribution t, define the expectation of loss as:
282
+ R(t, a) = ED|T=t [ℓ(D, a)] =
283
+ M
284
+
285
+ d=0
286
+ pt(d)ℓ(d, a).
287
+ (4)
288
+ where R(t, a) is known as the Bayes risk. Since a decision maker (DM) does
289
+ not know the underlying “true” t∗ ∈ T data distribution, the minimum
290
+ Bayes risk, mina R(t∗, a), is likely unachievable.
291
+ For a DM, risk is evaluated on an average basis based on the probability
292
+ distribution the DM places over the simplex T . Without any sample obser-
293
+ vations, this distribution is the prior π over all possible data distributions
294
+ in T . The average risk of taking action a using prior π is
295
+ ¯R(π, a) = ET [R(T, a)] ,
296
+ (5)
297
+ with T ∼ π. The Bayes action for π is
298
+ a∗(π) = arg min
299
+ a∈A
300
+ ¯R(π, a).
301
+ (6)
302
+ The Bayes risk for π is
303
+ ¯R(π, a∗(π)) = min
304
+ a∈A
305
+ ¯R(π, a).
306
+ (7)
307
+ Access to a sample X ≡ (X1, . . . , Xn) results in a different decision with
308
+ different risk.
309
+ With sample observations, the DM applies Bayes’ rule to
310
+ update π to πX (the posterior) and calculates the associated optimal Bayes
311
+ action a∗(πX). Since X is unknown prior to actually collecting the sample,
312
+ the Bayes risk for πX is itself a random variable. Hence, we evaluate the
313
+ DM’s prior expectation of loss with sample information over all possible
314
+ samples X,
315
+ EX
316
+ � ¯R(πX, a∗(πX))
317
+
318
+ = ET EX|T [R(T, a∗(πX))] ,
319
+ (8)
320
+ with T ∼ π and the right-hand side expression derived by substituting πX
321
+ for π in eq. (5) and applying the law of total expectation.
322
+ Thus, the expected value of a sample of information (EVSI), Vn(π), is the
323
+ difference between the prior expectations of loss with and without sample
324
+ X under prior π:
325
+ Vn(π)
326
+ =
327
+ ¯R(π, a∗(π)) − EX
328
+ � ¯R(πX, a∗(πX))
329
+
330
+ (9)
331
+ =
332
+ ET [R(T, a∗(π))] − ET EX|T [R(T, a∗(πX))]
333
+ (10)
334
+ 7
335
+
336
+ where T ∼ π and eq. (10) follows from eqs. (5) and (8). Proposition 2.1 for-
337
+ malizes our intuition that this expected value of sample information should
338
+ be non-negative.
339
+ Proposition 2.1. Suppose data distribution T ≡ (T0, . . . , TM) is drawn
340
+ from a given prior π. Assume further that a DM is given n samples X ≡
341
+ (X1, . . . , Xn) and updates his/her prior to the posterior πX. Then, under
342
+ quadratic loss, the expected value of these n samples is non-negative, i.e.
343
+ Vn(π) = ET [R(T, a∗(π))] − ET EX|T [R(T, a∗(πX))] ≥ 0.
344
+ (11)
345
+ Proof. See Appendix.
346
+
347
+ Because the ordering within the sample X does not matter, the inner ex-
348
+ pectation in (11) is performed over (n0, n1, . . . , nM) ∼ Multinomial(t) con-
349
+ ditioned on T = t where nj is the number of times that j ∈ [M] appears in
350
+ the sample, and the outer expectation is performed over T ∼ π.
351
+ 3
352
+ Tractable Valuation of Sample Information
353
+ To arrive at a tractable valuation for (10), we leverage the Dirichlet distri-
354
+ bution as a prior for three reasons: 1) it is a conjugate prior to categori-
355
+ cal/multinomial outcomes, 2) its support is the M-dimensional simplex T ,
356
+ and 3) it has flexibility to model many types of prior information for the de-
357
+ cision maker. With the Dirichlet assumption, the main result of this paper,
358
+ Theorem 3.1, can be presented:
359
+ Theorem 3.1. For data distribution T with support [M] and prior π =
360
+ Dirichlet(α0, α1, . . . , αM), the expected reduction in quadratic loss after ob-
361
+ serving n data samples, also called the expected value of sample information
362
+ (EVSI), is given by:
363
+ Vn(π) =
364
+ kn(c2 − c2
365
+ 1)
366
+ (n + α)(1 + α).
367
+ (12)
368
+ where α = �M
369
+ d=0 αd is the precision/concentration parameter of the Dirich-
370
+ let distribution (see [16]) and c1 =
371
+ 1
372
+ α
373
+ �M
374
+ d=0 dαd and c2 =
375
+ 1
376
+ α
377
+ �M
378
+ d=0 d2αd
379
+ are the first and second moments of the data under the marginal likelihood
380
+ (α1, α2, . . . , αM)/α.
381
+ Proof. See Appendix.
382
+
383
+ 8
384
+
385
+ Theorem 3.1 gives the expected value of observing an n-trial multinomial
386
+ sample with Dirichlet prior where support of the underlying data-generating
387
+ process is the bounded sequence of integers [M] = {0, 1, . . . , M}. This is a
388
+ natural generalization of valuing an n-trial binomial sample with beta prior
389
+ where support of the underlying data-generating process is restricted such
390
+ that [M] = {0, 1}. With just a slight change of notation, we know from [32]
391
+ that EVSI for the beta-binomial case in closed-form is:
392
+ kn
393
+ n + α0 + α1
394
+ ·
395
+ α0α1
396
+ (α0 + α1)2(α0 + α1 + 1)
397
+ (13)
398
+ where π ∼ Beta(α0, α1). Replacing this prior with the equivalent Dirichlet
399
+ parameterization of π ∼ Dirichlet(α0, α1) and using Theorem 3.1 yields an
400
+ identical result:
401
+ Vn(π) =
402
+ kn(c2 − c2
403
+ 1)
404
+ (n + α)(1 + α)
405
+ =
406
+ kn
407
+ (n + α0 + α1) ·
408
+ α1
409
+ α0+α1 −
410
+ α2
411
+ 1
412
+ (α0+α1)2
413
+ (α0 + α1 + 1)
414
+ =
415
+ kn
416
+ (n + α0 + α1) ·
417
+ α0α1
418
+ (α0 + α1)2(α0 + α1 + 1)
419
+ (14)
420
+ As a direct consequence of Theorem 3.1, when n → ∞, we have an
421
+ expression for the expected value of distribution information (EVDI), as an
422
+ infinite sample gives the data distribution exactly:
423
+ V∞(π) = lim
424
+ n→∞ Vn(π) = k(c2 − c2
425
+ 1)
426
+ 1 + α
427
+ .
428
+ (15)
429
+ Lastly, we can express the efficiency η of the sample information as a function
430
+ of the number of sample points using the ratio of (12) to (15) as:
431
+ η =
432
+ n
433
+ n + α.
434
+ (16)
435
+ Hence, the percentage of value obtained through sampling is given by the
436
+ ratio of the number of data points n to the sum of the n data points
437
+ and the concentration parameter α of a Dirichlet distribution. This sam-
438
+ pling efficiency calculation directly simplifies to the known formula of the
439
+ beta-binomial case from [34](in our notation):
440
+ η = n/(α0 + α1) where
441
+ π ∼ Beta(α0, α1).
442
+ Again, we make the notation more concrete, by revisiting our forecasting
443
+ product demand example from the end of §2.1. Recall, we have a product
444
+ 9
445
+
446
+ that will sell between 0 and 5 units (M = 5) and loss is scaled by k =
447
+ 5. The decision maker is contemplating the value of n = 3 observations.
448
+ Introducing a zero-inflated prior π ∼ Dirichlet( 10
449
+ 6 , 1
450
+ 6, 1
451
+ 6, 1
452
+ 6, 1
453
+ 6, 1
454
+ 6) means α =
455
+ 15
456
+ 6 , c1 =
457
+ 6
458
+ 15 · (0 · 10
459
+ 6 + 1 · 1
460
+ 6 + 2 · 1
461
+ 6 + 3 · 1
462
+ 6 + 4 · 1
463
+ 6 + 5 · 1
464
+ 6) = 1, c2 =
465
+ 6
466
+ 15 ·
467
+ (0 · 10
468
+ 6 + 1 · 1
469
+ 6 + 4 · 1
470
+ 6 + 9 · 1
471
+ 6 + 16 · 1
472
+ 6 + 25 · 1
473
+ 6) = 11
474
+ 3 . Plugging into eq. (12)
475
+ yields EVSI V3(π) = 160
476
+ 77 ≈ 2.08. and EVDI V∞(π) = 80
477
+ 21 ≈ 3.81. From
478
+ eq. (16) we get η =
479
+ 6
480
+ 11 ≈ 54.5%, so the learning from n = 3 samples is
481
+ expected to provide more than half of the maximum possible reduction in
482
+ loss. Following from eqs. (26) - (31), a∗(π) = 1 and prior expected loss
483
+ ¯R(π, a∗(π)) = 5 · (−12 · 10
484
+ 15 + 02 · 1
485
+ 15 + 12 · 1
486
+ 15 + 22 · 1
487
+ 15 + 32 · 1
488
+ 15 + 42 · 1
489
+ 15) =
490
+ 40
491
+ 3 ≈ 13.33. And thus, we can also get the prior expectation of posterior loss
492
+ EX
493
+ � ¯R(πX, a∗(πX))
494
+
495
+ = ¯R(π, a∗(π)) − V3(π) = 40
496
+ 3 − 160
497
+ 77 ≈ 11.26.
498
+ 4
499
+ Notes on Richness and Interpretability of Mod-
500
+ elings Assumptions
501
+ In the previous section, we showed one of the three EVSI desiderata, tractabil-
502
+ ity, can be achieved for a multinomial data-generating process with Dirichlet
503
+ prior. The multinomial distribution is flexible enough to model any discrete
504
+ (finite) data distribution. Its prior, the Dirichlet distribution, is also flexible
505
+ in its ability to model a wide range of distributions over a simplex. Yet, some
506
+ sacrifice of richness in modeling prior beliefs is made in the name of tractabil-
507
+ ity. Most notably, a more rich/flexible alternative prior over a simplex is the
508
+ logistic-normal distribution [3, see discussion in]. The most glaring weak-
509
+ ness of the Dirichlet distribution is in modeling prior beliefs where there is
510
+ some type of correlation structure between data observations. For example,
511
+ observing a high data value, say 100, would make one think values of 101
512
+ and 99 are also more likely to occur than data values further away. How-
513
+ ever, the Dirichlet distribution, as a prior distribution to multinomial data,
514
+ is unable to capture this structure. Notably, the distribution-free underpin-
515
+ nings of the Kaplan-Meier estimator also ignore this potential correlation
516
+ among data observations, yet shows favorable results in a similar repeated
517
+ newsvendor setting [17] .
518
+ The richness of the Dirichlet prior is best seen through the lens of its intu-
519
+ itive reparameterization [16]. Let the concentration parameter α = �M
520
+ i=0 αi
521
+ and let the vector m =
522
+ � α0
523
+ α , α1
524
+ α , . . . , αM
525
+ α
526
+
527
+ represent the mean where the
528
+ expected mean of the data observations is given as c1 =
529
+ 1
530
+ α
531
+ �M
532
+ i=0 iαi =
533
+ �M
534
+ i=0 imi. When α is small, say α ≤ M, the prior distribution over the
535
+ simplex can differ greatly from m and reflect a decision maker’s uncertainty
536
+ 10
537
+
538
+ 0.00
539
+ 0.25
540
+ 0.50
541
+ 0.75
542
+ α0
543
+ α5
544
+ α10
545
+ α15
546
+ α20
547
+ Dirichlet Parameter
548
+ Parameter Value
549
+ Dirichlet Shape Parameter for M = 20
550
+ 0.00
551
+ 0.05
552
+ 0.10
553
+ 0.15
554
+ 0.20
555
+ p0
556
+ p5
557
+ p10
558
+ p15
559
+ p20
560
+ Multinomial Parameter
561
+ Parameter Value
562
+ 0.00
563
+ 0.05
564
+ 0.10
565
+ 0.15
566
+ 0.20
567
+ p0
568
+ p5
569
+ p10
570
+ p15
571
+ p20
572
+ Multinomial Parameter
573
+ Parameter Value
574
+ Sample Realizations of Multinomial Parameters
575
+ EVDI
576
+ 0.0
577
+ 0.5
578
+ 1.0
579
+ 1.5
580
+ 2.0
581
+ 0
582
+ 10
583
+ 20
584
+ 30
585
+ 40
586
+ 50
587
+ # of samples (n)
588
+ value
589
+ EVSI
590
+ EVSI as Function of n for M = 20
591
+ Concentration Parameter = 10
592
+ 0
593
+ 1
594
+ 2
595
+ 3
596
+ 4
597
+ α0
598
+ α5
599
+ α10
600
+ α15
601
+ α20
602
+ Dirichlet Parameter
603
+ Parameter Value
604
+ Dirichlet Shape Parameter for M = 20
605
+ 0.00
606
+ 0.05
607
+ 0.10
608
+ p0
609
+ p5
610
+ p10
611
+ p15
612
+ p20
613
+ Multinomial Parameter
614
+ Parameter Value
615
+ 0.000
616
+ 0.025
617
+ 0.050
618
+ 0.075
619
+ 0.100
620
+ 0.125
621
+ p0
622
+ p5
623
+ p10
624
+ p15
625
+ p20
626
+ Multinomial Parameter
627
+ Parameter Value
628
+ Sample Realizations of Multinomial Parameters
629
+ EVDI
630
+ 0.0
631
+ 0.1
632
+ 0.2
633
+ 0.3
634
+ 0.4
635
+ 0
636
+ 10
637
+ 20
638
+ 30
639
+ 40
640
+ 50
641
+ # of samples (n)
642
+ value
643
+ EVSI
644
+ EVSI as Function of n for M = 20
645
+ Concentration Parameter = 50
646
+ Figure 1:
647
+ Graphical depiction of the Dirichlet prior parameters, poten-
648
+ tial realizations for that prior (i.e. the multinomial parameters), and the
649
+ EVSI/EVDI calculations as a function of n samples for the given prior. Top
650
+ row for concentration parameter α = 10 and bottom row for concentration
651
+ parameter α = 50
652
+ 11
653
+
654
+ around their expectation. As α is made larger, the prior distribution will
655
+ concentrate probability density near m and reflect greater confidence. We
656
+ present a graphical overview of this in Figure 1 for two different concentra-
657
+ tion parameters. As seen, when α is smaller (top row of Figure 1) the real-
658
+ ized multinomial parameters (middle-top plot) can be further away from the
659
+ mean m (which is proportional to the parameters in the top-left plot). As α
660
+ increases (bottom-row) the prior distribution becomes much more informa-
661
+ tive and multinomial parameters will most likely mirror the prior Dirichlet
662
+ parameters.
663
+ In terms of interpretability, Theorem 3.1 formalizes our intuition about
664
+ what drives the value of data. Specifically, data is valuable when 1) the
665
+ sample contains a lot of data (high n), 2) the expected variance of the
666
+ data distribution is large (high c2 − c2
667
+ 1), and 3) there is a lot of uncertainty
668
+ regarding the true data distribution (α is small). Additionally, the calcu-
669
+ lation for EVDI (eq. 15) gives an interpretable upper bound on the value
670
+ of data where high variance pushes to make samples more valuable and a
671
+ high concentration parameter makes samples less valuable. Lastly, the equa-
672
+ tion for efficiency (16) adds further insight by stating how quickly the upper
673
+ bound on the value of data is approached; basically, the smaller the Dirichlet
674
+ concentration parameter, the more quickly EVDI is approached with each
675
+ subsequent data point.
676
+ 5
677
+ Illustrative Examples
678
+ In this section, we demonstrate how the tractable formulation for EVSI,
679
+ equation (12), can serve as a building block inside of other research initia-
680
+ tives. The first example explores sample size optimization and the second
681
+ example shows how a tractable EVSI calculation can lead to a tractable de-
682
+ cision policy in a two-stage production planning problem. In the third/last
683
+ example, the EVSI formula provides a foundation from which to benchmark
684
+ heuristic updating procedures that seek to estimate an underlying unknown
685
+ data distribution.
686
+ 5.1
687
+ The Choice of Sample Size
688
+ We now explore a decision maker’s objective to choose the number of sam-
689
+ ple points to collect in such a way as to minimize his expected loss when
690
+ assuming expected sampling cost, Cs(n), is a linear function of the number
691
+ of sampled points n:
692
+ Cs(n) = K + sn
693
+ (17)
694
+ 12
695
+
696
+ where s is the cost of one sample and K represents the fixed costs of sam-
697
+ pling.
698
+ The loss function to be minimized, ℓs(n), combines equations (12) and
699
+ (17):
700
+ ℓs(n) = −
701
+ kn(c2 − c2
702
+ 1)
703
+ (n + α)(1 + α) + K + sn
704
+ (18)
705
+ And assuming for practical purposes that n can be treated continuously,
706
+ we get the optimal sample size:
707
+ n∗ =
708
+
709
+ α
710
+ (1 + α)
711
+ k
712
+ s (c2 − c2
713
+ 1) − α
714
+ (19)
715
+ for cases where n∗ is positively valued and the fixed costs of sampling K
716
+ can be recovered, i.e. Vn(π) > Cs(n∗). In all other cases, n∗ = 0. Equation
717
+ (19) has a nice economic interpretation where the three terms represent the
718
+ strength of the prior, the ratio between the scaling of the quadratic loss
719
+ costs and the unit sampling costs, and the predicted variance of the data
720
+ distribution.
721
+ 5.2
722
+ Two-Stage Production Planning
723
+ The example shown here is a simple two-stage production planning problem
724
+ (see, e.g., [9]) where the decision maker seeks to optimally schedule the 2nd
725
+ production run.
726
+ Assume J periods make up a selling season. Each period, j ∈ J faces in-
727
+ dependent and identical categorical demand with Dirichlet prior and quadratic
728
+ loss (i.e. a repeated newsvendor setting with quadratic loss) with identical
729
+ shipments scheduled for each period. A decision maker can choose either 1)
730
+ to schedule the delivery quantity for each period in the entire selling season
731
+ or, 2) at cost K can specify a period j∗ after which the scheduled delivery
732
+ quantity can be changed. Assuming this change date will be contractually
733
+ set in advance of the selling season, find j∗ to minimize expected net costs
734
+ over the entire season J.
735
+ The net cost function for this problem is:
736
+ C(j) =
737
+
738
+
739
+
740
+ 0,
741
+ if j = 0,
742
+ K − (J − j)
743
+ kj(c2 − c2
744
+ 1)
745
+ (j + α)(1 + α),
746
+ if j ∈ (0, J]
747
+ (20)
748
+ 13
749
+
750
+ When j ∈ (0, J], the net cost function C(·) is strictly convex and has a
751
+ unique global minimum value. The optimal period j∗ is
752
+ j∗ = arg min
753
+ j∈{0,1,...,J}
754
+ C(j)
755
+ When min C(j) = 0 for 0 < j ≤ J, we choose j∗ = 0.
756
+ For the case when min C(j) < 0, we have
757
+ j∗ =
758
+
759
+ α(J + α) − α
760
+ Considering that j∗ must be a non-negative integer, summarizing differ-
761
+ ent cases we have the optimal j∗ as
762
+ j∗ =
763
+
764
+
765
+
766
+
767
+
768
+ 0,
769
+ if min
770
+ j∈[0,J]C(j) = 0,
771
+ arg min
772
+ j∈{⌊j0⌋,⌈j0⌉}
773
+ C(j),
774
+ if min
775
+ j∈[0,J]C(j) < 0.
776
+ (21)
777
+ where j0 =
778
+
779
+ α(J + α) − α.
780
+ 5.3
781
+ Benchmarking Data-Driven Algorithms
782
+ An active area of research is to propose algorithms for decisions in repeated
783
+ settings where minimal assumptions about the underlying data distribution
784
+ are known. These approaches include Sample Average Approximation(SAA)
785
+ [24, 23], concave adaptive value estimation (CAVE) [12], and Second Order
786
+ Belief Maximum Entropy (SOBME) [35]. When benchmarking these algo-
787
+ rithms, it is customary to pick a handful of “true” distributions where the
788
+ algorithm competes against a known optimal solution.
789
+ With the introduction of a closed-form EVSI calculation in the context
790
+ of a Dirichlet prior, a more robust benchmarking scenario can be achieved.
791
+ Instead of picking a “true” data distribution, we pick a “true prior” from the
792
+ Dirichlet family with support matching the problem of interest. This prior
793
+ can be used to then simulate “true” data distributions (as many as we want)
794
+ by which we can estimate the reduction in squared loss as a function of n,
795
+ the number of data samples. Given this setup, a comparison of a proposed
796
+ algorithm can be made against a known optimal updating procedure. After
797
+ all, it is the updating procedure that we are seeking to validate, and the opti-
798
+ mal updating procedure to benchmark new algorithms against is, therefore,
799
+ the Bayesian one detailed in the proof of Theorem 3.1 (see appendix).
800
+ As a proof of concept, Figure 2 is an example benchmarking the well-
801
+ known sample average approximation (SAA) (see [24]) against the known op-
802
+ timal Bayesian updating procedure (BAYES) using a Dirichlet(α0, α1, . . . , αM)
803
+ 14
804
+
805
+ GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG
806
+ Optimal Squared Loss (i.e. distribution known)
807
+ 18
808
+ 20
809
+ 22
810
+ 0
811
+ 10
812
+ 20
813
+ 30
814
+ 40
815
+ # of sample data points
816
+ expected quadratic loss
817
+ Updating
818
+ Method
819
+ G BAYES
820
+ SAA
821
+ Expected Loss as Function of n for M = 20
822
+ Figure 2: Comparing the sample average approximation(SAA) updating
823
+ procedure to the known Bayesian (BAYES) optimal updating procedure.
824
+ prior with M = 20, α = 10, and m ∝ {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 13, 11, 9, 7, 5, 3, 1}
825
+ (chosen to be slightly skewed). In this scenario, we see the value of prior in-
826
+ formation in small data settings as BAYES outperforms SAA. It also shows
827
+ how as the amount of data increases, the non-parametric SAA algorithm’s
828
+ performance improves and closely mimics that of the optimal Bayesian up-
829
+ dating procedure.
830
+ 6
831
+ Conclusion
832
+ The use of preposterior analysis in this paper provides a formal method for
833
+ valuing data prior to its collection and as such, should serve as a build-
834
+ ing block in many systems and models going forward. By expanding the
835
+ support of the underlying data-generating process from [M] = {0, 1} to
836
+ [M] = {0, 1, . . . , M}, the beta-binomial EVSI calculations are successfully
837
+ generalized to a Dirichlet-multinomial setting. Using this new EVSI com-
838
+ putation, three illustrative examples valuing data prior to its collection are
839
+ shown, there are potentially many other contexts where this tractable formu-
840
+ lation might also prove useful. Researchers in two particular areas, medical
841
+ decision making and active (machine) learning are known to be interested in
842
+ EVSI types of calculations (see, e.g., [2, 13, 18, 31]). And we look forward
843
+ to hearing of other useful deployments for this method of valuing data prior
844
+ to its collection.
845
+ 15
846
+
847
+ A
848
+ Proof of Proposition 2.1 and Theorem 3.1
849
+ A.1
850
+ Proof of Proposition 2.1
851
+ The expected value of sample information is
852
+ Vn (π) = ET [R(T, a∗(π))] − ET EX|T [R(T, a∗(πX))] .
853
+ (22)
854
+ For the first term in eq. (22), we have
855
+ ET [R(T, a∗(π))] = kET
856
+
857
+ ED|T
858
+
859
+ (D − a∗ (π))2��
860
+ = kET
861
+
862
+ ED|T
863
+
864
+ (D − E [D])2��
865
+ = kED
866
+
867
+ (D − E [D])2�
868
+ = kVar [D] .
869
+ (23)
870
+ The second line is due to the optimal action under squared loss being the
871
+ mean (see eq. (30)). The third line of equation (23) follows from the law of
872
+ total expectation. Thus, the optimal Bayes risk without sample information
873
+ under quadratic loss (3) is the marginal variance of D scaled by a factor k.
874
+ Similarly, for the second term in eq. (22) we find
875
+ ET
876
+
877
+ EX|T [R (T, a∗ (πX))]
878
+
879
+ = kET
880
+
881
+ EX|T
882
+
883
+ ED|T
884
+
885
+ (D − a∗ (πX))2���
886
+ = kET
887
+
888
+ EX|T
889
+
890
+ ED|T
891
+ ��
892
+ D − ED|X [D]
893
+ �2���
894
+ = kEX
895
+
896
+ ED|X
897
+ ��
898
+ D − ED|X [D]
899
+ �2��
900
+ = kEX
901
+
902
+ VarD|X [D]
903
+
904
+ .
905
+ (24)
906
+ The optimal Bayes risk under quadratic loss (3) if a sample of size n is to
907
+ be collected is the expected variance of the predictive posterior distribution
908
+ of D scaled by a factor k.
909
+ Combining (22),(23), and (24), we complete the proof:
910
+ Vn (π) = ET [R(T, a∗(π))] − ET
911
+
912
+ EX|T [R (T, a∗ (πX))]
913
+
914
+ = kVar [D] − kEX
915
+
916
+ VarD|X [D]
917
+
918
+ = k
919
+
920
+ Var [D] − EX
921
+
922
+ VarD|X [D]
923
+ ��
924
+ = kVarX
925
+
926
+ ED|X [D]
927
+
928
+ ≥ 0.
929
+ (25)
930
+ 16
931
+
932
+ The last equal sign in equation (25) follows from the law of total variance.
933
+ Since k > 0 and VarX
934
+
935
+ ED|X [D]
936
+
937
+ ≥ 0 for any X, we have Vn (π) ≥ 0 for any
938
+ sample size n.
939
+
940
+ A.2
941
+ Proof of Theorem 3.1
942
+ Consider the prior distribution for the data-generating process
943
+ π = Dirichlet(α0, α1, . . . , αM).
944
+ Suppose our information consists of n samples of the data distribution. Let
945
+ nj, j ∈ [M] be the frequency of the data being j so that nj are integers such
946
+ that �M
947
+ j=0 nj = n. Then, because the multinomial and Dirichlet distribu-
948
+ tions are conjugate,
949
+ πX
950
+ =
951
+ Dirichlet(α0 + n0, α1 + n1, . . . , αM + nM).
952
+ Because π and πX both belong to the same class of distributions, we derive
953
+ closed-form valuations for the information X. The corresponding marginal
954
+ likelihoods for π and πX are
955
+ qπ(d)
956
+ =
957
+ αd
958
+ α ,
959
+ qπX(d)
960
+ =
961
+ αd + nd
962
+ α + n ,
963
+ where α = �M
964
+ i=0 αi. If the information happens to occur in such a way
965
+ that nj ∝ αj for each j, then the updated marginal likelihood is unchanged:
966
+ qd(π) = qd(πX), d ∈ [M].
967
+ For convenience, define the quantities
968
+ Z
969
+ =
970
+ 1
971
+ n
972
+ M
973
+
974
+ d=0
975
+ dnd,
976
+ c1
977
+ =
978
+ 1
979
+ α
980
+ M
981
+
982
+ d=0
983
+ dαd,
984
+ c2
985
+ =
986
+ 1
987
+ α
988
+ M
989
+
990
+ d=0
991
+ d2αd,
992
+ where Z represents the average frequency of the sample, c1 the prior expec-
993
+ tation for a sample value, and c2 the prior second moment for the sample
994
+ value.
995
+ 17
996
+
997
+ Given the loss function in (3), the Bayes risk and action without sample
998
+ information can be explicitly calculated
999
+ ¯R(π, a)
1000
+ =
1001
+ ET∼π[R(T, a)],
1002
+ (26)
1003
+ =
1004
+ ET∼π
1005
+ � M
1006
+
1007
+ d=0
1008
+ pT (d)ℓ(d, a)
1009
+
1010
+ ,
1011
+ (27)
1012
+ =
1013
+ M
1014
+
1015
+ d=0
1016
+ ℓ(d, a)ET∼π[pT (d)],
1017
+ (28)
1018
+ =
1019
+ M
1020
+
1021
+ d=0
1022
+ ℓ(d, a)qπ(d),
1023
+ (29)
1024
+ where {qπ(0), qπ(1), . . . , qπ(M)} is the marginal likelihood. The Bayes action
1025
+ minimizes eq. (29):
1026
+ ∂ ¯R(π, a)
1027
+ ∂a
1028
+ =
1029
+ −2k
1030
+ M
1031
+
1032
+ d=0
1033
+ (d − a)qπ(d) = −2k
1034
+ � M
1035
+
1036
+ d=0
1037
+ dqπ(d) − a
1038
+ M
1039
+
1040
+ d=0
1041
+ qπ(d)
1042
+
1043
+ = 0,
1044
+ ⇒ a∗(π)
1045
+ =
1046
+ M
1047
+
1048
+ d=0
1049
+ dqπ(d),
1050
+ =
1051
+ Eqπ[D],
1052
+ (30)
1053
+ =
1054
+ c1.
1055
+ (31)
1056
+ the mean data outcome under the prior marginal likelihood.
1057
+ The corre-
1058
+ sponding Bayes Risk is
1059
+ ¯R(π, a∗(π))
1060
+ =
1061
+ k
1062
+ M
1063
+
1064
+ d=0
1065
+ (d − a∗(π))2qπ(d),
1066
+ =
1067
+ kVarqπ[D],
1068
+ =
1069
+ k(c2 − c2
1070
+ 1).
1071
+ 18
1072
+
1073
+ Similarly, with sample information we have
1074
+ ∂ ¯R(πX, a)
1075
+ ∂a
1076
+ =
1077
+ −2k
1078
+ M
1079
+
1080
+ d=0
1081
+ (d − a)qπX(d),
1082
+ =
1083
+ −2k
1084
+ � M
1085
+
1086
+ d=0
1087
+ dqπX(d) − a
1088
+ M
1089
+
1090
+ d=0
1091
+ qπX(d)
1092
+
1093
+ = 0,
1094
+ ⇒ a∗(πX)
1095
+ =
1096
+ M
1097
+
1098
+ d=0
1099
+ dqπX(d),
1100
+ =
1101
+ EqπX [D],
1102
+ =
1103
+ αc1 + nZ
1104
+ α + n
1105
+ ,
1106
+ (32)
1107
+ which is the mean data outcome under the posterior marginal likelihood.
1108
+ Now expressing EVSI as
1109
+ Vn(π)
1110
+ =
1111
+ ¯R(π, a∗(π)) − ET EX|T R(T, a∗(πX)),
1112
+ (33)
1113
+ note the inner expectation is taken over the data frequency, which follows a
1114
+ multinomial distribution: (n0, . . . , nM) ∼ Multinomial(pt(0), . . . , pt(M))),
1115
+ and the outer expectation is taken over all possible distributions pt∗ ∼
1116
+ Dir(α0, . . . , αM).
1117
+ The first term in (33) has already been evaluated as k(c2 − c2
1118
+ 1). We now
1119
+ calculate the second term.
1120
+ R(t, a∗(πX))
1121
+ =
1122
+ k
1123
+ M
1124
+
1125
+ d=0
1126
+ pt(d)(d − a∗(πX))2
1127
+ =
1128
+ k
1129
+ M
1130
+
1131
+ d=0
1132
+ pt(d)
1133
+
1134
+ d − αc1 + nZ
1135
+ α + n
1136
+ �2
1137
+ ⇒ EX|T=t [R(t, a∗(πX))]
1138
+ =
1139
+ k
1140
+ M
1141
+
1142
+ d=0
1143
+ pt(d)
1144
+
1145
+ d2 −
1146
+ � 2nd
1147
+ α + n −
1148
+ 2nαc1
1149
+ (α + n)2
1150
+
1151
+ EX[Z]
1152
+ −2dαc1
1153
+ α + n +
1154
+ α2c2
1155
+ 1
1156
+ (α + n)2 +
1157
+ n2
1158
+ (α + n)2 EX[Z2]
1159
+
1160
+ .(34)
1161
+ 19
1162
+
1163
+ Since Z(n0, . . . , nM) = 1
1164
+ n
1165
+ �M
1166
+ d=0 dnd,
1167
+ EX|T=t[Z]
1168
+ =
1169
+ M
1170
+
1171
+ d=0
1172
+ dpt(d),
1173
+ EX|T=t[Z2]
1174
+ =
1175
+ VarX|T=t[Z] +
1176
+
1177
+ EX|T=t[Z]
1178
+ �2
1179
+ =
1180
+ 1
1181
+ n
1182
+ M
1183
+
1184
+ d=0
1185
+ d2pt(d) + (n − 1)
1186
+ n
1187
+ � M
1188
+
1189
+ d=0
1190
+ dpt(d)
1191
+ �2
1192
+ ,
1193
+ where the last line follows from the fact
1194
+ VarX|T=t[Z] = VarX|T=t
1195
+
1196
+ 1
1197
+ n
1198
+ M
1199
+
1200
+ d=0
1201
+ dnd
1202
+
1203
+ = 1
1204
+ n2 VarX|T=t
1205
+ � M
1206
+
1207
+ d=0
1208
+ dnd
1209
+
1210
+ = 1
1211
+ n2
1212
+
1213
+
1214
+
1215
+ M
1216
+
1217
+ d=0
1218
+ d2VarX|T=t [nd] + 2
1219
+ M
1220
+
1221
+ 0=i<j
1222
+ ijCovX|T=t (ni, nj)
1223
+
1224
+
1225
+
1226
+ = 1
1227
+ n2
1228
+
1229
+
1230
+
1231
+ M
1232
+
1233
+ d=0
1234
+ d2npt(d) (1 − pt(d)) − 2
1235
+ M
1236
+
1237
+ 0=i<j
1238
+ ijnpt(i)pt(j)
1239
+
1240
+
1241
+
1242
+ = 1
1243
+ n
1244
+ M
1245
+
1246
+ d=0
1247
+ d2pt(d) − 1
1248
+ n
1249
+
1250
+
1251
+
1252
+ M
1253
+
1254
+ d=0
1255
+ d2p2
1256
+ t (d) + 2
1257
+ M
1258
+
1259
+ 0=i<j
1260
+ ijpt(i)pt(j)
1261
+
1262
+
1263
+
1264
+ = 1
1265
+ n
1266
+ M
1267
+
1268
+ d=0
1269
+ d2pt(d) − 1
1270
+ n
1271
+ � M
1272
+
1273
+ d=0
1274
+ dpt(d)
1275
+ �2
1276
+ .
1277
+ (35)
1278
+ Eq. (34) becomes
1279
+ EX|T=t[R(t, a∗(πX))]
1280
+ =
1281
+ k
1282
+ ��
1283
+ 1 +
1284
+ n
1285
+ (α + n)2
1286
+ � M
1287
+
1288
+ d=0
1289
+ d2pt(d) +
1290
+ � 2αnc1
1291
+ (α + n)2 − 2αc1
1292
+ α + n
1293
+ � M
1294
+
1295
+ d=0
1296
+ dpt(d)
1297
+ +
1298
+ �n(n − 1)
1299
+ (α + n)2 −
1300
+ 2n
1301
+ α + n
1302
+ � � M
1303
+
1304
+ d=0
1305
+ dpt(d)
1306
+ �2
1307
+ +
1308
+ α2c2
1309
+ 1
1310
+ (α + n)2
1311
+
1312
+
1313
+ � .
1314
+ The final step is to take the expectation over all possible beliefs pt ∼
1315
+ 20
1316
+
1317
+ Dirichlet(α0, . . . , αM). Using the fact that
1318
+ ET∼π[pT (i)]
1319
+ =
1320
+ αi
1321
+ α ,
1322
+ ET∼π[pT (i)2]
1323
+ =
1324
+ Var[pT (i)] + ET [pT (i)]2
1325
+ =
1326
+ αi(α − αi)
1327
+ α2(α + 1) + α2
1328
+ i
1329
+ α2
1330
+ =
1331
+ αi(αi + 1)
1332
+ α(α + 1) ,
1333
+ ET∼π[pT (i)pT (j)]
1334
+ =
1335
+ Cov[pT (i), pT (j)] + ET [pT (i)]ET [pT (j)],
1336
+ i ̸= j,
1337
+ =
1338
+
1339
+ αiαj
1340
+ α2(α + 1) + αiαj
1341
+ α2
1342
+ =
1343
+ αiαj
1344
+ α(α + 1),
1345
+ and
1346
+ ET∼π
1347
+
1348
+
1349
+ � M
1350
+
1351
+ d=0
1352
+ dpT (d)
1353
+ �2�
1354
+ � = ET∼π
1355
+
1356
+
1357
+ M
1358
+
1359
+ d=0
1360
+ d2p2
1361
+ T (d) + 2
1362
+ M
1363
+
1364
+ 0=i<j
1365
+ ijpT (i)pT (j)
1366
+
1367
+
1368
+ =
1369
+ M
1370
+
1371
+ d=0
1372
+ d2ET∼π
1373
+
1374
+ p2
1375
+ T (d)
1376
+
1377
+ + 2
1378
+ M
1379
+
1380
+ 0=i<j
1381
+ ijET∼π [pT (i)pT (j)]
1382
+ =
1383
+ M
1384
+
1385
+ d=0
1386
+ d2 αd (αd + 1)
1387
+ α (α + 1) + 2
1388
+ M
1389
+
1390
+ 0=i<j
1391
+ ij
1392
+ αiαj
1393
+ α (α + 1)
1394
+ =
1395
+ 1
1396
+ α + 1
1397
+ M
1398
+
1399
+ d=0
1400
+ d2αd
1401
+ α
1402
+ +
1403
+ 1
1404
+ α (α + 1)
1405
+
1406
+
1407
+ M
1408
+
1409
+ d=0
1410
+ d2α2
1411
+ d + 2
1412
+ M
1413
+
1414
+ 0=i<j
1415
+ ijαiαj
1416
+
1417
+
1418
+ =
1419
+ c2
1420
+ α + 1 +
1421
+ 1
1422
+ α (α + 1)
1423
+ � M
1424
+
1425
+ d=0
1426
+ dαd
1427
+ �2
1428
+ =
1429
+ c2
1430
+ α + 1 + αc2
1431
+ 1
1432
+ α + 1,
1433
+ (36)
1434
+ 21
1435
+
1436
+ we obtain
1437
+ ET EX|T [R(T, a∗(πX))]
1438
+ =
1439
+ k
1440
+ ��
1441
+ 1 +
1442
+ n
1443
+ (α + n)2
1444
+ � M
1445
+
1446
+ d=0
1447
+ d2αd
1448
+ α
1449
+ +
1450
+ � 2αnc1
1451
+ (α + n)2 − 2αc1
1452
+ α + n
1453
+ � M
1454
+
1455
+ d=0
1456
+ dαd
1457
+ α
1458
+ +
1459
+ �n(n − 1)
1460
+ (α + n)2 −
1461
+ 2n
1462
+ α + n
1463
+ � �
1464
+ c2
1465
+ α + 1 + αc2
1466
+ 1
1467
+ α + 1
1468
+
1469
+ +
1470
+ α2c2
1471
+ 1
1472
+ (α + n)2
1473
+
1474
+ =
1475
+ k(c2 − c2
1476
+ 1) α(1 + α + n)
1477
+ (1 + α)(n + α).
1478
+ The value of n samples from the data distribution is therefore
1479
+ Vn(π)
1480
+ =
1481
+ k(c2 − c2
1482
+ 1) − k(c2 − c2
1483
+ 1) α(1 + α + n)
1484
+ (1 + α)(n + α)
1485
+ =
1486
+ kn(c2 − c2
1487
+ 1)
1488
+ (n + α)(1 + α).
1489
+ (37)
1490
+
1491
+ References
1492
+ [1] C. J. Adcock. An improved bayesian procedure for calculating sample
1493
+ sizes in multinomial sampling. Journal of the Royal Statistical Society.
1494
+ Series D (The Statistician), 42(2):91–95, 1993.
1495
+ [2] AE Ades, G Lu, and K Claxton. Expected value of sample information
1496
+ calculations in medical decision modeling. Medical Decision Making,
1497
+ 24(2):207–227, 2004.
1498
+ [3] J. Aitchison and S. M. Shen. Logistic-normal distributions: Some prop-
1499
+ erties and uses. Biometrika, 67(2):261–272, 1980.
1500
+ [4] Debarun Bhattacharjya, Jo Eidsvik, and Tapan Mukerji. The value of
1501
+ information in portfolio problems with dependent projects. Decision
1502
+ Analysis, 10(4):341–351, 2013.
1503
+ [5] J Eric Bickel. The relationship between perfect and imperfect informa-
1504
+ tion in a two-action risk-sensitive problem. Decision Analysis, 5(3):116–
1505
+ 128, 2008.
1506
+ [6] Jing Cao, J Jack Lee, and Susan Alber. Comparison of bayesian sample
1507
+ size criteria: Acc, alc, and woc.
1508
+ Journal of statistical planning and
1509
+ inference, 139(12):4111–4122, 2009.
1510
+ 22
1511
+
1512
+ [7] Li Chen and Erica L Plambeck.
1513
+ Dynamic inventory management
1514
+ with learning about the demand distribution and substitution probabil-
1515
+ ity. Manufacturing & Service Operations Management, 10(2):236–256,
1516
+ 2008.
1517
+ [8] Gary D. Eppen and Ananth. V. Iyer. Improved fashion buying with
1518
+ bayesian updates. Operations Research, 45(6):805–819, 1997.
1519
+ [9] Marshall Fisher, Kumar Rajaram, and Ananth Raman.
1520
+ Optimizing
1521
+ inventory replenishment of retail fashion products. Manufacturing &
1522
+ Service Operations Management, 3(3):230–241, 2001.
1523
+ [10] Adam J. Fleischhacker and Pak-Wing Fok. An entropy-based method-
1524
+ ology for valuation of demand uncertainty reduction. Decision Sciences,
1525
+ 46(6):1165–1198, 2015.
1526
+ [11] Yigal Gerchak and David Mossman.
1527
+ On the effect of demand ran-
1528
+ domness on inventories and costs. Operations Research, 40(4):804–807,
1529
+ 1992.
1530
+ [12] Gregory A. Godfrey and Warren B. Powell. An adaptive, distribution-
1531
+ free algorithm for the newsvendor problem with censored demands,
1532
+ with applications to inventory and distribution. Management Science,
1533
+ 47:1101–1112, 2001.
1534
+ [13] Robbie A Haertel, Kevin D Seppi, Eric K Ringger, and James L Carroll.
1535
+ Return on investment for active learning. In Proceedings of the NIPS
1536
+ Workshop on Cost-Sensitive Learning, volume 72, 2008.
1537
+ [14] Ronald W Hilton. The determinants of information value: Synthesizing
1538
+ some general results. Management Science, 27(1):57–64, 1981.
1539
+ [15] Ronald A Howard. Decision analysis: Perspectives on inference, de-
1540
+ cision, and experimentation. Proceedings of the IEEE, 58(5):632–643,
1541
+ 1970.
1542
+ [16] Jonathan Huang.
1543
+ Maximum likelihood estimation of dirichlet dis-
1544
+ tribution parameters.
1545
+ CMU Technique Report, 2005.
1546
+ http://
1547
+ jonathan-huang.org/research/dirichlet/dirichlet.pdfaccessed
1548
+ on 2017-09-19.
1549
+ [17] Woonghee Tim Huh,
1550
+ Retsef Levi,
1551
+ Paat Rusmevichientong,
1552
+ and
1553
+ James B. Orlin.
1554
+ Adaptive data-driven inventory control with cen-
1555
+ 23
1556
+
1557
+ sored demand based on kaplan-meier estimator. Operations Research,
1558
+ 59(4):929–941, 2011.
1559
+ [18] Christopher Jackson, Anne Presanis, Stefano Conti, and Daniela
1560
+ De Angelis.
1561
+ Value of information: Sensitivity analysis and research
1562
+ design in bayesian evidence synthesis. Journal of the American Statis-
1563
+ tical Association, 2019.
1564
+ [19] Hawre Jalal and Fernando Alarid-Escudero. A gaussian approximation
1565
+ approach for value of information analysis. Medical Decision Making,
1566
+ 38(2):174–188, 2018.
1567
+ [20] Jeffrey Keisler. Comparative static analysis of information value in a
1568
+ canonical decision problem.
1569
+ The Engineering Economist, 49(4):339–
1570
+ 349, 2004.
1571
+ [21] Jeffrey M Keisler. Additivity of information value in two-act linear loss
1572
+ decisions with normal priors. Risk Analysis: An International Journal,
1573
+ 25(2):351–359, 2005.
1574
+ [22] Jin Kyung Kwak and Srinagesh Gavirneni. Retailer policy, uncertainty
1575
+ reduction, and supply chain performance. International Journal of Pro-
1576
+ duction Economics, 132(2):271 – 278, 2011.
1577
+ [23] Retsef Levi, Georgia Perakis, and Joline Uichanco. The data-driven
1578
+ newsvendor problem: new bounds and insights. Operations Research,
1579
+ 63(6):1294–1306, 2015.
1580
+ [24] Retsef Levi, Robin O. Roundy, and David B. Shmoys.
1581
+ Provably
1582
+ Near-Optimal Sampling-Based Policies for Stochastic Inventory Control
1583
+ Models. Mathematics of Operations Research, 32(4):821–839, 2007.
1584
+ [25] Chi-Yuan Lin. Numerical techniques for evaluating sample information.
1585
+ Technometrics, 16(3):447–454, 1974.
1586
+ [26] Dennis V Lindley.
1587
+ The choice of sample size.
1588
+ Journal of the Royal
1589
+ Statistical Society: Series D (The Statistician), 46(2):129–138, 1997.
1590
+ [27] Abraham Mehrez. The effect of risk aversion on the expected value of
1591
+ perfect information. Operations research, 33(2):455–458, 1985.
1592
+ [28] Avraham Mehrez and Alan Stulman. Some aspects of the distributional
1593
+ properties of the expected value of perfect information (evpi). Journal
1594
+ of the Operational Research Society, 33(9):827–836, 1982.
1595
+ 24
1596
+
1597
+ [29] Joseph M. Milner and Panos Kouvelis. Order quantity and timing flexi-
1598
+ bility in supply chains: The role of demand characteristics. Management
1599
+ Science, 51(6):970–985, 2005.
1600
+ [30] Satoshi Morita, Peter F Thall, and Peter M¨uller.
1601
+ Determining the
1602
+ effective sample size of a parametric prior. Biometrics, 64(2):595–602,
1603
+ 2008.
1604
+ [31] Jeremy Muesing, Nisar Ahmed, Luke Burks, Michael Iuzzolino, and
1605
+ Danielle Albers Szafir. Fully bayesian human–machine data fusion for
1606
+ robust online dynamic target characterization. Journal of Aerospace
1607
+ Information Systems, 18(2):26–49, 2021.
1608
+ [32] T. Pham-Gia and N. Turkkan. Sample size determination in bayesian
1609
+ analysis. Journal of the Royal Statistical Society. Series D (The Statis-
1610
+ tician), 41(4):389–397, 1992.
1611
+ [33] Howard Raiffa and Stephen E Fienberg. The early statistical years:
1612
+ 1947-1967 a conversation with howard raiffa. Statistical Science, pages
1613
+ 136–149, 2008.
1614
+ [34] Howard Raiffa and Robert Schlaifer. Applied Statistical Decision The-
1615
+ ory. MIT press, Cambridge, Massachusetts, 1961.
1616
+ [35] Soroush Saghafian and Brian Tomlin. The newsvendor under demand
1617
+ ambiguity: Combining data with moment and tail information. Oper-
1618
+ ations Research, 64(1):167–185, 2016.
1619
+ [36] Mark Strong, Jeremy E. Oakley, Alan Brennan, and Penny Breeze. Es-
1620
+ timating the expected value of sample information using the probabilis-
1621
+ tic sensitivity analysis sample: A fast, nonparametric regression-based
1622
+ method. Medical Decision Making, 35(5):570–583, 2015.
1623
+ [37] Zhiyuan Wang,
1624
+ Zhiqiang Zheng,
1625
+ Wei Jiang,
1626
+ and Shaojie Tang.
1627
+ Blockchain-enabled data sharing in supply chains:
1628
+ Model, opera-
1629
+ tionalization, and tutorial. Production and Operations Management,
1630
+ 30(7):1965–1985, 2021.
1631
+ [38] Wei Xiang and Wenxing Zhou.
1632
+ Optimal sample size determination
1633
+ based on bayesian reliability and value of information. In 13th Interna-
1634
+ tional Conference on Applications of Statistics and Probability in Civil
1635
+ Engineering, 2019.
1636
+ 25
1637
+
1638
+ [39] Phillip M Yelland. Bayesian forecasting for low-count time series using
1639
+ state-space models: An empirical evaluation for inventory management.
1640
+ International Journal of Production Economics, 118(1):95–103, 2009.
1641
+ [40] Jinfeng Yue, Bintong Chen, and Min-Chiang Wang. Expected value
1642
+ of distribution information for the newsvendor problem.
1643
+ Operations
1644
+ Research, 54(6):1128–1136, 2006.
1645
+ 26
1646
+
0NAyT4oBgHgl3EQf0_na/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
19E3T4oBgHgl3EQfngq-/content/tmp_files/2301.04626v1.pdf.txt ADDED
@@ -0,0 +1,1251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Deep Axial Hypercomplex Networks
2
+ Nazmul Shahadat, Anthony S. Maida
3
+ University of Louisiana at Lafayette
4
+ Lafayette LA 70504, USA
5
+ nazmul.ruet@gmail.com, maida@louisiana.edu
6
+ Abstract
7
+ Over the past decade, deep hypercomplex-inspired
8
+ networks have enhanced feature extraction for image
9
+ classification by enabling weight sharing across input
10
+ channels.
11
+ Recent works make it possible to improve
12
+ representational capabilities by using hypercomplex-
13
+ inspired networks which consume high computational
14
+ costs.
15
+ This paper reduces this cost by factorizing
16
+ a quaternion 2D convolutional module into two con-
17
+ secutive vectormap 1D convolutional modules.
18
+ Also,
19
+ we use 5D parameterized hypercomplex multiplication
20
+ based fully connected layers. Incorporating both yields
21
+ our proposed hypercomplex network, a novel architec-
22
+ ture that can be assembled to construct deep axial-
23
+ hypercomplex networks (DANs) for image classifica-
24
+ tions. We conduct experiments on CIFAR benchmarks,
25
+ SVHN, and Tiny ImageNet datasets and achieve bet-
26
+ ter performance with fewer trainable parameters and
27
+ FLOPS. Our proposed model achieves almost 2% higher
28
+ performance for CIFAR and SVHN datasets, and more
29
+ than 3% for the ImageNet-Tiny dataset and takes six
30
+ times fewer parameters than the real-valued ResNets.
31
+ Also, it shows state-of-the-art performance on CIFAR
32
+ benchmarks in hypercomplex space.
33
+ 1. Introduction
34
+ Convolutional neural networks (CNNs) and hyper-
35
+ complex CNNs (HCNNs) for image classification form
36
+ a hierarchical design where different layers extract dif-
37
+ ferent levels of feature representation.
38
+ CNNs have
39
+ shown significant success in recent decades [2, 9]. In
40
+ vision tasks, these CNN-based feature extraction de-
41
+ signs can be improved in regard to working with multi-
42
+ dimensional data. To enhance the CNNs ability, HC-
43
+ NNs have been used which treat the multi-dimensional
44
+ data as a cohesive entity by applying cross-channel
45
+ weight sharing to discover cross-channel relationships
46
+ [4, 5, 14, 15]. Also, implementations in hypercomplex
47
+ space provide more advantages [1, 3, 7, 13]. It has also
48
+ been shown that the HCNNs could create better output
49
+ representations [13,16,17].
50
+ Recently, HCNNs with various dimensions like 2D
51
+ HCNNs [21], 4D HCNNs [4, 14, 15], 8D HCNNs [20],
52
+ or generalized HCNNs [5], have been studied and have
53
+ hypercomplex properties. The reason behind the suc-
54
+ cess of HCNNs is that they capture the cross-channel
55
+ relationships [4,5,14,15,17]. Among them, quaternion
56
+ networks have a set of algebra operations and they have
57
+ outperformed than the other HCNNs. Stacking quater-
58
+ nion convolutional coherent layers have achieved better
59
+ representational feature maps and have shown promis-
60
+ ing results in vision tasks [4,15,17]. These networks are
61
+ cost-effective compared to real-valued CNNs and fully
62
+ connected networks. But still, they are very expensive
63
+ for large inputs like vision tasks.
64
+ This work uses an axial hypercomplex network that:
65
+ 1) handles multidimensional inputs; 2) applies weight
66
+ sharing across input channels; 3) captures cross-channel
67
+ correlations; 4) reduces computational costs; and 5) in-
68
+ creases validation accuracy performance for image clas-
69
+ sification datasets.
70
+ The main idea of this work is to
71
+ decompose hypercomplex 2D convolutional operation
72
+ into two consecutive vectormap 1D convolutional oper-
73
+ ations . By splitting 2D spatial convolution operation
74
+ into height-axis and width-axis spatial convolution, it
75
+ enables the model to reduce cost once again. Addition-
76
+ ally, we apply a quaternion-based stem layer, and pa-
77
+ rameterized hypercomplex multiplication (PHM) based
78
+ fully connected layer to get better representation and
79
+ better generalization performance.
80
+ This paper conducts extensive experiments that show
81
+ the effectiveness of our novel axial hypercomplex net-
82
+ works on four image classification datasets. Our novel
83
+ contribution is a new model that factorizes the two-
84
+ dimensional spatial hypercomplex convolutional oper-
85
+ ation into two one-dimensional operations along the
86
+ height-axis and width-axis sequentially. Our contribu-
87
+ tions are:
88
+ • Replacing the spatial 3 × 3 QCNN in the bottle-
89
+ arXiv:2301.04626v1 [cs.CV] 11 Jan 2023
90
+
91
+ Figure 1. Proposed axial-hypercomplex network with PHM-based fully-connected layer in backend. “AHNN” stands for axial-
92
+ hypercomplex neural network bottleneck block which is described in Figure 2. Here, Qin = Qr + Qw + Qx + Qy + Qz,
93
+ H = Hr + Hw + Hx + Hy + Hz, and Qout = Qro + Qwo + Qxo + Qyo + Qzo are the input, hypercomplex parameterized
94
+ weight, and output, respectively. For the calculation of H see the “PHM Layer” section.
95
+ neck block of quaternion ResNets using two VC-
96
+ NNs and showing the effectiveness of the proposed
97
+ networks.
98
+ • Applying QCNN in the stem layer (the first layer of
99
+ the network), resulting in a quaternion-stem model.
100
+ • Like QPHM [16], applying PHM-based dense layer
101
+ in the backend of the network.
102
+ This proposed axial hypercomplex ResNets outper-
103
+ formed the baseline networks for classification datasets
104
+ which is shown in Tables 2, 3, and 4. Our experiments
105
+ show that the proposed model achieves state-of-the-art
106
+ results with far fewer trainable parameters, and FLOPS
107
+ for CIFAR benchmarks in hypercomplex space.
108
+ 2. Background and Related Work
109
+ 2.1. Quaternion Convolution
110
+ The deep quaternion CNN extends of complex CNNs
111
+ [18]. This section explains cross channel weight sharing.
112
+ [4] and [15] extended the principles of quaternion con-
113
+ volution operations, and weight initialization. Quater-
114
+ nion number system is formed as, Q = r + ix + jy +
115
+ kz ; r, x, y, z ∈ R where, r, x, y, and z are real values
116
+ and i, j, and k are imaginary. Quaternion convolution
117
+ between quaternion filter matrix F and quaternion input
118
+ vector M, is defined as [4]:
119
+ M ⊛ F = (Or, Oi, Oj, Ok)
120
+ = (Mr ∗ Fr − Mi ∗ Fi − Mj ∗ Fj − Mk ∗ Fk,
121
+ Mi ∗ Fr + Mr ∗ Fi + Mj ∗ Fk − Mk ∗ Fj,
122
+ Mj ∗ Fr + Mr ∗ Fj + Mk ∗ Fi − Mi ∗ Fk,
123
+ Mk ∗ Fr + Mr ∗ Fk + Mi ∗ Fj − Mj ∗ Fi)
124
+ (1)
125
+ where, M ⊛ F, and all others are quaternion numbers.
126
+ Or is the real part, and Oi, Oj, and Ok are the imag-
127
+ inary parts. Although there are 16 real-valued convo-
128
+ lutions in Equation 1, there are only four kernels that
129
+ are reused. The weight sharing happens this way [14]
130
+ which forces the model to learn cross-channel interre-
131
+ lationships. According to the quaternion definition, a
132
+ quaternion layer can accept four or m numbers of input
133
+ channels, where m is divisible by four. To process m
134
+ input channels (m ≥ 4), m/4 number of independent
135
+ quaternion convolution modules is required. Also, there
136
+ are m/4 weight sets where each module has its own
137
+ weight sets. Cross-channel weight sharing allows dis-
138
+ covering of cross-channel input correlations. Our weight
139
+ initialization was the same as [4].
140
+ 2.2. Vectormap Convolution
141
+ We explain 3D generalized hypercomplex networks
142
+ or VCNNs as VCNNs are used in our proposed mod-
143
+ els. The VCNN is more flexible as it doesn’t require 4D.
144
+
145
+ Parameterized
146
+ STAGE 1
147
+ STAGE 2
148
+ STAGE 3
149
+ STAGE 4
150
+ Weight
151
+ Input
152
+ Horse
153
+ Cat
154
+ 224x224
155
+ Classification
156
+ PHM
157
+ Stem Layer
158
+ based FC Layer
159
+ QCNN
160
+ 64 Filters
161
+ 1st Layer
162
+ 2nd Layer
163
+ 3rd Layer
164
+ 4rth Layer
165
+ AHNN
166
+ AHNN
167
+ AHNN
168
+ Flatten
169
+ Filter size 7
170
+ AHNN
171
+ 128 Filters
172
+ 256 Filters
173
+ 512 Filters
174
+ with stride 2
175
+ 64 Filters
176
+ Layer
177
+ max-pooling
178
+ Stride 2
179
+ Stride 2
180
+ Stride 2
181
+ Stride 1
182
+ 5-dimensional PHM laye
183
+ VPHM
184
+ TimerFigure 2. AHNN bottleneck block used in our proposed axial-hypercomplex networks. “bn”, “quat”, and “VCNN” stand for batch
185
+ normalization, quaternion CNN, and vectormap CNN, respectively.
186
+ However, still using cross channel weight sharing this is
187
+ seen in 3 × 3 matrix used in Equation 2, only three fil-
188
+ ters A, B, and C are used. The Vectormap convolution
189
+ operation is defined as:
190
+
191
+
192
+ R(M ∗ F)
193
+ I(M ∗ F)
194
+ J (M ∗ F)
195
+
196
+ � = L ⊙
197
+
198
+
199
+ A
200
+ B
201
+ C
202
+ C
203
+ A
204
+ B
205
+ B
206
+ C
207
+ A
208
+
209
+ � ∗
210
+
211
+
212
+ x
213
+ y
214
+ z
215
+
216
+
217
+ (2)
218
+ where, A, B, and C are real-valued kernels, x, y, and
219
+ z being real-valued vectors, and L is a learnable matrix,
220
+ L ∈ RD3×D3; where D3 stands for 3-dimensional input
221
+ channels. The initial value of this matrix L is defined as:
222
+ L =
223
+
224
+
225
+ 1
226
+ 1
227
+ 1
228
+ -1
229
+ 1
230
+ 1
231
+ -1
232
+ 1
233
+ 1
234
+
235
+
236
+ (3)
237
+ Our weight initialization follows [5].
238
+ 2.3. PHM Layer
239
+ Parameterized hypercomplex multiplication is an-
240
+ other form of generalized hypercomplex network, ex-
241
+ plained in [22].
242
+ As we use this PHM layer only in
243
+ the fully connected (FC) layer, our explanation is re-
244
+ stricted to this PHM-based dense layer. It is defined as,
245
+ y = Hx + b, where H ∈ Rk×d represents the PHM
246
+ layer and it is calculated as, H = �n
247
+ i=1 Ii ⊗ Ai, where
248
+ Ii ∈ Rn×n and Ai ∈ Rk/n×d/n are learnable parameter
249
+ matrices and i = 1 . . . n (n = 4 or 5). These matrices
250
+ can be reused which leads to parameter reduction. Also,
251
+ the ⊗ represents the Kronecker product. The flattened
252
+ layer which is the output of the CNN network is used as
253
+ an input to the PHM FC layer. These inputs are split as,
254
+ Qin = Qr + Qw + Qx + Qy + Qz and the outputs are
255
+ merged into Qout as, Qout = Qro+Qwo+Qxo+Qyo+
256
+ Qzo for 5D hypercomplex. The 4D hypercomplex pa-
257
+ rameter matrix is discussed in [22] which expresses the
258
+ Hamiltonian product, and the 5D hypercomplex param-
259
+ eter matrix of PHM operation is explained in [16]. This
260
+ 5D parameter matrix is used to construct a 5D PHM FC
261
+ layer which preserves all properties of the PHM layer
262
+ and hypercomplex networks. This work uses 5D PHM
263
+ layer.
264
+ 3. Proposed Axial Hypercomplex Networks
265
+ Complex convolutional neural networks (CCNNs),
266
+ QCNNs, Octonions convolutional neural networks (OC-
267
+ NNs), VCNNs, and PHM are the versions of HCNNs
268
+ that provide all advantages of HCNNs like weight shar-
269
+ ing across input channels, and the ability to discover
270
+ cross channel correlations. These HCNNs perform bet-
271
+ ter with fewer trainable parameters for vision applica-
272
+ tions. But, they are still computationally expensive. For
273
+ vision tasks, these HCNNs take O(N 2) resources for
274
+ an image of length N where N is the flattened pixel
275
+ set. For a 2D image of height h and width w, where
276
+ N
277
+ = hw, and h = w, the computational cost is
278
+ O((hw)2) = O(h2w2) = O(h4).
279
+ This
280
+ section
281
+ describes
282
+ our
283
+ proposed
284
+ axial-
285
+ hypercomplex model in Figures 1 and 2 to reduce
286
+ the computational cost. Axial networks were first used
287
+ in [8, 19].
288
+ To implement our proposed model, we
289
+ followed the assumption that images are approximately
290
+ square where the pixel count of h and w are the same,
291
+ and both are much less than the pixel count of hw [19].
292
+ To translate a quaternion convolutional bottleneck block
293
+ to an axial-hypercomplex bottleneck block, we replace
294
+ the 3 × 3 spatial quaternion convolutional operation
295
+ by two axial vectormap convolutional neural network
296
+ (VCNN) layers. These layers are applied to the height
297
+
298
+ m
299
+ -.-.-.-.-.-.-.
300
+ b
301
+ rel
302
+ bn
303
+ 3x1 VCNN
304
+ Width-Axis
305
+ enb
306
+ nb
307
+ relu
308
+ 1
309
+ X
310
+ X
311
+ X
312
+ S
313
+ H
314
+ -.
315
+ S
316
+ channel
317
+ s
318
+ sLayer
319
+ Output
320
+ size
321
+ Deep Quaternion
322
+ ResNet
323
+ Vectormap
324
+ ResNet
325
+ QPHM
326
+ Axial
327
+ Hypercomplex
328
+ Stem
329
+ 32x32
330
+ 3x3Q, 120, std=1
331
+ 3x3V, 120, std=1
332
+ 3x3Q, 120, std=1
333
+ 3x3Q, 120, std=1
334
+ Bottleneck
335
+ group 1
336
+ 32x32
337
+
338
+
339
+ 1x1Q, 120
340
+ 3x3Q, 120
341
+ 1x1Q, 480
342
+
343
+ �×3
344
+
345
+
346
+ 1x1V, 120
347
+ 3x3V, 120
348
+ 1x1V, 480
349
+
350
+ �×3
351
+
352
+
353
+ 1x1QP, 120
354
+ 3x3QP, 120
355
+ 1x1QP, 480
356
+
357
+ �×3
358
+
359
+ ���
360
+ 1x1Q, 120
361
+ 3x1AV, 120
362
+ 1x3AV, 120
363
+ 1x1Q, 480
364
+
365
+ ���×3
366
+ Bottleneck
367
+ group 2
368
+ 16x16
369
+
370
+
371
+ 1x1Q, 240
372
+ 3x3Q, 240
373
+ 1x1Q, 960
374
+
375
+ �×4
376
+
377
+
378
+ 1x1V, 240
379
+ 3x3V, 240
380
+ 1x1V, 960
381
+
382
+ �×4
383
+
384
+
385
+ 1x1QP, 240
386
+ 3x3QP, 240
387
+ 1x1QP, 960
388
+
389
+ �×4
390
+
391
+ ���
392
+ 1x1Q, 240
393
+ 3x1AV, 240
394
+ 1x3AV, 240
395
+ 1x1Q, 960
396
+
397
+ ���×4
398
+ Bottleneck
399
+ group 3
400
+ 8x8
401
+
402
+
403
+ 1x1Q, 480
404
+ 3x3Q, 480
405
+ 1x1Q, 1920
406
+
407
+ �×6
408
+
409
+
410
+ 1x1V, 480
411
+ 3x3V, 480
412
+ 1x1V, 1920
413
+
414
+ �×6
415
+
416
+
417
+ 1x1QP, 480
418
+ 3x3QP, 480
419
+ 1x1QP, 1920
420
+
421
+ �×6
422
+
423
+ ���
424
+ 1x1Q, 480
425
+ 3x1AV, 480
426
+ 1x3AV, 480
427
+ 1x1Q, 1920
428
+
429
+ ���×6
430
+ Bottleneck
431
+ group 4
432
+ 4x4
433
+
434
+
435
+ 1x1Q, 960
436
+ 3x3Q, 960
437
+ 1x1Q, 3840
438
+
439
+ �×3
440
+
441
+
442
+ 1x1V, 960
443
+ 3x3V, 960
444
+ 1x1V, 3840
445
+
446
+ �×3
447
+
448
+
449
+ 1x1QP, 960
450
+ 3x3QP, 960
451
+ 1x1QP, 3840
452
+
453
+ �×3
454
+
455
+ ���
456
+ 1x1Q, 960
457
+ 3x1AV, 960
458
+ 1x3AV, 960
459
+ 1x1Q, 3840
460
+
461
+ ���×3
462
+ Pooling layer
463
+ 1x1x100
464
+ global average-pool, 100 outputs
465
+ Output
466
+ 1x1x100
467
+ fully connected layer, softmax
468
+ 5D PHM layer
469
+ Table 1. The 50-layer architectures tested on CIFAR-100: quaternion ResNet [4, 5], vectormap ResNet [5], QPHM [16], and our
470
+ proposed axial-hypercomplex networks. Input is a 32x32x3 color image for CIFAR benchmarks. The number of stacked bottleneck
471
+ modules is specified by multipliers. “Q”, “V”, “QP”, “AV” and “std” denote quaternion convolution, 3D vectormap convolution,
472
+ QPHM (quaternion networks with 4D PHM layer), axial vectormap convolution, and stride correspondingly. Integers (e.g., 120,
473
+ 240) denote the number of output channels. PHM layer stands for parameterized hypercomplex multiplication layer. This work
474
+ uses 5D PHM based FC layer.
475
+ axis (3 channels 3x1 VCNN layer) and width axis (3
476
+ channels 1x3 VCNN layer) sequentially. The two 1 × 1
477
+ quaternion convolutional layers remain unchanged
478
+ like the original QCNNs [4]. The 1 × 1 QCNNs are
479
+ responsible to reduce and then increase the number of
480
+ channels. This forms our proposed axial-hypercomplex
481
+ bottleneck block seen in Figure 2.
482
+ This block is
483
+ stacked multiple times to construct axial-hypercomplex
484
+ ResNets.
485
+ Axial-hypercomplex models only work on one di-
486
+ mension at a time but the input images are 2-
487
+ dimensional. For two-dimensional vision tasks, a square
488
+ 2D input where h = w, so w2 = N, where N is the se-
489
+ quence length of the flattened pixel set, is split into two
490
+ 1D vectors. The 3-channel VCNN operation is first ap-
491
+ plied along the 1D input image region of length h and
492
+ then applied along the 1D input image region of length
493
+ w. These two 1D operations finally merged together re-
494
+ duces cost to O(h · h2) = O(h3) from the HCNNs cost
495
+ of O(h4).
496
+ Each quaternion convolution accepts four channels
497
+ of input and produces four channels of output. Hence,
498
+ the required number of 1 × 1 quaternion conv2d mod-
499
+ ules equals the number of input channels divided by
500
+ four. The set of output channels of down-sampled 1 × 1
501
+ quaternion is merged into input to the axial VCNN mod-
502
+ ules, and the output channels of axial VCNN modules
503
+ are split into groups of four again for 1 × 1 up-sampled
504
+ quaternion conv2d layer [4,17]. One quaternion 2D con-
505
+ volution is applied to each group of four channels and
506
+ one vectormap 2D convolution is applied to each group
507
+ of three channels. Like vectormap, each axial vectormap
508
+ module takes three input channels. Thus, the weight-
509
+ sharing is compartmentalized into groups of four input
510
+ channels and then groups of three input channels.
511
+ For better representation, a quaternion convolution
512
+ layer is also used in the stem layer (first layer of the net-
513
+ work) as a quaternion-based frontend layer and the fully-
514
+ connected dense layer as a PHM-based backend layer of
515
+ deep axial-hypercomplex networks (DANs). Figure 1
516
+ illustrates our proposed axial-hypercomplex network ar-
517
+ chitecture.
518
+ 4. Experiment
519
+ We conduct an extensive experiment on four classi-
520
+ fication datasets to analyze the effectiveness of our pro-
521
+ posed axial-hypercomplex model. As QCNNs, VCNNs,
522
+ residual networks (ResNets), QPHM [16], and VPHM
523
+ [16] all are performed 2D spatial convolution operation,
524
+
525
+ Model Name
526
+ Layers
527
+ Dataset
528
+ Params
529
+ FLOPS
530
+ Latency
531
+ Validation
532
+ Accuracy
533
+ ResNet [6]
534
+ 40.9M
535
+ 2.56G
536
+ 0.86ms
537
+ 94.68
538
+ ResNet-with-QPHM [16]
539
+ 40.8M
540
+ 2.55G
541
+ 0.64ms
542
+ 95.32
543
+ Quaternion [4]
544
+ 10.2M
545
+ 1.11G
546
+ 0.65ms
547
+ 94.89
548
+ Vectormap [5]
549
+ 26
550
+ CIFAR10
551
+ 13.6M
552
+ 1.09G
553
+ 0.65ms
554
+ 94.76
555
+ QPHM [16]
556
+ 10.2M
557
+ 1.10G
558
+ 0.64ms
559
+ 95.26
560
+ VPHM [16]
561
+ 13.6M
562
+ 1.08G
563
+ 0.67ms
564
+ 95.15
565
+ Axial-Hypercomplex
566
+ 6.2M
567
+ 1.06G
568
+ 0.68ms
569
+ 95.91-95.85
570
+ ResNet [6]
571
+ 57.8M
572
+ 3.31G
573
+ 1.08ms
574
+ 94.95
575
+ ResNet-with-QPHM [16]
576
+ 57.7M
577
+ 3.31G
578
+ 0.81ms
579
+ 95.80
580
+ Quaternion [4]
581
+ 14.5M
582
+ 1.47G
583
+ 0.82ms
584
+ 95.33
585
+ Vectormap [5]
586
+ 35
587
+ CIFAR10
588
+ 19.3M
589
+ 1.45G
590
+ 0.84ms
591
+ 95.06
592
+ QPHM [16]
593
+ 14.5M
594
+ 1.46G
595
+ 0.79ms
596
+ 95.55
597
+ VPHM [16]
598
+ 19.3M
599
+ 1.44G
600
+ 0.82ms
601
+ 95.60
602
+ Axial-Hypercomplex
603
+ 9.2M
604
+ 1.36G
605
+ 0.84ms
606
+ 96.49-96.45
607
+ ResNet [6]
608
+ 82.5M
609
+ 4.57G
610
+ 1.32ms
611
+ 94.08
612
+ ResNet-with-QPHM [16]
613
+ 82.5M
614
+ 4.57G
615
+ 0.81ms
616
+ 95.86
617
+ Quaternion [4]
618
+ 21.09M
619
+ 1.93G
620
+ 1.06ms
621
+ 95.42
622
+ Vectormap [5]
623
+ 50
624
+ CIFAR10
625
+ 27.6M
626
+ 1.93G
627
+ 1.13ms
628
+ 95.37
629
+ QPHM [16]
630
+ 20.7M
631
+ 1.92G
632
+ 1.06ms
633
+ 95.75
634
+ VPHM [16]
635
+ 27.5M
636
+ 1.92G
637
+ 1.08ms
638
+ 95.76
639
+ Axial-Hypercomplex
640
+ 13.6M
641
+ 1.75G
642
+ 1.09ms
643
+ 96.79-96.71
644
+ ResNet [6]
645
+ 41.2M
646
+ 2.56G
647
+ 0.89ms
648
+ 78.21
649
+ ResNet-with-QPHM [16]
650
+ 40.9M
651
+ 2.56G
652
+ 0.64ms
653
+ 79.14
654
+ Quaternion [4]
655
+ 10.6M
656
+ 1.15G
657
+ 0.64ms
658
+ 77.65
659
+ Vectormap [5]
660
+ 26
661
+ CIFAR100
662
+ 13.6M
663
+ 1.15G
664
+ 0.64ms
665
+ 77.65
666
+ QPHM [16]
667
+ 10.3M
668
+ 1.11G
669
+ 0.65ms
670
+ 78.15
671
+ VPHM [16]
672
+ 13.7M
673
+ 1.09G
674
+ 0.66ms
675
+ 78.14
676
+ Axial-Hypercomplex
677
+ 6.2M
678
+ 1.06G
679
+ 0.69ms
680
+ 79.42-79.24
681
+ ResNet [6]
682
+ 58.1M
683
+ 3.31G
684
+ 1.07ms
685
+ 78.72
686
+ ResNet-with-QPHM [16]
687
+ 57.8M
688
+ 3.31G
689
+ 0.81ms
690
+ 79.65
691
+ Quaternion [4]
692
+ 14.5M
693
+ 1.51G
694
+ 0.81ms
695
+ 78.96
696
+ Vectormap [5]
697
+ 35
698
+ CIFAR100
699
+ 19.3M
700
+ 1.48G
701
+ 0.84ms
702
+ 79.52
703
+ QPHM [16]
704
+ 14.5M
705
+ 1.47G
706
+ 0.82ms
707
+ 78.46
708
+ VPHM [16]
709
+ 19.6M
710
+ 1.45G
711
+ 0.82ms
712
+ 79.86
713
+ Axial-Hypercomplex
714
+ 9.2M
715
+ 1.36G
716
+ 0.85ms
717
+ 79.93-79.63
718
+ ResNet [6]
719
+ 82.9M
720
+ 4.57G
721
+ 1.36ms
722
+ 78.95
723
+ ResNet-with-QPHM [16]
724
+ 82.6M
725
+ 4.57G
726
+ 1.09ms
727
+ 79.89
728
+ Quaternion [4]
729
+ 21.09M
730
+ 1.96G
731
+ 1.06ms
732
+ 79.17
733
+ Vectormap [5]
734
+ 50
735
+ CIFAR100
736
+ 27.6M
737
+ 1.93G
738
+ 1.13ms
739
+ 79.39
740
+ QPHM [16]
741
+ 20.7M
742
+ 1.93G
743
+ 1.05ms
744
+ 78.22
745
+ VPHM [16]
746
+ 27.5M
747
+ 1.92G
748
+ 1.08ms
749
+ 79.49
750
+ Axial-Hypercomplex
751
+ 13.6M
752
+ 1.75G
753
+ 1.09ms
754
+ 80.81-80.75
755
+ Table 2. Image classification performance on the CIFAR benchmarks for 26, 35, and 50-layer architectures. Here, QPHM, and
756
+ VPHM define the quaternion networks with PHM FC layer, and vectormap networks with the PHM FC layer, respectively.
757
+ therefore we compare our proposed axial hypercomplex
758
+ networks performance with the above-mentioned base-
759
+ line models. Among them, all models perform Hamilto-
760
+ nian products like our proposed model except ResNets.
761
+ 4.1. Method
762
+ We conducted our experiments by using five-
763
+ dimensional PHM dense layer in the backend of the
764
+ network, quaternion network at the beginning of the
765
+
766
+ Model Name
767
+ Layers
768
+ Params
769
+ FLOPS
770
+ Latency
771
+ Validation Accuracy
772
+ ResNet [6]
773
+ 40.9M
774
+ 2.56G
775
+ 0.82ms
776
+ 96.04
777
+ ResNet-with-QPHM [16]
778
+ 40.8M
779
+ 2.56G
780
+ 0.62ms
781
+ 96.64
782
+ Quaternion [4]
783
+ 10.2M
784
+ 1.11G
785
+ 0.66ms
786
+ 95.88
787
+ Vectormap [5]
788
+ 26
789
+ 13.6M
790
+ 1.10G
791
+ 0.66ms
792
+ 95.93
793
+ QPHM [16]
794
+ 10.2M
795
+ 1.10G
796
+ 0.62ms
797
+ 95.97
798
+ VPHM [16]
799
+ 13.6M
800
+ 1.08G
801
+ 0.64ms
802
+ 96.24
803
+ Axial-Hypercomplex
804
+ 6.2M
805
+ 1.06G
806
+ 0.69ms
807
+ 97.21-97.05
808
+ ResNet [6]
809
+ 57.8M
810
+ 3.31G
811
+ 0.98ms
812
+ 95.74
813
+ ResNet-with-QPHM [16]
814
+ 57.7M
815
+ 3.31G
816
+ 0.79ms
817
+ 96.22
818
+ Quaternion [4]
819
+ 14.5M
820
+ 1.47G
821
+ 0.84ms
822
+ 95.95
823
+ Vectormap [5]
824
+ 35
825
+ 19.5M
826
+ 1.45G
827
+ 0.84ms
828
+ 95.97
829
+ QPHM [16]
830
+ 14.5M
831
+ 1.45G
832
+ 0.82ms
833
+ 95.99
834
+ VPHM [16]
835
+ 19.3M
836
+ 1.44G
837
+ 0.82ms
838
+ 96.34
839
+ Axial-Hypercomplex
840
+ 9.2M
841
+ 1.36G
842
+ 0.85ms
843
+ 97.25-96.90
844
+ ResNet [6]
845
+ 82.5M
846
+ 4.57G
847
+ 1.19ms
848
+ 95.76
849
+ ResNet-with-QPHM [16]
850
+ 82.5M
851
+ 4.57G
852
+ 1.04ms
853
+ 96.78
854
+ Quaternion [4]
855
+ 20.7M
856
+ 1.94G
857
+ 1.04ms
858
+ 96.24
859
+ Vectormap [5]
860
+ 50
861
+ 27.6M
862
+ 1.93G
863
+ 1.11ms
864
+ 96.39
865
+ QPHM [16]
866
+ 20.7M
867
+ 1.93G
868
+ 1.04ms
869
+ 96.46
870
+ VPHM [16]
871
+ 27.5M
872
+ 1.92G
873
+ 1.09ms
874
+ 96.49
875
+ Axial-Hypercomplex
876
+ 13.6M
877
+ 1.75G
878
+ 1.11ms
879
+ 97.47-97.25
880
+ Table 3. Image classification performance on the SVHN benchmarks for 26, 35, and 50-layer architectures. Here, QPHM, and
881
+ VPHM define the quaternion networks with PHM FC layer, and vectormap networks with PHM FC layer, respectively.
882
+ Model Name
883
+ Layers
884
+ Params
885
+ FLOPS
886
+ Latency
887
+ Validation Accuracy
888
+ ResNet [6]
889
+ 41.6M
890
+ 10.2G
891
+ 3.06ms
892
+ 57.21
893
+ ResNet-with-QPHM [16]
894
+ 41M
895
+ 2.56G
896
+ 2.31ms
897
+ 57.84
898
+ Quaternion [4]
899
+ 11.02M
900
+ 4.54G
901
+ 2.48ms
902
+ 53.84
903
+ Vectormap [5]
904
+ 26
905
+ 14.4M
906
+ 4.56G
907
+ 2.88ms
908
+ 56.15
909
+ QPHM [16]
910
+ 10.4M
911
+ 1.11G
912
+ 2.31ms
913
+ 54.02
914
+ VPHM [16]
915
+ 13.8M
916
+ 4.44G
917
+ 3.27ms
918
+ 53.11
919
+ Axial-Hypercomplex
920
+ 6.3M
921
+ 1.06G
922
+ 2.49ms
923
+ 58.56-58.06
924
+ ResNet [6]
925
+ 58.5M
926
+ 13.2G
927
+ 3.21ms
928
+ 57.80
929
+ ResNet-with-QPHM [16]
930
+ 57.9M
931
+ 3.31G
932
+ 2.85ms
933
+ 59
934
+ Quaternion [4]
935
+ 15.2M
936
+ 5.98G
937
+ 3.52ms
938
+ 54.53
939
+ Vectormap [5]
940
+ 35
941
+ 20.07M
942
+ 5.98G
943
+ 3.76ms
944
+ 55.99
945
+ QPHM [16]
946
+ 14.6M
947
+ 1.47G
948
+ 2.88ms
949
+ 56.42
950
+ VPHM [16]
951
+ 19.4M
952
+ 5.88G
953
+ 4.08ms
954
+ 56.10
955
+ Axial-Hypercomplex
956
+ 9.3M
957
+ 1.36G
958
+ 2.97ms
959
+ 60.06-59.87
960
+ ResNet [6]
961
+ 83.2M
962
+ 18.2G
963
+ 3.77ms
964
+ 59.06
965
+ ResNet-with-QPHM [16]
966
+ 82.6M
967
+ 4.57G
968
+ 3.66ms
969
+ 60.30
970
+ Quaternion [4]
971
+ 21.4M
972
+ 7.87G
973
+ 4.14ms
974
+ 56.63
975
+ Vectormap [5]
976
+ 50
977
+ 28.3M
978
+ 7.87G
979
+ 4.34ms
980
+ 57.52
981
+ QPHM [16]
982
+ 20.8M
983
+ 1.93G
984
+ 3.88ms
985
+ 59.42
986
+ VPHM [16]
987
+ 27.7M
988
+ 7.75G
989
+ 4.51ms
990
+ 58.96
991
+ Axial-Hypercomplex
992
+ 13.7M
993
+ 1.75G
994
+ 3.93ms
995
+ 62.73-62.07
996
+ Table 4. Image classification performance on the Tiny ImageNet benchmarks for 26, 35, and 50-layer architectures. Here, QPHM,
997
+ and VPHM define the quaternion networks with PHM FC layer, and vectormap networks with PHM FC layer, respectively.
998
+ network, and axial-hypercomplex residual bottleneck
999
+ block on CIFAR benchmark datasets [10], Street View
1000
+ House Numbers (SVHN) [12], and Tiny ImageNet [11]
1001
+ datasets.
1002
+
1003
+ The models we tested to compare with our proposed
1004
+ model, are: the standard DCNNs [6], the DQNNs [4],
1005
+ the axial-ResNet with QPHM [16], QPHM [16], VPHM
1006
+ [16], and our proposed method. CIFAR-10 and CIFAR-
1007
+ 100 datasets consist of 60,000 color images of size 32
1008
+ × 32 pixels. These datasets fall into 10 and 100 dis-
1009
+ tinct classes and are split into a training set with 50,000
1010
+ images and a test set with 10,000 images. We perform
1011
+ standard data augmentation schemes for these datasets
1012
+ like [4–6,16]. Both datasets were normalized using per-
1013
+ channel mean and standard deviation. We perform hori-
1014
+ zontal flips and take random crops from images padded
1015
+ by 4 pixels on each side to obtain a 40 × 40 pixel image,
1016
+ then a 32 × 32 crop is randomly extracted.
1017
+ SVHN contains about 600,000 digit images [12]. For
1018
+ experiments on SVHN we don’t do any image pre-
1019
+ processing, except simple mean/std normalization. We
1020
+ use similar augmentation for the Tiny ImageNet dataset
1021
+ which contains 100,000 training images of 200 classes
1022
+ (500 for each class) downsized to 64×64 colored images.
1023
+ The test set contains 10,000 images [11].
1024
+ All baseline models were trained using the same
1025
+ components as the real-valued networks, the original
1026
+ quaternion network, the original vectormap network,
1027
+ the QPHM, and the VPHM networks using the same
1028
+ datasets.
1029
+ All models in Table 2 were trained using
1030
+ the same hyperparameters and the same number of out-
1031
+ put channels. The 50-layer architectural details of the
1032
+ above-mentioned models are depicted in Table 1 for the
1033
+ CIFAR-100 dataset. Due to space limitation, the deep
1034
+ ResNets and VPHM network architectures are not de-
1035
+ picted in the architecture Table 1.
1036
+ In the stem layer, the 3 × 3 convolution network
1037
+ is used for deep ResNets [6], 3 × 3 quaternion net-
1038
+ work is used for the deep quaternion ResNets [4, 18],
1039
+ for the QPHM [16], and axial-hypercomplex networks
1040
+ (our proposed method), and 3 × 3 vectormap network is
1041
+ used for the deep vectormap ResNets [5], and the VPHM
1042
+ [16] networks with stride 1 & 120 output filters. We
1043
+ use parameterized hypercomplex multiplication (PHM)
1044
+ for the dense layer in the backend of deep ResNets,
1045
+ QPHM, VPHM, and our proposed axial-hypercomplex
1046
+ networks. In the bottleneck block, the number of out-
1047
+ put channels of bottleneck groups are 120, 240, 480, &
1048
+ 960 for all networks. In this experiment, we analyze 26-
1049
+ layer, 35-layer, and 50-layer architectures with the bot-
1050
+ tleneck block multipliers “[1, 2, 4, 1]”, “[2, 3, 4, 2]”, and
1051
+ “[3, 4, 6, 3]”. These are depicted in Table 1.
1052
+ We ran all of the models using stochastic gradient de-
1053
+ cent optimizer. We used linearly warmed-up learning
1054
+ from zero to 0.1 for the first 10 epochs and then used
1055
+ cosine learning rate scheduling from epochs 11 to 150.
1056
+ All models were trained for 128 batch sizes.
1057
+ 4.2. Results
1058
+ The overall results of all models (base models and
1059
+ our proposed networks) appear in Tables 2, 3, and 4.
1060
+ The top half of Table 2 shows the results for the CI-
1061
+ FAR10 dataset and the bottom half presents the results
1062
+ for the CIFAR100. Both datasets have been tested by
1063
+ the 26, 35, and 50 layers architectures. These are the pa-
1064
+ rameter count, FLOPS count (number of multiply-add
1065
+ operations), inference time or Latency (time required
1066
+ to process a single image), and the percentage accu-
1067
+ racy of validation results for each model. We evaluate
1068
+ original ResNets [6], ResNet with QPHM [16], orig-
1069
+ inal quaternion networks [4], original vectormap net-
1070
+ works [5], QPHM [16], and VPHM [16] with the same
1071
+ configuration like our proposed axial-hypercomplex net-
1072
+ works. Our proposed axial-hypercomplex networks per-
1073
+ form better in validation accuracy with lower param-
1074
+ eter count and FLOPS for CIFAR-10 and CIFAR-100
1075
+ datasets than the baseline networks.
1076
+ More precisely,
1077
+ our proposed method takes almost 6 times, 1/3 times,
1078
+ 1/2 times, 1/3 times, and 1/2 times fewer parameters
1079
+ than the ResNets, quaternion networks, vectormap net-
1080
+ works, QPHM, and VPHM respectively.
1081
+ Moreover,
1082
+ axial-hypercomplex networks achieved state-of-the-art
1083
+ results for these CIFAR benchmarks in hypercomplex
1084
+ space.
1085
+ The performances for SVHN and Tiny ImageNet
1086
+ datasets are shown in Tables 3 and 4 for all architec-
1087
+ tures. The axial-hypercomplex network’s validation ac-
1088
+ curacies outperform the other base networks with fewer
1089
+ trainable parameters and FLOPS like CIFAR datasets.
1090
+ The result Tables 2, 3, and 4 show our proposed model
1091
+ performance ranges of three runs. However, the latency
1092
+ of axial-hypercomplex networks is a little bit higher in
1093
+ some cases than the quaternion-based networks. This
1094
+ may be due to the use of vectormap networks along with
1095
+ quaternion networks as the latency for vectormap net-
1096
+ works is higher.
1097
+ 5. Discussion and Conclusions
1098
+ This paper proposes axial-hypercomplex convolu-
1099
+ tions to reduce the cost of 2D convolutional opera-
1100
+ tions and shows the effectiveness of image classifi-
1101
+ cation tasks.
1102
+ We also applied 4D PHM in the net-
1103
+ work’s backend.
1104
+ On CIFAR benchmarks, our pro-
1105
+ posed Axial-hypercomplex network, formed by stacking
1106
+ axial-vectormap convolution (three-dimensional) in the
1107
+ quaternion bottleneck blocks, achieved state-of-the-art
1108
+ results among hypercomplex networks.
1109
+ Our main conclusion is that using quaternion convo-
1110
+ lutions as the frontend stem layer, four/five-dimensional
1111
+ PHM-based densely connected backend layer, and axial-
1112
+ hypercomplex bottleneck block improves classification
1113
+
1114
+ performance on the CIFAR benchmarks, SVHN, and
1115
+ Tiny ImageNet datasets in comparison to the other
1116
+ models we tested. Our proposed method factorizes a
1117
+ channel-wise 2D convolution (hypercomplex convolu-
1118
+ tion which works along the channels) to a column con-
1119
+ volution and a row convolution. Extensive experiments
1120
+ show that this leads to systematic improvement with far
1121
+ fewer trainable parameters on image classification. This
1122
+ proposed method can save 33%, and 50% trainable pa-
1123
+ rameters compared to original quaternion and vectormap
1124
+ networks and QPHM and VPHM networks, respectively.
1125
+ Although our proposed axial-hypercomplex design
1126
+ reduced parameter counts and FLOPS, it exhibited
1127
+ higher latency than real-valued and hypercomplex-
1128
+ valued convolutional networks.
1129
+ This is because the
1130
+ model performs convolution twice (height-axis and
1131
+ width-axis) and it takes transition time from 2D convo-
1132
+ lution to two consecutive 1D convolutions. As we re-
1133
+ placed spatial quaternion (four-dimensional hypercom-
1134
+ plex network) 2D convolution using two axial vec-
1135
+ tormap (three-dimensional hypercomplex network) 1D
1136
+ convolutions, the number of output channels are re-
1137
+ stricted to 120 or a multiple of 120 which are divisible by
1138
+ three and four. Our investigation concludes that the per-
1139
+ formance comparison between the hypercomplex net-
1140
+ works and our proposed axial-hypercomplex networks
1141
+ shows that the axial-hypercomplex convolution provides
1142
+ better validation performance with fewer trainable pa-
1143
+ rameters and FLOPS for image classification tasks.
1144
+ Further work may be directed toward the architecture
1145
+ of the axial quaternion network and axial vectormap net-
1146
+ work. Moreover, other datasets will be tested to check
1147
+ whether these proposed architectures can perform in a
1148
+ similar manner or not.
1149
+ Finally, axial-quaternion and
1150
+ axial-vectormap convolutional methods will help to re-
1151
+ move the number of output channels constrained as it
1152
+ will divisible by four for axial-quaternion networks and
1153
+ three for axial-vectormap networks.
1154
+ References
1155
+ [1] Martin Arjovsky, Amar Shah, and Yoshua Bengio. Uni-
1156
+ tary evolution recurrent neural networks.
1157
+ In Interna-
1158
+ tional Conference on Machine Learning, pages 1120–
1159
+ 1128. PMLR, 2016. 1
1160
+ [2] Pierre Buyssens, Abderrahim Elmoataz, and Olivier
1161
+ L´ezoray. Multiscale convolutional neural networks for
1162
+ vision–based classification of cells. In Asian Conference
1163
+ on Computer Vision, pages 342–352. Springer, 2012. 1
1164
+ [3] Ivo Danihelka, Greg Wayne, Benigno Uria, Nal Kalch-
1165
+ brenner, and Alex Graves. Associative long short-term
1166
+ memory. In International Conference on Machine Learn-
1167
+ ing, pages 1986–1994. PMLR, 2016. 1
1168
+ [4] Chase J Gaudet and Anthony S Maida.
