jackkuo commited on
Commit
a9d2e92
·
verified ·
1 Parent(s): 00811da

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. -tAzT4oBgHgl3EQf_P79/content/tmp_files/2301.01947v1.pdf.txt +1113 -0
  2. -tAzT4oBgHgl3EQf_P79/content/tmp_files/load_file.txt +0 -0
  3. -tE2T4oBgHgl3EQfQgaK/content/tmp_files/2301.03771v1.pdf.txt +1284 -0
  4. -tE2T4oBgHgl3EQfQgaK/content/tmp_files/load_file.txt +0 -0
  5. -tE3T4oBgHgl3EQfSwk7/vector_store/index.faiss +3 -0
  6. .gitattributes +69 -0
  7. 0NE2T4oBgHgl3EQf4gjD/vector_store/index.pkl +3 -0
  8. 0dAzT4oBgHgl3EQfRPuW/content/2301.01213v1.pdf +3 -0
  9. 0dAzT4oBgHgl3EQfRPuW/vector_store/index.pkl +3 -0
  10. 0tAzT4oBgHgl3EQf8v4C/content/tmp_files/2301.01907v1.pdf.txt +436 -0
  11. 0tAzT4oBgHgl3EQf8v4C/content/tmp_files/load_file.txt +338 -0
  12. 19E2T4oBgHgl3EQfiwfE/content/2301.03962v1.pdf +3 -0
  13. 1dFKT4oBgHgl3EQfOy1w/content/2301.11760v1.pdf +3 -0
  14. 1dFKT4oBgHgl3EQfOy1w/vector_store/index.pkl +3 -0
  15. 3NE1T4oBgHgl3EQflwRz/content/2301.03289v1.pdf +3 -0
  16. 3dFQT4oBgHgl3EQf3TZi/vector_store/index.faiss +3 -0
  17. 3tE2T4oBgHgl3EQf6AiT/content/2301.04196v1.pdf +3 -0
  18. 3tE2T4oBgHgl3EQf6AiT/vector_store/index.pkl +3 -0
  19. 3tFKT4oBgHgl3EQfRC0N/vector_store/index.faiss +3 -0
  20. 49E3T4oBgHgl3EQfpApV/content/tmp_files/2301.04638v1.pdf.txt +1707 -0
  21. 49E3T4oBgHgl3EQfpApV/content/tmp_files/load_file.txt +0 -0
  22. 4dE0T4oBgHgl3EQfvQGV/content/2301.02616v1.pdf +3 -0
  23. 4dFRT4oBgHgl3EQfpDe0/content/tmp_files/2301.13612v1.pdf.txt +0 -0
  24. 4dFRT4oBgHgl3EQfpDe0/content/tmp_files/load_file.txt +0 -0
  25. 5tAyT4oBgHgl3EQfQPbi/content/tmp_files/2301.00043v1.pdf.txt +1077 -0
  26. 5tAyT4oBgHgl3EQfQPbi/content/tmp_files/load_file.txt +0 -0
  27. 6dAzT4oBgHgl3EQfEfrh/content/tmp_files/2301.00996v1.pdf.txt +1472 -0
  28. 6dAzT4oBgHgl3EQfEfrh/content/tmp_files/load_file.txt +0 -0
  29. 6dE1T4oBgHgl3EQfBQKx/content/2301.02850v1.pdf +3 -0
  30. 6dE1T4oBgHgl3EQfBQKx/vector_store/index.faiss +3 -0
  31. 8dAzT4oBgHgl3EQfgfzo/content/2301.01471v1.pdf +3 -0
  32. 99AzT4oBgHgl3EQfFfoM/content/tmp_files/2301.01010v1.pdf.txt +2493 -0
  33. 99AzT4oBgHgl3EQfFfoM/content/tmp_files/load_file.txt +0 -0
  34. A9AzT4oBgHgl3EQfF_vM/vector_store/index.faiss +3 -0
  35. A9E1T4oBgHgl3EQfVgQy/content/2301.03102v1.pdf +3 -0
  36. C9AyT4oBgHgl3EQf4fqw/content/2301.00788v1.pdf +3 -0
  37. D9E4T4oBgHgl3EQfGQye/content/2301.04893v1.pdf +3 -0
  38. ENE1T4oBgHgl3EQfEQP0/vector_store/index.faiss +3 -0
  39. FNE0T4oBgHgl3EQfzAKR/vector_store/index.pkl +3 -0
  40. G9AyT4oBgHgl3EQfrfnZ/content/tmp_files/2301.00563v1.pdf.txt +1961 -0
  41. G9AyT4oBgHgl3EQfrfnZ/content/tmp_files/load_file.txt +0 -0
  42. G9AzT4oBgHgl3EQfUvyD/content/2301.01272v1.pdf +3 -0
  43. G9AzT4oBgHgl3EQfUvyD/vector_store/index.faiss +3 -0
  44. G9AzT4oBgHgl3EQfUvyD/vector_store/index.pkl +3 -0
  45. GtFIT4oBgHgl3EQfWyuf/content/tmp_files/2301.11241v1.pdf.txt +0 -0
  46. GtFIT4oBgHgl3EQfWyuf/content/tmp_files/load_file.txt +0 -0
  47. J9E1T4oBgHgl3EQfYgQR/content/tmp_files/2301.03138v1.pdf.txt +2168 -0
  48. J9E1T4oBgHgl3EQfYgQR/content/tmp_files/load_file.txt +0 -0
  49. JdAzT4oBgHgl3EQfj_0H/content/tmp_files/2301.01523v1.pdf.txt +1534 -0
  50. JdAzT4oBgHgl3EQfj_0H/content/tmp_files/load_file.txt +0 -0