1169
+ Deep quater-
1170
+ nion networks. In 2018 International Joint Conference
1171
+ on Neural Networks (IJCNN), pages 1–8. IEEE, 2018. 1,
1172
+ 2, 4, 5, 6, 7
1173
+ [5] Chase J Gaudet and Anthony S Maida. Removing di-
1174
+ mensional restrictions on complex/hyper-complex neural
1175
+ networks. In 2021 IEEE International Conference on Im-
1176
+ age Processing (ICIP), pages 319–323. IEEE, 2021. 1,
1177
+ 3, 4, 5, 6, 7
1178
+ [6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian
1179
+ Sun. Deep residual learning for image recognition. In
1180
+ Proceedings of the IEEE conference on computer vision
1181
+ and pattern recognition, pages 770–778, 2016. 5, 6, 7
1182
+ [7] Akira Hirose and Shotaro Yoshida. Generalization char-
1183
+ acteristics of complex-valued feedforward neural net-
1184
+ works in relation to signal coherence. IEEE Transactions
1185
+ on Neural Networks and learning systems, 23(4):541–
1186
+ 551, 2012. 1
1187
+ [8] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and
1188
+ Tim Salimans. Axial attention in multidimensional trans-
1189
+ formers. arXiv preprint arXiv:1912.12180, 2019. 3
1190
+ [9] Shima Javanmardi,
1191
+ Seyed-Hassan Miraei Ashtiani,
1192
+ Fons J Verbeek, and Alex Martynenko. Computer-vision
1193
+ classification of corn seed varieties using deep convolu-
1194
+ tional neural network. Journal of Stored Products Re-
1195
+ search, 92:101800, 2021. 1
1196
+ [10] Alex Krizhevsky, Geoffrey Hinton, et al. Learning mul-
1197
+ tiple layers of features from tiny images. 2009. 6
1198
+ [11] Ya Le and Xuan S. Yang. Tiny imagenet visual recogni-
1199
+ tion challenge. 2015. 6, 7
1200
+ [12] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bis-
1201
+ sacco, Bo Wu, and Andrew Y Ng. Reading digits in nat-
1202
+ ural images with unsupervised feature learning. 2011. 6,
1203
+ 7
1204
+ [13] Tohru Nitta.
1205
+ On the critical points of the complex-
1206
+ valued neural network. In Proceedings of the 9th Inter-
1207
+ national Conference on Neural Information Processing,
1208
+ 2002. ICONIP’02., volume 3, pages 1099–1103. IEEE,
1209
+ 2002. 1
1210
+ [14] Titouan Parcollet, Mohamed Morchid, and Georges
1211
+ Linar`es. Quaternion convolutional neural networks for
1212
+ heterogeneous image processing. In ICASSP 2019-2019
1213
+ IEEE International Conference on Acoustics, Speech and
1214
+ Signal Processing (ICASSP), pages 8514–8518. IEEE,
1215
+ 2019. 1, 2
1216
+ [15] Titouan Parcollet, Mirco Ravanelli, Mohamed Morchid,
1217
+ Georges Linar`es, Chiheb Trabelsi, Renato De Mori, and
1218
+ Yoshua Bengio. Quaternion recurrent neural networks.
1219
+ arXiv preprint arXiv:1806.04418, 2018. 1, 2
1220
+ [16] Nazmul Shahadat and Anthony Maida. Enhancing resnet
1221
+ image classification performance by using parameterized
1222
+ hypercomplex multiplication, Nov 2021. 1, 2, 3, 4, 5, 6,
1223
+ 7
1224
+ [17] Nazmul Shahadat and Anthony S Maida. Adding quater-
1225
+ nion representations to attention networks for classifica-
1226
+ tion. arXiv preprint arXiv:2110.01185, 2021. 1, 4
1227
+ [18] Chiheb Trabelsi, Olexa Bilaniuk, Ying Zhang, Dmitriy
1228
+ Serdyuk, Sandeep Subramanian, Joao Felipe Santos,
1229
+ Soroush Mehri, Negar Rostamzadeh, Yoshua Bengio,
1230
+ and Christopher J Pal. Deep complex networks. arXiv
1231
+ preprint arXiv:1705.09792, 2017. 2, 7
1232
+
1233
+ [19] Huiyu Wang, Yukun Zhu, Bradley Green, Hartwig
1234
+ Adam, Alan Yuille, and Liang-Chieh Chen.
1235
+ Axial-
1236
+ deeplab: Stand-alone axial-attention for panoptic seg-
1237
+ mentation. In European Conference on Computer Vision,
1238
+ pages 108–126. Springer, 2020. 3
1239
+ [20] Jiasong Wu, Ling Xu, Fuzhi Wu, Youyong Kong, Lotfi
1240
+ Senhadji, and Huazhong Shu. Deep octonion networks.
1241
+ Neurocomputing, 397:179–191, 2020. 1
1242
+ [21] Ruyue Xin, Jiang Zhang, and Yitong Shao. Complex net-
1243
+ work classification with convolutional neural network.
1244
+ Tsinghua Science and technology, 25(4):447–457, 2020.
1245
+ 1
1246
+ [22] Aston Zhang, Yi Tay, Shuai Zhang, Alvin Chan,
1247
+ Anh Tuan Luu, Siu Cheung Hui, and Jie Fu. Beyond
1248
+ fully-connected layers with quaternions: Parameteriza-
1249
+ tion of hypercomplex multiplications with 1/n parame-
1250
+ ters. arXiv preprint arXiv:2102.08597, 2021. 3
1251
+
19E3T4oBgHgl3EQfngq-/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
1dAzT4oBgHgl3EQfDfo-/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1ee1a55b107ef0baa2c6f3bfaf9d3d649e3fe3d6f54e2706f76038ca1ae167fb
3
+ size 3932205
29FST4oBgHgl3EQfYTgs/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f77ecc606b722b64c5cee732e6833a61274b9dca86e8bcdfce00e5fbded4f9ef
3
+ size 3997741
2NE4T4oBgHgl3EQfagwY/content/2301.05064v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ad46b3db0d08b09bb4a44bb3bdc530a4eb85bf95a206b63fc4510061ef73ca6
3
+ size 1473776
2NE4T4oBgHgl3EQfagwY/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45b58acfcda79e70a75d2a4ac9f16136845f1a34777af98eb208d748034d8711
3
+ size 2031661
2NE4T4oBgHgl3EQfagwY/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ea54ad38f832cd24183496cc976c7ad0037911202cbd9a66c6bd6fd0a7c12da5
3
+ size 70924
4tE4T4oBgHgl3EQf1A0X/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4ba40e328a12de8e3a5d290ea3b3a3665b1c9a0122c64fdfc198494bc2e334e
3
+ size 2752557
5NAyT4oBgHgl3EQfpPjF/content/tmp_files/2301.00523v1.pdf.txt ADDED
@@ -0,0 +1,1189 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Bayesian Generalized Kernel Inference for Exploration
2
+ of Autonomous Robots
3
+ Yang Xu, Student Member, IEEE, Ronghao Zheng†, Member, IEEE, Senlin Zhang, Member, IEEE,
4
+ and Meiqin Liu, Senior Member, IEEE
5
+ Abstract— This paper concerns realizing highly efficient
6
+ information-theoretic robot exploration with desired perfor-
7
+ mance in complex scenes. We build a continuous lightweight
8
+ inference model to predict the mutual information (MI) and
9
+ the associated prediction confidence of the robot’s candidate
10
+ actions which have not been evaluated explicitly. This allows
11
+ the decision-making stage in robot exploration to run with
12
+ a logarithmic complexity approximately, this will also benefit
13
+ online exploration in large unstructured, and cluttered places
14
+ that need more spatial samples to assess and decide. We also
15
+ develop an objective function to balance the local optimal action
16
+ with highest MI value and the global choice with high prediction
17
+ variance. Extensive numerical and dataset simulations show
18
+ the desired efficiency of our proposed method without losing
19
+ exploration performance in different environments. We also
20
+ provide our open-source implementation codes released on
21
+ GitHub for the robot community.
22
+ I. INTRODUCTION
23
+ Robot exploration gains its prevalence recently in pri-
24
+ ori unknown environments such as subterranean, marine,
25
+ and planetary tasks [1]–[3]. Among the literature, state-
26
+ of-the-art exploration methods prefer to use information-
27
+ theoretic metrics in each iteration, such as Shannon mutual
28
+ information (MI) [4] and its derivatives [5]–[8], to evaluate
29
+ the information gain brought by candidate control actions
30
+ accurately and choose and execute the most informative
31
+ action, thus the exploration problem becomes a sequential
32
+ optimal decision-making one naturally. A typical exploration
33
+ example is in Fig. 1.
34
+ Intuitively, the way to tackle this problem is to use a
35
+ greedy strategy and add more candidate actions, including
36
+ sampled nodes [9], [10], available viewpoints [11], [12],
37
+ or special motion primitives [13], [14], in the discrete ac-
38
+ tion space. However, the exploration performance of greedy
39
+ selection is closely related to the discrete sampling reso-
40
+ lution/method of action space over the map grid, i.e., a
41
+ coarse resolution may lead to sub-optimal actions/paths,
42
+ and a fine one may generate more samples and be more
43
+ likely to choose the optimal action, but the computational
44
+ cost of the information gain evaluation of all candidate
45
+ actions will become expensive in this case since the forward
46
+ 1Yang Xu, Ronghao Zheng and Senlin Zhang are with the College
47
+ of Electrical Engineering, Zhejiang University, Hangzhou 310027, China.
48
+ {xuyang94,rzheng,slzhang}@zju.edu.cn
49
+ 2Meiqin
50
+ Liu
51
+ is
52
+ with
53
+ the
54
+ Institute
55
+ of
56
+ Artificial
57
+ Intelligence
58
+ and
59
+ Robotics,
60
+ Xi’an
61
+ Jiaotong
62
+ University,
63
+ Xi’an
64
+ 710049,
65
+ China.
66
+ liumeiqin@zju.edu.cn
67
+ 3All authors are also with the State Key Laboratory of Industrial Control
68
+ Technology, Zhejiang University, Hangzhou 310027, China.
69
+ †Corresponding author
70
+ (a)
71
+ 20
72
+ 40
73
+ 60
74
+ 80
75
+ 100
76
+ 120
77
+ X(grid)
78
+ 20
79
+ 40
80
+ 60
81
+ Y(grid)
82
+ 0
83
+ 0.1
84
+ 0.2
85
+ 0.3
86
+ 0.4
87
+ 0.5
88
+ 0.6
89
+ (b)
90
+ Fig. 1.
91
+ MI-based active robot exploration in an unknown unstructured
92
+ environment. (a) Informative trajectory and the resulting occupancy map,
93
+ (b) Resulting MI surface. Note that the coincident yellow squares mean the
94
+ start and end points. A minimum information threshold is set to select more
95
+ informative exploration actions, e.g. the middle left and top right areas are
96
+ less informative than the threshold and thus unexplored. Note that the scale
97
+ of MI is in [0,1] bit in this paper.
98
+ simulation in the evaluation requires extensive raycasting and
99
+ MI calculation. Notably, these consequences will be more
100
+ distinct in 3D environments because the increased dimension
101
+ needs much more samples.
102
+ In this paper, we aim to realize a more efficient and
103
+ accurate approach to find the most informative action without
104
+ evaluating all candidate actions exhaustively and expensively
105
+ in robot exploration. Specifically, our main contributions are
106
+ three-fold:
107
+ 1) We propose a Bayesian kernel spatial MI inference
108
+ method to construct a continuous surrogate evaluation model
109
+ between robot actions and MI values using only partial ex-
110
+ plicitly evaluated samples, which can perform highly efficient
111
+ MI prediction of control actions in logarithm time;
112
+ 2) We develop a reward function comprising the predicted
113
+ MI values and uncertainties to find the best action for realiz-
114
+ ing the trade-off between exploration and exploitation, which
115
+ has been validated in numerical and dataset simulations;
116
+ 3) Meanwhile, we release an open-source implementation
117
+ of our proposed method here1 for the robotics community.
118
+ The paper organization is as follows. Related works about
119
+ the recent learning-based robot exploration methods are pre-
120
+ sented in Section II. We formulate the problem in Section III
121
+ and present our Bayesian kernel-based MI inference method
122
+ in Section IV. Simulation results using synthetic data and
123
+ the real world dataset and discussions are given in Section
124
+ V, followed by conclusions in Section VI.
125
+ II. RELATED WORK
126
+ In the context of robot exploration, supervised learning
127
+ techniques provide a powerful tool to find the global op-
128
+ 1https://github.com/Shepherd-Gregory/BKI-exploration
129
+ arXiv:2301.00523v1 [cs.RO] 2 Jan 2023
130
+
131
+ 0.12020C.C
132
+ 0.2(grid)
133
+ 400.440
134
+ 60
135
+ X(gric0.6
136
+ 0.56080
137
+ )0.7100
138
+ 120timum approximately by training predictive models using
139
+ minor parts of actions in continuous action spaces, without
140
+ evaluating the objective function expensively, which also has
141
+ better interpretability in black-box inference [15]–[17].
142
+ In [18], Bai et al. used the Gaussian process (GP) to model
143
+ the relationship between control actions and the explicitly
144
+ evaluated MI for the robot exploring priori unknown areas.
145
+ In [19], they further introduced Bayesian optimization (BO)
146
+ into the information-theoretic robot exploration to optimize
147
+ the GP prediction in multiple iterations, which provides rapid
148
+ map entropy reduction and ensures computational efficiency.
149
+ Generally, BO assumes a prior distribution on the objective
150
+ function and constructs predictive models to describe the
151
+ underlying relationship between robot actions and their MI.
152
+ It also assesses the acquisition function derived from the
153
+ GP prior and samples, then chooses the next query point
154
+ maximizing the acquisition function and balancing the trade-
155
+ off between exploration (global) and exploitation (local).
156
+ Iteratively, BO presents more precise results on the posterior
157
+ distribution as the observations (training samples) increase.
158
+ Rather than evaluating discrete viewpoints, Francis et al. [17]
159
+ modeled the autonomous exploration and mapping task as a
160
+ constrained BO aiming to find optimal continuous paths.
161
+ However, the main bottleneck of the above BO-based robot
162
+ exploration methods is that the number of the training actions
163
+ N will affect the resulting prediction accuracy directly, as
164
+ well as the computational cost. That implies one needs to
165
+ pay expensive computations to achieve higher exploration
166
+ performance. Typically, updating and querying the GP mod-
167
+ els (the engine behind BO) have an overall O(N 3) time
168
+ complexity. This compromises the inference efficiency and
169
+ real-time performance of robot exploration tasks inevitably,
170
+ especially in large-scale and 3D scenes.
171
+ More recently, deep neural networks (DNNs) have been
172
+ introduced to realize predicting optimal sensing actions more
173
+ efficiently. Bai et al. [20] trained the DNN with plenty of
174
+ randomly generated 2D maps to generate suggested action
175
+ and ensure inferring in constant time. Graph neural networks
176
+ (GNNs) have also been combined with reinforcement learn-
177
+ ing methods to learn the best action from an exploration
178
+ graph, rather than metric maps or visual images [21], [22].
179
+ Nevertheless, the neural network-based robot exploration
180
+ methods require numerous training samples beforehand and
181
+ are also limited to the adaptability and generalization ca-
182
+ pability in different environments, which may need further
183
+ studies in the future.
184
+ Encouragingly, the Bayesian kernel inference (BKI) tech-
185
+ nique proposed in [23] gives us a chance to perform ef-
186
+ ficient exact inference on a simplified model, rather than
187
+ approximating inference on an exact generative model (e.g.
188
+ GP) expensively. BKI extends local kernel estimation to
189
+ Bayesian inference for exponential family likelihood func-
190
+ tions, enabling only O(log Nq) (Nq: the number of querying
191
+ samples) run time for inference. These significant merits
192
+ enhance BKI’s application in robotics, including sensor un-
193
+ certainty estimation [24], high-speed navigation [25], as well
194
+ as environment mapping using sparse sensor measurements
195
+ such as terrain traversability mapping [26], 3D occupancy
196
+ mapping [27], semantic mapping [28].
197
+ Motivated by [19] and [23], use BKI to infer the spatial MI
198
+ in an efficient and closed-form way for the control actions
199
+ whose MI values have not been explicitly evaluated via
200
+ expensive computation (e.g. [4]). Our method keeps similar
201
+ accuracy to previous approaches compared with existing
202
+ works such as [18] and [19], but shows more efficient and
203
+ suitable performance for complex scenes requiring numerous
204
+ explicitly evaluated samples.
205
+ III. PRELIMINARIES AND NOTIONS
206
+ In this paper, for simplicity of discussion, we mainly
207
+ consider the information-theoretic exploration using a mobile
208
+ robot equipped with a beam-based range sensor of limited
209
+ field of view (FOV) in a 2D environment. The results here
210
+ can also be extended to 3D cases expediently.
211
+ A. Information-Theoretic Exploration
212
+ Generally, the robot generates a set of candidate actions
213
+ Xaction in the robot’s feasible configuration space X ⊆
214
+ SE(2). We also assume this configuration space has been
215
+ discretized by a fixed resolution over the 2D static grid map.
216
+ The set of values m ∈ [0, 1] is the occupancy level over the
217
+ independent grid cells and can be updated and queried by
218
+ the classic log-odds method [29]. The occupancy value of
219
+ an unobserved map cell ξ is assumed to be uniform, i.e.,
220
+ p(mξ) = 0.5.
221
+ Here we use the classic Shannon MI [4] as the information
222
+ measure of candidate configuration xi = [px
223
+ i , py
224
+ i , ψi] ∈
225
+ Xaction, where px
226
+ i and py
227
+ i denote the robot’s position on the
228
+ map, and ψi denotes the heading angle of the robot. From
229
+ the view of information theory, the expected information
230
+ gain of xi can be evaluated by the current map entropy and
231
+ conditional entropy given a new measurement at xi:
232
+ I(m; xi) = H(m) − H(m|xi).
233
+ (1)
234
+ The aim of information-theoretic robot exploration is to
235
+ select the best action xbest maximizing the expected MI:
236
+ xbest = argmax
237
+ x∈Xaction
238
+ I(m; xi).
239
+ (2)
240
+ Notably, the MI of each configuration can be decomposed
241
+ over independent beams and then to cells via raycasting, then
242
+ accumulated MI over cells to approximate, which owns a
243
+ squared time complexity in map resolution λm at worst [7].
244
+ This also brings more evaluation costs for robot exploration.
245
+ B. Bayesian Generalized Kernel Inference
246
+ Consider a supervised learning-based inference problem
247
+ on predictive stochastic models p(y|x) given a sequence of
248
+ N observations D = {(x = {xi}, y = {yi})}N
249
+ i=1, where x
250
+ and y represent the set of evaluated configurations and the
251
+ resulting MI values I(m; x), respectively. The main objective
252
+ is to infer the posterior distribution p(y∗|x∗, D) for the target
253
+ inputs x∗ to be evaluated. This problem can be solved by
254
+ associating latent parameters θ = {θi}N
255
+ i=1 ∈ Θ with input x
256
+ in the latent space Θ, where the likelihood p(y|θ) is known.
257
+
258
+ Thus the inference on y∗ can be formulated as an inference
259
+ on target parameters θ∗ related to x∗:
260
+ p(y∗|x∗, D) =
261
+
262
+ Θ
263
+ p(y∗|θ∗)p(θ∗|x∗, D)dθ∗,
264
+ (3)
265
+ where the posterior distribution of the latent variables
266
+ can be characterized using Bayes’ rule: p(θ∗|x∗, D) ∝
267
+
268
+ Θ
269
+ �N
270
+ i=1 p(yi|θi)p(θ1:N, θ∗|x1:N, x∗)dθ1:N.
271
+ By strongly assuming latent parameters θ1:N are con-
272
+ ditionally independent given the target parameters θ∗:
273
+ p(θ1:N, θ∗|a1:N, x∗) = �N
274
+ i=1 p(θi|θ∗, xi, x∗)p(θ∗|x∗), one
275
+ can marginalize the latent variables θ1:N and then obtain
276
+ p(θ∗|x∗, D) ∝ �N
277
+ i=1 p(yi|θ∗, xi, x∗)p(θ∗|x∗).
278
+ BKI further defines a distribution that has a special
279
+ smoothness constraint and bounded Kullback-Leibler diver-
280
+ gence (KLD) DKL(g||f) between the extended likelihood
281
+ p(yi|θ∗, xi, x∗) represented by g and the likelihood p(yi|θi)
282
+ represented by f, i.e., the maximum entropy distribution
283
+ g satisfying DKL(g||f) ≤ ρ(x∗, x) has the form g(y) ∝
284
+ f(y)k(x∗,x), where ρ(·, ·) : X × X → R+ is a smoothness
285
+ bound and k(·, ·) : X ×X → [0, 1] is a kernel function which
286
+ can be uniquely determined by ρ. Substituting into Eq. (3),
287
+ we can get:
288
+ p(θ∗|x∗, D) ∝
289
+ N
290
+
291
+ i=1
292
+ p(yi|θ∗)k(x∗,x)p(θ∗|x∗)
293
+ (4)
294
+ Thus the posterior distribution can be exactly inferred
295
+ by using the likelihood from the exponential family and
296
+ assuming the corresponding conjugate prior.
297
+ IV. BAYESIAN KERNEL INFERENCE FOR ROBOT
298
+ EXPLORATION
299
+ To efficiently evaluate the exact MI of unknown robot
300
+ configurations sampled in the spatial action space, we solve
301
+ this problem by a Bayesian kernel inference way.
302
+ A. Bayesian Kernel Spatial MI Inference
303
+ As mentioned in Section III.B, we assume the underlying
304
+ likelihood model between the MI values y and the latent
305
+ parameters θ follows Gaussian distribution with unknown
306
+ mean µ ∈ RN and fixed, known covariance Σ:
307
+ p(y|µ) = N(µ, Σ), Σ = diag(σ2) ∈ RN×N,
308
+ (5)
309
+ thus its conjugate prior can also be described by a Gaussian
310
+ distribution using the hyperparameter ζ and target samples
311
+ input x∗:
312
+ p(µ|x∗) = N
313
+
314
+ µ0(x∗),
315
+ 1
316
+ ζ(x∗)Σ(x∗)
317
+
318
+ ,
319
+ (6)
320
+ where µ0 and ζ are the initial belief of the mean and the
321
+ uncertainty of that belief, respectively. ζ = 0 means no
322
+ confidence and ζ → ∞ indicates full prior knowledge. Here
323
+ we assume ζ is a quite small positive constant since we
324
+ do not have much prior information about the belief when
325
+ exploring unknown areas.
326
+ Therefore, we can substitute Eq. (6) and Eq. (5) into
327
+ Eq. (4) given observations D:
328
+ p(µ∗|x∗, D) ∝
329
+ N
330
+
331
+ i=1
332
+ exp
333
+
334
+ −1
335
+ 2
336
+ (yi − µi)2
337
+ σ2
338
+ k(x∗, xi)
339
+
340
+ (7)
341
+ · exp
342
+
343
+ −1
344
+ 2
345
+ (µi − µ0)2
346
+ σ2
347
+ ζ
348
+
349
+ ,
350
+ and the posterior over mean and covariance of the MI can
351
+ be derived as follows:
352
+ I(x∗) = E[y∗|x∗, D] = E[µ∗|x∗, D] = y + ζµ0
353
+ ζ + k
354
+ ≃ y
355
+ k ,
356
+ σI(x∗) = V[µ∗|x∗, D] =
357
+ Σ
358
+ ζ + k ≃ Σ
359
+ k ,
360
+ (8)
361
+ where y and k can be computed by kernel functions:
362
+ k = ΣN
363
+ i=1k(x∗, x), y = ΣN
364
+ i=1k(x∗, x)yi.
365
+ (9)
366
+ Give a set of observations D evaluated explicitly as the
367
+ input, then we can easily compute the MI and the corre-
368
+ sponding confidence for the test spatial configurations x∗ by
369
+ using Eq. (8) and Eq. (9).
370
+ B. Kernel Selection
371
+ The kernel function of the BKI method will directly affect
372
+ the computational efficiency and accuracy, thus selecting
373
+ an appropriate kernel is quite significant. In [26]–[28], the
374
+ chosen sparse kernels remove the training points far away
375
+ from the queried points, which allows efficient and exact
376
+ evaluation (e.g. occupancy, traversability, semantic class)
377
+ over the observations in logarithm run time using k-d trees.
378
+ Unlike the sufficient training data obtained from onboard
379
+ sensors in mapping tasks, robot exploration always generates
380
+ and evaluates relatively fewer candidate configurations in a
381
+ limited space at each time instance, so there is no need to
382
+ reject the rare training samples in robot exploration tasks.
383
+ Among the exponential kernel functions, we prefer the
384
+ Mat´ern kernel for its capability of handling sudden transi-
385
+ tions of terrain [30], [31], since the potential obstacles and
386
+ unknown structures in application scenes that have never
387
+ been seen before will vary the MI values greatly. The typical
388
+ Mat´ern kernel function is as follows:
389
+ k(x∗, x) = 21−ν
390
+ Γ(ν) (
391
+
392
+ 2νr
393
+
394
+ )νKν(
395
+
396
+ 2νr
397
+
398
+ ), r = ||x∗−x||, (10)
399
+ where the positive parameters ν and ℓ are the smoothness
400
+ constant and characteristic length scale respectively, Γ(·) and
401
+ Kν are the gamma and modified Bassel function, respec-
402
+ tively. In practice, we choose a Mat´ern 3/2 kernel (ν = 3/2)
403
+ with the form as k(x∗, x) = (1 +
404
+
405
+ 3r
406
+ ℓ ) exp(−
407
+
408
+ 3r
409
+ ℓ ).
410
+ C. BKI-based Robot Exploration
411
+ In robot exploration, we expect the robot moves toward
412
+ the places with high predicted MI values to maximize the
413
+ information gain locally, but this greedy “exploration” may
414
+ lead to undesired paths or even worse such as getting stuck
415
+ in cluttered areas. Instead, the unexplored places with high
416
+ predicted uncertainty are also worth exploring, since they
417
+
418
+ may guide an optimal path for the robot globally in a prior
419
+ unknown area, which is also characterized as “exploitation”.
420
+ Therefore, we integrate the prediction confidence of MI
421
+ values with the predicted MI to realize a trade-off between
422
+ the exploration and exploitation, then we can get the sug-
423
+ gested action maximizing the information objective function
424
+ based on Eq. (2) and Eq. (8):
425
+ xs = argmax
426
+ x∈Xaction
427
+ αI(m; xi) + (1 − α)σI(xi),
428
+ (11)
429
+ where α ∈ [0, 1] is the trade-off factor.
430
+ The autonomous exploration framework based on our
431
+ BKI MI inference method is given in Algorithm 1, where
432
+ Algorithm 2 is the BKI optimization module.
433
+ Algorithm 1 BKI Exploration( )
434
+ Require: Occupancy map at kth time step mk, previous
435
+ robot poses xhist = x0:k−1 and current pose xk, the
436
+ number of explicit evaluated samples N, information
437
+ threshold Ith, the number of querying samples Nq,
438
+ while-loop counts limit Nloop
439
+ 1: iter = 0
440
+ 2: while xhist ̸= ∅ AND iter < Nloop do
441
+ 3:
442
+ iter = iter + 1
443
+ 4:
444
+ // Sample N training actions
445
+ 5:
446
+ x ← Sampling(xk, mk, N);
447
+ 6:
448
+ // Evaluate these actions explicitly Eq. (1)
449
+ 7:
450
+ for each xi ∈ x do
451
+ 8:
452
+ mvirtual ← Raycasting(xi, mk);
453
+ 9:
454
+ Ii ← ComputeMI(mvirtual);
455
+ 10:
456
+ y ← y ∪ Ii;
457
+ 11:
458
+ end for
459
+ 12:
460
+ x∗ ← Sampling(xk, mk, Nq);
461
+ 13:
462
+ // Find the suggested action using Algorithm 2
463
+ 14:
464
+ {xbest, Ibest} ← BKIOptimization({x, y}, x∗);
465
+ 15:
466
+ if max(Ibest) > Ith then
467
+ 16:
468
+ xk+1 ← xbest(MaxInfoIndex);
469
+ 17:
470
+ xhist ← xhist ∪ xk+1;
471
+ 18:
472
+ else
473
+ 19:
474
+ xk+1 ← xk−1; // Back to previous action
475
+ 20:
476
+ Remove xk−1 from xhist;
477
+ 21:
478
+ end if
479
+ 22:
480
+ // Execute the action and update the map
481
+ 23:
482
+ Plocal ← Astar(xk, xk+1) // Plan local path by A*
483
+ 24:
484
+ mk+1 ← OccupancyGridMapping(Plocal);
485
+ 25: end while
486
+ Proposition 1 The time complexity of our proposed method
487
+ at each while-loop step in Algorithm 1 is:
488
+ O(NNzN 2
489
+ c )
490
+
491
+ ��
492
+
493
+ explicit MI evaluation
494
+ +
495
+ O(NepochN log Nq)
496
+
497
+ ��
498
+
499
+ BKI MI inference
500
+ (12)
501
+ where Nepoch is the number of training epoch, Nz and Nc
502
+ are the numbers of beams per sensor scan, and the number
503
+ of cells that a beam intersects with the grid map at worst,
504
+ respectively.
505
+ Algorithm 2 BKI Optimization( )
506
+ Require: Training set D = {(xi, yi)}N
507
+ i=1, current action set
508
+ to be evaluated x∗, training epoch Nepoch, factor α
509
+ 1: xbest ← {}, Ibest ← {};
510
+ 2: for each epoch do
511
+ 3:
512
+ // Compute the kernel function using Eq. (11)
513
+ 4:
514
+ k ← KernelFunction(x∗, x);
515
+ 5:
516
+ // Compute MI and uncertainty using Eq. (8)
517
+ 6:
518
+ k ← Σk, y ← k · y;
519
+ 7:
520
+ I∗ ← y/k, σ∗
521
+ I ← Σ/k;
522
+ 8:
523
+ ObjFunc ← αI∗ + (1 − α)σ∗
524
+ I;
525
+ 9:
526
+ xs = max(ObjFunc);
527
+ 10:
528
+ if xs ∈ x then
529
+ 11:
530
+ xbest ← xbest ∪ xs, Ibest ← Ibest ∪ ys
531
+ 12:
532
+ else
533
+ 13:
534
+ // Evaluate MI explicitly using Eq. (1)
535
+ 14:
536
+ Is = CalculateMI(xs);
537
+ 15:
538
+ // Add into D
539
+ 16:
540
+ xbest ← xbest ∪ xs, x ← x ∪ xs;
541
+ 17:
542
+ Ibest ← Ibest ∪ Is, y ← y ∪ Is;
543
+ 18:
544
+ end if
545
+ 19: end for
546
+ 20: return xbest, Ibest
547
+ Significantly, the GP-based robot exploration in [18] and
548
+ BO-based method in [19] have the same time cost of ours in
549
+ explicit MI evaluation, but these two methods have computa-
550
+ tional complexities of O(N 3 +N 2Nq) and O(Nepoch(N 3 +
551
+ N 2Nq)) to perform the expensive GP inference for MI,
552
+ respectively. This comparative theoretic result indicates our
553
+ BKI-based exploration method outperforms the GP-based
554
+ methods in time efficiency, especially in large-scale and
555
+ cluttered places which need more samples N and Nq to
556
+ evaluate rapidly.
557
+ V. RESULTS AND DISCUSSIONS
558
+ In this section, we run numerical simulations and dataset
559
+ experiments on a desktop PC with a 3.6 GHz Intel i3-
560
+ 9100F CPU and 32G RAM to verify the effectiveness of
561
+ proposed BKI-based robot exploration method. The infor-
562
+ mation threshold is Ith = 0.05 bit and the trade-off factor is
563
+ α = 0.5. We adopt a Mat´ern kernel for GP and the kernel
564
+ parameters are ℓ = 1 and ν = 3/2 for all simulations. We
565
+ also choose the parameters of ζ = 0.001 and σ = 0.01 for
566
+ BKI method. The robot poses are assumed to be known and
567
+ the robot’s candidate actions are sampled uniformly in the
568
+ FOV of range sensors. We conduct 20 Monte Carlo trials for
569
+ all maps.
570
+ We use greedy-based optimization (named “NBO” in
571
+ simulations), batch GP with only 1 epoch for optimization
572
+ (“bacth GP”) [18], and GP-based BO with multiple epochs
573
+ (“GP-BO”) [19] to compare with our methods, one named
574
+ “bacth BKI” with only 1 optimization epoch and another
575
+ one named “BKI-BO” with multiple epochs. Meanwhile, to
576
+ validate the time efficiencies, we apply 2 cases of N = 30
577
+ and N = 60 samples for each method, where GP-BO 30
578
+
579
+ Informative trajectory in occupancy map
580
+ 20
581
+ 40
582
+ 60
583
+ 80
584
+ 100
585
+ 120
586
+ X(grid)
587
+ 10
588
+ 20
589
+ 30
590
+ 40
591
+ 50
592
+ 60
593
+ 70
594
+ Y(grid)
595
+ 0.1
596
+ 0.2
597
+ 0.3
598
+ 0.4
599
+ 0.5
600
+ 0.6
601
+ 0.7
602
+ (a) Informative trajectory
603
+ MI surface
604
+ 20
605
+ 40
606
+ 60
607
+ 80
608
+ 100
609
+ 120
610
+ X(grid)
611
+ 10
612
+ 20
613
+ 30
614
+ 40
615
+ 50
616
+ 60
617
+ 70
618
+ Y(grid)
619
+ 0.2
620
+ 0.3
621
+ 0.4
622
+ 0.5
623
+ 0.6
624
+ (b) MI surface
625
+ Fig. 2.
626
+ An example of BKI-based robot exploration in an unknown
627
+ structured environment. Yellow square: start point; yellow star: end point;
628
+ red line: robot direction at each action.
629
+ and BKI-BO 30 use Nepoch = 15 iterations, GP-BO 60 and
630
+ BKI-BO 60 use 30 epochs in BKI optimization. We also set
631
+ Nq = 8N in all simulations.
632
+ A. Synthetic Environments Results
633
+ To simulate the indoor and field scenes, we generate 2
634
+ 24 m × 14 m synthetic maps, one structured maze map
635
+ surrounded by several walls (shown in Fig. 2, Nloop = 50),
636
+ and one unstructured map consisting of circles and ellipses
637
+ (shown in Fig. 1, Nloop = 150). The map resolutions are both
638
+ 0.2 m. The simulated range sensor has a FOV of ±1.5 rad
639
+ with a resolution of 0.05 rad, and a maximum sensing range
640
+ of 6 m. The robot is initially at [1.2 m, 1.2 m] with 0 rad
641
+ heading and trying to explore the prior unknown map. The
642
+ representative resulting paths maximizing the information
643
+ objective function are in Fig. 1 and Fig. 2.
644
+ The qualitative results of structured and unstructured maps
645
+ are shown in Fig. 3 and Fig. 4, respectively. To compare the
646
+ exploration performance using different methods intuitively,
647
+ we present the evolution of map entropy and coverage rate of
648
+ each method in the figures, where the solid and dashed lines
649
+ depict the means of Monte Carlo trials for each method, and
650
+ the shaded regions represent the standard deviations.
651
+ Fig. 3 shows the BKI and GP methods have similar
652
+ performance to the NBO methods since this structured scene
653
+ is relatively small and simple, especially in the beginning
654
+ stage where there is only one corridor to move forward.
655
+ Differently, Fig. 4 indicates that the NBO methods spend
656
+ more time (about 50∼70 steps) to converge and end the
657
+ exploration, while BKI and GP methods complete the ex-
658
+ ploration with comparable entropy reduction and coverage
659
+ rates to NBOs.
660
+ Moreover, as in Fig. 5, we use the explicitly evaluated MI
661
+ as the ground truth and compute the MI prediction errors
662
+ using BKI-BO and GP-BO methods with small training
663
+ samples in a randomly selected step, which implies the
664
+ BKI-based approach can resemble the GP-based one in MI
665
+ inference accuracy when facing challenging cases.
666
+ In short, these results validate that our BKI methods have
667
+ competitive properties with GP-based exploration ones in the
668
+ typical structured and unstructured scenes.
669
+ B. Dataset Results
670
+ To test our method in a more complex environment, we
671
+ choose the Seattle map [32] containing narrow long corridors
672
+ 0
673
+ 10
674
+ 20
675
+ 30
676
+ 40
677
+ 50
678
+ 60
679
+ Steps
680
+ 2000
681
+ 2200
682
+ 2400
683
+ 2600
684
+ 2800
685
+ 3000
686
+ Map entropy (bits)
687
+ NBO 30
688
+ NBO 60
689
+ GP-BO 30
690
+ GP-BO 60
691
+ BKI-BO 30
692
+ BKI-BO 60
693
+ (a) Map entropy
694
+ 0
695
+ 10
696
+ 20
697
+ 30
698
+ 40
699
+ 50
700
+ 60
701
+ Steps
702
+ 2000
703
+ 2200
704
+ 2400
705
+ 2600
706
+ 2800
707
+ 3000
708
+ Map entropy (bits)
709
+ NBO 30
710
+ NBO 60
711
+ batch GP 30
712
+ batch GP 60
713
+ batch BKI 30
714
+ batch BKI 60
715
+ (b) Map entropy
716
+ 0
717
+ 10
718
+ 20
719
+ 30
720
+ 40
721
+ 50
722
+ 60
723
+ Steps
724
+ 0
725
+ 0.1
726
+ 0.2
727
+ 0.3
728
+ 0.4
729
+ 0.5
730
+ 0.6
731
+ 0.7
732
+ 0.8
733
+ Coverage
734
+ NBO 30
735
+ NBO 60
736
+ GP-BO 30
737
+ GP-BO 60
738
+ BKI-BO 30
739
+ BKI-BO 60
740
+ (c) Coverage
741
+ 0
742
+ 10
743
+ 20
744
+ 30
745
+ 40
746
+ 50
747
+ 60
748
+ Steps
749
+ 0
750
+ 0.1
751
+ 0.2
752
+ 0.3
753
+ 0.4
754
+ 0.5
755
+ 0.6
756
+ 0.7
757
+ 0.8
758
+ Coverage
759
+ NBO 30
760
+ NBO 60
761
+ batch GP 30
762
+ batch GP 60
763
+ batch BKI 30
764
+ batch BKI 60
765
+ (d) Coverage
766
+ Fig. 3.
767
+ Map entropy and coverage results of the synthetic structured map.
768
+ and cluttered rooms, as in Fig. 6. The map size is 24 m×14 m
769
+ with a resolution of 0.2 m. We use a simulated laser scanner
770
+ emitting 20 beams uniformly within a FOV of ±π/3 rad at
771
+ a maximum range of 4 m. The robot starts at [13, 57]m with
772
+ a −π/2 initial heading angle. The Nloop is set to 100.
773
+ Fig. 7 presents the comparative curves of map entropy
774
+ and coverage rates, whereas Fig. 7(a) shows the BKI-BO
775
+ methods have more rapid reduction rates of map entropy
776
+ after the exploration starts and arrive at relatively lower levels
777
+ than other methods, among them, BKI-BO 60 performs the
778
+ best. In this typical cluttered map, GP-BO methods perform
779
+ slightly inferior to our BKI-BO methods but almost catch
780
+ up with ours, which also are much better than the NBO
781
+ methods. The curves in Fig. 7(b) imply that batch GP and
782
+ batch BKI have similar performance. We also can get an
783
+ insight from Fig. 7(c) and (d), i.e., the coverage curves
784
+ of BKI-BO methods converge slightly earlier than GP-BO
785
+ methods and reach higher values, and all BO-based methods
786
+ explore the unknown place much faster than the NBO ones.
787
+ This result evidences our BKI methods are more suitable for
788
+ large cluttered environments.
789
+ C. Time Efficiency
790
+ We have presented the exploration results in the previous
791
+ simulations of typical scenes, and our BKI-based method
792
+ has shown desired exploration performance in efficiency and
793
+ accuracy compared with state-of-the-art methods. To put
794
+ more intuitive and specific comparison, we further analyze
795
+ the time cost of each method per exploration step in all
796
+ maps. As in Table I, the results show the time cost of the
797
+ whole exploration process per step in the form of means
798
+ and standard deviations, as well as the average percent
799
+ of evaluation and decision-making time spent by different
800
+ methods in each step.
801
+
802
+ 0
803
+ 50
804
+ 100
805
+ 150
806
+ Steps
807
+ 1800
808
+ 2000
809
+ 2200
810
+ 2400
811
+ 2600
812
+ 2800
813
+ 3000
814
+ Map entropy (bits)
815
+ NBO 30
816
+ NBO 60
817
+ GP-BO 30
818
+ GP-BO 60
819
+ BKI-BO 30
820
+ BKI-BO 60
821
+ (a)
822
+ 0
823
+ 50
824
+ 100
825
+ 150
826
+ Steps
827
+ 1800
828
+ 2000
829
+ 2200
830
+ 2400
831
+ 2600
832
+ 2800
833
+ 3000
834
+ Map entropy (bits)
835
+ NBO 30
836
+ NBO 60
837
+ batch GP 30
838
+ batch GP 60
839
+ batch BKI 30
840
+ batch BKI 60
841
+ (b)
842
+ 0
843
+ 50
844
+ 100
845
+ 150
846
+ Steps
847
+ 0
848
+ 0.2
849
+ 0.4
850
+ 0.6
851
+ 0.8
852
+ 1
853
+ Coverage
854
+ NBO 30
855
+ NBO 60
856
+ GP-BO 30
857
+ GP-BO 60
858
+ BKI-BO 30
859
+ BKI-BO 60
860
+ (c)
861
+ 0
862
+ 50
863
+ 100
864
+ 150
865
+ Steps
866
+ 0
867
+ 0.2
868
+ 0.4
869
+ 0.6
870
+ 0.8
871
+ 1
872
+ Coverage
873
+ NBO 30
874
+ NBO 60
875
+ batch GP 30
876
+ batch GP 60
877
+ batch BKI 30
878
+ batch BKI 60
879
+ (d)
880
+ Fig. 4.
881
+ Map entropy and coverage results of the synthetic unstructured map.
882
+ TABLE I
883
+ TIME COST COMPARISON OF DIFFERENT EXPLORATION METHODS
884
+ Methods
885
+ Synthetic structured map
886
+ Synthetic unstructured map
887
+ Seattle map [32]
888
+ NBO 30
889
+ 95.29% / 10.4455 ± 0.9409
890
+ 96% / 12.1683 ± 1.3856
891
+ 96.95% / 4.8434 ± 0.7311
892
+ NBO 60
893
+ 95.38% / 10.9967 ± 1.0676
894
+ 95.93% / 12.4971 ± 2.1583
895
+ 96.93% / 5.3502 ± 0.9009
896
+ batch GP 30
897
+ 5.15% / 0.4387 ± 0.0246
898
+ 4.66% / 0.2805 ± 0.0232
899
+ 12.68% / 0.2134 ± 0.0169
900
+ batch GP 60
901
+ 6.44% / 0.4444 ± 0.0487
902
+ 5.89% / 0.3021 ± 0.0362
903
+ 14.94% / 0.2291 ± 0.0226
904
+ batch BKI 30 (ours)
905
+ 3.05 / 0.4324 ± 0.0276
906
+ 2.93% / 0.2485 ± 0.0346
907
+ 7.67% / 0.2036 ± 0.0254
908
+ batch BKI 60 (ours)
909
+ 3.87% / 0.4407 ± 0.0384
910
+ 3.56% / 0.2731 ± 0.0356
911
+ 9.05% / 0.2065 ± 0.0229
912
+ GP-BO 30
913
+ 49.71% / 0.9435 ± 0.0609
914
+ 48.09% / 0.6083 ± 0.1121
915
+ 67.97% / 0.5203 ± 0.0554
916
+ GP-BO 60
917
+ 74.03% / 1.8265 ± 0.1189
918
+ 72.99% / 1.3558 ± 0.1190
919
+ 84.26% / 1.0528 ± 0.1124
920
+ BKI-BO 30 (ours)
921
+ 39% / 0.7518 ± 0.0683
922
+ 39.14% / 0.514 ± 0.0966 30
923
+ 54.03% / 0.3903 ± 0.1175
924
+ BKI-BO 60 (ours)
925
+ 53.74% / 0.9952 ± 0.1061
926
+ 54.31% / 0.7363 ± 0.1186
927
+ 62.45% / 0.4955 ± 0.1775
928
+ Note: Time cost of inference per step (in percentage) / Total time cost of exploration per step of each method (in sec.)
929
+ 0
930
+ 100
931
+ 200
932
+ 300
933
+ 400
934
+ 500
935
+ -5
936
+ 0
937
+ 5
938
+ MI Error (bits)
939
+ BKI prediction
940
+ GP prediction
941
+ Fig. 5.
942
+ A challenging example of MI prediction error comparison using
943
+ BKI and GP methods trained with fewer samples in a randomly selected
944
+ exploration step.
945
+ (a) Exploration trajectory
946
+ 50
947
+ 100
948
+ 150
949
+ 200
950
+ 250
951
+ X(grid)
952
+ 20
953
+ 40
954
+ 60
955
+ 80
956
+ 100
957
+ Y(grid)
958
+ 0
959
+ 0.2
960
+ 0.4
961
+ 0.6
962
+ (b) MI surface
963
+ Fig. 6.
964
+ An example of BKI-based robot exploration in the large cluttered
965
+ Seattle map [32]. White square: start point; White star: end point.
966
+ Among the 10 methods, the basic NBO methods have the
967
+ most expensive time consumption (more than about 8∼50
968
+ times to BKI and GP methods) per step, while other methods
969
+ based on GP and BKI cost much less time, showing the
970
+ efficiency of Bayesian optimization-based approaches. We
971
+ can further analyze these results from 2 aspects of view.
972
+ From the top row to the bottom, our BKI-based methods get
973
+ better time efficiency performance of decision-making and
974
+ inference than the corresponding GP-based ones in all maps
975
+ 0
976
+ 20
977
+ 40
978
+ 60
979
+ 80
980
+ 100
981
+ Steps
982
+ 1.22
983
+ 1.23
984
+ 1.24
985
+ 1.25
986
+ 1.26
987
+ 1.27
988
+ 1.28
989
+ 1.29
990
+ 1.3
991
+ Map entropy (bits)
992
+ 104
993
+ NBO 30
994
+ NBO 60
995
+ GP-BO 30
996
+ GP-BO 60
997
+ BKI-BO 30
998
+ BKI-BO 60
999
+ (a) Map entropy rate
1000
+ 0
1001
+ 20
1002
+ 40
1003
+ 60
1004
+ 80
1005
+ 100
1006
+ Steps
1007
+ 0
1008
+ 0.02
1009
+ 0.04
1010
+ 0.06
1011
+ 0.08
1012
+ 0.1
1013
+ 0.12
1014
+ 0.14
1015
+ 0.16
1016
+ Coverage
1017
+ NBO 30
1018
+ NBO 60
1019
+ GP-BO 30
1020
+ GP-BO 60
1021
+ BKI-BO 30
1022
+ BKI-BO 60
1023
+ (b) Coverage rate
1024
+ Fig. 7.
1025
+ Map entropy and coverage results of the Seattle map results (batch
1026
+ methods omitted).
1027
+ when the number of samples increases, e.g. batch BKI 30/60
1028
+ vs batch GP 30/60 and BKI-BO 30/60 vs GP-BO 30/60. We
1029
+ also can observe that BKI methods run faster than GP ones
1030
+ when using more training epochs. BKI methods also bring
1031
+ significant time savings for exploration, such as decreasing
1032
+ by about 20% and 45% time compared with GP-BO 30 and
1033
+ GP-BO 60 respectively in the structured map.
1034
+ From the left column to the right, these above-mentioned
1035
+ differences get more distinct in unstructured and large clut-
1036
+ tered maps, e.g. the time costs per step of GP-BO 30 and GP-
1037
+ BO 60 decrease by about 25% and 53% in the Seattle map
1038
+ respectively, which also verifies our proposed BKI-based
1039
+ robot exploration methods can improve the time efficiency
1040
+ considerably without losing overall exploration performance
1041
+ compared with other methods.
1042
+
1043
+ 0.140H50
1044
+ 10C.C
1045
+ 0.2(grid
1046
+ 600.4800
1047
+ 150
1048
+ X(gri0.6
1049
+ 0.5100200
1050
+ d)0.725020VI. CONCLUSIONS
1051
+ This paper mainly contributed to a new efficient learning-
1052
+ based approach for information-theoretic robot exploration in
1053
+ unknown environments. In particular, a continuous informa-
1054
+ tion gain evaluation model for predicting the MI of numerous
1055
+ sampled robot actions is built by introducing the Bayesian
1056
+ kernel inference method. The time complexity of MI pre-
1057
+ diction is decreased to logarithm level in comparison with
1058
+ state-of-the-art methods. An objective function integrating
1059
+ the predicted MI and uncertainty is also designed to balance
1060
+ exploration and exploitation. The proposed method also
1061
+ gets verified under an autonomous exploration framework
1062
+ by extensive simulations of different scenes, which reveals
1063
+ our method outperforms the greedy-based and GP-based
1064
+ exploration methods overall in efficiency without loss of
1065
+ exploration performance, especially in unstructured and large
1066
+ cluttered scenes. Future work mainly involves studying the
1067
+ exploration performance using different α values and kernels,
1068
+ as well as extending our method to 3D scenes.
1069
+ REFERENCES
1070
+ [1] H. Azp´urua, M. F. M. Campos, and D. G. Macharet, “Three-
1071
+ dimensional terrain aware autonomous exploration for subterranean
1072
+ and confined spaces,” in 2021 IEEE International Conference on
1073
+ Robotics and Automation (ICRA).
1074
+ IEEE, 2021, pp. 2443–2449.
1075
+ [2] J. Strader, K. Otsu, and A.-a. Agha-mohammadi, “Perception-aware
1076
+ autonomous mast motion planning for planetary exploration rovers,”
1077
+ Journal of Field Robotics, vol. 37, no. 5, pp. 812–829, 2020.
1078
+ [3] P. Stankiewicz, Y. T. Tan, and M. Kobilarov, “Adaptive sampling with
1079
+ an autonomous underwater vehicle in static marine environments,”
1080
+ Journal of Field Robotics, vol. 38, no. 4, pp. 572–597, 2021.
1081
+ [4] B. J. Julian, S. Karaman, and D. Rus, “On mutual information-
1082
+ based control of range sensing robots for mapping applications,” The
1083
+ International Journal of Robotics Research, vol. 33, no. 10, pp. 1375–
1084
+ 1392, 2014.
1085
+ [5] B. Charrow, S. Liu, V. Kumar, and N. Michael, “Information-theoretic
1086
+ mapping using cauchy-schwarz quadratic mutual information,” in 2015
1087
+ IEEE International Conference on Robotics and Automation (ICRA).
1088
+ IEEE, 2015, pp. 4791–4798.
1089
+ [6] Z. Zhang, T. Henderson, S. Karaman, and V. Sze, “FSMI: Fast
1090
+ computation of shannon mutual information for information-theoretic
1091
+ mapping,” The International Journal of Robotics Research, vol. 39,
1092
+ no. 9, pp. 1155–1177, 2020.
1093
+ [7] Y. Xu, R. Zheng, M. Liu, and S. Zhang, “CRMI: Confidence-rich
1094
+ mutual information for information-theoretic mapping,” IEEE Robotics
1095
+ and Automation Letters, vol. 6, no. 4, pp. 6434–6441, 2021.
1096
+ [8] Y. Xu, R. Zheng, S. Zhang, and M. Liu, “Confidence-rich localization
1097
+ and mapping based on particle filter for robotic exploration,” in 2022
1098
+ IEEE/RSJ International Conference on Intelligent Robots and Systems
1099
+ (IROS).
1100
+ IEEE, 2022, pp. 1–7.
1101
+ [9] G. A. Hollinger and G. S. Sukhatme, “Sampling-based robotic infor-
1102
+ mation gathering algorithms,” The International Journal of Robotics
1103
+ Research, vol. 33, no. 9, pp. 1271–1287, 2014.
1104
+ [10] M. G. Jadidi, J. V. Mir´o, and G. Dissanayake, “Sampling-based incre-
1105
+ mental information gathering with applications to robotic exploration
1106
+ and environmental monitoring,” The International Journal of Robotics
1107
+ Research, vol. 38, no. 6, pp. 658–685, 2019.
1108
+ [11] A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, and R. Siegwart,
1109
+ “Receding horizon” next-best-view” planner for 3d exploration,” in
1110
+ 2016 IEEE International Conference on Robotics and Automation
1111
+ (ICRA).
1112
+ IEEE, 2016, pp. 1462–1468.
1113
+ [12] B. Charrow, V. Kumar, and N. Michael, “Approximate representations
1114
+ for multi-robot control policies that maximize mutual information,”
1115
+ Autonomous Robots, vol. 37, no. 4, pp. 383–400, 2014.
1116
+ [13] K. Yang, S. Keat Gan, and S. Sukkarieh, “A gaussian process-based rrt
1117
+ planner for the exploration of an unknown and cluttered environment
1118
+ with a uav,” Advanced Robotics, vol. 27, no. 6, pp. 431–443, 2013.
1119
+ [14] B. Charrow, G. Kahn, S. Patil, S. Liu, K. Goldberg, P. Abbeel,
1120
+ N. Michael, and V. Kumar, “Information-theoretic planning with
1121
+ trajectory optimization for dense 3d mapping.” in Robotics: Science
1122
+ and Systems, vol. 11, 2015, pp. 3–12.
1123
+ [15] R. Marchant and F. Ramos, “Bayesian optimisation for informative
1124
+ continuous path planning,” in 2014 IEEE International Conference on
1125
+ Robotics and Automation (ICRA).
1126
+ IEEE, 2014, pp. 6136–6143.
1127
+ [16] R. Oliveira, L. Ott, V. Guizilini, and F. Ramos, “Bayesian optimisation
1128
+ for safe navigation under localisation uncertainty,” in International
1129
+ Symposium of Robotics Research.
1130
+ Springer, 2020, pp. 489–504.
1131
+ [17] G. Francis, L. Ott, R. Marchant, and F. Ramos, “Occupancy map
1132
+ building through bayesian exploration,” The International Journal of
1133
+ Robotics Research, vol. 38, no. 7, pp. 769–792, 2019.
1134
+ [18] S. Bai, J. Wang, K. Doherty, and B. Englot, “Inference-enabled
1135
+ information-theoretic exploration of continuous action spaces,” in
1136
+ International Symposium of Robotics Research.
1137
+ Springer, 2015, pp.
1138
+ 419–433.
1139
+ [19] S. Bai, J. Wang, F. Chen, and B. Englot, “Information-theoretic ex-
1140
+ ploration with bayesian optimization,” in 2016 IEEE/RSJ International
1141
+ Conference on Intelligent Robots and Systems (IROS).
1142
+ IEEE, 2016,
1143
+ pp. 1816–1822.
1144
+ [20] S. Bai, F. Chen, and B. Englot, “Toward autonomous mapping and
1145
+ exploration for mobile robots through deep supervised learning,” in
1146
+ 2017 IEEE/RSJ International Conference on Intelligent Robots and
1147
+ Systems (IROS).
1148
+ IEEE, 2017, pp. 2379–2384.
1149
+ [21] F. Chen, J. D. Martin, Y. Huang, J. Wang, and B. Englot, “Autonomous
1150
+ exploration under uncertainty via deep reinforcement learning on
1151
+ graphs,” in 2020 IEEE/RSJ International Conference on Intelligent
1152
+ Robots and Systems (IROS).
1153
+ IEEE, 2020, pp. 6140–6147.
1154
+ [22] T. Wang, R. Liao, J. Ba, and S. Fidler, “Nervenet: Learning structured
1155
+ policy with graph neural networks,” in 2018 International Conference
1156
+ on Learning Representations (ICLR), 2018.
1157
+ [23] W. R. Vega-Brown, M. Doniec, and N. G. Roy, “Nonparametric
1158
+ bayesian inference on multivariate exponential families,” Advances in
1159
+ Neural Information Processing Systems, vol. 27, 2014.
1160
+ [24] V. Peretroukhin, W. Vega-Brown, N. Roy, and J. Kelly, “PROBE-GK:
1161
+ Predictive robust estimation using generalized kernels,” in 2016 IEEE
1162
+ International Conference on Robotics and Automation (ICRA), 2016,
1163
+ pp. 817–824.
1164
+ [25] C. Richter, W. Vega-Brown, and N. Roy, “Bayesian learning for safe
1165
+ high-speed navigation in unknown environments,” in International
1166
+ Symposium on Robotics Research.
1167
+ Springer, 2015, pp. 325–341.
1168
+ [26] T. Shan, J. Wang, B. Englot, and K. Doherty, “Bayesian generalized
1169
+ kernel inference for terrain traversability mapping,” in Conference on
1170
+ Robot Learning.
1171
+ PMLR, 2018, pp. 829–838.
1172
+ [27] K. Doherty, T. Shan, J. Wang, and B. Englot, “Learning-aided 3-d
1173
+ occupancy mapping with bayesian generalized kernel inference,” IEEE
1174
+ Transactions on Robotics, vol. 35, no. 4, pp. 953–966, 2019.
1175
+ [28] L. Gan, R. Zhang, J. W. Grizzle, R. M. Eustice, and M. Ghaffari,
1176
+ “Bayesian spatial kernel smoothing for scalable dense semantic map-
1177
+ ping,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 790–
1178
+ 797, 2020.
1179
+ [29] S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics. MIT Press,
1180
+ 2005.
1181
+ [30] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for
1182
+ Machine Learning.
1183
+ The MIT Press, 2005.
1184
+ [31] S. T. O’Callaghan and F. T. Ramos, “Gaussian process occupancy
1185
+ maps,” The International Journal of Robotics Research, vol. 31, no. 1,
1186
+ pp. 42–62, 2012.
1187
+ [32] A. Howard and N. Roy, “The robotics data set repository (radish),”
1188
+ 2003. [Online]. Available: http://radish.sourceforge.net/
1189
+
5NAyT4oBgHgl3EQfpPjF/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
6tE4T4oBgHgl3EQfCAsz/content/tmp_files/2301.04856v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
6tE4T4oBgHgl3EQfCAsz/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
79A0T4oBgHgl3EQfOf-b/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1e4ff6a6f17b8faf51eb35dae9349593157db9a2bec0b0bcfae856566c15c876
3
+ size 3145773
7tE1T4oBgHgl3EQfBwI8/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f52fb04a50790f61d400007e7e0b3b1284416965b9bda347ecd5d55fb4360b96
3
+ size 102289
7tE2T4oBgHgl3EQfPgY3/content/tmp_files/2301.03759v1.pdf.txt ADDED
@@ -0,0 +1,1407 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Title: Strain-programmable van der Waals magnetic tunnel junctions
2
+ Authors: John Cenker1, Dmitry Ovchinnikov1, Harvey Yang1, Daniel G. Chica2, Catherine Zhu1,
3
+ Jiaqi Cai1, Geoffrey Diederich1,3, Zhaoyu Liu1, Xiaoyang Zhu2, Xavier Roy2, Ting Cao4,
4
+ Matthew W. Daniels5, Jiun-Haw Chu1, Di Xiao4,1, Xiaodong Xu1,4,*
5
+ 1 Department of Physics, University of Washington, Seattle, Washington 98195, USA
6
+ 2 Department of Chemistry, Columbia University, New York, NY 10027 USA
7
+ 3 Intelligence Community Postdoctoral Research Fellowship Program, University of Washington,
8
+ Seattle, WA, USA
9
+ 4 Department of Materials Science and Engineering, University of Washington, Seattle,
10
+ Washington 98195, USA
11
+ 5 Physical Measurement Laboratory, National Institute of Standards and Technology,
12
+ Gaithersburg, MD, 20899, USA
13
+
14
+ *Corresponding author’s email: xuxd@uw.edu
15
+
16
+
17
+ Abstract: The magnetic tunnel junction (MTJ) is a backbone device for spintronics.
18
+ Realizing next generation energy efficient MTJs will require operating mechanisms beyond
19
+ the standard means of applying magnetic fields or large electrical currents. Here, we
20
+ demonstrate a new concept for programmable MTJ operation via strain control of the
21
+ magnetic states of CrSBr, a layered antiferromagnetic semiconductor used as the tunnel
22
+ barrier. Switching the CrSBr from antiferromagnetic to ferromagnetic order generates a
23
+ giant tunneling magnetoresistance ratio without external magnetic field at temperatures up
24
+ to ≈ 140 K. When the static strain is set near the phase transition, applying small strain
25
+ pulses leads to active flipping of layer magnetization with controlled layer number and thus
26
+ magnetoresistance states. Further, finely adjusting the static strain to a critical value turns
27
+ on stochastic switching between metastable states, with a strain-tunable sigmoidal response
28
+ curve akin to the stochastic binary neuron. Our results highlight the potential of strain-
29
+ programmable van der Waals MTJs towards spintronic applications, such as magnetic
30
+ memory, random number generation, and probabilistic and neuromorphic computing.
31
+
32
+
33
+
34
+
35
+
36
+
37
+
38
+ Main Text:
39
+
40
+ The control and readout of discrete magnetic states lies at the foundation of the fields of
41
+ spintronics and modern information storage1-4. Standard spintronic devices utilize the spin filtering
42
+ phenomenon, where spin-selective transport processes, such as electron tunneling through
43
+ magnetic layers, create spin polarization and magnetoresistance5-10. Controlling the energetics and
44
+ stability of the magnets in such devices, known as magnetic tunnel junctions (MTJ), has enabled
45
+ many important technological advancements. For instance, switching the orientation of the
46
+ magnets from anti-parallel (AP) to parallel (P) in stable MTJs results in large changes to the
47
+ tunneling magnetoresistance (TMR). This behavior is the conceptual basis for magnetic random-
48
+ access memory (MRAM). On the other hand, when the magnetic layers are thinned so that the
49
+ energy difference between P and AP states is small, the magnetic order becomes unstable and
50
+ stochastic switching between the two states is observed11-15. Such stochastic MTJs can serve as
51
+ probabilistic bits (p-bits), the fundamental building blocks for the emerging fields of probabilistic
52
+ and neuromorphic computing11,16. Despite the great successes of conventional MTJs in both
53
+ conventional and probabilistic computing schemes, writing the magnetic memory bits in current
54
+ MRAM schemes tends to rely on energy-intensive means such as the application of large magnetic
55
+ fields or currents17. Moreover, since the stability of the MTJ is fixed by the growth thickness, it is
56
+ difficult to switch from stable MRAM operation to unstable p-bit functionality in the same device.
57
+
58
+ The recent discovery18 of a reversible strain-induced magnetic phase transition in the air
59
+ stable A-type layered antiferromagnetic (AFM) semiconductor CrSBr could offer both a new
60
+ material platform and operating principle for controlling atomically thin MTJs. The A-type AFM
61
+ configuration consists of van der Waals (vdW) layers with intralayer ferromagnetic (FM) order
62
+ and interlayer AFM coupling along the stacking direction, forming intrinsic spin filters that can
63
+ generate exceptionally large TMR19-22. These previous works have demonstrated that applying an
64
+ external magnetic field to A-type antiferromagnets with weak interlayer exchange switches the
65
+ magnetic state from the AFM, high resistance configuration to intermediate states with layer-
66
+ dependent interlayer coupling, and then finally to a low resistance, field-induced FM state (Fig.
67
+ 1a). In comparison to the previously studied devices which require continuous application of
68
+ magnetic field to control the magnetic states, strain could provide an exceptionally energy-efficient
69
+ operating mechanism as it requires essentially no current. Moreover, the fine, continuous, and
70
+ reversible tuning of the interlayer exchange could enable unprecedented control of the layer-
71
+ dependent magnetic structure.
72
+ Here, we demonstrate a strain-controlled vdW MTJ with programmable magneto-
73
+ resistance states and stochastic switching, charting a path towards new memory and computing
74
+ technologies. The schematic for our strain device is shown in Figure 1b. The vdW MTJ
75
+ heterostructure is composed of a CrSBr tunnel barrier sandwiched between two narrow graphite
76
+ contacts. The whole MTJ is fixed to a stretchable polyimide substrate by a gold clamp with a small
77
+ (≈ 5 µm) window around the junction (Methods). This design ensures a highly efficient strain
78
+ transfer when the polyimide substrate is stretched by a home-built piezoelectric strain cell18,23,
79
+ while also allowing for optical spectroscopy measurements of the junction region. The strain is
80
+ applied along the crystallographic a axis for consistency with previous experiments18. The data in
81
+ the main text is taken on a MTJ with an ≈ 11 nm tunnel barrier, but the technique is compatible
82
+ with CrSBr flakes of any thickness.
83
+
84
+ Figure 1c shows the tunneling magnetoresistance as a function of magnetic field (µ0H)
85
+ applied along the c axis. In the low strain condition with piezo voltage (Vp) of -5 V, CrSBr is in
86
+ the AFM state at µ0H = 0 T. As |µ0H| increases, the spins cant from the AFM configuration,
87
+ gradually increasing the conductivity of the MTJ until it reaches the field-induced FM state with
88
+ |µ0H| > 1 T. This behavior is consistent with the in-plane A-type layered AFM order in CrSBr24.
89
+ We note that the saturating field is lower than standard exfoliated CrSBr samples due to a built-in
90
+ strain which we determine to be ≈ 0.9 % from the Raman spectra (Methods, and Extended Data
91
+ Fig. 1). Using the difference in resistance between the FM (Rp) and AFM (Rap) states, we find the
92
+ tunneling magnetoresistance ratio to be TMR (%) =
93
+ ������
94
+ ��
95
+ ≈ 3100 %, on par with other 2D A-
96
+ type AFM tunnel junctions19-21, albeit at much higher operating temperature.
97
+ Strain Switching MTJ
98
+ When the piezo voltage is increased, the TMR decreases dramatically (Fig. 2a).
99
+ Furthermore, the shape of the tunneling magnetoresistance curves evolves from a giant, purely
100
+ negative magnetoresistance (i.e., decreasing resistance with increasing field) at low strain to small
101
+ positive MR at high strain (Extended Data Fig. 2), with complex, hysteretic behavior in between,
102
+ e.g., the curve at 5 V in Fig. 2A. The large decrease in TMR and switching from negative to
103
+ positive magnetoresistance implies that the interlayer magnetic coupling is switched from AFM to
104
+ FM at large strain. This picture is confirmed by comparison of the strain-dependent
105
+ photoluminescence (PL) with the magnetoresistance. The PL shows the characteristic red-shift
106
+ from the strain induced AFM to FM phase transition, as demonstrated in a previous report18, which
107
+ is concurrent with the large changes in tunneling magnetoresistance (Figs. 2b-c). The close
108
+ correspondence between the magneto-PL and tunneling magnetoresistance is a consequence of the
109
+ coupling of spin and charge in magnetic semiconductors, which forbids or allows interlayer
110
+ electronic hybridization and tunneling in the AFM and FM states, respectively.
111
+ In the low-strain state, the A-type AFM structure creates tunnel barriers composed of spin
112
+ filters with alternating spin orientation. In the FM state, however, the tunnel barrier is uniform for
113
+ all layers, i.e., all spin filters are aligned in the same direction. As a result, applying a saturating
114
+ magnetic field at low strains strongly enhances the tunneling current with respect to the AFM state
115
+ (top panel, Fig. 2d). At high strains, however, there is little difference between the zero- and high
116
+ magnetic field tunneling behavior, as expected for a FM tunnel barrier25 (Fig. 2d, bottom). The
117
+ combination of optical and tunneling measurements unambiguously prove that the strain-induced
118
+ AFM to FM phase transition is the cause of the large tunneling magnetoresistance switching,
119
+ excluding trivial origins such as contact failure during the straining process.
120
+ We realized strain switching of the MTJ at zero magnetic field. Figure 3a shows the
121
+ tunneling resistance as the piezo voltage is continually increased. At around 5 V, the sample
122
+ experiences a switch from AFM to FM states accompanied by a sharp drop in resistance. This
123
+ strain-induced phase transition generates a TMR ratio of ≈ 2700 %, comparable to the field-
124
+ induced TMR in the AFM state. When the tension is released, the resistance recovers to its original
125
+ value. The observed hysteresis between up and down strain sweeps is likely due to a combination
126
+ of the piezo stack hysteresis and hysteresis in the first-order magnetic phase transition itself. This
127
+ switching operation is robust over many cycles, with no obvious slipping or degradation over the
128
+ entire measurement (> 50 strain sweeps).
129
+ The strain-switching operation of the MTJ persists to much higher temperature than other
130
+ 2D MTJs19-22,25-27. Figure 3b shows tunneling magnetoresistance vs strain cycles at select
131
+
132
+ temperatures. At higher temperatures, the transition between low and high tunneling
133
+ magnetoresistance states becomes broader, but a large strain switching ratio is maintained. As
134
+ shown in Fig. 3c, the zero-field strain-induced TMR exceeds 10,000 % at 30 K and remains above
135
+ 100 % up to ≈ 140 K. Interestingly, a dome of positive magnetoresistance as a function of field
136
+ can still be induced by a large strain at 155 K, well above the Neel temperature of 132 K reported
137
+ in previous studies24,28,29 (Fig. 3d). A likely explanation is that the enhancement of the interlayer
138
+ FM exchange induces a long-range ordering of the previously reported intermediate FM (iFM)
139
+ phase where the individual layers are ferromagnetically ordered, but the interlayer coupling
140
+ remains paramagnetic29.
141
+ Strain programmable layer-dependent magnetism
142
+ An intriguing feature of the strain-dependent TMR sweeps is that there are multiple
143
+ resistance jumps during the AFM-FM phase transition, indicating the formation of multiple
144
+ magnetic domains in the junction area of about 500 x 500 nm2. These domains are also evident
145
+ from the complex, hysteretic behavior observed in the field dependent TMR measurements (Fig.
146
+ 2a, 5 V). Similar magnetic domain behavior is observed in both the nanoscale junction region and
147
+ across several microns of the sample in magneto-PL (Extended Data Fig. 3). These results suggest
148
+ the formation of vertical instead of lateral magnetic domains during the phase transition. The
149
+ domains may arise from small vertical strain gradients. Thus, near the critical strain of the magnetic
150
+ phase transition, the interlayer coupling can be FM for some layers and AFM for others. These
151
+ layer-wise magnetic domains could serve as individual magnetic memory states which can be
152
+ precisely manipulated by strain.
153
+ To explore active control of layer magnetization flipping, we set the static strain near the
154
+ phase transition and then apply strain pulses with a small and controllable amplitude VPAC (see
155
+ Figure 4a inset). Figure 4a shows the tunneling current over time as VPAC is increased from 5 mV
156
+ to 0.25 V. As the pulse reaches an amplitude of ≈ 24 mV, corresponding to a strain of only ≈
157
+ 0.0008 %, the amplitude of tunneling current pulses jumps into a distinctly stable state (left-most
158
+ purple arrow in Fig. 4a). This indicates the MTJ switches between two magnetization states with
159
+ the strain pulse actively flipping the magnetization direction of individual layers. Calculating the
160
+ gauge factor, GF =
161
+ ∆�
162
+
163
+ � , gives an exceptionally large value of ≈ 3500, among the largest value
164
+ reported in any system30,31.
165
+ By increasing the magnitude of the strain pulse, the number of layers whose magnetization
166
+ can be flipped also increases. This is evidenced by the additional distinct jumps in tunneling current
167
+ with increasing pulse amplitude (purple arrows in Fig. 4a). With a large enough strain pulse, the
168
+ static state current abruptly increases, indicating a change in the static magnetic configuration.
169
+ This behavior is completely different than what is observed in the purely FM or purely AFM states,
170
+ where increasing strain pulse magnitude only produces small, continuous changes at a gauge factor
171
+ three orders of magnitude smaller, and with no change in the static current (Extended Data Fig. 4).
172
+ Therefore, we conclude that the strain pulse switching observed in Fig. 4a arises from changing
173
+ the vertical domain structure of the mixed magnetic states. These results demonstrate that multiple
174
+ individual magnetic domains, including the static magnetic state, can be controlled by applying
175
+ extremely small strain pulses.
176
+ Stochastic domain switching
177
+
178
+
179
+ The demonstrated ability to switch the layer-dependent magnetization suggests that strain
180
+ can tune the MTJ into a regime where the AFM and FM interlayer couplings are extremely close
181
+ in energy. Starting from a stable magnetic domain structure, we increase the static strain, VPDC, by
182
+ 14 mV, as indicated by the red arrow in top panel of Fig. 4b. In such a condition, the tunneling
183
+ current proceeds to fluctuate between two values (Fig. 4b, bottom). By decreasing the piezo
184
+ voltage back to the original value (blue arrow in Fig. 4b, top), the tunneling current returns to a
185
+ stable value. The current fluctuations can be reliably turned on and off, as demonstrated. To our
186
+ knowledge, this is the first realization of p-bit type operation using a vdW MTJ. This functionality
187
+ is enabled by the unique ability of strain to finely and continuously tune the energy barrier between
188
+ parallel and anti-parallel spin configurations, enabling in-situ switching from stable, MRAM type
189
+ to stochastic, p-bit type domains (Fig. 4c).
190
+ By defining the lower current state as a 0 and the higher current state as a 1, we can convert
191
+ the data to a binary sequence and analyze how the statistics of the domain switching respond to
192
+ external control knobs, i.e. applied bias voltage and strain. We find that increasing the bias voltage
193
+ applied to the tunnel junction leads to a large increase in the switching rate (Fig. 4d). Intriguingly,
194
+ no switching is observed when a current of similar magnitude flows in the opposite direction
195
+ (Extended Data Fig. 5). This bias-polarity dependence implies that heating is not the origin of the
196
+ increased switching rate. Instead, the data suggests that the sample has an asymmetric vertical
197
+ magnetic domain structure which creates a difference in spin polarization and thus spin transfer
198
+ torque effects when the current is passed in opposite directions12 (Extended Data Fig. 5). Whether
199
+ such an asymmetric domain structure can give rise to exchange bias32, magnetic ratchet effect33,
200
+ and other spintronics physics within a single crystal is a fascinating direction for future studies.
201
+
202
+ The relatively high Neel temperature (TN =132 K) of CrSBr in comparison to other 2D A-
203
+ type AFMs creates opportunities for potential device applications operating above liquid nitrogen
204
+ temperature. Figure 4e shows the response function (ρ) of the MTJ as a function of the static piezo
205
+ voltage with a starting value near the strain-induced phase transition at 85 K. The response function
206
+ is calculated by converting the MTJ output to a binary sequence and calculating the average over
207
+ the entire time window. Therefore, a response function value of 0 or 1 indicates a stable magnetic
208
+ domain, while a value of 0.5 indicates equal fluctuations between the two stable states. The ability
209
+ to finely tune the response function should enable both random number generation at ρ = 0.5 and
210
+ a biased Bernoulli sequence at higher or lower values, which can be important for applications
211
+ dealing with Ising and probabilistic computing12. We further note that the applied bias voltage may
212
+ also be used to tune the response function by increasing or decreasing the switching rate,
213
+ potentially providing fine control near the edges of the sigmoidal curve, while also enabling
214
+ interaction between multiple p-bits. In principle, the two independent control parameters (strain
215
+ and bias voltage) could also offer independent tuning of the effective temperature and energy
216
+ landscape of the p-bit, thereby allowing direct stochastic annealing of a p-bit system. Such a
217
+ scheme could significantly reduce the circuit complexity required to realize a large-scale analog
218
+ p-bit annealer, though additional study is needed to establish the full mapping between our two-
219
+ dimensional voltage landscape and the statistical mechanical state space of the p-bit dynamics.
220
+ To test the stochasticity of our device, we analyze the switching data taken when ρ ≈ 0.5,
221
+ generating a binary sequence with near equal 1s and 0s, as shown in Figs. 4f-g. Since the lock-in
222
+ detection scheme reads the current much faster than the domain switching rate, we sample the raw
223
+ data at a frequency which is slower than the calculated switching rate to prevent non-random runs
224
+ of 1s and 0s (see discussion in Supplementary Information). We tested the data using the NIST
225
+
226
+ test suite (Fig. 4g) and by analyzing the rise and dwell time of the switching events, which shows
227
+ that the device spends equal amounts of time in the 0 and 1 state within the experimental error
228
+ (Supplementary Information). These analyses combined with their physical origin strongly
229
+ suggests that the metastable states switch stochastically, thereby acting as a random number
230
+ generator.
231
+
232
+ In conclusion, we have demonstrated that strained single crystal CrSBr offers a powerful
233
+ platform for realizing zero-field programmable spintronic devices down to the atomically thin limit
234
+ (Extended Data Fig. 6). Due to the versatile nature of vdW heterostructures, our results create a
235
+ new path for various other programmable 2D quantum devices. For instance, replacing the graphite
236
+ contacts with superconducting ones could enable field-free control of magnetic Josephson
237
+ junctions34-37 and superconducting diode effects38-40. Moreover, the ability to switch the layer-
238
+ dependent magnetization and vertical magnetic domain structure creates unprecedented
239
+ opportunities to precisely vary the length of the FM and AFM tunnel barriers in-situ without
240
+ significantly changing the overall thickness of the insulating CrSBr barrier layer. This capability
241
+ could provide a new platform for exploring exotic phenomena that have been proposed in
242
+ superconductor/ferromagnetic junctions with inhomogeneous magnetization such as spin triplet
243
+ correlations. More generally, our clamping and strain technique greatly expands the accessible
244
+ strain range for cryogenic transport experiments on 2D devices, which could enable exciting
245
+ discoveries on the emergent quantum phenomena in vdW heterostructures including moiré systems.
246
+ Methods
247
+ Device fabrication and strain application
248
+ To prepare the strain substrate, we first cut transparent 20 µm thick polyimide into strips and
249
+ epoxied them onto 2D flexure sample plates produced by Razorbill instruments† using Stycast
250
+ 2850 FT epoxy. The distance between the edge of the epoxy on either side of the gap was less than
251
+ 200 µm to enable large strains.
252
+ Bulk CrSBr crystals were grown by the same method detailed previously28. The bulk CrSBr and
253
+ graphite crystals were exfoliated onto PDMS substrates using standard methods and thin (~ 10 nm)
254
+ flakes were identified by optical contrast. The MTJs were then assembled through a dry transfer
255
+ technique with a stamp consisting of a polypropylene carbonate (PPC) film spin coated onto a
256
+ polydimethylsiloxane (PDMS) cylinder. The flakes were picked up in the following order before
257
+ being deposited onto the polyimide substrate: top graphite, CrSBr, bottom graphite. The long axis
258
+ of the CrSBr flake was aligned with the strain axis for consistency with the previous studies18.
259
+ After depositing the MTJ heterostructure, the window clamping pattern and electrical contacts to
260
+ the two graphite contacts were fabricated using standard electron beam lithography techniques
261
+ with a metal thickness of 7 and 70 nm Cr and Au, respectively. Then, the sample plate was screwed
262
+ into the same symmetric three-piezo strain cell used previously18,23 for strain experiments on bulk
263
+ crystals and our previous experiments on strained CrSBr.
264
+ To calibrate the strain during the experiment, we used the same Raman shift rate of the mode near
265
+ ~ 346 cm-1 that we determined in the previous study2. We found that there was a rather large built-
266
+ in strain of ~ 0.9 %, which is consistent with the small saturating field in the out-of-plane direction.
267
+
268
+ † Certain commercial processes and software are identified in this article to foster understanding. Such identification
269
+ does not imply recommendation or endorsement by the National Institute of Standards and Technology, nor does it
270
+ imply that the processes and software identified are necessarily the best available for the purpose.
271
+
272
+ The observation that the strain-induced phase transition occurs at negative piezo voltages at lower
273
+ temperature is consistent with a thermally induced built-in strain which increases with cooling.
274
+ Optical measurements:
275
+ Optical measurements were performed using a backscattering geometry in a closed-cycle helium
276
+ cryostat (Opticool by Quantum Design) with a nominal sample temperature of 60 K. An objective
277
+ lens focused 632.8 nm light from a He/Ne laser to a spot size of ~ 1 µm. For Raman measurements,
278
+ a laser power of 200 µW was used and the collected signal was dispersed using a 1800 mm-1
279
+ groove-density grating and detected by an LN-cooled charge-coupled device (CCD) with an
280
+ integration time of 210 seconds. BragGrateTM notch filters were used to filter out Rayleigh
281
+ scattering down to ~10 cm-1. A roughly linear background originating from weak polyimide
282
+ photoluminescence was subtracted to increase the accuracy of the fitting results. For
283
+ photoluminescence measurements, we used a laser power of 50 µW focused by the same objective.
284
+ The collected light was dispersed by a 600 mm-1 groove-density grating and detected by the same
285
+ CCD with a 20 second integration time.
286
+ Transport measurements:
287
+ Except for the data presented in Extended Data Fig. 6, the transport measurements were performed
288
+ in the same measurement conditions (Opticool by Quantum Design) as the optical ones, enabling
289
+ direct comparison between the observed phenomena. The data shown in Figures 1-3 and 4b are
290
+ taken using standard two terminal DC measurements with a Keithley 2450, while the rest of the
291
+ data in Figure 4 are taken using AC detection with a DC offset voltage applied by a Zurich
292
+ Instruments HF2 lock-in amplifier. The current was amplified by a current preamplifier (DL
293
+ Instruments; Model 1211) with a sensitivity of 1 V/10−6 A. For the switching data used in Fig. 4E-
294
+ F and the stochasticity analysis, a time constant of 5.082 ms with a fourth-order filter was used,
295
+ which was found to give the best time resolution while maintaining a high signal to noise ratio.
296
+ The current was amplified by a current preamplifier (DL Instruments; Model 1211) with a
297
+ sensitivity of 1 V/10−6 A.
298
+ The 6L device in Extended Data Fig. 6 was measured in a PPMS DynaCool cryostat by Quantum
299
+ Design. The data in Fig. S6a-c were taken using the same AC detection scheme, but with an SR860
300
+ lock-in amplifier. The switching data in Fig. S6d-e were obtained using a constant current
301
+ measurement scheme, which was achieved by putting a 100 MΩ resistor in series with the device.
302
+ The resistance signal was then pre-amplified by the differential-ended mode of SR560 with 20
303
+ times amplification.
304
+ References:
305
+ 1
306
+ Baibich, M. N. et al. Giant Magnetoresistance of (001)Fe/(001)Cr Magnetic
307
+ Superlattices. Physical Review Letters 61, 2472-2475 (1988).
308
+ 2
309
+ Binasch, G., Grünberg, P., Saurenbach, F. & Zinn, W. Enhanced magnetoresistance in
310
+ layered magnetic structures with antiferromagnetic interlayer exchange. Physical Review
311
+ B 39, 4828-4830 (1989).
312
+ 3
313
+ Žutić, I., Fabian, J. & Das Sarma, S. Spintronics: Fundamentals and applications. Reviews
314
+ of Modern Physics 76, 323-410 (2004).
315
+ 4
316
+ Dieny, B. et al. Giant magnetoresistive in soft ferromagnetic multilayers. Physical
317
+ Review B 43, 1297-1300 (1991).
318
+
319
+ 5
320
+ Julliere, M. Tunneling between ferromagnetic films. Physics letters. 54, 225 (1975).
321
+ 6
322
+ Moodera, J. S., Kinder, L. R., Wong, T. M. & Meservey, R. Large Magnetoresistance at
323
+ Room Temperature in Ferromagnetic Thin Film Tunnel Junctions. Physical Review
324
+ Letters 74, 3273-3276 (1995).
325
+ 7
326
+ Miyazaki, T., Tezuka, N. Giant magnetic tunneling effect in Fe/Al2O3/Fe junction.
327
+ Journal of magnetism and magnetic materials 139, L231 (1995).
328
+ 8
329
+ Yuasa, S., Nagahama, T., Fukushima, A., Suzuki, Y. & Ando, K. Giant room-
330
+ temperature magnetoresistance in single-crystal Fe/MgO/Fe magnetic tunnel junctions.
331
+ Nature Materials 3, 868-871 (2004).
332
+ 9
333
+ Parkin, S. S. P. et al. Giant tunnelling magnetoresistance at room temperature with MgO
334
+ (100) tunnel barriers. Nature Materials 3, 862-867 (2004).
335
+ 10
336
+ Ikeda, S. et al. Tunnel magnetoresistance of 604% at 300K by suppression of Ta
337
+ diffusion in CoFeB∕MgO∕CoFeB pseudo-spin-valves annealed at high temperature.
338
+ Applied Physics Letters 93, 082508 (2008).
339
+ 11
340
+ Borders, W. A. et al. Integer factorization using stochastic magnetic tunnel junctions.
341
+ Nature 573, 390-393 (2019).
342
+ 12
343
+ Safranski, C. et al. Demonstration of Nanosecond Operation in Stochastic Magnetic
344
+ Tunnel Junctions. Nano Letters 21, 2040-2045 (2021).
345
+ 13
346
+ Bapna, M. et al. Magnetostatic effects on switching in small magnetic tunnel junctions.
347
+ Applied Physics Letters 108, 022406 (2016).
348
+ 14
349
+ Mizrahi, A. et al. Neural-like computing with populations of superparamagnetic basis
350
+ functions. Nature Communications 9 (2018).
351
+ 15
352
+ Rippard, W., Heindl, R., Pufall, M., Russek, S. & Kos, A. Thermal relaxation rates of
353
+ magnetic nanoparticles in the presence of magnetic fields and spin-transfer effects.
354
+ Physical Review B 84 (2011).
355
+ 16
356
+ Camsari, K. Y., Sutton, B. M. & Datta, S. p-bits for probabilistic spin logic. Applied
357
+ Physics Reviews 6, 011305 (2019).
358
+ 17
359
+ Bhatti, S. et al. Spintronics based random access memory: a review. Materials Today 20,
360
+ 530-548 (2017).
361
+ 18
362
+ Cenker, J. et al. Reversible strain-induced magnetic phase transition in a van der Waals
363
+ magnet. Nature Nanotechnology 17, 256-261 (2022).
364
+ 19
365
+ Song, T. et al. Giant tunneling magnetoresistance in spin-filter van der Waals
366
+ heterostructures. Science 360, 1214-1218 (2018).
367
+ 20
368
+ Wang, Z. et al. Very large tunneling magnetoresistance in layered magnetic
369
+ semiconductor CrI3. Nature Communications 9 (2018).
370
+ 21
371
+ Kim, H. H. et al. One Million Percent Tunnel Magnetoresistance in a Magnetic van der
372
+ Waals Heterostructure. Nano Letters 18, 4885-4890 (2018).
373
+ 22
374
+ Klein, D. R. et al. Probing magnetism in 2D van der Waals crystalline insulators via
375
+ electron tunneling. Science 360, 1218-1222 (2018).
376
+ 23
377
+ Hicks, C. W., Barber, M. E., Edkins, S. D., Brodsky, D. O. & Mackenzie, A. P.
378
+ Piezoelectric-based apparatus for strain tuning. Review of Scientific Instruments 85,
379
+ 065003 (2014).
380
+ 24
381
+ Telford, E. J. et al. Layered Antiferromagnetism Induces Large Negative
382
+ Magnetoresistance in the van der Waals Semiconductor CrSBr. Advanced Materials 32,
383
+ 2003240 (2020).
384
+
385
+ 25
386
+ Wang, Z. et al. Magnetization dependent tunneling conductance of ferromagnetic
387
+ barriers. Nature Communications 12 (2021).
388
+ 26
389
+ Cai, X. et al. Atomically Thin CrCl3: An In-Plane Layered Antiferromagnetic Insulator.
390
+ Nano Letters 19, 3993-3998 (2019).
391
+ 27
392
+ Wang, Z. et al. Determining the phase diagram of atomically thin layered
393
+ antiferromagnet CrCl3. Nature Nanotechnology 14, 1116-1122 (2019).
394
+ 28
395
+ Scheie, A. et al. Spin Waves and Magnetic Exchange Hamiltonian in CrSBr. Advanced
396
+ Science 9, 2202467 (2022).
397
+ 29
398
+ Lee, K. et al. Magnetic Order and Symmetry in the 2D Semiconductor CrSBr. (2020).
399
+ 30
400
+ Wu, J. M. et al. Ultrahigh Sensitive Piezotronic Strain Sensors Based on a ZnSnO3
401
+ Nanowire/Microwire. ACS Nano 6, 4369-4374 (2012).
402
+ 31
403
+ Yan, W. et al. Giant gauge factor of Van der Waals material based strain sensors. Nature
404
+ Communications 12 (2021).
405
+ 32
406
+ Meiklejohn, W. H. & Bean, C. P. New Magnetic Anisotropy. Physical Review 102, 1413-
407
+ 1414 (1956).
408
+ 33
409
+ Lavrijsen, R. et al. Magnetic ratchet for three-dimensional spintronic memory and logic.
410
+ Nature 493, 647-650 (2013).
411
+ 34
412
+ Gingrich, E. C. et al. Controllable 0–π Josephson junctions containing a ferromagnetic
413
+ spin valve. Nature Physics 12, 564-567 (2016).
414
+ 35
415
+ Ai, L. et al. Van der Waals ferromagnetic Josephson junctions. Nature Communications
416
+ 12 (2021).
417
+ 36
418
+ Idzuchi, H. et al. Unconventional supercurrent phase in Ising superconductor Josephson
419
+ junction with atomically thin magnetic insulator. Nature Communications 12 (2021).
420
+ 37
421
+ Kang, K. et al. van der Waals π Josephson Junctions. Nano Letters (2022).
422
+ 38
423
+ Narita, H. et al. Field-free superconducting diode effect in noncentrosymmetric
424
+ superconductor/ferromagnet multilayers. Nature Nanotechnology (2022).
425
+ 39
426
+ Ando, F. et al. Observation of superconducting diode effect. Nature 584, 373-376 (2020).
427
+ 40
428
+ Wu, H. et al. The field-free Josephson diode in a van der Waals heterostructure. Nature
429
+ 604, 653-656 (2022).
430
+
431
+ Acknowledgements: We thank Xuetao Ma and Yen-Cheng Kung for fabrication advice, G.C.
432
+ Adam, W.A. Borders, and J. J. Mcclelland for proofreading the paper, and John Stroud and
433
+ Heonjoon Park for their help during the initial stages of the project. The strain controlled optical
434
+ measurement is mainly supported by DE-SC0018171. The strain-controlled tunneling experiment
435
+ is mainly supported by Air Force Office of Scientific Research (AFOSR) Multidisciplinary
436
+ University Research Initiative (MURI) program, grant no. FA9550- 19-1-0390. CrSBr crystal
437
+ synthesis is supported by the Center on Programmable Quantum Materials, an Energy Frontier
438
+ Research Center funded by the U.S. Department of Energy (DOE), Office of Science, Basic Energy
439
+ Sciences (BES), under award DE-SC0019443. DGC is supported by the Columbia MRSEC on
440
+ Precision-Assembled Quantum Materials (PAQM) (DMR-2011738). XX acknowledges support
441
+ from the State of Washington funded Clean Energy Institute and from the Boeing Distinguished
442
+ Professorship in Physics. JC acknowledges the Graduate Fellowship from Clean Energy Institute
443
+ funded by the State of Washington. ZL and JHC acknowledge the support of the David and Lucile
444
+ Packard Foundation. This research was supported by an appointment to the Intelligence
445
+ Community Postdoctoral Research Fellowship Program at University of Washington,
446
+
447
+ administered by Oak Ridge Institute for Science and Education through an interagency agreement
448
+ between the U.S. Department of Energy and the Office of the Director of National Intelligence.
449
+
450
+ Author contributions: XX and John C conceived the project. John C performed the optical and
451
+ transport measurements with help from Jiaqi C and GD. DO supervised transport measurements
452
+ and contributed to fabrication development. John C fabricated the samples with assistance from
453
+ HY and ZL. John C, DO, TC, JHC, DX, and XX analyzed the data and interpreted the results. TC,
454
+ MWD and DX provided theoretical support. DGC grew the CrSBr crystals with supervision from
455
+ XR and XYZ. John C and XX wrote the manuscript with input from all authors. All authors
456
+ discussed the results.
457
+ Competing interests: John C and XX have applied for a patent based on this work.
458
+ Data availability: The datasets generated during and/or analyzed during this study are available
459
+ from the corresponding author upon reasonable request.
460
+
461
+
462
+ Figures:
463
+
464
+
465
+ Figure 1 | Straintronic van der Waals magnetic tunnel junction. a, Schematic of the
466
+ magnetic state evolution of the CrSBr tunnel barrier with the application of either magnetic
467
+ fields along the easy b axis or in-plane uniaxial strain. The changing magnetic configuration
468
+ creates different resistance states when bias is applied between the graphite contacts (grey).
469
+ The red and blue arrows denote the spin direction within each layer. b, Schematic of
470
+ straintronic MTJ consisting of graphite contacts sandwiching a CrSBr tunnel barrier (blue).
471
+ The whole device is fixed by gold clamps to a flexible polyimide substrate (purple) which is
472
+ then strained. c, Magnetic field dependence of a MTJ using an ≈ 11 nm CrSBr tunnel barrier
473
+ (optical image inset, scale bar 3 µm) at a temperature of 60K. The device is slightly strained
474
+ but remains in the AFM state at zero magnetic field. Magnetic field is applied along the hard c
475
+ axis, leading to spin canting (inset arrows).
476
+
477
+
478
+ Resistance
479
+ .H
480
+ YYY
481
+ KKK
482
+ 0.4
483
+ 0.2
484
+ 0
485
+ -2
486
+ -1
487
+ 0
488
+ 1
489
+ 2
490
+ Low resistance
491
+ μ.H (T)V
492
+ = 0.5 V
493
+ NN
494
+ Bias
495
+ T = 60 K
496
+ 0.8
497
+ V. = -5 V
498
+ C
499
+ pa
500
+ b
501
+ High resistance
502
+ Top Gr
503
+ CrSBr
504
+ Bot. Gr
505
+
506
+
507
+ Figure 2 | Strain switchable magnetic tunnel junctions. a, Magnetoresistance sweeps at
508
+ select piezo voltages with a fixed bias voltage across the MTJ of VB = 0.5 V. The sweeps are
509
+ offset for clarity. b, Full strain-dependent tunneling magnetoresistance with the magnetic field
510
+ swept from positive to negative. c, Strain dependent photoluminescence intensity plot. The
511
+ beam spot was kept fixed on the junction region while the strain was continuously swept. d-e,
512
+ Bias dependent tunneling current with magnetic fields of 0 T (blue) and 3 T (green) applied in
513
+ the low strain (d) and high strain (e) states. The magnetic state for each curve is depicted in the
514
+ inset. All measurements were performed at a temperature of 60 K.
515
+
516
+
517
+ Resista
518
+ 1.38
519
+ -5 V
520
+ (eV)
521
+ 1
522
+ Energy (
523
+ (nA)
524
+ OV
525
+ 1.34
526
+ 0
527
+ 5V
528
+ 0.2
529
+ Photon
530
+ 15 V
531
+ PL (a.u.)
532
+ -1
533
+ 1.3
534
+ V,= 25 V
535
+ 25 V
536
+ 600
537
+ 2000
538
+ 0
539
+ -2
540
+ -1
541
+ 0
542
+ 1
543
+ 2
544
+ 0
545
+ 5
546
+ 10
547
+ 15
548
+ 20
549
+ 25
550
+ -0.4
551
+ -0.2
552
+ 0
553
+ 0.2
554
+ 0.4
555
+ μ.H (T)
556
+ Piezo Voltage (V)
557
+ Bias Voltage (V)a
558
+ b
559
+ d
560
+ 2
561
+ -OT
562
+ -3 T
563
+ 2
564
+ 1
565
+ ↑个个
566
+ E
567
+ (nA)
568
+ 1.0
569
+ 0
570
+ 0
571
+ (GΩ)
572
+ -2
573
+ -1
574
+ R (GΩ)
575
+ V,= -5 V
576
+ 0.026
577
+ 0.89
578
+ ince
579
+ 2
580
+ 0.6
581
+ c
582
+ e
583
+
584
+
585
+ Figure 3 | Temperature dependent zero-field tunneling resistance switching. a, Tunneling
586
+ resistance as a function of piezo voltage. A large TMR change of ≈ 2700 % is observed between
587
+ the low and high strain states at 60 K. The change in magnetic state from AFM to FM interlayer
588
+ coupling is depicted by the inset spin diagram. b, Piezo-voltage-dependent tunneling resistance
589
+ at select temperatures from 30 K to 149 K. v, Temperature dependence of the tunneling
590
+ magnetoresistance ratio, defined as TMR (%) =
591
+ ������
592
+ ��
593
+ � 100. d, Magnetic-field dependent
594
+ tunneling resistance at 155 K in the low strain (blue) and high strain (red) states.
595
+
596
+
597
+ a
598
+ 10
599
+ d
600
+ 3.8
601
+ 4
602
+ (×105 (
603
+ VBias = 0.02 V
604
+ R
605
+ T = 155 K
606
+ T= 139 K
607
+ 6
608
+ 3.2
609
+ = 0.05 V
610
+ R
611
+ 2
612
+ -5 V
613
+ a
614
+ 2.6
615
+ (×105 (
616
+ R
617
+ 4
618
+ T = 149 K
619
+ 27.5 V
620
+ VBias = 0.03 V
621
+ 2
622
+ 0
623
+ R
624
+ 3
625
+ -5
626
+ 0
627
+ 10
628
+ 15
629
+ 20
630
+ -10
631
+ 0
632
+ 10
633
+ 20
634
+ -2
635
+ -1
636
+ 0
637
+ 1
638
+ 2
639
+ Piezo Voltage (V)
640
+ μ.H (T)
641
+ Piezo Voltage (V)a 10
642
+ b
643
+ C
644
+ 104
645
+ (×109 Q2)
646
+ T= 30 K
647
+ T= 60 K
648
+ (%)
649
+ = 0.7 V
650
+ μ.H= O T
651
+ Ratio
652
+ 8
653
+ R
654
+ 103
655
+ 0
656
+ =0.5V
657
+ TMR
658
+ a
659
+ (×107 (
660
+ 3
661
+ 6
662
+ T = 85 K
663
+ 102
664
+ 108 Q2)
665
+ V
666
+ = 0.35 V
667
+ 20
668
+ 60
669
+ 100
670
+ 140
671
+ R
672
+ Temperature (K)
673
+
674
+ Figure 4 | Strain control of multiple stable and stochastic layer-dependent magnetic
675
+ domains. a, Tunneling current over time as strain pulses of increasing amplitude are applied.
676
+ The inset shows the measurement scheme: a small pulse of amplitude VPAC is applied on top
677
+ of a static piezo voltage VPDC. The system is initialized by slowly increasing VPDC until the
678
+ magnetic phase transition starts to occur. As the pulse amplitude increases, the current
679
+ switching stabilizes into discrete states (denoted by the purple arrows). Additionally, the resting
680
+ current, i.e. ground state, can be changed by a sufficiently large pulse. b, Tunneling current
681
+ over time as the static piezo voltage, VPDC is increased (red arrow) and then decreased (blue
682
+ arrow) by .014 V. No strain pulse is applied. Bottom: Finer time resolution data of the domain
683
+ fluctuations observed in the top panel. c, Schematic of strain tuning between magnetic domains.
684
+ A sufficiently high pulse, VPAC, will flip between AFM and FM domains (left). The fine
685
+ adjustment of the static strain lowers the energy difference between AFM and FM domains,
686
+ creating a metastable state with stochastic domain switching (right). d, Bias dependence of the
687
+ switching rate in the metastable state. The piezo voltage is kept constant during the
688
+ measurement. Data from panels A-D are taken at 60 K. e, Response function of a sensitive
689
+ magnetic domain as a function of static piezo voltage at a temperature of 85 K. A value of
690
+ either 0 or 1 indicates a stable domain. f, Tunneling current (top) and converted binary
691
+ sequence (bottom) over time when the response function is near 0.5, indicating equal amount
692
+ of fluctuations between the parallel and antiparallel configuration. g, P-values returned by the
693
+ NIST random number test suite applied to the binary sequence from f. The black dashed line
694
+ indicates a p-value of .01, the threshold for passing the specific test. The sampling time was
695
+ .1760 seconds (see Supplementary Information).
696
+
697
+ Time (sec)
698
+ Time (sec)
699
+ c
700
+ e
701
+ 9
702
+ V
703
+ PAC
704
+ Value
705
+ 0.1
706
+ p
707
+ P
708
+ 8
709
+ T = 85 K
710
+ 0.01
711
+ 0
712
+ 0.02
713
+ 0.04
714
+ V
715
+ PDC
716
+ Stable
717
+ AV
718
+ (V)
719
+ Stochastic
720
+ PDC
721
+ TestTime (sec)
722
+ b
723
+ d
724
+ 48
725
+ 3
726
+ It (nA)
727
+ 4
728
+ (nA)
729
+ 2
730
+ 46
731
+ Current (
732
+ 500
733
+ 1500
734
+ 2500
735
+ 3500
736
+ 3.6
737
+ Switching
738
+ 6.5m
739
+ Binary
740
+ 6.4
741
+ 0
742
+ 0.54
743
+ 0.58
744
+ 0.62
745
+ 25
746
+ 0
747
+ 20
748
+ 40
749
+ 60
750
+ 80
751
+ 100a
752
+ V
753
+ PDC
754
+ 。= 250 mV
755
+ 6.4F
756
+ (nA)
757
+ PAC
758
+ 6.2
759
+ Current (
760
+ 9400
761
+ 9500
762
+ 9600
763
+ .=5mV
764
+ PAC
765
+ 6
766
+ 2000
767
+ 6000
768
+ 10000
769
+ 14000
770
+ 18000
771
+ 22000
772
+
773
+ 1
774
+
775
+ Extended Data for
776
+
777
+ Strain-programmable van der Waals magnetic tunnel junctions
778
+
779
+ Authors: John Cenker1, Dmitry Ovchinnikov1, Harvey Yang1, Daniel G. Chica2, Catherine Zhu1,
780
+ Jiaqi Cai1, Geoffrey Diederich1,3, Zhaoyu Liu1, Xiaoyang Zhu2, Xavier Roy2, Ting Cao4,
781
+ Matthew W. Daniels5, Jiun-Haw Chu1, Di Xiao4,1, Xiaodong Xu1,4,*
782
+
783
+
784
+ 1 Department of Physics, University of Washington, Seattle, Washington 98195, USA
785
+ 2 Department of Chemistry, Columbia University, New York, NY 10027 USA
786
+ 3 Intelligence Community Postdoctoral Research Fellowship Program, University of Washington,
787
+ Seattle, WA, USA
788
+ 4 Department of Materials Science and Engineering, University of Washington, Seattle,
789
+ Washington 98195, USA
790
+ 5 Physical Measurement Laboratory, National Institute of Standards and Technology,
791
+ Gaithersburg, MD, 20899, USA
792
+
793
+ *Correspondence to: xuxd@uw.edu
794
+
795
+
796
+
797
+
798
+
799
+
800
+ 2
801
+
802
+
803
+
804
+
805
+
806
+
807
+
808
+
809
+
810
+
811
+
812
+
813
+
814
+
815
+
816
+
817
+
818
+
819
+
820
+
821
+
822
+
823
+
824
+
825
+ Extended Data Fig. 1 | Calibration of strain through Raman spectroscopy. a, Raman scattering
826
+ from the P3 phonon taken on the tunnel junction region at a piezo voltage of 0 V. A linear
827
+ background originating from the polyimide photoluminescence is subtracted. The narrow linewidth
828
+ indicates a homogenous strain. b, Raman intensity plot as a function of piezo voltage. The beamspot
829
+ is kept on the junction as the piezo voltage is continually increased. c, Measured strain as a function
830
+ of the applied voltage to the strain cell. The strain is calculated by fitting the data from b with
831
+ Lorentzian fits and then comparing the peak position to the unstrained value of 346 cm-1 using a
832
+ strain shift rate of 4.2 cm-1/% as reported in previous studies. We found that there was a built-in
833
+ strain of ~ 0.9 % at the lowest piezo voltage used at this temperature.
834
+
835
+
836
+ Intens
837
+ V
838
+ Rama
839
+ S
840
+ 1.1
841
+ 336
842
+ 0
843
+ Intensity
844
+ 0
845
+ 300
846
+ 0.9
847
+ 332
848
+ 320
849
+ 340
850
+ 360
851
+ -5
852
+ 0
853
+ 5
854
+ 10
855
+ 15
856
+ 2025
857
+ -5
858
+ 0
859
+ 5
860
+ 1015
861
+ 25
862
+ Raman Shift (cm-1)
863
+ Piezo Voltage (V)
864
+ Piezo Voltage (V)a
865
+ b
866
+ c
867
+ 300
868
+ 1.7
869
+ 348
870
+ Shift (cm-1)
871
+ (counts)
872
+ 1.5
873
+ 344
874
+ 200
875
+ (%)
876
+ train
877
+ 1.3
878
+ 340
879
+
880
+ 3
881
+
882
+
883
+
884
+
885
+
886
+
887
+
888
+
889
+
890
+
891
+
892
+
893
+
894
+
895
+
896
+
897
+
898
+
899
+
900
+
901
+
902
+
903
+
904
+
905
+
906
+
907
+
908
+
909
+
910
+
911
+
912
+
913
+
914
+
915
+
916
+
917
+
918
+
919
+
920
+
921
+
922
+
923
+
924
+
925
+ Extended Data Fig. 2 | Magnetoresistance sweeps at select piezo voltages. a-d,
926
+ Magnetoresistance sweeps as the field is swept down (blue) and up (black) at select piezo
927
+ voltages through the strain-induced layered magnetization flipping. At low strain (a), large
928
+ negative magnetoresistance is observed, consistent with AFM order, while small positive
929
+ magnetoresistance is observed in the high strain induced FM state (d). In between, complex
930
+ and hysteretic magnetic domain behavior is observed.
931
+
932
+
933
+ 3.8
934
+ V = 10V
935
+ = 25 V
936
+ b
937
+ d
938
+ Resistance (× 107 ohm)
939
+ Resistance (× 107 ohm)
940
+ 6
941
+ 3.7
942
+ 5
943
+ 3.6
944
+ 4
945
+ 3.5
946
+ 3
947
+ 3.41
948
+ 2
949
+ -1
950
+ 0
951
+ 2
952
+ -2
953
+ -1
954
+ 0
955
+ μ。H(T)
956
+ μ。H(T)8
957
+ a
958
+ /=0V
959
+ 3.5
960
+ = 15V
961
+ Resistance (× 108 ohm)
962
+ 6
963
+ 3.4
964
+ 4
965
+ 3.3
966
+ -μ.H
967
+ +μ.H
968
+ 2
969
+ 3.2
970
+ 0
971
+ 3.1
972
+
973
+ 4
974
+
975
+
976
+
977
+
978
+
979
+
980
+
981
+
982
+
983
+
984
+
985
+
986
+
987
+
988
+
989
+
990
+
991
+
992
+
993
+
994
+
995
+
996
+
997
+
998
+
999
+
1000
+
1001
+
1002
+
1003
+
1004
+ Extended Data Fig. 3 | Magneto-photoluminescence mapping of magnetic domains. a-b,
1005
+ Comparison of tunneling magnetoresistance (a) and integrated intensity from magneto-
1006
+ photoluminescence (PL) (b) measurements at the same piezo voltage. The correlation of the
1007
+ curves highlights the connection of the interlayer magnetic coupling to both electronic
1008
+ tunneling and exciton luminescence. c, Optical image of the device with different spots labeled
1009
+ by different colors. d-g, Magneto-PL sweeps at each of the spots labeled in (c). The similarities
1010
+ between spots separated by several microns indicates the presence of vertical, rather than
1011
+ lateral, magnetic domains.
1012
+
1013
+ -1
1014
+ -0.5
1015
+ 0
1016
+ 0.5
1017
+ Energy (eV)
1018
+ 1.4
1019
+ g
1020
+ Energy (eV)
1021
+ 1.4
1022
+ μ.H (T)
1023
+ 1000
1024
+ 1030
1025
+ c
1026
+ PL (a.u.)
1027
+ PL (a.u.)
1028
+ 1.3
1029
+ Energy (eV)
1030
+ 1.4
1031
+ Energy (eV)
1032
+ 1.4
1033
+ 600
1034
+ 600
1035
+ 1.3
1036
+ 1.3
1037
+ -1
1038
+ 0
1039
+ 1
1040
+ -1
1041
+ 0
1042
+ 1
1043
+ μH (T)
1044
+ μ.H (T)Energy (eV)
1045
+ 1.4
1046
+ f
1047
+ Energy (eV)
1048
+ 1.4
1049
+ 6
1050
+ TMR
1051
+
1052
+ 910
1053
+ 1000
1054
+ 4
1055
+ PL
1056
+ PL (a.u.)
1057
+ 1.3
1058
+ (a.u.)
1059
+ 1.3
1060
+ R
1061
+ 3
1062
+ 1.4
1063
+ Energy (eV)
1064
+ 1.4
1065
+ 8
1066
+ Energy (eV)
1067
+ b
1068
+ (a. u.)
1069
+ PL
1070
+ 7
1071
+ 600
1072
+ 600
1073
+ Int. intensity
1074
+ 1.3
1075
+ 1.3
1076
+ 9
1077
+ -1
1078
+ 0
1079
+ 1
1080
+ -1
1081
+ 0
1082
+ μ,H (T)
1083
+ μ。H (T)
1084
+
1085
+ 5
1086
+
1087
+
1088
+
1089
+
1090
+
1091
+
1092
+
1093
+
1094
+
1095
+
1096
+
1097
+
1098
+
1099
+
1100
+
1101
+
1102
+
1103
+
1104
+
1105
+
1106
+
1107
+
1108
+
1109
+
1110
+
1111
+
1112
+
1113
+
1114
+
1115
+
1116
+
1117
+
1118
+
1119
+
1120
+
1121
+ Extended Data Fig. 4 | Strain pulse data in the purely FM and AFM states. a, Strain
1122
+ pulse amplitude dependence in the purely FM state. As VPAC is increased from 0 to 0.5 V, a
1123
+ continuous change in the current is observed. The calculated gauge factor is ~ 5. b, Change in
1124
+ tunneling current over time as a strain pulse of 0.5 V is applied in the AFM state. Due to the
1125
+ very large resistance, the effect of pulses with smaller amplitude cannot be resolved. A gauge
1126
+ factor of ~ 30 is calculated, but with a large uncertainty due to the high resistance in the AFM
1127
+ state. No changes to the static current are observed in either FM or AFM states.
1128
+
1129
+
1130
+ .= 0.5 V
1131
+ Current (nA)
1132
+ 3.2
1133
+ AFM State (Vppc = -4.5 V)
1134
+ 3.1
1135
+ 9800
1136
+ 9900
1137
+ 10000
1138
+ 10100
1139
+ Time (sec)a
1140
+ =OV
1141
+ : 0.5.V
1142
+ 75.2
1143
+ Current (nA)
1144
+ 75
1145
+ FM State (Vppc = 24.5 V)
1146
+ 1000
1147
+ 3000
1148
+ 5000
1149
+ b
1150
+
1151
+ 6
1152
+
1153
+
1154
+
1155
+
1156
+
1157
+
1158
+
1159
+
1160
+
1161
+
1162
+
1163
+
1164
+
1165
+
1166
+
1167
+
1168
+
1169
+
1170
+
1171
+
1172
+
1173
+
1174
+
1175
+
1176
+
1177
+
1178
+
1179
+
1180
+
1181
+
1182
+
1183
+
1184
+
1185
+ Extended Data Fig. 5 | Bias-polarity-dependent stochastic switching indicates asymmetric
1186
+ vertical magnetic domain structure. a-b, Tunneling current over time of a metastable domain
1187
+ with a positive (a) and negative (b) bias applied to the MTJ. Despite a similar magnitude of
1188
+ current, no switching is observed under negative bias, ruling out heating effects. Instead, the
1189
+ data is consistent with an asymmetric vertical domain structure, as illustrated in (c). A plausible
1190
+ scenario is that when a positive voltage is applied, the FM layers polarize the tunneling
1191
+ electrons. These spin polarized electrons apply a spin-transfer torque like effect to the AFM
1192
+ layers, enhancing the stochastic switching. On the other hand, when a negative bias is applied,
1193
+ the electrons are not highly polarized and do not exert a spin-transfer torque on the FM layers.
1194
+
1195
+
1196
+ Current (nA)
1197
+ 6.8
1198
+ 6.74
1199
+ -620.5 mV
1200
+ Bias
1201
+ 0
1202
+ 100
1203
+ 200
1204
+ 300
1205
+ Time (sec)a
1206
+ Current (nA)
1207
+ 6.5
1208
+ 6.4
1209
+ = 620.5 mV
1210
+ 6.3
1211
+ Bias
1212
+ 6
1213
+
1214
+ 7
1215
+
1216
+
1217
+
1218
+
1219
+
1220
+
1221
+ Extended Data Fig. 6 | Strain switching in a six-layer MTJ. a, Magnetoresistance sweeps
1222
+ of a MTJ with a six-layer CrSBr tunnel barrier as the piezo voltage Vp is increased from 32.5
1223
+ V to 75 V. The domain behavior at piezo voltages between the low-strain AFM (32.5 V) and
1224
+ high-strain FM (75 V) states is much simpler than the ~ 16-layer device presented in the main
1225
+ text, providing additional evidence that vertical, layer-dependent domains are the origin of the
1226
+ complex hysteretic domains behavior during the magnetic phase transition. The magnetic field
1227
+ is applied along the a axis at a temperature of 20 K. b-c, Magnetoresistance sweeps in the low
1228
+ strain AFM (b) and high strain FM (c) states, showing the characteristic switching from
1229
+ negative to positive MR. The optical image of the device is shown inset in b (scale bar 5 µm).
1230
+ d-e, Resistance over time at select piezo voltages during the magnetic phase transition.
1231
+ Stochastic domain switching (d) which can be stabilized by slightly increasing strain e) are
1232
+ observed. These results highlight the potential for extending the strain-programmable vdW
1233
+ MTJs to the 2D limit.
1234
+
1235
+
1236
+ Resistal
1237
+ (MQ
1238
+ 64
1239
+ V,= 54.3 V
1240
+ Resistance (M)
1241
+ Resistance (
1242
+ 2
1243
+ 1.68
1244
+ 32.5 V-
1245
+ 1.56
1246
+ R
1247
+ 7.5V
1248
+ 1.64
1249
+ 0
1250
+ 0
1251
+ 7880
1252
+ 7960
1253
+ 8040
1254
+ μ.H(T)
1255
+ μ。H(T)
1256
+ Time (sec)a
1257
+ b
1258
+ d
1259
+ 1.73
1260
+ 75 V
1261
+ = 32.5 V
1262
+ (MQ)
1263
+ V. = 52.9 V
1264
+ Resistance (
1265
+ 4
1266
+ Resistance
1267
+ 1.69
1268
+ 1.45
1269
+ (M2)
1270
+ 1.65
1271
+ +
1272
+ 1.35
1273
+ 3
1274
+ nce
1275
+ -1
1276
+ 0
1277
+ 1
1278
+ 3280
1279
+ 3360
1280
+ 3440
1281
+ c
1282
+ e
1283
+
1284
+ 1
1285
+
1286
+ Supplementary information for
1287
+
1288
+ Strain-programmable van der Waals magnetic tunnel junctions
1289
+
1290
+ Authors: John Cenker1, Dmitry Ovchinnikov1, Harvey Yang1, Daniel G. Chica2, Catherine Zhu1,
1291
+ Jiaqi Cai1, Geoffrey Diederich1,3, Zhaoyu Liu1, Xiaoyang Zhu2, Xavier Roy2, Ting Cao4,
1292
+ Matthew W. Daniels5, Jiun-Haw Chu1, Di Xiao4,1, Xiaodong Xu1,4,*
1293
+
1294
+
1295
+ 1 Department of Physics, University of Washington, Seattle, Washington 98195, USA
1296
+ 2 Department of Chemistry, Columbia University, New York, NY 10027 USA
1297
+ 3 Intelligence Community Postdoctoral Research Fellowship Program, University of Washington,
1298
+ Seattle, WA, USA
1299
+ 4 Department of Materials Science and Engineering, University of Washington, Seattle,
1300
+ Washington 98195, USA
1301
+ 5 Physical Measurement Laboratory, National Institute of Standards and Technology,
1302
+ Gaithersburg, MD, 20899, USA
1303
+
1304
+ *Correspondence to: xuxd@uw.edu
1305
+
1306
+
1307
+
1308
+
1309
+
1310
+
1311
+ 2
1312
+
1313
+ Supplementary Text: Additional stochasticity analysis of switching data taken near ρ = 0.5
1314
+ Since the tunneling current is sampled much faster than the switching rate (~ .14 sec), switching
1315
+ data collected over 200 seconds was downsampled and tested using 15 tests from the NIST test
1316
+ suite1. Maurer’s Universal Test was excluded since the binary sequence was not long enough. The
1317
+ full sampling time dependence is shown below, using a standard threshold p-value of .01. The grey
1318
+ line indicates the sequence passed all of the 15 considered tests. The red line indicates the average
1319
+ domain switching time obtained by dividing the total number of switches by the total time window.
1320
+
1321
+ In addition to the NIST test suite, we analyzed the dwell time, i.e. the time between switches, of
1322
+ the 0 and 1 states. The extracted dwell times are plotted as a histogram for the 0 and 1 states,
1323
+ following an exponential envelope as expected for a Poisson process.
1324
+
1325
+
1326
+ We then plot the logarithm of the histogram bin counts (N) versus the dwell time:
1327
+
1328
+ 80
1329
+ 80
1330
+ 40
1331
+ 40
1332
+ 0
1333
+ 0
1334
+ 0
1335
+ 0.2
1336
+ 0.4
1337
+ 0.6
1338
+ 0.8
1339
+ 0
1340
+ 0.2
1341
+ 0.4
1342
+ 0.6
1343
+ 0.8
1344
+ Time (sec)
1345
+ Time (sec)Zero State Dwell Time
1346
+ One State Dwell Time
1347
+ 200
1348
+ 200
1349
+ 160
1350
+ 160
1351
+ 120
1352
+ 120
1353
+ untsTests
1354
+ 5
1355
+ 0
1356
+ 0
1357
+ 0.05
1358
+ 0.1
1359
+ 0.15
1360
+ 0.2
1361
+ Sampling time (sec)15
1362
+ 10
1363
+ Passed
1364
+
1365
+ 3
1366
+
1367
+
1368
+ From the linear fits, we find that the characteristic lifetime, τ, of the 0 and 1 states are τ0 = 159 ±
1369
+ 9 ms and τ1 = 151 ± 9 ms, respectively, where the uncertainty is determined by the standard
1370
+ deviation of the linear fit. Based on this analysis and the NIST test suite results, we conclude that
1371
+ the strained MTJ can generate binary sequences with a high degree of randomness.
1372
+
1373
+
1374
+
1375
+
1376
+
1377
+
1378
+ 9
1379
+ 2
1380
+ 0
1381
+ 0.1
1382
+ 0.3
1383
+ 0.7
1384
+ 0.9
1385
+ 0.1
1386
+ 0.3
1387
+ 0.5
1388
+ 0.7
1389
+ 0.9
1390
+ 0.5
1391
+ Time (sec)
1392
+ Time (sec)Zero State
1393
+ One State
1394
+ 6
1395
+ 5
1396
+ 4
1397
+ 3
1398
+ g(N)
1399
+
1400
+ 4
1401
+
1402
+ References:
1403
+
1404
+ 1.
1405
+ Ang, S., Chuchill, S., NIST Test Suite, GitHub Repository,
1406
+ https://github.com/stevenang/randomness_testsuite (2017)
1407
+
7tE2T4oBgHgl3EQfPgY3/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
8dAzT4oBgHgl3EQfgfzo/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:875e25fa1e785b5707e936425d9f8232958a7f9fa738cdbb9d89dbc711282c8d
3
+ size 106544
8tFRT4oBgHgl3EQfpzcC/content/2301.13614v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5041e4cceb98e7c299f439c6214fd71d2cb6bb53c8801d392958f71f8880f6d7
3
+ size 2268832
8tFRT4oBgHgl3EQfpzcC/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0238f3291ec63c05f18305eabf803d35d46f36effa5e2a4da55dce85b7622c9
3
+ size 2490413
9dA0T4oBgHgl3EQfO_9E/content/tmp_files/2301.02168v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
9dA0T4oBgHgl3EQfO_9E/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
AdFQT4oBgHgl3EQfMTYr/content/tmp_files/2301.13267v1.pdf.txt ADDED
@@ -0,0 +1,1613 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ A R C H I S O U N D : A U D I O G E N E R AT I O N W I T H D I F F U S I O N
2
+ flavio schneider
3
+ Master’s Thesis
4
+ Supervised by Zhijing Jin, Prof. Bernhard Schölkopf
5
+ ETH Zurich
6
+ January 2023
7
+
8
+
9
+ A B S T R A C T
10
+ The recent surge in popularity of diffusion models for image gener-
11
+ ation has brought new attention to the potential of these models in
12
+ other areas of media generation. One area that has yet to be fully ex-
13
+ plored is the application of diffusion models to audio generation. Au-
14
+ dio generation requires an understanding of multiple aspects, such
15
+ as the temporal dimension, long term structure, multiple layers of
16
+ overlapping sounds, and the nuances that only trained listeners can
17
+ detect. In this work, we investigate the potential of diffusion models
18
+ for audio generation. We propose a set of models to tackle multiple
19
+ aspects, including a new method for text-conditional latent audio dif-
20
+ fusion with stacked 1D U-Nets, that can generate multiple minutes
21
+ of music from a textual description. For each model, we make an
22
+ effort to maintain reasonable inference speed, targeting real-time on
23
+ a single consumer GPU. In addition to trained models, we provide a
24
+ collection of open source libraries with the hope of simplifying future
25
+ work in the field. Samples can be found at bit.ly/audio-diffusion.
26
+ iii
27
+
28
+
29
+ C O N T E N T S
30
+ 1
31
+ introduction
32
+ 1
33
+ 1.1
34
+ Audio Generation
35
+ 1
36
+ 1.2
37
+ Challenges
38
+ 1
39
+ 1.3
40
+ Existing Methods
41
+ 2
42
+ 1.4
43
+ Research Questions
44
+ 2
45
+ 1.5
46
+ Contributions
47
+ 4
48
+ 1.5.1
49
+ Models
50
+ 4
51
+ 1.5.2
52
+ Libraries
53
+ 4
54
+ 1.6
55
+ Structure of the Thesis
56
+ 5
57
+ 2
58
+ audio representation
59
+ 7
60
+ 2.1
61
+ Desirable Properties
62
+ 7
63
+ 2.1.1
64
+ Compressibility
65
+ 7
66
+ 2.1.2
67
+ Decodability
68
+ 7
69
+ 2.1.3
70
+ Diffuseability
71
+ 7
72
+ 2.2
73
+ Waveform
74
+ 8
75
+ 2.3
76
+ Spectrograms
77
+ 8
78
+ 2.3.1
79
+ STFT
80
+ 8
81
+ 2.3.2
82
+ MEL
83
+ 10
84
+ 3
85
+ existing diffusion methods
86
+ 11
87
+ 3.1
88
+ DDPM-Diffusion
89
+ 11
90
+ 3.1.1
91
+ Noising (0 → t)
92
+ 12
93
+ 3.1.2
94
+ Denoising (t − 1 ← t)
95
+ 13
96
+ 3.1.3
97
+ Training Objective
98
+ 13
99
+ 3.1.4
100
+ Sampling
101
+ 14
102
+ 3.1.5
103
+ Limitations
104
+ 14
105
+ 3.2
106
+ DDIM
107
+ 14
108
+ 3.3
109
+ V-Diffusion
110
+ 15
111
+ 3.3.1
112
+ Noising (0 → σt)
113
+ 15
114
+ 3.3.2
115
+ Denoising (σt−1 ← σt)
116
+ 16
117
+ 3.3.3
118
+ Training Objective
119
+ 16
120
+ 3.3.4
121
+ Sampling (σ0 = 0 ← · · · ← σt−1 ← σt =
122
+ 1)
123
+ 16
124
+ 4
125
+ architectures
126
+ 17
127
+ 4.1
128
+ Our a-unet Library
129
+ 17
130
+ 4.1.1
131
+ Background of U-Net
132
+ 17
133
+ 4.1.2
134
+ U-Net Block
135
+ 17
136
+ 4.1.3
137
+ Items
138
+ 19
139
+ 4.1.4
140
+ Plugins
141
+ 20
142
+ 4.2
143
+ Our audio-encoders-pytorch Library
144
+ 21
145
+ 5
146
+ models
147
+ 23
148
+ 5.1
149
+ Overview
150
+ 23
151
+ 5.2
152
+ Diffusion Unconditional Generator
153
+ 23
154
+ v
155
+
156
+ vi
157
+ contents
158
+ 5.2.1
159
+ Motivation
160
+ 23
161
+ 5.2.2
162
+ Method
163
+ 23
164
+ 5.2.3
165
+ Diffusion Method
166
+ 24
167
+ 5.2.4
168
+ Transforms
169
+ 25
170
+ 5.2.5
171
+ Usage
172
+ 26
173
+ 5.2.6
174
+ Evaluation
175
+ 27
176
+ 5.3
177
+ Text-conditional Diffusion
178
+ 27
179
+ 5.3.1
180
+ Motivation
181
+ 27
182
+ 5.3.2
183
+ Method
184
+ 28
185
+ 5.3.3
186
+ Evaluation
187
+ 29
188
+ 5.4
189
+ Diffusion Auto-Encoders with Latent Diffusion
190
+ 30
191
+ 5.4.1
192
+ Motivation
193
+ 30
194
+ 5.4.2
195
+ Method
196
+ 30
197
+ 5.4.3
198
+ Evaluation
199
+ 31
200
+ 5.5
201
+ Diffusion Upsampler
202
+ 32
203
+ 5.5.1
204
+ Motivation
205
+ 32
206
+ 5.5.2
207
+ Method
208
+ 32
209
+ 5.5.3
210
+ Evaluation
211
+ 33
212
+ 5.6
213
+ Diffusion Vocoder
214
+ 34
215
+ 5.6.1
216
+ Motivation
217
+ 34
218
+ 5.6.2
219
+ Method
220
+ 34
221
+ 5.6.3
222
+ Evaluation
223
+ 35
224
+ 5.7
225
+ Training Info
226
+ 35
227
+ 5.7.1
228
+ Data
229
+ 35
230
+ 5.7.2
231
+ Training
232
+ 35
233
+ 6
234
+ future work
235
+ 37
236
+ 7
237
+ conclusion
238
+ 39
239
+ bibliography
240
+ 41
241
+
242
+ 1
243
+ I N T R O D U C T I O N
244
+ Music is an art of time at the intersection of fine-grained perception
245
+ and symbolic patter recognition. In this work, we will investigate the
246
+ use of diffusion model to generate music, or more broadly audio,
247
+ in order to gain a deeper understanding of this intersection using
248
+ modern deep learning diffusion models.
249
+ 1.1
250
+ audio generation
251
+ Audio generation refers to the process of automatically synthesizing
252
+ novel waveforms using deep learning models. Audio generation has
253
+ been commonly approached in two different ways: symbolically or
254
+ at the waveform level. Symbolically generating audio involves creat-
255
+ ing a representation of the audio using symbols, such as MIDI data,
256
+ which can then be converted into an audio waveform. This method
257
+ is often easier to work with, but it can be difficult to capture all the
258
+ nuanced details of a sound using symbols. Waveform-based audio
259
+ generation, on the other hand, involves generating the raw audio
260
+ waveform directly. This method is more complex, due to the sheer
261
+ amount of values that have to be generated per second, but it allows
262
+ for a more precise and detailed representation of sound, that includes
263
+ all of its intricacies. Furthermore, audio generation can be uncondi-
264
+ tional or conditional. Unconditional models are trained only on audio
265
+ data and are able to generate new samples without any additional in-
266
+ put. Conditional models, on the other hand, are trained on pairs of
267
+ audio data and some kind of conditioning information, such as a text
268
+ description, genre label, lyrics, speaker id, or some other description
269
+ of the audio. At inference time, this conditioning information can be
270
+ used to guide the generation of novel audio samples that match the
271
+ desired characteristics. In this thesis, we will explore methods of con-
272
+ ditional and unconditional waveform-level generation.
273
+ 1.2
274
+ challenges
275
+ Multiple tradeoffs have to be considered when generating audio at
276
+ the waveform level. To generate a single second of high quality 48kHz
277
+ stereo audio, 96000 values must be generated, which is comparable
278
+ in size to a medium resolution image. If the goal is to generate an
279
+ entire song (hundreds of seconds) maintaining high-quality and a rea-
280
+ sonable generation speed, this task becomes much more challenging.
281
+ A common approach to generating long audio sequences is to do so
282
+ 1
283
+
284
+ 2
285
+ introduction
286
+ in chunks, however, if the context length, or the amount of audio that
287
+ the model can consider at any given time is not sufficient, the result-
288
+ ing structure may not be consistent over multiple seconds or minutes
289
+ of generation. A longer context may allow for more consistent coarse
290
+ structure, but may also lead to lower overall quality of detail or vice-
291
+ versa.
292
+ 1.3
293
+ existing methods
294
+ In this section, we will review some of the most well-known or influ-
295
+ ential waveform-based methods that have been developed to date.
296
+ One of the pioneering waveform level generation models is WaveNet
297
+ (2016 [8]), a fully convolutional architecture that exploits dilated con-
298
+ volutions with various dilation factors in order to capture a large con-
299
+ text. It’s able to synthesize a few seconds of both speech and classical
300
+ piano music at 16kHz. Jukebox (2020 [2]) uses multiple quantized au-
301
+ toencoders to discretize sounds at 3 different resolutions, followed by
302
+ a cascade of transformer upsampler models to generate the quantized
303
+ representations autoregressively. Jukebox is able to generate 44kHz
304
+ music conditioned on lyrics, artists and genres. The stack of trans-
305
+ formers trades-off generation speed for structure and quality. Audi-
306
+ oLM (2022 [1]) uses a (residual) quantized autoencoder to compress
307
+ the waveform into discrete tokens and a semantic encoder, later a cas-
308
+ cade of transformer decoders (semantic, coarse, fine) is used to gener-
309
+ ate 16kHz audio continuations top-down from the semantic represen-
310
+ tation. Musika (2022) trains a set of 1D convolutional autoencoders to
311
+ compress log-magnitude spectrograms, and a vocoder to reconstruct
312
+ both phase and magnitude from the compressed representation, us-
313
+ ing a 2D GAN discriminator trained on sequential chunks of audio
314
+ exploits this process autoregressively to generate longer sequences of
315
+ 44kHz audio. This method has a limited context length, but is very
316
+ efficient given the 1D structure of convolutions. Riffusion1 (2022) fine-
317
+ tunes the Stable Diffusion model [12] on chunks of mel-spectrograms
318
+ of 5s at 44kHz, and uses style transfer to generate multiple coherent
319
+ concatenated images while conditioning on a textual description of
320
+ the song. This method has a limited 5s context length, and trades off
321
+ speed given the large 2D architecture, but works surprisingly well
322
+ considering that the original model is trained on images, not audio.
323
+ 1.4
324
+ research questions
325
+ Diffusion models have recently demonstrated exceptional capabilities
326
+ in the field of image generation [11, 12], leading to an explosion of
327
+ incredible AI generated art 2. Iteratively removing small amounts of
328
+ 1 https://www.riffusion.com/about
329
+ 2 https://www.midjourney.com/showcase/
330
+
331
+ 1.4 research questions
332
+ 3
333
+ noise from pure noise allows diffusion models to hallucinate novel
334
+ samples that have common attributes to the data in the training set.
335
+ Compared to GANs, diffusion models in the image domain don’t
336
+ suffer from training instability, scale well with parameter size, and
337
+ have good mode coverage.
338
+ As long as the training data can be progressively corrupted from a
339
+ clean to a fully covered state, diffusion models have the potential to
340
+ be applied to multiple domains to generate novel samples. This opens
341
+ up a wide range of possibilities beyond image generation, including
342
+ video and audio generation.
343
+ In this thesis, we explore the potential of diffusion models for audio
344
+ generation. We will explore whether diffusion models can be used
345
+ on audio as effectively as with images. The aim is to generate high-
346
+ quality 48kHz stereo audio as efficiently as possible and to control the
347
+ generation in different ways, with a focus on text-conditional audio
348
+ generation.
349
+
350
+ 4
351
+ introduction
352
+ 1.5
353
+ contributions
354
+ 1.5.1
355
+ Models
356
+ We introduce the following models, some of which are/will be acces-
357
+ sible in the archisound library:
358
+ • Long: a latent diffusion model for text-conditional music genera-
359
+ tion that is capable of generating audio with an extended con-
360
+ text of multiple minutes at 48kHz, targeting context length and
361
+ structure (∼857M parameters).
362
+ • Crisp: a text-conditional audio generation diffusion model with
363
+ a context of tens of seconds at 48kHz, targeting simplicity and
364
+ high-quality waveforms (∼419M parameters).
365
+ • Upsampler: a diffusion model to uspsample music from 3kHz to
366
+ 48kHz (∼238M parameters).
367
+ • Vocoder: A diffusion model to reconstruct 48kHz waveforms
368
+ from 80-channel mel-spectrograms, variable input length (∼178M
369
+ parameters).
370
+ 1.5.2
371
+ Libraries
372
+ Moreover, we open-source the following libraries, on which previous
373
+ models are based:
374
+ • archisound3, our library including trained models ready to use.
375
+ This repository doesn’t contain any modelling code, but acts as
376
+ a wrapper and documentation for our models hosted on Hug-
377
+ gingface 4.
378
+ • audio-diffusion-pytorch5 (ADP), the main library including
379
+ the proposed audio diffusion models. This library has both a-unet
380
+ and audio-encoders-pytorch as dependencies. At the time of
381
+ writing, this library has 550+ stars on GitHub, and has been
382
+ downloaded more than 50000 times on pip.
383
+ • a-unet6, a highly customizable library to build U-Net architec-
384
+ tures in any dimension, expansible with multiple blocks and
385
+ plugins. This library can be used for any type of grid data: 1D,
386
+ 2D, 3D.
387
+ • audio-encoders-pytorch7 (AEP), a set of encoders and autoen-
388
+ coders for 1D data.
389
+ 3 https://github.com/archinetai/archisound
390
+ 4 https://huggingface.co/archinetai
391
+ 5 https://github.com/archinetai/audio-diffusion-pytorch
392
+ 6 https://github.com/archinetai/a-unet
393
+ 7 https://github.com/archinetai/audio-encoders-pytorch
394
+
395
+ 1.6 structure of the thesis
396
+ 5
397
+ Some additional libraries we open-soruce that are not documented
398
+ in this thesis, but might nevertheless be interesting to the reader, in-
399
+ clude: cqt-pytorch8 for invertible CQT spectrograms using NSGT,
400
+ and bitcodes-pytorch9 a method for vector-quantization into binary
401
+ codes.
402
+ 1.6
403
+ structure of the thesis
404
+ In Chapter 2, we present the various audio representations and pro-
405
+ vide a set of tradeoffs that must be considered when selecting an ap-
406
+ propriate representation. In Chapter 3, we describe the general prin-
407
+ ciples of diffusion and then delve into the specific diffusion methods
408
+ that we have tested. In Chapter 4, we examine our custom architec-
409
+ tures, including the U-Net and autoencoder, and provide detailed de-
410
+ scriptions of each component and how they can be easily integrated
411
+ into our library. In Chapter 5, we propose a range of diffusion models
412
+ that combine the diffusion methods from Chapter 3 with our custom
413
+ architecture from Chapter 4. Finally, in Chapters 6 and 7, we discuss
414
+ potential future work and present our conclusions.
415
+ 8 https://github.com/archinetai/cqt-pytorch
416
+ 9 https://github.com/archinetai/bitcodes-pytorch
417
+
418
+
419
+ 2
420
+ A U D I O R E P R E S E N TAT I O N
421
+ In the following section, we will introduce the different types of au-
422
+ dio representation that we can choose from, and compare the differ-
423
+ ent tradeoffs. Before that, we’ll have a look at the different desirable
424
+ properties that should be considered.
425
+ 2.1
426
+ desirable properties
427
+ 2.1.1
428
+ Compressibility
429
+ We define compressibility as the approximate number of values per
430
+ second needed for high-quality audio compared to the original wave-
431
+ form, and how many can be easily removed without a significant loss
432
+ in fidelity, e.g. by applying a convolutional only autoencoder on the
433
+ representation.
434
+ 2.1.1.1
435
+ Perceptibility
436
+ Perceptibility implies how close is the representation to human hear-
437
+ ing, this part is important since if we are compressing a representa-
438
+ tion that carries a lot of information we are not able to perceive in
439
+ the first place we will lose a lot of useful capacity. More specifically,
440
+ humans hear sound in the range of frequency from 20Hz to 20kHz,
441
+ on a logarithmic scale, which means that the frequency resolution
442
+ decreases as we approach 20kHz.
443
+ 2.1.2
444
+ Decodability
445
+ Decodability refers to how simple and fast is to decode the given
446
+ representation back to the waveform domain that can be reproduced.
447
+ 2.1.3
448
+ Diffuseability
449
+ Diffusability is a set of desirable properties that are important in or-
450
+ der for a diffusion model to be applicable. In particular, (1) the values
451
+ should be approximately in the range [−1, 1], (2) the signal should ide-
452
+ ally have some inductive biases that can be exploited by the network
453
+ (primarily 1D or 2D convolutional blocks), (3) time-shift invarance if
454
+ we are doing inpainting or autoregressive generation, i.e. the repre-
455
+ sentation should look the same at different time steps for the same
456
+ 7
457
+
458
+
459
+ J6&' EgC COJJJU 16bI626Uf2 g 2JU9JI CJUK Ot fJ6 O11&JU9J MA6-
460
+ fuG UnupGI Ot csUuGje' E 1e fuG Unp6I Ot ednGucI6e gug F Je fuG
461
+ 26Uf gnqiO JUfO g cOIb]GX-A9JnGq fGU2O1 Ot 2gb6 [C'L I] MG16 C Je
462
+ - t W (TT) - T-
463
+ roa(I : ) to f
464
+ dtiw blse TT Igndo lgi o q eigemi bne Is : gg
465
+ TIT2I.E.S
466
+ S'3 2LECLKOCVW2
467
+ IgbIgJA OA6I fIJU6' JGUC6 f6 MJII fAbIC9JJA JSA6 JS1&61 g!1616UC62
468
+ COU2fLICf f9U JOM 116d6UC162: H& 16d6UC162 f6Ug fO Ag1 JO16
469
+ FS JO22 InUCfIOU OU MSAG1OLUe' JI& 1ed6UCI6 MJJI p6 SIgGL fO 16-
470
+ 62f coaiq61 b6icebfp!f: IU igcf It M6 gbbj g 2fgqgig Fi 1
471
+ 126 qiL6CI gUg fgf MgA61OLI gL6 g pg2IC 6b1626SfOU MIC
472
+ 26dnGUC62 fG qi2gqASUg62 SI6 FuSt JOu& 26dnGUC62 916 2JOM fO qlt-
473
+ ode lof ldeauib lie9 'fi bn bvlovi i iboogb o dt i
474
+ Ot 2f6160 gq1o (+8000xs): M6 p261A6g fgf Mif g 2f9ugig ID coU-
475
+ leix e i Exxa e gi
476
+ Jsi&6: e g &oog cobg2o' fG Jnp6 Ot Agj6e Ot g 2fguggg
477
+ Vi9 9d Iliw a9 on oibs ilsp-gid o aimabro9
478
+ 2OL MII COUf9IU I 26COUg Ot gHqIO' It M6 MgUf fO &UG1gf6 JJfbJ6
479
+ It 1 = 48000 Ug M6 9I6 LGbI626UIU& ghqiO 9f 48 KHE' FUGU FJ6 f6U-
480
+ ghqio) gug 1 fe nupel ot bojure eeg fo iebi62euf fue gnqio cg.
481
+ o = l dm [a
482
+ S'S MVAEEOBW
483
+ 2onUg' (+) FJ6 ASJnG2 2onJg Of gA6 fOO JUgua AgJ62 IU f6 fIU6 q!-
484
+ VNDIO ELKEEMIVIIOM
485
+ 8
486
+ M
487
+ i e
488
+ qICf 6 bg26 OI q!L6CIA fJG MSAGULOU LO fJG Jg&UIfng6 JUIOL-
489
+ 9Jfo&6f61 gUg fo fIgIU gU ggg!foUSJ JUog6J (c9JJ6g Aocog6L) fo b16-
490
+ bjoif qo of 6saa gbbja: coou bigcjc6 Je fo qiecsig fue buga6
491
+ Ug GUC6 fG IUCA6 pIg262 Ot 2bgFI9I JOC9JIfA fgf COUAOJIfOU 6X-
492
+ ol t ig liti T .
493
+ A61 2gI JO22 I dJf OU f6 of6 gug bg26 e A61 gig fO
494
+ Jg&uinq6 ebeco&ig cgu p6 6gea cobi6e26q b fo 3sx MIfy
495
+ Fanl6 S: Wsafqe abecoaisu gug bsee ot g euaj6 cusuel 2LEI ge-
496
+ (x(tr))
497
+ s 9rugi mi rwore 2
498
+ gLCfS
499
+ α(X(t『))
500
+ guq bjg26 96 q6uU6q g2 waa(x(t r)) := lx(t r)l- 9Uq bμe(x(t r)) :
501
+ fO 69J 9Ug JUg&JUga bgLfe' gUg J6UC6 pgCK fo MgA6uLo' Wg&uIfng6
502
+ IUfo JgIfn6 gU bg26' g 16bi626UfgfOU fSf c9U p6 JA66g pgK
503
+ .bgi9i li9
504
+ GCIGJA 2I& f6 S2f ONG IOL (EE) C 9J2O G
505
+ FJG 2J&USJ MIF I KGLUGJe Ot JGu&t M LG 2LEL CSU p6 CObrIIGq
506
+ lovo i d blow o
507
+ 2IUG MAGe tOI f6 J&USIA OUG SKI& FG GSJIg&JUSI 2LEL
508
+ b16g61!U6g Jf612 cObo26g Ot Co2JU6 MSA62 1OL fJ6 L69J bgLf' 9Ug
509
+ 92 M: JUfG1G2fIU&X MG C9U COU2Ig6I fJ6 2LEL 92 ID-COUAOJNFIOU MIfJ
510
+ MJG16 M J2 g MJUqoM tnUCfIOU Ot 2ugb6 M fugf 12 26 fo J2ojfG fuG
511
+ (s)
512
+ 2l上l[x)(tr) =X(r)
513
+ (1)
514
+ X-xI2 16b1626Uf2 f6' LJ6 2LEL 9f g bo1Uf (t ) 12 q66g g2:
515
+ 1O1 M616 F6 X-gx12 16b1626Uf2 JIU691JA 2c9J6g 116dn6UC162' gUq f6
516
+ S3 LECLEOCKV2
517
+ Q
518
+ IUAGL2IOU 9Ug bJg26 IGCOU2fLNICfIOU JU OUG &o:
519
+ l od oo o o d
520
+ L L6CoG1 f6 J!6J 2c9J6g Jg&UIfg6 bCfLo&ige' obIIIgOU
521
+ .foitofi o ol oi l il t t d lditi 'i
522
+ CJO1C6 JU gqiO gbbJIcgfIOU2' LG qi2gqAgUtg&6 Ot f6 J6J 2c9J6 12 fgf
523
+ C6bfIOU fJSU g JU69IJA 2C9J6g 2b6CfIo&1gJ' JSKIU& If 9 AG1A COJUJUOU
524
+ -1 iv i oiti H
525
+ - Isos i st rts boa o oi bvio t o baed
526
+ i gle rltireol Isioe eidt gle Ie gt ( + orgo d
527
+ E&6 3: WE-2cJG 2becftog" gfq 2c9JGg Mi roa() to AepI-
528
+ IHMS.E.S
529
+ b 9t tig lib 9i 9 bilo q
530
+ SUe geI&I2 1edUCIG2 OU g Jo&gIfIC 2C9J6 MIC I2 2fI Of
531
+ f6 Jgig fO cOUub1622 l1O f6 6g2^ fo cOb1622 bgLf2:
532
+ 2626' COUA6LfI& fJ6 f19U21OLJ f Jg&UIfg6 gUg bg26 g126Uf9U&J62
533
+ 21UC6 g JOf Ot 19UqOUU622 12 2fJI b1626Uf I f6 q6f9JJ2 Ot fJ6 SLig: IU g
534
+ IGbLG2GUTSfIOU'
535
+ VNDIO KELKEEMIVLIOM
536
+ 103
537
+ E X I S T I N G D I F F U S I O N M E T H O D S
538
+ Diffusion models, first proposed in [3, 17] are most commonly imple-
539
+ mented with U-Net [7, 13] that is repeatedly called during inference
540
+ for each denoising step. Since the same network is called multiple
541
+ times during sampling, the weights are shared, making it a recur-
542
+ rent model. Since the data can be progressively corrupted from a
543
+ clean to a fully covered state, we can use this trick to jump to any
544
+ intermediate noise level and denoise a single step, backpropagating
545
+ only once during training. From the perspective of recurrent models,
546
+ (forward) diffusion allows us to recover the memory at an intermedi-
547
+ ate state (which we can see as the corrupted datapoint) without the
548
+ need to backpropagate the entire chain. This is a useful technique
549
+ for efficiently generating intermediate states, and has the advantage
550
+ that it can be highly parallelized during training. Compared to recur-
551
+ rent models, the memory state is predefined by the (noise) corruption
552
+ process and not fully learned. Diffusion exploits very similar princi-
553
+ ples as autoregressive transformer models [19], namely a highly par-
554
+ allelizeable training process and repeated network call with weight-
555
+ sharing during sampling. Compared to other generative models like
556
+ GANs, diffusion models are easier to train and don’t suffer from in-
557
+ stability problems arsing from having to coordinate a generator and
558
+ discriminator.
559
+ Diffusion models are a category of powerful generative models first
560
+ introduced in [17] (2015), and later popularized in [3] (2020), thanks
561
+ to the impressive results obtained in image generation on CIFAR10.
562
+ In this section, we will examine different diffusion methods. First,
563
+ the seminal DDPM [3] method, which involves training the diffu-
564
+ sion process with a finite number of denoising steps. Following that,
565
+ DDIM [18] introduces a few changes that generalize DDPM to an arbi-
566
+ trary number of steps. Then we will introduce V-diffusion from [16],
567
+ a continuous diffusion method that aims to improve the mixing of
568
+ the signal-to-noise ratio from DDIM. For DDPM and V-diffusion, we
569
+ will highlight the most important operations, namely: (1) noising the
570
+ original datapoint (signal) to a desired noise, (2) denoising a single
571
+ step with the use of our (trained) network, (3) the training objective
572
+ used, and (4) a sampling technique that repeatedly applies (2).
573
+ 3.1
574
+ ddpm-diffusion
575
+ DDPM [3] is one of the seminal works in diffusion models. The method
576
+ starts by assuming that xxx(0)
577
+ 0 , . . . ,xxx(D)
578
+ 0
579
+ is a dataset of D i.i.d. points
580
+ 11
581
+
582
+
583
+ MJGIG ef M(O' I)
584
+ (2)
585
+ Xf
586
+ qGAIgfOU M6 9U 6g2JJA 2gbJ6 xf f6 JO12 A6121OU Ot ×o~ g2:
587
+ M616 f := II=( -f) q6b6Uq2 OU 9JI f 26J6Cf6q IU d(×f I×f-):
588
+ ×|hf :=^fx0f;=(-f)I
589
+ d(xf |x0) :=
590
+ (4)
591
+ :es bgtslumrof roitudirtaib
592
+ fG L6bgLGUGfLISSfIOU fLICK If C9U p6 2OMU fSf FJI2 J2 9J2O g JOLIU9J
593
+ JGaGj f Lia bioceqrL6 1e cJJeg f toMgig qilejOu bioc: eue
594
+ Msa o qiLCIa Inb o UO126 JGG o' (om CJ6S stboTt) fO UO1e6
595
+ B 2I f6 b16e g2nbOUa' M6 c q61I6 d(xf I xo)~ I'6' g
596
+ 3'I'I Voe& (0→ f)
597
+ tioq uoivi t
598
+ C9JJ6 00AONC6 2C6NJ6' MIC COUFOI f6 IUCL6926 IU JO126 JGA6I LOU
599
+ b6Ug6Uf OU fJ6 b16AOn2 bo1Uf 9Ug 2OUU6 ab6tbg19J61612 ↓: : :
600
+ fO 29UbJ6 g JOLU9J gI2fLJpHIfIOU MIfJ fJ6 69U 9Ug COUASLISUC6 g6-
601
+ F6 JO126 J6A6J O1 OnIL qggboIUf ×f-1 p OU6 2f6b fO J6A6J f M6.JI A6
602
+ M(×f I hrf = ^I -fxf-」f = fI): IU MOLq2' I M6 M9Uf fO JUCI6926
603
+ : (-x I x)p edt bns (T 1o mumixem o mor Igvl 9io f
604
+ 2gbj6g g UKoM g2Jpfou d(xo)(f6 2p2cbf Iugicf62
605
+ EI&M6 2: D!2JOU JU1616UC6'
606
+ oibuA
607
+ EI&n6 : D!L2IOU fIUIU&.
608
+ oibuA
609
+ IS
610
+ EXIIИ DIEEIOИ WEHOD3.1 ddpm-diffusion
611
+ 13
612
+ 3.1.2
613
+ Denoising (t − 1 ← t)
614
+ The reverse process distribution q(xxxt−1 | xxxt) is also a normal distri-
615
+ bution. However, it cannot be directly estimated as it depends on the
616
+ entire dataset. Instead, we train a neural network with parameters θ
617
+ as an approximation:
618
+ pθ(xxxt−1 | xxxt) ..= N (xxxt−1 | µµµθ(xxxt), Σθ(xxxt))
619
+ (6)
620
+ If our model is trained properly, similarly to the forward process,
621
+ we will be able to carry out a single denoising step by sampling the
622
+ normal distribution using the learned mean and variance.
623
+ 3.1.3
624
+ Training Objective
625
+ To train our model, we need a handle on the true mean and covari-
626
+ ance of the reverse process q(xxxt−1 | xxxt). As we have seen before, this
627
+ is not directly tractable, however, if we include additional informa-
628
+ tion about either xxx0 (the true data point), or ǫǫǫt (the noise used to
629
+ get xxxt from xxx0 in the forward process) we can compute a different
630
+ but tractable auxiliary distribution. In the case where xxx0 is given, the
631
+ distribution is:
632
+ q(xxxt−1 | xxxt,xxx0) = N
633
+
634
+ xxxt−1 | ˜µµµ(xxxt,xxx0), ˜Σ(xxxt,xxx0)
635
+
636
+ (7)
637
+ With mean ˜µµµ(xxxt,xxx0) ..=
638
+ √1−βt(1− ¯βt−1)
639
+ 1− ¯βt
640
+ xxxt +
641
+ √ ¯βt−1βt
642
+ 1− ¯βt
643
+ xxx0 and covariance
644
+ ˜Σ(xxxt,xxx0) ..= 1− ¯βt−1
645
+ 1− ¯βt βtI, as shown in [3]. To train our network, we will
646
+ then minimize the divergence between this tractable distribution and
647
+ the distribution estimated with our model:
648
+ Lt ..= DKL [q(xxxt−1 | xxxt,xxx0) || pθ(xxxt−1 | xxxt)]
649
+ (8)
650
+ = Exxx0
651
+
652
+ 1
653
+ 2 ∥Σθ(xxxt)∥2
654
+ 2
655
+ ∥˜µµµ(xxxt,xxx0) −µµµθ(xxxt)∥2
656
+ 2
657
+
658
+ (9)
659
+ Which amounts to a simple L2 loss between the auxiliary mean, and
660
+ the true mean estimated by the model, with some extra scaling factor
661
+ that is dependent on the covariance, in [3] the covariance is fixed to
662
+ Σθ(xxxt) = βtI. A more rigorous argument using variational inference
663
+ can be applied to show that this is a lower bound of the negative
664
+ log-likelihood of the data distribution. More concretely, our model
665
+ fθ will output an estimated mean given the noisy datapoint and the
666
+ noise level as input: µµµθ(xxxt) = fθ(xxxt, t), which we can then use to
667
+ sample the next xxxt−1 from a normal distribution.
668
+ If instead we assume ǫǫǫt is given, we can follow a similar procedure
669
+ to get the loss Lt:
670
+ Lt ..= DKL [q(xxxt−1 | xxxt,ǫǫǫt) || pθ(xxxt−1 | xxxt)]
671
+ (10)
672
+ = E
673
+
674
+ β2
675
+ t
676
+ 2βt(1 − ¯βt) ∥Σθ(xxxt)∥2
677
+ 2
678
+ ∥ǫǫǫt − ǫǫǫθ(xxxt)∥2
679
+ 2
680
+
681
+ (11)
682
+
683
+ 14
684
+ existing diffusion methods
685
+ In this case our model will estimate the noise instead of the mean of
686
+ the datapoint xxxt, i.e. ǫǫǫθ(xxxt) = fθ(xxxt, t), however we can still recover
687
+ the mean as: ˜µµµ =
688
+ 1
689
+ √1−βt
690
+
691
+ xxxt −
692
+ βt
693
+
694
+ 1− ¯βtǫǫǫt
695
+
696
+ . Empirically, it has been
697
+ shown in [3] that the objective can be simplified further by ignoring
698
+ the scaling factor:
699
+ Lt = Eǫǫǫt
700
+
701
+ ∥ǫǫǫt −ǫǫǫθ(xxxt)∥2
702
+ 2
703
+
704
+ (12)
705
+ The final objective function to train the model is then computed with
706
+ random noise levels t sampled from a uniform distribution.
707
+ L ..= Et∼[1,T][Lt]
708
+ (13)
709
+ 3.1.4
710
+ Sampling
711
+ Sampling in DDPM is very straightforward, we start with xxxT ∼ N(0, I)
712
+ and recursively call the model T times using at each step the esti-
713
+ mated means µµµθ(xxxt) (or noises ǫǫǫθ(xxxt)) of the T normal distributions
714
+ to get each subsequent sample: xxxT−1 ∼ pθ(· | xxxT), . . . , xxx1 ∼ pθ(· | xxx2) ,
715
+ xxx0 ∼ pθ(· | xxx1) where xxx0 will be our generated output data point. Note
716
+ that this is a stochastic sampling process, since at each step additional
717
+ noise is added from sampling the normal distribution.
718
+ 3.1.5
719
+ Limitations
720
+ This method requires on the order of hundreds of sampling steps to
721
+ get good quality samples. Compared to more modern methods that
722
+ follow, the number of steps T is a fixed hyperparameter both during
723
+ training and sampling, limiting its flexibility.
724
+ 3.2
725
+ ddim
726
+ DDIM [18], is another seminal work for diffusion models. By intro-
727
+ ducing a few changes to DDPM, the number of sampling steps used
728
+ during inference can be dynamically changed while maintaining the
729
+ same training procedure. This allows to sample between x10 and x100
730
+ faster, and to trade speed for quality at will. A direct implication of
731
+ having a variable number of steps during sampling is that we can
732
+ train with very large T, or even infinitely large T, leading to a contin-
733
+ uous diffusion process. The idea of DDIM is that if we know both xxx0
734
+ and xxxt, we can use q(xxxt−1 | xxxt,xxx0) to sample xxxt−1. There are two pos-
735
+ sibilities, either train our network to predict directly (i.e. no sampling)
736
+ xxx0, or train our network to predict the noise ǫǫǫt (as done in DDPM)
737
+ that combined with xxxt can be used to infer xxx0. A key observation is
738
+ that using this alternative method doesn’t change the training objec-
739
+ tive, as the objective only depends on the backward diffusion process.
740
+
741
+ 3.3 v-diffusion
742
+ 15
743
+ Importantly, we can use a different forward process to recover the
744
+ next step, for example use q(xxxt−2 | xxxt,xxx0) to jump directly to xxxt−2
745
+ instead of xxxt−1, essentially skipping a sampling step and speeding
746
+ up the process. If we make the time-step continuous, we can jump
747
+ to any intermediate step in (0, t]. Even more interestingly, this con-
748
+ tinuous sampling procedure can be viewed from the lens of ordinary
749
+ differential equations, allowing us to use a variety of existing sam-
750
+ plers, like the basic Euler methods or more advanced ODE samplers.
751
+ 3.3
752
+ v-diffusion
753
+ V-diffusion, or v-objective diffusion [16], is a diffusion method in-
754
+ spired from DDIM, trained with a continuous value σt ∈ [0, 1]. This
755
+ is the method we found to work best on a variety of audio tasks. In
756
+ v-diffusion, if σt = 0 then xxxσt represents a data point xxx from the data
757
+ distribution and if σt = 1, it will be Gaussian noise ǫǫǫ. In DDIM we
758
+ can choose to either use the model to predict xxx0, or use it to predict
759
+ ǫǫǫt, in this case however, a velocity value vvvσt is estimated from which
760
+ both xxx0 and ǫǫǫσt can be inferred.
761
+ 3.3.1
762
+ Noising (0 → σt)
763
+ ��
764
+ ���
765
+ ��
766
+ ���
767
+ ���
768
+ ���
769
+
770
+ Figure 6: V-Diffusion semicircle
771
+ The noising process uses a weighting on a circle:
772
+ xxxσt = ασtxxx0 + βσtǫǫǫ
773
+ (14)
774
+ Where ασt
775
+ ..= cos(φt), and βσt
776
+ ..= sin(φt), where φt ..= π
777
+ 2 σt. When
778
+ σt = 0, then xxxσt = xxx0, i.e. no noise is added, if instead σt = 1,
779
+ then xxxσt = xxx1 = ǫǫǫ, i.e. only noise ǫǫǫ ∼ N(0, I). Intuitively, using the
780
+ weighting on a circle makes sure that as we move σt linearly from
781
+ 0 to 1 the noising process slowly removes information from xxx0. By
782
+ sampling a random σt ∈ [0, 1], we are more likely to pick a value that
783
+ resembles xxx0 instead of pure noise ǫǫǫ, meaning that the model will
784
+ more often see data with smaller amount of noise. Empirically, this
785
+ has been shown to be beneficial over standard DDIM diffusion.
786
+
787
+ 16
788
+ existing diffusion methods
789
+ 3.3.2
790
+ Denoising (σt−1 ← σt)
791
+ To denoise a from noise level σt to noise level σt−1, we can use our
792
+ velocity-estimating model ˆvvvσt = fθ(xxxσt, σt), note that the velocity
793
+ here is defined as the derivative vvvσt
794
+ ..= ∂xxxσt
795
+ σt , i.e. how much does the
796
+ datapoint change with a small change in the noise level σt (see circle
797
+ figure). As mentioned before, using an estimate of vvvt, we can obtain
798
+ both xxx0 and ǫǫǫt, which in turn can be used to estimate xxxσt−1 in DDIM
799
+ style:
800
+ ˆvvvσt = fθ(xxxσt, σt)
801
+ (15)
802
+ ˆxxx0 = ασtxxxσt − βσt ˆvvvσt
803
+ (16)
804
+ ˆǫǫǫσt = βσtxxxσt + ασt ˆvvvσt
805
+ (17)
806
+ ˆxxxσt−1 = ασt−1 ˆxxx0 + βσt−1 ˆǫǫǫt
807
+ (18)
808
+ In the previous equations, the first 3 lines show how to recover
809
+ the clean datapoint xxx0 and the noise ǫǫǫt from vvvt, and the last line,
810
+ remixes the noise with the initial datapoint to get xxxσt−1. The previous
811
+ equations can be formally obtained by using trigonometric properties
812
+ on the definition of velocity (as shown in the appendix of [16]), and
813
+ intuitively understood by rearranging vectors on the semicircle.
814
+ 3.3.3
815
+ Training Objective
816
+ By taking the derivative of the noising formulation, we can compute
817
+ the true velocity vvvσt = ασtǫǫǫ − βσtxxxσt. The training objective is then:
818
+ L = Et∼[0,1],σt
819
+
820
+ ∥ˆvvvσt − vvvσt∥2
821
+ 2
822
+
823
+ (19)
824
+ = Et∼[0,1],σt
825
+
826
+ ∥fθ(xxxσt, σt) − ασtǫǫǫ − βσtxxxσt∥2
827
+ 2
828
+
829
+ (20)
830
+ (21)
831
+ Where ǫǫǫ ∼ N(0, I) and xxxσt is computed according to the noising for-
832
+ mulation.
833
+ 3.3.4
834
+ Sampling (σ0 = 0 ← · · · ← σt−1 ← σt = 1)
835
+ To obtain a new data point ˆxxx0, some starting random noise ǫǫǫ ∼ N(0, I)
836
+ is sampled, and the denoising procedure previously demonstrated is
837
+ iteratively applied over a linear sigma-schedule.
838
+
839
+
840
+ EI&LG : -V6f PJOCK
841
+ COL6 CObOUGU Ot fG CIf6CL6 (JgfGg U EI&L6 ):
842
+ IU O1gG fo pJJg g &6UGIC n-V6f PJOcK If 12 JGC622ga fO I6UF!ta fG
843
+ I'-Vf Brock
844
+ aei d bqe
845
+ IU& fO 9Jf61 f6 pg2IC f6UbJS6 2 fgf 6xb6LIUU6UfgfIOU gUg If61gfIOJ
846
+ Agiefa ot n-efe: Lue Bog Ot g-uef Je fo bioe fue af Jeaej Ot
847
+ bliud ot old gibliud gldsosd o xodloot 2gbulori sdt isrd
848
+ -il en- bivoq w eoiev gib iw Isb ot ab l
849
+ t o od t g t ew Is [ l oold oit
850
+ JUGU2' JUCJq!U&: UGM 2KI COUUGCIOU2' COUAOJFIOU9J PJOCK2' 9ffGU-
851
+ GLU AGL2IOU2 fSf IUCOIbOLgf6 UMIUGLON2 GUJSUC6UGUf2 9Uq IIbLOA6-
852
+ -bom rit it v bvlov ido -U T
853
+ b bs oe o id ibi
854
+ JIg&6e' f IU OnI C26 M6 MJJI gggbf If 1O ID COUAOJOU2 U O1q61
855
+ gICIf6CfI6 26g SD COUAOfIOU2 fO 6xbJOIf fG 2bgfIg 2LICfL6 O
856
+ FO JG9LU g bI62GLAG !UG qGF9IJ2 gf JJFbJG LG2OJIfIOU2' LG OLI&IUgJ
857
+ J6UrgfIOU' -V6fe cOU21ef Ot 9U 6UCoq61 U6fMOLK gUq g q6Coq1 J6f-
858
+ fAb6 O1 COUAOJIFIOU9J 9ICIf6CfL6 OLI&IU9JJA g6A6JOb6g 1OL Ug86 28-
859
+ [l V-U fiw bgmlqi lnommo lbom oai
860
+ 4'I'1Back&tong ot -Vsf
861
+ I O g-U IBV
862
+ VBCHILECLNBE218
863
+ architectures
864
+ These include a downsampling block that simultaneously reduces
865
+ the resolution and number of channels of the input (typically imple-
866
+ mented with a single convolution), a stack of customizable processing
867
+ items (see subsection 4.1.3 for details), an inner block that may contain
868
+ another instance of the block recursively, a second stack of processing
869
+ items that typically mirrors the first stack, an upsampling block that re-
870
+ verses the effects of the downsampling (typically implemented with a
871
+ single transposed convolution), and a skip block that merges the skip
872
+ connection using some operation.
873
+ Furthermore, we select 3 possible types of conditioning contexts
874
+ that can be injected in the processing items, namely: a feature-vector
875
+ based conditioning
876
+ typically used with diffusion to provide the
877
+ noise level, an embedding based conditioning
878
+ that injects multiple
879
+ embedding vectors as context, typically used for text/CLIP-embedding
880
+ based conditioning, and lastly a channel-based conditioning
881
+ used
882
+ to inject entire stacks of channels in the block. Depending on the task,
883
+ we might a different combination of conditioning methods.
884
+ All described characteristics can be defined and customized using
885
+ the following block:
886
+ from a_unet.apex import Block
887
+ block = Block(
888
+ dim=1,
889
+ in_channels=2,
890
+ channels=4,
891
+ factor=2,
892
+ items=[...],
893
+ # Optional
894
+ items_up=[...],
895
+ downsample_t=...,
896
+ upsample_t=...,
897
+ skipt_t=...,
898
+ inner_block=...
899
+ )
900
+ This is a building block for a U-Net, where we can customize the num-
901
+ ber of input/ouput channels (in_channels), the number of channels
902
+ post-downsampling, and the downsampling factor. The items list
903
+ will contain the different items that will be duplicated after the inner
904
+ block. Optionally, we can change the type of skip connection, down-
905
+ sampling and upsampling operations. The inner_block can be an-
906
+ other instance of Block to recursively nest multiple blocks.
907
+ Since a U-Net is usually composed of multiple nested blocks where
908
+ the number of in_channles of the inner block must match the num-
909
+ ber of channels of the outer block, we provide XUnet as a glue class,
910
+ and XBlock as a template class for Block to make this process more
911
+ convenient and automated:
912
+ from a_unet.apex import XUNet, XBlock
913
+ unet = XUNet(
914
+
915
+
916
+ procK = XBrocK('
917
+ ( +[ + []=
918
+ [ . A .M] = 6
919
+
920
+ L66qLolmglqIfw 2 "
921
+ CL022Vf+6U10UIf6W 92 C"
922
+
923
+
924
+ XBTOC
925
+ ) oq q.n_ mo
926
+ :wolof oJax ooja b lia d
927
+ LJG 6XSJUbJ6 COUUPIUSFIOU ILOUU EI&NIL6 8' OI 9UA OFUGI COUUPIUSFIOU'
928
+ eti t-U :8 gi
929
+ .O
930
+ Ot 6&1OU A6CfOL2' gU g IJ6cfIGW (I) tO1 U)6CfI& g 26f Ot 1OAIg6g
931
+ Gp6qq!& AGcfO12 g 66qLoLMg qIGw () 1O1 WIL JIk6 b1Oc622I
932
+ (C) tOL CIO2e-SfGUOU pef6GU I6&JOU AGCfOLe gg g bLOAIgeg 2f O1
933
+ 26Jt-ff6UFO LOC622I& IIf 6fM66U I681OU A6CO12" g C L22fGU+TOIfG
934
+ 1G1GUf CSUU6Je IOAI6g 1G9fIL6 A6CO1 g VfFGufTOuIfGW (v) tOL
935
+ CG22u& UIf' g WoqnrgfiouIfew (w) fo gbbja Jogngfiou Ot fG g!t
936
+ - Ioilovo () m bivo 9w xd t o
937
+ 3Ife
938
+ .lb gbivo ot oJax
939
+ 1OLMgiq6g fo 9I PJOck2' Lgig6f612 c9U 9J2o p6 b1OAig6q fo g 2b6cIC
940
+ 2KIb COUUGCFIOU 2KTb-f IU fG XnM6f' fUSf MJII IU fHU gnfOUSIC9JA
941
+ 2Kb-f=
942
+ 1
943
+ XBrock(cuguu6r2=je' 9cfol=5' 1f6w2=[-.-1)*
944
+ ([...]= = 8=J)
945
+ ([...=2 .= .=)
946
+ p『oc2=[
947
+ Tu-cge2=*
948
+ .I=mib
949
+ YAI - I.20
950
+ architectures
951
+ Additional customized items can be easily included without alter-
952
+ ing the template code, making experimentation very simple.
953
+ 4.1.4
954
+ Plugins
955
+ Plugins are used to augment the U-Net model with additional func-
956
+ tionality. It’s often the case that we have to wrap the U-Net model
957
+ with some pre- and post-transformation, or that we have to alter or
958
+ augment the inputs provided to the U-Net. In order to maintain a
959
+ modular structure, plugins can be used to directly modify the U-Net
960
+ type without having to change the model code.
961
+ 4.1.4.1
962
+ Time Conditioning Plugin
963
+ The time conditioning plugin is used to convert a floating point value
964
+ to a conditioning feature vector
965
+ , this is useful during diffusion to
966
+ provide the current noise level, or timestep. To obtain the time feature
967
+ vector from a floating point value, a learned weight is multiplied by
968
+ the time information to get a frequency vector that is then processed
969
+ using a pair of sin and cos to get Fourier features. The Fourier features
970
+ are then transformed to a learned feature vector of the desired size by
971
+ a stack of MLPs. This function can be easily added to the base U-Net
972
+ as:
973
+ UNetWithTime = TimeConditioningPlugin(UNet)
974
+ This extends the U-Net with an additional time parameter, which can
975
+ be one or more floating point values of each batch element.
976
+ 4.1.4.2
977
+ Embedding Classifier Free Guidance Plugin
978
+ Classifier free guidance is a method proposed in [4]. We provide
979
+ a ClassifierFreeGuidancePlugin used to increase the conditioning
980
+ strength of the provided embedding
981
+ . During training, the embed-
982
+ ding is masked with a fixed (learned) embedding with a small prob-
983
+ ability in order to ensure that the network is able to generate realistic
984
+ output without access to any conditioning information. During infer-
985
+ ence, the network is called twice, once with the conditioning embed-
986
+ ding to get ˆye, and once with the fixed embedding used as mask to
987
+ get ˆym. A scaling factor embedding_scale (λ) is then used to guide
988
+ the network to produce an output that gives more or less importance
989
+ to the conditioning embedding compared to the masked embedding:
990
+ ˆy = ˆym + (ˆym − ˆye) · λ
991
+ (22)
992
+ This plugin can be easily used by augmenting the U-Net as:
993
+ UNetCFG = ClassifierFreeGuidancePlugin(
994
+ net_t=UNet,
995
+ embedding_max_length=64
996
+ )
997
+
998
+ 4.2 our audio-encoders-pytorch library
999
+ 21
1000
+ Later the new UNetCFG model can be called with the additional param-
1001
+ eter embedding_mask_proba to probabilistically mask a batch of em-
1002
+ beddings during training (e.g. a value of 0.1 will mask 10% of the em-
1003
+ beddings with a fixed embedding of length embedding_max_length),
1004
+ or with an embedding_scale parameter during inference, to call the
1005
+ U-Net twice with and without masking, and apply the scaling factor.
1006
+ In both cases, the embedding parameter must be provided as well.
1007
+ 4.1.4.3
1008
+ Text Conditioning
1009
+ The text conditioning plugin augments the U-Net embedding condi-
1010
+ tioning information
1011
+ with a learned text embedding from a frozen
1012
+ pretrained language model. By default, the T5-base transformer model
1013
+ from [10] is used if no embedder is provided.
1014
+ UNetWithText = TextConditioningPlugin(
1015
+ net_t=UNet,
1016
+ embedder=T5Embedder()
1017
+ )
1018
+ This adds an additional text field to the U-Net forward method that
1019
+ automatically extends the embedding with text embeddings from a
1020
+ pretrained language model.
1021
+ 4.2
1022
+ our audio-encoders-pytorch library
1023
+ The autoencoder component has a similar structure to the U-Net,
1024
+ with a few changes: (1) to make it an autoencoder no skip connections
1025
+ will be used, (2) no attention blocks will be used to make it generic to
1026
+ any input sequence length (3) no conditioning blocks will be applied.
1027
+ We open-source the autoencoder library audio-encoders-pytorch (AEP)
1028
+ as a separate library from a-unet. AEP includes both encoders and
1029
+ decoders, and a set of bottlenecks that can be used to normalize
1030
+ the latent space, namely (1) a variational bottleneck in the style of
1031
+ VAEs [5], (2) a simple tanh bottleneck, (3) a quantizer bottleneck, sim-
1032
+ ilar to the one proposed by VQ-VAEs [9]. Furthermore, we propose
1033
+ two encoders that encode spectrograms channelwise into a 1D latent,
1034
+ namely a ME1d (magnitude spectrogram only encoder), or MelE1d
1035
+ (mel spectrogram encoder), both compatible with the different bottle-
1036
+ necks.
1037
+
1038
+
1039
+ 5
1040
+ M O D E L S
1041
+ 5.1
1042
+ overview
1043
+ In this section we describe various diffusion models and their under-
1044
+ lying structures. We investigate various diffusion models that serve
1045
+ different purposes and functions, including upsampling and autoen-
1046
+ coding. Although these models may have distinct applications, they
1047
+ are ultimately utilized with the goal of audio generation. All of the
1048
+ different models are implemented using variations and combinations
1049
+ of the previously described achitectures (i.e. U-Nets and auto-encoders).
1050
+ The models proposed are implemented in the audio-diffusion-pytorch
1051
+ (ADP) library.
1052
+ 5.2
1053
+ diffusion unconditional generator
1054
+ The diffusion generator is the simplest model we propose to syn-
1055
+ thetize unconditional audio and is implemented with a single 1D
1056
+ U-Net.
1057
+ 5.2.1
1058
+ Motivation
1059
+ The unconditional diffusion model is a good starting point to test the
1060
+ overall quality of the particular architecture and diffusion method
1061
+ used. It doesn’t include any type of conditioning, making the dataset
1062
+ and training procedure very simple, and at the same time can give a
1063
+ good idea of the generation quality.
1064
+ 5.2.2
1065
+ Method
1066
+ The diffusion generator takes a raw high-quality stereo audio source
1067
+ as input from the datasets, that is then corrupted to a random noise
1068
+ level based on the chosen diffusion method. Using a U-Net, the gen-
1069
+ erator then predicts the output, which may be the denoised input or
1070
+ a value that is used to compute the denoised input, depending on the
1071
+ type of diffusion method employed. The noise level (usually called
1072
+ time or σ) is provided as conditioning to the network thorugh as an
1073
+ encoded feature vector
1074
+ to tell the network how much noise must
1075
+ be removed from the provided input. For the diffusion generator nei-
1076
+ ther the embedding conditioning
1077
+ nor the cross attention blocks are
1078
+ used.
1079
+ 23
1080
+
1081
+
1082
+ 2bJCif 2b66g' gUg 2gbj6 dgJf^
1083
+ 2oupj gef fo 2gbj6 l1ou' Mif gionug 2o 2gbj!u& 2f6be biognc62
1084
+ rt d of oiaib-v bof W Igvl ebuol i iev tt le
1085
+ 6 Jong2-o1J6g' M6 &t &og 6f2 & gog 1ox
1086
+ Ji&y gaUSIC-Lgu&e' 6AGU MIf ggAgUC6 2gbJIu& Gfoge: It biob
1087
+ oe e a-df g
1088
+ 21b2 Mf g21C 2JbJ61 qnL& IU1616UC6 fo &6U61f6 16920U9pJ6
1089
+ qitteioU efoqe: Onf ot f6 pox' f6 og6J goueftgf6g &oog I6-
1090
+ Ms Gagjrgteg fue beltouguc ot fue bioboeeg Joqel Mify q!leieUf
1091
+ E!&G IO: D!!21OU JUOq6I IUIGIGUC6
1092
+ WO126
1093
+ .fioitudirtib stsb gt
1094
+ MIf AgJ& O126 J6A6J2 fO &6U61g16 g U6M bJg2JPJ6 29bJ6 1OU
1095
+ bglovni vlevitsigti ai t9V-U grlt brs bglqmse i glqmse oibus gri
1096
+ DLJ& IU1616UC6' g 9UqOUU A6CfO1 MIfJ fJ6 29JU6 2Jgb6 g2 9 fL9IU-
1097
+ EI&n6 O: D!2OU JqGI fIgUI&
1098
+ oibuA
1099
+ 54
1100
+ ИODE25.2 diffusion unconditional generator
1101
+ 25
1102
+ 5.2.4
1103
+ Transforms
1104
+ Independently of the diffusion method used, this model without any
1105
+ addition struggles to generate more than a few second of sound. If
1106
+ the raw waveform is provided to the network the initial convolutional
1107
+ blocks of the U-Net will have to process huge samples, e.g. even a
1108
+ single second of high-quality 48kHz audio requires 48000 values to be
1109
+ processed by the first convolutional block. This can be a speed issue
1110
+ if the audio is not downsampled quickly enough in the U-Net, as the
1111
+ inefficiency will compound over the number of sampling steps of the
1112
+ diffusion process. In addition to that, if attention blocks are used, we
1113
+ will have to downsample enough to make sure that the number of
1114
+ timesteps to be in the range of 1024 or 2048 values. Exceeding that
1115
+ will slow down self-attention drastically due to the n2 computational
1116
+ complexity for sequence length n. Hence, a lot of downsampling is
1117
+ required with long audio samples if we want to satisfy these criteria.
1118
+ To mitigate the challenges mentioned earlier, we investigate the use
1119
+ of various methods and audio transforms to convert the raw audio
1120
+ source into a representation that reduces the temporal dimension in
1121
+ exchange for additional channels.
1122
+ 5.2.4.1
1123
+ Patching
1124
+ The first transform is patching, proposed originally for the image do-
1125
+ main in [6]. We adapt patching to the 1D domain, where the idea is
1126
+ to group sequential time steps into chunks, that will then be trans-
1127
+ posed to channels. Given a patch size p, the length t is reduced by
1128
+ t
1129
+ p where the number of channels increases to c · p, at the end of the
1130
+ U-Net processing the channels are unchunked back to the full length.
1131
+ We found patching to give drastic speedups, almost at a factor of p
1132
+ for p = 2, 4, 8, 16, 32, ..., allowing to train models with much longer
1133
+ audio sources. However, even if the audio generation quality almost
1134
+ matches the non-patched version, audible aliasing is present with all
1135
+ factors. This drawback is likely due to the repeated unchunking pro-
1136
+ cess, which will have a repeating structure, creating a high-frequency
1137
+ sine wave in the signal. Furthermore, we found that patching with
1138
+ p ⩾ 64 started to degrade quality, probably due to some capacity
1139
+ constraint in the channel dimension. We can think of patching as a
1140
+ deterministic auto-encoding process, with a downsampling factor of
1141
+ p.
1142
+ 5.2.4.2
1143
+ STFT
1144
+ The second transform is the previously introduced STFT. We use the
1145
+ common setting of 1024 num fft and window length with 256 hop
1146
+ size. By wrapping the U-Net with STFT and iSTFT the transform
1147
+ downsamples the length of the audio by 1024 while equally increas-
1148
+
1149
+ 26
1150
+ models
1151
+ ing the channel count. STFT is implemented with the Fast-Fourier
1152
+ Transform, hence it’s efficient to apply. No normalization is required
1153
+ on the spectrogram, since the diffusion loss will still be applied on the
1154
+ reconstructed wave. This method gives great speedups thanks to the
1155
+ large downsampling, but similarly to patching suffers from degrada-
1156
+ tion in quality compared to the raw wave representation. Perceptible
1157
+ noise is present in the generations both when transforming to magni-
1158
+ tude+phase, or when using real+complex.
1159
+ 5.2.4.3
1160
+ Learned Transform
1161
+ Lastly, we propose a learned transformation with a single convolu-
1162
+ tional and transposed-convolutional block at the start and respec-
1163
+ tively end of the U-Net. The transform consists in using a large kernel
1164
+ size and stride of 64. This will down-sample the original signal in a
1165
+ single step, trading off small amounts of speed from the determinis-
1166
+ tic patching or FFT implemented STFT. However, since it’s a convo-
1167
+ lutional method, we can choose the number of channels and increase
1168
+ it to a larger value (e.g. 128, double the kernel size and stride) than
1169
+ used during patching, giving more capacity to be resilient to artifacts.
1170
+ At the same time, we can use ideas from STFT and have large over-
1171
+ lapping windows with learned kernels instead of fixed sine/cosine
1172
+ waves (e.g. kernel size 128, stride 64, 64 channels, with padding to
1173
+ preserve dimension), which can help to overcome aliasing. We found
1174
+ this to be the best quality/speed tradeoff method of pre-transforming
1175
+ audio.
1176
+ 5.2.5
1177
+ Usage
1178
+ The diffusion generation model proposed is constructed by first adding
1179
+ the LTPlugin to the default U-Net UNetV0. This plugin wraps the U-
1180
+ Net with the previously described learned transform. After that, we
1181
+ have to provide the U-Net type to the DiffusionModel class which is
1182
+ responsible for constructing the U-Net, the diffusion training method
1183
+ (by default V-Diffusion), and the diffusion sampler (by default DDIM).
1184
+ from audio_diffusion_pytorch import DiffusionModel, UNetV0,
1185
+ LTPlugin, VDiffusion, VSampler
1186
+ UNet = LTPlugin(
1187
+ UNetV0, num_filters=128, window_length=64, stride=64
1188
+ )
1189
+ model = DiffusionModel(
1190
+ net_t=UNet,
1191
+ in_channels=channels,
1192
+ channels=[256, 256, 512, 512, 1024, 1024],
1193
+ factors=[1, 2, 2, 2, 2, 2],
1194
+ items=[2, 2, 2, 2, 4, 4],
1195
+
1196
+ 5.3 text-conditional diffusion
1197
+ 27
1198
+ attentions=[0, 0, 0, 0, 1, 1],
1199
+ attention_features=64,
1200
+ attention_heads=12,
1201
+ diffusion_t=VDiffusion,
1202
+ sampler_t=VSampler
1203
+ )
1204
+ This model can be easily used to get the diffusion loss for train-
1205
+ ing (which automatically applies the entire diffusion process) or to
1206
+ sample a new element provided the starting noise.
1207
+ # Training
1208
+ x = torch.randn(1, 2, 2**21) # [batch, channels, length]
1209
+ loss = model(x)
1210
+ # Sampling
1211
+ noise = torch.randn(1, 2, 2**21)
1212
+ sample = model.sample(noise=x, num_steps=50)
1213
+ 5.2.6
1214
+ Evaluation
1215
+ We found that it’s important for quality to have a single non-downsampled
1216
+ block at the start to process the transformed audio at full resolution.
1217
+ Furthermore, attention blocks are crucial for temporal consistency of
1218
+ the generated audio, but can only be applied after the original wave-
1219
+ form is down sampled to around 1024-2048 length. For example, if
1220
+ the original audio has length 219 (i.e. ∼11s at 48kHz), we downsam-
1221
+ ple by 64 = 26 in the learned transform, and by 23 in the 4 blocks
1222
+ before the first attention block, hence the context length of the first
1223
+ attention blocks will be in the desired range of 210 = 1024.
1224
+ This model can generate high quality audio over tens of seconds,
1225
+ possibly more depending on the speed requirements. In general, a
1226
+ larger set of initial convolutional/resnet blocks (closer to the wave-
1227
+ form) will result in better audio quality, at the cost of generation
1228
+ speed.
1229
+ We found that the architecture is able to generalize to longer sam-
1230
+ ples than it was trained on, if attention blocks are used. The samples
1231
+ maintain good long-context awareness even when doubling or more
1232
+ the training length. Note that this increases the attention context size
1233
+ and hence needs to be considered for before training.
1234
+ 5.3
1235
+ text-conditional diffusion
1236
+ 5.3.1
1237
+ Motivation
1238
+ We used text as a mean of conditioning for several reasons. In Imagen
1239
+ [15] it has been shown that pretrained and frozen language models
1240
+ can be successfully applied to condition the diffusion process to gen-
1241
+ erate images matching a textual description, and that by increasing
1242
+
1243
+
1244
+ Clo22-9ff6uf1ou2=[1' {' {' {' {'{]
1245
+ o = (
1246
+ :uds Isoitibbs giwollo
1247
+ fOUIu& MIfJ L2 gUg CEC c9U p6 692JJA 9qq6q fo f6 Joq6I MIf fJ6
1248
+ -ibo x9T grribb9dm9 b9r169l b9xil s 1o 1ovs mi ,29mrit 9dt o oo1
1249
+ 1166 niggUc6 [+]: DLIu& fLgUI' f6 f6xf Gp6qgu& 12 qiobb6g
1250
+ IO JUCLG926 f6 2LG&ty Ot f6 fGxf COUqfOUJu& M6 gbA cJs22IGL-
1251
+ DibuA
1252
+ nibbedma
1253
+ Ewpeqqtua
1254
+ x9T
1255
+
1256
+ JI2f MIfJ 2bgC62 9Uq fJ6 OfJ61 2o0 Ot fJ6 fIUU62 M6 26 COJUIg2 fO
1257
+ t it .1.0 ilidd
1258
+ 1opn2f' M6 2nt6 f6 J!2f ot J6fggfg gug qiob 6gc 6j66Uf Mif g
1259
+ JUOUJA 2FSLf (I ot ) OL GUg (V ot ): LO JSKG FJG COUqIFIOUJU& JUO16
1260
+ fLgIUGg OU gUg Jom JSU fofSJ CJHUK2 fJG 2ou& 12 JSgG Ot (6'&: I ot +)
1261
+ d w
1262
+ 2JC &GUGISfIOU' MG fISJU OU JGTSgSg' JUCJg!U& FJG fIfJG Ot fUG 2OU&
1263
+ -m ron .Igbom oiauib gt noitibro ot bgau zi oidw gribbgdm9
1264
+ L2 fL9U21OLJ6 6UCOq61 f 6UCO6 f6 f6Xf9J L6b1626UtSFIOU JfO
1265
+ bii be o a w[l iwollo V-U t o oil
1266
+ boNtgMS..