-tAzT4oBgHgl3EQf_P79/content/tmp_files/2301.01947v1.pdf.txt ADDED
@@ -0,0 +1,1113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.01947v1 [cs.LG] 5 Jan 2023
2
+ StitchNet: Composing Neural Networks from Pre-Trained Fragments
3
+ Surat Teerapittayanon, Marcus Comiter, Brad McDanel, H.T. Kung
4
+ Abstract
5
+ We propose StitchNet, a novel neural network creation
6
+ paradigm that stitches together fragments (one or more con-
7
+ secutive network layers) from multiple pre-trained neural net-
8
+ works. StitchNet allows the creation of high-performing neu-
9
+ ral networks without the large compute and data requirements
10
+ needed under traditional model creation processes via back-
11
+ propagation training. We leverage Centered Kernel Align-
12
+ ment (CKA) as a compatibility measure to efficiently guide
13
+ the selection of these fragments in composing a network for a
14
+ given task tailored to specific accuracy needs and computing
15
+ resource constraints. We then show that these fragments can
16
+ be stitched together to create neural networks with compa-
17
+ rable accuracy to traditionally trained networks at a fraction
18
+ of computing resource and data requirements. Finally, we ex-
19
+ plore a novel on-the-fly personalized model creation and in-
20
+ ference application enabled by this new paradigm.
21
+ 1
22
+ Introduction
23
+ AI models have become increasingly more complex to sup-
24
+ port additional functionality, multiple modalities, and higher
25
+ accuracy. While the increased complexity has improved
26
+ model utility and performance, it has imposed significant
27
+ model training costs. Therefore, training complex models is
28
+ often infeasible for resource limited environments such as
29
+ those at the cloud edge.
30
+ In response to these challenges, in this paper we propose a
31
+ new paradigm for creating neural networks: rather than train-
32
+ ing networks from scratch or retraining them, we create neu-
33
+ ral networks through composition by stitching together frag-
34
+ ments of existing pre-trained neural networks. A fragment is
35
+ one or more consecutive layers of a neural network. We call
36
+ the resulting neural network composed of one or more frag-
37
+ ments a “StitchNet” (Figure 1). By significantly reducing the
38
+ amount of computation and data resources needed for creat-
39
+ ing neural networks, StitchNets enable an entire new set of
40
+ applications, such as rapid generation of personalized neural
41
+ networks at the edge.
42
+ StitchNet’s model creation is fundamentally different
43
+ from today’s predominant backpropagation-based method
44
+ for creating neural networks. Given a dataset and a task
45
+ as input, the traditional training method uses backpropaga-
46
+ tion with stochastic gradient descent (SGD) or other opti-
47
+ mization algorithms to adjust the weights of the neural net-
48
+ works. This training process iterates through the full dataset
49
+ StitchNets
50
+ Fragments
51
+ Existing Networks
52
+ F0 of N0
53
+ F1 of N0
54
+ 0
55
+ F2 of N0
56
+ 1
57
+ F3 of N0
58
+ 2
59
+ F4 of N0
60
+ 3
61
+ F5 of N0
62
+ 4
63
+ F6 of N0
64
+ 5
65
+ F7 of N0
66
+ 6
67
+ F0 of N1
68
+ F1 of N1
69
+ 0
70
+ F2 of N1
71
+ 1
72
+ F3 of N1
73
+ 2
74
+ F4 of N1
75
+ F0 of N1
76
+ F1 of N1
77
+ F2 