1267
+ J9KIU& fJ6 IU6L19C6 JOL6 &6U61IC Ug 692A fO I26'
1268
+ (---) i o
1269
+ o
1270
+ UgfCJIu&: LJI2 JIUf2 fO f6 1gCf fgf g 2JIUIJg1 J6fJog J1&f 9J2O MOLK
1271
+ Fu6 2s6 Ot fG J9u&ng&6 Joql MI 162nJf JU gU JubiO6g 6xf-Ig&6
1272
+ 58
1273
+ WODEF2
1274
+ F-
1275
+ s io qlgd gdt diw bvloa llsuau i tsdt 2TT i mldorq ommo
1276
+ i idT .igbo o s t ld o i d ilep ois b
1277
+ tiw 2brow wgf s 9ldmumr ot 9lds 2i Igbom 9dt tsdt bruo ud .(2TT)
1278
+ 86U61JC MO1q2 fgf 916 1onUq JU fIfJ62: M6 9J20 fLJ6g f6xf-fO-2b66CJ
1279
+ f6Xfn9J q62CLbIOU 2b6CI9JI 2Ju& FJ6 &6I6 Ot f6 2Ou& OL JUO16
1280
+ gd iw oibus otsm ot Iow iow ot rinoitibron xg briuof
1281
+ 2.3.3Eofom
1282
+ .0.=J62-pnibb9dm9
1283
+ Unw-2f6b2=20"
1284
+ [x9 m" ]=x
1285
+ Uo126'
1286
+ 2gwbr6 = woq6'29wbf6(
1287
+ # 29wbua
1288
+ ro22 = woq6(x* f6xf=[a f6xf、]* 6wpeqqiua-wg2-blopg=0'1)
1289
+ # ga
1290
+ wpeqq-wg-a=
1291
+ .9=p_pnbb9dm_92
1292
+ =_
1293
+ i o Ix :
1294
+ nust
1295
+ WOT26
1296
+ 23 IEX-COMDIIIOMV DIEE2IOM
1297
+ sd30
1298
+ models
1299
+ 5.4
1300
+ diffusion auto-encoders with latent diffusion
1301
+ 5.4.1
1302
+ Motivation
1303
+ Patching, STFT, and learned transforms can be used to reduce the in-
1304
+ put length during the diffusion process. Those approaches are advan-
1305
+ tageous if we want to train a single model end-to-end, however, this is
1306
+ suboptimal since the waveform is expanded to its original full-length
1307
+ shape multiple times during sampling, slowing down the process.
1308
+ A more appropriate way would be to first encode the waveform,
1309
+ then do the diffusion loop in the compressed representation, never
1310
+ expanding it to the full waveform until the end of the loop. This
1311
+ is the idea proposed in [12] (latent diffusion), where a variational
1312
+ autoencoder is first used to compress images by a few factors to a
1313
+ smaller latent space, and later diffusion is applied to that latent. By
1314
+ compressing the audio before applying diffusion, we can drastically
1315
+ speed up the diffusion sampling procedure, making an important
1316
+ case for an efficient and good quality autoencoder.
1317
+ 5.4.2
1318
+ Method
1319
+ There are different ways to implement the autoencoder, however an
1320
+ important property is that we must be able to apply the diffusion pro-
1321
+ cess to its latent space, hence some sort of normalization is required
1322
+ to make sure the values are in the range [−1, 1]. Furthermore, the
1323
+ autoencoder should compress as much as possible without a signifi-
1324
+ cant loss in quality. The smaller the latent, the faster will be the inner
1325
+ diffusion model to process and generate.
1326
+ We experimented with different autoencoders, and found that di-
1327
+ rectly compressing the waveform can only provide around 2x-4x com-
1328
+ pression without a significant loss in quality. On the other hand, as we
1329
+ have discussed in the representation section, compressing magnitude
1330
+ or mel spectrograms can provide much higher compression rates. The
1331
+ downside is that the spectrogram requires a model (vocoder) to recon-
1332
+ struct the original waveform, even from a non-compressed state.
1333
+ In this work, we propose to use a magnitude diffusion autoencoder,
1334
+ an encoder (ME1d) first encodes the waveform into a magnitude spec-
1335
+ trogram which is then encoded into a latent compressed 64x com-
1336
+ pared to the original waveform, and later uses a diffusion model to re-
1337
+ construct the waveform conditioned on the latent, acting both as a de-
1338
+ terministic compressing encoder and a diffusion vocoder at the same
1339
+ time. In order to make sure the latent space is normalized, we use
1340
+ a tanh function on the bottleneck. Since the decoding/vocoding pro-
1341
+ cess is a diffusion model, the waveform can be quickly reconstructed
1342
+ from the latent by using a small step count, if instead a more accu-
1343
+ rate reconstruction is desired a higher step count is required. To make
1344
+
1345
+
1346
+ .gribrid oibus-txgt boo 2sd brs
1347
+
1348
+ I69J-fIU6 &6U619fOU 2b66q OU g 2IU&J6 CLn' gUg J91&6 COUf6Xf J6U&f'
1349
+ 2·4·3
1350
+ q6COq61 fO 6xbsUg F6 I6bL626UfSIOU pSCk fO MgA61OLUU:
1351
+ q!2O: 2!UC6 f6 I6b626fO 12 gfCJJ COb6226' M6
1352
+ fO &6U61gf6 f6 JSf6Uf MIf f6Xf COUqIfIOUI U f6 2fAJ6 Ot JSf6Uf
1353
+ Lo &f f6 !UgJ Joq6J M6 gbbja g cgecsg& g!tn21ou &eUeifo
1354
+ I6UOA6 ff6UFOU PJOCK2 9Ug 26 COUAOJTIFOUSJ-OUJA 9ICJIf6CfJI6
1355
+ 2L6 f6 q!I21OU gIfOGUCOqGI I2 &GUGLIC fO U MSA61OL JGU&' M6
1356
+ E&nLG I: D!!2JOU gTfOGUCOGI JUIGLGUCG
1357
+ oibuA
1358
+ EI&ILG I3: D!LEIOU STfOGUCOGL fISJUU&
1359
+ oibuA
1360
+ IUI TMHTAI HTIW HO-OTUA OIUI
1361
+ 31
1362
+ .Igbom roieulib grdt lo tuqri
1363
+ 2u& JUf6IboJSfOU gUg gbb6Ug!u& fGJU g2 ggg!fIOUJ coUf6Xxf fo f6
1364
+ boNtM S..
1365
+ oib 1
1366
+ g 26coUqg b2gJbJG Joq6T (s) f JUCL6g26 fU6 2gbJ6 Igf6 O1 6X-
1367
+ g JOM 2gbj6 Igf6 ggIO MIf g bLJgia JUog6J gUg JgfGl bagJbJ6
1368
+ ie el e ()
1369
+ 2gbJI& obGifO: nbegbjG2 c9u p6 26 lOL gG1GUt bnbo262:
1370
+ gnfO6UCo612' MJ6L6 fJ6 6UCoq!u& 1UCfIOU J2 X6q fO p6 fJ6 qOMU-
1371
+ 1L2' D!tt21OU bagbj612 c9U p6 266 g2 g 2b6cC cg26 Ot qitt21OU
1372
+ JU& GIO2' LG JUOq6I M6 bIobO26' OMGAGI MOIKe qIIGCIA U MSA6-
1373
+ fO 6LO fG fob JgJt Ot fG &LIg (O1 IIg&6) 2SLfIU& gf 2OG l1GdnGUcA
1374
+ Ot 2b6cfLo&Lg2' goMU2gbJ!& g MgA61OL cOLL62boUg2 fo 26ffIu
1375
+ bIOAIg6g MSA61OLI (6&: ILOU 3KH fO 8KH) EIOU F6 b612b6CfIA6
1376
+ MA
1377
+ LI&G I2: IMO-2g&G gIt2JOu &GUGiSfOL MIf geJOU gGCogG1
1378
+ 00
1379
+ 92i0M
1380
+ ibu
1381
+ Ewpeqtue
1382
+ Ewpeqqtua
1383
+ 3s
1384
+ WODE2
1385
+ MSAG1LOU2: VffGUfIOU PJOCK gU JgL&GI COUfGXf JGU&f2 C9U JGJb b.
1386
+ COMIUf OL JSAGL2) OT fJ6 JUIfI9J COUAOJIFIOU9J PJOCK2 JU f6 -V6f COL-
1387
+ 161 JOgJ!f 29J fO OFG1 JO6J2' UC162I F6 26 (CUJ
1388
+ etx: M6 tonug nbegiubjG12 fo Gxc6j Ou 2b66cu stg' ga If,2 JkGja gu 69e-
1389
+ cu &ef AGiA 8oog 16enJfe pa begb!u& guamG16 pefMGG IQx gug
1390
+ Debeuqi& oU f6 cobj6xif ot f6 qgtg26t qittn2jou bagbj612
1391
+ 2.2.3E00mfo
1392
+ f6 I6COU2fLICFIOU 6eb6CI9JIA It b2IbI& 1O AG1 JOM 2gJbJ6
1393
+ qlgr ot gorsbig Isroitibbs ae bgbivorq gd e gribbadmg rs 1o
1394
+ E&L6 I: DI2O bL GGC6
1395
+ DoMU29wbr6q
1396
+ 36nb2gwbf6q
1397
+ .li t
1398
+ fu6 r8y 2suubj6 Isf6 JGu&ry 9ug 26 fs 29bjiu bloc622 fo 16cou-
1399
+ DHILI& IUIGLGUC6' M6 IUfGLbOJ9f6 fJ6 JOM 2gJbJ6 CJ9UUGJ2 fO JSfCJ
1400
+ .b9
1401
+ JgfCJ62 f6 Ontbnf Jr&j 29bJ6 CJ9UU6J2 9Ug J6UC6 c9U p6 bLob6IJA
1402
+ o
1403
+ EI&6 I: DI2OU beb6L fgIUI&
1404
+ b9Jqm62nwod
1405
+ oibuA
1406
+ 2 DIEEIOM LVLE
1407
+ 33
1408
+ 1 rrb2:\/&frpcoAIDIB&ACM
1409
+ EI&nI6 IO: D!!LI2JOU AOCOqGI IUIGGUC6
1410
+ W6 2b6cfLogw
1411
+ EI&nL6 I8: D!LeJOU AcoqGL FIJUJ&
1412
+ 92io
1413
+ CouAILgubo26J
1414
+ orbuA
1415
+ J9M J6J
1416
+ M6 2tgCK fJ6 gqqIfIOU9J cJ9UJ6J2 OU f6 IUbf CJgUU6J2 Ot fJ6 N-V16f'
1417
+ COUAOJfIOU pgCK fO Ife MSAGIOL 2JsbG: 2IIJSIJa fO fG begbJG
1418
+ LG q!IL2IOU ACOqGI I2 fIUGg pA IL2f COUAGLFI& FG MSAGIOLIU fO
1419
+ boNtoM s..2
1420
+ gICIf6CfL6 MIf 9JO2f UO CS&6 IUfO JI&-dngJ!fA UIC AOCog6I
1421
+ fIou' M6 brobo26 g 2bj6 gggbrsfIou fgf 9JJOMe fo fhlU Onl n-M6f
1422
+ dSJIfA 8KHS I2IC AOCoqi& g16 2I JCKIU& IU f6 tOJOMI& 26C-
1423
+ -g i - do-- f
1424
+ J6i& pg26g ocoq612: ig6g Aocoq612 c ioqc6 A61 &oog
1425
+ f6Ug fo b1ognC6 9Lfl1gCf2' JU9KIu& fJ6 c926 1O cOUOUJA 26g g66b-
1426
+ i-i 2 do bo viti o .e Isivi on i iof
1427
+ GAGr biobeuJa fhUI& g 2b6ctlo&igJ pgck fo g bjgagpJ6 gqi MgaG-
1428
+ C6IA6' J9KI& fJ6U gU Ig69J I6bL626UffOU 1O gHgO &6U6L9fOU: HOM-
1429
+ -
1430
+ oitoitoM I.d.
1431
+ 34
1432
+ WODE25.7 training info
1433
+ 35
1434
+ In order to flatten the spectrogram, we have to match the configu-
1435
+ ration of the STFT used to obtain the spectrogram, with the configu-
1436
+ ration of the 1d transposed convolution. The key insight is that the
1437
+ STFT operation can be viewed as a 1D convolution with large ker-
1438
+ nel sizes (or window size) of sine and cosine waves, which is then
1439
+ merged in-place using the absolute value, and later mel-scaled. The
1440
+ mel-scaling doesn’t alter the temporal positioning, only the frequency
1441
+ (or channels) of the spectrogram. Hence, if we set large kernel sizes
1442
+ equivalent to the STFT window length, strides equivalent to the STFT
1443
+ hop-length, and proper padding, the transposed convolution will fo-
1444
+ cus on the same context region of the waveform used to obtain the
1445
+ spectrogram. Similarly, we will set the input channels of the trans-
1446
+ posed convolution to match the number of channels used for the mel-
1447
+ spectrogram, and the output channels to 1. Stereo audio is decoded
1448
+ by batching. We used a window-length/kernel-size of 1024 and hop-
1449
+ length/stride of 256, similarly to popular vocoders we used 80 mel-
1450
+ spectrogram channels. With this configuration, the spectrogram has a
1451
+ default 3.2x compression factor over the initial waveform.
1452
+ 5.6.3
1453
+ Evaluation
1454
+ This model can produce high quality waveform, as with other mod-
1455
+ els, a good reconstruction of high-frequencies requires more convolu-
1456
+ tional blocks towards the start of the U-Net. Moreover, we hypothe-
1457
+ size that increasing the number of mel-channels might increase qual-
1458
+ ity for two reasons: first, mel-spectrogram would compress less infor-
1459
+ mation out of the initial waveform, and second, the transposed con-
1460
+ volution would have more channels to flatten the spectrogram and
1461
+ hence more capacity.
1462
+ 5.7
1463
+ training info
1464
+ 5.7.1
1465
+ Data
1466
+ We trained all of our models on a 2500h mix of audio at 48kHz. In
1467
+ the text-based model, we used metadata such as title, genre, album
1468
+ and artist as conditioning information. For the autoencoder, upsam-
1469
+ pler, vocoder, we trained on random crops of length 218 (∼5.5s at
1470
+ 48kHz). For the long-context text-conditional audio generation model,
1471
+ we trained on fixed crops of length 221 (∼44s at 48kHz), using the crop
1472
+ index as additional conditioning information.
1473
+ 5.7.2
1474
+ Training
1475
+ We trained all of our models with AdamW, using a learning rate of
1476
+ 10−4, β1 = 0.95, β2 = 0.999, ǫ = 10−6, and wight decay of 10−3. For
1477
+
1478
+ 36
1479
+ models
1480
+ all models, we used an exponential moving average with β = 0.995
1481
+ and power of 0.7. We trained all models for around 1M steps with
1482
+ a batch size of 32, this takes approximately 1 week on a single A100
1483
+ GPU for the largest, text-conditional model.
1484
+
1485
+ 6
1486
+ F U T U R E W O R K
1487
+ While our models can have a good generation quality on short few-
1488
+ second segments, or a good structure with longer segments, training
1489
+ an efficient model with both high quality and long context remains
1490
+ an open problem. A few promising future modelling approaches that
1491
+ need more experimentation include: (1) train diffusion models using
1492
+ perceptual losses on the waveforms instead of L2, this might help to
1493
+ decrease the initial size of the U-Net, as we wouldn’t have to pro-
1494
+ cess non-percieveable sounds, (2) stack multiple upsamplers to gen-
1495
+ erate a song top-down from low-sample rates to high sample rates,
1496
+ (3) improve the quality of the diffusion autoencoder by using mel-
1497
+ spectrograms instead of magnitude spectrograms as input, (4) other
1498
+ types of conditioning which are not text-based might be useful to nav-
1499
+ igate the audio latent space, which is often hard to describe in words
1500
+ - DreamBooth-like models [14] could be used to assign symbols to
1501
+ sounds, (5) compress mel-spectrograms to a quantized representation
1502
+ with diffusion autoencoders to allow for high compression ratios and
1503
+ later train an autoregressive transformer on top of that.
1504
+ Other simpler improvements on the current models include: (1) in-
1505
+ crease the training data from 2k hours to 60k-100k hours, (2) use
1506
+ more sophisticated diffusion samplers to get higher quality for the
1507
+ same number of sampling steps, (3) for text-based models, use larger
1508
+ pretrained language to obtain embeddings, which has been shown to
1509
+ be very important for quality in [15].
1510
+ 37
1511
+
1512
+
1513
+ 7
1514
+ C O N C L U S I O N
1515
+ Generating high-quality audio efficiently is a challenging task as it in-
1516
+ volves the generation of numerous values to accurately represent the
1517
+ sound waves, especially when aiming for high-fidelity stereo sound at
1518
+ a sample rate of 48kHz. In this work, we proposed different methods
1519
+ and models to generate high quality audio from a textual descrip-
1520
+ tion. From models targeting long-context audio with an emphasis on
1521
+ structure, short-context with an emphasis on quality, to other useful
1522
+ models such as the diffusion upsampler and vocoder. We introduced
1523
+ a new method that utilizes text-conditional diffusion models based on
1524
+ 1D U-Nets, allowing for the generation of multiple minutes of 48kHz
1525
+ audio on a single consumer GPU. Furthermore, we have provided a
1526
+ collection of open-source libraries to streamline future research, in-
1527
+ cluding potential improvements in audio autoencoders and diffusion
1528
+ models.
1529
+ 39
1530
+
1531
+
1532
+ B I B L I O G R A P H Y
1533
+ [1]
1534
+ Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov,
1535
+ Olivier Pietquin, Matt Sharifi, Olivier Teboul, David Grangier,
1536
+ Marco Tagliasacchi, and Neil Zeghidour. AudioLM: a Language
1537
+ Modeling Approach to Audio Generation. 2022. eprint: arXiv:2209.
1538
+ 03143.
1539
+ [2]
1540
+ Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook
1541
+ Kim, Alec Radford, and Ilya Sutskever. Jukebox: A Generative
1542
+ Model for Music. 2020. eprint: arXiv:2005.00341.
1543
+ [3]
1544
+ Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising diffu-
1545
+ sion probabilistic models.” In: Advances in Neural Information
1546
+ Processing Systems 33 (Dec. 2020), pp. 6840–6851.
1547
+ [4]
1548
+ Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guid-
1549
+ ance. 2022. eprint: arXiv:2207.12598.
1550
+ [5]
1551
+ Diederik P Kingma and Max Welling. Auto-Encoding Variational
1552
+ Bayes. 2013. eprint: arXiv:1312.6114.
1553
+ [6]
1554
+ Troy Luhman and Eric Luhman. Improving Diffusion Model Effi-
1555
+ ciency Through Patching. 2022. eprint: arXiv:2207.04316.
1556
+ [7]
1557
+ Ozan Oktay et al. Attention U-Net: Learning Where to Look for the
1558
+ Pancreas. 2018. eprint: arXiv:1804.03999.
1559
+ [8]
1560
+ Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Si-
1561
+ monyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew
1562
+ Senior, and Koray Kavukcuoglu. WaveNet: A Generative Model
1563
+ for Raw Audio. 2016. eprint: arXiv:1609.03499.
1564
+ [9]
1565
+ Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu.
1566
+ Neural Discrete Representation Learning. 2017. eprint: arXiv:1711.
1567
+ 00937.
1568
+ [10]
1569
+ Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sha-
1570
+ ran Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J.
1571
+ Liu. Exploring the Limits of Transfer Learning with a Unified Text-
1572
+ to-Text Transformer. 2019. eprint: arXiv:1910.10683.
1573
+ [11]
1574
+ Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu,
1575
+ and Mark Chen. Hierarchical Text-Conditional Image Generation
1576
+ with CLIP Latents. 2022. eprint: arXiv:2204.06125.
1577
+ [12]
1578
+ Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick
1579
+ Esser, and Björn Ommer. High-Resolution Image Synthesis with
1580
+ Latent Diffusion Models. 2021. eprint: arXiv:2112.10752.
1581
+ [13]
1582
+ Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net:
1583
+ Convolutional Networks for Biomedical Image Segmentation. 2015.
1584
+ eprint: arXiv:1505.04597.
1585
+ 41
1586
+
1587
+ 42
1588
+ bibliography
1589
+ [14]
1590
+ Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael
1591
+ Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-
1592
+ Image Diffusion Models for Subject-Driven Generation. 2022. eprint:
1593
+ arXiv:2208.12242.
1594
+ [15]
1595
+ Chitwan Saharia et al. Photorealistic Text-to-Image Diffusion Mod-
1596
+ els with Deep Language Understanding. 2022. eprint: arXiv:2205.
1597
+ 11487.
1598
+ [16]
1599
+ Tim Salimans and Jonathan Ho. Progressive Distillation for Fast
1600
+ Sampling of Diffusion Models. 2022. eprint: arXiv:2202.00512.
1601
+ [17]
1602
+ Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan,
1603
+ and Surya Ganguli. Deep Unsupervised Learning using Nonequi-
1604
+ librium Thermodynamics. 2015. eprint: arXiv:1503.03585.
1605
+ [18]
1606
+ Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Dif-
1607
+ fusion Implicit Models. 2020. eprint: arXiv:2010.02502.
1608
+ [19]
1609
+ Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
1610
+ Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polo-
1611
+ sukhin. Attention Is All You Need. 2017. eprint: arXiv : 1706 .
1612
+ 03762.
1613
+
AdFQT4oBgHgl3EQfMTYr/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
CdFAT4oBgHgl3EQftB4M/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6bdba68cccb191522643532c09923483c692dd2dc033f29bad3d7da940ab4e9c
3
+ size 152637
CtE1T4oBgHgl3EQf9wbT/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:955cce9f369460f2443c386e186e8e81551181abce0d58d7a4721fce2ca3a030
3
+ size 261912
GNE1T4oBgHgl3EQf_AZD/content/tmp_files/2301.03575v1.pdf.txt ADDED
@@ -0,0 +1,3147 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1
2
+ This work has been submitted to the IEEE for possible
3
+ publication. Copyright may be transferred without notice, after
4
+ which this version may no longer be accessible.
5
+ arXiv:2301.03575v1 [cs.IT] 9 Jan 2023
6
+
7
+ On the Coexistence of eMBB and URLLC in
8
+ Multi-cell Massive MIMO
9
+ Giovanni Interdonato, Member, IEEE, Stefano Buzzi, Senior Member, IEEE, Carmen D’Andrea, Member, IEEE,
10
+ Luca Venturino, Senior Member, IEEE, Ciro D’Elia, and Paolo Vendittelli
11
+ Abstract—The non-orthogonal coexistence between the en-
12
+ hanced mobile broadband (eMBB) and the ultra-reliable low-
13
+ latency communication (URLLC) in the downlink of a multi-
14
+ cell massive MIMO system is rigorously analyzed in this work.
15
+ We provide a unified information-theoretic framework blending
16
+ an infinite-blocklength analysis of the eMBB spectral efficiency
17
+ (SE) in the ergodic regime with a finite-blocklength analysis of
18
+ the URLLC error probability relying on the use of mismatched
19
+ decoding, and of the so-called saddlepoint approximation. Punc-
20
+ turing (PUNC) and superposition coding (SPC) are considered
21
+ as alternative downlink coexistence strategies to deal with the
22
+ inter-service interference, under the assumption of only statistical
23
+ channel state information (CSI) knowledge at the users. eMBB
24
+ and URLLC performances are then evaluated over different
25
+ precoding techniques and power control schemes, by accounting
26
+ for imperfect CSI knowledge at the base stations, pilot-based
27
+ estimation overhead, pilot contamination, spatially correlated
28
+ channels, the structure of the radio frame, and the characteristics
29
+ of the URLLC activation pattern. Simulation results reveal
30
+ that SPC is, in many operating regimes, superior to PUNC
31
+ in providing higher SE for the eMBB yet achieving the target
32
+ reliability for the URLLC with high probability. Moreover, PUNC
33
+ might cause eMBB service outage in presence of high URLLC
34
+ traffic loads. However, PUNC turns to be necessary to preserve
35
+ the URLLC performance in scenarios where the multi-user
36
+ interference cannot be satisfactorily alleviated.
37
+ Index Terms—Enhanced Mobile Broadband, Error Probabil-
38
+ ity, Massive MIMO, Mismatched Decoding, Network Availability,
39
+ Non-Orthogonal Multiple Access, Puncturing, Saddlepoint Ap-
40
+ proximation, Spectral Efficiency, Superposition Coding, Ultra-
41
+ Reliable Low-Latency Communications.
42
+ I. INTRODUCTION
43
+ W
44
+ ITH the advent of the mobile application ecosystem
45
+ and the resulting increase of the data-processing and
46
+ storage capabilities of the smart devices, several heterogeneous
47
+ services have emerged setting various stringent communication
48
+ requirements in terms of data rates, latency, reliability and
49
+ massive connectivity. These requirements and related use cases
50
+ have been summarized by the 3rd Generation Partnership
51
+ Project (3GPP) into three macro services, namely enhanced
52
+ This work was supported by the Ministero delle Imprese e del Made in
53
+ Italy (former MISE) within the project “Smart Urban Mobility Management”
54
+ (5G-SUMMA), Asse II, Supporto alle Tecnologie Emergenti.
55
+ G. Interdonato, S. Buzzi, C. D’Andrea, L. Venturino and C. D’Elia are
56
+ with the Department of Electrical and Information Engineering, University of
57
+ Cassino and Southern Latium, 03043 Cassino, Italy. They are also affiliated
58
+ with Consorzio Nazionale Interuniversitario per le Telecomunicazioni (CNIT),
59
+ 43124 Parma, Italy. P. Vendittelli is with TIM S.p.A., 20133 Milan, Italy. S.
60
+ Buzzi is also affiliated with Politecnico di Milano, 20133 Milan, Italy.
61
+ Corresponding author: Giovanni Interdonato.
62
+ mobile broadband (eMBB), ultra-reliable low-latency com-
63
+ munications (URLLC) and massive machine-type communica-
64
+ tions (mMTC) [1]. eMBB services require high-peak data-rate
65
+ and stable connectivity, and include most of the everyday us-
66
+ age applications: entertainment, multimedia, communication,
67
+ collaboration, mapping, web-surfing, etc. URLLC services
68
+ require an one-way radio latency of 1 ms with 99.999%
69
+ success probability, and include real-time and time-critical
70
+ applications, such as autonomous driving, automation control,
71
+ augmented reality, video and image processing, etc. mMTC
72
+ services enable connectivity between a vast number of miscel-
73
+ laneous devices, and include applications such as smart grids,
74
+ traffic management systems, environmental monitoring, etc.
75
+ 5G started to roll out variously as an eMBB service,
76
+ essentially like a faster version of LTE, whereas mMTC
77
+ and URLLC requirements continue to be refined and will
78
+ materialize within the next decade, although some experimen-
79
+ tal activities are already taking place in many parts of the
80
+ world1. Academic research and industrial standardization is
81
+ currently interested at different coexistence mechanisms for
82
+ such heterogeneous services, apparently moving apart from
83
+ the initial vision of a sliced network [2]. Slicing the network
84
+ basically means allocating orthogonal resources (storage, com-
85
+ puting, radio communications, etc.) to heterogeneous services
86
+ so that to guarantee their mutual isolation. This approach
87
+ is, in broad sense, generally known as orthogonal multiple
88
+ access (OMA). As an interesting alternative to orthogonal
89
+ resource allocation, non-orthogonal OMA (NOMA) is gaining
90
+ increasing importance especially with respect to the allocation
91
+ of the radio access network (RAN) communication resources.
92
+ The conventional approach to slice the RAN is to separate
93
+ eMBB, mMTC, and URLLC services in time and/or frequency
94
+ domains, whereas NOMA relies on efficient coexistence strate-
95
+ gies wherein heterogeneous services share the same time-
96
+ frequency resources, being separated in the power and spatial
97
+ domain. In this regard, the terminology Heterogeneous OMA
98
+ (H-OMA) is often adopted [2] to distinguish the orthogonal
99
+ resource allocation of heterogeneous services from that of the
100
+ same type, referred to as OMA. (The same distinction applies
101
+ to H-NOMA with respect to NOMA.)
102
+ Massive MIMO [3]–[5] is a technology that uses a very
103
+ large number of co-located antennas at the base stations (BSs)
104
+ to coherently and simultaneously serve multiple users over
105
+ 1See, e.g., the funding programs from the Italian former Ministry of
106
+ Economic Development, as well as those of other European Countries, the
107
+ EU, USA, China and Japan.
108
+
109
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
110
+ 3
111
+ the same radio resources. The users are multiplexed in the
112
+ spatial domain by using beamforming techniques that enable
113
+ high-directivity transmission and reception. The use of many
114
+ antennas also triggers the favorable propagation which further
115
+ reduces the multi-user interference and the channel hardening
116
+ which reduces the random fluctuations of the effective channel
117
+ gain. As a consequence, there is no need to adopt intri-
118
+ cate signal processing techniques to deal with the multi-user
119
+ interference. Such an aggressive spatial multiplexing along
120
+ with the intrinsic practicality and scalability of the massive
121
+ MIMO technology leads to high levels of energy and spectral
122
+ efficiency, spatial diversity, link reliability and connectivity.
123
+ The primary focus of the massive MIMO research has
124
+ been on increasing the user data rates, thereby targeting the
125
+ eMBB requirements. Lately, some studies have highlighted the
126
+ significant benefits that massive MIMO is able to provide to
127
+ URLLC [6]–[8] by reducing the outage and error probability,
128
+ and therefore increasing the link reliability. Higher reliability
129
+ results to less retransmissions which, in turn, translates to
130
+ a lower latency. mMTC also benefits from massive MIMO
131
+ technology [7], [9] by capitalizing on the high energy effi-
132
+ ciency to increase devices’ battery lifetime. Besides, favorable
133
+ propagation enables an aggressive spatial multiplexing of the
134
+ mMTC devices, facilitating the detection and the random
135
+ access procedures.
136
+ A. RELATED WORKS
137
+ Coexistence between heterogeneous services has been ini-
138
+ tially studied in systems wherein a single-antenna BS serves
139
+ multiple heterogeneous users. In [2], Popovski et al. proposed
140
+ a first tractable communication-theoretic model that captures
141
+ the key features of eMBB, URLLC and mMTC traffic. (These
142
+ features are summarized in Table I.) Specifically, [2] analyzes
143
+ two scenarios for a single-cell model: (i) slicing for URLLC
144
+ and eMBB, and (ii) slicing for mMTC and eMBB. The
145
+ downlink multiplexing of URLLC and eMBB is studied in [10]
146
+ with the goal of maximizing the utility of the eMBB traffic
147
+ while satisfying the quality of service requirements of the
148
+ URLLC traffic, and by abstracting the operation at the physical
149
+ layer. Coexistence mechanisms between URLLC and eMBB
150
+ traffic, based on the puncturing technique, have been proposed
151
+ in [11] for the uplink of a multi-cell network wherein a
152
+ simplified Wyner channel model with no fading was assumed.
153
+ As for multi-user MIMO systems, in [12] a null-space-based
154
+ spatial preemptive scheduler for joint URLLC and eMBB
155
+ traffic is proposed for cross-objective optimization, where the
156
+ critical URLLC quality-of-service (QoS) is guaranteed while
157
+ maximizing the eMBB ergodic capacity. The spatial degrees of
158
+ freedom at the BS are leveraged to fulfill the URLLC decoding
159
+ requirements without jeopardizing the performance of the
160
+ eMBB users. A similar study but for a distributed setup was
161
+ conducted in [13] where a joint user association and resource
162
+ allocation problem is formulated for the downlink of a fog
163
+ network, considering the coexistence of URLLC and eMBB
164
+ services for internet-of-things (IoT) applications. An analytic
165
+ hierarchy process was proposed for setting the priorities of the
166
+ services and to formulate a two-sided matching game where a
167
+ stable association between the fog network infrastructure and
168
+ IoT devices is established.
169
+ The coexistence between eMBB and URLLC is of most
170
+ interest [14]–[18], and is mainly handled with three alternative
171
+ techniques, herein listed in descending order of complexity:
172
+ • successive interference cancellation (SIC), with which
173
+ the receiver iteratively decode and remove the contribu-
174
+ tions of a specific service from the cumulative received
175
+ signal. This approach requires that the receiver has access
176
+ to the channel state information (CSI) to be able to
177
+ perform the multi-stage decoding, with decreasing lev-
178
+ els of interference, to the required successful decoding
179
+ probability.
180
+ • puncturing (PUNC), consisting in preventing the inter-
181
+ service interference. In the downlink, whenever the trans-
182
+ mitter has to transmit a URLLC signal, then the eMBB
183
+ signals are dropped over the channel uses involved by th
184
+ URLLC transmission. In the uplink, the receiver uses an
185
+ erasure decoder to discard the eMBB signals, provided
186
+ that it is able to detect the presence of URLLC transmis-
187
+ sions, e.g., via energy detection.
188
+ • superposition coding (SPC), with which the transmitter
189
+ simply sends a linear combination of eMBB and URLLC
190
+ signals. At the receiver, both for the uplink and the
191
+ downlink, the inter-service interference is treated as un-
192
+ correlated noise (TIN). Again, this approach requires the
193
+ receiver to be able to detect the presence of the undesired
194
+ transmissions.
195
+ In [14] the coexistence of URLLC and eMBB services in the
196
+ uplink of a C-RAN architecture with shared analog fronthaul
197
+ links is analyzed, accounting for SIC, puncturing, and TIN.
198
+ This work provides an information-theoretic study in the
199
+ performance of URLLC and eMBB traffic under both H-OMA
200
+ and H-NOMA, by considering standard cellular models with
201
+ additive Gaussian noise links and a finite inter-cell interfer-
202
+ ence. The main conclusions are that NOMA achieves higher
203
+ eMBB rates with respect to H-OMA, while guaranteeing
204
+ reliable low-rate URLLC communication with minimal access
205
+ latency. Moreover, H-NOMA under SIC is seen to achieve
206
+ the best performance, while, unlike the case with digital
207
+ capacity-constrained fronthaul links, TIN always outperforms
208
+ puncturing. A similar analysis is conducted in [11] including
209
+ both uplink and downlink of C-RAN without analog fronthaul
210
+ but considering practical aspects, such as fading, the lack
211
+ of CSI for URLLC transmitters, rate adaptation for eMBB
212
+ transmitters and finite fronthaul capacity. Abreu et al. in [16]
213
+ analyzes both the H-OMA and H-NOMA options for eMBB
214
+ traffic, and grant-free URLLC in the uplink accounting for
215
+ minimum mean square error (MMSE) receivers with and
216
+ without SIC, and under the assumption of Rayleigh fading
217
+ channels. The resulting outage probability and achievable
218
+ rates show that TIN is mostly beneficial in sufficiently high-
219
+ SNR regime when SIC is employed or, in some cases, with
220
+ low URLLC load. Otherwise, H-OMA supports higher loads
221
+ for both services simultaneously. Recently, [17] proposed an
222
+ approach to improve the supported loads for URLLC in the
223
+ uplink, for both H-OMA and H-NOMA in presence of eMBB
224
+
225
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
226
+ 4
227
+ TABLE I
228
+ FEATURES OF THE 5G USE CASES
229
+ eMBB
230
+ URLLC
231
+ mMTC
232
+ characteristics
233
+ high rate, moderate reliability
234
+ low latency, ultra reliability, low rate
235
+ low rate, large connectivity
236
+ traffic
237
+ large payload, several devices
238
+ small payload, few devices
239
+ small payload, massive devices
240
+ activation pattern
241
+ stable
242
+ intermittent
243
+ intermittent
244
+ time span
245
+ long, multiple resources
246
+ short, slot
247
+ long, multiple resources
248
+ frequency span
249
+ single/multiple resources
250
+ multiple resources, diversity
251
+ single resource
252
+ scheduling
253
+ to prevent access collision
254
+ for high reliability
255
+ infeaseable
256
+ random access
257
+ if needed
258
+ to support intermittency
259
+ fundamental
260
+ target
261
+ maximize data rate
262
+ meet latency and reliability requirements
263
+ maximize supported arrival rate
264
+ reliability requirement
265
+ ∼ 10−3
266
+ ∼ 10−5
267
+ ∼ 10−1
268
+ applications
269
+ video streaming, augmented reality,
270
+ entertainment
271
+ connected factories, traffic safety, au-
272
+ tonomous vehicles, telemedicine
273
+ internet of things, low-power sensors,
274
+ smart cities
275
+ traffic, showing the superiority of H-NOMA in ensuring the
276
+ reliability requirements of both the services. A similar analysis
277
+ but for the downlink is conducted in [18], [19] where optimal
278
+ resource allocation strategies and H-NOMA are combined to
279
+ satisfy the eMBB and URLLC QoS constraints, under the
280
+ assumption of perfect eMBB CSI and statistical URLLC CSI
281
+ knowledge.
282
+ The information-theoretic framework used by the afore-
283
+ mentioned works to characterize the performance achieved by
284
+ eMBB and URLLC users cannot be applied to massive MIMO
285
+ scenarios, for different reasons. Establishing the rate (or the
286
+ spectral efficiency) of the eMBB users in the ergodic (infinite-
287
+ blocklength) regime, upon the block-fading channel model, is
288
+ sound as the eMBB codewords spans an infinite number of
289
+ independent fading realizations. Nevertheless, as per the per-
290
+ formance of the URLLC users in a quasi-static fading scenario,
291
+ the use of the outage capacity, whose analysis includes infinite-
292
+ blocklength assumptions, leads to an inaccurate evaluation
293
+ of the error probability, as demonstrated in [8]. In addition,
294
+ outage capacity analyses do not capture the effects of the
295
+ CSI acquisition overhead when pilots are used to estimate the
296
+ uplink channel. As an alternative, finite-blocklength analyses
297
+ have been proposed for URLLC in conventional cellular
298
+ networks [18], [19], co-located massive MIMO networks [20],
299
+ [21] and cell-free massive MIMO networks [22], and rely on
300
+ the information-theoretic bounds and tools developed in [23],
301
+ e.g., the well known normal approximation. However, the
302
+ work in [8] proved that the normal approximation is not
303
+ accurate in the region of low error probabilities of interest
304
+ in URLLC (<10−4), especially as the number of antennas at
305
+ the BS increases, and in presence of imperfect CSI. Impor-
306
+ tantly, ¨Ostman et al. in [8] provided a more rigorous finite-
307
+ blocklength information-theoretic framework relying on the
308
+ use of a mismatched decoding [24], and of the saddlepoint
309
+ approximation [25] for evaluating the error probability of the
310
+ URLLC users in co-located massive MIMO systems. This
311
+ framework, priory developed for wireless fading channels
312
+ in [26]–[28], accounts for linear signal processing, imperfect
313
+ CSI and instantaneous channel estimation error, and additive
314
+ uncorrelated noise including multi-user interference. However,
315
+ the analysis of [8] is limited to the URLLC regime, and the
316
+ coexistence with the eMBB is yet to be investigated under a
317
+ unified information-theoretic framework.
318
+ B. CONTRIBUTIONS
319
+ Our contributions can be summarized as follows.
320
+ • We investigate the non-orthogonal multiplexing of the
321
+ eMBB and the URLLC, in the downlink of a multi-
322
+ cell
323
+ massive
324
+ MIMO
325
+ system,
326
+ by
327
+ providing
328
+ a
329
+ uni-
330
+ fied information-theoretic framework that combines an
331
+ infinite-blocklength analysis to assess the SE of the
332
+ eMBB and a finite-blocklength analysis to assess the error
333
+ probability of the URLLC.
334
+ • Unlike prior works wherein the URLLC performance
335
+ is inappropriately evaluated by the use of the outage
336
+ capacity analysis or the error probability obtained via the
337
+ normal approximation, in this work the finite-blocklength
338
+ information-theoretic analysis relies on the results and
339
+ tools established in [8], where mismatched receivers
340
+ and saddlepoint approximation are assumed, but the
341
+ coexistence between and URLLC and eMBB was not
342
+ investigated.
343
+ • The proposed unified framework accommodates two al-
344
+ ternative coexistence strategies: PUNC and SPC. The
345
+ former prevents the inter-service interference to protect
346
+ the URLLC reliability, whereas the latter accepts it to
347
+ maintain the eMBB service. In addition, the analytical
348
+ framework accounts for imperfect CSI acquisition at the
349
+ BSs via uplink pilot transmissions, pilot contamination
350
+ and pilot overhead, spatially correlated channels and the
351
+ lack of CSI at the users.
352
+ • We numerically evaluate the performance achieved by
353
+ PUNC and SPC under different precoding schemes,
354
+ namely maximum ratio, regularized zero-forcing and
355
+ multi-cell MMSE, and different power allocation strate-
356
+ gies, i.e., equal power allocation, weighted fractional
357
+ power allocation and optimal power allocation maxi-
358
+ mizing the product SINR throughout the network. The
359
+ coexistence between eMBB and URLLC is explored
360
+ in various scenarios, including different configurations
361
+ of the time-division duplex radio frame, and different
362
+ URLLC random activation patterns.
363
+ • The results of our comprehensive simulation campaign
364
+ highlight the clear superiority of SPC over PUNC in
365
+ most of the considered operating regimes. The main
366
+ limitation of SPC, namely the caused multi-user inter-
367
+ ference, is often overcome by using regularized zero-
368
+ forcing and multi-cell MMSE, which in turn hinge on a
369
+
370
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
371
+ 5
372
+ high-quality CSI acquisition. Whenever these precoding
373
+ techniques cannot be implemented due to complexity or
374
+ hardware constraints, the URLLC reliability requirements
375
+ can be met by fine-tuning the parameters of the pro-
376
+ posed weighted fractional power allocation. Conversely,
377
+ performing PUNC is necessary to preserve the URLLC
378
+ performance if the interference cancellation via precoding
379
+ is ineffective, for instance, when pilot contamination is
380
+ high or the multi-user interference is excessive.
381
+ • Pilot contamination among URLLC users is particularly
382
+ destructive. This led us to devise a pilot assignment policy
383
+ that prioritizes the URLLC users. In our approach, we
384
+ primarily assign unique orthogonal pilots to the URLLC
385
+ users, admitting pilot reuse only among eMBB users.
386
+ If doable, orthogonal pilots are assigned within cells
387
+ to prevent the intra-cell pilot contamination, and if the
388
+ uplink training length is sufficiently large, then mutually
389
+ orthogonal pilots are guaranteed to everyone.
390
+ C. PAPER OUTLINE
391
+ The remainder of this paper is organized as follows. In Sec-
392
+ tion II, we introduce the system model of the multi-cell
393
+ massive MIMO system, including the description of the uplink
394
+ training and a unified framework for the data transmission
395
+ stage accounting for both puncturing and superposition cod-
396
+ ing techniques. In Section III we present the information-
397
+ theoretic analyses in the infinite-blocklength regime and finite-
398
+ blocklength regime for the eMBB and the URLLC perfor-
399
+ mance evaluation, respectively. Section IV details the precod-
400
+ ing techniques and power allocation strategies to deal with the
401
+ coexistence of eMBB and URLLC users. Simulation results
402
+ and discussions are provided in Section V, while the main
403
+ findings of this work are discussed in Section VI.
404
+ D. NOTATION
405
+ Vectors and matrices are denoted by boldface lowercase and
406
+ boldface uppercase letters, respectively. Calligraphy uppercase
407
+ letters denote sets, while C and R represent the sets of complex
408
+ and real numbers, respectively. E {·} indicates the expectation
409
+ operator, while Pr {·} denotes the probability of a set. x+
410
+ represents the positive part function, namely x+ =max{x, 0},
411
+ and ⌊·⌋ denotes the floor function. The natural logarithm
412
+ is indicated by log(·) and Q(·) describes the Gaussian Q-
413
+ function. CN (µ, Σ) describes a circularly symmetric complex
414
+ Gaussian distribution with mean µ and covariance matrix Σ.
415
+ The superscripts (·)T, (·)∗ and (·)H denote the transpose, the
416
+ conjugate and the conjugate transpose (Hermitian) operators,
417
+ respectively. tr(A) indicates the trace of the matrix A, while
418
+ ∥a∥ denotes the ℓ2-norm of the vector a. The notation [A]:,i
419
+ indicates the ith column of the matrix A. IN represents the
420
+ identity matrix of size N×N. Table II introduces the notation
421
+ definition used in the system model of this paper.
422
+ II. SYSTEM MODEL
423
+ Let us consider a multi-cell massive MIMO system with
424
+ L cells, each one served by a BS that is placed at the cell-
425
+ center and equipped with M co-located antennas. Each cell
426
+ TABLE II
427
+ SYSTEM MODEL NOTATION
428
+ Symbol Description
429
+ Symbol Description
430
+ L
431
+ n. of cells
432
+ K
433
+ n. of users/cell
434
+ M
435
+ n. of BS antennas
436
+ Ku
437
+ n. of URLLC users/cell
438
+ α
439
+ Ku/K ∈ (0, 1)
440
+ Ke
441
+ n. of eMBB users/cell
442
+ τc
443
+ TDD frame length
444
+ Ku
445
+ j
446
+ URLLC users set in cell j
447
+ τp
448
+ UL training length
449
+ Ke
450
+ j
451
+ eMBB users set in cell j
452
+ τd
453
+ DL data trans. length
454
+ T
455
+ n. of slots in a TDD frame
456
+ hj
457
+ lk
458
+ channel between BS j and user k in cell l, vector CM
459
+ �hj
460
+ lk
461
+ estimate of hj
462
+ lk
463
+ �hj
464
+ lk
465
+ estimation error hj
466
+ lk−�hj
467
+ lk
468
+ Rj
469
+ lk
470
+ correl. matrix of hj
471
+ lk
472
+ βj
473
+ lk
474
+ average channel gain of �hj
475
+ lk
476
+ Cj
477
+ lk
478
+ correl. matrix of �hj
479
+ lk
480
+ f
481
+ pilot reuse factor
482
+ pp
483
+ jk
484
+ UL pilot power
485
+ ρmax
486
+ j
487
+ max transmit power at BS j
488
+ Pjk
489
+ set of all the users using the same pilot as user k in cell j
490
+ At
491
+ jk
492
+ 1 if URLLC user k in cell j is active in slot t, 0 otherwise
493
+ au
494
+ parameter of the Bernoulli distribution that draws At
495
+ jk
496
+ ςe
497
+ jk[n]
498
+ data transmitted by BS j to eMBB user k in channel use n
499
+ ςu
500
+ ji[n]
501
+ data transmitted by BS j to URLLC user i in channel use n
502
+ wjk
503
+ precoding vector, CM, used by BS j to its user k
504
+ σ2
505
+ u
506
+ UL noise variance
507
+ ρu
508
+ ji
509
+ DL power to URLLC user i
510
+ σ2
511
+ d
512
+ DL noise variance
513
+ ρe
514
+ jk
515
+ DL power to eMBB user k
516
+ gli
517
+ jk
518
+ precoded DL channel from BS l using wli to user k in cell j
519
+ �gli
520
+ jk
521
+ estimate of gli
522
+ jk
523
+ nd
524
+ URLLC codeword length
525
+ ϵdl
526
+ jk
527
+ DL error probability
528
+ ηdl
529
+ DL network availability
530
+ ν
531
+ exponent characterizing the fractional power allocation (FPA)
532
+ ω
533
+ FPA weight tuning the power allocated to the URLLC users
534
+ covers a square area of D × D km2, and provide service
535
+ to K users. It holds that M ≫ K so that interference
536
+ suppression can be efficiently carried out by exploiting the
537
+ spatial degrees of freedom. A fraction 0≤α≤1 of the K users
538
+ requests a URLLC service, e.g., a vehicle in cellular vehicle-
539
+ to-everything (C-V2X) use cases for intelligent transportation
540
+ systems, or a machine in factory automation use cases for
541
+ “Industry 4.0”. Letting Ku = αK be the number of URLLC
542
+ users per cell, then Ke = K −Ku is the number of eMBB
543
+ users per cell. The set including the indices of the eMBB and
544
+ URLLC users in cell j is denoted as Ke
545
+ j and Ku
546
+ j , respectively.
547
+ A. TDD PROTOCOL AND FRAME STRUCTURE
548
+ The considered system operates in time-division duplex
549
+ (TDD) mode to facilitate CSI acquisition and limit the es-
550
+ timation overhead. In addition, we assume that the channel is
551
+ reciprocal as a result of a perfect calibration of the RF chains.
552
+ By leveraging the channel reciprocity, the channel estimates
553
+ acquired by the BS in the uplink are then utilized in the
554
+ downlink to design the transmit precoding vectors. As channel
555
+ hardening holds for co-located massive MIMO systems with
556
+ sufficiently large antenna arrays in most of the propagation
557
+ environments, we assume that the users do not estimate the
558
+ downlink channels, and reliably decode downlink data solely
559
+ relying on the knowledge of the statistical CSI. Hence, the
560
+ TDD protocol consists of three phases: (i) pilot-based uplink
561
+ training, (ii) uplink data transmission, and (iii) downlink data
562
+ transmission.
563
+ The time-frequency resources are structured in TDD frames,
564
+ each one grouping a set of subcarriers and time samples over
565
+ which the channel response is assumed being frequency-flat
566
+ and time-invariant. The TDD frame must accommodate the
567
+
568
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
569
+ 6
570
+ URLLC
571
+ eMBB
572
+ URLLC
573
+ eMBB
574
+ PUNC
575
+ or
576
+ =
577
+ +
578
+ SPC
579
+ one slot, nd
580
+ τd
581
+ DL Data Transmission
582
+ UL training
583
+ τp
584
+ Coherence
585
+ bandwidth
586
+ Bc
587
+ one channel use
588
+ Coherence time, Tc
589
+ Fig. 1. An illustration of the TDD frame assuming no uplink data transmission
590
+ phase, and representing the resource allocation in case of puncturing (PUNC)
591
+ and superposition coding (SPC) operation.
592
+ aforementioned protocol phases and supporting all the users,
593
+ thus its size is designed to match that of the smallest user’s
594
+ coherence block in the network. As shown in Fig. 1, the
595
+ TDD frame consists of τc = TcBc samples (or channel uses)
596
+ where Tc is the coherence time and Bc is the coherence
597
+ bandwidth. τp channel uses out of τc are spent for the uplink
598
+ CSI acquisition, whereas the remaining channel uses are
599
+ devoted to the uplink and downlink data transmission. Since,
600
+ in this paper, we only focus on the downlink operation, we
601
+ assume that τd = τc −τp is the length of the downlink data
602
+ transmission phase, without loss of generality. The latter is
603
+ divided in T slots of equal length. As conventionally assumed
604
+ in the ergodic regime, an eMBB transmission spans multiple
605
+ (theoretically an infinite number of) TDD frames, wherein
606
+ the channel realizations evolve independently according to
607
+ the block-fading model. To evaluate the spectral efficiency
608
+ achieved by the eMBB users, we look at a single TDD frame
609
+ and resort to the information-theoretic bounds and tools in
610
+ the infinite-blocklength regime [4], [5]. Whereas, URLLC
611
+ transmissions are confined in time to meet the very strict
612
+ latency requirements and are allowed to span only one slot.
613
+ Hence, the number of channel uses in a slot equals the URLLC
614
+ codeword length. We assume a random activation pattern of
615
+ the URLLC users. Within a TDD frame, a URLLC user may
616
+ be active in multiple slots. To characterize the error probability
617
+ of the URLLC transmissions, we look separately at each
618
+ single slot of a TDD frame and resort to the finite-blocklength
619
+ information-theoretic bounds and tools presented in [8].
620
+ B. CHANNEL MODEL AND UPLINK TRAINING
621
+ The channel response between the k-th user in cell l and
622
+ the BS in cell j is denoted by the M-dimensional complex-
623
+ valued vector hj
624
+ lk. We assume correlated Rayleigh fading
625
+ channels, that is hj
626
+ lk ∼ CN
627
+
628
+ 0M, Rj
629
+ lk
630
+
631
+ , where Rj
632
+ lk ∈ CM×M
633
+ is the positive semi-definite spatial correlation matrix. The
634
+ corresponding average channel gain (or large-scale fading
635
+ coefficient) is given by βj
636
+ lk = tr(Rj
637
+ lk)/M. Large-scale fading
638
+ quantities are assumed to be known at the BS.
639
+ In the uplink training phase, each user transmits a pilot
640
+ sequence that spans τp channel uses. The pilot sequence of
641
+ user k in cell j is denoted by φjk ∈ Cτp. All the pilot
642
+ sequences are drawn from a set of τp mutually orthogonal
643
+ pilots, thereby the inner product between two pilots equals
644
+ either τp if the sequences are identical or 0 if they are mutually
645
+ orthogonal. Notice that re-using the pilots throughout the
646
+ network might be unavoidable as the share of the TDD frame
647
+ reserved to the training is limited and, importantly, as the
648
+ CSI acquisition overhead significantly degrades the spectral
649
+ efficiency. Pilot reuse gives rise to additional interference,
650
+ known as pilot contamination [3], that degrades the quality
651
+ of the acquired CSI and correlates the channel estimates.
652
+ The cumulative uplink signal received at BS j, denoted by
653
+ Yp
654
+ j ∈ CM×τp, reads
655
+ Yp
656
+ j =
657
+ K
658
+
659
+ k=1
660
+
661
+ pp
662
+ jkhj
663
+ jkφT
664
+ jk +
665
+ L
666
+
667
+ l=1
668
+ l̸=j
669
+ K
670
+
671
+ i=1
672
+
673
+ pp
674
+ lihj
675
+ liφT
676
+ li + Np
677
+ j ,
678
+ (1)
679
+ where pp
680
+ jk is the transmit pilot power, and Np
681
+ j is the additive
682
+ receiver noise with i.i.d. elements distributed as CN
683
+
684
+ 0, σ2
685
+ u
686
+
687
+ ,
688
+ with σ2
689
+ u being the receiver noise variance in the uplink. To
690
+ estimate the channel of user k in its own cell, hj
691
+ jk, BS j
692
+ correlates Yp
693
+ j with the known pilot sequence φjk as
694
+ yp
695
+ jjk =Yp
696
+ jφ∗
697
+ jk
698
+ =
699
+
700
+ pp
701
+ jkτphj
702
+ jk +
703
+ K
704
+
705
+ i=1
706
+ i̸=k
707
+
708
+ pp
709
+ jihj
710
+ jiφT
711
+ jiφ∗
712
+ jk
713
+ +
714
+ L
715
+
716
+ l=1
717
+ l̸=j
718
+ K
719
+
720
+ i=1
721
+
722
+ pp
723
+ lihj
724
+ liφT
725
+ liφ∗
726
+ jk + Np
727
+ jφ∗
728
+ jk .
729
+ (2)
730
+ In (2), the second term of the rightmost right-hand side
731
+ represents the intra-cell pilot contamination term, while the
732
+ third term quantifies the inter-cell pilot contamination. A
733
+ conventional pilot allocation strategy consists in assigning
734
+ mutually orthogonal pilots to users within the same cell, and
735
+ re-using the pilot sequences over different cells [5]. This
736
+ is a reasonable choice as intra-cell pilot contamination is
737
+ presumably stronger than inter-cell pilot contamination. We
738
+ let τp = fK where f is referred to as pilot reuse factor.
739
+ Importantly, in order not to jeopardize the ultra-reliability of
740
+ the URLLC transmissions, we assume that unique orthogonal
741
+ pilot sequences are assigned to all the URLLC users in the
742
+ network, if doable (namely when τp >LKe). Summarizing, the
743
+ pilot allocation strategy we propose primarily aims to prevent
744
+ URLLC users from being affected of pilot contamination, and
745
+ secondarily to prevent intra-cell pilot contamination. Finally,
746
+ if τp is sufficiently large, that is τp ≥ LK, then mutually
747
+ orthogonal pilots can be guaranteed to everyone. Let us define
748
+ the set
749
+ Pjk =
750
+
751
+ (l, i) : φli =φjk, l=1, . . . , L, i=1, . . . , K
752
+
753
+ ,
754
+ (3)
755
+ including the indices of all the users (and of the corresponding
756
+ cells) that use the same pilot as user k in cell j. Hence, we
757
+ can rewrite (2) as
758
+ yp
759
+ jjk =
760
+
761
+ pp
762
+ jkτphj
763
+ jk + τp
764
+
765
+ (l,i)∈Pjk\(j,k)
766
+
767
+ pp
768
+ lihj
769
+ li + Np
770
+ jφ∗
771
+ jk.
772
+ (4)
773
+
774
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
775
+ 7
776
+ The processed uplink signal, yp
777
+ jjk, is a sufficient statistic for
778
+ the estimation of hj
779
+ jk. Upon the knowledge of the spatial
780
+ correlation matrices, BS j can compute the minimum mean-
781
+ squared error (MMSE) estimate of hj
782
+ jk, denoted by �hj
783
+ jk, based
784
+ on the observation yp
785
+ jjk as [5]
786
+ �hj
787
+ jk =
788
+
789
+ pp
790
+ jkRj
791
+ jkΨj
792
+ jkyp
793
+ jjk
794
+ (5)
795
+ where
796
+ Ψj
797
+ jk =
798
+
799
+
800
+
801
+ (l,i)∈Pjk
802
+ pp
803
+ liτpRj
804
+ li + σ2
805
+ ulIMj
806
+
807
+
808
+ −1
809
+ .
810
+ (6)
811
+ The estimation error is given by �hj
812
+ jk = hj
813
+ jk − �hj
814
+ jk, and has
815
+ correlation matrix
816
+ Cj
817
+ jk =E
818
+
819
+ �hj
820
+ jk(�hj
821
+ jk)H�
822
+ = Rj
823
+ jk−pp
824
+ jkτpRj
825
+ jkΨj
826
+ jkRj
827
+ jk.
828
+ It follows that �hj
829
+ jk and �hj
830
+ jk are independent random variables
831
+ distributed as
832
+ �hj
833
+ jk ∼ CN
834
+
835
+ 0M, Cj
836
+ jk
837
+
838
+ ,
839
+ �hj
840
+ jk ∼ CN
841
+
842
+ 0M, Rj
843
+ jk−Cj
844
+ jk
845
+
846
+ .
847
+ C. DOWNLINK TRANSMISSION
848
+ In the downlink transmission phase, each BS transmits
849
+ payload data to all the active users of its cell. Let At
850
+ jk be
851
+ a coefficient that equals 1 if a URLLC transmission takes
852
+ place at the t-th slot for URLLC user k in cell j, and
853
+ 0 otherwise. This coefficient models the random activation
854
+ pattern of the URLLC users which follows a Bernoulli dis-
855
+ tribution with parameter au, At
856
+ jk ∼ Bern(au). To handle the
857
+ coexistence of eMBB and URLLC users in the downlink, we
858
+ consider two transmission techniques: (i) puncturing, and (ii)
859
+ superposition coding. Under puncturing, whenever a URLLC
860
+ transmission is triggered by a BS in a certain slot, all the
861
+ eMBB transmissions therein are dropped. However, the eMBB
862
+ service can be still guaranteed in the remaining slots of the
863
+ frame where no URLLC users are active. Under superposition
864
+ coding, eMBB transmissions occur in all the slots and each
865
+ BS linearly combines eMBB and URLLC signals whenever
866
+ URLLC transmissions are triggered.
867
+ The analytical framework detailed next is generalized,
868
+ namely holds for both the aforementioned transmission tech-
869
+ niques upon setting, for an arbitrary BS j and slot t, the
870
+ coefficient
871
+ ˜At
872
+ j =
873
+
874
+
875
+
876
+
877
+
878
+
879
+
880
+
881
+ 1 − �
882
+ i∈Ku
883
+ j
884
+ At
885
+ ji
886
+ �+
887
+ ,
888
+ for puncturing,
889
+ 1 ,
890
+ for superposition coding.
891
+ Let ςe
892
+ jk[n] or ςu
893
+ jk[n] be the data symbol transmitted by BS
894
+ j to user k over an arbitrary channel use n, if k is an
895
+ eMBB user or a URLLC user, respectively. We assume that
896
+ ςs
897
+ jk[n] ∼ CN (0, 1), with s = {e, u}. A slot consists of nd
898
+ channel uses, with nd =⌊τd/T⌋, and equals the length of the
899
+ URLLC codeword. The data symbol is precoded by using the
900
+ M-dimensional precoding vector wjk, which is function of
901
+ the CSI acquired at the BS during the uplink training. It also
902
+ holds E
903
+
904
+ ∥wjk∥2�
905
+ = 1. The data signal transmitted by BS j
906
+ over an arbitrary channel use n of slot t is given by
907
+ xt
908
+ j[n] = ˜At
909
+ j
910
+
911
+ k∈Ke
912
+ j
913
+
914
+ ρe
915
+ jkwjkςe
916
+ jk[n]+
917
+
918
+ i∈Ku
919
+ j
920
+ At
921
+ ji
922
+
923
+ ρu
924
+ jiwjiςu
925
+ ji[n],
926
+ (7)
927
+ with n = 1, . . . , nd, and where ρe
928
+ jk and ρu
929
+ ji are the downlink
930
+ transmit powers used by BS j to its eMBB user k and URLLC
931
+ user i, respectively, satisfying the following per-BS power
932
+ constraint
933
+ E
934
+ ���xt
935
+ j[n]
936
+ ��2�
937
+ = ˜At
938
+ j
939
+
940
+ k∈Ke
941
+ j
942
+ ρe
943
+ jk+
944
+
945
+ i∈Ku
946
+ j
947
+ At
948
+ jiρu
949
+ ji ≤ ρmax
950
+ j
951
+ ,
952
+ (8)
953
+ with j = 1, . . . , L, and where ρmax
954
+ j
955
+ is the maximum transmit
956
+ power at BS j. The data signal received at user k in cell
957
+ j over an arbitrary channel use n of slot t is denoted as
958
+ yt,s
959
+ jk[n], with s = {e, u}. In line with the conventional massive
960
+ MIMO operation, we assume that the users do not acquire
961
+ the instantaneous downlink CSI, but rather rely on a mean
962
+ value approximation of their downlink precoded channels.
963
+ Such approximation is accurate if channel hardening occurs.
964
+ If user k in cell j is an eMBB user, namely k ∈ Ke
965
+ j, then
966
+ its received data signal over an arbitrary channel use n of
967
+ slot t can be written as in (9) at the top of the next page,
968
+ where wjk[n] ∼ CN
969
+
970
+ 0, σ2
971
+ d
972
+
973
+ is the i.i.d. receiver noise with
974
+ variance σ2
975
+ d, and we have defined gli
976
+ jk =(hl
977
+ jk)Hwli, namely the
978
+ precoded downlink (scalar) channel between the BS in cell l,
979
+ using the precoding vector intended for its user i, and the k-th
980
+ user in cell j. If user k in cell j is a URLLC user, its received
981
+ data signal over an arbitrary channel use n in slot t can be
982
+ written as in (10) at the top of the next page. Equation (9)
983
+ emphasizes the fact that user k in cell j solely knows the
984
+ statistical CSI of the downlink channel, that is E
985
+
986
+ gjk
987
+ jk
988
+
989
+ . The
990
+ second term in (9) represents the self-interference due to this
991
+ lack of instantaneous CSI, referred to as beamforming gain
992
+ uncertainty. Going forward, the intra-cell inter-service inter-
993
+ ference and intra-cell intra-service interference terms represent
994
+ the interference caused by the URLLC and eMBB users of
995
+ cell j, respectively. This is presumably stronger than the inter-
996
+ cell interference caused by the eMBB users (i.e., intra-service)
997
+ and the URLLC users (i.e., inter-service) in the other cells.
998
+ A similar distinction of the various signal contributions is
999
+ reported in (10) for URLLC user k in cell j. In this case,
1000
+ the lack of instantaneous CSI at the user will be highlighted
1001
+ in the next section.
1002
+ III. PERFORMANCE ANALYSIS
1003
+ In this section, we evaluate the downlink performance
1004
+ of eMBB and URLLC users. As per the eMBB users, we
1005
+ consider the spectral efficiency (SE) by applying the infinite-
1006
+ blocklength information-theoretic results established in the
1007
+ ergodic regime [4], [5], [29]. An achievable downlink SE,
1008
+ namely a lower-bound on the ergodic downlink capacity,
1009
+ can be obtained by applying the popular hardening bound
1010
+ technique [4], [5] on the signal model in (9), by treating all
1011
+ the interference sources as uncorrelated noise. Specifically, an
1012
+
1013
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
1014
+ 8
1015
+ yt,e
1016
+ jk [n] = E
1017
+
1018
+ gjk
1019
+ jk
1020
+
1021
+ ˜At
1022
+ j
1023
+
1024
+ ρe
1025
+ jkςe
1026
+ jk[n]
1027
+
1028
+ ��
1029
+
1030
+ desired signal
1031
+ +
1032
+
1033
+ gjk
1034
+ jk −E
1035
+
1036
+ gjk
1037
+ jk
1038
+ ��
1039
+ ˜At
1040
+ j
1041
+
1042
+ ρe
1043
+ jkςe
1044
+ jk[n]
1045
+
1046
+ ��
1047
+
1048
+ self-interference
1049
+ +
1050
+
1051
+ i∈Ku
1052
+ j
1053
+ gji
1054
+ jkAt
1055
+ ji
1056
+
1057
+ ρu
1058
+ jiςu
1059
+ ji[n]
1060
+
1061
+ ��
1062
+
1063
+ intra-cell inter-service interference
1064
+ +
1065
+
1066
+ i∈Ke
1067
+ j\{k}
1068
+ gji
1069
+ jk ˜At
1070
+ j
1071
+
1072
+ ρe
1073
+ jiςe
1074
+ ji[n]
1075
+
1076
+ ��
1077
+
1078
+ intra-cell intra-service interference
1079
+ +
1080
+ L
1081
+
1082
+ l=1
1083
+ l̸=j
1084
+
1085
+ i∈Ke
1086
+ l
1087
+ gli
1088
+ jk ˜At
1089
+ l
1090
+
1091
+ ρe
1092
+ liςe
1093
+ li[n]
1094
+
1095
+ ��
1096
+
1097
+ inter-cell intra-service interference
1098
+ +
1099
+ L
1100
+
1101
+ l=1
1102
+ l̸=j
1103
+
1104
+ i∈Ku
1105
+ l
1106
+ gli
1107
+ jkAt
1108
+ li
1109
+
1110
+ ρu
1111
+ liςu
1112
+ li[n]
1113
+
1114
+ ��
1115
+
1116
+ inter-cell inter-service interference
1117
+ + wjk[n]
1118
+ � �� �
1119
+ noise
1120
+ (9)
1121
+ yt,u
1122
+ jk [n]=gjk
1123
+ jkAt
1124
+ jk
1125
+
1126
+ ρu
1127
+ jkςu
1128
+ jk[n]
1129
+
1130
+ ��
1131
+
1132
+ desired signal
1133
+ +
1134
+
1135
+ i∈Ku
1136
+ j \{k}
1137
+ gji
1138
+ jkAt
1139
+ ji
1140
+
1141
+ ρu
1142
+ jiςu
1143
+ ji[n]
1144
+
1145
+ ��
1146
+
1147
+ intra-cell intra-service interference
1148
+ +
1149
+ L
1150
+
1151
+ l=1
1152
+ l̸=j
1153
+
1154
+ i∈Ku
1155
+ l
1156
+ gli
1157
+ jkAt
1158
+ li
1159
+
1160
+ ρu
1161
+ liςu
1162
+ li[n]
1163
+
1164
+ ��
1165
+
1166
+ inter-cell intra-service interference
1167
+ +
1168
+ L
1169
+
1170
+ l=1
1171
+
1172
+ i∈Ke
1173
+ l
1174
+ gli
1175
+ jk ˜At
1176
+ l
1177
+
1178
+ ρe
1179
+ liςe
1180
+ li[n]
1181
+
1182
+ ��
1183
+
1184
+ inter-service interference
1185
+ + wjk[n]
1186
+ � �� �
1187
+ noise
1188
+ (10)
1189
+ achievable downlink spectral efficiency of an arbitrary eMBB
1190
+ user k in cell j, is given by
1191
+ SEe
1192
+ jk = τd
1193
+ τc
1194
+ 1
1195
+ T
1196
+ T
1197
+
1198
+ t=1
1199
+ log2(1 + SINRt,e
1200
+ jk), [bits/s/Hz] ,
1201
+ (11)
1202
+ where τd/τc accounts for the estimation overhead,
1203
+ SINRt,e
1204
+ jk =
1205
+ ˜At
1206
+ jρe
1207
+ jk
1208
+ ���E
1209
+
1210
+ gjk
1211
+ jk
1212
+ ����
1213
+ 2
1214
+ L
1215
+
1216
+ l=1
1217
+ K
1218
+
1219
+ i=1
1220
+ ϱt
1221
+ li E
1222
+
1223
+ |gli
1224
+ jk|2
1225
+
1226
+ − ˜At
1227
+ jρe
1228
+ jk
1229
+ ���E
1230
+
1231
+ gjk
1232
+ jk
1233
+ ����
1234
+ 2
1235
+ +σ2
1236
+ d
1237
+ ,
1238
+ (12)
1239
+ is the effective SINR of user k ∈ Ke
1240
+ j, where the expectations
1241
+ are taken with respect to the random channel realizations, and
1242
+ ϱt
1243
+ li =
1244
+
1245
+ At
1246
+ liρu
1247
+ li,
1248
+ if i ∈ Ku
1249
+ l ,
1250
+ ˜At
1251
+ lρe
1252
+ li,
1253
+ if i ∈ Ke
1254
+ l .
1255
+ (13)
1256
+ The expression of the achievable SE shown in (11) holds
1257
+ for any choice of precoding scheme, any channel estimator
1258
+ and any channel distributions. Importantly, it accounts for
1259
+ any choice of coexistence technique between heterogeneous
1260
+ services, namely puncturing or superposition coding. The
1261
+ infinite-blocklength analysis above is established upon the
1262
+ assumption of block-fading channel model, entailing that each
1263
+ eMBB codeword has infinite length that spans a large number
1264
+ of independent fading realizations. This assumption cannot
1265
+ be applied to the URLLC case. As per the URLLC user,
1266
+ we consider a nonasymptotic analysis of the downlink error
1267
+ probability on a slot basis by applying the finite-blocklength
1268
+ information-theoretic results established in [8]. Firstly, we
1269
+ rewrite (10) as
1270
+ yt,u
1271
+ jk [n] = gjk
1272
+ jkqjk[n] + zjk[n],
1273
+ n = 1, . . . , nd,
1274
+ (14)
1275
+ where qjk[n]=At
1276
+ jk
1277
+ �ρu
1278
+ jkςu
1279
+ jk[n], and
1280
+ zjk[n] =
1281
+
1282
+ i∈Ku
1283
+ j \{k}
1284
+ gji
1285
+ jkqji[n] +
1286
+
1287
+ i∈Ke
1288
+ j
1289
+ gji
1290
+ jk ˜At
1291
+ j
1292
+
1293
+ ρe
1294
+ jiςe
1295
+ ji[n]
1296
+ +
1297
+ L
1298
+
1299
+ l=1
1300
+ l̸=j
1301
+
1302
+ � �
1303
+ i∈Ku
1304
+ l
1305
+ gli
1306
+ jkqli[n] +
1307
+
1308
+ i∈Ke
1309
+ l
1310
+ gli
1311
+ jk ˜At
1312
+ l
1313
+
1314
+ ρe
1315
+ liςe
1316
+ li[n]
1317
+
1318
+
1319
+ + wjk[n] .
1320
+ (15)
1321
+ However, URLLC user k in cell j has not access to gjk
1322
+ jk, but
1323
+ performs data decoding by only leveraging its mean value,
1324
+ �gjk
1325
+ jk = E
1326
+
1327
+ (hj
1328
+ jk)Hwjk
1329
+
1330
+ , which is treated as perfect. This
1331
+ estimate is accurate if channel hardening holds. Notice that, the
1332
+ precoded channel gjk
1333
+ jk is frequency-flat and time-invariant over
1334
+ the transmission of the nd-length URLLC codeword in slot t.
1335
+ Moreover, gjk
1336
+ jk remains constant for any other transmission
1337
+ from BS j to user k over slots in the same TDD frame.
1338
+ Given all channels and precoding vectors, the effective noise
1339
+ terms {zjk[n] ∈ C; n = 1, . . . , nd} are random variables
1340
+ conditionally i.i.d. with variance σ2
1341
+ jk, i.e., CN
1342
+
1343
+ 0, σ2
1344
+ jk
1345
+
1346
+ , given
1347
+ by
1348
+ σ2
1349
+ jk =
1350
+
1351
+ i∈Ku
1352
+ j \{k}
1353
+ At
1354
+ jiρu
1355
+ ji|gji
1356
+ jk|2 +
1357
+
1358
+ i∈Ke
1359
+ j
1360
+ ˜At
1361
+ jρe
1362
+ ji|gji
1363
+ jk|2
1364
+ +
1365
+ L
1366
+
1367
+ l=1
1368
+ l̸=j
1369
+
1370
+ � �
1371
+ i∈Ku
1372
+ l
1373
+ At
1374
+ liρu
1375
+ li|gli
1376
+ jk|2+
1377
+
1378
+ i∈Ke
1379
+ l
1380
+ ˜At
1381
+ lρe
1382
+ li|gli
1383
+ jk|2
1384
+
1385
+ � + σ2
1386
+ d .
1387
+ (16)
1388
+ To determine the transmitted codeword
1389
+ qjk = [qjk[1], . . . , qjk[nd]]T ,
1390
+ user k in cell j employs a mismatched scaled nearest-neighbor
1391
+ (SNN) decoder [30], with which selects the codeword �qjk
1392
+ from the codebook C by applying the rule
1393
+ �qjk = arg min
1394
+ �qjk∈C
1395
+ ���yt,u
1396
+ jk − �gjk
1397
+ jk�qjk
1398
+ ���
1399
+ 2
1400
+ ,
1401
+ (17)
1402
+
1403
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
1404
+ 9
1405
+ where yt,u
1406
+ jk =[yt,u
1407
+ jk [1], . . . , yt,u
1408
+ jk [nd]]T ∈Cnd is the received data
1409
+ vector.
1410
+ Let ϵdl
1411
+ jk = Pr {�qjk ̸= qjk} be the downlink error probability
1412
+ experienced by the URLLC user k in cell j achieved by the
1413
+ SNN decoding. An upper bound on ϵdl
1414
+ jk is obtained by using
1415
+ the standard random-coding approach [31],
1416
+ ϵdl
1417
+ jk ≤Egjk
1418
+ jk
1419
+
1420
+ Pr
1421
+ � nd
1422
+
1423
+ n=1
1424
+ ıs(qjk[n], yt,u
1425
+ jk [n]) ≤ log m−1
1426
+ r
1427
+ ����gjk
1428
+ jk
1429
+ ��
1430
+ ,
1431
+ (18)
1432
+ where m=2b is the number of codewords with length nd that
1433
+ convey b information bits, r is a random variable uniformly
1434
+ distributed in the interval [0, 1] and ıs(qjk[n], yt,u
1435
+ jk [n]) is the
1436
+ generalized information density, given by
1437
+ ıs(qjk[n], yt,u
1438
+ jk [n])
1439
+ = −s
1440
+ ���yt,u
1441
+ jk [n] −�gjk
1442
+ jkqjk[n]
1443
+ ���
1444
+ 2
1445
+ +
1446
+ s|yt,u
1447
+ jk [n]|2
1448
+ 1+sρu
1449
+ jk|�gjk
1450
+ jk|2
1451
+ + log(1+sρu
1452
+ jk|�gjk
1453
+ jk|2) ,
1454
+ (19)
1455
+ for all s > 0. In (18) the expectation is taken over the
1456
+ distribution of gjk
1457
+ jk, and the probability is computed with
1458
+ respect to the downlink data symbol {qjk[n]}nd
1459
+ n=1, the effective
1460
+ additive noise {zjk[n]}nd
1461
+ n=1, and the random variable r. The
1462
+ evaluation of the upper bound in (18) entails a very demanding
1463
+ numerical computation to firstly obtain the probability, and
1464
+ then to numerically tighten the upper bound value to the low
1465
+ error probability target of the URLLC use case by optimizing
1466
+ with respect to s.
1467
+ Luckily, we can reliably approximate the right-hand side
1468
+ of (18) in closed form, hence with a significant relief of the
1469
+ computational burden, by using the saddlepoint approximation
1470
+ provided in [8, Th. 2].
1471
+ The existence of a saddlepoint approximation is guaranteed
1472
+ by the fact that the third derivative of the moment-generating
1473
+ function of −ıs(qjk[n], yt,u
1474
+ jk [n]) exists in a neighborhood of
1475
+ zero delimited by the values ε<0<ε given by [8, Appendix
1476
+ B]
1477
+ ε = −
1478
+
1479
+ (ζb − ζa)2 + 4ζaζb(1 − µ) + ζa − ζb
1480
+ 2ζaζb(1 − µ)
1481
+ ,
1482
+ (20)
1483
+ ε =
1484
+
1485
+ (ζb − ζa)2 + 4ζaζb(1 − µ) − ζa + ζb
1486
+ 2ζaζb(1 − µ)
1487
+ ,
1488
+ (21)
1489
+ where
1490
+ ζa = s(ρu
1491
+ jk|gjk
1492
+ jk − �gjk
1493
+ jk|2 + σ2) ,
1494
+ (22)
1495
+ ζb =
1496
+ s
1497
+ 1 + sρu
1498
+ jk|�gjk
1499
+ jk|2 (ρu
1500
+ jk|gjk
1501
+ jk|2 + σ2) ,
1502
+ (23)
1503
+ µ =
1504
+ s2 ���ρu
1505
+ jk|gjk
1506
+ jk|2 + σ2 − (gjk
1507
+ jk)
1508
+ ∗�gjk
1509
+ jkρu
1510
+ jk
1511
+ ���
1512
+ 2
1513
+ ζaζb(1 + sρu
1514
+ jk|�gjk
1515
+ jk|2)
1516
+ .
1517
+ (24)
1518
+ The saddlepoint approximation hinges on the cumulant-
1519
+ generating function of −ıs(qjk[n], yt,u
1520
+ jk [n]) given by
1521
+ υ(ε) = log E
1522
+
1523
+ e−εıs(qjk[n],yt,u
1524
+ jk [n])�
1525
+ ,
1526
+ (25)
1527
+ on its first derivative υ′(ζ), and second derivative υ′′(ζ), for
1528
+ all ε ∈ (ε, ε)
1529
+ υ(ε) =−ε log(1 + sρu
1530
+ jk|�gjk
1531
+ jk|2)
1532
+ − log(1 + (ζb − ζa)ε − ζaζb(1 − µ)ε2)
1533
+ (26)
1534
+ υ′(ε) =− log(1 + sρu
1535
+ jk|�gjk
1536
+ jk|2)
1537
+
1538
+ (ζb − ζa) − 2ζaζb(1 − µ)ε
1539
+ 1 + (ζb − ζa)ε − ζaζb(1 − µ)ε2
1540
+ (27)
1541
+ υ′′(ε) =
1542
+
1543
+ (ζb − ζa) − 2ζaζb(1 − µ)ε
1544
+ 1 + (ζb − ζa)ε − ζaζb(1 − µ)ε2
1545
+ �2
1546
+ +
1547
+ 2ζaζb(1 − µ)
1548
+ 1 + (ζb − ζa)ε − ζaζb(1 − µ)ε2 .
1549
+ (28)
1550
+ Let m = endR for some strictly positive transmission rate
1551
+ R = (log m)/nd, and let ε ∈ (ε, ε) be the solution to the
1552
+ equation R=−υ′(ε). Let Is be the generalized mutual infor-
1553
+ mation [30] defined as Is = E {ıs(qjk[1], vjk[1])} = −υ′(0).
1554
+ Lastly, consider the critical rate [31, Eq. (5.6.30)] given by
1555
+ Rcr
1556
+ s
1557
+ = −υ′(1). Then, we have three possible saddlepoint
1558
+ approximations for the error probability upper bound [8].
1559
+ If ε ∈ [0, 1], then Rcr
1560
+ s ≤ R ≤ Is and
1561
+ Pr
1562
+ � nd
1563
+
1564
+ n=1
1565
+ ıs(qjk[n], yt,u
1566
+ jk [n]) ≤ log endR − 1
1567
+ r
1568
+
1569
+ ≈end[υ(ε)+εR] [Ψnd,ε(ε)+Ψnd,ε(1−ε)] ,
1570
+ (29)
1571
+ where
1572
+ Ψnd,ε(ℓ) ≜ e
1573
+ 1
1574
+ 2 ndℓ2υ′′(ε)Q
1575
+
1576
+
1577
+
1578
+ ndυ′′(ε)
1579
+
1580
+ .
1581
+ (30)
1582
+ If ε > 1, then R < Rcr
1583
+ s and
1584
+ Pr
1585
+ � nd
1586
+
1587
+ n=1
1588
+ ıs(qjk[n], yt,u
1589
+ jk [n]) ≤ log endR − 1
1590
+ r
1591
+
1592
+ ≈ end[υ(1)+R] �
1593
+ �Ψnd(1, 1) + �Ψnd(0, −1)
1594
+
1595
+ ,
1596
+ (31)
1597
+ where
1598
+ �Ψnd(ℓ1, ℓ2)≜endℓ1[Rcr
1599
+ s −R+ 1
1600
+ 2 υ′′(1)]
1601
+ ×Q
1602
+
1603
+ ℓ1
1604
+
1605
+ ndυ′′(1)+ℓ2
1606
+ nd(Rcr
1607
+ s −R)
1608
+
1609
+ ndυ′′(1)
1610
+
1611
+ .
1612
+ (32)
1613
+ If ε < 0, then R > Is and
1614
+ Pr
1615
+ � nd
1616
+
1617
+ n=1
1618
+ ıs(qjk[n], yt,u
1619
+ jk [n]) ≤ log endR−1
1620
+ r
1621
+
1622
+ ≈ 1−end[υ(ε)+εR] [Ψnd,ε(−ε)−Ψnd,ε(1−ε)] .
1623
+ (33)
1624
+ The saddlepoint approximation is more accurate in the URLLC
1625
+ massive MIMO regime than the conventionally-used normal
1626
+ approximation [23] as the former characterizes the exponential
1627
+ decay of the error probability, i.e., the error-exponent, as a
1628
+ function of the URLLC codeword length, and the transmission
1629
+ rate requirement R, while uses the Berry-Esseen central-limit
1630
+ theorem (used in the normal approximation) to only charac-
1631
+ terize the multiplicative factor following the error-exponent
1632
+ term. The normal approximation, whose formulation directly
1633
+ involves the generalized mutual information, Is, but does not
1634
+
1635
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
1636
+ 10
1637
+ R, is accurate only when Is is close to R. This operating
1638
+ regime does not hold for URLLC wherein R is typically lower
1639
+ than Is to accomplish the very low error probability targets.
1640
+ Once that the approximate upper bounds on the downlink error
1641
+ probability are obtained via saddlepoint approximation, we
1642
+ compute the downlink network availability [8], ηdl, as
1643
+ ηdl = Pr
1644
+
1645
+ ϵdl
1646
+ jk ≤ ϵdl
1647
+ target
1648
+
1649
+ (34)
1650
+ which measures the probability that the target error probability
1651
+ ϵdl
1652
+ target is satisfied by an arbitrary user k in cell j, in presence of
1653
+ interfering users. While the expectation in the error probability
1654
+ definition is taken with respect to the small-scale fading and
1655
+ the effective additive noise, given a large-scale fading real-
1656
+ ization, the probability in the network availability definition is
1657
+ computed with respect to the large-scale fading (i.e., path loss,
1658
+ shadowing etc.). The expression of the network availability
1659
+ shown in (34) holds for any choice of precoding scheme, any
1660
+ channel estimator and any channel distributions. Importantly,
1661
+ it accounts for any choice of coexistence technique between
1662
+ heterogeneous services, namely puncturing or superposition
1663
+ coding.
1664
+ IV. PRECODING AND POWER CONTROL
1665
+ The choice of the precoding scheme and of the downlink
1666
+ power allocation deeply affects the SE of the eMBB users and
1667
+ the network availability for the URLLC users. For the sake of
1668
+ comparison, we herein consider three precoding schemes and
1669
+ three power allocation strategies. The general expression for
1670
+ the precoding vector intended for user k in cell j is given by
1671
+ wjk =
1672
+ vjk
1673
+ ∥vjk∥ ,
1674
+ (35)
1675
+ where the denominator serves to make the average power of
1676
+ the precoding vector unitary, and vjk is next characterized.
1677
+ Multi-cell MMSE (M-MMSE):
1678
+ vM−MMSE
1679
+ jk
1680
+ =
1681
+
1682
+
1683
+ � L
1684
+
1685
+ l=1
1686
+ �Hj
1687
+ l Pl( �Hj
1688
+ l )H+Υj+σ2
1689
+ uIM
1690
+ �−1
1691
+ �Hj
1692
+ jPj
1693
+
1694
+
1695
+ :,k
1696
+ where Pl =diag(pli, . . . , plK)∈RK×K is the matrix with the
1697
+ uplink transmit powers of all the users in cell l as diagonal
1698
+ elements, Υj = �L
1699
+ l=1
1700
+ �K
1701
+ i=1 pliCj
1702
+ li, and �Hj
1703
+ l = [�hj
1704
+ l1 . . . �hj
1705
+ lK].
1706
+ M-MMSE precoding provides a nearly optimal downlink SE
1707
+ but requires each BS to acquire CSI and statistical CSI of all
1708
+ the users of the multi-cell system. Moreover, the computation
1709
+ of the precoding vector, which entails inverting a matrix
1710
+ M ×M, may be demanding for large BS arrays. Although
1711
+ impractical, M-MMSE precoding will serve as benchmark.
1712
+ Regularized zero-forcing (RZF):
1713
+ vRZF
1714
+ jk
1715
+ =
1716
+
1717
+ �Hj
1718
+ j
1719
+
1720
+ ( �Hj
1721
+ j)H �Hj
1722
+ j + σ2
1723
+ uP−1
1724
+ j
1725
+ �−1�
1726
+ :,k
1727
+ .
1728
+ Compared to M-MMSE, RZF precoding requires each BS to
1729
+ estimate the channels of only its users. Moreover, computing
1730
+ the RZF precoding vector is computationally cheaper since
1731
+ the size of the matrix to be inverted is K×K. However, RZF
1732
+ does only suppress the intra-cell interference while, unlike M-
1733
+ MMSE, does not provide to the users any protection mech-
1734
+ anism against inter-cell interference and channel estimation
1735
+ error.
1736
+ Maximum Ratio (MR): vMR
1737
+ jk = �hj
1738
+ jk. It is computationally the
1739
+ cheapest but performance-wise the worst precoding scheme.
1740
+ MR only aims at maximizing the power of the desired signal,
1741
+ providing no interference-suppression mechanism. MR will
1742
+ serve as lower bound on the performance.
1743
+ Properly allocating the downlink power can make all the
1744
+ difference to meet the strict reliability requirements of the
1745
+ URLLC and to improve the SE of the eMBB users. Next, we
1746
+ provide three power allocation schemes that take into account
1747
+ the power budget at the BSs, the adopted eMBB-URLLC
1748
+ coexistence strategy and the URLLC activation pattern, which
1749
+ is known at the BS in the downlink operation.
1750
+ Equal power allocation (EPA): It consists in setting
1751
+ ρu
1752
+ ji = ρmax
1753
+ j
1754
+ At
1755
+ ji
1756
+ ˜At
1757
+ jKe + �
1758
+ k∈Ku
1759
+ j
1760
+ At
1761
+ jk
1762
+ , i ∈ Ku
1763
+ j
1764
+ (36)
1765
+ ρe
1766
+ jk = ρmax
1767
+ j
1768
+ ˜At
1769
+ j
1770
+ ˜At
1771
+ jKe + �
1772
+ i∈Ku
1773
+ j
1774
+ At
1775
+ ji
1776
+ , k ∈ Ke
1777
+ j
1778
+ (37)
1779
+ to satisfy the per-BS power constraint in (8) with equality and
1780
+ allocate the same share of power to each user, regardless of
1781
+ its channel conditions and its service requirements.
1782
+ Weighted fractional power allocation (FPA): it consists in
1783
+ setting the powers as
1784
+ ρu
1785
+ ji =
1786
+ ωρmax
1787
+ j
1788
+ At
1789
+ ji(βj
1790
+ ji)
1791
+ ν
1792
+ (1−ω) ˜At
1793
+ j
1794
+
1795
+ k∈Ke
1796
+ j
1797
+ (βj
1798
+ jk)
1799
+ ν + ω �
1800
+ u∈Ku
1801
+ j
1802
+ At
1803
+ ju(βj
1804
+ ju)
1805
+ ν , i ∈ Ku
1806
+ j
1807
+ (38)
1808
+ ρe
1809
+ jk =
1810
+ (1 − ω)ρmax
1811
+ j
1812
+ ˜At
1813
+ j(βj
1814
+ jk)
1815
+ ν
1816
+ (1−ω) ˜At
1817
+ j
1818
+
1819
+ e∈Ke
1820
+ j
1821
+ (βj
1822
+ je)
1823
+ ν + ω �
1824
+ i∈Ku
1825
+ j
1826
+ At
1827
+ ji(βj
1828
+ ji)
1829
+ ν , k ∈ Ke
1830
+ j
1831
+ (39)
1832
+ where the weight ω ∈ (0, 1) adjusts the amount of downlink
1833
+ power to be allocated to the URLLC users, while ν establishes
1834
+ the power control policy as a function of the average channel
1835
+ gain. An opportunistic power allocation is attained by setting
1836
+ ν > 0, with which more power is allocated to the users with
1837
+ better channel conditions. Conversely, fairness is supported
1838
+ by setting ν < 0, with which more power is allocated to the
1839
+ users with worse channel conditions. If ω ∈ (0.5, 1) a larger
1840
+ share of power is allocated to the URLLC users rather than
1841
+ to the eMBB users, whereas it is the other way around if
1842
+ ω ∈ (0, 0.5). Notice that, if ν = 0 and ω = 0.5, then the FPA
1843
+ reduces to the EPA.
1844
+ Optimal power allocation (OPA) for max product SINR:
1845
+
1846
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
1847
+ 11
1848
+ The powers are the solution of the optimization problem
1849
+ maximize
1850
+ {ρs
1851
+ jk}
1852
+ L
1853
+
1854
+ j=1
1855
+ K
1856
+
1857
+ k=1
1858
+ SINRt,s
1859
+ jk
1860
+ (40a)
1861
+ s.t.
1862
+ K
1863
+
1864
+ k=1
1865
+ ϱt
1866
+ jk ≤ ρmax
1867
+ j
1868
+ , ∀j,
1869
+ (40b)
1870
+ where the superscript s = e if user k ∈ Ke
1871
+ j, s = u otherwise,
1872
+ and ϱt
1873
+ jk is given in (13). Without further entangling the
1874
+ notation in (40), we remark that the SINR of inactive users
1875
+ is fictitiously set to 1 to preserve the optimization problem
1876
+ formulation. This power allocation strategy treats all the users
1877
+ as eMBB users, hence it would be optimal if there would
1878
+ be no URLLC users active in a given slot, by maximizing a
1879
+ lower bound on the sum SE of the multi-cell system. Although
1880
+ the SINR expression in (12) is meaningless when applied to
1881
+ a URLLC user, we can still heuristically plug the URLLC
1882
+ powers resulting from (40) into the error probability analysis
1883
+ and motivate this approach by looking at the performance.
1884
+ All the considered power allocation schemes, in principle, run
1885
+ on a slot-basis in order to adapt the power coefficients to the
1886
+ URLLC activation pattern. Fortunately, these schemes only
1887
+ rely on the knowledge of the statistical CSI which allows to
1888
+ pre-compute some power coefficients or to keep the power
1889
+ allocation for multiple slots/frames in case of no macroscopic
1890
+ changes in the propagation environment. Unlike the EPA and
1891
+ the FPA schemes, the OPA scheme requires a certain degree
1892
+ of cooperation among the BSs which must send statistical CSI
1893
+ to let a central processing unit (e.g., a master BS) compute the
1894
+ SINR of all the users and solve the optimization problem, and
1895
+ feed them back with the power coefficients to use. This would
1896
+ introduce intolerable delay for the URLLC users. Moreover,
1897
+ solving problem (40), although efficiently as a geometric
1898
+ program [5, Th. 7.2], is unlikely to be doable within a time-
1899
+ slot, especially for crowded networks. Hence, the OPA scheme
1900
+ is of limited practical use, but will serve for benchmarking
1901
+ purposes.
1902
+ V. SIMULATION RESULTS
1903
+ In this section, we present and discuss the results of our
1904
+ simulations in which the coexistence of eMBB and URLLC
1905
+ is deeply analyzed under different setups. Specifically, we shed
1906
+ the light on the impact of different factors on the performance,
1907
+ such as the transmission technique and the precoding scheme,
1908
+ the power control strategy, the imperfect CSI and estimation
1909
+ overhead, the pilot contamination, the length and number of
1910
+ slots in a TDD frame, and the characteristics of the URLLC
1911
+ activation pattern.
1912
+ Our simulation scenario consists of a multi-cell massive
1913
+ MIMO system with L = 4 cells. Each cell covers a nominal
1914
+ area of 500×500 squared meters, and is served by a BS, placed
1915
+ at the cell center, equipped with a uniform linear array (ULA)
1916
+ with M = 100 equispaced half-wavelength antenna elements.
1917
+ A wrap-around topology is implemented as in [5, Sec. 4.1.3].
1918
+ The users are dropped uniformly at random over the coverage
1919
+ area but at a minimum distance of 25 m from the BS. In
1920
+ addition, we assume that the URLLC users are distributed
1921
+ uniformly at random in an area of 125×125 squared meters
1922
+ that surrounds the BS. A random realization of the user
1923
+ locations determines a set of large-scale fading coefficients
1924
+ and constitutes a snapshot of the network. For a given network
1925
+ snapshot the achievable downlink SEs of the active eMBB
1926
+ users are computed according to (11), while the downlink
1927
+ error probabilities of the URLLC users are obtained according
1928
+ to the approximations (29)-(33). The cumulative distribution
1929
+ function (CDF) of the SE and the network availability are then
1930
+ drawn over many network snapshots. The channel correlation
1931
+ matrices are generated according to the popular local scatter-
1932
+ ing spatial correlation model [5, Sec. 2.6], and we assume that
1933
+ the scattering is only localized around the users and uniformly
1934
+ distributed at random with delay spread 25◦ degrees [8]. The
1935
+ average channel gain is obtained according to the non-line-
1936
+ of-sight macro cell 3GPP model for 2 GHz carriers [32], and
1937
+ given in dB by
1938
+ βk =−35.3 − 37.6 log10
1939
+ � dk
1940
+ 1 m
1941
+
1942
+ + Fk
1943
+ for an arbitrary user k placed at a distance dk from its BS, and
1944
+ where Fk ∼ N
1945
+
1946
+ 0, σ2
1947
+ sh
1948
+
1949
+ models the log-normal shadowing as
1950
+ an i.i.d. random variable with standard deviation σsh = 4 dB.
1951
+ The transmission bandwidth is 20 MHz, and the receiver noise
1952
+ power equals -94 dBm both for the uplink and the downlink.
1953
+ Moreover, we let ρmax
1954
+ j
1955
+ =46 dBm, j =1, . . . , L, and the uplink
1956
+ transmit power, both for pilot and payload data, be 23 dBm
1957
+ for all the users. We assume that the URLLC packet consists
1958
+ of b=160 bits, yielding a transmission rate R=b/nd, which
1959
+ is suitable for factory automation use cases, such as motion
1960
+ controls, and in line with the low latency requirements [33,
1961
+ Annex A]. Lastly, without loss of generality, we set τu =0 as
1962
+ we only focus on the downlink performance. Unless otherwise
1963
+ stated, we consider TDD frames with length τc = 580 channel
1964
+ uses, given by Tc = 2 ms and Bc = 290 kHz, which supports
1965
+ user mobility up to 67.50 km/h.
1966
+ In the first set of simulations we consider the following
1967
+ setup: K = 20, α = 0.2, au = 10−0.5, τp = 80 (no pilot
1968
+ contamination), T = 5 slots of length nd = 100 channel
1969
+ uses. In Fig. 2 we plot the CDFs of the achievable downlink
1970
+ SE per “active” eMBB user obtained for different precoding
1971
+ and power allocation strategies, both for superposition coding
1972
+ (top subfigure) and puncturing technique (bottom subfigure).
1973
+ Under these assumptions, SPC is greatly superior than PUNC,
1974
+ precoding and power allocation strategies being equal. M-
1975
+ MMSE with OPA gives, as expected, the best SE but EPA per-
1976
+ forms almost equally well, regardless of the precoding scheme.
1977
+ RZF provides a practical excellent trad-off between M-MMSE
1978
+ and MR. These results suggest that we are approximately
1979
+ operating in an interference-free scenario, thanks to the full
1980
+ and partial interference-suppression mechanism provided by
1981
+ M-MMSE and RZF, respectively. As per the FPA strategy,
1982
+ in these simulations we have selected ν = 0.5 to promote
1983
+ an opportunistic power allocation and ω = 0.6 to prioritize
1984
+ the URLLC users. Such a choice does not favor the eMBB
1985
+ users and justify the worst performance of FPA among the
1986
+ considered strategies when SPC is applied.
1987
+
1988
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
1989
+ 12
1990
+ Fig. 2.
1991
+ CDFs of the achievable downlink SE per active eMBB user, for
1992
+ different transmission, precoding and power allocation strategies. Settings:
1993
+ K =20, α=0.2, au =10−0.5, τp =80, T =5, nd =100.
1994
+ Fig. 3.
1995
+ CDFs of the achievable downlink sum SE per cell, for different
1996
+ transmission, precoding and power allocation strategies. Settings: K = 20,
1997
+ α=0.2, au =10−0.5, τp =80, T =5, nd =100.
1998
+ Same conclusions hold for the results shown in Fig. 3 where
1999
+ the CDFs of the corresponding sum SE per cell are illustrated.
2000
+ In these figures, we mainly emphasize the eMBB service
2001
+ outage likely occurring when PUNC is adopted. We define
2002
+ the eMBB service outage, under PUNC operation, as
2003
+ ς out = Pr
2004
+
2005
+
2006
+
2007
+
2008
+ k∈Ke
2009
+ j
2010
+ SEe
2011
+ jk = 0
2012
+
2013
+
2014
+ � ,
2015
+ j =1, . . . , L ,
2016
+ where the probability is computed with respect to the large-
2017
+ scale fading. This probability for a BS to provide no service
2018
+ in a TDD frame to its eMBB users depends on the activation
2019
+ pattern of the URLLC users and the number of slots per
2020
+ frame. We will discuss this aspect in detail later. Under the
2021
+ SPC
2022
+ M-MMSE
2023
+ RZF
2024
+ MR
2025
+ 0
2026
+ 0.2
2027
+ 0.4
2028
+ 0.6
2029
+ 0.8
2030
+ 1
2031
+ PUNC
2032
+ M-MMSE
2033
+ RZF
2034
+ MR
2035
+ 0
2036
+ 0.2
2037
+ 0.4
2038
+ 0.6
2039
+ 0.8
2040
+ 1
2041
+ Fig. 4.
2042
+ Network availability for different transmission, precoding and power
2043
+ allocation strategies. Settings: K = 20, α = 0.2, au = 10−0.5, τp = 80,
2044
+ T =5, nd =100.
2045
+ Fig. 5.
2046
+ Downlink per-user error probability for different transmission and
2047
+ precoding strategies. Settings: EPA, K =20, α=0.2, au =10−0.5, τp =80,
2048
+ T =5, nd =100.
2049
+ settings considered in Fig. 3, the eMBB service outage is quite
2050
+ significant as amounts to about 30%.
2051
+ In Fig. 4 we move to the URLLC performance by showing
2052
+ the downlink network availability achieved when ϵdl
2053
+ target =
2054
+ 10−5. Despite the interference caused by the eMBB users
2055
+ when SPC is performed, both M-MMSE and RZF are able to
2056
+ provide levels of network availability close to one, in line with
2057
+ PUNC, revealing a great ability of suppressing the interference
2058
+ and supporting high reliability. Conversely, MR provides poor
2059
+ performance in SPC when EPA or OPA (which is optimal for
2060
+ the eMBB users) schemes are used. Notice that, our choice for
2061
+ the parameters of the FPA scheme pays off for the combination
2062
+ SPC/MR. The network availability values shown in Fig. 4 are
2063
+ obtained by the error probabilities whose CDFs are illustrated
2064
+ in Fig. 5. To better understand its meaning, the network
2065
+ availability is given by the cross-point between the CDF of the
2066
+ per-user error probability and the vertical line representing the
2067
+ error probability target value, as Fig. 5 highlights (blue circle
2068
+ markers). From this set of simulations, we conclude that SPC
2069
+ is clearly superior to PUNC in terms of SE yet providing very
2070
+ high network availability, when M-MMSE or RZF are carried
2071
+ out. If MR is the only viable option (for instance due to strict
2072
+ complexity or hardware constraints), then SPC with FPA, upon
2073
+ properly setting the design parameters ν and ω, is an effective
2074
+ choice to keep the network availability high while preventing
2075
+
2076
+ 0.5
2077
+ SPC0
2078
+ 1
2079
+ 2
2080
+ 3
2081
+ 4
2082
+ per-user SE [bit/
2083
+ 1
2084
+ 0.5
2085
+ PUNC
2086
+ 0
2087
+ 0.5
2088
+ 1
2089
+ 1.5
2090
+ 2
2091
+ per-user SE [bit/s5
2092
+ 6
2093
+ 7
2094
+ 8
2095
+ s/Hzl
2096
+ MR
2097
+ RZF
2098
+ M-MIMSE
2099
+ -FPA,V=0.5.w=0.6
2100
+ -EPA
2101
+ --OPA
2102
+ 2.5
2103
+ 3
2104
+ 3.5
2105
+ 4
2106
+ s/Hzl0.5SPC0
2107
+ 0
2108
+ 10
2109
+ 20
2110
+ 30
2111
+ 40
2112
+ 50
2113
+ per-cell sum SE [bi
2114
+ 1
2115
+ 0.5
2116
+ PUNC
2117
+ eMBB service outage
2118
+ 0
2119
+ 0
2120
+ 10
2121
+ 20
2122
+ 30
2123
+ per-cell sum SE [bi60
2124
+ 70
2125
+ 80
2126
+ 90
2127
+ /s/Hz)
2128
+ MR
2129
+ RZF
2130
+ M-MMSE
2131
+ -FPA,V=0.5.W=0.6
2132
+ -EPA
2133
+ -OPA
2134
+ 40
2135
+ 50
2136
+ 60
2137
+ /s/HzlPUNC/M-MMSE
2138
+ PUNC/RZF
2139
+ 0.8
2140
+ PUNC/MRT
2141
+ SPC/M-MMSE
2142
+ SPC/RZFutage
2143
+ error
2144
+ robabilityCDF
2145
+ SPC/MRT
2146
+ 0.4
2147
+ FPA. v=0.5. w=0.6
2148
+ M-MMSE
2149
+ andRZF
2150
+ 0.2
2151
+ 0
2152
+ 10-50
2153
+ 10-40
2154
+ 10-30
2155
+ 10-2
2156
+ DL error probabtarget
2157
+ URLLC service
2158
+ MRT
2159
+ 20
2160
+ 10-10
2161
+ 10-5
2162
+ 100
2163
+ ilityON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
2164
+ 13
2165
+ Fig. 6.
2166
+ Average per-user SE achieved by SPC with FPA, for different
2167
+ precoding schemes and values of ν, ω. The average is taken over 200 network
2168
+ snapshots. Settings: K = 20, α = 0.2, au = 10−0.5, τp = 80, T = 5,
2169
+ nd =100.
2170
+ any eMBB service outage.
2171
+ In this regard, we now focus on how to select ν and ω
2172
+ appropriately. By using the same settings as in the first set
2173
+ of simulations, in Fig. 6 we plot the average per-user SE
2174
+ assuming SPC and different precoding schemes with FPA
2175
+ as ν and ω vary. From the eMBB user perspective, it is
2176
+ preferable setting a small value for ω, and ν in the interval
2177
+ [−0.5, 0]. While the former is trivial, the latter needs further
2178
+ discussions. Indeed, recall that positive values for ν enable
2179
+ allocating more power to users with better channel conditions.
2180
+ Since we assume the URLLC users are uniformly distributed
2181
+ in a smaller area surrounding the BSs, it is very likely that they
2182
+ are closer to the BS than most of the eMBB users. Therefore,
2183
+ negative values for ν increase the fairness and improve eMBB
2184
+ users performance. Large values for both ω and ν excessively
2185
+ unbalance the power distribution in favor of the URLLC users,
2186
+ degrading the SE of the eMBB users.
2187
+ Conversely, small values for both ω and ν break down the
2188
+ network availability of the URLLC users in SPC operation,
2189
+ as clearly seen in Fig. 7. Nevertheless, both M-MMSE and
2190
+ RZF are able to provide levels of network availability close
2191
+ to 1 except when ν = −1, while MR is quite sensitive to
2192
+ this parameters tuning. Suppressing the multi-user interference
2193
+ is of a vital importance when SPC is adopted, and RZF,
2194
+ although not dealing with the inter-cell interference, is an
2195
+ excellent trade-off between performance and practicality. Fine-
2196
+ tuning the parameters of the FPA scheme yields satisfying
2197
+ performance when using MR. FPA becomes a valid, heuristic
2198
+ alternative to combat the multi-user interference whenever the
2199
+ latter cannot be removed by the precoding technique.
2200
+ Setting ω becomes pointless when using PUNC with FPA as
2201
+ only URLLC transmissions take place in the considered slot.
2202
+ Hence, in Fig. 8 and Fig. 9 we focus on the average SE per user
2203
+ and the network availability as only ν varies. For both cases we
2204
+ notice that an equal power allocation, i.e., ν =0, is desirable.
2205
+ As per the SE of the eMBB users, negative values of ν support
2206
+ Fig. 7.
2207
+ Network availability achieved by SPC with FPA, for different
2208
+ precoding schemes and values of ν, ω. Settings: K = 20, α = 0.2,
2209
+ au =10−0.5, τp =80, T =5, nd =100.
2210
+ -1
2211
+ -0.5
2212
+ 0
2213
+ 0.5
2214
+ 1
2215
+ 0.2
2216
+ 0.4
2217
+ 0.6
2218
+ 0.8
2219
+ 1
2220
+ 1.2
2221
+ 1.4
2222
+ Fig. 8.
2223
+ Average per-user SE (with 95% confidence interval) achieved by
2224
+ PUNC with FPA, for different precoding schemes and values of ν. The average
2225
+ is taken over 200 network snapshots. Settings: K =20, α=0.2, au =10−0.5,
2226
+ τp =80, T =5, nd =100.
2227
+ lower SEs (e.g., the 95%-likely SE per user), hence the fairness
2228
+ among the users, while large positive values of ν support the
2229
+ peak SE in a greedy fashion, neglecting lower SEs. Therefore,
2230
+ ν =0 is sound if the average SE is targeted, especially when
2231
+ the multi-user interference is partially or fully canceled. As per
2232
+ the network availability of the URLLC users, any choice of
2233
+ ν ∈ [−1, 1] is solid as long as M-MMSE or RZF are employed,
2234
+ while the performance of MR is relatively penalized whenever
2235
+ a non-neutral choice for ν is taken. Presumably, the number of
2236
+ URLLC users simultaneously active in the same slot (resulting
2237
+ from the chosen values of α and au) is such that the multi-user
2238
+ interference is not significant.
2239
+ Next, we evaluate the performance as a function of the
2240
+ number of the slots in a TDD frame, T, and the size of
2241
+ the slot, nd, which in turn determines the URLLC codeword
2242
+ length. In this set of simulations and hereafter, we omit the
2243
+ results achieved by MR and only consider FPA with ν = 0
2244
+ and ω = α motivated by the previous results. Fig. 10 shows
2245
+ the CDFs of the sum SE per cell, for three different setups:
2246
+
2247
+ St
2248
+ 4.5
2249
+ 4uperpositionCoding3.5
2250
+ 3
2251
+ 2.5
2252
+ 2
2253
+ RZF
2254
+ 1.5
2255
+ 1
2256
+ 0.5
2257
+ MR
2258
+ 0
2259
+ 0.2
2260
+ 0.4 0.5 0.6
2261
+ 0.8
2262
+ 0.95
2263
+ -1
2264
+ -0M-MMSE
2265
+ 0.5
2266
+ 1
2267
+ 0
2268
+ .5
2269
+ VSuperposition Coding
2270
+ 10.8
2271
+ M-MMS
2272
+ 0.6
2273
+ MR
2274
+ 0.4
2275
+ 0.2
2276
+ 0
2277
+ 0.95
2278
+ 0.8
2279
+ 0.6
2280
+ 0.5
2281
+ m
2282
+ 0.4
2283
+ 0.2
2284
+ 1E
2285
+ RZF
2286
+ -1
2287
+ -0.5
2288
+ 0
2289
+ 0.5ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
2290
+ 14
2291
+ -1
2292
+ -0.5
2293
+ 0
2294
+ 0.5
2295
+ 1
2296
+ 0.8
2297
+ 0.85
2298
+ 0.9
2299
+ 0.95
2300
+ 1
2301
+ Fig. 9.
2302
+ Network availability achieved by PUNC with FPA, for different
2303
+ precoding schemes and values of ν. Settings: K =20, α=0.2, au =10−0.5,
2304
+ τp =80, T =5, nd =100.
2305
+ 0
2306
+ 10
2307
+ 20
2308
+ 30
2309
+ 40
2310
+ 50
2311
+ 60
2312
+ 70
2313
+ 80
2314
+ 90
2315
+ 0
2316
+ 0.5
2317
+ 1
2318
+ 0
2319
+ 10
2320
+ 20
2321
+ 30
2322
+ 40
2323
+ 50
2324
+ 60
2325
+ 70
2326
+ 80
2327
+ 90
2328
+ 0
2329
+ 0.5
2330
+ 1
2331
+ 0
2332
+ 10
2333
+ 20
2334
+ 30
2335
+ 40
2336
+ 50
2337
+ 60
2338
+ 70
2339
+ 80
2340
+ 90
2341
+ 0
2342
+ 0.5
2343
+ 1
2344
+ SPC
2345
+ SPC
2346
+ PUNC
2347
+ PUNC
2348
+ PUNC
2349
+ SPC
2350
+ Fig. 10.
2351
+ CDFs of the achievable downlink sum SE per cell, for different
2352
+ transmission and precoding strategies, as the number of slots per frame varies.
2353
+ Settings: FPA with ν = 0 and ω = 0.2, K = 20, α = 0.2, au = 10−0.5,
2354
+ τp =80.
2355
+ (i) nd =25, T =20, (ii) nd =50, T =10, and (iii) nd =100,
2356
+ T = 5. The structure of the TDD frame has not a significant
2357
+ impact on the SE of the eMBB users when SPC is used.
2358
+ Conversely, that deeply affects the per-cell sum SE in case
2359
+ of PUNC. Indeed, increasing the number of slots per frame
2360
+ makes the probability of having eMBB service outage smaller
2361
+ as it increases the opportunities for an eMBB user to find slots
2362
+ with no active URLLC users. This argument is supported by
2363
+ the results in Fig. 10 in which the eMBB service outage equals
2364
+ 0.01, 0.0725 and 0.2875 when T = 20, T = 10 and T = 5,
2365
+ respectively. On the other hand, with fewer slots, eMBB users
2366
+ might be active for longer time, thereby experiencing higher
2367
+ SE. This explains the larger variations of the per-cell sum SE
2368
+ as T is decreased.
2369
+ The length of the slot directly affects the performance of
2370
+ the URLLC users. As we can see in Fig. 11, the network
2371
+ availability increases drastically with the length of the slot
2372
+ (i.e., the URLLC codeword length). In fact, the length of
2373
+ the URLLC codeword determines the transmission rate of the
2374
+ URLLC users as R=b/nd, thus the shorter the codeword the
2375
+ Fig. 11.
2376
+ Network availability, for different transmission and precoding
2377
+ strategies, as the length of the slot varies. Settings: FPA with ν = 0 and
2378
+ ω =0.2, K =20, α=0.2, au =10−0.5, τp =80.
2379
+ higher the rate requirement to be reliably achieved and, in turn,
2380
+ the larger the error probability.2 Again, SPC is the technique
2381
+ that overall guarantees the best performance to both the eMBB
2382
+ and URLLC users as its main limitation, namely the caused
2383
+ multi-user interference, is overcome by using interference-
2384
+ suppression-based precoding schemes. Lastly, although letting
2385
+ the URLLC transmissions span many channel uses is benefi-
2386
+ cial in terms of network availability, the latency requirements
2387
+ impose to localize the transmissions in time.
2388
+ Now, we move our focus on the impact of the pilot
2389
+ contamination and estimation overhead on the performance.
2390
+ By fixing the TDD frame length and the number of slots
2391
+ per frame, we vary the length of the uplink training, hence
2392
+ the number of available orthogonal pilots, and the length of
2393
+ each slot accordingly. In Fig. 12 we show how the average
2394
+ sum SE per cell evolves in different operating regimes with
2395
+ respect to the uplink training length. In these simulations,
2396
+ we assume K = 20, α = 0.2, τc = 580 and T = 5. Small
2397
+ values of τp entails low channel estimation overhead but high
2398
+ levels of pilot contamination which reduces the effectiveness
2399
+ of the precoding. Our pilot assignment scheme preserves the
2400
+ performance of the URLLC users by assigning them unique
2401
+ pilots if available, otherwise pilots are assigned randomly and
2402
+ contamination hits any user indiscriminately. The maximum
2403
+ number of URLLC users potentially active in this scenario
2404
+ is, according to the chosen parameter, 16. Hence, pilots are
2405
+ assigned randomly when τp = 10 causing both intra- and
2406
+ inter-cell pilot contamination and providing a low sum SE
2407
+ per cell, namely about 30 bit/s/Hz with SPC and less than 10
2408
+ bit/s/Hz with PUNC. The performance worsens when τp =20
2409
+ as the eMBB users have to share only 4 orthogonal pilots
2410
+ since the protection mechanism of the URLLC users is now
2411
+ triggered. As we increase the value of τp, the intra-cell pilot
2412
+ 2The random-coding union bound in (18) defines the error probability as
2413
+ the probability that the average generalized information density is smaller than
2414
+ the transmission rate requirement.
2415
+
2416
+ 0.8Network Availability,
2417
+ 0.6
2418
+ 0.4
2419
+ 0.2
2420
+ 0
2421
+ RZF/SPC
2422
+ M-MMSE/SPC
2423
+ nd=25, T=20
2424
+ RZF/PU
2425
+ nd=50,T=10
2426
+ nd=100, T=5NC
2427
+ M-MMSE/PUNCON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
2428
+ 15
2429
+ Fig. 12.
2430
+ Average SE per cell (with 95% confidence interval), for different
2431
+ transmission and precoding strategies, as τp (and nd) varies. The average is
2432
+ taken over 200 network snapshots. Settings: FPA with ν = 0 and ω = 0.2,
2433
+ K =20, α=0.2, au =10−0.5, τc =580, T =5.
2434
+ contamination is primarily reduced by assigning orthogonal
2435
+ pilots to eMBB users of the same cell. If τp ≥32 then intra-cell
2436
+ pilot contamination is prevented and the inter-cell interference
2437
+ among the eMBB users remains the only impairment. The
2438
+ sum SE per cell keep growing up to τp = 80, when all
2439
+ the users in the network are assigned mutual orthogonal
2440
+ pilots and the benefits of having no pilot contamination at all
2441
+ overcome the penalty from increasing the estimation overhead.
2442
+ Trivially, there are no benefits in the channel estimation when
2443
+ further increasing τp, while the estimation overhead turns to
2444
+ be expensive and drastically lowers the sum SE per cell.
2445
+ Finally, notice that RZF and M-MMSE provide essentially
2446
+ the same performance when both the intra- and inter-cell pilot
2447
+ contamination occur, because the ability of suppressing the
2448
+ multi-user interference is poor for both the schemes.
2449
+ As per the URLLC users, pilot contamination heavily affects
2450
+ the network availability when τp <16, especially when SPC is
2451
+ employed and despite a long slot lowers the rate requirements,
2452
+ as we can observe in Fig. 13. Pilot contamination among
2453
+ URLLC users is destructive mainly because they are likely to
2454
+ be close to the BS and to each other, experiencing strong inter-
2455
+ ference that cannot be resolved when their channel estimates
2456
+ are correlated. Hence, our approach aiming at prioritizing the
2457
+ URLLC users in the pilot assignment is technically sound. In
2458
+ addition, increasing the estimation overhead deeply penalizes
2459
+ the network availability since more resources are subtracted to
2460
+ the data transmission, namely the slot length reduces and, as
2461
+ already explained earlier, the rate requirements of the URLLC
2462
+ users increase.
2463
+ Next we study how the performance are affected by the
2464
+ random activation pattern and the number of potentially active
2465
+ URLLC users per frame. Fig. 14 shows the average sum SE
2466
+ per cell as au and α vary, assuming different transmission
2467
+ and precoding schemes, and FPA with ν = 0 and ω = α.
2468
+ Notice that, proportionally increasing ω to α is a reasonable
2469
+ Fig. 13.
2470
+ Network availability, for different transmission and precoding
2471
+ strategies, as τp (and nd) varies. Settings: FPA with ν = 0 and ω = 0.2,
2472
+ K =20, α=0.2, au =10−0.5, τc =580, T =5.
2473
+ Fig. 14.
2474
+ Average SE per cell, for different transmission and precoding
2475
+ strategies, as au and α vary. The average is taken over 200 network snapshots.
2476
+ Settings: FPA with ν = 0 and ω = α, K = 20, τc = 580, f = 4, T = 5,
2477
+ nd =100.
2478
+ approach for SPC as more power is allocated to an increasing
2479
+ number of potentially active URLLC users, especially for
2480
+ large values of au. In these simulations, we assume two TDD
2481
+ frame configurations: (i) f = 4, T = 5, nd = 100, and (ii)
2482
+ f = 3, T = 8, nd = 65 (whose results are instead shown
2483
+ in Fig. 15). First, we observe that similar average sum SE per
2484
+ cell can be achieved by adopting the considered TDD frame
2485
+ configurations: pilot contamination is what slightly degrades
2486
+ the performance of the eMBB users when using the second
2487
+ frame configuration. The performance of PUNC converges to
2488
+ that of SPC when au ≥ 10−2, hence for sparse activation
2489
+ patterns, as expected. Again, the performance gap between
2490
+ RZF and M-MMSE reduces in the second scenario (Fig. 15)
2491
+ as the inter-cell pilot contamination decreases the ability of
2492
+ M-MMSE in suppressing the multi-user interference. PUNC
2493
+ provides eMBB service outage for large values of au, whereas
2494
+
2495
+ 70
2496
+ --.--RZF
2497
+ ZH
2498
+ I--- M-MMSE
2499
+ 60
2500
+ [bit/s/
2501
+ everyone
2502
+ contam.
2503
+ 50
2504
+ ellSPCper
2505
+ 40
2506
+ inter-and intra-cell pilot contar
2507
+ intra-cell pilo
2508
+ inter-cellpilot
2509
+ SE
2510
+ contam.
2511
+ 30
2512
+ eMBB users
2513
+ eMBB users
2514
+ 20
2515
+ inter- and
2516
+ 10
2517
+ PL
2518
+ 0
2519
+ 114
2520
+ 112
2521
+ 110
2522
+ 108
2523
+ 104
2524
+ 100
2525
+ 10
2526
+ T
2527
+ 20
2528
+ 30
2529
+ 40
2530
+ 60
2531
+ 80
2532
+ pno pilot
2533
+ contamination
2534
+ JNC
2535
+ 96
2536
+ 86
2537
+ 76
2538
+ 66
2539
+ 56
2540
+ 100
2541
+ 150
2542
+ 200
2543
+ 250
2544
+ 300AL
2545
+ PUNC
2546
+ tam.everyone
2547
+ 0.95
2548
+ SPC
2549
+ ility,
2550
+ ....Network Availal
2551
+ inter-and intra-cell pilot co
2552
+ inter-and intra-cell p
2553
+ contam. eMBB users
2554
+ inter-cell pilot
2555
+ contam.
2556
+ 0.85
2557
+ eMBB users
2558
+ 0.8
2559
+ 0.75
2560
+ 114
2561
+ 112
2562
+ 110
2563
+ 108
2564
+ 104
2565
+ 100
2566
+ 10
2567
+ 20
2568
+ 30
2569
+ 40
2570
+ 60
2571
+ 80
2572
+ KQno pilot
2573
+ contamination
2574
+ -.----RZF
2575
+ ------ M-MMSE
2576
+ 96
2577
+ 86
2578
+ 76
2579
+ 66
2580
+ 56
2581
+ 100
2582
+ 150
2583
+ 200
2584
+ 250
2585
+ 300f-4,T=5,
2586
+ SPC
2587
+ 70
2588
+ M-MMSIna = 100bit/s/Hz
2589
+ 60
2590
+ 50
2591
+ SE per cell [
2592
+ 40
2593
+ 30
2594
+ 20
2595
+ 10
2596
+ PUNC
2597
+ 0
2598
+ 100
2599
+ 10-1
2600
+ 10-2
2601
+ 10-3
2602
+ 10-4RZF
2603
+ 0.8
2604
+ 0.6
2605
+ 0.4
2606
+ 0.2ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
2607
+ 16
2608
+ Fig. 15.
2609
+ Average SE per cell, for different transmission and precoding
2610
+ strategies, as au and α vary. The average is taken over 200 network snapshots.
2611
+ Settings: FPA with ν = 0 and ω = α, K = 20, τc = 580, f = 3, T = 8,
2612
+ nd =65.
2613
+ Fig. 16.
2614
+ Network availability, for different precoding strategies, as au and
2615
+ α vary. The average is taken over 200 network snapshots. Settings: SPC and
2616
+ FPA with ν =0 and ω =α, K =20, τc =580. Two TDD frame configurations
2617
+ are considered.
2618
+ SPC is still able to cancel the URLLC user interference and to
2619
+ provide excellent SEs. Lastly, we observe that if the 80% of
2620
+ the users requests URLLC, then the performance of the eMBB
2621
+ users is reduced of almost one third with respect to the case
2622
+ α=0.2. This result is mainly due to the chosen value of ω in
2623
+ the FPA scheme that aims to favor the URLLC performance
2624
+ as the number of URLLC users increases.
2625
+ The performance achieved by the two considered TDD
2626
+ frame configurations appreciably differ in terms of network
2627
+ availability as shown in Fig. 16 for SPC and Fig. 17 for PUNC.
2628
+ In both cases, reducing the length of the slot leads to about
2629
+ a 10% performance loss, while the pilot contamination only
2630
+ concerns the eMBB users. This performance gap is slightly
2631
+ more pronounced when using PUNC because the entire BS
2632
+ power is distributed among the URLLC users causing stronger
2633
+ mutual interference. Overall, the first TDD frame configuration
2634
+ turns to be quite robust to any of the considered transmission
2635
+ Fig. 17.
2636
+ Network availability, for different precoding strategies, as au and
2637
+ α vary. Settings: PUNC and FPA with ν =0 and ω =α, K =20, τc =580.
2638
+ Two TDD frame configurations are considered.
2639
+ Fig. 18.
2640
+ eMBB service outage, for different precoding strategies, as au and
2641
+ α vary. Settings: PUNC and FPA with ν =0 and ω =α, K =20, τc =580.
2642
+ Two TDD frame configurations are considered.
2643
+ and precoding strategies, considered random URLLC activa-
2644
+ tion pattern and URLLC user load.
2645
+ A final aspect to be analyzed for this set of simulations
2646
+ is how the probability of eMBB service outage varies with
2647
+ au and α when PUNC is adopted. This would complete
2648
+ the picture on which operating points PUNC is an effective
2649
+ choice for the eMBB users too, and importantly, further remark
2650
+ the relevance of properly structuring the TDD frame. As
2651
+ we can see in Fig. 18, the advantage of adopting the TDD
2652
+ frame configuration with T = 8 slots, when using PUNC,
2653
+ consists in better preventing the eMBB service outage than the
2654
+ configuration with T = 5. For instance, when au = 10−1 and
2655
+ α=0.8 or α=0.6, partitioning the share of the frame devoted
2656
+ to the data transmission in 8 slots enables to halve the eMBB
2657
+ outage service compared to the case where 5 slots are adopted.
2658
+ Overall, PUNC can compete with SPC only in scenarios with
2659
+ low URLLC traffic loads, upon properly structuring the TDD
2660
+ frame, as long as a moderate eMBB performance loss is
2661
+ tolerated, either in terms of sum SE per cell or of eMBB
2662
+ service outage. On the other hand, SPC hinges on precoding
2663
+ schemes able to suppress the multi-user interference which, in
2664
+ turn, leverages the spatial degrees of freedom available at the
2665
+
2666
+ f-3,T= 8
2667
+ 70
2668
+ SPCnd = 65[bit/s/Hz
2669
+ V-
2670
+ 60
2671
+ 50
2672
+ SE per cell [
2673
+ 40
2674
+ 30
2675
+ 20
2676
+ 10
2677
+ PUNC
2678
+ 0
2679
+ 100
2680
+ 10~1
2681
+ au
2682
+ 10-2
2683
+ 10-3
2684
+ 10-4RZF
2685
+ 0.8
2686
+ 0.6
2687
+ 0.4
2688
+ 0.2Superposition Coding0.98
2689
+ Availability,
2690
+ 0.96
2691
+ 0.94
2692
+ M-MMSE
2693
+ 0.92
2694
+ Network
2695
+ 0.9
2696
+ 0.88
2697
+ 0.86
2698
+ RZF
2699
+ 100.0
2700
+ 10-0.1
2701
+ f =3, T =8, nd=6
2702
+ 10-0.5
2703
+ 10-1.0
2704
+ 0.24, T = 5, nd = 100
2705
+ 0.8
2706
+ 0.6
2707
+ 0.4
2708
+ 3=0Puncturing0.98
2709
+ Network Availability,
2710
+ 0.96
2711
+ f
2712
+ 0.94
2713
+ 0.92
2714
+ 0.9
2715
+ f = 3,T - 8,nd = 65
2716
+ 0.88
2717
+ 0.86
2718
+ RZF
2719
+ 100.0
2720
+ 10-0.1
2721
+ 10-0.2
2722
+ 10-0.5
2723
+ 10-1.0
2724
+ 0.2, T = 5,nd = 100
2725
+ M-MMSE
2726
+ 0.8
2727
+ 0.6
2728
+ 0.4
2729
+ α=wf =3,T =eMBB service outage
2730
+ eMBB service outage
2731
+ 0.8
2732
+ 0.8
2733
+ 0.6
2734
+ 0.6
2735
+ 0.4
2736
+ 0.2
2737
+ 0.2
2738
+ 0
2739
+ 0
2740
+ 10-1.0
2741
+ du
2742
+ 10-2.0
2743
+ 0.8
2744
+ 0.8
2745
+ 0.6
2746
+ 0.6
2747
+ 10-3.0
2748
+ 0.4
2749
+ 0.4
2750
+ 10-4.0
2751
+ 0.2
2752
+ 310-0.2
2753
+ 10-0.5
2754
+ 10-1.0
2755
+ du
2756
+ 10-3.0
2757
+ 10-4.0
2758
+ 0.2ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
2759
+ 17
2760
+ 10
2761
+ 20
2762
+ 30
2763
+ 40
2764
+ 50
2765
+ 60
2766
+ 0
2767
+ 20
2768
+ 40
2769
+ 60
2770
+ 80
2771
+ 100
2772
+ 120
2773
+ SPC
2774
+ SPC
2775
+ PUNC
2776
+ PUNC
2777
+ Fig. 19.
2778
+ Average SE per cell (with 95% confidence interval), for different
2779
+ transmission and precoding strategies, as K and τc vary. The average is taken
2780
+ over 200 network snapshots. Settings: FPA with ν =0 and ω =0.2, α=0.2,
2781
+ au =10−1, f =3, T =5.
2782
+ BS and the high accuracy of the acquired CSI.
2783
+ Finally, we evaluate the performance varying the total
2784
+ number of users and the TDD frame length. Fig. 19 shows
2785
+ the average sum SE per cell, for different transmission and
2786
+ precoding strategies, as the number of users per cell, K, grows
2787
+ from 10 to 60, and considering two different TDD frame
2788
+ lengths, namely 580 and 300 channel uses. The latter may
2789
+ support a shorter coherence time and a narrower coherence
2790
+ bandwidth as well as a higher user mobility compared to the
2791
+ case with 580 channel uses. However, a shorter frame entails
2792
+ less resources that can be allocated to the data transmission
2793
+ and uplink training. In these simulations we assume FPA with
2794
+ ν = 0 and ω = 0.2, α = 0.2, au = 10−1, T = 5 and pilot reuse
2795
+ factor f = 3. Moreover, as τp = fK and τc is fixed, for each
2796
+ value of K we have different configurations of uplink training
2797
+ and slot length, i.e., τp and nd, respectively. From Fig. 19
2798
+ we observe the average sum SE per cell increasing with K,
2799
+ which demonstrates the great ability of SPC with M-MMSE
2800
+ and RZF to spatially multiplex the users. The average sum SE
2801
+ per cell saturates for values of K larger than 60 for τc =580,
2802
+ and around 40 for τc = 300 wherein the channel estimation
2803
+ overhead heavily burden the SE. PUNC is far inferior to
2804
+ SPC because allocates less resources to the eMBB users and
2805
+ the performance gap increases with K as the number of
2806
+ URLLC users per cell grows proportionally. Therefore, letting
2807
+ K increase makes punctured slots more likely, which not only
2808
+ subtracts resources to the eMBB user reducing its SE but also
2809
+ increases the eMBB service outage, as shown in Table III.
2810
+ Notice that, the eMBB service outage does not change when
2811
+ varying τc as long as T is fixed.
2812
+ Table III and Table IV show the network availability for
2813
+ different transmission and precoding strategies, and different
2814
+ values of K, also emphasizing how τp and nd vary accordingly
2815
+ to meet the TDD frame length. In particular, Table III shows
2816
+ the performance achieved by considering τc = 580, while
2817
+ TABLE III
2818
+ NETWORK AVAILABILITY AND EMBB SERVICE OUTAGE, τc =580
2819
+ K
2820
+ τp
2821
+ nd
2822
+ ηdl
2823
+ ς out
2824
+ SPC
2825
+ PUNC
2826
+ PUNC
2827
+ M-MMSE
2828
+ RZF
2829
+ M-MMSE
2830
+ RZF
2831
+ 10
2832
+ 30
2833
+ 110
2834
+ 0.9989
2835
+ 0.9966
2836
+ 1
2837
+ 0.9989
2838
+ 0.0012
2839
+ 20
2840
+ 60
2841
+ 104
2842
+ 0.9988
2843
+ 0.9957
2844
+ 0.9944
2845
+ 0.9906
2846
+ 0.0038
2847
+ 30
2848
+ 90
2849
+ 98
2850
+ 0.9988
2851
+ 0.9950
2852
+ 0.9934
2853
+ 0.9893
2854
+ 0.0225
2855
+ 40
2856
+ 120
2857
+ 92
2858
+ 0.9969
2859
+ 0.9885
2860
+ 0.9881
2861
+ 0.9819
2862
+ 0.0625
2863
+ 50
2864
+ 150
2865
+ 86
2866
+ 0.9864
2867
+ 0.9787
2868
+ 0.9790
2869
+ 0.9672
2870
+ 0.1050
2871
+ 60
2872
+ 180
2873
+ 80
2874
+ 0.9807
2875
+ 0.9697
2876
+ 0.9728
2877
+ 0.9601
2878
+ 0.1737
2879
+ TABLE IV
2880
+ NETWORK AVAILABILITY AND EMBB SERVICE OUTAGE, τc =300
2881
+ K
2882
+ τp
2883
+ nd
2884
+ ηdl
2885
+ ς out
2886
+ SPC
2887
+ PUNC
2888
+ PUNC
2889
+ M-MMSE
2890
+ RZF
2891
+ M-MMSE
2892
+ RZF
2893
+ 10
2894
+ 30
2895
+ 54
2896
+ 0.7936
2897
+ 0.7683
2898
+ 0.7844
2899
+ 0.7534
2900
+ 0.0012
2901
+ 20
2902
+ 60
2903
+ 48
2904
+ 0.6786
2905
+ 0.6353
2906
+ 0.6905
2907
+ 0.6685
2908
+ 0.0038
2909
+ 30
2910
+ 90
2911
+ 42
2912
+ 0.4796
2913
+ 0.4296
2914
+ 0.5646
2915
+ 0.5435
2916
+ 0.0225
2917
+ 40
2918
+ 120
2919
+ 36
2920
+ 0.1813
2921
+ 0.1457
2922
+ 0.3192
2923
+ 0.3192
2924
+ 0.0625
2925
+ 50
2926
+ 150
2927
+ 30
2928
+ 0.0021
2929
+ 0
2930
+ 0.0250
2931
+ 0.0250
2932
+ 0.1050
2933
+ 60
2934
+ 180
2935
+ 24
2936
+ 0
2937
+ 0
2938
+ 0
2939
+ 0
2940
+ 0.1737
2941
+ Table IV shows the performance achieved with τc = 300.
2942
+ The TDD frame with τc = 580 allows to achieve a network
2943
+ availability above 96% up to 60 users per cell (of which 12
2944
+ are URLLC users) with any of the considered transmission
2945
+ and precoding techniques, meaning that such an amount of
2946
+ resources are sufficient to excellently support the considered
2947
+ URLLC user loads and their activation pattern. Conversely,
2948
+ the network availability supported by the TDD frame with
2949
+ τc = 300, reported in Table IV, is considerably lower, even
2950
+ close (or equal) to zero for K ≥50, emphasizing how sensitive
2951
+ the network availability is to the length of the TDD frame,
2952
+ hence to the amount of available resources. Importantly, we
2953
+ observe the decreasing trend of the network availability as
2954
+ K increases, which for PUNC is milder and mainly due to
2955
+ the shorter URLLC codeword length, but for SPC is severe
2956
+ and mainly due to the increase of the multi-user interference.
2957
+ Indeed, the results in Table IV clearly confirms that PUNC is
2958
+ more robust than SPC when K ≥20.
2959
+ VI. CONCLUSION
2960
+ In this paper, we considered the non-orthogonal multiplex-
2961
+ ing of heterogeneous services, namely the enhanced mobile
2962
+ broadband (eMBB) and the ultra-reliable low-latency commu-
2963
+ nication (URLLC), in the downlink of a multi-cell massive
2964
+ MIMO system. eMBB and URLLC have opposite characteris-
2965
+ tics and diverse requirements. eMBB transmissions involve a
2966
+ large payload that spans multiple radio frames, and demand for
2967
+ high spectral efficiency. While, URLLC users intermittently
2968
+ transmit small payloads in a very short time demanding for
2969
+ low latency and successful probability in the order of 10−5.
2970
+ Such a heterogeneity calls for effective resource allocation
2971
+ strategies to let eMBB and URLLC peacefully coexist. Firstly,
2972
+ we provided a unified information-theoretic framework to
2973
+ assess the spectral efficiency (SE) of the eMBB in the infinite-
2974
+ blocklength ergodic regime, and the error probability of the
2975
+ URLLC in the nonasymptotic finite-blocklength regime. Both
2976
+
2977
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
2978
+ 18
2979
+ analyses encompass imperfect channel state information (CSI)
2980
+ acquisition at the base stations (BSs) via uplink pilot trans-
2981
+ missions, pilot contamination and pilot overhead, spatially
2982
+ correlated channels and the lack of CSI at the users. Secondly,
2983
+ we generalized the proposed framework to accommodate two
2984
+ alternative coexistence strategies: puncturing (PUNC) and
2985
+ superposition coding (SPC). The former prevents the inter-
2986
+ service interference aiming to protect the URLLC reliability,
2987
+ while the latter accepts it aiming to maintain the eMBB
2988
+ service. Thirdly, we numerically evaluated the performance
2989
+ achieved by PUNC and SPC under different precoding and
2990
+ power allocation schemes, and subject to different configu-
2991
+ rations of the time-division duplex radio frame and URLLC
2992
+ random activation pattern. Simulation results revealed that
2993
+ the spatial degrees of freedom available at the BSs, when
2994
+ fully exploited by interference-suppression-based precoding
2995
+ schemes, and upon a high-quality CSI acquisition, enable to
2996
+ significantly resolve the multi-user interference caused by the
2997
+ SPC operation, providing way higher eMBB SE than PUNC,
2998
+ yet ensuring similar great levels of error probability for the
2999
+ URLLC. However, whenever these conditions does not hold,
3000
+ e.g., when a severe pilot contamination degrades the channel
3001
+ estimates or the degrees of freedom are not sufficient to
3002
+ handle the interference between many users, PUNC turns to
3003
+ be a necessary operation to preserve the URLLC performance,
3004
+ although it might cause eMBB service outage. Unlike prior
3005
+ works wherein the URLLC performance is inappropriately
3006
+ assessed by using the outage capacity analysis or the error
3007
+ probability obtained by the normal approximation, in this work
3008
+ the finite-blocklength information-theoretic analysis relies on
3009
+ mismatched receivers and on the saddlepoint approximation
3010
+ which is proper of URLLC scenarios in massive MIMO
3011
+ operation. This work can be extended by including mas-
3012
+ sive machine-type communication (mMTC) in the coexis-
3013
+ tence strategies, and by including the study of the uplink
3014
+ in the proposed generalized framework. Finally, investigating
3015
+ the non-orthogonal multiplexing of heterogeneous services
3016
+ in distributed user-centric systems, such as cell-free massive
3017
+ MIMO [34]–[36], able to provide user’s proximity, macrodi-
3018
+ versity and ubiquitous connectivity, is certainly an appealing
3019
+ future research direction.
3020
+ REFERENCES
3021
+ [1] IMT Vision – Framework and overall objectives of the future develop-
3022
+ ment of IMT for 2020 and beyond, ITU-R Std. M.2083-0, 2015.
3023
+ [2] P. Popovski, K. F. Trillingsgaard, O. Simeone, and G. Durisi, “5G wire-
3024
+ less network slicing for eMBB, URLLC, and mMTC: A communication-
3025
+ theoretic view,” IEEE Access, vol. 6, pp. 55 765–55 779, 2018.
3026
+ [3] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
3027
+ bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
3028
+ no. 11, pp. 3590–3600, 2010.
3029
+ [4] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals
3030
+ of Massive MIMO.
3031
+ Cambridge University Press, 2016.
3032
+ [5] E. Bj¨ornson, J. Hoydis, and L. Sanguinetti, “Massive MIMO networks:
3033
+ Spectral, energy, and hardware efficiency,” Foundations and Trends® in
3034
+ Signal Processing, vol. 11, no. 3-4, pp. 154–655, 2017.
3035
+ [6] P. Popovski, J. J. Nielsen, ˇC. Stefanovi´c, E. d. Carvalho, E. Str¨om, K. F.
3036
+ Trillingsgaard, A. Bana, D. M. Kim, R. Kotaba, J. Park, and R. B.
3037
+ Sørensen, “Wireless access for ultra-reliable low-latency communica-
3038
+ tion: Principles and building blocks,” IEEE Network, vol. 32, no. 2, pp.
3039
+ 16–23, Mar. 2018.
3040
+ [7] A.-S. Bana, E. de Carvalho, B. Soret, T. Abr˜ao, J. C. Marinello,
3041
+ E. G. Larsson, and P. Popovski, “Massive MIMO for internet of
3042
+ things (IoT) connectivity,” Physical Communication, vol. 37, p. 100859,
3043
+ 2019. [Online]. Available: http://www.sciencedirect.com/science/article/
3044
+ pii/S1874490719303891
3045
+ [8] J. ¨Ostman, A. Lancho, G. Durisi, and L. Sanguinetti, “URLLC with
3046
+ massive MIMO: Analysis and design at finite blocklength,” IEEE Trans.
3047
+ Wireless Commun., vol. 20, no. 10, pp. 6387–6401, Oct. 2021.
3048
+ [9] E. Bj¨ornson, E. de Carvalho, J. H. Sørensen, E. G. Larsson, and
3049
+ P. Popovski, “A random access protocol for pilot allocation in crowded
3050
+ massive MIMO systems,” IEEE Transactions on Wireless Communica-
3051
+ tions, vol. 16, no. 4, pp. 2220–2234, Apr. 2017.
3052
+ [10] A. Anand, G. De Veciana, and S. Shakkottai, “Joint scheduling of
3053
+ URLLC and eMBB traffic in 5G wireless networks,” in Proc. of IEEE
3054
+ Conference on Computer Communications (INFOCOM), Apr. 2018, pp.
3055
+ 1970–1978.
3056
+ [11] R. Kassab, O. Simeone, P. Popovski, and T. Islam, “Non-orthogonal
3057
+ multiplexing of ultra-reliable and broadband services in fog-radio archi-
3058
+ tectures,” IEEE Access, vol. 7, pp. 13 035–13 049, 2019.
3059
+ [12] A. A. Esswie and K. I. Pedersen, “Opportunistic spatial preemptive
3060
+ scheduling for URLLC and eMBB coexistence in multi-user 5G net-
3061
+ works,” IEEE Access, vol. 6, pp. 38 451–38 463, 2018.
3062
+ [13] S. F. Abedin, M. G. R. Alam, S. M. A. Kazmi, N. H. Tran, D. Niyato,
3063
+ and C. S. Hong, “Resource allocation for ultra-reliable and enhanced
3064
+ mobile broadband IoT applications in fog network,” IEEE Transactions
3065
+ on Communications, vol. 67, no. 1, pp. 489–502, Jan. 2019.
3066
+ [14] A. Matera, R. Kassab, O. Simeone, and U. Spagnolini, “Non-orthogonal
3067
+ eMBB-URLLC radio access for cloud radio access networks with analog
3068
+ fronthauling,” Entropy, vol. 20, no. 9, 2018.
3069
+ [15] M. Alsenwi, N. H. Tran, M. Bennis, A. Kumar Bairagi, and C. S.
3070
+ Hong, “eMBB-URLLC resource slicing: A risk-sensitive approach,”
3071
+ IEEE Communications Letters, vol. 23, no. 4, pp. 740–743, Apr. 2019.
3072
+ [16] R. Abreu, T. Jacobsen, G. Berardinelli, K. Pedersen, N. H. Mahmood,
3073
+ I. Z. Kovacs, and P. Mogensen, “On the multiplexing of broadband traffic
3074
+ and grant-free ultra-reliable communication in uplink,” in Proc. of IEEE
3075
+ Vehicular Technology Conference (VTC-Spring), Apr. 2019, pp. 1–6.
3076
+ [17] E. N. Tominaga, H. Alves, R. D. Souza, J. Luiz Rebelatto, and
3077
+ M. Latva-aho, “Non-orthogonal multiple access and network slicing:
3078
+ Scalable coexistence of eMBB and URLLC,” in Proc. of IEEE Vehicular
3079
+ Technology Conference (VTC-Spring), Apr. 2021, pp. 1–6.
3080
+ [18] F. Saggese, M. Moretti, and P. Popovski, “Power minimization of
3081
+ downlink spectrum slicing for eMBB and URLLC users,” IEEE Trans.
3082
+ Wireless Commun., vol. 21, no. 12, pp. 11 051–11 065, Dec. 2022.
3083
+ [19] M. Almekhlafi, M. A. Arfaoui, C. Assi, and A. Ghrayeb, “Joint resource
3084
+ and power allocation for URLLC-eMBB traffics multiplexing in 6G
3085
+ wireless networks,” in Proc. IEEE Int. Conf. on Commun. (ICC), Jun.
3086
+ 2021, pp. 1–6.
3087
+ [20] J. Zeng, T. Lv, R. P. Liu, X. Su, Y. J. Guo, and N. C. Beaulieu, “Enabling
3088
+ ultrareliable and low-latency communications under shadow fading by
3089
+ massive MU-MIMO,” IEEE Internet Things J., vol. 7, no. 1, pp. 234–
3090
+ 246, Jan. 2020.
3091
+ [21] H. Ren, C. Pan, Y. Deng, M. Elkashlan, and A. Nallanathan, “Joint pilot
3092
+ and payload power allocation for massive-MIMO-enabled URLLC IIoT
3093
+ networks,” IEEE J. Sel. Areas Commun., vol. 38, no. 5, pp. 816–830,
3094
+ May 2020.
3095
+ [22] A. A. Nasir, H. D. Tuan, H. Q. Ngo, T. Q. Duong, and H. V. Poor,
3096
+ “Cell-free massive MIMO in the short blocklength regime for URLLC,”
3097
+ IEEE Trans. Wireless Commun., vol. 20, no. 9, pp. 5861–5871, Sep.
3098
+ 2021.
3099
+ [23] Y. Polyanskiy, H. V. Poor, and S. Verdu, “Channel coding rate in the
3100
+ finite blocklength regime,” IEEE Trans. Inf. Theory, vol. 56, no. 5, pp.
3101
+ 2307–2359, May 2010.
3102
+ [24] J. Scarlett, A. Martinez, and A. Guill´en i F`abregas, “Mismatched
3103
+ decoding: Error exponents, second-order rates and saddlepoint approxi-
3104
+ mations,” IEEE Trans. Inf. Theory, vol. 60, no. 5, pp. 2647–2666, May
3105
+ 2014.
3106
+ [25] A. Martinez and A. Guill´en i F`abregas, “Saddlepoint approximation of
3107
+ random-coding bounds,” in Proc. of Inf. Theory and Applicat. Workshop
3108
+ (ITA), Feb. 2011, pp. 1–6.
3109
+ [26] W. Yang, G. Durisi, T. Koch, and Y. Polyanskiy, “Quasi-static multiple-
3110
+ antenna fading channels at finite blocklength,” IEEE Trans. Inf. Theory,
3111
+ vol. 60, no. 7, pp. 4232–4265, Jul. 2014.
3112
+ [27] G. Durisi, T. Koch, J. ¨Ostman, Y. Polyanskiy, and W. Yang, “Short-
3113
+ packet communications over multiple-antenna rayleigh-fading channels,”
3114
+ IEEE Trans. Commun., vol. 64, no. 2, pp. 618–629, Feb. 2016.
3115
+ [28] J. ¨Ostman, G. Durisi, E. G. Str¨om, M. C. Coc¸kun, and G. Liva,
3116
+ “Short packets over block-memoryless fading channels: Pilot-assisted
3117
+
3118
+ ON THE COEXISTENCE OF EMBB AND URLLC IN MULTI-CELL MASSIVE MIMO
3119
+ 19
3120
+ or noncoherent transmission?” IEEE Trans. Commun., vol. 67, no. 2,
3121
+ pp. 1521–1536, Feb. 2019.
3122
+ [29] D. Tse and P. Viswanath, Fundamentals of wireless communication.
3123
+ Cambridge University Press, 2005.
3124
+ [30] A. Lapidoth and S. Shamai, “Fading channels: how perfect need ”perfect
3125
+ side information” be?” IEEE Trans. Inf. Theory, vol. 48, no. 5, pp. 1118–
3126
+ 1134, May 2002.
3127
+ [31] R. G. Gallager, Information Theory and Reliable Communication. New
3128
+ York, NY, U.S.A: John Wiley & Sons, 1968.
3129
+ [32] 3GPP, Further advancements for E-UTRA physical layer aspects (Re-
3130
+ lease 9).
3131
+ 3GPP TS 36.814, Mar. 2017.
3132
+ [33] 3rd Generation Partnership Project 3GPP, Service requirements for
3133
+ cyber-physical control applications in vertical domains.
3134
+ 3GPP TS
3135
+ 22.104 v. 17.2.0, Dec. 2019.
3136
+ [34] S. Buzzi and C. D’Andrea, “User-centric communications versus cell-
3137
+ free massive MIMO for 5G cellular networks,” in WSA 2017; 21th
3138
+ International ITG Workshop on Smart Antennas.
3139
+ VDE, 2017, pp. 1–6.
3140
+ [35] G. Interdonato, E. Bj¨ornson, H. Q. Ngo, P. Frenger, and E. G. Larsson,
3141
+ “Ubiquitous cell-free massive MIMO communications,” EURASIP J.
3142
+ Wireless Commun. and Netw., vol. 2019, no. 1, p. 197, 2019.
3143
+ [36] S. Buzzi, C. D’Andrea, A. Zappone, and C. D’Elia, “User-centric 5G
3144
+ cellular networks: Resource allocation and comparison with the cell-
3145
+ free massive MIMO approach,” IEEE Trans. Wireless Commun., vol. 19,
3146
+ no. 2, pp. 1250–1264, Feb. 2020.
3147
+
GNE1T4oBgHgl3EQf_AZD/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
HdAzT4oBgHgl3EQfUvz7/content/tmp_files/2301.01274v1.pdf.txt ADDED
@@ -0,0 +1,801 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Activity Detection for Grant-Free NOMA in Massive
2
+ IoT Networks
3
+ Mehrtash Mehrabi, Student Member, IEEE, Mostafa Mohammadkarimi, Member, IEEE,
4
+ and Masoud Ardakani, Senior Member, IEEE
5
+ Department of Electrical and Computer Engineering, University of Alberta, Edmonton, AB, T6G 1H9, Canada
6
+ Email: {mehrtash, mostafa.mohammadkarimi, ardakani}@ualberta.ca
7
+ Abstract—Recently, grant-free transmission paradigm has been
8
+ introduced for massive Internet of Things (IoT) networks to save
9
+ both time and bandwidth and transmit the message with low
10
+ latency. In order to accurately decode the message of each device at
11
+ the base station (BS), first, the active devices at each transmission
12
+ frame must be identified. In this work, first we investigate the
13
+ problem of activity detection as a threshold comparing problem.
14
+ We show the convexity of the activity detection method through
15
+ analyzing its probability of error which makes it possible to
16
+ find the optimal threshold for minimizing the activity detection
17
+ error. Consequently, to achieve an optimum solution, we propose
18
+ a deep learning (DL)-based method called convolutional neural
19
+ network (CNN)-activity detection (AD). In order to make it more
20
+ practical, we consider unknown and time-varying activity rate
21
+ for the IoT devices. Our simulations verify that our proposed
22
+ CNN-AD method can achieve higher performance compared to the
23
+ existing non-Bayesian greedy-based methods. This is while existing
24
+ methods need to know the activity rate of IoT devices, while our
25
+ method works for unknown and even time-varying activity rates.
26
+ Index Terms—Activity detection, IoT, deep learning, NOMA,
27
+ massive MIMO.
28
+ I. INTRODUCTION
29
+ W
30
+ IRELESS technology recent advances provide massive
31
+ connectivity for machines and objects resulting in the
32
+ Internet of Things (IoT) [1]. The demand for the IoT is
33
+ expected to grow drastically in the near future with numerous
34
+ applications in health care systems, education, businesses and
35
+ governmental services [2]–[4].
36
+ As the demand for connectivity in IoT systems is growing
37
+ rapidly, it is crucial to improve the spectrum efficiency [5].
38
+ Hence, the non-orthogonal multiple access (NOMA) has been
39
+ introduced [6]. To address the main challenges of IoT, including
40
+ access collisions and massive connectivity, NOMA allows
41
+ devices to access the channel non-orthogonally by either power-
42
+ domain [7] or code-domain [8] multiplexing. Meanwhile, this
43
+ massive connectivity is highly affected by the conventional
44
+ grant-based NOMA transmission scheme, where the exchange
45
+ of control signaling between the base station (BS) and IoT
46
+ devices is needed for channel access. The excessive signaling
47
+ overhead causes spectral deficiency and large transmission
48
+ latency. Grant-free NOMA has been introduced to make a
49
+ flexible transmission mechanism for the devices and save time
50
+ and bandwidth by removing the need for the exchange of
51
+ control signaling between the BS and devices. Hence, devices
52
+ can transmit data randomly at any time slot without any request-
53
+ grant procedure.
54
+ In many IoT applications, a few devices become active for a
55
+ short period of time to communicate with the BS while others
56
+ are inactive [9]. In IoT networks with a large number of nodes
57
+ each with a small probability of activity, multiuser detection
58
+ (MUD) methods heavily rely on activity detection (AD) prior to
59
+ detection and decoding [4], [10]–[13]. For uplink transmission
60
+ in IoT systems with grant-free NOMA transmission scheme,
61
+ where the performance of MUD can be severely affected by
62
+ the multi-access interference, the reliable detection of both
63
+ activity and transmitted signal is very challenging and can be
64
+ computationally expensive [10], [12].
65
+ There have been many studies in the literature suggesting
66
+ compressive sensing (CS) methods for joint activity and data
67
+ detection [12]–[16]. Although CS methods can achieve a re-
68
+ liable MUD, they only work in networks with sporadic traffic
69
+ pattern, and are expensive in terms of computational complexity
70
+ [12]. Recently, deep learning (DL) models have observed a lot
71
+ of interests in communication systems and more specifically
72
+ in signal detection [17]–[19]. A study in [19] suggests to use
73
+ DL for activity and data detection, however they consider a
74
+ deterministic traffic pattern for the activity which is not valid
75
+ in all environments.
76
+ In this work, we first formulate the problem of IoT activity
77
+ detection as a threshold comparing problem. We then analyze
78
+ the probability of error of this activity detection method.
79
+ Observing that this probability of error is a convex function
80
+ of the decision threshold, we raise the question of finding the
81
+ optimal threshold for minimizing the activity detection error. To
82
+ achieve this goal, we propose a convolutional neural network
83
+ (CNN)-based AD algorithm for grant-fee code-domain uplink
84
+ NOMA. Unlike existing CS-based AD algorithms, our solution
85
+ does not need to know the exact number of active devices or
86
+ even the activity rate of IoT devices. In fact, in our system
87
+ model we assume a time-varying and unknown activity rate and
88
+ a heterogeneous network. Simulation results verify the success
89
+ the proposed algorithm.
90
+ The rest of this paper is organized as follows. We present
91
+ the system model in Section II. In Section III we formulate the
92
+ device AD problem and derive its probability of error. Section
93
+ IV introduces our CNN-based solution for device AD. The
94
+ simulation results are presented in Section V. Finally, the paper
95
+ is concluded in Section VI.
96
+ arXiv:2301.01274v1 [eess.SP] 23 Dec 2022
97
+
98
+ f
99
+ N
100
+ 1
101
+ 2
102
+ tT
103
+
104
+ 2
105
+ 1
106
+ s
107
+ N
108
+ s
109
+ T
110
+
111
+
112
+
113
+ i-th Transmission Frame
114
+ (i+1)-th Transmission Frame
115
+ Channel
116
+ Estimation
117
+ Channel
118
+ Estimation
119
+ Fig. 1: CDMA slotted ALOHA transmission frame
120
+ A. Notations
121
+ Throughout this paper, (·)∗ represents the complex conjugate.
122
+ Matrix transpose and Hermitian operators are shown by (·)T
123
+ and (·)H, respectively. The operator diag(b) returns a square
124
+ diagonal matrix with the elements of vector b on the main
125
+ diagonal. Furthermore, E[·] is the statistical expectation, ˆa
126
+ denotes an estimated value for a, and size of set S is shown
127
+ by |S|. The constellation and m-dimensional complex spaces
128
+ are denoted by D and Cm, respectively. Finally, the circularly
129
+ symmetric complex Gaussian distribution with mean vector µ
130
+ and covariance matrix Σ is denoted by CN(µ, Σ).
131
+ II. SYSTEM MODEL
132
+ We consider a code-division multiple access (CDMA) uplink
133
+ transmission, where K IoT devices communicate with a single
134
+ IoT BS equipped with M receive antennas. This commonly
135
+ used model [3], [6], [19], also considers a frame structure for
136
+ uplink transmission composed of a channel estimation phase
137
+ followed by CDMA slotted ALOHA data transmissions as
138
+ shown in Fig. 1. In each frame, let Nf short packets of length
139
+ Tt = NsTs, where Ns is the number of symbols per IoT packet
140
+ and Ts is the symbol duration. It is assumed that the channel
141
+ is fixed during each frame, but it varies from one frame to
142
+ another. The channel state information (CSI) is acquired at the
143
+ BS during the channel estimation phase. As it is common in
144
+ massive machine-type communications (mMTC), we assume
145
+ that the IoT devices are only active on occasion and transmit
146
+ short data packets during each frame. The activity rate of the
147
+ IoT devices is denoted by Pa ∈ [0, Pmax], which is assumed
148
+ to be unknown and time-varying from one packet transmission
149
+ to another. Let bk ∈ A be the transmitted symbol of the k-
150
+ th device and chosen from a finite alphabet A, when the k-th
151
+ device is active; otherwise, bk = 0. Consequently, bk can take
152
+ values from an augmented alphabet ¯
153
+ A = A ∪ {0}. We also
154
+ denote the set of all devices and the set of active devices by
155
+ St = {1, 2, . . . , K} and Sa, respectively, where Sa ⊂ St.1
156
+ A unique spreading code is dedicated to each IoT device
157
+ which is simultaneously used for the spreading purpose and de-
158
+ vice identification. This removes the need for control signaling
159
+ associated with IoT device identification. Control signals are
160
+ inefficient for short packet mMTC. The spreading sequence for
161
+ the k-th IoT device is denoted by ck = [c1,k c2,k · · · cNc,k]T
162
+ where ci,k ∈ {−1, +1} and Nc is the spreading factor. To
163
+ 1For the simplicity of notation, we remove the index of frame and packet.
164
+ support a large number of devices, non-orthogonal spreading
165
+ sequences are employed; resulting in NOMA transmission.
166
+ For a single frame, the complex channel coefficient between
167
+ the k-th IoT device and the m-th receive antenna at the BS is
168
+ denoted as gm,k. The active IoT device k, k ∈ Sa transmits Ns
169
+ symbols denoted by bk = [bk,1, · · · , bk,Ns]T during a packet.
170
+ The received baseband signal over Rayleigh flat fading channel
171
+ in a single slot of the slotted ALOHA frame at the m-th receive
172
+ antenna of the BS is expressed as
173
+ Ym =
174
+ K
175
+
176
+ k=1
177
+ gm,kckbT
178
+ k + Wm,
179
+ (1)
180
+ where Wm
181
+
182
+ CNc×Ns
183
+ with wi,j
184
+
185
+ CN(0, σ2
186
+ w) and
187
+ E[wi,jw∗
188
+ u,v]
189
+ =
190
+ σ2
191
+ wδ[i − u]δ[j − v] is the additive white
192
+ Gaussian noise (AWGN) matrix at the m-th receive an-
193
+ tenna. The equivalent channel matrix between all IoT devices
194
+ and the m-th receive antenna can be expressed as Φm =
195
+ [gm,1c1, · · · , gm,KcK] ∈ CNc×K. Thus, the received packet
196
+ at the m-th (m = 1, 2, · · · , M) receive antenna is given by
197
+ Ym = ΦmB + Wm,
198
+ (2)
199
+ where B = [b1, · · · , bK]T ∈ DK×Ns.
200
+ Let the total set of all IoT devices be decomposed into a
201
+ finite number of disjoint groups G1, G2, · · · , GS. Within group
202
+ Gj, the power of every IoT device is given by Pj. The
203
+ powers of the devices are equal within each group, but differ
204
+ from group to group. The fraction of devices in group Gj is
205
+ therefore |Gj|/K. It is assumed that Pj is known at the BS.
206
+ This configuration captures heterogeneous IoT networks, where
207
+ groups of IoT devices capture different phenomenon in a given
208
+ geographical area. A single group of IoT devices with equal
209
+ power transmission, resulting in a homogeneous network, is
210
+ also studied in this paper.
211
+ III. PROBLEM FORMULATION
212
+ In this section, we present the problem of IoT device AD in
213
+ the cases of known CSI at the receiver and in the presence of
214
+ sparse or non-sparse transmission. In order to detect the active
215
+ devices, it is assumed that the BS is equipped with a match filter
216
+ and the precoding matrix and CSI Φm is available. Before AD,
217
+ the observation matrix at the m-th receive antenna ym is passed
218
+ through the decorrelator to obtain
219
+ Ym = ΦH
220
+ mYm ∈ CK×Ns.
221
+ (3)
222
+ In the following, we investigate the details of the AD problem
223
+ based on the Gaussian detection to show how a threshold can be
224
+ computed to distinguish active IoT devices from inactive ones.
225
+ The output of the decorrelator receiver for the m-th receive
226
+ antenna is expressed as
227
+ Ym = ΦH
228
+ mΦmB + ΦH
229
+ mWm,
230
+ =
231
+
232
+ �����
233
+ �K
234
+ k=1 g∗
235
+ m,1gm,kcT
236
+ 1 ckbT
237
+ k + g∗
238
+ m,1cT
239
+ 1 Wm
240
+ �K
241
+ k=1 g∗
242
+ m,2gm,kcT
243
+ 2 ckbT
244
+ k + g∗
245
+ m,2cT
246
+ 2 Wm
247
+ ...
248
+ �K
249
+ k=1 g∗
250
+ m,Kgm,kcT
251
+ KckbT
252
+ k + g∗
253
+ m,KcT
254
+ KWm
255
+
256
+ �����
257
+ .
258
+ (4)
259
+
260
+ Consequently, the received signal from the k-th user at the m-th
261
+ receive antenna is
262
+ rm
263
+ k = ||gm,kck||2
264
+ 2bT
265
+ k +
266
+ K
267
+
268
+ i=1(i̸=k)
269
+ g∗
270
+ m,kgm,icT
271
+ k cibT
272
+ i +g∗
273
+ m,kcT
274
+ k Wm,
275
+ (5)
276
+ where the second and third terms are multi user interference and
277
+ additive noise, respectively. Since an IoT device is either active
278
+ or inactive for the entire packet transmission, we determine the
279
+ activity status of a device based on each received symbol and
280
+ then use the results in [20] for spectrum sensing and combine
281
+ the obtained results from all Ns symbols. The device AD in
282
+ the case of single symbol transmission is studied in [12], and
283
+ we follow that to determine the status of each device based on
284
+ each received symbol and then combine the results. The j-th
285
+ received symbol from device k at receive antenna m, denoted
286
+ as rm
287
+ k,j, is
288
+ rm
289
+ k,j =||gm,kck||2
290
+ 2bk,j+
291
+ K
292
+
293
+ i=1(i̸=k)
294
+ g∗
295
+ m,kgm,icT
296
+ k cibi,j + g∗
297
+ m,kcT
298
+ k wj,
299
+ (6)
300
+ where the first term is the main signal, the second term is multi
301
+ user interference from other devices, and the third term is the
302
+ additive noise. For the sake of simplicity we assume that BPSK
303
+ modulation is used, i.e., the transmitted symbols are drawn from
304
+ A = {−1, +1} and p(bk,j = +1) = p(bk,j = −1) = 1/2. The
305
+ multi user interference plus noise in rm
306
+ k,j has variance
307
+ σ2
308
+ k,j = var
309
+
310
+ K
311
+
312
+ i=1(i̸=k)
313
+ g∗
314
+ m,kgm,icT
315
+ k cibi,j + g∗
316
+ m,kcT
317
+ k wj
318
+
319
+ =
320
+ K
321
+
322
+ i=1(i̸=k)
323
+ |g∗
324
+ m,kgm,icT
325
+ k ci|2Pa + ||g∗
326
+ m,kcT
327
+ k ||2
328
+ 2.
329
+ (7)
330
+ Now we can approximate rm
331
+ k,j by a Gaussian distribution
332
+ as N(||gm,kck||2
333
+ 2, σ2
334
+ k,j) [20]. In order to identify the activity of
335
+ device k, our goal is to propose an algorithm to define threshold
336
+ τ and set device k as active if |rm
337
+ k,j| > τ. Then the probability
338
+ of error, Pe, is computed as
339
+ P k,j
340
+ e
341
+ =Pap(|rm
342
+ k,j| < τ|bk,j ̸= 0)
343
+ + 2(1 − Pa)p(|rm
344
+ k,j| > τ|bk,j = 0),
345
+ (8)
346
+ where we have p(rm
347
+ k,j|bk,j ̸= 0) ∼ N(||gm,kck||2
348
+ 2, σ2
349
+ k,j) and
350
+ p(rm
351
+ k,j|bk,j = 0) ∼ N(0, σ2
352
+ k,j). We can rewrite (8) as
353
+ P k,j
354
+ e
355
+ = 2(1 − Pa)Q( τ
356
+ σk,j
357
+ ) + PaQ(||gm,kck||2
358
+ 2 − τ
359
+ σk,j
360
+ ),
361
+ (9)
362
+ where Q(x) = (1/
363
+
364
+ 2π)
365
+ � ∞
366
+ x exp(−t2/2)dt denotes the Gaus-
367
+ sian tail function. The probability of error in (9) is a convex
368
+ function of τ and hence, a fine tuned neural network is capable
369
+ of solving this problem and detect the active devices by finding
370
+ the optimum τ. In the following section, we define our DL-
371
+ based algorithm to find the optimum τ and minimize the
372
+ probability of error.
373
+ IV. DL-BASED AD
374
+ Device AD is the first step toward effective MUD in a grant-
375
+ free uplink multiple access. The recent studies on AD suggest to
376
+ use CS methods to identify the set of active devices [14], [15].
377
+ However, these methods fail in the practical scenarios, where
378
+ the activity rate is time-varying and/or unknown. Moreover,
379
+ these methods are mainly effective for low device activity rate
380
+ scenarios, i.e., when sparsity level is high [14]. In this section,
381
+ we propose our AD algorithms called CNN-AD by employing a
382
+ CNN for heterogeneous IoT networks. By employing a suitably
383
+ designed CNN, the underlying pattern in device activity can be
384
+ easily learnt.
385
+ A. CNN-AD Algorithm
386
+ Fig. 2 illustrates the structure of the proposed CNN-AD algo-
387
+ rithm. As seen, it is composed of there blocks: 1) preprocessing,
388
+ 2) CNN processing, and 3) hypothesis testing.
389
+ In the preprocessing step after sequence matched filtering, we
390
+ first sort the observation matrix from all M receive antennas
391
+ in a 3D Tensor as
392
+ R =
393
+
394
+ ����
395
+
396
+ P ¯Y1
397
+
398
+
399
+ P ¯Y2
400
+
401
+ ...
402
+
403
+ P ¯YM
404
+
405
+
406
+ ����
407
+ (10)
408
+ where PYm ∈ CK×Ns, Ym = ΦH
409
+ mYm ∈ CK×Ns for
410
+ m
411
+ =
412
+ 1, 2, · · · , M, and P
413
+
414
+ diag(p1, · · · , pK), pk
415
+
416
+ {1/P1, · · · , 1/PS} for k = 1, 2, · · · , K.
417
+ In the CNN processing block, the 3D Tensor R is used
418
+ as the input of a suitable designed CNN. The CNN models
419
+ benefit from the convolutional layers performing convolution
420
+ operations between matrices instead of multiplication. Thus, it
421
+ leads to dimension reduction for feature extraction and provides
422
+ a new input to the next network layers which includes only
423
+ the useful features of the original high-dimensional input. The
424
+ IoT device AD can be formulated as a binary classification or
425
+ regression problem. Formulating device AD as a classification
426
+ problem is straightforward, but it requires the accurate design
427
+ of the CNN’s structure and parameters.
428
+ In the hypothesis testing block, the K outputs of the CNN’s
429
+ Sigmoid layer is compared with a predefined threshold to
430
+ determine the activity status of the IoT devices in the network.
431
+ If the k-th node of the Sigmoid layer exceeds the threshold,
432
+ the k-th IoT device is identified as active.
433
+ B. Training Phase
434
+ In order to train the designed CNN, we define the activity
435
+ vector a as
436
+ a = [a1 a2
437
+ · · ·
438
+ aK]T ,
439
+ (11)
440
+ where ak is 1 when the k-th IoT device is active and 0
441
+ otherwise. We train our model with N independent training
442
+ samples (R
443
+ (j),a(j)), where j = 1, 2, · · · , N and a(j) and
444
+ R
445
+ (j) are the activity vector and observation matrix of the
446
+ j-th training sample, respectively. Our objective is to train
447
+ the designed CNN to generate the desired output vector a(j)
448
+
449
+ Preprocessing
450
+ ]
451
+ ,
452
+ ,
453
+ ,
454
+ [
455
+ 2
456
+ 1
457
+ M
458
+ Y
459
+ Y
460
+ Y
461
+
462
+ Received Message
463
+ at M Antennas
464
+
465
+
466
+
467
+
468
+
469
+
470
+
471
+
472
+
473
+
474
+
475
+
476
+
477
+ ]
478
+ [
479
+ ]
480
+ [
481
+ ]
482
+ [
483
+ 2
484
+ 1
485
+ M
486
+ Y
487
+ P
488
+ Y
489
+ P
490
+ Y
491
+ P
492
+ R
493
+
494
+ CNN
495
+ Input
496
+ M
497
+ K
498
+ s
499
+ N
500
+ CONV 3*3,
501
+ stride=3,
502
+ pad=same
503
+ 128 kernels
504
+ 128
505
+ MAX_POOL,
506
+ 2*2,
507
+ stride=2,
508
+ 2
509
+ M
510
+ 2
511
+ K
512
+ FC
513
+
514
+ 1024
515
+ ReLU
516
+ FC
517
+
518
+ K
519
+ Sigmoid
520
+ Hypothesis Testing
521
+
522
+ ?
523
+ 5.0
524
+
525
+ a
526
+ S
527
+ 128
528
+ M
529
+ K
530
+ Fig. 2: Model structure of the proposed CNN-AD algorithm
531
+ for input matrix R
532
+ (j). The model tries to learns non-linear
533
+ transformation Ψ such that
534
+ ˆa(j) = Ψ(R
535
+ (j); Θ),
536
+ (12)
537
+ where Θ is the set of parameters learned during the training
538
+ by minimizing the loss function. The output of model, i.e.
539
+ ˆa determines the activity probabilities of the IoT devices.
540
+ Here since there are two classes (active or inactive) for each
541
+ IoT device, the loss function is chosen as the binary cross-
542
+ entropy. For each training sample j, the binary cross-entropy
543
+ loss function compares the probability that the IoT devices are
544
+ active (ˆa(j)) with the true activity vector a(j) as
545
+ Loss(Θ) = 1
546
+ N
547
+ N
548
+
549
+ j=1
550
+
551
+
552
+ a(j) log(ˆa(j))+(1−a(j)) log(1−ˆa(j))
553
+
554
+ ,
555
+ (13)
556
+ where log(·) performs an element-wise log operation on ˆa(j),
557
+ and the vector multiplication is also element-wise.
558
+ V. EXPERIMENTS
559
+ In this section, we evaluate the performance of the proposed
560
+ CNN-AD algorithm through various simulation experiments
561
+ and compare it with some of the existing methods.
562
+ A. Simulation Setup
563
+ We consider an IoT network with K devices where K > Nc
564
+ and pseudo-random codes are used as the spreading sequences
565
+ for IoT devices. The probability of activity Pa is considered
566
+ to be unknown and time-varying from one packet to another
567
+ in the range of Pa ∈ [0, Pmax], where Pmax = 0.1. The
568
+ BPSK modulation is used for uplink transmission. Without
569
+ loss of generality, the channel coefficient between IoT devices
570
+ and the BS is modeled as independent zero-mean complex
571
+ Gaussian random variables with variance σ2
572
+ k,m = 1, k ∈ St
573
+ and m ∈ {1, · · · , M}. The additive white noise is modeled as
574
+ zero-mean complex Gaussian random variables with variance
575
+ σ2
576
+ w, and the signal-to-noise ratio (SNR) in dB is defined as
577
+ γ ≜ 10 log(σ2
578
+ s /σ2
579
+ w), where σ2
580
+ s = PaPt is the average transmit
581
+ power with Pt = �K
582
+ k=1 pk as the total transmission power.
583
+ Unless otherwise mentioned, we consider spreading sequences
584
+ with spreading factor Nc = 32.
585
+ In order to train CNN-AD, we generate 105 independent
586
+ samples and use 80% for training and the rest for validation
587
+ and test. Adam optimizer [21] with learning rate of 10−3 is
588
+ used to minimize cross-entropy loss function in (13).
589
+ 0
590
+ 2
591
+ 4
592
+ 6
593
+ 8
594
+ 10
595
+ 12
596
+ 14
597
+ 16
598
+ 18
599
+ 20
600
+ SNR
601
+ 10
602
+ 3
603
+ 10
604
+ 2
605
+ 10
606
+ 1
607
+ AER
608
+ OMP (Uniform Power)
609
+ AMP (Uniform Power)
610
+ CNN_Based (Uniform Power)
611
+ OMP (non-Uniform Power)
612
+ AMP (non-Uniform Power)
613
+ CNN_Based (non-Uniform Power)
614
+ Fig. 3: Achieved BER with MMSE with a priory AD of OMP, AMP,
615
+ and CNN-AD without knowing the number of active devices.
616
+ 0.1
617
+ 0.2
618
+ 0.3
619
+ 0.4
620
+ 0.5
621
+ 0.6
622
+ 0.7
623
+ 0.8
624
+ Activity Rate
625
+ 10
626
+ 2
627
+ 10
628
+ 1
629
+ BER
630
+ OMP (Uniform Power)
631
+ AMP (Uniform Power)
632
+ CNN-AD (Uniform Power)
633
+ Fig. 4: Impact of Pa on the performance of different methods as the
634
+ priory AD for MMSE in terms of achieved BER.
635
+ B. Simulation Results
636
+ 1) Performance Evaluation of CNN-AD: We assess CNN-
637
+ AD through various simulations and compare it with the exist-
638
+ ing CS-based methods including orthogonal matching pursuit
639
+ (OMP) [22] and approximate message passing (AMP) [23].
640
+ The impact of SNR on the activity error rate (AER) achieved
641
+ by different AD algorithms in both homogeneous and hetero-
642
+ geneous IoT networks with uniform and non-uniform power
643
+ allocation is shown in Fig. 3. The AER of different methods
644
+ are compared for a wide range of SNRs in an IoT system with
645
+ total K = 40 IoT devices and a single BS with M = 100
646
+ receive antennas. As expected, the AER of all AD algorithms
647
+ decreases with increasing SNR. However, CNN-AD achieves
648
+
649
+ IoT Device
650
+ Model
651
+ Precision
652
+ Recall
653
+ F1-score
654
+ OMP
655
+ 28%
656
+ 32%
657
+ 30%
658
+ Device A
659
+ AMP
660
+ 31%
661
+ 35%
662
+ 33%
663
+ CNN-AD
664
+ 73%
665
+ 92%
666
+ 81%
667
+ OMP
668
+ 33%
669
+ 32%
670
+ 32%
671
+ Device B
672
+ AMP
673
+ 38%
674
+ 35%
675
+ 36%
676
+ CNN-AD
677
+ 100%
678
+ 83%
679
+ 91%
680
+ TABLE I: Performance analysis different algorithms for two typical
681
+ IoT devices for Pmax = 0.1 at γ = 10 dB.
682
+ the best performance since unlike the non-Bayesian greedy
683
+ algorithms OMP and AMP, our method relies on the statistical
684
+ distributions of device activities and channels and exploit them
685
+ in the training process.
686
+ Fig. 4 illustrates the effect of activity rate on the bit error
687
+ rate (BER) for minimum mean square error (MMSE)-MUD
688
+ with different AD algorithms at γ = 10 dB SNR. As seen,
689
+ as the activity rate increases, the number of active devices
690
+ also increases accordingly and thus it becomes difficult to
691
+ detect all the active devices. This results in a higher BER. We
692
+ use Pmax = 0.1 to train CNN-AD. Thus, the MMSE-MUD
693
+ with CNN-AD shows performance degradation for the activity
694
+ rates of larger than Pmax = 0.1. However, it still outperforms
695
+ the performance of the MMSE-MUD with OMP and AMP
696
+ AD algorithms. It should be mentioned that this performance
697
+ improves when CNN-AD is trained for a larger value of Pmax.
698
+ We further investigate the AD algorithms in terms of other
699
+ metrics for two typical IoT devices for Pmax = 0.1 at γ = 10
700
+ dB SNR, presented in Table I. In this table we compare the
701
+ precision, recall, and F1-score, defined in [24], achieved by
702
+ CNN-AD with OMP and AMP AD algorithms. As seen, all
703
+ metrics are improved by using CNN-AD.
704
+ VI. CONCLUSIONS
705
+ In this paper, we consider the problem of AD in IoT networks
706
+ in grant-free NOMA systems. Based on the application, IoT
707
+ devices can be inactive for a long period of time and only active
708
+ in the time of transmission to the BS. Hence, identifying the
709
+ active devices is required for an accurate data detection. Some
710
+ studies propose CS-based method for AD. However, high level
711
+ of message sparsity is necessary for those methods. In order
712
+ to remove this need and exploit the statistical properties of the
713
+ channels we propose a CNN-based method called CNN-AD to
714
+ detect active IoT devices. Comparison with available methods
715
+ shows the strength of our algorithm.
716
+ ACKNOWLEDGMENT
717
+ The study presented in this paper is supported by Alberta
718
+ Innovates and Natural Sciences and Engineering Research
719
+ Council of Canada (NSERC).
720
+ REFERENCES
721
+ [1] G. Durisi, T. Koch, and P. Popovski, “Toward massive, ultrareliable, and
722
+ low-latency wireless communication with short packets,” Proceedings of
723
+ the IEEE, vol. 104, no. 9, pp. 1711–1726, 2016.
724
+ [2] L. D. Xu, W. He, and S. Li, “Internet of things in industries: A survey,”
725
+ IEEE Transactions on Industrial Informatics, vol. 10, no. 4, pp. 2233–
726
+ 2243, 2014.
727
+ [3] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash,
728
+ “Internet of Things: A survey on enabling technologies, protocols, and
729
+ applications,” IEEE Communications Surveys Tutorials, vol. 17, no. 4,
730
+ pp. 2347–2376, 2015.
731
+ [4] C. Bockelmann, N. Pratas, H. Nikopour, K. Au, T. Svensson, C. Ste-
732
+ fanovic, P. Popovski, and A. Dekorsy, “Massive machine-type communi-
733
+ cations in 5G: Physical and MAC-layer solutions,” IEEE Communications
734
+ Magazine, vol. 54, no. 9, pp. 59–65, 2016.
735
+ [5] W. Ejaz and M. Ibnkahla, “Multiband spectrum sensing and resource
736
+ allocation for IoT in cognitive 5G networks,” IEEE Internet of Things
737
+ Journal, vol. 5, no. 1, pp. 150–163, 2018.
738
+ [6] Z. Ding, P. Fan, and H. V. Poor, “Impact of user pairing on 5G nonorthog-
739
+ onal multiple-access downlink transmissions,” IEEE Transactions on
740
+ Vehicular Technology, vol. 65, no. 8, pp. 6010–6023, 2016.
741
+ [7] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and
742
+ K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future
743
+ radio access,” in 2013 IEEE 77th Vehicular Technology Conference (VTC
744
+ Spring), 2013, pp. 1–5.
745
+ [8] K. Au, L. Zhang, H. Nikopour, E. Yi, A. Bayesteh, U. Vilaipornsawai,
746
+ J. Ma, and P. Zhu, “Uplink contention based SCMA for 5G radio access,”
747
+ in 2014 IEEE Globecom Workshops (GC Wkshps), 2014, pp. 900–905.
748
+ [9] L. Liu, E. G. Larsson, W. Yu, P. Popovski, C. Stefanovic, and E. de
749
+ Carvalho, “Sparse signal processing for grant-free massive connectivity:
750
+ A future paradigm for random access protocols in the Internet of Things,”
751
+ IEEE Signal Processing Magazine, vol. 35, no. 5, pp. 88–99, Sep. 2018.
752
+ [10] S. Verdu et al., Multiuser detection.
753
+ Cambridge university press, 1998.
754
+ [11] Y. Zhang, Q. Guo, Z. Wang, J. Xi, and N. Wu, “Block sparse bayesian
755
+ learning based joint user activity detection and channel estimation for
756
+ grant-free noma systems,” IEEE Transactions on Vehicular Technology,
757
+ vol. 67, no. 10, pp. 9631–9640, 2018.
758
+ [12] H. Zhu and G. B. Giannakis, “Exploiting sparse user activity in multiuser
759
+ detection,” IEEE Transactions on Communications, vol. 59, no. 2, pp.
760
+ 454–465, Feb. 2011.
761
+ [13] H. F. Schepker, C. Bockelmann, and A. Dekorsy, “Coping with CDMA
762
+ asynchronicity in compressive sensing multi-user detection,” in 2013
763
+ IEEE 77th Vehicular Technology Conference (VTC Spring), Jun. 2013,
764
+ pp. 1–5.
765
+ [14] Z. Chen, F. Sohrabi, and W. Yu, “Sparse activity detection for massive
766
+ connectivity,” IEEE Transactions on Signal Processing, vol. 66, no. 7,
767
+ pp. 1890–1904, Apr. 2018.
768
+ [15] K. Takeuchi, T. Tanaka, and T. Kawabata, “Performance improvement
769
+ of iterative multiuser detection for large sparsely spread CDMA systems
770
+ by spatial coupling,” IEEE Transactions on Information Theory, vol. 61,
771
+ no. 4, pp. 1768–1794, Apr. 2015.
772
+ [16] Y. Wang, X. Zhu, E. G. Lim, Z. Wei, Y. Liu, and Y. Jiang, “Compressive
773
+ sensing based user activity detection and channel estimation in uplink
774
+ noma systems,” in 2020 IEEE Wireless Communications and Networking
775
+ Conference (WCNC).
776
+ IEEE, 2020, pp. 1–6.
777
+ [17] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning.
778
+ MIT press
779
+ Cambridge, 2016, vol. 1.
780
+ [18] M. Mohammadkarimi, M. Mehrabi, M. Ardakani, and Y. Jing, “Deep
781
+ learning based sphere decoding,” IEEE Trans. Wireless Commun., pp.
782
+ 1–1, 2019.
783
+ [19] X. Miao, D. Guo, and X. Li, “Grant-free NOMA with device activity
784
+ learning using long short-term memory,” IEEE Wireless Communications
785
+ Letters, pp. 1–1, 2020.
786
+ [20] W. Zhang, R. K. Mallik, and K. B. Letaief, “Cooperative spectrum sensing
787
+ optimization in cognitive radio networks,” in 2008 IEEE International
788
+ Conference on Communications, 2008, pp. 3411–3415.
789
+ [21] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
790
+ arXiv preprint arXiv:1412.6980, 2014.
791
+ [22] T. T. Cai and L. Wang, “Orthogonal matching pursuit for sparse signal
792
+ recovery with noise,” IEEE Transactions on Information theory, vol. 57,
793
+ no. 7, pp. 4680–4688, 2011.
794
+ [23] D. L. Donoho, A. Maleki, and A. Montanari, “Message-passing algo-
795
+ rithms for compressed sensing,” Proceedings of the National Academy of
796
+ Sciences, vol. 106, no. 45, pp. 18 914–18 919, 2009.
797
+ [24] C. Goutte and E. Gaussier, “A probabilistic interpretation of precision, re-
798
+ call and f-score, with implication for evaluation,” in European conference
799
+ on information retrieval.
800
+ Springer, 2005, pp. 345–359.
801
+
HdAzT4oBgHgl3EQfUvz7/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,387 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf,len=386
2
+ page_content='Activity Detection for Grant-Free NOMA in Massive IoT Networks Mehrtash Mehrabi, Student Member, IEEE, Mostafa Mohammadkarimi, Member, IEEE, and Masoud Ardakani, Senior Member, IEEE Department of Electrical and Computer Engineering, University of Alberta, Edmonton, AB, T6G 1H9, Canada Email: {mehrtash, mostafa.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
3
+ page_content='mohammadkarimi, ardakani}@ualberta.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
4
+ page_content='ca Abstract—Recently, grant-free transmission paradigm has been introduced for massive Internet of Things (IoT) networks to save both time and bandwidth and transmit the message with low latency.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
5
+ page_content=' In order to accurately decode the message of each device at the base station (BS), first, the active devices at each transmission frame must be identified.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
6
+ page_content=' In this work, first we investigate the problem of activity detection as a threshold comparing problem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
7
+ page_content=' We show the convexity of the activity detection method through analyzing its probability of error which makes it possible to find the optimal threshold for minimizing the activity detection error.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
8
+ page_content=' Consequently, to achieve an optimum solution, we propose a deep learning (DL)-based method called convolutional neural network (CNN)-activity detection (AD).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
9
+ page_content=' In order to make it more practical, we consider unknown and time-varying activity rate for the IoT devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
10
+ page_content=' Our simulations verify that our proposed CNN-AD method can achieve higher performance compared to the existing non-Bayesian greedy-based methods.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
11
+ page_content=' This is while existing methods need to know the activity rate of IoT devices, while our method works for unknown and even time-varying activity rates.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
12
+ page_content=' Index Terms—Activity detection, IoT, deep learning, NOMA, massive MIMO.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
13
+ page_content=' I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
14
+ page_content=' INTRODUCTION W IRELESS technology recent advances provide massive connectivity for machines and objects resulting in the Internet of Things (IoT) [1].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
15
+ page_content=' The demand for the IoT is expected to grow drastically in the near future with numerous applications in health care systems, education, businesses and governmental services [2]–[4].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
16
+ page_content=' As the demand for connectivity in IoT systems is growing rapidly, it is crucial to improve the spectrum efficiency [5].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
17
+ page_content=' Hence, the non-orthogonal multiple access (NOMA) has been introduced [6].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
18
+ page_content=' To address the main challenges of IoT, including access collisions and massive connectivity, NOMA allows devices to access the channel non-orthogonally by either power- domain [7] or code-domain [8] multiplexing.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
19
+ page_content=' Meanwhile, this massive connectivity is highly affected by the conventional grant-based NOMA transmission scheme, where the exchange of control signaling between the base station (BS) and IoT devices is needed for channel access.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
20
+ page_content=' The excessive signaling overhead causes spectral deficiency and large transmission latency.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
21
+ page_content=' Grant-free NOMA has been introduced to make a flexible transmission mechanism for the devices and save time and bandwidth by removing the need for the exchange of control signaling between the BS and devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
22
+ page_content=' Hence, devices can transmit data randomly at any time slot without any request- grant procedure.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
23
+ page_content=' In many IoT applications, a few devices become active for a short period of time to communicate with the BS while others are inactive [9].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
24
+ page_content=' In IoT networks with a large number of nodes each with a small probability of activity, multiuser detection (MUD) methods heavily rely on activity detection (AD) prior to detection and decoding [4], [10]–[13].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
25
+ page_content=' For uplink transmission in IoT systems with grant-free NOMA transmission scheme, where the performance of MUD can be severely affected by the multi-access interference, the reliable detection of both activity and transmitted signal is very challenging and can be computationally expensive [10], [12].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
26
+ page_content=' There have been many studies in the literature suggesting compressive sensing (CS) methods for joint activity and data detection [12]–[16].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
27
+ page_content=' Although CS methods can achieve a re- liable MUD, they only work in networks with sporadic traffic pattern, and are expensive in terms of computational complexity [12].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
28
+ page_content=' Recently, deep learning (DL) models have observed a lot of interests in communication systems and more specifically in signal detection [17]–[19].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
29
+ page_content=' A study in [19] suggests to use DL for activity and data detection, however they consider a deterministic traffic pattern for the activity which is not valid in all environments.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
30
+ page_content=' In this work, we first formulate the problem of IoT activity detection as a threshold comparing problem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
31
+ page_content=' We then analyze the probability of error of this activity detection method.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
32
+ page_content=' Observing that this probability of error is a convex function of the decision threshold, we raise the question of finding the optimal threshold for minimizing the activity detection error.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
33
+ page_content=' To achieve this goal, we propose a convolutional neural network (CNN)-based AD algorithm for grant-fee code-domain uplink NOMA.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
34
+ page_content=' Unlike existing CS-based AD algorithms, our solution does not need to know the exact number of active devices or even the activity rate of IoT devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
35
+ page_content=' In fact, in our system model we assume a time-varying and unknown activity rate and a heterogeneous network.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
36
+ page_content=' Simulation results verify the success the proposed algorithm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
37
+ page_content=' The rest of this paper is organized as follows.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
38
+ page_content=' We present the system model in Section II.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
39
+ page_content=' In Section III we formulate the device AD problem and derive its probability of error.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
40
+ page_content=' Section IV introduces our CNN-based solution for device AD.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
41
+ page_content=' The simulation results are presented in Section V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
42
+ page_content=' Finally, the paper is concluded in Section VI.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
43
+ page_content=' arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
44
+ page_content='01274v1 [eess.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
45
+ page_content='SP] 23 Dec 2022 f N 1 2 tT \uf04c 2 1 s N s T \uf04c \uf04c \uf04c i-th Transmission Frame (i+1)-th Transmission Frame Channel Estimation Channel Estimation Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
46
+ page_content=' 1: CDMA slotted ALOHA transmission frame A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
47
+ page_content=' Notations Throughout this paper, (·)∗ represents the complex conjugate.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
48
+ page_content=' Matrix transpose and Hermitian operators are shown by (·)T and (·)H, respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
49
+ page_content=' The operator diag(b) returns a square diagonal matrix with the elements of vector b on the main diagonal.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
50
+ page_content=' Furthermore, E[·] is the statistical expectation, ˆa denotes an estimated value for a, and size of set S is shown by |S|.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
51
+ page_content=' The constellation and m-dimensional complex spaces are denoted by D and Cm, respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
52
+ page_content=' Finally, the circularly symmetric complex Gaussian distribution with mean vector µ and covariance matrix Σ is denoted by CN(µ, Σ).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
53
+ page_content=' II.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
54
+ page_content=' SYSTEM MODEL We consider a code-division multiple access (CDMA) uplink transmission, where K IoT devices communicate with a single IoT BS equipped with M receive antennas.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
55
+ page_content=' This commonly used model [3], [6], [19], also considers a frame structure for uplink transmission composed of a channel estimation phase followed by CDMA slotted ALOHA data transmissions as shown in Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
56
+ page_content=' 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
57
+ page_content=' In each frame, let Nf short packets of length Tt = NsTs, where Ns is the number of symbols per IoT packet and Ts is the symbol duration.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
58
+ page_content=' It is assumed that the channel is fixed during each frame, but it varies from one frame to another.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
59
+ page_content=' The channel state information (CSI) is acquired at the BS during the channel estimation phase.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
60
+ page_content=' As it is common in massive machine-type communications (mMTC), we assume that the IoT devices are only active on occasion and transmit short data packets during each frame.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
61
+ page_content=' The activity rate of the IoT devices is denoted by Pa ∈ [0, Pmax], which is assumed to be unknown and time-varying from one packet transmission to another.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
62
+ page_content=' Let bk ∈ A be the transmitted symbol of the k- th device and chosen from a finite alphabet A, when the k-th device is active;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
63
+ page_content=' otherwise, bk = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
64
+ page_content=' Consequently, bk can take values from an augmented alphabet ¯ A = A ∪ {0}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
65
+ page_content=' We also denote the set of all devices and the set of active devices by St = {1, 2, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
66
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
67
+ page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
68
+ page_content=' , K} and Sa, respectively, where Sa ⊂ St.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
69
+ page_content='1 A unique spreading code is dedicated to each IoT device which is simultaneously used for the spreading purpose and de- vice identification.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
70
+ page_content=' This removes the need for control signaling associated with IoT device identification.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
71
+ page_content=' Control signals are inefficient for short packet mMTC.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
72
+ page_content=' The spreading sequence for the k-th IoT device is denoted by ck = [c1,k c2,k · · · cNc,k]T where ci,k ∈ {−1, +1} and Nc is the spreading factor.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
73
+ page_content=' To 1For the simplicity of notation, we remove the index of frame and packet.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
74
+ page_content=' support a large number of devices, non-orthogonal spreading sequences are employed;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
75
+ page_content=' resulting in NOMA transmission.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
76
+ page_content=' For a single frame, the complex channel coefficient between the k-th IoT device and the m-th receive antenna at the BS is denoted as gm,k.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
77
+ page_content=' The active IoT device k, k ∈ Sa transmits Ns symbols denoted by bk = [bk,1, · · · , bk,Ns]T during a packet.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
78
+ page_content=' The received baseband signal over Rayleigh flat fading channel in a single slot of the slotted ALOHA frame at the m-th receive antenna of the BS is expressed as Ym = K � k=1 gm,kckbT k + Wm, (1) where Wm ∈ CNc×Ns with wi,j ∼ CN(0, σ2 w) and E[wi,jw∗ u,v] = σ2 wδ[i − u]δ[j − v] is the additive white Gaussian noise (AWGN) matrix at the m-th receive an- tenna.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
79
+ page_content=' The equivalent channel matrix between all IoT devices and the m-th receive antenna can be expressed as Φm = [gm,1c1, · · · , gm,KcK] ∈ CNc×K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
80
+ page_content=' Thus, the received packet at the m-th (m = 1, 2, · · · , M) receive antenna is given by Ym = ΦmB + Wm, (2) where B = [b1, · · · , bK]T ∈ DK×Ns.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
81
+ page_content=' Let the total set of all IoT devices be decomposed into a finite number of disjoint groups G1, G2, · · · , GS.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
82
+ page_content=' Within group Gj, the power of every IoT device is given by Pj.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
83
+ page_content=' The powers of the devices are equal within each group, but differ from group to group.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
84
+ page_content=' The fraction of devices in group Gj is therefore |Gj|/K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
85
+ page_content=' It is assumed that Pj is known at the BS.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
86
+ page_content=' This configuration captures heterogeneous IoT networks, where groups of IoT devices capture different phenomenon in a given geographical area.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
87
+ page_content=' A single group of IoT devices with equal power transmission, resulting in a homogeneous network, is also studied in this paper.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
88
+ page_content=' III.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
89
+ page_content=' PROBLEM FORMULATION In this section, we present the problem of IoT device AD in the cases of known CSI at the receiver and in the presence of sparse or non-sparse transmission.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
90
+ page_content=' In order to detect the active devices, it is assumed that the BS is equipped with a match filter and the precoding matrix and CSI Φm is available.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
91
+ page_content=' Before AD, the observation matrix at the m-th receive antenna ym is passed through the decorrelator to obtain Ym = ΦH mYm ∈ CK×Ns.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
92
+ page_content=' (3) In the following, we investigate the details of the AD problem based on the Gaussian detection to show how a threshold can be computed to distinguish active IoT devices from inactive ones.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
93
+ page_content=' The output of the decorrelator receiver for the m-th receive antenna is expressed as Ym = ΦH mΦmB + ΦH mWm, = � ����� �K k=1 g∗ m,1gm,kcT 1 ckbT k + g∗ m,1cT 1 Wm �K k=1 g∗ m,2gm,kcT 2 ckbT k + g∗ m,2cT 2 Wm .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
94
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
95
+ page_content=' �K k=1 g∗ m,Kgm,kcT KckbT k + g∗ m,KcT KWm � ����� .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
96
+ page_content=' (4) Consequently, the received signal from the k-th user at the m-th receive antenna is rm k = ||gm,kck||2 2bT k + K � i=1(i̸=k) g∗ m,kgm,icT k cibT i +g∗ m,kcT k Wm, (5) where the second and third terms are multi user interference and additive noise, respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
97
+ page_content=' Since an IoT device is either active or inactive for the entire packet transmission, we determine the activity status of a device based on each received symbol and then use the results in [20] for spectrum sensing and combine the obtained results from all Ns symbols.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
98
+ page_content=' The device AD in the case of single symbol transmission is studied in [12], and we follow that to determine the status of each device based on each received symbol and then combine the results.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
99
+ page_content=' The j-th received symbol from device k at receive antenna m, denoted as rm k,j, is rm k,j =||gm,kck||2 2bk,j+ K � i=1(i̸=k) g∗ m,kgm,icT k cibi,j + g∗ m,kcT k wj, (6) where the first term is the main signal, the second term is multi user interference from other devices, and the third term is the additive noise.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
100
+ page_content=' For the sake of simplicity we assume that BPSK modulation is used, i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
101
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
102
+ page_content=', the transmitted symbols are drawn from A = {−1, +1} and p(bk,j = +1) = p(bk,j = −1) = 1/2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
103
+ page_content=' The multi user interference plus noise in rm k,j has variance σ2 k,j = var � K � i=1(i̸=k) g∗ m,kgm,icT k cibi,j + g∗ m,kcT k wj � = K � i=1(i̸=k) |g∗ m,kgm,icT k ci|2Pa + ||g∗ m,kcT k ||2 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
104
+ page_content=' (7) Now we can approximate rm k,j by a Gaussian distribution as N(||gm,kck||2 2, σ2 k,j) [20].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
105
+ page_content=' In order to identify the activity of device k, our goal is to propose an algorithm to define threshold τ and set device k as active if |rm k,j| > τ.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
106
+ page_content=' Then the probability of error, Pe, is computed as P k,j e =Pap(|rm k,j| < τ|bk,j ̸= 0) + 2(1 − Pa)p(|rm k,j| > τ|bk,j = 0), (8) where we have p(rm k,j|bk,j ̸= 0) ∼ N(||gm,kck||2 2, σ2 k,j) and p(rm k,j|bk,j = 0) ∼ N(0, σ2 k,j).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
107
+ page_content=' We can rewrite (8) as P k,j e = 2(1 − Pa)Q( τ σk,j ) + PaQ(||gm,kck||2 2 − τ σk,j ), (9) where Q(x) = (1/ √ 2π) � ∞ x exp(−t2/2)dt denotes the Gaus- sian tail function.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
108
+ page_content=' The probability of error in (9) is a convex function of τ and hence, a fine tuned neural network is capable of solving this problem and detect the active devices by finding the optimum τ.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
109
+ page_content=' In the following section, we define our DL- based algorithm to find the optimum τ and minimize the probability of error.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
110
+ page_content=' IV.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
111
+ page_content=' DL-BASED AD Device AD is the first step toward effective MUD in a grant- free uplink multiple access.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
112
+ page_content=' The recent studies on AD suggest to use CS methods to identify the set of active devices [14], [15].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
113
+ page_content=' However, these methods fail in the practical scenarios, where the activity rate is time-varying and/or unknown.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
114
+ page_content=' Moreover, these methods are mainly effective for low device activity rate scenarios, i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
115
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
116
+ page_content=', when sparsity level is high [14].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
117
+ page_content=' In this section, we propose our AD algorithms called CNN-AD by employing a CNN for heterogeneous IoT networks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
118
+ page_content=' By employing a suitably designed CNN, the underlying pattern in device activity can be easily learnt.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
119
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
120
+ page_content=' CNN-AD Algorithm Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
121
+ page_content=' 2 illustrates the structure of the proposed CNN-AD algo- rithm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
122
+ page_content=' As seen, it is composed of there blocks: 1) preprocessing, 2) CNN processing, and 3) hypothesis testing.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
123
+ page_content=' In the preprocessing step after sequence matched filtering, we first sort the observation matrix from all M receive antennas in a 3D Tensor as R = � ���� � P ¯Y1 � � P ¯Y2 � .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
124
+ page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
125
+ page_content=' � P ¯YM � � ���� (10) where PYm ∈ CK×Ns, Ym = ΦH mYm ∈ CK×Ns for m = 1, 2, · · · , M, and P ≜ diag(p1, · · · , pK), pk ∈ {1/P1, · · · , 1/PS} for k = 1, 2, · · · , K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
126
+ page_content=' In the CNN processing block, the 3D Tensor R is used as the input of a suitable designed CNN.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
127
+ page_content=' The CNN models benefit from the convolutional layers performing convolution operations between matrices instead of multiplication.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
128
+ page_content=' Thus, it leads to dimension reduction for feature extraction and provides a new input to the next network layers which includes only the useful features of the original high-dimensional input.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
129
+ page_content=' The IoT device AD can be formulated as a binary classification or regression problem.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
130
+ page_content=' Formulating device AD as a classification problem is straightforward, but it requires the accurate design of the CNN’s structure and parameters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
131
+ page_content=' In the hypothesis testing block, the K outputs of the CNN’s Sigmoid layer is compared with a predefined threshold to determine the activity status of the IoT devices in the network.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
132
+ page_content=' If the k-th node of the Sigmoid layer exceeds the threshold, the k-th IoT device is identified as active.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
133
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
134
+ page_content=' Training Phase In order to train the designed CNN, we define the activity vector a as a = [a1 a2 · · aK]T , (11) where ak is 1 when the k-th IoT device is active and 0 otherwise.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
135
+ page_content=' We train our model with N independent training samples (R (j),a(j)), where j = 1, 2, · · · , N and a(j) and R (j) are the activity vector and observation matrix of the j-th training sample, respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
136
+ page_content=' Our objective is to train the designed CNN to generate the desired output vector a(j) Preprocessing ] , , , [ 2 1 M Y Y Y \uf04c Received Message at M Antennas \uf0fa \uf0fa \uf0fa \uf0fa \uf0fb \uf0f9 \uf0ea \uf0ea \uf0ea \uf0ea \uf0eb \uf0e9 \uf03d ] [ ] [ ] [ 2 1 M Y P Y P Y P R \uf04d CNN Input M K s N CONV 3*3, stride=3, pad=same 128 kernels 128 MAX_POOL, 2*2, stride=2, 2 M 2 K FC \uf04d 1024 ReLU FC \uf04d K Sigmoid Hypothesis Testing \uf04d ?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
137
+ page_content=' 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
138
+ page_content='0 \uf0b3 a S 128 M K Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
139
+ page_content=' 2: Model structure of the proposed CNN-AD algorithm for input matrix R (j).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
140
+ page_content=' The model tries to learns non-linear transformation Ψ such that ˆa(j) = Ψ(R (j);' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
141
+ page_content=' Θ), (12) where Θ is the set of parameters learned during the training by minimizing the loss function.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
142
+ page_content=' The output of model, i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
143
+ page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
144
+ page_content=' ˆa determines the activity probabilities of the IoT devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
145
+ page_content=' Here since there are two classes (active or inactive) for each IoT device, the loss function is chosen as the binary cross- entropy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
146
+ page_content=' For each training sample j, the binary cross-entropy loss function compares the probability that the IoT devices are active (ˆa(j)) with the true activity vector a(j) as Loss(Θ) = 1 N N � j=1 − � a(j) log(ˆa(j))+(1−a(j)) log(1−ˆa(j)) � , (13) where log(·) performs an element-wise log operation on ˆa(j), and the vector multiplication is also element-wise.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
147
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
148
+ page_content=' EXPERIMENTS In this section, we evaluate the performance of the proposed CNN-AD algorithm through various simulation experiments and compare it with some of the existing methods.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
149
+ page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
150
+ page_content=' Simulation Setup We consider an IoT network with K devices where K > Nc and pseudo-random codes are used as the spreading sequences for IoT devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
151
+ page_content=' The probability of activity Pa is considered to be unknown and time-varying from one packet to another in the range of Pa ∈ [0, Pmax], where Pmax = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
152
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
153
+ page_content=' The BPSK modulation is used for uplink transmission.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
154
+ page_content=' Without loss of generality, the channel coefficient between IoT devices and the BS is modeled as independent zero-mean complex Gaussian random variables with variance σ2 k,m = 1, k ∈ St and m ∈ {1, · · · , M}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
155
+ page_content=' The additive white noise is modeled as zero-mean complex Gaussian random variables with variance σ2 w, and the signal-to-noise ratio (SNR) in dB is defined as γ ≜ 10 log(σ2 s /σ2 w), where σ2 s = PaPt is the average transmit power with Pt = �K k=1 pk as the total transmission power.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
156
+ page_content=' Unless otherwise mentioned, we consider spreading sequences with spreading factor Nc = 32.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
157
+ page_content=' In order to train CNN-AD, we generate 105 independent samples and use 80% for training and the rest for validation and test.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
158
+ page_content=' Adam optimizer [21] with learning rate of 10−3 is used to minimize cross-entropy loss function in (13).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
159
+ page_content=' 0 2 4 6 8 10 12 14 16 18 20 SNR 10 3 10 2 10 1 AER OMP (Uniform Power) AMP (Uniform Power) CNN_Based (Uniform Power) OMP (non-Uniform Power) AMP (non-Uniform Power) CNN_Based (non-Uniform Power) Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
160
+ page_content=' 3: Achieved BER with MMSE with a priory AD of OMP, AMP, and CNN-AD without knowing the number of active devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
161
+ page_content=' 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
162
+ page_content='1 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
163
+ page_content='2 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
164
+ page_content='3 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
165
+ page_content='4 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
166
+ page_content='5 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
167
+ page_content='6 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
168
+ page_content='7 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
169
+ page_content='8 Activity Rate 10 2 10 1 BER OMP (Uniform Power) AMP (Uniform Power) CNN-AD (Uniform Power) Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
170
+ page_content=' 4: Impact of Pa on the performance of different methods as the priory AD for MMSE in terms of achieved BER.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
171
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
172
+ page_content=' Simulation Results 1) Performance Evaluation of CNN-AD: We assess CNN- AD through various simulations and compare it with the exist- ing CS-based methods including orthogonal matching pursuit (OMP) [22] and approximate message passing (AMP) [23].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
173
+ page_content=' The impact of SNR on the activity error rate (AER) achieved by different AD algorithms in both homogeneous and hetero- geneous IoT networks with uniform and non-uniform power allocation is shown in Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
174
+ page_content=' 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
175
+ page_content=' The AER of different methods are compared for a wide range of SNRs in an IoT system with total K = 40 IoT devices and a single BS with M = 100 receive antennas.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
176
+ page_content=' As expected, the AER of all AD algorithms decreases with increasing SNR.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
177
+ page_content=' However, CNN-AD achieves IoT Device Model Precision Recall F1-score OMP 28% 32% 30% Device A AMP 31% 35% 33% CNN-AD 73% 92% 81% OMP 33% 32% 32% Device B AMP 38% 35% 36% CNN-AD 100% 83% 91% TABLE I: Performance analysis different algorithms for two typical IoT devices for Pmax = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
178
+ page_content='1 at γ = 10 dB.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
179
+ page_content=' the best performance since unlike the non-Bayesian greedy algorithms OMP and AMP, our method relies on the statistical distributions of device activities and channels and exploit them in the training process.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
180
+ page_content=' Fig.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
181
+ page_content=' 4 illustrates the effect of activity rate on the bit error rate (BER) for minimum mean square error (MMSE)-MUD with different AD algorithms at γ = 10 dB SNR.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
182
+ page_content=' As seen, as the activity rate increases, the number of active devices also increases accordingly and thus it becomes difficult to detect all the active devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
183
+ page_content=' This results in a higher BER.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
184
+ page_content=' We use Pmax = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
185
+ page_content='1 to train CNN-AD.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
186
+ page_content=' Thus, the MMSE-MUD with CNN-AD shows performance degradation for the activity rates of larger than Pmax = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
187
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
188
+ page_content=' However, it still outperforms the performance of the MMSE-MUD with OMP and AMP AD algorithms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
189
+ page_content=' It should be mentioned that this performance improves when CNN-AD is trained for a larger value of Pmax.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
190
+ page_content=' We further investigate the AD algorithms in terms of other metrics for two typical IoT devices for Pmax = 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
191
+ page_content='1 at γ = 10 dB SNR, presented in Table I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
192
+ page_content=' In this table we compare the precision, recall, and F1-score, defined in [24], achieved by CNN-AD with OMP and AMP AD algorithms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
193
+ page_content=' As seen, all metrics are improved by using CNN-AD.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
194
+ page_content=' VI.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
195
+ page_content=' CONCLUSIONS In this paper, we consider the problem of AD in IoT networks in grant-free NOMA systems.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
196
+ page_content=' Based on the application, IoT devices can be inactive for a long period of time and only active in the time of transmission to the BS.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
197
+ page_content=' Hence, identifying the active devices is required for an accurate data detection.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
198
+ page_content=' Some studies propose CS-based method for AD.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
199
+ page_content=' However, high level of message sparsity is necessary for those methods.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
200
+ page_content=' In order to remove this need and exploit the statistical properties of the channels we propose a CNN-based method called CNN-AD to detect active IoT devices.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
201
+ page_content=' Comparison with available methods shows the strength of our algorithm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
202
+ page_content=' ACKNOWLEDGMENT The study presented in this paper is supported by Alberta Innovates and Natural Sciences and Engineering Research Council of Canada (NSERC).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
203
+ page_content=' REFERENCES [1] G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
204
+ page_content=' Durisi, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
205
+ page_content=' Koch, and P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
206
+ page_content=' Popovski, “Toward massive, ultrareliable, and low-latency wireless communication with short packets,” Proceedings of the IEEE, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
207
+ page_content=' 104, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
208
+ page_content=' 9, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
209
+ page_content=' 1711–1726, 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
210
+ page_content=' [2] L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
211
+ page_content=' D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
212
+ page_content=' Xu, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
213
+ page_content=' He, and S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
214
+ page_content=' Li, “Internet of things in industries: A survey,” IEEE Transactions on Industrial Informatics, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
215
+ page_content=' 10, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
216
+ page_content=' 4, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
217
+ page_content=' 2233– 2243, 2014.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
218
+ page_content=' [3] A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
219
+ page_content=' Al-Fuqaha, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
220
+ page_content=' Guizani, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
221
+ page_content=' Mohammadi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
222
+ page_content=' Aledhari, and M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
223
+ page_content=' Ayyash, “Internet of Things: A survey on enabling technologies, protocols, and applications,” IEEE Communications Surveys Tutorials, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
224
+ page_content=' 17, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
225
+ page_content=' 4, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
226
+ page_content=' 2347–2376, 2015.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
227
+ page_content=' [4] C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
228
+ page_content=' Bockelmann, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
229
+ page_content=' Pratas, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
230
+ page_content=' Nikopour, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
231
+ page_content=' Au, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
232
+ page_content=' Svensson, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
233
+ page_content=' Ste- fanovic, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
234
+ page_content=' Popovski, and A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
235
+ page_content=' Dekorsy, “Massive machine-type communi- cations in 5G: Physical and MAC-layer solutions,” IEEE Communications Magazine, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
236
+ page_content=' 54, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
237
+ page_content=' 9, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
238
+ page_content=' 59–65, 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
239
+ page_content=' [5] W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
240
+ page_content=' Ejaz and M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
241
+ page_content=' Ibnkahla, “Multiband spectrum sensing and resource allocation for IoT in cognitive 5G networks,” IEEE Internet of Things Journal, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
242
+ page_content=' 5, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
243
+ page_content=' 1, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
244
+ page_content=' 150–163, 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
245
+ page_content=' [6] Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
246
+ page_content=' Ding, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
247
+ page_content=' Fan, and H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
248
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
249
+ page_content=' Poor, “Impact of user pairing on 5G nonorthog- onal multiple-access downlink transmissions,” IEEE Transactions on Vehicular Technology, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
250
+ page_content=' 65, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
251
+ page_content=' 8, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
252
+ page_content=' 6010–6023, 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
253
+ page_content=' [7] Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
254
+ page_content=' Saito, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
255
+ page_content=' Kishiyama, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
256
+ page_content=' Benjebbour, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
257
+ page_content=' Nakamura, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
258
+ page_content=' Li, and K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
259
+ page_content=' Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future radio access,” in 2013 IEEE 77th Vehicular Technology Conference (VTC Spring), 2013, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
260
+ page_content=' 1–5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
261
+ page_content=' [8] K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
262
+ page_content=' Au, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
263
+ page_content=' Zhang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
264
+ page_content=' Nikopour, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
265
+ page_content=' Yi, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
266
+ page_content=' Bayesteh, U.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
267
+ page_content=' Vilaipornsawai, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
268
+ page_content=' Ma, and P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
269
+ page_content=' Zhu, “Uplink contention based SCMA for 5G radio access,” in 2014 IEEE Globecom Workshops (GC Wkshps), 2014, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
270
+ page_content=' 900–905.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
271
+ page_content=' [9] L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
272
+ page_content=' Liu, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
273
+ page_content=' G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
274
+ page_content=' Larsson, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
275
+ page_content=' Yu, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
276
+ page_content=' Popovski, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
277
+ page_content=' Stefanovic, and E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
278
+ page_content=' de Carvalho, “Sparse signal processing for grant-free massive connectivity: A future paradigm for random access protocols in the Internet of Things,” IEEE Signal Processing Magazine, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
279
+ page_content=' 35, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
280
+ page_content=' 5, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
281
+ page_content=' 88–99, Sep.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
282
+ page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
283
+ page_content=' [10] S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
284
+ page_content=' Verdu et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
285
+ page_content=', Multiuser detection.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
286
+ page_content=' Cambridge university press, 1998.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
287
+ page_content=' [11] Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
288
+ page_content=' Zhang, Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
289
+ page_content=' Guo, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
290
+ page_content=' Wang, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
291
+ page_content=' Xi, and N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
292
+ page_content=' Wu, “Block sparse bayesian learning based joint user activity detection and channel estimation for grant-free noma systems,” IEEE Transactions on Vehicular Technology, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
293
+ page_content=' 67, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
294
+ page_content=' 10, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
295
+ page_content=' 9631–9640, 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
296
+ page_content=' [12] H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
297
+ page_content=' Zhu and G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
298
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
299
+ page_content=' Giannakis, “Exploiting sparse user activity in multiuser detection,” IEEE Transactions on Communications, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
300
+ page_content=' 59, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
301
+ page_content=' 2, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
302
+ page_content=' 454–465, Feb.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
303
+ page_content=' 2011.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
304
+ page_content=' [13] H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
305
+ page_content=' F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
306
+ page_content=' Schepker, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
307
+ page_content=' Bockelmann, and A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
308
+ page_content=' Dekorsy, “Coping with CDMA asynchronicity in compressive sensing multi-user detection,” in 2013 IEEE 77th Vehicular Technology Conference (VTC Spring), Jun.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
309
+ page_content=' 2013, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
310
+ page_content=' 1–5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
311
+ page_content=' [14] Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
312
+ page_content=' Chen, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
313
+ page_content=' Sohrabi, and W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
314
+ page_content=' Yu, “Sparse activity detection for massive connectivity,” IEEE Transactions on Signal Processing, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
315
+ page_content=' 66, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
316
+ page_content=' 7, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
317
+ page_content=' 1890–1904, Apr.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
318
+ page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
319
+ page_content=' [15] K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
320
+ page_content=' Takeuchi, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
321
+ page_content=' Tanaka, and T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
322
+ page_content=' Kawabata, “Performance improvement of iterative multiuser detection for large sparsely spread CDMA systems by spatial coupling,” IEEE Transactions on Information Theory, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
323
+ page_content=' 61, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
324
+ page_content=' 4, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
325
+ page_content=' 1768–1794, Apr.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
326
+ page_content=' 2015.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
327
+ page_content=' [16] Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
328
+ page_content=' Wang, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
329
+ page_content=' Zhu, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
330
+ page_content=' G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
331
+ page_content=' Lim, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
332
+ page_content=' Wei, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
333
+ page_content=' Liu, and Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
334
+ page_content=' Jiang, “Compressive sensing based user activity detection and channel estimation in uplink noma systems,” in 2020 IEEE Wireless Communications and Networking Conference (WCNC).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
335
+ page_content=' IEEE, 2020, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
336
+ page_content=' 1–6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
337
+ page_content=' [17] I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
338
+ page_content=' Goodfellow, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
339
+ page_content=' Bengio, and A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
340
+ page_content=' Courville, Deep learning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
341
+ page_content=' MIT press Cambridge, 2016, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
342
+ page_content=' 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
343
+ page_content=' [18] M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
344
+ page_content=' Mohammadkarimi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
345
+ page_content=' Mehrabi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
346
+ page_content=' Ardakani, and Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
347
+ page_content=' Jing, “Deep learning based sphere decoding,” IEEE Trans.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
348
+ page_content=' Wireless Commun.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
349
+ page_content=', pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
350
+ page_content=' 1–1, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
351
+ page_content=' [19] X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
352
+ page_content=' Miao, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
353
+ page_content=' Guo, and X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
354
+ page_content=' Li, “Grant-free NOMA with device activity learning using long short-term memory,” IEEE Wireless Communications Letters, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
355
+ page_content=' 1–1, 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
356
+ page_content=' [20] W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
357
+ page_content=' Zhang, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
358
+ page_content=' K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
359
+ page_content=' Mallik, and K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
360
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
361
+ page_content=' Letaief, “Cooperative spectrum sensing optimization in cognitive radio networks,” in 2008 IEEE International Conference on Communications, 2008, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
362
+ page_content=' 3411–3415.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
363
+ page_content=' [21] D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
364
+ page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
365
+ page_content=' Kingma and J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
366
+ page_content=' Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
367
+ page_content='6980, 2014.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
368
+ page_content=' [22] T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
369
+ page_content=' T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
370
+ page_content=' Cai and L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
371
+ page_content=' Wang, “Orthogonal matching pursuit for sparse signal recovery with noise,” IEEE Transactions on Information theory, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
372
+ page_content=' 57, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
373
+ page_content=' 7, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
374
+ page_content=' 4680–4688, 2011.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
375
+ page_content=' [23] D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
376
+ page_content=' L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
377
+ page_content=' Donoho, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
378
+ page_content=' Maleki, and A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
379
+ page_content=' Montanari, “Message-passing algo- rithms for compressed sensing,” Proceedings of the National Academy of Sciences, vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
380
+ page_content=' 106, no.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
381
+ page_content=' 45, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
382
+ page_content=' 18 914–18 919, 2009.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
383
+ page_content=' [24] C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
384
+ page_content=' Goutte and E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
385
+ page_content=' Gaussier, “A probabilistic interpretation of precision, re- call and f-score, with implication for evaluation,” in European conference on information retrieval.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
386
+ page_content=' Springer, 2005, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
387
+ page_content=' 345–359.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/HdAzT4oBgHgl3EQfUvz7/content/2301.01274v1.pdf'}
JNFIT4oBgHgl3EQfZCuh/content/2301.11251v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:615d060741c38e062c0ac71fb4ed52ad7f261a247a90b42f7044d142d49de416
3
+ size 5642271
JdE2T4oBgHgl3EQfUgew/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:15cc58fa3787a33564790da1b006ea35075e5166608ec5539f64846362378e9e
3
+ size 655405
KdFOT4oBgHgl3EQfzDRI/content/2301.12930v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d7423488e44e8f2586f6c1205cc378cd6d115fc280a8580b628733d4d97d429
3
+ size 1666446
MNE4T4oBgHgl3EQfiw29/content/tmp_files/2301.05137v1.pdf.txt ADDED
@@ -0,0 +1,2136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Springer Nature 2021 LATEX template
2
+ Density functions of periodic sequences of continuous events
3
+ Olga Anosova1 and Vitaliy Kurlin1*
4
+ 1*Computer Science, University of Liverpool, Ashton street, Liverpool, L69 3BX, UK.
5
+ *Corresponding author(s). E-mail(s): vitaliy.kurlin@gmail.com;
6
+ Contributing authors: oanosova@liverpool.ac.uk;
7
+ Abstract
8
+ Periodic Geometry studies isometry invariants of periodic point sets that are also continuous under
9
+ perturbations. The motivations come from periodic crystals whose structures are determined in
10
+ a rigid form but any minimal cells can discontinuously change due to small noise in measure-
11
+ ments. For any integer k ≥ 0, the density function of a periodic set S was previously defined
12
+ as the fractional volume of all k-fold intersections (within a minimal cell) of balls that have a
13
+ variable radius t and centers at all points of S. This paper introduces the density functions for
14
+ periodic sets of points with different initial radii motivated by atomic radii of chemical elements
15
+ and by continuous events occupying disjoint intervals in time series. The contributions are explicit
16
+ descriptions of the densities for periodic sequences of intervals. The new densities are strictly
17
+ stronger and distinguish periodic sequences that have identical densities in the case of zero radii.
18
+ Keywords: computational geometry, periodic set, periodic time series, isometry invariant, density function
19
+ MSC Classification: 68U05 , 51K05 , 51N20 , 51F30 , 51F20
20
+ 1 Motivations for the density
21
+ functions of periodic sets
22
+ This work substantially extends the previous
23
+ conference paper [3] in Discrete Geometry and
24
+ Mathematical Morphology 2022. The past work
25
+ explicitly described the density functions for peri-
26
+ odic sequences of zero-sized points. The new work
27
+ extends these analytic descriptions to periodic
28
+ sequences whose points have non-negative radii.
29
+ The proposed extension to the weighted case is
30
+ motivated by crystallography and materials chem-
31
+ istry [1] because all chemical elements have differ-
32
+ ent atomic radii. In dimension 1, the key motiva-
33
+ tion is the study of periodic time series consisting
34
+ of continuous and sequential (non-overlapping)
35
+ events represented by disjoint intervals. Any such
36
+ interval [a, b] ⊂ R for a ≤ b is the one-dimensional
37
+ ball with the center a + b
38
+ 2
39
+ and radius b − a
40
+ 2
41
+ .
42
+ The point-set representation of periodic crys-
43
+ tals is the most fundamental mathematical model
44
+ for crystalline materials because nuclei of atoms
45
+ are well-defined physical objects, while chemical
46
+ bonds are not real sticks or strings but abstractly
47
+ represent inter-atomic interactions depending on
48
+ many thresholds for distances and angles.
49
+ Since crystal structures are determined in a
50
+ rigid form, their most practical equivalence is rigid
51
+ motion (a composition of translations and rota-
52
+ tions) or isometry that maintains all inter-point
53
+ distances and includes also mirror reflections [20].
54
+ Now we introduce the key concepts. Let Rn be
55
+ Euclidean space, Z be the set of all integers.
56
+ 1
57
+ arXiv:2301.05137v1 [cs.CG] 12 Jan 2023
58
+
59
+ Springer Nature 2021 LATEX template
60
+ 2
61
+ Density functions of periodic sequences
62
+ 0
63
+ 0.2
64
+ 0.4
65
+ 0.6
66
+ 0.8
67
+ 1
68
+ 0
69
+ 0.2
70
+ 0.4
71
+ 0.6
72
+ 0.8
73
+ 1
74
+ 1.2
75
+ 1.4
76
+ 1.6
77
+ ψkA(t)
78
+ Radius of Balls
79
+ ψ0A
80
+ ψ1A
81
+ ψ2A
82
+ ψ3A
83
+ ψ4A
84
+ ψ5A
85
+ ψ6A
86
+ ψ7A
87
+ ψ8A
88
+ 0
89
+ 0.2
90
+ 0.4
91
+ 0.6
92
+ 0.8
93
+ 1
94
+ 0
95
+ 0.2
96
+ 0.4
97
+ 0.6
98
+ 0.8
99
+ 1
100
+ 1.2
101
+ 1.4
102
+ 1.6
103
+ Σ1nψkA(t)
104
+ Radius of Balls
105
+ n = 1
106
+ n = 2
107
+ n = 3
108
+ n = 4
109
+ n = 5
110
+ n = 6
111
+ n = 7
112
+ n = 8
113
+ Fig. 1 Illustration of Definition 1.2 for the hexagonal lattice. Left: subregions Uk(t) are covered by k disks for the radii
114
+ t = 0.25, 0.55, 0.75, 1. Right: the densities ψk are above the corresponding densigram of accumulated functions
115
+ k
116
+
117
+ i=1
118
+ ψi(t).
119
+ Definition 1.1 (a lattice Λ, a unit cell U, a
120
+ motif M, a periodic point set S = M + Λ).
121
+ For any linear basis v1, . . . , vn of Rn, a lattice
122
+ is Λ
123
+ =
124
+ {
125
+ n�
126
+ i=1
127
+ civi
128
+ :
129
+ ci
130
+
131
+ Z}. The unit cell
132
+ U(v1, . . . , vn) =
133
+ � n�
134
+ i=1
135
+ civi : ci ∈ [0, 1)
136
+
137
+ is the par-
138
+ allelepiped spanned by the basis above. A motif
139
+ M ⊂ U is any finite set of points p1, . . . , pm ∈ U.
140
+ A periodic point set [20] is the Minkowski sum
141
+ S = M + Λ = {u + v | u ∈ M, v ∈ Λ}.
142
+
143
+ In dimension n = 1, a lattice is defined by any
144
+ non-zero vector v ∈ R, any periodic point set S
145
+ is a periodic sequence {p1, . . . , pm} + vZ with the
146
+ period v equal to the length of the vector v.
147
+ Definition 1.2 (density functions for periodic
148
+ sets of points with radii). Let a periodic set S =
149
+ Λ + M ⊂ Rn have a unit cell U. For every point
150
+ p ∈ M, fix a radius r(p) ≥ 0. For any integer
151
+ k ≥ 0, let Uk(t) be the region within the cell U
152
+ covered by exactly k closed balls ¯B(p; r(p) + t)
153
+ for t ≥ 0 and all points p ∈ M and their transla-
154
+ tions by Λ. The k-th density function ψk[S](t) =
155
+ Vol[Uk(t)]/Vol[U] is the fractional volume of the
156
+ k-fold intersections of these balls within U.
157
+
158
+ The density ψk[S](t) can be interpreted as the
159
+ probability that a random (uniformly chosen in U)
160
+ point q is at a maximum distance t to exactly k
161
+ balls with initial radii r(p) and all centers p ∈ S.
162
+ For k = 0, the 0-th density ψ0[S](t) mea-
163
+ sures the fractional volume of the empty space not
164
+ covered by any expanding balls ¯B(p; r(p) + t)
165
+ In the simplest case of radii r(p) = 0, the infi-
166
+ nite sequence Ψ[S] = {ψk(t)}+∞
167
+ k=0 was called in
168
+ [8, section 3] the density fingerprint of a periodic
169
+ point set S. For k = 1 and small t > 0 while
170
+ all equal-sized balls ¯B(p; t) remain disjoint, the
171
+ 1st density ψ1[S](t) increases proportionally to tn
172
+ but later reaches a maximum and eventually drops
173
+ back to 0 when all points of Rn are covered of by at
174
+ least two balls. See the densities ψk, k = 0, . . . , 8
175
+ for the square and hexagonal lattices in [8, Fig. 2].
176
+ The original densities helped find a missing
177
+ crystal in the Cambridge Structural Database,
178
+ which was accidentally confused with a slight per-
179
+ turbation (measured at a different temperature)
180
+ of another crystal (polymorph) with the same
181
+ chemical composition, see [8, section 7].
182
+ The new weighted case with radii r(p) ≥ 0 in
183
+ Definition 1.2 is even more practically important
184
+ due to different Van der Waals radii, which are
185
+ individually defined for all chemical elements.
186
+ The key advantage of density functions over
187
+ other isometry invariants of periodic crystals
188
+
189
+ Springer Nature 2021 LATEX template
190
+ Density functions of periodic sequences
191
+ 3
192
+ (such as symmetries or conventional representa-
193
+ tions based on a geometry of a minimal cell) is
194
+ their continuity under perturbations, see details in
195
+ section 2 reviewing the related past work.
196
+ The only limitation is the infinite size of den-
197
+ sities ψk(t) due to the unbounded parameters:
198
+ integer index k ≥ 0 and continuous radius t ≥ 0.
199
+ We state the following problem in full general-
200
+ ity to motivate future work on these densities.
201
+ Problem 1.3 (computation of ψk). Verify if the
202
+ density functions ψk[S](t) from Definition 1.2 can
203
+ be computed in a polynomial time (in the size m
204
+ of a motif of S) for a fixed dimension n.
205
+
206
+ The main contribution is the full solution of
207
+ Problem 1.3 for n = 1. Theorems 3.2, 4.2, 5.2, 6.2,
208
+ and Corollary 6.4 efficiently compute all ψk[S](t)
209
+ depending on infinitely many values of k and t.
210
+ 2 Review of related past work
211
+ Periodic Geometry was initiated in 2020 by the
212
+ problem [14, section 2.3] to design a computable
213
+ metric on isometry classes of lattices, which is
214
+ continuous under perturbations of a lattice basis.
215
+ Though a Voronoi domain is combinatorially
216
+ unstable under perturbations, its geometric shape
217
+ was used to introduce two continuous metrics [14,
218
+ Theorems 2, 4] requiring approximations due to a
219
+ minimization over infinitely many rotations.
220
+ Similar minimizations over rotations or other
221
+ continuous parameters are required for the com-
222
+ plete invariant isosets [2, 4] and density functions,
223
+ which can be practically computed in low dimen-
224
+ sions [16], whose completeness was proved for
225
+ generic periodic point sets in R3 [8, Theorem 2].
226
+ The density fingerprint Ψ[S] turned out to be
227
+ incomplete [8, section 5] in the example below.
228
+ Example 2.1 (periodic sequences S15, Q15 ⊂ R).
229
+ Widdowson et al. [20, Appendix B] discussed
230
+ homometric sets that can be distinguished by
231
+ the invariant AMD (Average Minimum Distances)
232
+ and not by diffraction patterns. The sequences
233
+ S15 = {0, 1, 3, 4, 5, 7, 9, 10, 12} + 15Z,
234
+ Q15 = {0, 1, 3, 4, 6, 8, 9, 12, 14} + 15Z
235
+ have the unit cell [0, 15] shown as a circle in Fig. 2.
236
+ Fig. 2 Circular versions of the periodic sets S15, Q15.
237
+ These periodic sequences [9] are obtained as
238
+ Minkowski sums S15 = U + V + 15Z and Q15 =
239
+ U − V + 15Z for U = {0, 4, 9}, V = {0, 1, 3}.
240
+
241
+ For rational-valued periodic sequences, [9,
242
+ Theorem 4] proved that r-th order invariants
243
+ (combinations of r-factor products) up to r = 6
244
+ are enough to distinguish such sequences up to a
245
+ shift (a rigid motion of R without reflections).
246
+ The AMD invariant was extended to the Point-
247
+ wise Distance Distribution (PDD), whose generic
248
+ completeness [19, Theorem 4.4] was proved in
249
+ any dimension n ≥ 1. However there are finite
250
+ sets in R3 [15, Fig. S4] with the same PDD,
251
+ which were distinguished by more sophisticated
252
+ distance-based invariants in [18, appendix C].
253
+ The
254
+ subarea
255
+ of
256
+ Lattice
257
+ Geometry
258
+ devel-
259
+ oped continuous parameterizations for the moduli
260
+ spaces of lattices considered up to isometry in
261
+ dimension two [7, 13] and three [6, 10].
262
+ For 1-periodic sequences of points in Rn, com-
263
+ plete isometry invariants with continuous and
264
+ computable metrics appeared in [12], see related
265
+ results for finite clouds of unlabeled points [11, 17].
266
+ 3 The 0-th density function ψ0
267
+ This section proves Theorem 3.2 explicitly describ-
268
+ ing the 0-th density function ψ0[S](t) for any
269
+ periodic sequence S ⊂ R of disjoint intervals.
270
+ For convenience, scale any periodic sequence
271
+ S to period 1 so that S is given by points
272
+ 0 ≤ p1 < · · · < pm < 1 with radii r1, . . . , rm,
273
+ respectively. Since the expanding balls in R are
274
+ growing intervals, volumes of their intersections
275
+ linearly change with respect to the variable radius
276
+ t. Hence any density function ψk(t) is piecewise
277
+ linear and uniquely determined by corner points
278
+ (aj, bj) where the gradient of ψk(t) changes.
279
+
280
+ Springer Nature 2021 LATEX template
281
+ 4
282
+ Density functions of periodic sequences
283
+ To prepare the proof of Theorem 3.2, we first
284
+ consider Example 3.1 for the simple sequence S.
285
+ Example 3.1 (0-th density function ψ0). Let the
286
+ periodic sequence S =
287
+
288
+ 0, 1
289
+ 3, 1
290
+ 2
291
+
292
+ + Z have three
293
+ points p1 = 0, p2 = 1
294
+ 3, p3 = 1
295
+ 2 of radii r1 = 1
296
+ 12,
297
+ r2 = 0, r3 = 1
298
+ 12, respectively. Fig. 3 shows each
299
+ point pi and its growing interval
300
+ Li(t) = [(pi−ri)−t, (pi+ri)+t] of the length 2ri+2t
301
+ for i = 1, 2, 3 in its own color: red, green, blue.
302
+ By
303
+ Definition
304
+ 1.2
305
+ each
306
+ density
307
+ function
308
+ ψk[S](t) measures a fractional length covered by
309
+ exactly k intervals within the unit cell [0, 1]. We
310
+ periodicaly map the endpoints of each growing
311
+ interval to the unit cell [0, 1]. For instance, the
312
+ interval [− 1
313
+ 12 − t, 1
314
+ 12 + t] of the point p1 = 0 ≡ 1
315
+ (mod 1) maps to the red intervals [0, 1
316
+ 12 +t]∪[11
317
+ 12 −
318
+ t, 1] shown by solid red lines in Fig. 3. The same
319
+ image shows the green interval [1
320
+ 3 − t, 1
321
+ 3 + t] by
322
+ dashed lines and the blue interval [ 5
323
+ 12 − t, 7
324
+ 12 + t]
325
+ by dotted lines.
326
+ At the moment t = 0, since the starting inter-
327
+ vals are disjoint, they cover the length l = 2( 1
328
+ 12 +
329
+ 0 + 1
330
+ 12) = 1
331
+ 3. The non-covered part of [0, 1] has
332
+ length 1 − 1
333
+ 3 = 2
334
+ 3. So the graph of ψ0(t) at t = 0
335
+ starts from the point (0, 2
336
+ 3), see Fig. 4.
337
+ At the first critical moment t = 1
338
+ 24 when the
339
+ green and blue intervals collide at p = 3
340
+ 8, only
341
+ the intervals [1
342
+ 8, 7
343
+ 24] ∪ [5
344
+ 8, 7
345
+ 8] of total length
346
+ 5
347
+ 12
348
+ remain uncovered. Hence ψ0(t) linearly drops to
349
+ the point ( 1
350
+ 12, 5
351
+ 12). At the next critical moment
352
+ t = 1
353
+ 8 when the red and green intervals collide at
354
+ p =
355
+ 5
356
+ 24, only the interval [17
357
+ 24, 19
358
+ 24] of length
359
+ 1
360
+ 12
361
+ remain uncovered, so ψ0(t) continues to (1
362
+ 8, 1
363
+ 12).
364
+ The graph of ψ0(t) finally returns to the t-axis
365
+ at the point (1
366
+ 6, 0) and remains there for t ≥ 1
367
+ 6.
368
+ The piecewise linear behavior of ψ0(t) can be
369
+ described by specifying the corner points in Fig. 4:
370
+
371
+ 0, 2
372
+ 3
373
+
374
+ ,
375
+ � 1
376
+ 24, 5
377
+ 12
378
+
379
+ ,
380
+ �1
381
+ 8, 1
382
+ 12
383
+
384
+ ,
385
+ �1
386
+ 6, 0
387
+
388
+ .
389
+
390
+ Theorem 3.2 extends Example 3.1 to any peri-
391
+ odic sequence S and implies that the 0-th density
392
+ function ψ0(t) is uniquely determined by the
393
+ ordered gap lengths between successive intervals.
394
+ Theorem 3.2 (description of ψ0). Let a periodic
395
+ sequence S = {p1, . . . , pm} + Z consist of disjoint
396
+ intervals with centers 0 ≤ p1 < · · · < pm < 1 and
397
+ radii r1, . . . , rm ≥ 0. Consider the total length l =
398
+ 2
399
+ m
400
+
401
+ i=1
402
+ ri and gaps between successive intervals gi =
403
+ (pi − ri) − (pi−1 + ri−1), where i = 1, . . . , m and
404
+ p0 = pm − 1, r0 = rm. Put the gaps in increasing
405
+ order: g[1] ≤ g[2] ≤ · · · ≤ g[m].
406
+ Then
407
+ the
408
+ 0-th
409
+ density
410
+ ψ0[S](t)
411
+ is
412
+ piecewise
413
+ linear
414
+ with
415
+ the
416
+ following
417
+ (unordered)
418
+ corner
419
+ points:
420
+ (0, 1 − l)
421
+ and
422
+
423
+ g[i]
424
+ 2 , 1 − l −
425
+ i−1
426
+
427
+ j=1
428
+ g[j] − (m − i + 1)g[i]
429
+
430
+ for
431
+ i = 1, . . . , m, so the last corner is
432
+ �g[m]
433
+ 2 , 0
434
+
435
+ .
436
+ If any corners are repeated, e.g. when g[i−1] =
437
+ g[i], these corners are collapsed into one corner. ■
438
+ Proof By Definition 1.2 the 0-th density function
439
+ ψ0(t) measures the total length of subintervals in the
440
+ unit cell [0, 1] that are not covered by any of the grow-
441
+ ing intervals Li(t) = [pi−ri−t, pi+ri+t], i = 1, . . . , m.
442
+ For t = 0, since all initial intervals Li(0) are disjoint,
443
+ they cover the total length 2
444
+ m
445
+
446
+ i=1
447
+ ri = l.
448
+ Then the graph of ψ0(t) at t = 0 starts from the
449
+ point (0, 1 − l). So ψ0(t) linearly decreases from the
450
+ initial value ψ0(0) = 1 − l except for m critical values
451
+ of t where one of the gap intervals [pi + ri + t, pi+1 −
452
+ ri+1−t] between successive growing intervals Li(t) and
453
+ Li+1(t) shrinks to a point. These critical radii t are
454
+ ordered according to the gaps g[1] ≤ g[2] ≤ · · · ≤ g[m].
455
+ The first critical radius is t =
456
+ 1
457
+ 2g[1], when a
458
+ shortest gap interval of the length g[1] is covered
459
+ by the growing successive intervals. At this moment
460
+
461
+ Springer Nature 2021 LATEX template
462
+ Density functions of periodic sequences
463
+ 5
464
+ Fig. 3 The sequence S =
465
+
466
+ 0, 1
467
+ 3 , 1
468
+ 2
469
+
470
+ + Z has the points of weights
471
+ 1
472
+ 12 , 0, 1
473
+ 12 , respectively. The growing intervals around
474
+ the red point 0 ≡ 1 (mod 1), green point 1
475
+ 3 , blue point 1
476
+ 2 have the same color for various radii t, see Examples 3.1, 4.1, 5.1.
477
+ t = 1
478
+ 2g[1], all m growing intervals Li(t) have the total
479
+ length l + mg[1]. Then the 0-th density ψ0(t) has the
480
+ first corner points (0, 1−l) and
481
+ �g[1]
482
+ 2 , 1 − l − mg[1]
483
+
484
+ .
485
+ The second critical radius is t
486
+ =
487
+ g[2]
488
+ 2 , when
489
+ all
490
+ intervals
491
+ Li(t)
492
+ have
493
+ the
494
+ total
495
+ length
496
+ l +
497
+ g[1] + (m − 1)g[2], i.e. the next corner point is
498
+ �g[2]
499
+ 2 , 1 − l − g[1] − (m − 1)g[2]
500
+
501
+ . If g[1] = g[2], then
502
+ both corner points coincide, so ψ0(t) will continue
503
+ from the joint corner point.
504
+ The above pattern generalizes to the i-th critical
505
+ radius t = 1
506
+ 2g[i], when all covered intervals have the
507
+ total length
508
+ i−1
509
+
510
+ j=1
511
+ g[j] (for the fully covered intervals)
512
+ plus (m − i + 1)g[i] (for the still growing intervals).
513
+ For the final critical radius t =
514
+ g[m]
515
+ 2 , the whole
516
+ unit cell [0, 1] is covered by the grown intervals because
517
+ m
518
+
519
+ j=1
520
+ g[j] = 1 − l. The final corner is (
521
+ g[m]
522
+ 2 , 0).
523
+
524
+ Example 3.3 applies Theorem 3.2 to get ψ0
525
+ found for the periodic sequence S in Example 3.1.
526
+
527
+ 1/3Springer Nature 2021 LATEX template
528
+ 6
529
+ Density functions of periodic sequences
530
+ Fig. 4 The 0-th density function ψ0(t) for the 1-period
531
+ sequence S whose points 0, 1
532
+ 3 , 1
533
+ 2
534
+ have radii
535
+ 1
536
+ 12 , 0, 1
537
+ 12 ,
538
+ respectively, see Example 3.1.
539
+ Example 3.3 (using Theorem 3.2). The sequence
540
+ S =
541
+
542
+ 0, 1
543
+ 3, 1
544
+ 2
545
+
546
+ + Z in Example 3.1 with points
547
+ p1 = 0, p2 = 1
548
+ 3, p3 = 1
549
+ 2 of radii r1 = 1
550
+ 12, r2 = 0,
551
+ r3 = 1
552
+ 12, respectively, has l = 2(r1 + r2 + r3) = 1
553
+ 3
554
+ and the initial gaps between successive intervals
555
+ g1 = p1 − r1 − p3 − r3 = (1 − 1
556
+ 12) − (1
557
+ 2 + 1
558
+ 12) = 1
559
+ 3,
560
+ g2 = p2 − r2 − p1 − r1 = (1
561
+ 3 − 0) − (0 + 1
562
+ 12) = 1
563
+ 4,
564
+ g3 = p3 − r3 − p2 − r2 = (1
565
+ 2 − 1
566
+ 12) − (1
567
+ 3 + 0) = 1
568
+ 12.
569
+ Order the gaps: g[1] = 1
570
+ 12 < g[2] = 1
571
+ 4 < g[3] = 1
572
+ 3.
573
+ 1 − l = 1 − 1
574
+ 3 = 2
575
+ 3,
576
+ 1 − l − 3g[1] = 2
577
+ 3 − 3
578
+ 12 = 5
579
+ 12,
580
+ 1 − l − g[1] − 2g[2] = 2
581
+ 3 − 1
582
+ 12 − 2
583
+ 4 = 1
584
+ 12,
585
+ 1 − l − g[1] − g[2] − g[3] = 2
586
+ 3 − 1
587
+ 12 − 1
588
+ 4 − 1
589
+ 3 = 0.
590
+ By Theorem 3.2 ψ0(t) has the corner points
591
+ (0, 1 − l) =
592
+
593
+ 0, 2
594
+ 3
595
+
596
+ ,
597
+ �1
598
+ 2g[1], 1 − l − 3g[1]
599
+
600
+ =
601
+ � 1
602
+ 24, 5
603
+ 12
604
+
605
+ ,
606
+ �1
607
+ 2g[2], 1 − l − g[1] − 2g[2]
608
+
609
+ =
610
+ �1
611
+ 8, 1
612
+ 12
613
+
614
+ ,
615
+ �1
616
+ 2g[3], 1 − l − g[1] − g[2] − g[3]
617
+
618
+ =
619
+ �1
620
+ 6, 0
621
+
622
+ . See
623
+ the graph of the 0-th density ψ0(t) in Fig. 4.
624
+
625
+ By Theorem 3.2 any 0-th density function
626
+ ψ0(t) is uniquely determined by the (unordered)
627
+ set of gap lengths between successive intervals.
628
+ Hence we can re-order these intervals with-
629
+ out changing ψ0(t). For instance, the periodic
630
+ sequence Q = {0, 1
631
+ 2, 2
632
+ 3} + Z with points 0, 1
633
+ 2, 2
634
+ 3 of
635
+ weights
636
+ 1
637
+ 12, 1
638
+ 12, 0 has the same set ordered gaps
639
+ g[1] =
640
+ 1
641
+ 12, d[2] = 1
642
+ 3, d[3] = 1
643
+ 2 as the periodic
644
+ sequence S =
645
+
646
+ 0, 1
647
+ 3, 1
648
+ 2
649
+
650
+ + Z in Example 3.1.
651
+ The above sequences S, Q are related by the
652
+ mirror reflection t
653
+ �→
654
+ 1 − t. One can eas-
655
+ ily construct many non-isometric sequences with
656
+ ψ0[S](t) = ψ0[Q](t). For any 1 ≤ i ≤ m − 3,
657
+ the sequences Sm,i = {0, 2, 3, . . . , i + 2, i + 4, i +
658
+ 5, . . . , m + 2} + (m + 2)Z have the same interval
659
+ lengths d[1] = · · · = d[m−2] = 1, d[m−1] = d[m] = 2
660
+ but are not related by isometry (translations and
661
+ reflections in R) because the intervals of length 2
662
+ are separated by i−1 intervals of length 1 in Sm,i.
663
+ 4 The 1st density function ψ1
664
+ This section proves Theorem 4.2 explicitly describ-
665
+ ing the 1st density ψ1[S](t) for any periodic
666
+ sequence S of disjoint intervals. To prepare the
667
+ proof of Theorem 4.2, Example 4.1 finds ψ1[S] for
668
+ the sequence S from Example 3.1.
669
+ Example 4.1 (ψ1 for S =
670
+
671
+ 0, 1
672
+ 3, 1
673
+ 2
674
+
675
+ + Z). The
676
+ 1st density function ψ1(t) can be obtained as a
677
+ sum of the three trapezoid functions ηR, ηG, ηB,
678
+ each measuring the length of a region covered by
679
+ a single interval of one color, see Fig. 3.
680
+ At the initial moment t = 0, the red intervals
681
+ [0, 1
682
+ 12] ∪ [11
683
+ 12, 1] have the total length ηR(0) = 1
684
+ 6.
685
+ These red intervals [0, 1
686
+ 12 + t] ∪ [11
687
+ 12 − t, 1] for
688
+ t ∈ [0, 1
689
+ 8] grow until they touch the green interval
690
+ [ 7
691
+ 24, 3
692
+ 8] and have the total length ηR(1
693
+ 8) = 1
694
+ 6 + 2
695
+ 8 =
696
+
697
+ Springer Nature 2021 LATEX template
698
+ Density functions of periodic sequences
699
+ 7
700
+ Fig. 5 The trapezoid functions ηR, ηG, ηB and the 1st
701
+ density function ψ1(t) for the 1-period sequence S whose
702
+ points 0, 1
703
+ 3 , 1
704
+ 2 have radii 1
705
+ 12 , 0, 1
706
+ 12 , see Example 4.1.
707
+ 5
708
+ 12 in the second picture of Fig. 3. So the graph of
709
+ the red length ηR(t) linearly grows with gradient 2
710
+ from the point (0, 1
711
+ 6) to the corner point (1
712
+ 8, 5
713
+ 12).
714
+ For t ∈ [1
715
+ 8, 1
716
+ 6], the left red interval is shrink-
717
+ ing at the same rate (due to the overlapping green
718
+ interval) as the right red interval continues to grow
719
+ until t = 1
720
+ 6, when it touches the blue interval
721
+ [1
722
+ 4, 3
723
+ 4]. Hence the graph of ηR(t) remains constant
724
+ for t ∈ [1
725
+ 8, 1
726
+ 6] up to the corner point (1
727
+ 6, 5
728
+ 12).
729
+ After
730
+ that,
731
+ the
732
+ graph
733
+ of
734
+ ηR(t)
735
+ linearly
736
+ decreases (with gradient −2) until all red intervals
737
+ are fully covered by the green and blue intervals
738
+ at moment t = 3
739
+ 8, see the 6th picture in Fig. 3.
740
+ Hence the trapezoid function ηR has the piece-
741
+ wise linear graph through the corner points (0, 1
742
+ 6),
743
+ (1
744
+ 8, 5
745
+ 12), (1
746
+ 6, 5
747
+ 12), (3
748
+ 8, 0). After that, ηR(t) = 0
749
+ remains constant for t ≥ 3
750
+ 8. Fig. 5 shows the
751
+ graphs of ηR, ηG, ηB and ψ1 = ηR + ηG + ηB.
752
+
753
+ Theorem 4.2 extends Example 4.1 and proves
754
+ that any ψ1(t) is a sum of trapezoid functions
755
+ whose corners are explicitly described. We con-
756
+ sider any index i = 1, . . . , m (of a point pi or a
757
+ gap gi) modulo m so that m + 1 ≡ 1 (mod m).
758
+ Theorem 4.2 (description of ψ1). Let a periodic
759
+ sequence S = {p1, . . . , pm} + Z consist of disjoint
760
+ intervals with centers 0 ≤ p1 < · · · < pm < 1 and
761
+ radii r1, . . . , rm ≥ 0, respectively.
762
+ Consider the gaps gi = (pi−ri)−(pi−1+ri−1),
763
+ where i = 1, . . . , m and p0 = pm − 1, r0 = rm.
764
+ Then the 1st density ψ1(t) is the sum of m
765
+ trapezoid functions ηi, i = 1, . . . , m, with the
766
+ corners (0, 2ri),
767
+ �gi
768
+ 2 , g + 2ri
769
+
770
+ ,
771
+ �gi+1
772
+ 2 , g + 2ri
773
+
774
+ ,
775
+ �gi + gi+1
776
+ 2
777
+ + ri, 0
778
+
779
+ , where g = min{gi, gi+1}.
780
+ Hence ψ1(t) is determined by the unordered
781
+ set of unordered pairs (gi, gi+1), i = 1, . . . , m.
782
+
783
+ Proof The 1st density ψ1(t) equals the total length
784
+ of subregions covered by exactly one of the intervals
785
+
786
+ RSpringer Nature 2021 LATEX template
787
+ 8
788
+ Density functions of periodic sequences
789
+ Li(t) = [pi − ri − t, pi + ri + t], i = 1, . . . , m, where all
790
+ intervals are taken modulo 1 within [0, 1].
791
+ Hence ψ1(t) is the sum of the functions η1i, each
792
+ measuring the length of the subinterval of Li(t) not
793
+ covered by other intervals Lj(t), j ∈ {1, . . . , m}−{i}.
794
+ Since the initial intervals Li(0) are disjoint, each
795
+ function η1i(t) starts from the value η1i(0) = 2ri and
796
+ linearly grows (with gradient 2) up to ηi(1
797
+ 2g) = 2ri+g,
798
+ where g = min{gi, gi+1}, when the growing interval
799
+ Li(t) of the length 2ri+2t = 2ri+g touches its closest
800
+ neighboring interval Li±1(t) with a shortest gap g.
801
+ If (say) gi < gi+1, then the subinterval covered
802
+ only by Li(t) is shrinking on the left and is grow-
803
+ ing at the same rate on the right until Li(t) touches
804
+ the growing interval Li+1(t) on the right. During
805
+ this growth, when t is between 1
806
+ 2gi and 1
807
+ 2gi+1, the
808
+ trapezoid function ηi(t) = g remains constant.
809
+ If gi = gi+1, this horizontal line collapses to one
810
+ point in the graph of ηi(t). For t ≥ max{gi, gi+1},
811
+ the subinterval covered only by Li(t) is shrinking on
812
+ both sides until the neighboring intervals Li±1(t) meet
813
+ at a mid-point between their initial closest endpoints
814
+ pi−1 + ri−1 and pi+1 − ri+1. This meeting time is
815
+ t = 1
816
+ 2(pi+1 −ri+1 −pi−1 −ri−1) = 1
817
+ 2(gi +2ri +gi+1),
818
+ which is also illustrated by Fig. 6. So the trapezoid
819
+ function ηi has the corners (0, 2ri),
820
+ �gi
821
+ 2 , 2ri + g
822
+
823
+ ,
824
+ �gi+1
825
+ 2
826
+ , 2ri + g
827
+
828
+ ,
829
+ �gi + gi+1
830
+ 2
831
+ + ri, 0
832
+
833
+ as expected.
834
+
835
+ Example 4.3 applies Theorem 4.2 to get ψ1
836
+ found for the periodic sequence S in Example 4.1.
837
+ Example 4.3 (using Theorem 4.2 for ψ1). The
838
+ sequence S =
839
+
840
+ 0, 1
841
+ 3, 1
842
+ 2
843
+
844
+ + Z in Example 4.1 with
845
+ points p1 = 0, p2 = 1
846
+ 3, p3 = 1
847
+ 2 of radii r1 = 1
848
+ 12,
849
+ r2 = 0, r3 = 1
850
+ 12, respectively, has the initial gaps
851
+ between successive intervals g1 = 1
852
+ 3, g2 = 1
853
+ 4, g3 =
854
+ 1
855
+ 12, see all the computations in Example 3.3.
856
+ Case (R). In Theorem 4.2 for the trapezoid func-
857
+ tion ηR = η1 measuring the fractional length
858
+ covered only by the red interval, we set k = 1 and
859
+ i = 1. Then ri = 1
860
+ 12, gi = 1
861
+ 3 and gi+1 = 1
862
+ 4, so
863
+ gi + gi+1
864
+ 2
865
+ + ri = 1
866
+ 2
867
+ �1
868
+ 3 + 1
869
+ 4
870
+
871
+ + 1
872
+ 12 = 3
873
+ 8,
874
+ g = min{gi, gi+1} = 1
875
+ 4, g + 2ri = 1
876
+ 4 + 2
877
+ 12 = 5
878
+ 12.
879
+ Then ηR = η1 has the following corner points:
880
+ (0, 2ri) =
881
+
882
+ 0, 1
883
+ 6
884
+
885
+ ,
886
+ �gi
887
+ 2 , g + 2ri
888
+
889
+ =
890
+ �1
891
+ 6, 5
892
+ 12
893
+
894
+ ,
895
+ �gi+1
896
+ 2 , g + 2ri
897
+
898
+ =
899
+ �1
900
+ 8, 5
901
+ 12
902
+
903
+ ,
904
+ �gi + gi+1
905
+ 2
906
+ + ri, 0
907
+
908
+ =
909
+ �3
910
+ 8, 0
911
+
912
+ ,
913
+ where the two middle corners are accidentally
914
+ swapped due to gi > gi+1 but they define the same
915
+ trapezoid function as in the first picture of Fig. 5.
916
+ Case (G). In Theorem 4.2 for the trapezoid func-
917
+ tion ηG = η2 measuring the fractional length
918
+ covered only by the green interval, we set k = 1
919
+ and i = 2. Then ri = 0, gi = 1
920
+ 4 and gi+1 = 1
921
+ 12, so
922
+ gi + gi+1
923
+ 2
924
+ + ri = 1
925
+ 2
926
+ �1
927
+ 4 + 1
928
+ 12
929
+
930
+ + 0 = 1
931
+ 6,
932
+ g = min{gi, gi+1} = 1
933
+ 12, g + 2ri = 1
934
+ 12 + 0 = 1
935
+ 12.
936
+ Then ηG = η2 has the following corner points
937
+ exactly as shown in the second picture of Fig. 5:
938
+ (0, 2ri) = (0, 0) ,
939
+ �gi
940
+ 2 , g + 2ri
941
+
942
+ =
943
+ �1
944
+ 8, 1
945
+ 12
946
+
947
+ ,
948
+ �gi+1
949
+ 2 , g + 2ri
950
+
951
+ =
952
+ � 1
953
+ 24, 5
954
+ 12
955
+
956
+ ,
957
+ �gi + gi+1
958
+ 2
959
+ + ri, 0
960
+
961
+ =
962
+ �1
963
+ 6, 0
964
+
965
+ .
966
+ Case (B). In Theorem 4.2 for the trapezoid func-
967
+ tion ηB = η3 measuring the fractional length
968
+ covered only by the blue interval, we set k = 1 and
969
+
970
+ Springer Nature 2021 LATEX template
971
+ Density functions of periodic sequences
972
+ 9
973
+ Fig. 6 The distances g, s, g′ between line intervals used in the proofs of Theorems 4.2 and 5.2, shown here for k = 3.
974
+ i = 3. Then ri = 1
975
+ 12, gi = 1
976
+ 12 and gi+1 = 1
977
+ 3, so
978
+ gi + gi+1
979
+ 2
980
+ + ri = 1
981
+ 2
982
+ � 1
983
+ 12 + 1
984
+ 3
985
+
986
+ + 1
987
+ 12 = 7
988
+ 24,
989
+ g = min{gi, gi+1} = 1
990
+ 12, g + 2ri = 1
991
+ 12 + 2
992
+ 12 = 1
993
+ 4.
994
+ Then ηB = η3 has the following corner points:
995
+ (0, 2ri) =
996
+
997
+ 0, 1
998
+ 6
999
+
1000
+ ,
1001
+ �gi
1002
+ 2 , g + 2ri
1003
+
1004
+ =
1005
+ � 1
1006
+ 24, 1
1007
+ 4
1008
+
1009
+ ,
1010
+ �gi+1
1011
+ 2 , g + 2ri
1012
+
1013
+ =
1014
+ �1
1015
+ 6, 1
1016
+ 4
1017
+
1018
+ ,
1019
+ �gi + gi+1
1020
+ 2
1021
+ + ri, 0
1022
+
1023
+ =
1024
+ � 7
1025
+ 24, 0
1026
+
1027
+ exactly as shown in the third picture of Fig. 5. ■
1028
+ 5 Higher density functions ψk
1029
+ This section proves Theorem 5.2 describing the k-
1030
+ th density function ψk[S](t) for any k ≥ 2 and a
1031
+ periodic sequence S of disjoint intervals.
1032
+ To prepare the proof of Theorem 5.2, Exam-
1033
+ ple 5.1 computes ψ2[S] for S from Example 3.1.
1034
+ Example 5.1 (ψ2 for S =
1035
+
1036
+ 0, 1
1037
+ 3, 1
1038
+ 2
1039
+
1040
+ + Z). The
1041
+ density ψ2(t) can be found as the sum of the trape-
1042
+ zoid functions ηGB, ηBR, ηRG, each measuring the
1043
+ length of a double intersection, see Fig. 3.
1044
+ For the green interval [1
1045
+ 3 −t, 1
1046
+ 3 +t] and the blue
1047
+ interval [ 5
1048
+ 12 − t, 7
1049
+ 12 + t], the graph of the function
1050
+ ηGB(t) is piecewise linear and starts at the point
1051
+ ( 1
1052
+ 24, 0) because these intervals touch at t = 1
1053
+ 24.
1054
+ The green-blue intersection [ 5
1055
+ 12 −t, 1
1056
+ 3 +t] grows
1057
+ until t = 1
1058
+ 6, when the resulting interval [1
1059
+ 4, 1
1060
+ 2]
1061
+ touches the red interval on the left. At the same
1062
+ time, the graph of ηGB(t) is linearly growing (with
1063
+ gradient 2) to the corner (1
1064
+ 6, 1
1065
+ 4), see Fig, 7.
1066
+ For t ∈ [1
1067
+ 6, 7
1068
+ 24], the green-blue intersection
1069
+ interval becomes shorter on the left, but grows at
1070
+ the same rate on the right until t = 7
1071
+ 24 when [1
1072
+ 8, 5
1073
+ 8]
1074
+ touches the red interval [5
1075
+ 8, 1] on the right, see
1076
+ the 5th picture in Fig. 3. So the graph of ηGB(t)
1077
+ remains constant up to the point ( 7
1078
+ 24, 1
1079
+ 4).
1080
+ For t ∈ [ 7
1081
+ 24, 5
1082
+ 12] the green-blue intersection
1083
+ interval is shortening from both sides. So the
1084
+ graph of ηGB(t) linearly decreases (with gradient
1085
+ −2) and returns to the t-axis at the corner ( 5
1086
+ 12, 0),
1087
+ then remains constant ηGB(t) = 0 for t ≥ 5
1088
+ 12.
1089
+ Fig. 7 shows all trapezoid functions for double
1090
+ intersections and ψ2 = ηGB + ηBR + ηRG.
1091
+
1092
+ Theorem 5.2 (description of ψk for k ≥ 2). Let
1093
+ a periodic sequence S = {p1, . . . , pm} + Z consist
1094
+ of disjoint intervals with centers 0 ≤ p1 < · · · <
1095
+ pm < 1 and radii r1, . . . , rm ≥ 0, respectively.
1096
+ Consider the gaps gi = (pi − ri) − (pi−1 + ri−1)
1097
+ between the successive intervals of S, where i =
1098
+ 1, . . . , m and p0 = pm − 1, r0 = rm.
1099
+ For k ≥ 2, the density function ψk(t) equals
1100
+ the sum of m trapezoid functions ηk,i(t), i =
1101
+ 1, . . . , m, each having the following corner points:
1102
+ �s
1103
+ 2, 0
1104
+
1105
+ ,
1106
+ �g + s
1107
+ 2
1108
+ , g
1109
+
1110
+ ,
1111
+ �s + g′
1112
+ 2
1113
+ , g
1114
+
1115
+ ,
1116
+ �g + s + g′
1117
+ 2
1118
+ , 0
1119
+
1120
+ ,
1121
+
1122
+ +K
1123
+ P
1124
+ i+k-1
1125
+ 1+1
1126
+ distance
1127
+ distance
1128
+ CSpringer Nature 2021 LATEX template
1129
+ 10
1130
+ Density functions of periodic sequences
1131
+ Fig. 7 The trapezoid functions ηGB, ηBR, ηRG and the
1132
+ 2nd density function ψ2(t) for the 1-period sequence S
1133
+ whose points 0, 1
1134
+ 3 , 1
1135
+ 2 have radii 1
1136
+ 12 , 0, 1
1137
+ 12 , see Example 5.1.
1138
+ where g, g′ are the minimum and maximum values
1139
+ in the pair {gi + 2ri, gi+k + 2ri+k−1}, and s =
1140
+ i+k−1
1141
+
1142
+ j=i+1
1143
+ gj + 2
1144
+ i+k−2
1145
+
1146
+ j=i+1
1147
+ rj, so s = gi+1 for k = 2.
1148
+ Hence ψk(t) is determined by the unordered
1149
+ set of the ordered tuples (g, s, g′), i = 1, . . . , m. ■
1150
+ Proof The k-th density function ψk(t) measures the
1151
+ total fractional length of k-fold intersections among m
1152
+ intervals Li(t) = [pi − ri − t, pi + ri + t], i = 1, . . . , m.
1153
+ Now we visualize all such intervals Li(t) in the line R
1154
+ without mapping them modulo 1 to the unit cell [0, 1].
1155
+ Since all radii ri ≥ 0, only k successive inter-
1156
+ vals can contribute to k-fold intersections. So a k-fold
1157
+ intersection of growing intervals emerges only when
1158
+ two intervals Li(t) and Li+k−1(t) overlap because
1159
+ their intersection should be also covered by all the
1160
+ intermediate intervals Li(t), Li+1(t), . . . , Li+k−1(t).
1161
+ Then the density ψk(t) equals the sum of the m
1162
+ trapezoid functions ηk,i, i = 1, . . . , m, each equal to
1163
+ the length of the k-fold intersection ∩i+k−1
1164
+ j=i
1165
+ Lj(t) not
1166
+ covered by other intervals. Then ηk,i(t) remains 0 until
1167
+ the first critical moment t when 2t equals the distance
1168
+ between the points pi + ri and pi+k−1 − ri+k−1 in R,
1169
+ see Fig. 6, so 2t =
1170
+ i+k−1
1171
+
1172
+ j=i+1
1173
+ gj + 2
1174
+ i+k−2
1175
+
1176
+ j=i+1
1177
+ rj = s. Hence
1178
+ t = s
1179
+ 2 and ( s
1180
+ 2, 0) is the first corner point of ηk,i(t).
1181
+ At t = s
1182
+ 2, the interval of the k-fold intersection
1183
+ ∩i+k−1
1184
+ j=i
1185
+ Lj(t) starts expanding on both sides. Hence
1186
+ ηk,i(t) starts increasing (with gradient 2) until the
1187
+ k-fold intersection touches one of the neighboring
1188
+ intervals Li−1(t) or Li+k(t) on the left or on the right.
1189
+ The left interval Li−1(t) touches the k-fold inter-
1190
+ section ∩i+k−1
1191
+ j=i
1192
+ Lj(t) when 2t equals the distance from
1193
+ pi−1 + ri−1 (the right endpoint of Li−1) to pi+k−1 −
1194
+ ri+k−1 (the left endpoint of Li+k−1), see Fig. 6, so
1195
+ 2t =
1196
+ i+k−1
1197
+
1198
+ j=i
1199
+ gj + 2
1200
+ i+k−2
1201
+
1202
+ j=i
1203
+ rj = gi + 2ri + s.
1204
+ The right interval Li+k−1(t′) touches the k-fold
1205
+ intersection ∩i+k−1
1206
+ j=i
1207
+ Lj(t′) when 2t′ equals the distance
1208
+ from pi + ri (the right endpoint of Li) to pi+k − ri+k
1209
+ (the left endpoint of Li+k), see Fig. 6, so
1210
+ 2t′ =
1211
+ i+k
1212
+
1213
+ j=i+1
1214
+ gj + 2
1215
+ i+k−1
1216
+
1217
+ j=i+1
1218
+ rj = s + gi+k + 2ri+k−1.
1219
+ If (say) gi + 2ri = g < g′ = gi+k + 2ri+k−1, the
1220
+ k-fold intersection ∩i+k−1
1221
+ j=i
1222
+ Lj(t) first touches Li��1 at
1223
+ the earlier moment t before reaching Li+k(t′) at the
1224
+ later moment t′. At the earlier moment, ηk,i(t) equals
1225
+ 2(t − s
1226
+ 2) = gi + 2ri = g and has the corner (g + s
1227
+ 2
1228
+ , g).
1229
+ After that, the k-fold intersection is shrinking on
1230
+ the left and is expanding at the same rate on the right.
1231
+ So the function ηk,i(t) = g remains constant until the
1232
+ k-fold intersection touches the right interval Li+k(t′).
1233
+
1234
+ GB
1235
+ BR
1236
+ RGSpringer Nature 2021 LATEX template
1237
+ Density functions of periodic sequences
1238
+ 11
1239
+ At this later moment t′ = s + gi+k
1240
+ 2
1241
+ + ri+k−1 = g′,
1242
+ ηk,i(t′) still equals g and has the corner (s + g′
1243
+ 2
1244
+ , g).
1245
+ If gi + 2ri = g′ > g = gi+k + 2ri+k−1, the grow-
1246
+ ing intervals Li−1(t) and Li+k−1(t) touch the k-fold
1247
+ intersection ∩i+k−1
1248
+ j=i
1249
+ Lj(t) in the opposite order. How-
1250
+ ever, the above arguments lead to the same corners
1251
+ (g + s
1252
+ 2
1253
+ , g) and (s + g′
1254
+ 2
1255
+ , g) of ηk,i(t). If g = g′, the two
1256
+ corners collapse to one corner in the graph of ηk,i(t).
1257
+ The k-fold intersection ∩i+k−1
1258
+ j=i
1259
+ Lj(t) becomes fully
1260
+ covered when the intervals Li−1(t), Li+k(t). At this
1261
+ moment, 2t equals the distance from pi−1 + ri−1 (the
1262
+ right endpoint of Li−1) to pi+k − ri+k (the left end-
1263
+ point of Li+k), see Fig. 6, so 2t =
1264
+ i+k
1265
+
1266
+ j=i
1267
+ gj +2
1268
+ i+k−1
1269
+
1270
+ j=i
1271
+ rj =
1272
+ gi + 2ri + s + gi+k + 2ri+k−1 = g + s + g′. The graph
1273
+ of ηk,i(t) has the final corner
1274
+ �g + s + g′
1275
+ 2
1276
+ , 0
1277
+
1278
+ .
1279
+
1280
+ Example 5.3 applies Theorem 5.2 to get ψ2
1281
+ found for the periodic sequence S in Example 3.1.
1282
+ Example 5.3 (using Theorem 5.2 for ψ2). The
1283
+ sequence S =
1284
+
1285
+ 0, 1
1286
+ 3, 1
1287
+ 2
1288
+
1289
+ + Z in Example 4.1 with
1290
+ points p1 = 0, p2 = 1
1291
+ 3, p3 = 1
1292
+ 2 of radii r1 = 1
1293
+ 12,
1294
+ r2 = 0, r3 = 1
1295
+ 12, respectively, has the initial gaps
1296
+ g1 = 1
1297
+ 3, g2 = 1
1298
+ 4, g3 = 1
1299
+ 12, see Example 3.3.
1300
+ In Theorem 5.2, the 2nd density function
1301
+ ψ2[S](t) is expressed as a sum of the trapezoid
1302
+ functions computed via their corners below.
1303
+ Case (GB). For the function ηGB measuring the
1304
+ double intersections of the green and blue intervals
1305
+ centered at p2 = pi and p3 = pi+k−1, we set k = 2
1306
+ and i = 2. Then we have the radii ri = 0 and
1307
+ ri+1 = 1
1308
+ 12, the gaps gi = 1
1309
+ 4, gi+1 = 1
1310
+ 12, gi+2 = 1
1311
+ 3,
1312
+ and the sum s = gi+1 = 1
1313
+ 12. The pair
1314
+ {gi + 2ri, gi+2 + 2ri+1} =
1315
+ �1
1316
+ 4 + 0, 1
1317
+ 3 + 2
1318
+ 12
1319
+
1320
+ has the minimum value g = 1
1321
+ 4 and maximum value
1322
+ g′ = 1
1323
+ 2. Then η2,2[S](t) = ηGB has the following
1324
+ corners as expected in the top picture of Fig. 7:
1325
+ �s
1326
+ 2, 0
1327
+
1328
+ =
1329
+ � 1
1330
+ 24, 0
1331
+
1332
+ ,
1333
+ �g + s
1334
+ 2
1335
+ , g
1336
+
1337
+ =
1338
+ �1
1339
+ 2
1340
+ �1
1341
+ 4 + 1
1342
+ 12
1343
+
1344
+ , 1
1345
+ 4
1346
+
1347
+ =
1348
+ �1
1349
+ 6, 1
1350
+ 4
1351
+
1352
+ ,
1353
+ �s + g′
1354
+ 2
1355
+ , g
1356
+
1357
+ =
1358
+ �1
1359
+ 2
1360
+ � 1
1361
+ 12 + 1
1362
+ 2
1363
+
1364
+ , 1
1365
+ 4
1366
+
1367
+ =
1368
+ � 7
1369
+ 24, 1
1370
+ 4
1371
+
1372
+ ,
1373
+ �g + s + g′
1374
+ 2
1375
+ , 0
1376
+
1377
+ =
1378
+ �1
1379
+ 2(1
1380
+ 4 + 1
1381
+ 12 + 1
1382
+ 2), 0
1383
+
1384
+ =
1385
+ � 5
1386
+ 12, 0
1387
+
1388
+ .
1389
+ Case (BR). For the trapezoid function ηBR mea-
1390
+ suring the double intersections of the blue and red
1391
+ intervals centered at p3 = pi and p1 = pi+k−1,
1392
+ we set k = 2 and i = 3. Then we have the radii
1393
+ ri =
1394
+ 1
1395
+ 12 = ri+1, the gaps gi =
1396
+ 1
1397
+ 12, gi+1 = 1
1398
+ 3,
1399
+ gi+2 = 1
1400
+ 4, and s = gi+1 = 1
1401
+ 3. The pair
1402
+ {gi + 2ri, gi+2 + 2ri+1} =
1403
+ � 1
1404
+ 12 + 2
1405
+ 12, 1
1406
+ 4 + 2
1407
+ 12
1408
+
1409
+ has the minimum g = 1
1410
+ 4 and maximum g′ = 5
1411
+ 12.
1412
+ Then η2,3[S](t) = ηBR has the following corners
1413
+ as expected in the second picture of Fig. 7:
1414
+ �s
1415
+ 2, 0
1416
+
1417
+ =
1418
+ �1
1419
+ 6, 0
1420
+
1421
+ ,
1422
+ �g + s
1423
+ 2
1424
+ , g
1425
+
1426
+ =
1427
+ �1
1428
+ 2
1429
+ �1
1430
+ 4 + 1
1431
+ 3
1432
+
1433
+ , 1
1434
+ 4
1435
+
1436
+ =
1437
+ � 7
1438
+ 24, 1
1439
+ 4
1440
+
1441
+ ,
1442
+ �s + g′
1443
+ 2
1444
+ , g
1445
+
1446
+ =
1447
+ �1
1448
+ 2
1449
+ �1
1450
+ 3 + 5
1451
+ 12
1452
+
1453
+ , 1
1454
+ 4
1455
+
1456
+ =
1457
+ �3
1458
+ 8, 1
1459
+ 4
1460
+
1461
+ ,
1462
+ �g + s + g′
1463
+ 2
1464
+ , 0
1465
+
1466
+ =
1467
+ �1
1468
+ 2(1
1469
+ 4 + 1
1470
+ 3 + 5
1471
+ 12), 0
1472
+
1473
+ =
1474
+ �1
1475
+ 2, 0
1476
+
1477
+ .
1478
+ Case (RG). For the trapezoid function ηRG mea-
1479
+ suring the double intersections of the red and
1480
+ green intervals centered at p1 = pi and p2 =
1481
+ pi+k−1, we set k = 2 and i = 1. Then we have
1482
+ the radii ri = 1
1483
+ 12 and ri+1 = 0, the gaps gi = 1
1484
+ 3,
1485
+
1486
+ Springer Nature 2021 LATEX template
1487
+ 12
1488
+ Density functions of periodic sequences
1489
+ gi+1 = 1
1490
+ 4, gi+2 = 1
1491
+ 12, and s = gi+1 = 1
1492
+ 4. The pair
1493
+ {gi + 2ri, gi+2 + 2ri+1} =
1494
+ �1
1495
+ 3 + 2
1496
+ 12, 1
1497
+ 12 + 0
1498
+
1499
+ has the minimum g = 1
1500
+ 12 and maximum g′ = 1
1501
+ 2.
1502
+ Then η2,1[S](t) = ηRG has the following corners:
1503
+ �s
1504
+ 2, 0
1505
+
1506
+ =
1507
+ �1
1508
+ 8, 0
1509
+
1510
+ ,
1511
+ �g + s
1512
+ 2
1513
+ , g
1514
+
1515
+ =
1516
+ �1
1517
+ 2
1518
+ � 1
1519
+ 12 + 1
1520
+ 4
1521
+
1522
+ , 1
1523
+ 12
1524
+
1525
+ =
1526
+ �1
1527
+ 6, 1
1528
+ 12
1529
+
1530
+ ,
1531
+ �s + g′
1532
+ 2
1533
+ , g
1534
+
1535
+ =
1536
+ �1
1537
+ 2
1538
+ �1
1539
+ 4 + 1
1540
+ 2
1541
+
1542
+ , 1
1543
+ 12
1544
+
1545
+ =
1546
+ �3
1547
+ 8, 1
1548
+ 12
1549
+
1550
+ ,
1551
+ �g + s + g′
1552
+ 2
1553
+ , 0
1554
+
1555
+ =
1556
+ �1
1557
+ 2( 1
1558
+ 12 + 1
1559
+ 4 + 1
1560
+ 2), 0
1561
+
1562
+ =
1563
+ � 5
1564
+ 12, 0
1565
+
1566
+ .
1567
+ as expected in the third picture of Fig. 7.
1568
+
1569
+ 6 Properties of new densities
1570
+ This section proves the periodicity of the sequence
1571
+ ψk with respect to the index k ≥ 0 in Theorem 6.2,
1572
+ which was a bit unexpected from original Defini-
1573
+ tion 1.2. We start with the simpler example for
1574
+ the familiar 3-point sequence in Fig. 3.
1575
+ Example 6.1 (periodicity of ψk in the index k).
1576
+ Let the periodic sequence S =
1577
+
1578
+ 0, 1
1579
+ 3, 1
1580
+ 2
1581
+
1582
+ +Z have
1583
+ three points p1 = 0, p2 = 1
1584
+ 3, p3 = 1
1585
+ 2 of radii
1586
+ r1 = 1
1587
+ 12, r2 = 0, r3 = 1
1588
+ 12, respectively. The ini-
1589
+ tial intervals L1(0) = [− 1
1590
+ 12, 1
1591
+ 12], L2(0) = [ 1
1592
+ 3, 1
1593
+ 3],
1594
+ L3(0) = [ 5
1595
+ 12, 7
1596
+ 12] have the 0-fold intersection mea-
1597
+ sured by ψ0(0) = 2
1598
+ 3 and the 1-fold intersection
1599
+ measured by ψ1(0) = 1
1600
+ 3, see Fig. 4 and 5.
1601
+ By the time t = 1
1602
+ 2 the initial intervals will grow
1603
+ to L1( 1
1604
+ 2) = [− 7
1605
+ 12, 7
1606
+ 12], L2( 1
1607
+ 2) = [− 1
1608
+ 6, 5
1609
+ 6], L3( 1
1610
+ 2) =
1611
+ [− 1
1612
+ 12, 13
1613
+ 12]. The grown intervals at the radius t = 1
1614
+ 2
1615
+ have the 3-fold intersection [− 1
1616
+ 12, 7
1617
+ 12] of the length
1618
+ ψ3( 1
1619
+ 2) = 2
1620
+ 3, which coincides with ψ0(0) = 2
1621
+ 3.
1622
+ With the extra interval L4( 1
1623
+ 2) = [ 5
1624
+ 12, 19
1625
+ 12] cen-
1626
+ tered at p4 = 1, the 4-fold intersection is L1 ∩
1627
+ L2 ∩ L3 ∩ L4 = [ 5
1628
+ 12, 7
1629
+ 12]. With the extra inter-
1630
+ val L5( 1
1631
+ 2) = [ 5
1632
+ 6, 11
1633
+ 6 ] centered at p5 =
1634
+ 4
1635
+ 3, the
1636
+ 4-fold intersection L2 ∩ L3 ∩ L4 ∩ L5 is the single
1637
+ point 5
1638
+ 6. With the extra interval L6( 1
1639
+ 2) = [ 11
1640
+ 12, 13
1641
+ 12]
1642
+ centered at p6 = 3
1643
+ 2, the 4-fold intersection is
1644
+ L3∩L4∩L5∩L6 = [ 11
1645
+ 12, 13
1646
+ 12]. Hence the total length
1647
+ of the 4-fold intersection at t = 1
1648
+ 2 is ψ4( 1
1649
+ 2) = 1
1650
+ 3,
1651
+ which coincides with ψ1(0) = 1
1652
+ 3.
1653
+ For the larger t = 1, the six grown intervals
1654
+ L1(1) =
1655
+
1656
+ −13
1657
+ 12, 13
1658
+ 12
1659
+
1660
+ , L2(1) =
1661
+
1662
+ −2
1663
+ 3, 4
1664
+ 3
1665
+
1666
+ ,
1667
+ L3(1) =
1668
+
1669
+ − 7
1670
+ 12, 19
1671
+ 12
1672
+
1673
+ , L4(1) =
1674
+
1675
+ − 1
1676
+ 12, 25
1677
+ 12
1678
+
1679
+ ,
1680
+ L5(1) =
1681
+ �1
1682
+ 3, 7
1683
+ 3
1684
+
1685
+ ,
1686
+ L6(1) =
1687
+ � 5
1688
+ 12, 31
1689
+ 12
1690
+
1691
+ have the 6-fold intersection
1692
+ � 5
1693
+ 12, 13
1694
+ 12
1695
+
1696
+ of length
1697
+ ψ6(1) = 2
1698
+ 3 coinciding with ψ0(0) = ψ3( 1
1699
+ 2) = 2
1700
+ 3. ■
1701
+ Corollary 6.2 proves that the coincidences in
1702
+ Example 6.1 are not accidental. The periodicity of
1703
+ ψk with respect to k is illustrated by Fig. 8.
1704
+ Theorem 6.2 (periodicity of ψk in the index k).
1705
+ The density functions ψk[S] of a periodic sequence
1706
+ S = {p1, . . . , pm} + Z consist of disjoint intervals
1707
+ with centers 0 ≤ p1 < · · · < pm < 1 and radii
1708
+ r1, . . . , rm ≥ 0, respectively, satisfy the periodicity
1709
+ ψk+m(t + 1
1710
+ 2) = ψk(t) for any k ≥ 0 and t ≥ 0.
1711
+
1712
+ Proof Since the initial intervals are disjoint, for k ≥ 0,
1713
+ any (k +m)-fold intersection involves k +m successive
1714
+ intervals Li(t), . . . , Li+k+m−1(t) centered around the
1715
+ points of S. Then we can find an interval [x, x + 1]
1716
+ covering exactly m of these initial intervals of S.
1717
+ By collapsing [x, x+1] to the point x, any (k+m)-
1718
+ fold intersection of k + m intervals grown by a radius
1719
+ r ≥ 1
1720
+ 2 becomes a k-fold intersection of k intervals
1721
+
1722
+ Springer Nature 2021 LATEX template
1723
+ Density functions of periodic sequences
1724
+ 13
1725
+ Fig. 8 The densities ψk, k = 0, . . . , 9 for the 1-period sequence S whose points 0, 1
1726
+ 3 , 1
1727
+ 2 have radii 1
1728
+ 12 , 0, 1
1729
+ 12 , respectively. The
1730
+ densities ψ0, ψ1, ψ2 are described in Examples 3.1, 4.1, 5.1 and determine all other densities by periodicity in Theorem 6.2.
1731
+ grown by t = r− 1
1732
+ 2. Both k-fold and (k+m)-fold inter-
1733
+ sections within any unit cell have the same fractional
1734
+ length, so ψk+m(t + 1
1735
+ 2) = ψk(t) for any t ≥ 0.
1736
+
1737
+ The symmetry ψm−k( 1
1738
+ 2 − t) = ψk(t) for k =
1739
+ 0, . . . , [ m
1740
+ 2 ], and t ∈ [0, 1
1741
+ 2] from [3, Theorem 8]
1742
+ no longer holds for points with different radii.
1743
+ For example, ψ1(t) ̸= ψ2( 1
1744
+ 2 − t) for the periodic
1745
+ sequence S =
1746
+
1747
+ 0, 1
1748
+ 3, 1
1749
+ 2
1750
+
1751
+ + Z, see Fig. 5, 7. If
1752
+ all points have the same radius r, [3, Theorem 8]
1753
+ implies the symmetry after replacing t by t + 2r.
1754
+ The main results of [3] implied that all den-
1755
+ sity functions cannot distinguish the non-isometric
1756
+ sequences S15 = {0, 1, 3, 4, 5, 7, 9, 10, 12} + 15Z
1757
+ and Q15 = {0, 1, 3, 4, 6, 8, 9, 12, 14}+15Z of points
1758
+ with zero radii. Example 6.3 shows that the den-
1759
+ sities for sequences with non-zero radii are strictly
1760
+ stronger and distinguish the sequences S15 ̸∼= Q15.
1761
+ Example 6.3 (ψk for S15, Q15 with neighbor
1762
+ radii). For any point p in a periodic sequence S ⊂
1763
+ R, define its neighbor radius as the half-distance
1764
+ to a closest neighbor of p within the sequence S.
1765
+ This choice of radii respects the isometry in the
1766
+ sense that periodic sequences S, Q with zero-sized
1767
+ radii are isometric if and only if S, Q with neighbor
1768
+ radii are isometric. Fig. 9 shows that the densi-
1769
+ ties ψk for k ≥ 2 distinguish the non-isometric
1770
+ sequences S15 and Q15 scaled down by factor 15
1771
+ to the unit cell [0, 1], see Example 2.1.
1772
+
1773
+ Corollary
1774
+ 6.4
1775
+ (computation
1776
+ of
1777
+ ψk(t)). Let
1778
+ S, Q ⊂ R be periodic sequences with at most m
1779
+ motif points. For k ≥ 1, one can draw the graph
1780
+ of the k-th density function ψk[S] in time O(m2).
1781
+ One can check in time O(m3) if Ψ[S] = Ψ[Q].
1782
+
1783
+ Proof To draw the graph of ψk[S] or evaluate the k-
1784
+ th density function ψk[S](t) at any radius t, we first
1785
+ use the periodicity from Theorem 6.2 to reduce k to
1786
+ the range 0, 1, . . . , m. In time O(m log m) we put the
1787
+ points from a unit cell U (scaled to [0, 1] for conve-
1788
+ nience) in the increasing (cyclic) order p1, . . . , pm. In
1789
+ time O(m) we compute the gaps gi = (pi−ri)−(pi−1+
1790
+ ri−1) between successive intervals.
1791
+ For k = 0, we put the gaps in the increasing order
1792
+ g[1] ≤ · · · ≤ g[m] in time O(m log m). By Theorem 3.2
1793
+ in time O(m2), we write down the O(m) corner points
1794
+ whose horizontal coordinates are the critical radii
1795
+ where ψ0(t) can change its gradient.
1796
+ We evaluate ψ0 at every critical radius t by sum-
1797
+ ming up the values of m trapezoid functions at t, which
1798
+ needs O(m2) time. It remains to plot the points at all
1799
+
1800
+ 0.8
1801
+ psi 0
1802
+ 0.6
1803
+ psi_1
1804
+ psi_2
1805
+ psi_3
1806
+ psi_4
1807
+ s
1808
+ 0.4 -
1809
+ p
1810
+ psi_5
1811
+ psi_6
1812
+ psi_7
1813
+ 0.2
1814
+ psi_8
1815
+ psi_9
1816
+ 0.0
1817
+ 0.0
1818
+ 0.5
1819
+ 1.0
1820
+ 1.5
1821
+ TSpringer Nature 2021 LATEX template
1822
+ 14
1823
+ Density functions of periodic sequences
1824
+ Fig. 9 The densities ψk, k = 0, . . . , 10, distinguish (already for k ≥ 2) the sequences (scaled down by period 15) S15 =
1825
+ {0, 1, 3, 4, 5, 7, 9, 10, 12} + 15Z (top) and Q15 = {0, 1, 3, 4, 6, 8, 9, 12, 14} + 15Z (bottom), where the radius ri of any point
1826
+ is the half-distance to its closest neighbor. These sequences with zero radii have identical ψk for all k, see [3, Example 10].
1827
+ O(m) critical radii t and connect the successive points
1828
+ by straight lines, so the total time is O(m2).
1829
+ For any larger fixed index k = 1, . . . , m, in time
1830
+ O(m2) we write down all O(m) corner points from
1831
+ Theorems 4.2 and 5.2, which leads to the graph of
1832
+ ψk(t) similarly to the above argument for k = 0.
1833
+ To decide if the infinite sequences of density func-
1834
+ tions coincide: Ψ[S] = Ψ[Q], by Theorem 6.2 it suffices
1835
+ to check only if O(m) density functions coincide:
1836
+ ψk[S](t) = ψk[Q](t) for k = 0, 1, . . . , [ m
1837
+ 2 ].
1838
+ To check if two piecewise linear functions coincide,
1839
+ it remains to compare their values at all O(m) critical
1840
+ radii t from the corner points in Theorems 3.2, 4.2, 5.2.
1841
+ Since these values were found in time O(m2) above,
1842
+ the total time for k = 0, 1, . . . , [ m
1843
+ 2 ] is O(m3).
1844
+
1845
+
1846
+ 0.75
1847
+ psi_0
1848
+ psi_1
1849
+ psi_2
1850
+ psi_3
1851
+ 0.50
1852
+ K
1853
+ psi_4
1854
+ psi
1855
+ psi_5
1856
+ psi_6
1857
+ psi_7
1858
+ 0.25
1859
+ psi_8
1860
+ psi_ 9
1861
+ psi_10
1862
+ 0.00
1863
+ 0.0
1864
+ 0.2
1865
+ 0.4
1866
+ 0.6
1867
+ T0.75
1868
+ psi_ 0
1869
+ psi_1
1870
+ psi_2
1871
+ psi_3
1872
+ 0.50
1873
+ K
1874
+ psi_4
1875
+ psi_5
1876
+ psi_6
1877
+ psi_7
1878
+ 0.25
1879
+ psi_8
1880
+ psi_ 9
1881
+ psi_10
1882
+ 0.00
1883
+ 0.0
1884
+ 0.2
1885
+ 0.4
1886
+ 0.6
1887
+ TSpringer Nature 2021 LATEX template
1888
+ Density functions of periodic sequences
1889
+ 15
1890
+ All previous examples show densities with a
1891
+ single local maximum. However, the new R code
1892
+ [5] helped us discover the opposite examples.
1893
+ Fig. 10 For the periodic sequence S =
1894
+
1895
+ 0, 1
1896
+ 8 , 1
1897
+ 4 , 3
1898
+ 4
1899
+
1900
+ + Z
1901
+ whose all points have radii 0, the 2nd density ψ2[S](t) has
1902
+ the local minimum at t = 1
1903
+ 4 between two local maxima.
1904
+ Example 6.5 (densities with multiple maxima).
1905
+ Fig. 10 shows a simple 4-point sequence S whose
1906
+ 2nd density ψ2[S] has two local maxima. Fig. 11
1907
+ and 12 show more complicated sequences whose
1908
+ density functions have more than two maxima. ■
1909
+ Fig. 11 For the sequence S =
1910
+
1911
+ 0, 1
1912
+ 81 , 1
1913
+ 27 , 1
1914
+ 9 , 1
1915
+ 3
1916
+
1917
+ +Z whose
1918
+ all points have radii 0, ψ2[S] equal to the sum of the shown
1919
+ five trapezoid functions has three maxima.
1920
+ Fig. 12 For the sequence S =
1921
+
1922
+ 0, 1
1923
+ 64 , 1
1924
+ 16 , 1
1925
+ 8 , 1
1926
+ 4 , 3
1927
+ 4
1928
+
1929
+ + Z
1930
+ whose all points have radii 0, ψ3[S] has 5 local maxima.
1931
+ 7 Conclusions and future work
1932
+ In comparison with the past work [3], the key
1933
+ contributions of this paper are the following.
1934
+ • Definition 1.2 extends density functions ψk to
1935
+ any periodic sets of points with radii ri ≥ 0.
1936
+ • Theorems 3.2, 4.2, 5.2 explicitly describe all ψk
1937
+ for any periodic sequence S of points with radii.
1938
+ • The descriptions of ψk allowed us to justify the
1939
+ periodicity of ψk in Theorem 6.2 and a quadratic
1940
+ algorithm computing any ψk in Corollary 6.4.
1941
+ • The code [5] helped us distinguish S15 ̸∼= Q15 in
1942
+ Example 6.3 and find sequences whose densities
1943
+ have multiple local maxima in Example 6.5.
1944
+ Here are the open problems for future work.
1945
+ • Verify if density functions ψk[S](t) for small
1946
+ values of k distinguish all non-isometric periodic
1947
+ point sets S ⊂ Rn at least with radii 0.
1948
+ • Characterize the periodic sequences S ⊂ R
1949
+ whose all density functions ψk for k ≥ 1 have a
1950
+ unique local maximum, not as in Example 6.5.
1951
+ • Similar to Theorems 3.2, 4.2, 5.2, analytically
1952
+ describe the density function ψk[S] for periodic
1953
+ point sets S ⊂ Rn in higher dimensions n > 1.
1954
+ This research was supported by the grants of
1955
+ the UK Engineering Physical Sciences Research
1956
+
1957
+ 1.00
1958
+ 0.75
1959
+ psi_0
1960
+ K
1961
+ 0.50
1962
+ psi_1
1963
+ Isd
1964
+ psi_2
1965
+ psi_3
1966
+ 0.25
1967
+ 0.00
1968
+ 0.0
1969
+ 0.1
1970
+ 0.2
1971
+ 0.3
1972
+ 0.4
1973
+ 0.5
1974
+ t0.12
1975
+ -
1976
+ 0.08
1977
+ eta_2_1
1978
+ eta_2_2
1979
+ eta_2_3
1980
+ .2
1981
+ eta_2_4
1982
+ eta_2_5
1983
+ psi_2
1984
+ 0.04 -
1985
+ 0.00 -
1986
+ 0.0
1987
+ 0.1
1988
+ 0.2
1989
+ 0.3
1990
+ 0.4
1991
+ 0.5
1992
+ t0.09-
1993
+ eta_3_1
1994
+ eta_3_2
1995
+ 3
1996
+ 0.06
1997
+ eta_3_3
1998
+ eta
1999
+ eta_3_4
2000
+ 3
2001
+ eta_3_5
2002
+ eta_3_6
2003
+ eta_3_7
2004
+ psi_3
2005
+ 0.03 -
2006
+ 0.00-
2007
+ 0.0
2008
+ 0.1
2009
+ 0.2
2010
+ 0.3
2011
+ 0.4
2012
+ tSpringer Nature 2021 LATEX template
2013
+ 16
2014
+ Density functions of periodic sequences
2015
+ Council (EP/R018472/1, EP/X018474/1) and the
2016
+ Royal Academy of Engineering Industrial Fellow-
2017
+ ship (IF2122/186) of the last author. We thank all
2018
+ reviewers for their time and helpful advice.
2019
+ References
2020
+ [1] Anosova,
2021
+ O.,
2022
+ Kurlin,
2023
+ V.:
2024
+ Introduction
2025
+ to
2026
+ periodic
2027
+ geometry
2028
+ and
2029
+ topology.
2030
+ arxiv:2103.02749 (2021)
2031
+ [2] Anosova, O., Kurlin, V.: An isometry clas-
2032
+ sification of periodic point sets. In: Lecture
2033
+ Notes in Computer Science (Proceedings of
2034
+ DGMM). vol. 12708, pp. 229–241 (2021)
2035
+ [3] Anosova, O., Kurlin, V.: Density functions of
2036
+ periodic sequences. In: Lecture Notes in Com-
2037
+ puter Science (Proceedings of DGMM). vol.
2038
+ 13493, pp. 395–408 (2022)
2039
+ [4] Anosova, O., Kurlin, V.: Recognition of near-
2040
+ duplicate periodic patterns in polynomial
2041
+ time. arxiv:2205.15298 (2022)
2042
+ [5] Anosova, O.: R code for density functions
2043
+ of periodic sequences (2023), https://github.
2044
+ com/oanosova/DensityFunctions1D
2045
+ [6] Bright, M., Cooper, A.I., Kurlin, V.: Wel-
2046
+ come to a continuous world of 3-dimensional
2047
+ lattices. arxiv:2109.11538 (2021)
2048
+ [7] Bright,
2049
+ M.,
2050
+ Cooper,
2051
+ A.I.,
2052
+ Kurlin,
2053
+ V.:
2054
+ Geographic-style maps for 2-dimensional lat-
2055
+ tices.
2056
+ Acta
2057
+ Crystallographica
2058
+ Section
2059
+ A
2060
+ 79(1), 1–13 (2023)
2061
+ [8] Edelsbrunner, H., Heiss, T., Kurlin, V.,
2062
+ Smith, P., Wintraecken, M.: The density fin-
2063
+ gerprint of a periodic point set. In: SoCG.
2064
+ vol. 189, pp. 32:1–32:16 (2021)
2065
+ [9] Gr¨unbaum, F., Moore, C.: The use of higher-
2066
+ order invariants in the determination of
2067
+ generalized patterson cyclotomic sets. Acta
2068
+ Cryst. A 51, 310–323 (1995)
2069
+ [10] Kurlin, V.: A complete isometry classification
2070
+ of 3D lattices. arxiv:2201.10543 (2022)
2071
+ [11] Kurlin, V.: Computable complete invari-
2072
+ ants for finite clouds of unlabeled points.
2073
+ arxiv:2207.08502 (2022), http://kurlin.org/
2074
+ projects/complete-isometry-invariants.pdf
2075
+ [12] Kurlin, V.: Exactly computable and contin-
2076
+ uous metrics on isometry classes of finite
2077
+ and 1-periodic sequences. arXiv:2205.04388
2078
+ (2022),
2079
+ http://kurlin.org/projects/
2080
+ periodic-geometry-topology/metric1D.pdf
2081
+ [13] Kurlin, V.: Mathematics of 2-dimensional lat-
2082
+ tices. Foundations of Computational Math-
2083
+ ematics (2022), http://kurlin.org/projects/
2084
+ lattice-geometry/lattices2Dmaths.pdf
2085
+ [14] Mosca, M., Kurlin, V.: Voronoi-based sim-
2086
+ ilarity distances between arbitrary crystal
2087
+ lattices. Crystal Research and Technology
2088
+ 55(5), 1900197 (2020)
2089
+ [15] Pozdnyakov, S., et al.: Incompleteness of
2090
+ atomic structure representations. Phys. Rev.
2091
+ Let. 125, 166001 (2020)
2092
+ [16] Smith, P., Kurlin, V.: A practical algo-
2093
+ rithm for degree-k Voronoi domains of three-
2094
+ dimensional periodic point sets. In: Lecture
2095
+ Notes in Computer Science (Proceedings of
2096
+ ISVC). vol. 13599, pp. 377–391 (2022)
2097
+ [17] Smith, P., Kurlin, V.: Families of point
2098
+ sets
2099
+ with
2100
+ identical
2101
+ 1D
2102
+ persistence,.
2103
+ arxiv:2202.00577
2104
+ (2022),
2105
+ http://kurlin.
2106
+ org/projects/periodic-geometry-topology/
2107
+ trivial-persistence.pdf
2108
+ [18] Widdowson,
2109
+ D.,
2110
+ Kurlin,
2111
+ V.:
2112
+ Pointwise
2113
+ distance
2114
+ distributions
2115
+ of
2116
+ periodic
2117
+ sets.
2118
+ arXiv:2108.04798 (version 1) (2021)
2119
+ [19] Widdowson,
2120
+ D.,
2121
+ Kurlin,
2122
+ V.:
2123
+ Resolving
2124
+ the data ambiguity for periodic crystals.
2125
+ Advances in Neural Information Process-
2126
+ ing
2127
+ Systems
2128
+ (arXiv:2108.04798,
2129
+ v2)
2130
+ 35
2131
+ (2022), http://kurlin.org/projects/periodic+
2132
+ geometry/NeurIPS2022PDD.pdf
2133
+ [20] Widdowson, D., et al.: Average minimum dis-
2134
+ tances of periodic point sets. MATCH Comm.
2135
+ Math. Comp. Chemistry 87, 529–559 (2022)
2136
+
MNE4T4oBgHgl3EQfiw29/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
MdE1T4oBgHgl3EQftQUl/content/2301.03374v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bc0c51773605073b29f17ccb50987dfb951d1560bcc4e9776b3c30a6806a4a5c
3
+ size 8692727
OtAyT4oBgHgl3EQfUfen/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac37ec030752deb69937cdae794be524b673730286f54de13f8e5fb734a9a796
3
+ size 3276845
OtAyT4oBgHgl3EQfUfen/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5e80552674d606850c456ed200b9a6f3e9357bda630ffa7365fa2057a43939d1
3
+ size 120390
PdE0T4oBgHgl3EQfkAEs/content/2301.02466v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:de288ecf10fb888fdb10618959b96999da403d8696e47ec0725b2fb388a38307
3
+ size 3971536
PdE0T4oBgHgl3EQfkAEs/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4266c243207217c1333dc9bfa1ad90865ee43030a73d7c4ca37d66294e855026
3
+ size 139122
PtAyT4oBgHgl3EQf7fqt/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05f9fdfb61d46533f74d817bdd89532d48cc80026270358847cd434095918ccb
3
+ size 927998
QNFJT4oBgHgl3EQf2i1I/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fc34add24b7d56a8a21acae64aa0eea37d346480b1f6e1f3c457fe8a148843ef
3
+ size 188787
QNFRT4oBgHgl3EQfJje8/content/tmp_files/2301.13496v1.pdf.txt ADDED
@@ -0,0 +1,1447 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.13496v1 [math.AP] 31 Jan 2023
2
+ Conditional regularity for the Navier–Stokes–Fourier system
3
+ with Dirichlet boundary conditions
4
+ Danica Basari´c ∗
5
+ Eduard Feireisl ∗
6
+ Hana Mizerov´a ∗,†
7
+ ∗ Institute of Mathematics of the Czech Academy of Sciences
8
+ ˇZitn´a 25, CZ-115 67 Praha 1, Czech Republic
9
+ † Department of Mathematical Analysis and Numerical Mathematics, Comenius University
10
+ Mlynsk´a dolina, 842 48 Bratislava, Slovakia
11
+ Abstract
12
+ We consider the Navier–Stokes–Fourier system with the inhomogeneous boundary condi-
13
+ tions for the velocity and the temperature. We show that solutions emanating from sufficiently
14
+ regular data remain regular as long as the density ̺, the absolute temperature ϑ, and the
15
+ modulus of the fluid velocity |u| remain bounded.
16
+ Keywords: Navier–Stokes–Fourier system, conditional regularity, blow–up criterion, regular
17
+ solution
18
+ 1
19
+ Introduction
20
+ Standard systems of equations in fluid mechanics including the Navier–Stokes–Fourier system
21
+ governing the motion of a compressible, viscous, and heat conducting fluid are well posed in the
22
+ class of strong solutions on a possibly short time interval [0, Tmax). The recent results of Merle at
23
+ al. [16], [17] strongly indicate that Tmax may be finite, at least in the idealized case of “isentropic”
24
+ viscous flow. Conditional regularity results guarantee that a blow up will not occur as soon as
25
+ some lower order norms of solutions are controlled.
26
+ We consider the Navier–Stokes–Fourier system governing the time evolution of the mass density
27
+ ̺ = ̺(t, x), the (absolute) temperature ϑ = ϑ(t, x), and the velocity u = u(t, x) of a compressible,
28
+ viscous, and heat conducting fluid:
29
+ ∗The work of D.B., E.F., and H.M. was supported by the Czech Sciences Foundation (GAˇCR), Grant Agreement
30
+ 21–02411S. The Institute of Mathematics of the Czech Academy of Sciences is supported by RVO:67985840.
31
+ 1
32
+
33
+ ∂t̺ + divx(̺u) = 0,
34
+ (1.1)
35
+ ∂t(̺u) + divx(̺u ⊗ u) + ∇xp(̺, ϑ) = divxS(Dxu) + ̺f, Dxu = 1
36
+ 2
37
+
38
+ ∇xu + ∇t
39
+ xu
40
+
41
+ ,
42
+ (1.2)
43
+ ∂t(̺e(̺, ϑ)) + divx(̺e(̺, ϑ)u) + divxq(∇xϑ) = S(Dxu) : Dxu − p(̺, ϑ)divxu.
44
+ (1.3)
45
+ The fluid is Newtonian, the viscous stress S is given by Newton’s rheological law
46
+ S(Dxu) = 2µ
47
+
48
+ Dxu − 1
49
+ 3divxuI
50
+
51
+ + ηdivxuI, µ > 0, η ≥ 0.
52
+ (1.4)
53
+ The heat flux obeys Fourier’s law
54
+ q(∇xϑ) = −κ∇xϑ, κ > 0.
55
+ (1.5)
56
+ The equation of state for the pressure p and the internal energy e is given by the standard Boyle–
57
+ Mariotte law of perfect gas,
58
+ p(̺, ϑ) = ̺ϑ, e(̺, ϑ) = cvϑ, cv > 0.
59
+ (1.6)
60
+ For the sake of simplicity, we suppose that the viscosity coefficients µ, η, the heat conductivity
61
+ coefficient κ as well as the specific heat at constant volume cv are constant.
62
+ There is a large number of recent results concerning conditional regularity for the Navier–
63
+ Stokes–Fourier system in terms of various norms. Fan, Jiang, and Ou [4] consider a bounded fluid
64
+ domain Ω ⊂ R3 with the conservative boundary conditions
65
+ u|∂Ω = 0, ∇xϑ · n|∂Ω = 0.
66
+ (1.7)
67
+ The same problem is studied by Sun, Wang, and Zhang [19] and later by Huang, Li, Wang [14].
68
+ There are results for the Cauchy problem Ω = R3 by Huang and Li [13], and Jiu, Wang and Ye
69
+ [15]. Possibly the best result so far has been established in [11], where the blow up criterion for
70
+ both the Cauchy problem and the boundary value problem (1.7) is formulated in terms of the
71
+ maximum of the density and a Serrin type regularity for the temperature:
72
+ lim sup
73
+ t→Tmax−
74
+
75
+ ∥̺(t, ·)∥L∞ + ∥ϑ − ϑ∞∥Ls(0,t)(Lr)
76
+
77
+ = ∞, 3
78
+ 2 < r ≤ ∞, 1 ≤ s ≤ ∞, 2
79
+ s + 3
80
+ r ≤ 2,
81
+ where ϑ∞ denotes the far field temperature in the Cauchy problem, cf. also the previous results
82
+ by Wen and Zhu [23], [24].
83
+ Much less is known in the case of the Dirichlet boundary conditions
84
+ u|∂Ω = uB, ϑ|∂Ω = ϑB.
85
+ (1.8)
86
+ 2
87
+
88
+ Fan, Zhi, and Zhang [5] showed that a strong solution of the Navier–Stokes–Fourier system remains
89
+ regular up to a time T > 0 if (i) Ω ⊂ R2 is a bounded domain, (ii) uB = 0, ϑB = 0, and (iii)
90
+ lim sup
91
+ t→T−
92
+ (∥̺∥L∞ + ∥ϑ∥L∞) < ∞.
93
+ (1.9)
94
+ All results mentioned above describe fluids in a conservative regime, meaning solutions are
95
+ close to equilibrium in the long run. However, many real world applications concern fluids out of
96
+ equilibrium driven by possibly large driving forces f and/or inhomogeneous boundary conditions.
97
+ The iconic examples are the Rayleigh–B´enard and Taylor–Couette flows where the fluid is driven
98
+ to a turbulent regime by a large temperature gradient and large boundary velocity, respectively,
99
+ see Davidson [3].
100
+ Motivated by these physically relevant examples, we consider a fluid confined to a bounded
101
+ domain Ω ⊂ R3 with impermeable boundary, where the temperature and the (tangential) velocity
102
+ are given on ∂Ω,
103
+ ϑ|∂Ω = ϑB, ϑB = ϑB(x), ϑB > 0 on ∂Ω,
104
+ (1.10)
105
+ u|∂Ω = uB, uB = uB(x), uB · n = 0 on ∂Ω.
106
+ (1.11)
107
+ The initial state of the fluid is prescribed:
108
+ ̺(0, ·) = ̺0, ̺0 > 0 in Ω, ϑ(0, ·) = ϑ0, ϑ0 > 0 in Ω, u(0, ·) = u0.
109
+ (1.12)
110
+ The initial and boundary data are supposed to satisfy suitable compatibility conditions specified
111
+ below.
112
+ The existence of local in time strong solutions for the problem (1.1)–(1.6), endowed with the
113
+ inhomogeneous boundary conditions (1.10), (1.11) was established by Valli [20], [21] , see also Valli
114
+ and Zajaczkowski [22]. The solution exists on a maximal time interval [0, Tmax), Tmax > 0. Our
115
+ goal is to show that if Tmax < ∞, then necessarily
116
+ lim sup
117
+ t→Tmax−
118
+
119
+ ∥̺(t, ·)∥L∞(Ω) + ∥ϑ(t, ·)∥L∞(Ω) + ∥u(t, ·)∥L∞(Ω;R3)
120
+
121
+ = ∞.
122
+ (1.13)
123
+ The proof is based on deriving suitable a priori bounds assuming boundedness of all norms involved
124
+ in (1.13) as well as the norm of the initial/boundary data in a suitable function space. Although
125
+ approach shares some similarity with Fang, Zi, and Zhang [5], essential modifications must be
126
+ made to accommodate the inhomogeneous boundary data as well as the driving force f.
127
+ The
128
+ importance of conditional regularity results in numerical analysis of flows with uncertain initial
129
+ data was discussed recently in [7].
130
+ 3
131
+
132
+ The paper is organized as follows. In Section 2, we introduce the class of strong solutions to the
133
+ Navier–Stokes–Fourier system and state our main result concerning conditional regularity. The
134
+ remaining part of the paper is devoted to the proof of the main result – deriving suitable a priori
135
+ bounds. In Section 3 we recall the standard energy estimates that hold even in the class of weak
136
+ solutions. Section 4 is the heart of the paper. We establish the necessary estimates on the velocity
137
+ gradient by means of the celebrated Gagliardo–Nirenberg interpolation inequality. In Section 5,
138
+ higher order estimates on the velocity gradient are derived, and, finally, the estimates are closed
139
+ by proving bounds on the temperature time derivative in Section 6. This last part borrows the
140
+ main ideas from [9].
141
+ 2
142
+ Strong solutions, main result
143
+ We start the analysis by recalling the concept of strong solution introduced by Valli [21]. Similarly
144
+ to the boundary data uB, ϑB we suppose that the driving force f = f(x) is independent of time,
145
+ meaning we deal with an autonomous problem. Following [21], we suppose that Ω ⊂ R3 is a
146
+ bounded domain with ∂Ω of class C4.
147
+ We assume the data belong to the following class:
148
+ ̺0 ∈ W 3,2(Ω), 0 < ̺0 ≤ min
149
+ x∈Ω ̺0(x),
150
+ ϑ0 ∈ W 3,2(Ω), 0 < ϑ0 ≤ min
151
+ x∈Ω ϑ0(x),
152
+ u0 ∈ W 3,2(Ω; R3),
153
+ ϑB ∈ W
154
+ 7
155
+ 2(∂Ω), 0 < ϑB ≤ min
156
+ x∈∂Ω ϑB(x),
157
+ uB ∈ W
158
+ 7
159
+ 2(∂Ω; R3), uB · n = 0,
160
+ f ∈ W 2,2(Ω; R3).
161
+ (2.1)
162
+ In addition, the data must satisfy the compatibility conditions
163
+ ϑ0 = ϑB, u0 = uB on ∂Ω,
164
+ ̺0u0 · ∇xu0 + ∇xp(̺0, ϑ0) = divxS(Dxu0) + ̺0f on ∂Ω,
165
+ ̺0u0 · ∇xϑ0 + divxq(ϑ0) = S(Dxu0) : Dxu0 − p(̺0, ϑ0)divxu0 on ∂Ω.
166
+ (2.2)
167
+ We set
168
+ D0 = max
169
+
170
+ ∥(̺0, ϑ0, u0)∥W 3,2(Ω;R5), 1
171
+ ̺0
172
+ ,
173
+ 1
174
+ ϑ0
175
+ , 1
176
+ ϑB
177
+ , ∥ϑB∥W
178
+ 7
179
+ 2 (∂Ω), ∥uB∥W
180
+ 7
181
+ 2 (∂Ω;R3), ∥f∥W 2,2(Ω;R3)
182
+
183
+ .
184
+ (2.3)
185
+ 4
186
+
187
+ 2.1
188
+ Local existence
189
+ The following result was proved by Valli [21, Theorem A] (see also [20]).
190
+ Theorem 2.1. (Local existence of strong solutions) Let Ω ⊂ R3 be a bounded domain of
191
+ class C4. Suppose that the data (̺0, ϑ0, u0), (ϑB, uB) and f belong to the class (2.1) and satisfy
192
+ the compatibility conditions (2.2).
193
+ Then there exists a maximal time Tmax > 0 such that the Navier–Stokes–Fourier system (1.1)–
194
+ (1.6), with the boundary conditions (1.10), (1.11), and the initial conditions (1.12) admits a solu-
195
+ tion (̺, ϑ, u) in [0, Tmax) × Ω unique in the class
196
+ ̺, ϑ ∈ C([0, T]; W 3,2(Ω)), u ∈ C([0, T]; W 3,2(Ω; R3)),
197
+ ϑ ∈ L2(0, T; W 4,2(Ω)), u ∈ L2(0, T; W 4,2(Ω; R3))
198
+ (2.4)
199
+ for any 0 < T < Tmax. The existence time Tmax is bounded below by a quantity c(D0) depending
200
+ solely on the norms of the data specified in (2.3). In particular,
201
+ lim
202
+ τ→Tmax− ∥(̺, ϑ, u)(τ, ·)∥W 3,2(Ω;R5) = ∞.
203
+ (2.5)
204
+ 2.2
205
+ Blow up criterion, conditional regularity
206
+ Our goal is to show the following result.
207
+ Theorem 2.2. (Blow up criterion) Under the hypotheses of Theorem 2.1, suppose that the
208
+ maximal existence time Tmax < ∞ is finite.
209
+ Then
210
+ lim sup
211
+ τ→Tmax−
212
+ ∥(̺, ϑ, u)(τ, ·)∥L∞(Ω;R5) = ∞.
213
+ (2.6)
214
+ Theorem 2.2 is in the spirit of the blow up criteria for general parabolic systems – the solution
215
+ remains regular as long as it is bounded. Of course, our problem in question is of mixed hyperbolic–
216
+ parabolic type.
217
+ The proof of Theorem 2.2 follows from suitable a priori bounds applied on a compact time
218
+ interval.
219
+ Proposition 2.3. (Conditional regularity)
220
+ Under the hypotheses of Theorem 2.1, let (̺, ϑ, u) be the strong solution of the Navier–Stokes–
221
+ Fourier system belonging to the class (2.4) and satisfying
222
+ sup
223
+ (τ,x)∈[0,T)×Ω
224
+ ̺(τ, x) ≤ ̺,
225
+ sup
226
+ (τ,x)∈[0,T)×Ω
227
+ ϑ(τ, x) ≤ ϑ,
228
+ sup
229
+ (τ,x)∈[0,T)×Ω
230
+ |u(τ, x)| ≤ u
231
+ (2.7)
232
+ 5
233
+
234
+ for some T < Tmax.
235
+ Then there is a quantity c(T, D0, ̺, ϑ, u), bounded for bounded arguments, such that
236
+ sup
237
+ τ∈[0,T)
238
+ max
239
+
240
+ ∥(̺, ϑ, u)(τ, ·)∥W 3,2(Ω;R5); sup
241
+ x∈Ω
242
+ 1
243
+ ̺(τ, x); sup
244
+ x∈Ω
245
+ 1
246
+ ϑ(τ, x)
247
+
248
+ ≤ c(T, D0, ̺, ϑ, u).
249
+ (2.8)
250
+ In view of Theorem 2.1, the conclusion of Theorem 2.2 follows from Proposition 2.3. The rest
251
+ of the paper is therefore devoted to the proof of Proposition 2.3.
252
+ Remark 2.4. As observed in [8], the conditional regularity results established in Proposition 2.3
253
+ gives rise to stability with respect to the data. More specifically, the maximal existence time Tmax
254
+ is a lower semicontinuous function of the data with respect to the topologies in (2.1).
255
+ Remark 2.5. Conditional regularity results in combination with the weak–strong uniqueness
256
+ principle in the class of measure–valued solutions is an efficient tool for proving convergence of
257
+ numerical schemes, see [6, Chapter 11]. The concept of measure–valued solutions to the Navier–
258
+ Stokes–Fourier system with inhomogeneous Dirichlet boundary conditions has been introduced
259
+ recently by Chaudhuri [1].
260
+ 3
261
+ Energy estimates
262
+ To begin, it is suitable to extend the boundary data into Ω. For definiteness, we consider the
263
+ (unique) solutions of the Dirichlet problem
264
+ ∆x ˜ϑ = 0 in Ω, ˜ϑ|∂Ω = ϑB,
265
+ divxS(Dx˜u) = 0 in Ω, ˜u|∂Ω = uB.
266
+ (3.1)
267
+ By abuse of notation, we use the same symbol ϑB, uB for both the boundary values and their C1
268
+ extensions ˜ϑ = ˜ϑ(x), ˜u = ˜u(x) inside Ω.
269
+ We start with the ballistic energy equality, see [2, Section 2.4],
270
+ d
271
+ dt
272
+
273
+
274
+ �1
275
+ 2̺|u − uB|2 + ̺e − ϑB̺s
276
+
277
+ dx +
278
+
279
+
280
+ ϑB
281
+ ϑ
282
+
283
+ S(Dxu) : Dxu + κ|∇xϑ|2
284
+ ϑ
285
+
286
+ dx
287
+ = −
288
+
289
+
290
+
291
+ ̺u ⊗ u + pI − S(Dxu)
292
+
293
+ : DxuB dx + 1
294
+ 2
295
+
296
+
297
+ ̺u · ∇x|uB|2 dx
298
+ +
299
+
300
+
301
+ ̺(u − uB) · f dx −
302
+
303
+
304
+ ̺su · ∇xϑB dx + κ
305
+
306
+
307
+ ∇xϑ
308
+ ϑ
309
+ · ∇xϑB dx,
310
+ (3.2)
311
+ where we have introduced the entropy
312
+ s = cv log(ϑ) − log(̺).
313
+ 6
314
+
315
+ Thus the choice (3.1) yields the following bounds
316
+ sup
317
+ t∈[0,T)
318
+
319
+
320
+ ̺| log(ϑ)|(t, ·) dx ≤ c(T, D0, ̺, ϑ, u),
321
+ (3.3)
322
+ � T
323
+ 0
324
+
325
+
326
+ |∇xu|2 dx dt ≤ C(̺, ϑ, u; data) ⇒
327
+ � T
328
+ 0
329
+ ∥u∥2
330
+ W 1,2(Ω;R3) dt ≤ c(T, D0, ̺, ϑ, u),
331
+ (3.4)
332
+ � T
333
+ 0
334
+
335
+
336
+
337
+ |∇xϑ|2 + |∇x log(ϑ)|2�
338
+ dx dt ≤ c(T, D0, ̺, ϑ, u),
339
+
340
+ � T
341
+ 0
342
+ ∥ϑ∥2
343
+ W 1,2(Ω) dt +
344
+ � T
345
+ 0
346
+ ∥ log(ϑ)∥2
347
+ W 1,2(Ω) dt ≤ c(T, D0, ̺, ϑ, u).
348
+ (3.5)
349
+ 4
350
+ Estimates of the velocity gradient
351
+ This section is the heart of the paper. In principle, we follow the arguments similar to Fang, Zi,
352
+ and Zhang [5, Section 3] but here adapted to the inhomogeneous boundary conditions.
353
+ 4.1
354
+ Estimates of the velocity material derivative
355
+ Let us introduce the material derivative of a function g,
356
+ Dtg = ∂tg + u · ∇xg.
357
+ Accordingly, we may rewrite the momentum equation (1.2) as
358
+ ̺Dtu + ∇xp = divxS + ̺f.
359
+ (4.1)
360
+ Now, consider the scalar product of the momentum equation (4.1) with Dt(u − uB),
361
+ ̺|Dtu|2 + ∇xp · Dt(u − uB) = divxS(Dxu) · Dt(u − uB) + ̺f · Dt(u − uB) + ̺Dtu · DtuB. (4.2)
362
+ The next step is integrating (4.2) over Ω. Here and hereafter we use the hypothesis uB·n|∂Ω = 0
363
+ yielding
364
+ Dt(u − uB)|∂Ω = (∂tu − u · ∇x(u − uB)) |∂Ω = −uB · ∇x(u − uB)|∂Ω = 0.
365
+ (4.3)
366
+ Writing
367
+ divxS(Dxu) = µ∆xu +
368
+
369
+ η + µ
370
+ 3
371
+
372
+ ∇xdivxu,
373
+ and making use of (4.3) we obtain
374
+
375
+
376
+ divxS(Dxu) · Dt(u − uB) dx
377
+ 7
378
+
379
+ = −
380
+
381
+
382
+ S(Dxu) : ∇x∂tu dx
383
+ − µ
384
+
385
+
386
+ ∇xu : ∇x
387
+
388
+ u · ∇x(u − uB)
389
+
390
+ dx −
391
+
392
+ η + µ
393
+ 3
394
+ � �
395
+
396
+ divxu divx
397
+
398
+ u · ∇x(u − uB)
399
+
400
+ dx
401
+ = − 1
402
+ 2
403
+ d
404
+ dt
405
+
406
+
407
+ S(Dxu) : Dxu dx
408
+ − µ
409
+
410
+
411
+ ∇xu : ∇x
412
+
413
+ u · ∇x(u − uB)
414
+
415
+ dx −
416
+
417
+ η + µ
418
+ 3
419
+ � �
420
+
421
+ divxu divx
422
+
423
+ u · ∇x(u − uB)
424
+
425
+ dx,
426
+ (4.4)
427
+ where, furthermore,
428
+
429
+
430
+ ∇xu : ∇x(u · ∇xu) dx =
431
+
432
+
433
+ ∇xu : (∇xu · ∇xu) dx + 1
434
+ 2
435
+
436
+
437
+ u · ∇x|∇xu|2 dx
438
+ =
439
+
440
+
441
+ ∇xu : (∇xu · ∇xu) dx − 1
442
+ 2
443
+
444
+
445
+ divxu|∇xu|2 dx
446
+ (4.5)
447
+ Note carefully we have used u · n|∂Ω = 0 in the last integration. Similarly,
448
+
449
+
450
+ divxu divx(u · ∇xu) dx =
451
+
452
+
453
+ divxu ∇xu : ∇t
454
+ xu dx − 1
455
+ 2
456
+
457
+
458
+ (divxu)3 dx.
459
+ (4.6)
460
+ Thus summing up the previous observations, we get
461
+ 1
462
+ 2
463
+ d
464
+ dt
465
+
466
+
467
+ S(Dxu) : Dxu dx + 1
468
+ 2
469
+
470
+
471
+ ̺|Dtu|2 dx +
472
+
473
+
474
+ ∇xp · Dt(u − uB) dx
475
+ ≤ c(T, D0, ̺, ϑ, u)
476
+
477
+ 1 +
478
+
479
+
480
+ |∇xu|3 dx
481
+
482
+ .
483
+ (4.7)
484
+ Moreover,
485
+
486
+
487
+ ∇xp · Dt(u − uB) dx = −
488
+
489
+
490
+ p divx(Dt(u − uB)) dx
491
+ = −
492
+
493
+
494
+ p divxDtu dx +
495
+
496
+
497
+ p divx(u · ∇xuB) dx,
498
+ (4.8)
499
+ where
500
+ p divxDtu = ∂t(p divxu) −
501
+
502
+ ∂tp + divx(pu)
503
+
504
+ divxu + divx(pu)divxu + p divx(u · ∇xu)
505
+ = ∂t(p divxu) −
506
+
507
+ ∂tp + divx(pu)
508
+
509
+ divxu + p∇xu : ∇t
510
+ xu + divx
511
+
512
+ pu divxu
513
+
514
+ .
515
+ As u · n|∂Ω = 0, we have
516
+
517
+
518
+ divx
519
+
520
+ pu divxu
521
+
522
+ dx = 0,
523
+ 8
524
+
525
+ and the above estimates together with (4.7) give rise to
526
+ 1
527
+ 2
528
+ d
529
+ dt
530
+
531
+
532
+ S(Dxu) : Dxu dx − d
533
+ dt
534
+
535
+
536
+ pdivxu dx + 1
537
+ 2
538
+
539
+
540
+ ̺|Dtu|2 dx
541
+ ≤ c(T, D0, ̺, ϑ, u)
542
+
543
+ 1 +
544
+
545
+
546
+ |∇xu|3 dx
547
+
548
+
549
+
550
+
551
+
552
+ ∂tp + divx(pu)
553
+
554
+ divxu dx.
555
+ Finally, we realize
556
+ ∂tp + divx(pu) = ̺Dtϑ
557
+ to conclude
558
+ 1
559
+ 2
560
+ d
561
+ dt
562
+
563
+
564
+ S(Dxu) : Dxu dx − d
565
+ dt
566
+
567
+
568
+ pdivxu dx + 1
569
+ 2
570
+
571
+
572
+ ̺|Dtu|2 dx
573
+ ≤ c(T, D0, ̺, ϑ, u)
574
+
575
+ 1 +
576
+
577
+
578
+ ̺|Dtϑ||∇xu| dx +
579
+
580
+
581
+ |∇xu|3 dx
582
+
583
+ .
584
+ (4.9)
585
+ 4.2
586
+ Higher order velocity material derivative estimates
587
+ Following [5, Section 3, Lemma 3.3], see also Hoff [12], we deduce
588
+ ̺D2
589
+ t u + ∇x∂tp + divx(∇xp ⊗ u)
590
+ = µ
591
+
592
+ ∆x∂tu + divx(∆xu ⊗ u)
593
+
594
+ +
595
+
596
+ η + µ
597
+ 3
598
+ � �
599
+ ∇xdivx∂tu + divx ((∇xdivxu) ⊗ u)
600
+
601
+ + ̺u · ∇xf.
602
+ (4.10)
603
+ Next, we compute
604
+ DtuB = u · ∇xuB,
605
+ D2
606
+ t uB = ∂tu · ∇xuB + u · ∇x(u · ∇xuB)
607
+ = Dtu · ∇xuB − (u · ∇xu) · ∇xuB + u · ∇x(u · ∇xuB)
608
+ = Dtu · ∇xuB + (u ⊗ u) : ∇2
609
+ xuB.
610
+ (4.11)
611
+ Consequently, we may rewrite (4.10) in the form
612
+ ̺D2
613
+ t (u − uB) + ∇x∂tp + divx(∇xp ⊗ u)
614
+ = µ
615
+
616
+ ∆x∂tu + divx(∆xu ⊗ u)
617
+
618
+ +
619
+
620
+ η + µ
621
+ 3
622
+ � �
623
+ ∇xdivx∂tu + divx ((∇xdivxu) ⊗ u)
624
+
625
+ + ̺u · ∇xf
626
+ − ̺Dtu · ∇xuB − ̺(u ⊗ u) : ∇2
627
+ xuB.
628
+ (4.12)
629
+ The next step is considering the scalar product of (4.12) with Dt(u − uB) and integrating over
630
+ Ω. The resulting integrals can be handled as follows:
631
+ ̺D2
632
+ t (u − uB) · Dt(u − uB) = ̺1
633
+ 2Dt|Dt(u − uB)|2
634
+ 9
635
+
636
+ = 1
637
+
638
+
639
+ ∂t|Dt(u − uB)|2 + u · ∇x|Dt(u − uB)|2�
640
+ = 1
641
+ 2∂t
642
+
643
+ ̺|Dt(u − uB)|2�
644
+ + 1
645
+ 2divx
646
+
647
+ ̺u|Dt(u − uB)|2�
648
+ ,
649
+ where we have used the equation of continuity (1.1). Seeing that u · n|∂Ω = 0 we get
650
+
651
+
652
+ ̺D2
653
+ t (u − uB) · Dt(u − uB) dx = d
654
+ dt
655
+ 1
656
+ 2
657
+
658
+
659
+ ̺|Dt(u − uB)|2 dx.
660
+ (4.13)
661
+ Similarly,
662
+
663
+
664
+
665
+ ∇x∂tp + divx(∇xp ⊗ u)
666
+
667
+ · Dt(u − uB) dx
668
+ = −
669
+
670
+
671
+
672
+ ∂tp + divx(pu)
673
+
674
+ divxDt(u − uB) dx
675
+ +
676
+
677
+
678
+
679
+ divx(pu)divxDt(u − uB) − ∇xp ⊗ u : ∇xDt(u − uB)
680
+
681
+ dx,
682
+ (4.14)
683
+ where
684
+
685
+
686
+ ∇xp ⊗ u : ∇xDt(u − uB) dx
687
+ = −
688
+
689
+
690
+ p∇xu : ∇xDt(u − uB) dx +
691
+
692
+
693
+ ∇x(pu) : ∇xDt(u − uB) dx.
694
+ In addition, as Dt(u−uB) vanishes on ∂Ω, we can perform by parts integration in the last integral
695
+ obtaining
696
+
697
+
698
+ ∇x(pu) : ∇xDt(u − uB) dx =
699
+
700
+
701
+ divx(pu)divxDt(u − uB) dx.
702
+ Thus, similarly to the preceding section, we conclude
703
+
704
+
705
+
706
+ ∇x∂tp + divx(∇xp ⊗ u)
707
+
708
+ · Dt(u − uB) dx
709
+ = −
710
+
711
+
712
+ ̺DtϑdivxDt(u − uB) dx +
713
+
714
+
715
+ p∇xu : ∇xDt(u − uB) dx.
716
+ (4.15)
717
+ Analogously,
718
+
719
+
720
+
721
+ ∆x∂tu + divx(∆xu ⊗ u)
722
+
723
+ · Dt(u − uB) dx
724
+ = −
725
+
726
+
727
+ ∇x∂tu : ∇xDt(u − uB) dx −
728
+
729
+
730
+ (∆xu ⊗ u) : ∇xDt(u − uB) dx
731
+ = −
732
+
733
+
734
+ ∇xDtu : ∇xDt(u − uB) dx −
735
+
736
+
737
+
738
+ ∆xu ⊗ u − ∇x(u · ∇xu)
739
+
740
+ : ∇xDt(u − uB) dx, (4.16)
741
+ 10
742
+
743
+ where, using summation convention,
744
+
745
+
746
+
747
+ ∆xu ⊗ u
748
+
749
+ : ∇xDt(u − uB) dx
750
+ =
751
+
752
+
753
+ ∂xk
754
+
755
+ uj∂xkui
756
+
757
+ ∂xjDt(u − uB)i dx −
758
+
759
+
760
+ ∂xkui∂xkuj∂xjDt(u − uB)i dx
761
+ =
762
+
763
+
764
+ ∂xj
765
+
766
+ uj∂xkui
767
+
768
+ ∂xkDt(u − uB)i dx −
769
+
770
+
771
+ ∂xkui∂xkuj∂xjDt(u − uB)i dx
772
+ =
773
+
774
+
775
+ divxu ∇xu : ∇xDt(u − uB) dx
776
+ +
777
+
778
+
779
+
780
+ uj∂xk∂xjui
781
+
782
+ ∂xkDt(u − uB)i dx −
783
+
784
+
785
+ ∂xkui∂xkuj∂xjDt(u − uB)i dx
786
+ =
787
+
788
+
789
+ ∇x(u · ∇xu) : ∇xDt(u − uB) dx +
790
+
791
+
792
+ divxu ∇xu : ∇xDt(u − uB) dx
793
+
794
+
795
+
796
+ ∂xjui∂xkuj∂xkDt(u − uB)i dx −
797
+
798
+
799
+ ∂xkui∂xkuj∂xjDt(u − uB)i dx.
800
+ (4.17)
801
+ Summing up (4.16), (4.17) we conclude
802
+
803
+
804
+
805
+ ∆x∂tu + divx(∆xu ⊗ u)
806
+
807
+ · Dt(u − uB) dx
808
+ = −
809
+
810
+
811
+ ∇xDtu : ∇xDt(u − uB) dx −
812
+
813
+
814
+ divxu ∇xu : ∇xDt(u − uB) dx
815
+ +
816
+
817
+
818
+ ∂xjui∂xkuj∂xkDt(u − uB)i dx +
819
+
820
+
821
+ ∂xkui∂xkuj∂xjDt(u − uB)i dx.
822
+ (4.18)
823
+ Estimating the remaining integrals in (4.12) in a similar manner we may infer
824
+ 1
825
+ 2
826
+ d
827
+ dt
828
+
829
+
830
+ ̺|Dt(u − uB)|2 dx + µ
831
+
832
+
833
+ |∇xDt(u − uB)|2 dx +
834
+
835
+ η + µ
836
+ 3
837
+ � �
838
+
839
+ |divxDt(u − uB)|2 dx
840
+ ≤ c(T, D0, ̺, ϑ, u)
841
+
842
+ 1 +
843
+
844
+
845
+ ̺|Dtϑ|2 dx +
846
+
847
+
848
+ |∇xu|4 dx +
849
+
850
+
851
+ ̺|Dtu|2 dx
852
+
853
+ .
854
+ (4.19)
855
+ cf. [5, Section 3, Lemma 3.3].
856
+ 4.3
857
+ Velocity decomposition
858
+ Following the original idea of Sun, Wang, and Zhang [18], we decompose the velocity field in the
859
+ form:
860
+ u = v + w,
861
+ (4.20)
862
+ divxS(Dxv) = ∇xp in (0, T) × Ω, v|∂Ω = 0,
863
+ (4.21)
864
+ 11
865
+
866
+ divxS(Dxw) = ̺Dtu − ̺f in (0, T) × Ω, w|∂Ω = uB.
867
+ (4.22)
868
+ Since
869
+ divxS(Dx∂tv) = ∇x∂tp in (0, T) × Ω, v|∂Ω = 0,
870
+ we get
871
+
872
+
873
+ ∂tp divxv dx = −
874
+
875
+
876
+ ∇x∂tp · v dx = 1
877
+ 2
878
+ d
879
+ dt
880
+
881
+
882
+ S(Dxv) : Dxv dx.
883
+ (4.23)
884
+ Moreover, the standard elliptic estimates for the Lam´e operator yield:
885
+ ∥v∥W 1,q(Ω;R3) ≤ c(q, ̺, ϑ) for all 1 ≤ q < ∞,
886
+ (4.24)
887
+ ∥v∥W 2,q(Ω;R3) ≤ c(q, ̺, ϑ)
888
+
889
+ ∥∇x̺∥Lq(Ω;R3) + ∥∇xϑ∥Lq(Ω;R3)
890
+
891
+ , 1 < q < ∞.
892
+ (4.25)
893
+ Similarly,
894
+ ∥w∥W 2,2(Ω;R3) ≤ c(T, D0, ̺, ϑ, u)
895
+
896
+ 1 + ∥√̺∂tu∥L2(Ω;R3) + ∥∇xu∥L2(Ω;R3×3)
897
+
898
+ .
899
+ (4.26)
900
+ The estimates (4.24)–(4.26) are uniform in the time interval [0, T).
901
+ 4.4
902
+ Temperature estimates
903
+ Similarly to Fang, Zi, Zhang [5, Section 3, Lemma 3.4] we multiply the internal energy equation
904
+ (1.3) on ∂tϑ and integrate over Ω obtaining
905
+ cv
906
+
907
+
908
+ ̺|Dtϑ|2 dx + κ
909
+ 2
910
+ d
911
+ dt
912
+
913
+
914
+ |∇xϑ|2 dx
915
+ = cv
916
+
917
+
918
+ ̺Dtϑ u · ∇xϑ dx −
919
+
920
+
921
+ ̺ϑ divxu Dtϑ dx +
922
+
923
+
924
+ ̺ϑ divxu u · ∇xϑ dx
925
+ + d
926
+ dt
927
+
928
+
929
+ ϑ S(Dxu) : ∇xu dx
930
+ − µ
931
+
932
+
933
+ ϑ
934
+
935
+ ∇xu + ∇t
936
+ xu − 2
937
+ 3divxuI
938
+
939
+ :
940
+
941
+ ∇x∂tu + ∇t
942
+ x∂tu − 2
943
+ 3divx∂tuI
944
+
945
+ dx
946
+ − 2η
947
+
948
+
949
+ ϑ divxu divx∂tu dx.
950
+ (4.27)
951
+ Indeed the term involving the boundary integral is handled as
952
+ −κ
953
+
954
+
955
+ ∆xϑ ∂tϑ dx = −κ
956
+
957
+ ∂Ω
958
+ ∂tϑB∇xϑ · n dSx + κ
959
+ 2
960
+ d
961
+ dt
962
+
963
+
964
+ |∇xϑ|2 dx,
965
+ where
966
+
967
+ ∂Ω
968
+ ∂tϑB∇xϑ · n dSx = 0
969
+ 12
970
+
971
+ as the boundary temperature is independent of t.
972
+ Similarly to Fang, Zi, Zhang [5, Section 3, Lemma 3.4], we have to show that the intergrals
973
+
974
+
975
+ ϑ ∇xu : ∇x∂tu dx,
976
+
977
+
978
+ ϑ ∇xu : ∇t
979
+ x∂tu dx, and
980
+
981
+
982
+ ϑ divxu divx∂tu dx
983
+ can be rewritten in the form compatible with (4.19), meaning with the time derivatives replaced
984
+ by material derivatives. Fortunately, this step can be carried out in the present setting using only
985
+ the boundary condition u · n|∂Ω = 0. Indeed we get
986
+
987
+
988
+ ϑ ∇xu : ∇x∂tu dx =
989
+
990
+
991
+ ϑ ∇xu : ∇x(Dtu) dx −
992
+
993
+
994
+ ϑ ∇xu : ∇x(u · ∇xu) dx,
995
+ where
996
+
997
+
998
+ ϑ ∇xu : ∇x(u · ∇xu) dx
999
+ =
1000
+
1001
+
1002
+ ϑ ∇xu : (∇xu · ∇xu) dx + 1
1003
+ 2
1004
+
1005
+
1006
+ ϑ u · ∇x|∇xu|2 dx
1007
+ =
1008
+
1009
+
1010
+ ϑ ∇xu : (∇xu · ∇xu) dx − 1
1011
+ 2
1012
+
1013
+
1014
+ |∇xu|2 ∇xϑ · u dx − 1
1015
+ 2
1016
+
1017
+
1018
+ |∇xu|2 ϑdivxu dx.
1019
+ Similarly,
1020
+
1021
+
1022
+ ϑ ∇xu : ∇t
1023
+ x∂tu dx =
1024
+
1025
+
1026
+ ϑ ∇xu : ∇t
1027
+ x(Dtu) dx −
1028
+
1029
+
1030
+ ϑ ∇xu : ∇t
1031
+ x(u · ∇xu) dx,
1032
+ where
1033
+
1034
+
1035
+ ϑ ∇xu : ∇t
1036
+ x(u · ∇xu) dx
1037
+ =
1038
+
1039
+
1040
+ ϑ ∇xu : (∇t
1041
+ xu · ∇t
1042
+ xu) dx + 1
1043
+ 2
1044
+
1045
+
1046
+ ϑ u · ∇x(∇xu : ∇t
1047
+ xu) dx
1048
+ =
1049
+
1050
+
1051
+ ϑ ∇xu : (∇t
1052
+ xu · ∇t
1053
+ xu) dx − 1
1054
+ 2
1055
+
1056
+
1057
+ (∇xu : ∇t
1058
+ xu) ∇xϑ · u dx − 1
1059
+ 2
1060
+
1061
+
1062
+ (∇xu : ∇t
1063
+ xu) ϑdivxu dx.
1064
+ Finally,
1065
+
1066
+
1067
+ ϑ divxu divx∂tu dx =
1068
+
1069
+
1070
+ ϑ divxu divxDtu dx −
1071
+
1072
+
1073
+ ϑ divxu divx(u · ∇xu) dx,
1074
+ where
1075
+
1076
+
1077
+ ϑ divxu divx(u · ∇xu) dx
1078
+ 13
1079
+
1080
+ =
1081
+
1082
+
1083
+ ϑ divxu (∇xu : ∇t
1084
+ xu) dx + 1
1085
+ 2
1086
+
1087
+
1088
+ ϑu · ∇x|divxu|2 dx
1089
+ =
1090
+
1091
+
1092
+ ϑ divxu (∇xu : ∇t
1093
+ xu) dx − 1
1094
+ 2
1095
+
1096
+
1097
+ |divxu|2 ∇xϑ · u dx − 1
1098
+ 2
1099
+
1100
+
1101
+ |divxu|2 ϑdivxu dx.
1102
+ We conclude, using (4.7), (4.19), and (4.27),
1103
+
1104
+
1105
+ |∇xϑ|2(τ, ·) dx +
1106
+ � τ
1107
+ 0
1108
+
1109
+
1110
+ ̺|Dtϑ|2 dx dt
1111
+ ≤ c(T, D0, ̺, ϑ, u)
1112
+
1113
+ 1 +
1114
+ � τ
1115
+ 0
1116
+
1117
+
1118
+ |∇xu|4 dx dt
1119
+
1120
+ .
1121
+ (4.28)
1122
+ Next, by virtue of the decomposition u = v + w and the bound (4.24),
1123
+
1124
+
1125
+ |∇xu|4 dx
1126
+ <∼
1127
+
1128
+
1129
+ |∇xv|4 dx +
1130
+
1131
+
1132
+ |∇xw|4 dx ≤ c(T, D0, ̺, ϑ, u)
1133
+
1134
+ 1 +
1135
+
1136
+
1137
+ |∇xw|4 dx
1138
+
1139
+ ,
1140
+ (4.29)
1141
+ and, similarly,
1142
+ ∥w∥L∞(Ω;R3) ≤ ∥u∥L∞(Ω;R3) + ∥v∥L∞(Ω;R3) ≤ c(T, D0, ̺, ϑ, u).
1143
+ (4.30)
1144
+ Recalling the Gagliardo–Nirenberg interpolation inequality in the form
1145
+ ∥∇xU∥2
1146
+ L4(Ω;R3) ≤ ∥U∥L∞(Ω)∥∆xU∥L2(Ω) whenever U|∂Ω = 0,
1147
+ (4.31)
1148
+ we may use (4.29), (4.30) to rewrite (4.28) in the form
1149
+
1150
+
1151
+ |∇xϑ|2(τ, ·) dx +
1152
+ � τ
1153
+ 0
1154
+
1155
+
1156
+ ̺|Dtϑ|2 dx dt
1157
+ ≤ c(T, D0, ̺, ϑ, u)
1158
+
1159
+ 1 +
1160
+ � τ
1161
+ 0
1162
+
1163
+
1164
+ |∇xϑ|2 dx dt +
1165
+ � τ
1166
+ 0
1167
+ ∥w∥2
1168
+ W 2,2(Ω;R3) dt
1169
+
1170
+ .
1171
+ (4.32)
1172
+ Finally, we use the elliptic estimates (4.26) to conclude
1173
+
1174
+
1175
+ |∇xϑ|2(τ, ·) dx +
1176
+ � τ
1177
+ 0
1178
+
1179
+
1180
+ ̺|Dtϑ|2 dx dt
1181
+ ≤ c(T, D0, ̺, ϑ, u)
1182
+
1183
+ 1 +
1184
+ � τ
1185
+ 0
1186
+
1187
+
1188
+
1189
+ |∇xϑ|2 + |∇xu|2�
1190
+ dx dt +
1191
+ � τ
1192
+ 0
1193
+ ∥√̺∂tu∥2
1194
+ L2(Ω;R3) dt
1195
+
1196
+ .
1197
+ (4.33)
1198
+ Summing up (4.7), (4.19), and (4.33) we may apply Gronwall’s lemma to obtain the following
1199
+ bounds:
1200
+ sup
1201
+ t∈[0,T)
1202
+ ∥u(t, ·)∥W 1,2(Ω;R3) ≤ c(T, D0, ̺, ϑ, u),
1203
+ (4.34)
1204
+ sup
1205
+ t∈[0,T)
1206
+ ∥√̺Dtu(t, ·)∥L2(Ω;R3) ≤ c(T, D0, ̺, ϑ, u),
1207
+ (4.35)
1208
+ 14
1209
+
1210
+ sup
1211
+ t∈[0,T)
1212
+ ∥ϑ(t, ·)∥W 1,2(Ω) ≤ c(T, D0, ̺, ϑ, u),
1213
+ (4.36)
1214
+ � T
1215
+ 0
1216
+
1217
+
1218
+ |∇xDtu|2 dx dt ≤ c(T, D0, ̺, ϑ, u),
1219
+ (4.37)
1220
+ � T
1221
+ 0
1222
+
1223
+
1224
+ ̺|Dtϑ|2 dx dt ≤ c(T, D0, ̺, ϑ, u).
1225
+ (4.38)
1226
+ Moreover, it follows from (4.24), (4.31), (4.35)
1227
+ sup
1228
+ t∈[0,T)
1229
+ ∥∇xu(t, ·)∥L4(Ω;R3×3) ≤ c(T, D0, ̺, ϑ, u).
1230
+ (4.39)
1231
+ In addition, (4.38), (4.39) and the standard parabolic estimates applied to the internal energy
1232
+ balance (1.3) yield
1233
+ � T
1234
+ 0
1235
+ ∥ϑ∥2
1236
+ W 2,2(Ω) dt ≤ c(T, D0, ̺, ϑ, u).
1237
+ (4.40)
1238
+ 5
1239
+ Second energy bound
1240
+ It follows from (4.26), (4.35) that
1241
+ sup
1242
+ t∈[0,T)
1243
+ ∥w(t, ·)∥W 2,2(Ω;R3) ≤ c(T, D0, ̺, ϑ, u);
1244
+ (5.1)
1245
+ whence, by virtue of (4.24) and Sobolev embedding W 1,2(Ω) ֒→ L6(Ω),
1246
+ sup
1247
+ t∈[0,T)
1248
+ ∥∇xu(t, ·)∥2
1249
+ L6(Ω;R3×3) ≤ c(T, D0, ̺, ϑ, u).
1250
+ (5.2)
1251
+ Moreover, as a consequence of (4.37), Dtu is bounded in L2(L6), which, combined with (5.2), gives
1252
+ rise to
1253
+ � T
1254
+ 0
1255
+ ∥∂tu∥2
1256
+ L6(Ω;R3) dt ≤ c(T, D0, ̺, ϑ, u).
1257
+ (5.3)
1258
+ Finally, going back to (4.22) we conclude
1259
+ � T
1260
+ 0
1261
+ ∥w∥2
1262
+ W 2,6(Ω;R3) dt ≤ c(T, D0, ̺, ϑ, u),
1263
+ (5.4)
1264
+ and
1265
+ � T
1266
+ 0
1267
+ ∥u∥2
1268
+ W 1,q(Ω;R3) dt ≤ c(T, D0, ̺, ϑ, u, q) for any 1 ≤ q < ∞.
1269
+ (5.5)
1270
+ 15
1271
+
1272
+ 6
1273
+ Estimates of the derivatives of the density
1274
+ Using (5.4), (5.5), we may proceed as in [19, Section 5] to deduce the bounds
1275
+ supt∈[0,T)
1276
+
1277
+ ∥∂t̺(t, ·)∥L6(Ω) + ∥̺(t, ·)∥W 1,6(Ω)
1278
+
1279
+ ≤ c(T, D0, ̺, ϑ, u).
1280
+ (6.1)
1281
+ Revisiting the momentum equation (1.2) we use (6.1) together with the other bounds established
1282
+ above to obtain
1283
+ � T
1284
+ 0
1285
+ ∥u∥2
1286
+ W 2,6(Ω;R3) dt ≤ c(T, D0, ̺, ϑ, u).
1287
+ (6.2)
1288
+ 6.1
1289
+ Positivity of the density and temperature
1290
+ It follows from (6.2) that divxu is bounded in L1(0, T; L∞(Ω)). Thus the equation of continuity
1291
+ (1.1) yields a positive lower bound on the density
1292
+ inf
1293
+ (t,x)∈[0,T)×Ω ̺(t, x) ≥ ̺ > 0,
1294
+ (6.3)
1295
+ where the lower bound depends on the data as well as on the length T of the time interval.
1296
+ Similarly, rewriting the internal energy balance equation (1.3) in the form
1297
+ cv (∂tϑ + u · ∇xϑ) − κ
1298
+ ̺∆xϑ = 1
1299
+ ̺S : Dxu − ϑdivxu
1300
+ (6.4)
1301
+ we may apply the standard parabolic maximum/minimum principle to deduce
1302
+ inf
1303
+ (t,x)∈[0,T)×Ω ϑ(t, x) ≥ ϑ > 0.
1304
+ (6.5)
1305
+ 7
1306
+ Parabolic regularity for the heat equation
1307
+ We rewrite the parabolic equation (6.4) in terms of Θ = ϑ − ϑB. Recalling ∆xϑB = 0 we get
1308
+ cv (∂tΘ + u · ∇xϑ) − κ
1309
+ ̺∆xΘ = 1
1310
+ ̺S : Dxu − ϑdivxu
1311
+ (7.1)
1312
+ with the homogeneous Dirichlet boundary conditions
1313
+ Θ|∂Ω = 0.
1314
+ (7.2)
1315
+ Now, we can apply all arguments of [10, Sections 4.6, 4.7] to Θ obtaining the bounds
1316
+ ∥ϑ∥Cα([0,T]×Ω) ≤ c(T, D0, ̺, ϑ, u) for some α > 0,
1317
+ (7.3)
1318
+ ∥ϑ∥Lp(0,T;W 2,3(Ω)) + ∥∂tϑ∥Lp(0,T;L3(Ω)) ≤ c(T, D0, ̺, ϑ, u) for all 1 ≤ p < ∞,
1319
+ (7.4)
1320
+ together with
1321
+ ∥u∥Lp(0,T;W 2,6(Ω;R3)) + ∥∂tu∥Lp(0,T;L6(Ω;R3)) ≤ c(T, D0, ̺, ϑ, u) for any 1 ≤ p < ∞.
1322
+ (7.5)
1323
+ 16
1324
+
1325
+ 8
1326
+ Final estimates
1327
+ The bounds (7.5) imply, in particular,
1328
+ sup
1329
+ (t,x)∈[0,T)×Ω
1330
+ |∇xu(t, x)| ≤ c(T, D0, ̺, ϑ, u).
1331
+ (8.1)
1332
+ Thus the desired higher order estimates can be obtained exactly as in [9, Section 4.6]. Indeed
1333
+ the arguments of [9, Section 4.6] are based on differentiating the equation (7.1) with respect to
1334
+ time which gives rise to a parabolic problem for ∂tϑ with the homogeneous Dirichlet boundary
1335
+ conditions ∂tϑ|∂Ω = 0. Indeed we get
1336
+ cv∂2
1337
+ ttϑ + cvu · ∇x∂tϑ − κ
1338
+ ̺∆x∂tϑ =−cv∂tu · ∇xϑ − 1
1339
+ ̺2∂t̺ (κ∆xϑ + S(Dxu) : Dxu)
1340
+ +2
1341
+ ̺ S(Dxu) : Dx∂tu − ∂tϑ divxu − ϑ divx∂tu.
1342
+ The estimates obtained in the previous sections imply that the right–hand side of the above
1343
+ equation is bounded in L2(0, T; L2(Ω)). Thus multiplying the equation on ∆x∂tϑ and performing
1344
+ the standard by parts integration, we get the desired estimates as in [9, Section 4.6].
1345
+ The remaining estimates are obtained exactly as in [9, Section 4.6] :
1346
+ sup
1347
+ t∈[0,T)
1348
+ ∥ϑ(t, ·)∥W 3,2(Ω) + sup
1349
+ t∈[0,T)
1350
+ ∥∂tϑ(t, ·)∥W 1,2(Ω) ≤ c(T, D0, ̺, ϑ, u),
1351
+ (8.2)
1352
+ � T
1353
+ 0
1354
+
1355
+ ∥∂tϑ∥2
1356
+ W 2,2(Ω) + ∥ϑ∥2
1357
+ W 4,2(Ω)
1358
+
1359
+ dt ≤ c(T, D0, ̺, ϑ, u),
1360
+ (8.3)
1361
+ sup
1362
+ t∈[0,T)
1363
+ ∥u(t, ·)∥W 3,2(Ω;R3) + sup
1364
+ t∈[0,T)
1365
+ ∥∂tu(t, ·)∥W 1,2(Ω;R3) ≤ c(T, D0, ̺, ϑ, u),
1366
+ (8.4)
1367
+ � T
1368
+ 0
1369
+
1370
+ ∥∂tu∥2
1371
+ W 2,2(Ω;R3) + ∥u∥2
1372
+ W 4,2(Ω;R3)
1373
+
1374
+ dt ≤ c(T, D0, ̺, ϑ, u),
1375
+ (8.5)
1376
+ and
1377
+ sup
1378
+ t∈[0,T)
1379
+ ∥̺(t, ·)∥W 3,2(Ω) ≤ c(T, D0, ̺, ϑ, u).
1380
+ (8.6)
1381
+ We have completed the proof of Proposition 2.3.
1382
+ References
1383
+ [1] N. Chaudhuri. On weak(measure valued)–strong uniqueness for Navier–Stokes–Fourier system
1384
+ with Dirichlet boundary condition.
1385
+ Archive Preprint Series, 2022.
1386
+ arxiv preprint No.
1387
+ 2207.00991.
1388
+ [2] N. Chaudhuri and E. Feireisl. Navier-Stokes-Fourier system with Dirichlet boundary condi-
1389
+ tions. Appl. Anal., 101(12):4076–4094, 2022.
1390
+ 17
1391
+
1392
+ [3] P. A. Davidson. Turbulence:An introduction for scientists and engineers. Oxford University
1393
+ Press, Oxford, 2004.
1394
+ [4] J. Fan, S. Jiang, and Y. Ou. A blow-up criterion for compressible viscous heat-conductive
1395
+ flows. Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, 27(1):337–350, 2010.
1396
+ [5] D. Fang, R. Zi, and T. Zhang. A blow-up criterion for two dimensional compressible viscous
1397
+ heat-conductive flows. Nonlinear Anal., 75(6):3130–3141, 2012.
1398
+ [6] E. Feireisl, M. Luk´aˇcov´a-Medviˇdov´a, H. Mizerov´a, and B. She. Numerical analysis of com-
1399
+ pressible fluid flows. Springer-Verlag, Cham, 2022.
1400
+ [7] E. Feireisl and M. Luk´aˇcov´a-Medviˇdov´a. Convergence of a stochastic collocation finite volume
1401
+ method for the compressible Navier–Stokes system. Archive Preprint Series, 2021. arxiv
1402
+ preprint No.2111.07435.
1403
+ [8] E. Feireisl and M. Luk´aˇcov´a-Medviˇdov´a. Statistical solutions for the Navier–Stokes–Fourier
1404
+ system. Archive Preprint Series, 2022. arxiv preprint No. 2212.06784.
1405
+ [9] E. Feireisl, A. Novotn´y, and Y. Sun. A regularity criterion for the weak solutions to the
1406
+ Navier-Stokes-Fourier system. Arch. Ration. Mech. Anal., 212(1):219–239, 2014.
1407
+ [10] E. Feireisl and Y. Sun. Conditional regularity of very weak solutions to the Navier-Stokes-
1408
+ Fourier system.
1409
+ In Recent advances in partial differential equations and applications, vol-
1410
+ ume 666 of Contemp. Math., pages 179–199. Amer. Math. Soc., Providence, RI, 2016.
1411
+ [11] E. Feireisl, H. Wen, and C. Zhu. On Nash’s conjecture for models of viscous, compressible,
1412
+ and heat conducting fluids. IM ASCR Prague, preprint No. IM 2022 6, 2022.
1413
+ [12] D. Hoff. Global solutions of the Navier-Stokes equations for multidimensional compressible
1414
+ flow with discontinuous initial data. J. Differential Equations, 120:215–254, 1995.
1415
+ [13] X. Huang and J. Li. Serrin-type blowup criterion for viscous, compressible, and heat conduct-
1416
+ ing Navier-Stokes and magnetohydrodynamic flows. Comm. Math. Phys., 324(1):147–171,
1417
+ 2013.
1418
+ [14] X. Huang, J. Li, and Y. Wang. Serrin-type blowup criterion for full compressible Navier-Stokes
1419
+ system. Arch. Ration. Mech. Anal., 207(1):303–316, 2013.
1420
+ [15] Q. Jiu, Y. Wang, and Y. Ye. Refined blow-up criteria for the full compressible Navier-Stokes
1421
+ equations involving temperature. J. Evol. Equ., 21(2):1895–1916, 2021.
1422
+ [16] F. Merle, P. Rapha¨el, I. Rodnianski, and J. Szeftel. On the implosion of a compressible fluid
1423
+ I: smooth self-similar inviscid profiles. Ann. of Math. (2), 196(2):567–778, 2022.
1424
+ 18
1425
+
1426
+ [17] F. Merle, P. Rapha¨el, I. Rodnianski, and J. Szeftel. On the implosion of a compressible fluid
1427
+ II: singularity formation. Ann. of Math. (2), 196(2):779–889, 2022.
1428
+ [18] Y. Sun, C. Wang, and Z. Zhang. A Beale-Kato-Majda criterion for the 3-D compressible
1429
+ Navier-Stokes equations. J. Math. Pures Appl., 95(1):36–47, 2011.
1430
+ [19] Y. Sun, C. Wang, and Z. Zhang. A Beale-Kato-Majda criterion for three dimensional com-
1431
+ pressible viscous heat-conductive flows. Arch. Ration. Mech. Anal., 201(2):727–742, 2011.
1432
+ [20] A. Valli. A correction to the paper: “An existence theorem for compressible viscous fluids”
1433
+ [Ann. Mat. Pura Appl. (4) 130 (1982), 197–213; MR 83h:35112]. Ann. Mat. Pura Appl. (4),
1434
+ 132:399–400 (1983), 1982.
1435
+ [21] A. Valli. An existence theorem for compressible viscous fluids. Ann. Mat. Pura Appl. (4),
1436
+ 130:197–213, 1982.
1437
+ [22] A. Valli and M. Zajaczkowski. Navier-Stokes equations for compressible fluids: Global exis-
1438
+ tence and qualitative properties of the solutions in the general case. Commun. Math. Phys.,
1439
+ 103:259–296, 1986.
1440
+ [23] H. Wen and C. Zhu. Blow-up criterions of strong solutions to 3D compressible Navier-Stokes
1441
+ equations with vacuum. Adv. Math., 248:534–572, 2013.
1442
+ [24] H. Wen and C. Zhu.
1443
+ Global solutions to the three-dimensional full compressible Navier-
1444
+ Stokes equations with vacuum at infinity in some classes of large data. SIAM J. Math. Anal.,
1445
+ 49(1):162–221, 2017.
1446
+ 19
1447
+
QNFRT4oBgHgl3EQfJje8/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
QtFJT4oBgHgl3EQfJyy0/content/tmp_files/2301.11462v1.pdf.txt ADDED
@@ -0,0 +1,2310 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ How poor is the stimulus? Evaluating hierarchical generalization in
2
+ neural networks trained on child-directed speech
3
+ Aditya Yedetore∗1, Tal Linzen2, Robert Frank3, R. Thomas McCoy∗4
4
+ 1Boston University, 2New York University, 3Yale University, 4Princeton University
5
+ yedetore@bu.edu, linzen@nyu.edu, robert.frank@yale.edu,
6
+ tom.mccoy@princeton.edu
7
+ Abstract
8
+ When acquiring syntax, children consistently
9
+ choose hierarchical rules over competing non-
10
+ hierarchical possibilities.
11
+ Is this preference
12
+ due to a learning bias for hierarchical struc-
13
+ ture, or due to more general biases that in-
14
+ teract with hierarchical cues in children’s lin-
15
+ guistic input?
16
+ We explore these possibili-
17
+ ties by training LSTMs and Transformers—
18
+ two types of neural networks without a hi-
19
+ erarchical bias—on data similar in quantity
20
+ and content to children’s linguistic input: text
21
+ from the CHILDES corpus. We then evaluate
22
+ what these models have learned about English
23
+ yes/no questions, a phenomenon for which hi-
24
+ erarchical structure is crucial.
25
+ We find that,
26
+ though they perform well at capturing the sur-
27
+ face statistics of child-directed speech (as mea-
28
+ sured by perplexity), both model types general-
29
+ ize in a way more consistent with an incorrect
30
+ linear rule than the correct hierarchical rule.
31
+ These results suggest that human-like general-
32
+ ization from text alone requires stronger biases
33
+ than the general sequence-processing biases of
34
+ standard neural network architectures.
35
+ 1
36
+ Introduction
37
+ Syntax is driven by hierarchical structure, yet we
38
+ typically encounter sentences as linear sequences
39
+ of words. How do children come to recognize the
40
+ hierarchical nature of the languages they acquire?
41
+ Some argue that humans must have a hierarchical
42
+ inductive bias—an innate predisposition for hierar-
43
+ chical structure (Chomsky, 1965, 1980). An alter-
44
+ native view (e.g., Lewis and Elman, 2001) is that
45
+ no such bias is necessary: there may be clear evi-
46
+ dence for hierarchical structure in children’s input,
47
+ so that children would choose hierarchical rules
48
+ even without a hierarchical bias.
49
+ ∗ Work done while at Johns Hopkins University.
50
+ At first blush, recent work in natural language
51
+ processing (NLP) may seem to indicate that no hier-
52
+ archical bias is necessary. Neural networks trained
53
+ on naturally-occurring text perform impressively
54
+ on syntactic evaluations even though they have no
55
+ explicit syntactic structure built into them (e.g., Gu-
56
+ lordava et al., 2018; Wilcox et al., 2018; Warstadt
57
+ et al., 2020a). However, these results do not pro-
58
+ vide strong evidence about the learning biases re-
59
+ quired to learn language from the data available
60
+ to humans because these models receive very dif-
61
+ ferent training data than humans do (Warstadt and
62
+ Bowman, 2022). First, NLP models are typically
63
+ trained on far more data than children receive, so
64
+ models have more opportunities to encounter rare
65
+ syntactic structures (Linzen, 2020). Second, most
66
+ training sets in NLP are built from Internet text
67
+ (e.g., Wikipedia), which differs qualitatively from
68
+ the utterances that children typically hear; e.g., sen-
69
+ tences in Wikipedia are on average 25 words long
70
+ (Yasseri et al., 2012), compared to 5 words for
71
+ sentences in the North American English subset
72
+ of the CHILDES corpus of child-directed speech
73
+ (MacWhinney, 2000).
74
+ In this work, to evaluate if neural networks with-
75
+ out a hierarchical bias generalize like children do,
76
+ we train models on text1 comparable to the sen-
77
+ tences in children’s linguistic input: English data
78
+ from CHILDES. We then analyze what they have
79
+ learned about the relationship between declarative
80
+ sentences, such as (1a), and their corresponding
81
+ yes/no questions, such as (1b):
82
+ (1)
83
+ a. Those are your checkers.
84
+ b. Are those your checkers?
85
+ Crucially, nearly all naturally-occurring yes/no
86
+ questions are consistent with two rules: one based
87
+ 1Section 6.5 discusses other input types (e.g., visual input).
88
+ arXiv:2301.11462v1 [cs.CL] 26 Jan 2023
89
+
90
+ on hierarchical structure (2), and one based on lin-
91
+ ear order (3):2,3
92
+ (2)
93
+ HIERARCHICALQ: The auxiliary at the start
94
+ of a yes/no question corresponds to the main
95
+ auxiliary of the corresponding declarative.
96
+ (3)
97
+ LINEARQ: The auxiliary at the start of a
98
+ yes/no question corresponds to the first auxil-
99
+ iary of the corresponding declarative.
100
+ Despite the scarcity of evidence disambiguating
101
+ these rules, children reliably favor HIERARCHI-
102
+ CALQ (Crain and Nakayama, 1987), albeit with
103
+ occasional errors consistent with LINEARQ (Am-
104
+ bridge et al., 2008). Yes/no questions thus are a
105
+ prime candidate for an aspect of English syntax
106
+ for which human-like generalization requires a hi-
107
+ erarchical bias. We evaluate yes/no question per-
108
+ formance in LSTMs and Transformers, two neural-
109
+ network architectures that have no inherent hierar-
110
+ chical inductive bias (McCoy et al., 2020; Petty and
111
+ Frank, 2021). These architectures employ different
112
+ computational mechanisms, so consistent results
113
+ across both would indicate that our results are not
114
+ due to idiosyncrasies of one particular architecture.
115
+ To investigate if models generalize more con-
116
+ sistently with the hierarchical or linear rule, we
117
+ evaluate them on cases where the rules make dif-
118
+ ferent predictions, such as (4): under HIERARCHI-
119
+ CALQ, the question that corresponds to (4a) is (4b),
120
+ whereas under LINEARQ it is (4c).
121
+ (4)
122
+ a. The boy who has talked can read.
123
+ b. Can the boy who has talked
124
+ read?
125
+ c. *Has the boy who
126
+ talked can read?
127
+ We find that across several ways of framing the
128
+ learning task, models fail to learn HIERARCHI-
129
+ CALQ. Instead, they generalize in ways that de-
130
+ pend on linear order and on the identities of spe-
131
+ cific words. These results suggest that children’s
132
+ training data, if taken to be words alone, may not
133
+ contain enough hierarchical cues to encourage hier-
134
+ archical generalization in a learner without a hierar-
135
+ chical bias. Thus, explaining human acquisition of
136
+ syntax may require postulating that humans have
137
+ stronger inductive biases than those of LSTMs and
138
+ 2In past work these rules have been framed as transforma-
139
+ tions named MOVE-FIRST and MOVE-MAIN (McCoy et al.,
140
+ 2020). We instead follow Berwick et al. (2011) and frame the
141
+ child’s knowledge as a relationship between sentences.
142
+ 3Though these two rules are the most prominent in prior
143
+ literature, other rules are possible; see Section 5.2.
144
+ Transformers, or that information other than word
145
+ sequences plays a crucial role.4
146
+ 2
147
+ Background
148
+ Though HIERARCHICALQ and LINEARQ often
149
+ make the same predictions, the evidence in chil-
150
+ dren’s input may still favor HIERARCHICALQ.
151
+ The most straightforward evidence would be ut-
152
+ terances that directly disambiguate the rules, such
153
+ as (4b). Pullum and Scholz (2002) show that disam-
154
+ biguating examples appear in the Wall Street Jour-
155
+ nal, in literature, and arguably in child-directed
156
+ speech, but direct evidence may still be too rare to
157
+ robustly support HIERARCHICALQ (Legate and
158
+ Yang, 2002). Nonetheless, children might con-
159
+ clude that yes/no questions obey HIERARCHI-
160
+ CALQ rather than LINEARQ based on indirect
161
+ evidence—evidence that other syntactic phenom-
162
+ ena are hierarchical (Mulligan et al., 2021).
163
+ To test if the cues favoring HIERARCHICALQ
164
+ render a hierarchical bias unnecessary, we study
165
+ how well non-hierarchically-biased models acquire
166
+ English yes/no questions. Several prior papers have
167
+ used this approach, but their training data differed
168
+ from children’s input in important ways: some used
169
+ synthetic datasets (Lewis and Elman, 2001; Frank
170
+ and Mathis, 2007; Clark and Eyraud, 2007; McCoy
171
+ et al., 2020), others used massive Internet corpora
172
+ (Lin et al., 2019; Warstadt and Bowman, 2020),
173
+ and those that used child-directed speech simpli-
174
+ fied the data by replacing each word with its part
175
+ of speech (Perfors et al., 2011; Bod et al., 2012).
176
+ We used training data closer to children’s input,
177
+ namely sentences from CHILDES with word iden-
178
+ tities preserved, rather than being converted to parts
179
+ of speech. Two other recent works have also trained
180
+ neural networks on CHILDES data (Pannitto and
181
+ Herbelot, 2020; Huebner et al., 2021), but neither
182
+ investigated yes/no questions.
183
+ One particularly important reason for training
184
+ models on CHILDES is that, in prior work, differ-
185
+ ent types of training data have yielded diverging
186
+ results: Recent models trained on synthetic data
187
+ failed to properly acquire yes/no questions (McCoy
188
+ et al., 2020; Petty and Frank, 2021), whereas ones
189
+ trained on large Internet corpora scored well on
190
+ evaluations of yes/no questions (Lin et al., 2019;
191
+ Warstadt and Bowman, 2020). Given these differ-
192
+ ing results, it is not clear from past work how these
193
+ 4Our datasets and models will be uploaded online soon to
194
+ facilitate further research.
195
+
196
+ models would generalize when faced with the type
197
+ of data that children receive.
198
+ 3
199
+ Overview of Experimental Setup
200
+ We evaluated models on yes/no questions in two
201
+ ways. First, we used relative acceptability judg-
202
+ ments (Experiment 1): We trained neural networks
203
+ on the task of language modeling (predicting the
204
+ next word at every point in the sentence) and evalu-
205
+ ated whether they assigned a higher probability to
206
+ sentences consistent with LINEARQ or HIERAR-
207
+ CHICALQ. Our second approach was based on text
208
+ generation (Experiment 2): We trained networks
209
+ to take in a declarative sentence and output the
210
+ corresponding question, and tested whether they
211
+ generalized in a way more consistent with LIN-
212
+ EARQ or HIERARCHICALQ. Under both framings,
213
+ we trained models on data from CHILDES and
214
+ evaluated them on targeted datasets constructed to
215
+ differentiate LINEARQ and HIERARCHICALQ.
216
+ 4
217
+ Experiment 1: Relative Acceptability
218
+ 4.1
219
+ Dataset
220
+ To train models on data as similar as possible to
221
+ the sentences children receive, we extracted data
222
+ from CHILDES (MacWhinney, 2000). We used
223
+ the North American English portion. We wished
224
+ to replicate children’s input, so we excluded the
225
+ children’s own utterances, leaving a 9.6-million-
226
+ word corpus. We allocated 90% of the data to
227
+ training, 5% to validation, and 5% to testing. We
228
+ replaced words that appeared two or fewer times in
229
+ the training set with <unk>, giving a replacement
230
+ rate of 0.3%. See Appendix A for more details.
231
+ 4.2
232
+ Task: Next-Word Prediction
233
+ We trained models on next-word prediction, also
234
+ known as language modeling. We chose this task
235
+ for two reasons. First, it is clear empirically that
236
+ next-word prediction can teach neural networks a
237
+ substantial amount about syntax (e.g., Hu et al.,
238
+ 2020). Second, it is plausible that humans per-
239
+ form some version of next-word prediction during
240
+ sentence processing (Altmann and Kamide, 1999;
241
+ Hale, 2001; Levy, 2008; Kutas et al., 2011) and
242
+ that such prediction may play a role in acquisition
243
+ (Elman, 1991). Thus, while next-word prediction
244
+ is certainly not the only goal of human language
245
+ learners, we view this task as a reasonable first step
246
+ in emulating human language acquisition.
247
+ 4.3
248
+ Architectures
249
+ We used two neural network architectures: LSTMs
250
+ (Hochreiter and Schmidhuber, 1997) and Trans-
251
+ formers (Vaswani et al., 2017). We chose these
252
+ models for two reasons. First, they have been the
253
+ most successful architectures in NLP. Thus, we
254
+ have reason to believe that, of the types of low-bias
255
+ models invented, these two are the ones most likely
256
+ to discover linguistic regularities in our CHILDES
257
+ training data. Second, the two architectures pro-
258
+ cess sequences very differently (via recurrence vs.
259
+ via attention). Thus, if both generalize similarly,
260
+ we would have evidence that what was learned is
261
+ strongly evidenced in the data, rather than due to a
262
+ quirk of one particular architecture.
263
+ For our LSTMs, we used 2 layers, a hidden and
264
+ embedding size of 800, a batch size of 20, a dropout
265
+ rate of 0.4, and a learning rate of 10. For our Trans-
266
+ formers, the corresponding values were 4, 800, 10,
267
+ 0.2, and 5, and we used 4 attention heads. We chose
268
+ these values based on a hyperparameter search de-
269
+ scribed in Appendix B. All following results are av-
270
+ eraged across 10 runs with different random seeds.
271
+ 4.4
272
+ Results: Language Model Quality
273
+ Before testing models on questions, we used per-
274
+ plexity to evaluate how well they captured the basic
275
+ structure of their training domain. As a baseline,
276
+ we used a 5-gram model with Kneser-Ney smooth-
277
+ ing (Kneser and Ney, 1995) trained with KenLM
278
+ (Heafield, 2011). The test set perplexity for the
279
+ 5-gram baseline was 24.37, while the average test
280
+ set perplexity for the LSTMs and Transformers
281
+ was 20.05 and 19.69, respectively. For perplexity,
282
+ lower is better. Thus, both neural network types
283
+ outperformed the strong baseline of a smoothed
284
+ 5-gram model, showing that they performed well
285
+ at capturing the basic statistics of their training
286
+ domain.5
287
+ 4.5
288
+ General Syntactic Evaluation
289
+ As an additional way to check the validity of our
290
+ setup, we evaluated our models on the Zorro dataset
291
+ (Huebner et al., 2021), which is based on BLiMP
292
+ (Warstadt et al., 2020a). Zorro contains 24 evalu-
293
+ ations, each of which targets one syntactic phe-
294
+ nomenon (e.g., subject-verb agreement) and in-
295
+ volves sentence pairs for which one sentence is
296
+ grammatical, and the other is minimally different
297
+ 5For an intuitive illustration of our model quality, see the
298
+ sample text generated by them in Appendix H.
299
+
300
+ but ungrammatical (e.g., by violating subject verb
301
+ agreement). A model is said to get a sentence
302
+ pair correct if it assigns a higher probability to the
303
+ grammatical sentence than the ungrammatical one.
304
+ Huebner et al. (2021) showed that Transformers
305
+ trained on CHILDES data can perform well on
306
+ many of the Zorro categories, so if our setup is
307
+ sound, our own models should also perform well
308
+ on Zorro.
309
+ See Appendix D for full results. For each syntac-
310
+ tic phenomenon, most model re-runs scored above
311
+ 0.9, though at least one scored near the chance level
312
+ of 0.5. For each re-run of each architecture there
313
+ is at least one phenomenon for which the model
314
+ scores over 0.97, and many models score 1.00 on
315
+ some phenomena. Thus, all models score well on
316
+ at least some syntactic evaluations, attaining results
317
+ comparable to those of Huebner et al. (2021) and
318
+ providing additional support for the validity of our
319
+ setup. We now test whether these models have also
320
+ successfully learned the specific phenomenon that
321
+ we focus on, yes/no questions—a phenomenon not
322
+ included in the Zorro dataset.
323
+ 4.6
324
+ Yes/No Questions
325
+ Evaluation Dataset: Forced-Choice Acceptabil-
326
+ ity Judgments
327
+ As a first way to test whether our
328
+ models have learned HIERARCHICALQ, we eval-
329
+ uate whether they assign higher probabilities to
330
+ sentences consistent with HIERARCHICALQ than
331
+ to minimally different sentences that are ungram-
332
+ matical. For this purpose, we create an evaluation
333
+ dataset containing groups of 6 questions, each cre-
334
+ ated by starting with a declarative sentence, such
335
+ as (5), and then deleting the first, main, or neither
336
+ auxiliary, and inserting the first or main auxiliary
337
+ at the front of the sentence.6 For instance, in (6b),
338
+ the first auxiliary has been preposed, and the main
339
+ auxiliary has been deleted.
340
+ (5)
341
+ The dog who has seen a boy did try.
342
+ (6)
343
+ a. Has the dog who seen a boy did try?
344
+ b. Has the dog who has seen a boy try?
345
+ c. Has the dog who has seen a boy did try ?
346
+ d. Did the dog who seen a boy did try?
347
+ e. Did the dog who has seen a boy try?
348
+ f. Did the dog who has seen a boy did try?
349
+ 6It would be possible to also use a ‘prepose other’ category,
350
+ where an auxiliary not in the input is inserted (McCoy et al.,
351
+ 2018). We excluded this category because using it would raise
352
+ complications about which ‘other’ auxiliary to choose.
353
+ Within each group, we evaluate which question
354
+ the model assigned the highest probability to. If a
355
+ model has correctly learned HIERARCHICALQ, it
356
+ should assign the highest probability to the question
357
+ consistent with this rule, such as (6e).
358
+ Several past papers about yes/no questions have
359
+ used the same general approach (Lewis and El-
360
+ man, 2001; Reali and Christiansen, 2005). How-
361
+ ever, these papers considered only pairs of sen-
362
+ tences, whereas we consider groups of 6 to allow
363
+ for a wider range of possible generalizations that a
364
+ model might have learned.
365
+ To generate the declaratives from which we
366
+ formed groups of 6 questions, we used the context-
367
+ free grammar (CFG) in Appendix F, which has a vo-
368
+ cabulary selected from the most common words in
369
+ CHILDES. Each declarative generated by the CFG
370
+ (e.g., (5)) contains two auxiliary verbs: one before
371
+ the sentence’s main verb and one inside a relative
372
+ clause modifying the subject. One potential prob-
373
+ lem is that some questions are consistent with both
374
+ HIERARCHICALQ and LINEARQ. For instance,
375
+ (7a) can be formed from (7b) with the HIERARCHI-
376
+ CALQ-consistent steps PREPOSE-MAIN,DELETE-
377
+ MAIN, or from (7c) with the LINEARQ-consistent
378
+ steps PREPOSE-FIRST,DELETE-MAIN.
379
+ (7)
380
+ a. Did the boy who did see the person laugh?
381
+ b. The boy who did see the person did laugh.
382
+ c. The boy who did see the person can laugh.
383
+ To avoid this problem, we required that the aux-
384
+ iliary before the main verb must select for a dif-
385
+ ferent verb inflection than the one in the relative
386
+ clause. For instance in (5), did selects for the verb’s
387
+ bare form, while has selects for the past participle
388
+ form. Thus, the auxiliary at the start of the question
389
+ could only correspond to whichever auxiliary in the
390
+ declarative has the same selectional properties.7
391
+ Results: Relative Question Acceptability
392
+ For
393
+ each sentence group, we used per-word perplex-
394
+ ity to see which of the 6 candidates the models
395
+ scored most highly.8 For both LSTMs and Trans-
396
+ formers, the correct category (PREPOSE MAIN,
397
+ DELETE MAIN) was the second-rarest choice, and
398
+ 7A model could succeed on this dataset with a rule that
399
+ relates the auxiliary at the start of a question with the last
400
+ auxiliary in the declarative form. Since our models fail on this
401
+ dataset, this consideration is not relevant here.
402
+ 8We also explored evaluation of the models with a more
403
+ complex measure called SLOR where we additionally nor-
404
+ malized scores by word frequency (Pauls and Klein, 2012).
405
+ Both metrics produced qualitatively similar results, so we only
406
+ report the simpler metric here. See Appendix C.1.
407
+
408
+ Prepose First
409
+ Prepose Main
410
+ Delete First
411
+ Delete Main
412
+ Delete none
413
+ LSTM
414
+ Transformer
415
+ LSTM
416
+ Transformer
417
+ 0.0
418
+ 0.5
419
+ 1.0
420
+ 0.0
421
+ 0.5
422
+ 1.0
423
+ 0.0
424
+ 0.5
425
+ 1.0
426
+ Preference for question type
427
+ Declarative sentence: The person who has seen this boy did try.
428
+ Has the person who seen
429
+ this boy did try?
430
+ Did the person who seen
431
+ this boy did try?
432
+ Has the person who has
433
+ seen this boy try?
434
+ Did the person who has
435
+ seen this boy try?
436
+ Has the person who has
437
+ seen this boy did try?
438
+ Did the person who has
439
+ seen this boy did try?
440
+ Figure 1: The question types that models prefer when
441
+ offered a choice between 6 questions. These 6 ques-
442
+ tions are formed by modifying a declarative with a rel-
443
+ ative clause on the subject according to ‘prepose’ and
444
+ ‘delete’ rules. The correct category is PREPOSE MAIN,
445
+ DELETE MAIN. Within each architecture, the propor-
446
+ tions across all 6 question types necessarily sum to 1.
447
+ Each bar shows the average across 10 model re-runs,
448
+ with single-standard-deviation error bars.
449
+ the most frequent preference was for PREPOSE
450
+ FIRST, DELETE MAIN, a category that is only par-
451
+ tially correct because it references linear order in
452
+ addition to hierarchical structure. (Figure 1).
453
+ Thus, neither model displays preferences con-
454
+ sistent with the correct, fully-hierarchical gener-
455
+ alization. The two model types showed similar
456
+ scores, which may mean that these results are
457
+ largely driven by the statistics of the training data
458
+ that both models share, rather than the models’ dif-
459
+ fering inductive biases.
460
+ One of the incorrect categories—PREPOSE
461
+ MAIN, DELETE NONE, such as (6f)—only re-
462
+ quires reference to hierarchical structure, so it
463
+ could be said to capture the hierarchical nature of
464
+ yes/no questions. Nonetheless, this category was
465
+ also relatively rare: combining the two fully hier-
466
+ archical possibilities (PREPOSE MAIN, DELETE
467
+ MAIN and PREPOSE MAIN, DELETE NONE) ac-
468
+ counts for only 26% of LSTM preferences and
469
+ 27% of Transformer preferences, meaning that both
470
+ models over 70% of the time favored a sentence
471
+ generated at least partially based on linear order.
472
+ There are two likely reasons for why our models
473
+ performed so poorly on yes-no questions when they
474
+ performed well on many of the phenomena in the
475
+ Zorro dataset (Section 4.5). First, yes/no questions
476
+ may simply be harder to learn than the other phe-
477
+ nomena; indeed, yes/no questions are often singled
478
+ out as being likely to pose difficulties for a general-
479
+ purpose learner (Section 1). Alternatively, it might
480
+ be that the six-way evaluation we used for yes/no
481
+ questions is stricter than the binary judgments used
482
+ for the Zorro dataset.
483
+ 5
484
+ Experiment 2: Question Formation
485
+ The previous experiment was designed to operate
486
+ entirely in the next-word-prediction paradigm, mo-
487
+ tivated by arguments from past literature about
488
+ the strength and relative ecological validity of
489
+ next-word-prediction as a training objective (see
490
+ Section 4.2).
491
+ However, one of this setup’s
492
+ shortcomings is that HIERARCHICALQ describes
493
+ correspondences between questions and declara-
494
+ tives, but Experiment 1 focused on questions alone,
495
+ with no consideration of declaratives.
496
+ In this second experiment, to better capture that
497
+ HIERARCHICALQ is defined over sentence pairs,
498
+ we trained models on a sentence-pair task: trans-
499
+ forming a declarative into a question (McCoy et al.,
500
+ 2020). For instance, given the child did learn the
501
+ model must produce did the child learn ?
502
+ We evaluated models in two ways. First, we
503
+ checked if the models’ predictions fully matched
504
+ the correct questions. This full-sentence evaluation
505
+ is demanding, and models might fail this evalua-
506
+ tion for reasons unrelated to our core hypotheses.
507
+ For instance, given the child did learn the model
508
+ might produce did the baby learn, which would be
509
+ marked as incorrect, even though this lexical error
510
+ is not relevant to HIERARCHICALQ.
511
+ As a metric that is less demanding and that also
512
+ more directly targets HIERARCHICALQ, we mea-
513
+ sured if the first word of the output question corre-
514
+ sponded to the first or main auxiliary of the input.
515
+ Critically, LINEARQ and HIERARCHICALQ make
516
+ different predictions for the first word of a question
517
+ so long as the two auxiliaries are distinct: see (4).
518
+ Because this framing lets the model freely generate
519
+ its output (instead of choosing one option from a
520
+ pre-specified set), we allow for the possibility that
521
+ the rule learned by models may not be identical to
522
+ any of our manually-generated hypotheses.
523
+ Solely training models to perform this transfor-
524
+ mation involves the implicit assumption that, when
525
+ children acquire English yes/no questions, the only
526
+ evidence they leverage is English yes/no questions.
527
+ However, other types of sentences may also pro-
528
+ vide useful evidence (Pearl and Mis, 2016): e.g.,
529
+ wh-questions also illustrate subject-auxiliary in-
530
+
531
+ version (Pullum and Scholz, 2002), while, more
532
+ generally, many types of sentences could provide
533
+ evidence that the syntax as a whole is hierarchical
534
+ (Perfors et al., 2011). To explore this possibility,
535
+ we compared a condition in which models were
536
+ only trained to perform question formation (the
537
+ QUESTION FORMATION condition) to another in
538
+ which models were first pre-trained on next-word
539
+ prediction with the exact same setup as in Experi-
540
+ ment 1 before being further trained to perform ques-
541
+ tion formation (the NEXT-WORD PREDICTION +
542
+ QUESTION FORMATION condition).
543
+ 5.1
544
+ Dataset
545
+ Training Set
546
+ Our question formation dataset con-
547
+ sisted of the yes/no questions in the CHILDES
548
+ Treebank (Pearl and Sprouse, 2013a,b), a parsed
549
+ subset of CHILDES containing 189,359 sentences.
550
+ We used these parses to extract all yes/no ques-
551
+ tions from the CHILDES Treebank and derive their
552
+ corresponding declarative forms. The resulting
553
+ declarative was concatenated with the question. An
554
+ example declarative/question pair is:
555
+ (8)
556
+ you can spell your name .
557
+ can you
558
+ spell your name ?
559
+ The training set consisted of 10,870 declara-
560
+ tive/question pairs, the validation set 1,360 pairs,
561
+ and the test set 1,358 pairs (we will call this test
562
+ set the randomly-partitioned test set to distinguish
563
+ it from two other evaluation sets discussed below).
564
+ We trained models to perform next-word prediction
565
+ on such concatenated sentence pairs.
566
+ The first-word accuracy of the trained model
567
+ was then computed based on the model’s predic-
568
+ tion for the word after the period in each test exam-
569
+ ple, while the full-sentence accuracy was computed
570
+ based on its predictions for all tokens after the pe-
571
+ riod. All questions in the randomly-partitioned test
572
+ set were withheld from both the question-formation
573
+ training set and the next-word-prediction training
574
+ set. Thus, models had not seen these test examples
575
+ in their training, even in the NEXT-WORD PRE-
576
+ DICTION + QUESTION FORMATION condition in
577
+ which they were trained on both tasks.
578
+ Evaluation Sets
579
+ In addition to the randomly-
580
+ partitioned test set, we used CFGs to generate two
581
+ targeted evaluation sets. As in Experiment 1, we se-
582
+ lected the CFGs’ vocabulary from common words
583
+ in our CHILDES data. In sentences generated from
584
+ the first CFG, the sentence’s first auxiliary was also
585
+ its main auxiliary, so LINEARQ and HIERARCHI-
586
+ CALQ make the same predictions. (8) exemplifies
587
+ the type of declarative-question pair in this dataset.
588
+ We call this dataset FIRST-AUX = MAIN-AUX. For
589
+ sentences generated by the second CFG, the main
590
+ auxiliary was the second auxiliary in the sentence;
591
+ thus, these examples disambiguate LINEARQ and
592
+ HIERARCHICALQ. Example (9) is a declarative-
593
+ question pair from this evaluation set.
594
+ (9) a boy who is playing can try .
595
+ can a
596
+ boy who is playing try ?
597
+ We call this dataset FIRST-AUX ̸= MAIN-AUX.
598
+ See Appendix F for the CFGs used. We sampled
599
+ 10,000 declarative sentences from these grammars
600
+ and transformed them into questions according to
601
+ HIERARCHICALQ to create our evaluation sets.
602
+ 5.2
603
+ Results
604
+ Randomly-Partitioned Test Set
605
+ The LSTMs
606
+ and Transformers in the QUESTION FORMA-
607
+ TION condition performed well on the randomly-
608
+ partitioned test set, with a full-question accuracy
609
+ of 0.68 ± 0.014 and 0.87 ± 0.005 (averaged across
610
+ 10 reruns with margins indicating one standard de-
611
+ viation). The models in the NEXT-WORD PRE-
612
+ DICTION + QUESTION FORMATION condition per-
613
+ formed similarly well, with a full-question accu-
614
+ racy of 0.66 ± 0.008 for the LSTMs and 0.93 ±
615
+ 0.004 for the Transformers. For both model types,
616
+ the first-word accuracy for the question was nearly
617
+ 1.00 across re-runs. We suspect that Transform-
618
+ ers have a stronger full-question accuracy because
619
+ producing the question requires copying all words
620
+ from the declarative (but in a different order). Copy-
621
+ ing is likely easy for Transformers because they can
622
+ attend to specific words in the prior context, while
623
+ our LSTMs must compress the entire context into a
624
+ fixed-size vector, which may degrade the individual
625
+ word representations. Because both model types
626
+ achieved near-perfect performance on the crucial
627
+ first-word accuracy metric, we conclude that our
628
+ models have successfully learned how to handle
629
+ the types of declarative/question pairs that we ex-
630
+ tracted from the CHILDES Treebank.
631
+ Targeted Evaluation Sets
632
+ On our two targeted
633
+ evaluation sets, models almost never produced the
634
+ complete question correctly. Turning to the more
635
+ lenient measure of first-word accuracy, for exam-
636
+ ples on which LINEARQ and HIERARCHICALQ
637
+ predict the same first output word (FIRST-AUX =
638
+ MAIN-AUX), the Transformer trained only on ques-
639
+ tion formation performed strongly, while the Trans-
640
+
641
+ LSTM
642
+ Transformer
643
+ First-Aux = Main-Aux
644
+ First-Aux ≠ Main-Aux
645
+ HierarchicalQ
646
+ & LinearQ
647
+ HierarchicalQ
648
+ Only
649
+ LinearQ
650
+ Only
651
+ HierarchicalQ
652
+ & LinearQ
653
+ HierarchicalQ
654
+ Only
655
+ LinearQ
656
+ Only
657
+ 0.00
658
+ 0.25
659
+ 0.50
660
+ 0.75
661
+ 1.00
662
+ 0.00
663
+ 0.25
664
+ 0.50
665
+ 0.75
666
+ 1.00
667
+ Consistency with rule(s),
668
+ based on first word of question
669
+ Condition
670
+ Question Formation
671
+ Next-Word Prediction
672
+ + Question Formation
673
+ Figure 2: Proportion of model-produced questions that
674
+ were consistent with the linear rule LINEARQ and/or
675
+ the hierarchical rule HIERARCHICALQ. In the FIRST-
676
+ AUX = MAIN-AUX dataset, the first auxiliary is the
677
+ main auxiliary, so both LINEARQ and HIERARCHI-
678
+ CALQ produce the correct question string. The FIRST-
679
+ AUX ̸= MAIN-AUX dataset disambiguates the two
680
+ rules. Each bar shows the average across 10 model re-
681
+ runs, with error bars showing one standard deviation.
682
+ former trained on both tasks, and both LSTMs,
683
+ performed reasonably well (Figure 2; note mod-
684
+ els could choose any word in their vocabulary to
685
+ begin the output, so chance performance is near
686
+ 0.00). For the crucial cases that disambiguate the
687
+ two rules (FIRST-AUX ̸= MAIN-AUX), both mod-
688
+ els in both conditions performed more consistently
689
+ with LINEARQ than HIERARCHICALQ. Training
690
+ on next-word prediction before question formation
691
+ had inconsistent effects: it modestly increased the
692
+ likelihood of hierarchical generalization in LSTMs,
693
+ yet it decreased that likelihood in Transformers.
694
+ Lexical Specificity
695
+ In Appendix G, we further
696
+ break down the FIRST-AUX ̸= MAIN-AUX results
697
+ based the auxiliaries’ identity. The generalization
698
+ pattern varied considerably across auxiliary pairs.
699
+ For some auxiliary pairs, the auxiliary chosen to
700
+ begin the question was usually neither auxiliary
701
+ in the input (Figure 3, left facet). For other pairs,
702
+ models usually chose the first auxiliary, regardless
703
+ of lexical identity (Figure 3, middle facet). Finally,
704
+ for some pairs, the auxiliary chosen was usually
705
+ the same one, regardless of whether it was the first
706
+ or main auxiliary (Figure 3, right facet).
707
+ Generalization based on lexical identity is rarely
708
+ considered in past discussions of English yes/no
709
+ question acquisition. Of the papers on this phe-
710
+ nomenon (see Clark and Lappin (2010), Lasnik
711
+ and Lidz (2017), and Pearl (2021) for overviews),
712
+ the only one to our knowledge that discusses lexi-
713
+ have and has
714
+ can and do
715
+ have and did
716
+ Move−first−aux
717
+ Move−main−aux
718
+ Move−have
719
+ Move−has
720
+ Move−first−aux
721
+ Move−main−aux
722
+ Move−can
723
+ Move−do
724
+ Move−first−aux
725
+ Move−main−aux
726
+ Move−have
727
+ Move−did
728
+ 0.0
729
+ 0.5
730
+ 1.0
731
+ First word behavior
732
+ consistent with rule
733
+ Comparison
734
+ First−vs−main
735
+ Aux−vs−Aux
736
+ Figure 3: Lexical specificity in model behavior. Each
737
+ facet considers only the evaluation examples contain-
738
+ ing the two auxiliaries in the facet heading; e.g., the
739
+ can and do facet includes, for example, the inputs the
740
+ children who can play do learn and the children who
741
+ do play can learn. The bars show the proportion of
742
+ model predictions for the first word of the output that
743
+ are consistent with four potential movement rules, aver-
744
+ aged across 10 model re-runs and with error bars show-
745
+ ing one standard deviation above and below the mean.
746
+ This plot only shows an illustrative subset of auxiliary
747
+ pairs for one model type (Transformers in the NEXT-
748
+ WORD PREDICTION + QUESTION FORMATION con-
749
+ dition); see Appendix G for the full results.
750
+ cal specificity is Frank and Mathis (2007), which
751
+ studied models trained on synthetic data. Our re-
752
+ sults highlight the importance of testing for a broad
753
+ range of generalizations: Lexically-specific hy-
754
+ potheses appear attractive for our low-bias learners,
755
+ so an account of what biases can yield human-like
756
+ learning should rule out these lexically-specific hy-
757
+ potheses along with linear ones.
758
+ 6
759
+ Discussion
760
+ We have found that, when trained on child-directed
761
+ speech, two types of standard neural networks per-
762
+ formed reasonably well at capturing the statistical
763
+ properties of the dataset, yet their handling of En-
764
+ glish yes/no questions was more consistent with
765
+ a linear rule LINEARQ than the correct hierarchi-
766
+ cal rule HIERARCHICALQ. These results support
767
+ the hypothesis that a learner requires a hierarchical
768
+ bias to consistently learn hierarchical rules when
769
+ learning from the linguistic data children receive.
770
+ 6.1
771
+ Takeaways for LSTMs and Transformers
772
+ When trained on massive corpora, LSTMs and
773
+ Transformers perform impressively on some syn-
774
+ tactic evaluations. Based on such results, it is tempt-
775
+ ing to conclude that the general-purpose biases of
776
+ these architectures suffice to yield human-like syn-
777
+
778
+ tax acquisition. Our results caution against this
779
+ interpretation: When we trained the same architec-
780
+ tures on data more similar to children’s input, they
781
+ failed to learn the structure of English yes/no ques-
782
+ tions. Thus, at least when learning from text alone,
783
+ LSTMs and Transformers do not display human-
784
+ like language learning—they do not generalize as
785
+ humans do from the data that humans receive.
786
+ 6.2
787
+ Takeaways for the Poverty of the
788
+ Stimulus Debate
789
+ Below we specify four possible positions in the
790
+ poverty-of-the-stimulus debate about the adequacy
791
+ of children’s input for inducing hierarchical rules in
792
+ low-bias learners, arranged from assuming the most
793
+ limited to the most expansive innate component:
794
+ (10) Any inductive biases: Any learner trained on
795
+ CHILDES will generalize like humans do.
796
+ (11) Any
797
+ inductive
798
+ biases
799
+ that
800
+ enable
801
+ in-
802
+ distribution learning: Any learner that cap-
803
+ tures the statistical patterns of the training dis-
804
+ tribution will generalize to HIERARCHICALQ.
805
+ (12) Some non-hierarchical inductive biases:
806
+ Some general-purpose learners will generalize
807
+ as humans do, but others will not.
808
+ (13) Only a hierarchical inductive bias:
809
+ No
810
+ general-purpose learners will generalize as
811
+ humans do: hierarchical biases are necessary.
812
+ Position (10) is clearly false: many learners can-
813
+ not learn certain aspects of syntax, no matter their
814
+ training data (e.g., bigram models cannot capture
815
+ long-distance dependencies). Our work shows that
816
+ position (11) is also false: Though our models per-
817
+ formed well on the in-distribution test sets of Exper-
818
+ iments 1 and 2, they did not generalize in human-
819
+ like ways. This leaves positions (12) and (13),
820
+ which our existing results cannot differentiate. It is
821
+ possible that only learners with hierarchical induc-
822
+ tive biases can demonstrate human-like language
823
+ learning (position (13)), but also that some learners
824
+ without this bias can succeed (position (12))—just
825
+ not the learners we tested. For further discussion
826
+ of how computational modeling can bear on learn-
827
+ ability arguments, see Wilcox et al. (2021).
828
+ One potential solution supporting position (12)
829
+ would be that learners leverage the hierarchical
830
+ structure of some syntactic phenomenon to help
831
+ conclude that other, impoverished phenomena are
832
+ hierarchical (Perfors et al., 2011; Mulligan et al.,
833
+ 2021). However, our results from Experiment 2
834
+ show that giving learners access to a wider range
835
+ of phenomena does not automatically improve hi-
836
+ erarchical generalization: Models’ performance on
837
+ question formation was not substantially improved
838
+ (and in some cases was even harmed) when they
839
+ were trained not just on question formation but also
840
+ on next-word prediction on the entire CHILDES
841
+ corpus. Thus, although training on text that con-
842
+ tains many linguistic phenomena can give mod-
843
+ els a hierarchical inductive bias when the training
844
+ is done over large Internet corpora (Warstadt and
845
+ Bowman, 2020; Mueller et al., 2022), our results
846
+ provide evidence that this conclusion does not ex-
847
+ tend to models trained on child-directed speech.
848
+ Though both (12) and (13) remain as possibil-
849
+ ities, we believe that our results more strongly
850
+ support (13). Of all currently available general-
851
+ purpose learners, LSTMs and Transformers are the
852
+ best at modeling the probabilistic structure of lin-
853
+ guistic data. Therefore, if child-directed speech
854
+ contains clear evidence for the hierarchical nature
855
+ of yes/no questions—evidence so clear that at least
856
+ some general-purpose learners could recognize it—
857
+ it is likely that LSTMs and Transformers would
858
+ be among the set of general-purpose learners that
859
+ could use this evidence to make hierarchical gener-
860
+ alizations in our experiments. The fact that these
861
+ architectures instead predominantly favored linear
862
+ generalizations therefore supports position (13).
863
+ 6.3
864
+ How to test for HIERARCHICALQ
865
+ We have argued that an ideal simulation of the
866
+ acquisition of English yes/no questions would have
867
+ the following properties:
868
+ (14) The training data should be similar to chil-
869
+ dren’s linguistic input.
870
+ (15) The training task should be ecologically valid.
871
+ (16) The evaluation method should focus on corre-
872
+ spondences between pairs of sentences rather
873
+ than the acceptability of individual sentences.
874
+ Property (14) motivated our use of text from
875
+ CHILDES as the training data. We are not aware
876
+ of a single experimental setup that fully satisfies
877
+ both Property (15) and Property (16), so we instead
878
+ used two experiments, each one focusing on one
879
+ property at the cost of satisfying the other one less
880
+ well. Experiment 1 works entirely in the context
881
+ of the relatively ecologically valid task of next-
882
+ word prediction, motivated by Property (15), but its
883
+
884
+ evaluation is only based on the acceptability of in-
885
+ dividual sentences, failing to satisfy Property (16).
886
+ Experiment 2 fully satisfies Property (16) by using
887
+ an evaluation based on sentence pairs, at the cost of
888
+ including a less ecologically-valid training compo-
889
+ nent based on sentence transformations. Both ex-
890
+ periments yielded qualitatively similar conclusions
891
+ (failure of models to learn HIERARCHICALQ).
892
+ 6.4
893
+ Quantity of Training Data
894
+ The size of our training set was plausibly within
895
+ the range from which children can acquire HIER-
896
+ ARCHICALQ. Crain and Nakayama (1987) found
897
+ that children between ages 3 and 5 behaved much
898
+ more consistently with HIERARCHICALQ than
899
+ LINEARQ. Though these children made many er-
900
+ rors, their errors were usually compatible with a
901
+ hierarchical rule (e.g., PREPOSE MAIN, DELETE
902
+ NONE errors: see Section 4.6). By age 3, Ameri-
903
+ can children receive approximately 10 to 33 mil-
904
+ lion words of input (Hart and Risley, 1995), and
905
+ the 8.5 million words of our training set is close
906
+ to the lower end of that range. Thus, it is reason-
907
+ able to suppose that a learner that generalizes as
908
+ children do would favor HIERARCHICALQ after
909
+ being trained on our training set. Our models, in
910
+ contrast, regularly preferred sentences generated
911
+ in ways based on linear order (Figures 1 and 2), a
912
+ category of error that is very rare in children (Crain
913
+ and Nakayama, 1987; Ambridge et al., 2008).
914
+ In order to give our models the strongest chance
915
+ of generalizing correctly, it would have been ideal
916
+ to provide a quantity of data closer to 33 million
917
+ words, the high end of Hart and Risley’s range. Our
918
+ data source did not contain enough text to make this
919
+ possible, but future work could investigate ways to
920
+ augment the data using other sources.
921
+ 6.5
922
+ Type of Training Data
923
+ Our training set was both qualitatively and quanti-
924
+ tatively closer to children’s input than the massive
925
+ Internet corpora standardly used to train models in
926
+ NLP (Linzen, 2020). This difference is important:
927
+ Lin et al. (2019), Warstadt and Bowman (2020),
928
+ and Mueller et al. (2022) all found evidence that
929
+ models trained on large Internet corpora performed
930
+ well on yes/no questions evaluations, whereas our
931
+ models trained on CHILDES performed poorly—
932
+ though we cannot be certain the differences in re-
933
+ sults are solely due to differences in the training
934
+ data, since these prior papers used different model
935
+ architectures, training tasks, and evaluation setups.
936
+ Though our training data are more similar to
937
+ children’s input than massive Internet corpora are,
938
+ differences remain. Our experiments omit several
939
+ aspects of a child’s experience that might help them
940
+ acquire syntax, such as prosody (Morgan and De-
941
+ muth, 1996), visual information (Shi et al., 2019),
942
+ and meaning (Fitz and Chang, 2017; Abend et al.,
943
+ 2017), all of which might correlate with syntac-
944
+ tic structure and thus provide cues to the correct
945
+ hierarchical generalization. On the other hand,
946
+ our dataset might present an easier learning sce-
947
+ nario than children are faced with, because chil-
948
+ dren must learn to segment the speech stream into
949
+ words (Lakhotia et al., 2021), while our models do
950
+ not need to. Further, though real-world grounding
951
+ could provide helpful information, learners might
952
+ struggle to leverage this information due to diffi-
953
+ culty determining what is being discussed in the
954
+ physical world (Gleitman et al., 2005).
955
+ 7
956
+ Conclusion
957
+ In this work, we trained two types of neural net-
958
+ works (LSTMs and Transformers) on sentences of
959
+ the types available to children and then analyzed
960
+ what they had learned about English yes/no ques-
961
+ tions. Across several evaluation paradigms, these
962
+ models failed to generalize in human-like ways:
963
+ Humans display hierarchical generalization, while
964
+ the models’ generalization was instead based on
965
+ linear order and individual words’ identities. Our
966
+ results support the hypothesis that human-like lin-
967
+ guistic generalization requires biases stronger than
968
+ those of LSTMs and Transformers. Future work
969
+ should investigate what inductive biases enable suc-
970
+ cessful generalization. One approach would be to
971
+ test architectures with built-in hierarchical struc-
972
+ ture; past work has shown that such architectures
973
+ have a hierarchical bias (McCoy et al., 2020) and
974
+ generalize better on the hierarchical phenomenon
975
+ of subject-verb agreement (Kuncoro et al., 2018;
976
+ Lepori et al., 2020), so they may also generalize bet-
977
+ ter on English yes/no questions. A final direction
978
+ would be to expand the input beyond words alone
979
+ so that learners can leverage hierarchical structure
980
+ that is present in other modalities, such as hierar-
981
+ chical structure in visual scenes.
982
+ Ethics Statement
983
+ Use of human data:
984
+ While we did not collect
985
+ any new human data ourselves, many of our anal-
986
+ yses involved the use of prior datasets within the
987
+
988
+ CHILDES database. All of these datasets were
989
+ collected in accordance with IRB policies at the
990
+ institutions of the data collectors, and all followed
991
+ standard practices in obtaining informed consent
992
+ and deidentifying data.9
993
+ Risks and limitations:
994
+ The main risk of our pro-
995
+ posed analyses is that future work using the same
996
+ analyses might draw overly strong conclusions
997
+ based on increased model performance, leading
998
+ to overestimates of model strength. Such overesti-
999
+ mates are an issue because they can lead users to
1000
+ place more trust in a model than is warranted.
1001
+ To clarify, we view strong performance on our
1002
+ evaluation datasets as necessary but not sufficient to
1003
+ demonstrate human-like learning. Thus, if models
1004
+ perform poorly on our datasets (as the models we
1005
+ evaluated did), then we have strong reason to con-
1006
+ clude that models are not learning in human-like
1007
+ ways. If future models perform better, such results
1008
+ would be consistent with human-like learning but
1009
+ would not conclusively establish that models learn
1010
+ as humans do, as they might instead be using some
1011
+ shallow heuristic that is not controlled for in our
1012
+ datasets. In other words, a criterion that is neces-
1013
+ sary but not sufficient facilitates strong conclusions
1014
+ about failure but does not facilitate strong conclu-
1015
+ sions about success. If future papers are faced with
1016
+ models that are more successful, such papers would
1017
+ ideally supplement results based on our datasets
1018
+ with analyses of models’ internal strategies in order
1019
+ to more conclusively establish that what they have
1020
+ learned is not a spurious heuristic.
1021
+ References
1022
+ Omri Abend, Tom Kwiatkowski, Nathaniel J Smith,
1023
+ Sharon Goldwater, and Mark Steedman. 2017. Boot-
1024
+ strapping language acquisition. Cognition, 164:116–
1025
+ 143.
1026
+ Gerry TM Altmann and Yuki Kamide. 1999. Incremen-
1027
+ tal interpretation at verbs: Restricting the domain of
1028
+ subsequent reference. Cognition, 73(3):247–264.
1029
+ Ben Ambridge, Caroline F Rowland, and Julian M
1030
+ Pine. 2008. Is structure dependence an innate con-
1031
+ straint? New experimental evidence from children’s
1032
+ complex-question production.
1033
+ Cognitive Science,
1034
+ 32(1):222–255.
1035
+ Robert Berwick, Paul Pietroski, Beracah Yankama, and
1036
+ Noam Chomsky. 2011. Poverty of the stimulus re-
1037
+ visited. Cognitive science, 35:1207–42.
1038
+ 9https://talkbank.org/share/irb/
1039
+ Rens Bod, Margaux Smets, et al. 2012. Empiricist so-
1040
+ lutions to nativist problems using tree-substitution
1041
+ grammars. Workshop on Computational Models of
1042
+ Language Acquisition and Loss: EACL.
1043
+ Noam Chomsky. 1965. Aspects of the Theory of Syntax,
1044
+ 50 edition. The MIT Press.
1045
+ Noam Chomsky. 1980.
1046
+ Rules and representations.
1047
+ Columbia University Press.
1048
+ Alexander Clark and Rémi Eyraud. 2007. Polynomial
1049
+ identification in the limit of substitutable context-
1050
+ free languages.
1051
+ Journal of Machine Learning Re-
1052
+ search, 8(8).
1053
+ Alexander Clark and Shalom Lappin. 2010. Linguis-
1054
+ tic Nativism and the Poverty of the Stimulus. John
1055
+ Wiley & Sons.
1056
+ Stephen Crain and Mineharu Nakayama. 1987. Struc-
1057
+ ture dependence in grammar formation. Language,
1058
+ pages 522–543.
1059
+ Jeffrey L Elman. 1991.
1060
+ Distributed representations,
1061
+ simple recurrent networks, and grammatical struc-
1062
+ ture. Machine learning, 7(2):195–225.
1063
+ Hartmut Fitz and Franklin Chang. 2017. Meaningful
1064
+ questions: The acquisition of auxiliary inversion in
1065
+ a connectionist model of sentence production. Cog-
1066
+ nition, 166:225–250.
1067
+ Robert Frank and Donald Mathis. 2007. Transforma-
1068
+ tional networks. Models of Human Language Acqui-
1069
+ sition, 22.
1070
+ Lila R Gleitman, Kimberly Cassidy, Rebecca Nappa,
1071
+ Anna Papafragou, and John C Trueswell. 2005.
1072
+ Hard words. Language learning and development,
1073
+ 1(1):23–64.
1074
+ Kristina Gulordava, Piotr Bojanowski, Edouard Grave,
1075
+ Tal Linzen, and Marco Baroni. 2018.
1076
+ Colorless
1077
+ green recurrent networks dream hierarchically.
1078
+ John Hale. 2001. A probabilistic Earley parser as a psy-
1079
+ cholinguistic model. In Second Meeting of the North
1080
+ American Chapter of the Association for Computa-
1081
+ tional Linguistics.
1082
+ Betty Hart and Todd R Risley. 1995. Meaningful differ-
1083
+ ences in the everyday experience of young American
1084
+ children. Paul H Brookes Publishing.
1085
+ Kenneth Heafield. 2011. KenLM: Faster and smaller
1086
+ language model queries.
1087
+ In Proceedings of the
1088
+ Sixth Workshop on Statistical Machine Translation,
1089
+ pages 187–197, Edinburgh, Scotland. Association
1090
+ for Computational Linguistics.
1091
+ Sepp Hochreiter and Jürgen Schmidhuber. 1997.
1092
+ Long short-term memory.
1093
+ Neural computation,
1094
+ 9(8):1735–1780.
1095
+
1096
+ Jennifer Hu, Jon Gauthier, Peng Qian, Ethan Wilcox,
1097
+ and Roger Levy. 2020.
1098
+ A systematic assessment
1099
+ of syntactic generalization in neural language mod-
1100
+ els.
1101
+ In Proceedings of the 58th Annual Meeting
1102
+ of the Association for Computational Linguistics,
1103
+ pages 1725–1744, Online. Association for Compu-
1104
+ tational Linguistics.
1105
+ Philip A. Huebner, Elior Sulem, Cynthia Fisher, and
1106
+ Dan Roth. 2021. BabyBERTa: Learning more gram-
1107
+ mar with small-scale child-directed language.
1108
+ In
1109
+ Proceedings of CoNLL.
1110
+ Xuân-Nga Cao Kam, Iglika Stoyneshka, Lidiya Torny-
1111
+ ova, Janet D Fodor, and William G Sakas. 2008. Bi-
1112
+ grams and the richness of the stimulus. Cognitive
1113
+ Science, 32(4):771–787.
1114
+ Reinhard Kneser and Hermann Ney. 1995. Improved
1115
+ backing-off for m-gram language modeling. 1995
1116
+ International Conference on Acoustics, Speech, and
1117
+ Signal Processing, 1:181–184 vol.1.
1118
+ Adhiguna Kuncoro, Chris Dyer, John Hale, Dani Yo-
1119
+ gatama, Stephen Clark, and Phil Blunsom. 2018.
1120
+ LSTMs can learn syntax-sensitive dependencies
1121
+ well, but modeling structure makes them better. In
1122
+ Proceedings of the 56th Annual Meeting of the As-
1123
+ sociation for Computational Linguistics (Volume 1:
1124
+ Long Papers), pages 1426–1436, Melbourne, Aus-
1125
+ tralia. Association for Computational Linguistics.
1126
+ Marta Kutas, Katherine A DeLong, and Nathaniel J
1127
+ Smith. 2011. A look around at what lies ahead: Pre-
1128
+ diction and predictability in language processing. In
1129
+ Predictions in the brain: Using our past to generate
1130
+ a future.
1131
+ Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu,
1132
+ Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh
1133
+ Nguyen, Jade Copet, Alexei Baevski, Abdelrahman
1134
+ Mohamed, et al. 2021. On generative spoken lan-
1135
+ guage modeling from raw audio. Transactions of the
1136
+ Association for Computational Linguistics, 9:1336–
1137
+ 1354.
1138
+ Howard Lasnik and Jeffrey L Lidz. 2017. The argu-
1139
+ ment from the poverty of the stimulus. The Oxford
1140
+ handbook of universal grammar, pages 221–248.
1141
+ Julie Anne Legate and Charles D Yang. 2002.
1142
+ Em-
1143
+ pirical re-assessment of stimulus poverty arguments.
1144
+ The Linguistic Review, 19(1-2):151–162.
1145
+ Michael Lepori, Tal Linzen, and R. Thomas McCoy.
1146
+ 2020.
1147
+ Representations of syntax [MASK] useful:
1148
+ Effects of constituency and dependency structure in
1149
+ recursive LSTMs. In Proceedings of the 58th An-
1150
+ nual Meeting of the Association for Computational
1151
+ Linguistics, pages 3306–3316, Online. Association
1152
+ for Computational Linguistics.
1153
+ Roger Levy. 2008. Expectation-based syntactic com-
1154
+ prehension. Cognition, 106(3):1126–1177.
1155
+ John Lewis and Jeffrey Elman. 2001. Learnability and
1156
+ the statistical structure of language: Poverty of stim-
1157
+ ulus arguments revisited. Proceedings of the 26th
1158
+ Annual Boston University Conference on Language
1159
+ Development, 1.
1160
+ Yongjie Lin, Yi Chern Tan, and Robert Frank. 2019.
1161
+ Open sesame:
1162
+ Getting inside BERT’s linguistic
1163
+ knowledge. In Proceedings of the 2019 ACL Work-
1164
+ shop BlackboxNLP: Analyzing and Interpreting Neu-
1165
+ ral Networks for NLP, pages 241–253, Florence,
1166
+ Italy. Association for Computational Linguistics.
1167
+ Tal Linzen. 2020. How can we accelerate progress to-
1168
+ wards human-like linguistic generalization? In Pro-
1169
+ ceedings of the 58th Annual Meeting of the Asso-
1170
+ ciation for Computational Linguistics, pages 5210–
1171
+ 5217, Online. Association for Computational Lin-
1172
+ guistics.
1173
+ Brian MacWhinney. 2000.
1174
+ The CHILDES project:
1175
+ Tools for analyzing talk. Lawrence Erlbaum Asso-
1176
+ ciates.
1177
+ R. Thomas McCoy, Robert Frank, and Tal Linzen.
1178
+ 2018. Revisiting the poverty of the stimulus: hier-
1179
+ archical generalization without a hierarchical bias in
1180
+ recurrent neural networks.
1181
+ R. Thomas McCoy, Robert Frank, and Tal Linzen.
1182
+ 2020. Does syntax need to grow on trees? sources of
1183
+ hierarchical inductive bias in sequence-to-sequence
1184
+ networks.
1185
+ James L. Morgan and Katherine Demuth. 1996. Signal
1186
+ to syntax: Bootstrapping from speech to grammar in
1187
+ early acquisition. Psychology Press.
1188
+ Aaron Mueller, Robert Frank, Tal Linzen, Luheng
1189
+ Wang, and Sebastian Schuster. 2022. Coloring the
1190
+ blank slate: Pre-training imparts a hierarchical in-
1191
+ ductive bias to sequence-to-sequence models.
1192
+ In
1193
+ Findings of the Association for Computational Lin-
1194
+ guistics: ACL 2022, pages 1352–1368, Dublin, Ire-
1195
+ land. Association for Computational Linguistics.
1196
+ Karl Mulligan, Robert Frank, and Tal Linzen. 2021.
1197
+ Structure here, bias there: Hierarchical generaliza-
1198
+ tion by jointly learning syntactic transformations. In
1199
+ Proceedings of the Society for Computation in Lin-
1200
+ guistics 2021, pages 125–135, Online. Association
1201
+ for Computational Linguistics.
1202
+ Ludovica Pannitto and Aurélie Herbelot. 2020. Recur-
1203
+ rent babbling: evaluating the acquisition of gram-
1204
+ mar from limited input data.
1205
+ In Proceedings of
1206
+ the 24th Conference on Computational Natural Lan-
1207
+ guage Learning, pages 165–176, Online. Associa-
1208
+ tion for Computational Linguistics.
1209
+ Adam Pauls and Dan Klein. 2012. Large-scale syntac-
1210
+ tic language modeling with treelets. In Proceedings
1211
+ of the 50th Annual Meeting of the Association for
1212
+ Computational Linguistics (Volume 1: Long Papers),
1213
+ pages 959–968, Jeju Island, Korea. Association for
1214
+ Computational Linguistics.
1215
+
1216
+ Lisa Pearl. 2021. Poverty of the stimulus without tears.
1217
+ Language Learning and Development, pages 1–40.
1218
+ Lisa Pearl and Benjamin Mis. 2016. The role of in-
1219
+ direct positive evidence in syntactic acquisition: A
1220
+ look at anaphoric one. Language, 92:1–30.
1221
+ Lisa Pearl and Jon Sprouse. 2013a.
1222
+ Computational
1223
+ models of acquisition for islands. Experimental syn-
1224
+ tax and islands effects, pages 109–131.
1225
+ Lisa Pearl and Jon Sprouse. 2013b. Syntactic islands
1226
+ and learning biases: Combining experimental syntax
1227
+ and computational modeling to investigate the lan-
1228
+ guage acquisition problem. Language Acquisition,
1229
+ 20(1):23–68.
1230
+ Andrew Perfors, Josh Tenenbaum, and Terry Regier.
1231
+ 2011. The learnability of abstract syntactic princi-
1232
+ ples. Cognition, 118:306–338.
1233
+ Jackson Petty and Robert Frank. 2021.
1234
+ Trans-
1235
+ formers
1236
+ generalize
1237
+ linearly.
1238
+ arXiv
1239
+ preprint
1240
+ arXiv:2109.12036.
1241
+ Geoffrey K. Pullum and Barbara C. Scholz. 2002. Em-
1242
+ pirical assessment of stimulus poverty arguments.
1243
+ The Linguistic Review, 18(1-2):9–50.
1244
+ Florencia Reali and Morten H. Christiansen. 2005. Un-
1245
+ covering the richness of the stimulus: Structure de-
1246
+ pendence and indirect statistical evidence.
1247
+ Cogni-
1248
+ tive Science, 29(6):1007–1028.
1249
+ Haoyue Shi, Jiayuan Mao, Kevin Gimpel, and Karen
1250
+ Livescu. 2019. Visually grounded neural syntax ac-
1251
+ quisition. In Proceedings of the 57th Annual Meet-
1252
+ ing of the Association for Computational Linguistics,
1253
+ pages 1842–1861, Florence, Italy. Association for
1254
+ Computational Linguistics.
1255
+ Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
1256
+ Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
1257
+ Kaiser, and Illia Polosukhin. 2017. Attention is all
1258
+ you need. In Advances in neural information pro-
1259
+ cessing systems, pages 5998–6008.
1260
+ Alex Warstadt and Samuel R Bowman. 2020. Can neu-
1261
+ ral networks acquire a structural bias from raw lin-
1262
+ guistic data? Proceedings of the 42nd Annual Con-
1263
+ ference of the Cognitive Science Society.
1264
+ Alex Warstadt and Samuel R Bowman. 2022. What
1265
+ artificial
1266
+ neural
1267
+ networks
1268
+ can
1269
+ tell
1270
+ us
1271
+ about
1272
+ human language acquisition.
1273
+ arXiv preprint
1274
+ arXiv:2208.07998.
1275
+ Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mo-
1276
+ hananey, Wei Peng, Sheng-Fu Wang, and Samuel R
1277
+ Bowman. 2020a.
1278
+ BLiMP: The benchmark of lin-
1279
+ guistic minimal pairs for english.
1280
+ Transactions
1281
+ of the Association for Computational Linguistics,
1282
+ 8:377–392.
1283
+ Alex Warstadt, Yian Zhang, Xiaocheng Li, Haokun
1284
+ Liu, and Samuel R. Bowman. 2020b.
1285
+ Learning
1286
+ which features matter: RoBERTa acquires a prefer-
1287
+ ence for linguistic generalizations (eventually). In
1288
+ Proceedings of the 2020 Conference on Empirical
1289
+ Methods in Natural Language Processing (EMNLP),
1290
+ pages 217–235, Online. Association for Computa-
1291
+ tional Linguistics.
1292
+ Ethan Wilcox, Richard Futrell, and Roger Levy. 2021.
1293
+ Using computational models to test syntactic learn-
1294
+ ability. lingbuzz preprint lingbuzz/006327.
1295
+ Ethan Wilcox, Roger Levy, Takashi Morita, and
1296
+ Richard Futrell. 2018.
1297
+ What do RNN language
1298
+ models learn about filler–gap dependencies?
1299
+ In
1300
+ Proceedings of the 2018 EMNLP Workshop Black-
1301
+ boxNLP: Analyzing and Interpreting Neural Net-
1302
+ works for NLP, pages 211–221, Brussels, Belgium.
1303
+ Association for Computational Linguistics.
1304
+ Taha Yasseri, András Kornai, and János Kertész. 2012.
1305
+ A practical approach to language complexity: A
1306
+ Wikipedia case study. PLoS ONE, 7(11):e48386.
1307
+ A
1308
+ CHILDES preprocessing details
1309
+ The train, test, and validation split kept each docu-
1310
+ ment in the corpora intact to allow for learning of
1311
+ context. Since a document roughly correspond to
1312
+ a single recording session, and the sentence order
1313
+ within each document was not randomized, the net-
1314
+ works could utilize cross sentence context while
1315
+ predicting the next word.
1316
+ Generally, we kept the data as close to the actual
1317
+ input that the child receives as possible. However,
1318
+ in some cases we modified tokenization to match
1319
+ the CHILDES Treebank, a syntactically parsed sub-
1320
+ set of the CHILDES corpora. For instance, con-
1321
+ tractions were split, e.g. we replaced don’t with do
1322
+ n’t,
1323
+ The ages of the children vary by corpus, ranging
1324
+ from six months to twelve years. Almost 95%
1325
+ (49/52) of the corpora consist of transcriptions with
1326
+ children between one and six years of age.
1327
+ Note that for Experiment 2, we used the same vo-
1328
+ cabulary as we used in Experiment 1, which means
1329
+ that the words that were not present in the Exper-
1330
+ iment 1’s vocabulary were replaced with <unk>
1331
+ tokens.
1332
+ The unprocessed CHILDES datasets were down-
1333
+ loaded in XML format from the online XML ver-
1334
+ sion10 of the CHILDES database (MacWhinney,
1335
+ 2000).11
1336
+ A modified NLTK CHILDESCorpus-
1337
+ 10https://childes.talkbank.org/
1338
+ data-xml/
1339
+ 11https://childes.talkbank.org
1340
+
1341
+ Reader12 was used to parse the XML into plain
1342
+ text for training.
1343
+ The CHILDES dataset is licensed for use under
1344
+ a CC BY-NC-SA 3.0 license13. Under the terms of
1345
+ this license, the data can be freely used and adapted,
1346
+ as long as it is not used for commercial purposes
1347
+ and as long as attribution is provided.14 Our usage
1348
+ fits these criteria.
1349
+ Though CHILDES contains many corpora of
1350
+ many languages, we use only corpora from the
1351
+ North American English subset of CHILDES,
1352
+ which contains child-directed speech with many
1353
+ different North American children.
1354
+ See the
1355
+ CHILDES database for more details.
1356
+ By the CHILDES rules for data citation.15 re-
1357
+ search that relies on more than 6 of the corpora
1358
+ need only cite the overall database, not each indi-
1359
+ vidual corpus.
1360
+ All the data on CHILDES must adhere to
1361
+ IRB guidelines,16 including a requirement for
1362
+ anonymity.
1363
+ The final dataset will be included in our GitHub
1364
+ repository, to be released soon. This dataset is not
1365
+ intended for commercial use.
1366
+ CHILDES corpora included
1367
+ The CHILDES
1368
+ corpora that we used were: Bates, Bernstein, Bliss,
1369
+ Bloom70, Bloom73, Bohannon, Braunwald, Brent,
1370
+ Brown, Carterette, Clark, Cornell, Demetras1,
1371
+ Demetras2, EllisWeismer, Evans, Feldman, Garvey,
1372
+ Gathercole, Gelman, Gillam, Gleason, HSLLD,
1373
+ Haggerty, Hall, Higginson, Kuczaj, MacWhin-
1374
+ ney, McCune, McMillan, Morisset, NH, Nelson,
1375
+ NewEngland, NewmanRatner, Normal, POLER,
1376
+ Peters, Post, Rollins, Sachs, Sawyer, Snow, Soder-
1377
+ strom, Sprott, Suppes, Tardif, Valian, VanHouten,
1378
+ VanKleeck, Warren, Weist.
1379
+ B
1380
+ Hyperparameter Search and Model
1381
+ Implementation
1382
+ We conducted a hyperparameter search for each
1383
+ of the architectures we investigated (LSTMs and
1384
+ Transformers). Our broad goal in this paper is to
1385
+ investigate the extent to which capturing the statis-
1386
+ tical properties of the CHILDES dataset naturally
1387
+ 12https://www.nltk.org/howto/childes.
1388
+ html
1389
+ 13https://talkbank.org/share/rules.html
1390
+ 14https://creativecommons.org/licenses/
1391
+ by-nc-sa/3.0/
1392
+ 15https://talkbank.org/share/citation.
1393
+ html
1394
+ 16https://talkbank.org/share/irb/
1395
+ leads a learner to capture the structure of yes/no
1396
+ questions. Therefore, we sought to find the hyper-
1397
+ parameter settings that made models most effective
1398
+ at capturing the statistical properties of CHILDES
1399
+ data, a goal which we operationalized as finding
1400
+ the model with the lowest perplexity.
1401
+ B.1
1402
+ Hyperparameter search
1403
+ LSTMs
1404
+ For LSTMs we explored the following
1405
+ hyper-parameters via a grid search for a total of
1406
+ 144 models.
1407
+ 1. layers: 2
1408
+ 2. hidden and embedding size: 200, 800
1409
+ 3. batch size: 20, 80
1410
+ 4. dropout rate: 0.0, 0.2, 0.4, 0.6
1411
+ 5. learning rate: 5.0, 10.0, 20.0
1412
+ 6. random seed: 3 per parameter combination,
1413
+ unique for each LSTM
1414
+ The LSTM model with the lowest perplexity on the
1415
+ validation set after training had 2 layers, a hidden
1416
+ and embedding size of 800, a batch size of 20, a
1417
+ dropout rate of 0.4, and a learning rate of 10.17
1418
+ A LSTM model with these hyperparameters has
1419
+ 37,620,294 parameters.
1420
+ Transformers
1421
+ For the Transformers we per-
1422
+ formed a hyperparameter sweep over the following
1423
+ hyper-parameters for a total of 84 models.
1424
+ 1. layers: 2, 4, 8, 16
1425
+ 2. context size: 50, 100, 500
1426
+ 3. hidden and embedding size: 200, 800, 1600
1427
+ 4. heads: 2, 4, 8, 16
1428
+ 5. batch size: 20, 80, 160
1429
+ 6. dropout rate: 0.0, 0.2, 0.4, 0.6
1430
+ 7. learning rate: 0.5, 1.0, 5.0, 10.0, 20.0
1431
+ 8. random seed: 3 per parameter combination
1432
+ 17The hyperparameters we explored for the LSTMs
1433
+ were
1434
+ those
1435
+ of
1436
+ Gulordava
1437
+ et
1438
+ al.
1439
+ (2018),
1440
+ the
1441
+ code
1442
+ for which can be found at
1443
+ https://github.com/
1444
+ facebookresearch/colorlessgreenRNNs
1445
+
1446
+ LSTMs
1447
+ Prepose First
1448
+ Prepose Main
1449
+ Delete First
1450
+ 0.01
1451
+ 0.14
1452
+ Delete Main
1453
+ 0.39
1454
+ 0.12
1455
+ Delete None
1456
+ 0.20
1457
+ 0.14
1458
+ Table 1: Numerical results for LSTMs’ preference for
1459
+ questions consistent with combinations of ‘prepose’
1460
+ and ‘delete’ rules. Within each architecture, the propor-
1461
+ tion preferences across all 6 question types necessarily
1462
+ sum to 1.
1463
+ The Transformer model with the lowest perplexities
1464
+ after training had 4 layers, a context size of 500,
1465
+ a hidden size of 800, a batch size of 10, 4 heads,
1466
+ a dropout rate of 0.2, and a learning rate of 5.0.
1467
+ A Transformer model with these parameters has
1468
+ 42,759,494 parameters.
1469
+ B.2
1470
+ Comment on model size
1471
+ Although neural networks generally perform better
1472
+ as they increase in size, the best-performing models
1473
+ that we found were not the largest ones. This re-
1474
+ sult is consistent with the finding of Warstadt et al.
1475
+ (2020b) that, for small training sets, smaller lan-
1476
+ guage models sometimes outperform larger ones.
1477
+ Thus, it is unlikely that scaling up models beyond
1478
+ the range we investigated would have yielded bet-
1479
+ ter CHILDES language models than the ones we
1480
+ trained.
1481
+ B.3
1482
+ Implementation
1483
+ All
1484
+ models
1485
+ were
1486
+ implemented
1487
+ in
1488
+ Py-
1489
+ Torch
1490
+ by
1491
+ building
1492
+ on
1493
+ code
1494
+ from
1495
+ https:
1496
+ //github.com/facebookresearch/
1497
+ colorlessgreenRNNs
1498
+ and
1499
+ https:
1500
+ //github.com/pytorch/examples/
1501
+ tree/main/word_language_model,
1502
+ and
1503
+ trained using Nvidia k80 GPUs. The final models
1504
+ will be included in our GitHub repository, which
1505
+ will be released soon.
1506
+ These models are not
1507
+ intended for commercial use.
1508
+ C
1509
+ PREPOSE-ONE&DELETE-ONE Full
1510
+ Results
1511
+ See Table 1 and Table 2 for these results. See Table
1512
+ 1 and Table 2 for these results.
1513
+ C.1
1514
+ Results using SLOR
1515
+ See Table 3 and Table 4 for these results.
1516
+ Transformers
1517
+ Prepose First
1518
+ Prepose Main
1519
+ Delete First
1520
+ 0.01
1521
+ 0.16
1522
+ Delete Main
1523
+ 0.31
1524
+ 0.06
1525
+ Delete None
1526
+ 0.25
1527
+ 0.21
1528
+ Table 2: Numerical results for Transformers’ prefer-
1529
+ ence for questions consistent with combinations of ‘pre-
1530
+ pose’ and ‘delete’ rules. Within each architecture, the
1531
+ proportion preferences across all 6 question types nec-
1532
+ essarily sum to 1.
1533
+ LSTMs
1534
+ Prepose First
1535
+ Prepose Main
1536
+ Delete First
1537
+ 0.01
1538
+ 0.14
1539
+ Delete Main
1540
+ 0.33
1541
+ 0.80
1542
+ Delete None
1543
+ 0.26
1544
+ 0.18
1545
+ Table 3: Analysis of LSTMs’ preference for questions
1546
+ consistent with combinations of ‘prepose’ and ‘delete’
1547
+ rules, evaluated using SLOR. Within each architecture,
1548
+ the proportion preferences across all 6 question types
1549
+ necessarily sum to 1.
1550
+ Transformers
1551
+ Prepose First
1552
+ Prepose Main
1553
+ Delete First
1554
+ 0.01
1555
+ 0.15
1556
+ Delete Main
1557
+ 0.27
1558
+ 0.40
1559
+ Delete None
1560
+ 0.29
1561
+ 0.24
1562
+ Table 4: Analysis of Transformers’ preference for ques-
1563
+ tions consistent with combinations of ‘prepose’ and
1564
+ ‘delete’ rules, evaluated using SLOR. Within each ar-
1565
+ chitecture, the proportion preferences across all 6 ques-
1566
+ tion types necessarily sum to 1.
1567
+ D
1568
+ BabyBERTa dataset evaluation
1569
+ For an illustrative subset of the results on the Zorro
1570
+ evaluation dataset (discussed in Section 4.5), see
1571
+ Figure 4. For the full results, see Figure 5.
1572
+ E
1573
+ Move-One Dataset Results
1574
+ One approach used in several past papers (e.g.,
1575
+ Lewis and Elman (2001) and Reali and Chris-
1576
+ tiansen (2005)) is to evaluate models using pairs
1577
+ of sentences that can be formed by starting with a
1578
+ declarative sentence (e.g., (17)) and moving one of
1579
+ its auxiliaries to the front of the sentence. The first
1580
+ sentence in each pair (e.g., (18a) ) follows HIER-
1581
+ ARCHICALQ, because the main auxiliary is moved,
1582
+ while the second (e.g., (18b)), follows LINEARQ
1583
+ because the first auxiliary is moved.
1584
+ (17) The children who are talking are sleeping.
1585
+ (18) a. Are the children who are talking sleeping?
1586
+ b. Are the children who talking are sleeping?
1587
+
1588
+ 1
1589
+ 0.46
1590
+ 0.98
1591
+ 0.41
1592
+ 0.95
1593
+ 0.99
1594
+ 0.97
1595
+ 0.95
1596
+ 0.64
1597
+ 0.99
1598
+ 0.43
1599
+ 0.96
1600
+ 0.41
1601
+ 0.88
1602
+ 0.95
1603
+ 0.96
1604
+ 0.97
1605
+ 0.54
1606
+ LSTM 02
1607
+ LSTM 03
1608
+ LSTM 08
1609
+ Transformer 02
1610
+ Transformer 03
1611
+ Transformer 08
1612
+ irreg_v
1613
+ sv_agr_rc
1614
+ swap_arg
1615
+ Zorro Evaluation
1616
+ Model
1617
+ 0.5
1618
+ 0.6
1619
+ 0.7
1620
+ 0.8
1621
+ 0.9
1622
+ 1.0
1623
+ Proportion
1624
+ Correct
1625
+ Figure 4: The performance of a selected subset of
1626
+ model re-runs on a selected subset of the Zorro evalua-
1627
+ tions. Each Zorro evaluation targets a specific syntactic
1628
+ phenomenon—in the cases shown here, irregular verbs,
1629
+ subject-verb agreement across relative clauses, and cor-
1630
+ rect argument ordering.
1631
+ If a model assigns a higher probability to (18a)
1632
+ than (18b), that is evidence that the models favors
1633
+ HIERARCHICALQ over LINEARQ. While this pref-
1634
+ erence is a necessary component of correctly learn-
1635
+ ing HIERARCHICALQ, it is by no means sufficient:
1636
+ indeed, Kam et al. (2008) showed that models can
1637
+ prefer sentences consistent with HIERARCHICALQ
1638
+ over sentences consistent with LINEARQ due to
1639
+ shallow n-gram statistics rather than due to knowl-
1640
+ edge of hierarchical structure.
1641
+ More generally,
1642
+ there are infinitely many other incorrect hypotheses
1643
+ besides LINEARQ, and demonstrating successful
1644
+ learning of HIERARCHICALQ would require ruling
1645
+ out all of them. Investigating all possibilities is
1646
+ intractable, but we can at least investigate a few
1647
+ additional plausible ones. Thus, in the main paper
1648
+ we depart from prior work by considering a greater
1649
+ number of candidate sentences than just the pairs
1650
+ of sentences used in prior work.
1651
+ To create the MOVE-ONE dataset, we ran-
1652
+ domly sampled 10,000 declarative sentences from
1653
+ our CFGs for which the first and main auxiliary
1654
+ were identical and then modified them to give
1655
+ 10,000 sentence pairs. To create the PREPOSE-
1656
+ ONE&DELETE-ONE dataset, we randomly sam-
1657
+ pled a different 10,000 declarative sentences from
1658
+ our CFGs for which the first and main auxiliary
1659
+ were different and then we modified them to give
1660
+ 10,000 6-tuples of sentences. See Appendix F for
1661
+ more details about the CFGs.
1662
+ F
1663
+ Context Free Grammars
1664
+ Figure 6 contains the context-free grammar used
1665
+ for the analyses in Section 4.6. Figures 7 and 8 con-
1666
+ tain the context-free grammars used for the targeted
1667
+ evaluation sets in Section 5.2. Figure 9 contains
1668
+ the vocabulary used for all of these datasets.
1669
+ G
1670
+ Breakdown by lexical identity
1671
+ Here we further break down models’ predictions
1672
+ for the FIRST-AUX ̸= MAIN-AUX evaluation set
1673
+ based on the identities of the two auxiliaries in
1674
+ the input sentence. Figure 10 gives the results for
1675
+ the LSTM in the QUESTION FORMATION condi-
1676
+ tion; Figure 11 for the LSTM in the NEXT-WORD
1677
+ PREDICTION + QUESTION FORMATION condi-
1678
+ tion; Figure 12 for the Transformer in the QUES-
1679
+ TION FORMATION condition; and Figure 13 for the
1680
+ for the Transformer in the NEXT-WORD PREDIC-
1681
+ TION + QUESTION FORMATION condition.
1682
+ H
1683
+ Example generated text
1684
+ Figure 14 gives some example text generated by our
1685
+ models. Models trained on next-word prediction
1686
+ produce their predictions as a probability distribu-
1687
+ tion over the vocabulary. To use such models to
1688
+ generate text, we sample a word from this distribu-
1689
+ tion then use that word as the model’s input for the
1690
+ next time step.
1691
+
1692
+ 89 79
1693
+ 67 89 86 82 98 91 100 92 100 97 85
1694
+ 88 60 56 78 40
1695
+ 67 87 71 59 96
1696
+ 89 100 91 100 96 85 88 61
1697
+ 46 39 39 74 88
1698
+ 78 59 98 88 88
1699
+ 62 89 79 83 98
1700
+ 59 56
1701
+ 68 41 60 86 69 66
1702
+ 95 90 87 80 89
1703
+ 68 80 99 88 100 90 100 94 85 92
1704
+ 61 98
1705
+ 87 88 85 88 80 82
1706
+ 97 90 100 93 99
1707
+ 96 83 86 59 43
1708
+ 48 39 68 90 73
1709
+ 85 97
1710
+ 91 100 93 100 95 84
1711
+ 91 60 60 82 38
1712
+ 64 89 77 63 96
1713
+ 85 79 88 89 80
1714
+ 84 89
1715
+ 61 53 51 36 61 89
1716
+ 74 62 99 88 83
1717
+ 89 88 79 80 98
1718
+ 90 99 93 100 94
1719
+ 86 73
1720
+ 63 95 90 85 60 89
1721
+ 58 77 97 90 100 90 100 97 83 89
1722
+ 60 50 76 37 70
1723
+ 90 81
1724
+ 81 97 90 100 93 100 95 84 89 64 50
1725
+ 54 37 64 89 73
1726
+ 61 97 89 85 80
1727
+ 100 96
1728
+ 81 87 61 55 72 36
1729
+ 73 93 73 61 98
1730
+ 87 89 64 89 83
1731
+ 81 96 90 100 91
1732
+ 38 70
1733
+ 90 76 59 97 89 84
1734
+ 78 87 85 78 98
1735
+ 88 100 91 100 97
1736
+ 82 91 58 46 69
1737
+ 79 77
1738
+ 46 89 54 62 92 77
1739
+ 99 81 99 95 72
1740
+ 75 54 53 61 42
1741
+ 50 81 64 61 83
1742
+ 82 97
1743
+ 83 99 94 71 78 54
1744
+ 43 72 45 48 85
1745
+ 71 59 96 80 74
1746
+ 58 91 62 62 93
1747
+ 54 48
1748
+ 77 41 62 79 73 58
1749
+ 88 81 79 65 91
1750
+ 59 65 95 80 93
1751
+ 88 98 96 74 78
1752
+ 60 88
1753
+ 75 74 61 92 51 63
1754
+ 95 84 100 88 98
1755
+ 93 73 80 53 38
1756
+ 78 42 53 83 73
1757
+ 70 95
1758
+ 81 99 83 98 95 74
1759
+ 77 54 36 50 43
1760
+ 47 83 65 58 90
1761
+ 78 72 82 92 53
1762
+ 74 76
1763
+ 53 44 55 42 47 80
1764
+ 72 59 90 78 72
1765
+ 87 92 58 66 96
1766
+ 80 99 88 98 94
1767
+ 80 73
1768
+ 61 93 77 72 90 91
1769
+ 53 65 90 80 98
1770
+ 90 98 95 73 83
1771
+ 55 43 44 44 33
1772
+ 91 67
1773
+ 64 96 83 98 81 100 97 72 79 54 40
1774
+ 58 40 54 80 71
1775
+ 60 92 83 69 38
1776
+ 98 96
1777
+ 71 75 55 47 53 42
1778
+ 49 79 86 58 88
1779
+ 79 81 64 91 61
1780
+ 62 92 84 97 85
1781
+ 42 53
1782
+ 86 74 63 88 76 80
1783
+ 93 90 53 63 94
1784
+ 83 99 86 96 96
1785
+ 75 82 55 50 53
1786
+ LSTM 01
1787
+ LSTM 02
1788
+ LSTM 03
1789
+ LSTM 04
1790
+ LSTM 05
1791
+ LSTM 06
1792
+ LSTM 07
1793
+ LSTM 08
1794
+ LSTM 09
1795
+ LSTM 10
1796
+ Transformer 01
1797
+ Transformer 02
1798
+ Transformer 03
1799
+ Transformer 04
1800
+ Transformer 05
1801
+ Transformer 06
1802
+ Transformer 07
1803
+ Transformer 08
1804
+ Transformer 09
1805
+ Transformer 10
1806
+ agreement_determiner_noun−across_1_adjective
1807
+ agreement_determiner_noun−between_neighbors
1808
+ agreement_subject_verb−across_prepositional_phrase
1809
+ agreement_subject_verb−across_relative_clause
1810
+ agreement_subject_verb−in_question_with_aux
1811
+ agreement_subject_verb−in_simple_question
1812
+ anaphor_agreement−pronoun_gender
1813
+ argument_structure−dropped_argument
1814
+ argument_structure−swapped_arguments
1815
+ argument_structure−transitive
1816
+ binding−principle_a
1817
+ case−subjective_pronoun
1818
+ ellipsis−n_bar
1819
+ filler−gap−wh_question_object
1820
+ filler−gap−wh_question_subject
1821
+ irregular−verb
1822
+ island−effects−adjunct_island
1823
+ island−effects−coordinate_structure_constraint
1824
+ local_attractor−in_question_with_aux
1825
+ npi_licensing−matrix_question
1826
+ npi_licensing−only_npi_licensor
1827
+ quantifiers−existential_there
1828
+ quantifiers−superlative
1829
+ Evaluation
1830
+ Model
1831
+ 40
1832
+ 60
1833
+ 80
1834
+ 100
1835
+ % Correct
1836
+ Figure 5: Results on the targeted syntactic evaluations in Huebner et al. (2021) in percent accuracy. Evaluation
1837
+ names in Figure 4 were shortened.
1838
+
1839
+ S
1840
+ → {NP_S RC_S_BARE MAIN-AUX VP_S_PAST}
1841
+ S
1842
+ → {NP_S RC_S_PAST MAIN-AUX VP_S_BARE}
1843
+ S
1844
+ → {NP_S RC_S_BARE MAIN-AUX VP_S_PROG}
1845
+ S
1846
+ → {NP_S RC_S_PROG MAIN-AUX VP_S_BARE}
1847
+ S
1848
+ → {NP_S RC_S_PAST MAIN-AUX VP_S_PROG}
1849
+ S
1850
+ → {NP_S RC_S_PROG MAIN-AUX VP_S_PAST}
1851
+ S
1852
+ → {NP_P RC_P_BARE MAIN-AUX VP_P_PAST}
1853
+ S
1854
+ → {NP_P RC_P_PAST MAIN-AUX VP_P_BARE}
1855
+ S
1856
+ → {NP_P RC_P_BARE MAIN-AUX VP_P_PROG}
1857
+ S
1858
+ → {NP_P RC_P_PROG MAIN-AUX VP_P_BARE}
1859
+ S
1860
+ → {NP_P RC_P_PAST MAIN-AUX VP_P_PROG}
1861
+ S
1862
+ → {NP_P RC_P_PROG MAIN-AUX VP_P_PAST}
1863
+ NP_S
1864
+ → {Det_S N_S}
1865
+ NP_P
1866
+ → {Det_P N_P}
1867
+ NP_O
1868
+ → {Det_S N_S | Det_P N_P | Det_S N_S Prep Det_S N_S | Det_S N_S Prep
1869
+ Det_P N_P | Det_P N_P Prep Det_S N_S | Det_P N_P Prep Det_P N_P}
1870
+ VP_S_BARE
1871
+ → {Aux_S IV }
1872
+ VP_S_BARE
1873
+ → {Aux_S TV NP_O}
1874
+ VP_S_PROG
1875
+ → {Aux_S_BE IV_IS}
1876
+ VP_S_PROG
1877
+ → {Aux_S_BE TV_IS NP_O}
1878
+ VP_S_PAST
1879
+ → {Aux_S_HAS IV_HAS}
1880
+ VP_S_PAST
1881
+ → {Aux_S_HAS TV_HAS NP_O}
1882
+ VP_P_BARE
1883
+ → {Aux_P IV}
1884
+ VP_P_BARE
1885
+ → {Aux_P TV NP_O}
1886
+ VP_P_PROG
1887
+ → {Aux_P_BE IV_IS}
1888
+ VP_P_PROG
1889
+ → {Aux_P_BE TV_IS NP_O}
1890
+ VP_P_PAST
1891
+ → {Aux_P_HAS IV_HAS}
1892
+ VP_P_PAST
1893
+ → {Aux_P_HAS TV_HAS NP_O}
1894
+ RC_S_BARE
1895
+ → {Rel Aux_S IV | Rel Det_S N_S Aux_S TV | Rel Det_P N_P Aux_P TV |
1896
+ Rel Aux_S TV Det_S N_S | Rel Aux_S TV Det_P N_P}
1897
+ RC_S_PROG
1898
+ → {Rel Aux_S_BE IV_IS | Rel Det_S N_S Aux_S_BE TV_IS | Rel Det_P
1899
+ N_P Aux_P_BE TV_IS | Rel Aux_S_BE TV_IS Det_S N_S | Rel Aux_S_BE
1900
+ TV_IS Det_P N_P}
1901
+ RC_S_PAST
1902
+ → {Rel Aux_S_HAS IV_HAS | Rel Det_S N_S Aux_S_HAS TV_HAS | Rel
1903
+ Det_P N_P Aux_P_HAS TV_HAS | Rel Aux_S_HAS TV_HAS Det_S N_S |
1904
+ Rel Aux_S_HAS TV_HAS Det_P N_P}
1905
+ RC_P_BARE
1906
+ → {Rel Aux_P IV | Rel Det_S N_S Aux_S TV | Rel Det_P N_P Aux_P TV |
1907
+ Rel Aux_P TV Det_S N_S | Rel Aux_P TV Det_P N_P}
1908
+ RC_P_PROG
1909
+ → {Rel Aux_P_BE IV_IS | Rel Det_S N_S Aux_S_BE TV_IS | Rel Det_P
1910
+ N_P Aux_P_BE TV_IS | Rel Aux_P_BE TV_IS Det_S N_S | Rel Aux_P_BE
1911
+ TV_IS Det_P N_P}
1912
+ RC_P_PAST
1913
+ → {Rel Aux_P_HAS IV_HAS | Rel Det_S N_S Aux_S_HAS TV_HAS | Rel
1914
+ Det_P N_P Aux_P_HAS TV_HAS | Rel Aux_P_HAS TV_HAS Det_S N_S |
1915
+ Rel Aux_P_HAS TV_HAS Det_P N_P}
1916
+ Figure 6: CFG used to generate PREPOSE-ONE-AND-DELETE-ONE evaluation dataset
1917
+
1918
+ S
1919
+ → {NP_M_S VP_M_S | NP_M_P VP_M_P}
1920
+ NP_M_S→ {Det_S N_S | Det_S N_S Prep Det_S N_S | Det_S N_S Prep Det_P N_P}
1921
+ NP_M_P→ {Det_P N_P | Det_P N_P Prep Det_S N_S | Det_P N_P Prep Det_P N_P}
1922
+ NP_O
1923
+ → {Det_S N_S | Det_P N_P | Det_S N_S Prep Det_S N_S | Det_S N_S Prep
1924
+ Det_P N_P | Det_P N_P Prep Det_S N_S | Det_P N_P Prep Det_P N_P | Det_S
1925
+ N_S RC_S | Det_P N_P RC_P }
1926
+ VP_M_S→ {Aux_S IV }
1927
+ VP_M_S→ {Aux_S TV NP_O}
1928
+ VP_M_S→ {Aux_S_BE IV_IS}
1929
+ VP_M_S→ {Aux_S_BE TV_IS NP_O}
1930
+ VP_M_S→ {Aux_S_HAS IV_HAS}
1931
+ VP_M_S→ {Aux_S_HAS TV_HAS NP_O}
1932
+ VP_M_P→ {Aux_P IV}
1933
+ VP_M_P→ {Aux_P TV NP_O}
1934
+ VP_M_P→ {Aux_P_BE IV_IS}
1935
+ VP_M_P→ {Aux_P_BE TV_IS NP_O}
1936
+ VP_M_P→ {Aux_P_HAS IV_HAS}
1937
+ VP_M_P→ {Aux_P_HAS TV_HAS NP_O}
1938
+ RC_S
1939
+ → {Rel Aux_S IV | Rel Det_S N_S Aux_S TV | Rel Det_P N_P Aux_P TV |
1940
+ Rel Aux_S TV Det_S N_S | Rel Aux_S TV Det_P N_P}
1941
+ RC_S
1942
+ → {Rel Aux_S_BE IV_IS | Rel Det_S N_S Aux_S_BE TV_IS | Rel Det_P
1943
+ N_P Aux_P_BE TV_IS | Rel Aux_S_BE TV_IS Det_S N_S | Rel Aux_S_BE
1944
+ TV_IS Det_P N_P}
1945
+ RC_S
1946
+ → {Rel Aux_S_HAS IV_HAS | Rel Det_S N_S Aux_S_HAS TV_HAS | Rel
1947
+ Det_P N_P Aux_P_HAS TV_HAS | Rel Aux_S_HAS TV_HAS Det_S N_S |
1948
+ Rel Aux_S_HAS TV_HAS Det_P N_P}
1949
+ RC_P
1950
+ → {Rel Aux_P IV | Rel Det_S N_S Aux_S TV | Rel Det_P N_P Aux_P TV |
1951
+ Rel Aux_P TV Det_S N_S | Rel Aux_P TV Det_P N_P}
1952
+ RC_P
1953
+ → {Rel Aux_P_BE IV_IS | Rel Det_S N_S Aux_S_BE TV_IS | Rel Det_P
1954
+ N_P Aux_P_BE TV_IS | Rel Aux_P_BE TV_IS Det_S N_S | Rel Aux_P_BE
1955
+ TV_IS Det_P N_P}
1956
+ RC_P
1957
+ → {Rel Aux_P_HAS IV_HAS | Rel Det_S N_S Aux_S_HAS TV_HAS | Rel
1958
+ Det_P N_P Aux_P_HAS TV_HAS | Rel Aux_P_HAS TV_HAS Det_S N_S |
1959
+ Rel Aux_P_HAS TV_HAS Det_P N_P}
1960
+ Figure 7: CFG used to generate FIRST-AUX = MAIN-AUX evaluation dataset
1961
+
1962
+ S
1963
+ → {NP_M_S VP_M_S | NP_M_P VP_M_P}
1964
+ NP_M_S→ {Det_S N_S | Det_S N_S Prep Det_S N_S | Det_S N_S Prep Det_P N_P}
1965
+ NP_M_P→ {Det_P N_P | Det_P N_P Prep Det_S N_S | Det_P N_P Prep Det_P N_P}
1966
+ NP_O
1967
+ → {Det_S N_S | Det_P N_P | Det_S N_S Prep Det_S N_S | Det_S N_S Prep
1968
+ Det_P N_P | Det_P N_P Prep Det_S N_S | Det_P N_P Prep Det_P N_P | Det_S
1969
+ N_S RC_S | Det_P N_P RC_P }
1970
+ VP_M_S→ {Aux_S IV }
1971
+ VP_M_S→ {Aux_S TV NP_O}
1972
+ VP_M_S→ {Aux_S_BE IV_IS}
1973
+ VP_M_S→ {Aux_S_BE TV_IS NP_O}
1974
+ VP_M_S→ {Aux_S_HAS IV_HAS}
1975
+ VP_M_S→ {Aux_S_HAS TV_HAS NP_O}
1976
+ VP_M_P→ {Aux_P IV}
1977
+ VP_M_P→ {Aux_P TV NP_O}
1978
+ VP_M_P→ {Aux_P_BE IV_IS}
1979
+ VP_M_P→ {Aux_P_BE TV_IS NP_O}
1980
+ VP_M_P→ {Aux_P_HAS IV_HAS}
1981
+ VP_M_P→ {Aux_P_HAS TV_HAS NP_O}
1982
+ RC_S
1983
+ → {Rel Aux_S IV | Rel Det_S N_S Aux_S TV | Rel Det_P N_P Aux_P TV |
1984
+ Rel Aux_S TV Det_S N_S | Rel Aux_S TV Det_P N_P}
1985
+ RC_S
1986
+ → {Rel Aux_S_BE IV_IS | Rel Det_S N_S Aux_S_BE TV_IS | Rel Det_P
1987
+ N_P Aux_P_BE TV_IS | Rel Aux_S_BE TV_IS Det_S N_S | Rel Aux_S_BE
1988
+ TV_IS Det_P N_P}
1989
+ RC_S
1990
+ → {Rel Aux_S_HAS IV_HAS | Rel Det_S N_S Aux_S_HAS TV_HAS | Rel
1991
+ Det_P N_P Aux_P_HAS TV_HAS | Rel Aux_S_HAS TV_HAS Det_S N_S |
1992
+ Rel Aux_S_HAS TV_HAS Det_P N_P}
1993
+ RC_P
1994
+ → {Rel Aux_P IV | Rel Det_S N_S Aux_S TV | Rel Det_P N_P Aux_P TV |
1995
+ Rel Aux_P TV Det_S N_S | Rel Aux_P TV Det_P N_P}
1996
+ RC_P
1997
+ → {Rel Aux_P_BE IV_IS | Rel Det_S N_S Aux_S_BE TV_IS | Rel Det_P
1998
+ N_P Aux_P_BE TV_IS | Rel Aux_P_BE TV_IS Det_S N_S | Rel Aux_P_BE
1999
+ TV_IS Det_P N_P}
2000
+ RC_P
2001
+ → {Rel Aux_P_HAS IV_HAS | Rel Det_S N_S Aux_S_HAS TV_HAS | Rel
2002
+ Det_P N_P Aux_P_HAS TV_HAS | Rel Aux_P_HAS TV_HAS Det_S N_S |
2003
+ Rel Aux_P_HAS TV_HAS Det_P N_P}
2004
+ Figure 8: CFG used to generate FIRST-AUX ̸= MAIN-AUX evaluation dataset
2005
+
2006
+ Det_S
2007
+ → {the | some | this }
2008
+ Det_P
2009
+ → {the | some | those}
2010
+ N_S
2011
+ → {baby | girl | boy | animal | child | person | horse }
2012
+ N_P
2013
+ → {babies | girls | boys | animals | children | people | horses }
2014
+ IV
2015
+ → {play | read | draw | sit | fall | talk | sleep | try | work | walk}
2016
+ IV_IS
2017
+ → {playing | reading | drawing | sitting | falling | talking | sleeping | trying |
2018
+ working | walking}
2019
+ IV_HAS
2020
+ → {played | read | drawn | sat | fallen | talked | slept | tried | worked | walked}
2021
+ TV
2022
+ → {call | see | find | help | feed | know | pick | visit | watch | reach}
2023
+ TV_IS
2024
+ → {calling | seeing | finding | helping | feeding | knowing | picking | visiting |
2025
+ watching | reaching}
2026
+ TV_HAS
2027
+ → {called | seen | found | helped | fed | known | picked | visited | watched |
2028
+ reached}
2029
+ Aux_P
2030
+ → {do | did | can | would | shall}
2031
+ Aux_S
2032
+ → {does | did | can | would | shall}
2033
+ Aux_S_BE → {is | was}
2034
+ Aux_P_BE → {are | were}
2035
+ Aux_S_HAS→ {has}
2036
+ Aux_P_HAS→ {have}
2037
+ Prep
2038
+ → {by | behind }
2039
+ Rel
2040
+ → {who | that }
2041
+ Figure 9: Vocabulary used for the PREPOSE-ONE-AND-DELETE-ONE, FIRST-AUX ̸= MAIN-AUX, and FIRST-
2042
+ AUX = MAIN-AUX evaluation datasets
2043
+ Figure 10: Breakdown by the identities of the two auxiliaries for outputs in the FIRST-AUX ̸= MAIN-AUX eval-
2044
+ uation set for LSTMs first trained on next-word prediction and then question formation. The two leftmost bars in
2045
+ each cell show a First-vs-main comparison, while the two rightmost bars show an AuxY-vs-AuxX comparison.
2046
+
2047
+ AuxX =
2048
+ AuxX =
2049
+ Auxx =
2050
+ AuxX =
2051
+ Auxx =
2052
+ AuxX =
2053
+ AuxX =
2054
+ AuxX =
2055
+ AuxX =
2056
+ AuxX =
2057
+ AuxX =
2058
+ was
2059
+ have
2060
+ can
2061
+ were
2062
+ shall
2063
+ p!p
2064
+ would
2065
+ does
2066
+ op
2067
+ are
2068
+ s
2069
+ 1.0
2070
+ 0.5
2071
+ 0.0
2072
+ 1
2073
+ 11
2074
+ 1.0
2075
+ 0.5
2076
+ 0.0
2077
+ 1.0
2078
+ 0.5
2079
+ 0.0
2080
+ TIT
2081
+ 1.0
2082
+ behavior consistent
2083
+ 0.5
2084
+ 0.0
2085
+ 1.0
2086
+ 0.5
2087
+ 0.0
2088
+ 1.0
2089
+ Comparison
2090
+ 0.5
2091
+ First-vs-main
2092
+ 0.0
2093
+ AuxY-vs-AuxX
2094
+ 0.0
2095
+ 1.0
2096
+ word
2097
+ 0.5
2098
+ 0.0
2099
+ 1.0
2100
+ First
2101
+ 0.5
2102
+ 0.0
2103
+ 1.0
2104
+ 0.5
2105
+ 0.0
2106
+ 1.0
2107
+ 0.5
2108
+ 0.0Figure 11: Breakdown by the identities of the two auxiliaries for outputs in the FIRST-AUX ̸= MAIN-AUX evalu-
2109
+ ation set for LSTMs trained only on question formation. The two leftmost bars in each cell show a First-vs-main
2110
+ comparison, while the two rightmost bars show an AuxY-vs-AuxX comparison.
2111
+ Figure 12: Breakdown by the identities of the two auxiliaries for outputs in the FIRST-AUX ̸= MAIN-AUX evalua-
2112
+ tion set for Transformers first trained on next-word prediction and then question formation. The two leftmost bars
2113
+ in each cell show a First-vs-main comparison, while the two rightmost bars show an AuxY-vs-AuxX comparison.
2114
+
2115
+ AuxX =
2116
+ AuxX =
2117
+ Auxx =
2118
+ AuxX =
2119
+ AuxX =
2120
+ AuxX =
2121
+ AuxX =
2122
+ AuxX =
2123
+ AuxX =
2124
+ AuxX =
2125
+ AuxX =
2126
+ was
2127
+ have
2128
+ can
2129
+ were
2130
+ shall
2131
+ p!p
2132
+ would
2133
+ does
2134
+ op
2135
+ are
2136
+ S
2137
+ 1.0
2138
+ 0.5
2139
+ 0.0
2140
+ 1
2141
+ 1.0
2142
+ 0.5
2143
+ 0.0
2144
+ 1.0
2145
+ 0.5
2146
+ 0.0
2147
+ 1.0
2148
+ 0.5
2149
+ 1.0
2150
+ Comparison
2151
+ 0.5
2152
+ First-vs-main
2153
+ 0.0
2154
+ AuxY-vs-AuxX
2155
+ 1.0
2156
+ 0.5
2157
+ Aux
2158
+ 0.0
2159
+ 0.5
2160
+ 0.0
2161
+ 1.0
2162
+ 0.5
2163
+ 0'0
2164
+ 1.0
2165
+ 0.5
2166
+ 0.0
2167
+ 1.0
2168
+ 0.5
2169
+ 0.0 -AuxX =
2170
+ AuxX =
2171
+ AuxX =
2172
+ AuxX =
2173
+ Auxx =
2174
+ AuxX =
2175
+ AuxX =
2176
+ AuxX =
2177
+ AuxX =
2178
+ AuxX =
2179
+ AuxX =
2180
+ was
2181
+ have
2182
+ can
2183
+ were
2184
+ shall
2185
+ pIp
2186
+ would
2187
+ does
2188
+ op
2189
+ are
2190
+ s
2191
+ 1.0
2192
+ 0.5
2193
+ 0.0
2194
+ 1
2195
+ 1.0
2196
+ 0.5
2197
+ 0.0
2198
+ T
2199
+ 1.0
2200
+ 0.5
2201
+ 0.0
2202
+ 1.0
2203
+ behavior consistent
2204
+ 0.5
2205
+ Aux
2206
+ 0.0
2207
+ 05
2208
+ Comparison
2209
+ First-vs-main
2210
+ 0.0
2211
+ AuxY-vs-AuxX
2212
+ AuxY
2213
+ 0.0
2214
+ 1.0
2215
+ word
2216
+ 0.5
2217
+ 0.0
2218
+ 1.0
2219
+ First
2220
+ 0.5
2221
+ 0.0
2222
+ 1.0
2223
+ 0.5
2224
+ 0.0
2225
+ 1.0
2226
+ 0.5
2227
+ 0.0 -Figure 13: Breakdown by the identities of the two auxiliaries for outputs in the FIRST-AUX ̸= MAIN-AUX evalu-
2228
+ ation set for Transformers trained only on question formation. The two leftmost bars in each cell show a First-vs-
2229
+ main comparison, while the two rightmost bars show an AuxY-vs-AuxX comparison.
2230
+ Text
2231
+ from
2232
+ CHILDES
2233
+ do you eat in nursery school too ? oh . it ’s nursery school . xxx . what do you
2234
+ get to eat ? what else do you get to eat besides a book ? do n’t play with those
2235
+ books . what else do you have ? what do you have in nursery school ? alright -
2236
+ you can take it . why do n’t you put the turtle back . well - you ’re not going to
2237
+ change his water now . tonight . tonight - you change it about ten times a day .
2238
+ no - you can do it later . forget it now . too bad . too bad . that poor turtle . you
2239
+ do n’t let him sit there five minutes . that was five minutes ?
2240
+ LSTM gen-
2241
+ erated text
2242
+ do you eat in nursery school too ? oh . it ’s nursery school . no ! tomorrow ’s
2243
+ Friday . right . yeah - yeah . yeah I know what . I told you that he liked
2244
+ snipsnap that way . well this is address the grayish kings . ok . what kind of
2245
+ bird do you want to play with ? what else do you want to do ? your crayons .
2246
+ got a circle . I ’m going to use these . there . remember ? ok do you want to put
2247
+ all of these there ? ok let ’s put everything away . what do you want to do ? you
2248
+ want to play in the box ? it ’s just like a ball because it has a ball on it .
2249
+ Transformer
2250
+ generated
2251
+ text
2252
+ do you eat in nursery school too ? oh . it ’s nursery school . mm-hm . and it
2253
+ looks like a little girl . mm-hm . like a little girl . hm , mm-hm . looks like she
2254
+ ’s concerned about baseball this is a picture of a double football player . mm-hm
2255
+ . um , it has a uniform on it . mm-hm . anything more ? no . ok great . got a hat
2256
+ on . and a flower and a little flower . remember that ? do you remember that
2257
+ word ? what do you remember that word ? do you remember what it word says
2258
+ ? it says seven - eight . what ’s that word ? it says eight . look at that word .
2259
+ three - four - six ... eight - nine ...
2260
+ Figure 14: Comparison of text generated by the LSTM and Transformer models with a block of text chosen
2261
+ randomly from the training data. The LSTMs and Transformers were both seeded with the first three sentences
2262
+ of the text taken from CHILDES, which is the underlined in the two model generated texts. Note that neither of
2263
+ the model generated texts were cherry picked either for quality or to be representative of the models’ usual output:
2264
+ rather they were the first things they generated when seeded with the above underlined portion.
2265
+
2266
+ AuxX =
2267
+ AuxX =
2268
+ AuxX =
2269
+ AuxX =
2270
+ AuxX =
2271
+ AuxX =
2272
+ AuxX =
2273
+ AuxX =
2274
+ AuxX =
2275
+ AuxX =
2276
+ AuxX =
2277
+ was
2278
+ have
2279
+ can
2280
+ were
2281
+ shall
2282
+ pIp
2283
+ pinom
2284
+ does
2285
+ op
2286
+ are
2287
+ s
2288
+ 1.0
2289
+ 0.5
2290
+ 0.0
2291
+ 1.0
2292
+ 0.5
2293
+ rule
2294
+ 0.0
2295
+ 1.0
2296
+ I behavior consistent with
2297
+ 881
2298
+ 18
2299
+ AUXY
2300
+ 0.0
2301
+ Comparison
2302
+ First-vs-main
2303
+ AuxY-vs-AuxX
2304
+ 0.0
2305
+ 1.0
2306
+ 0.5
2307
+ 0.0
2308
+ 1.0
2309
+ 0.5
2310
+ 0.0
QtFJT4oBgHgl3EQfJyy0/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff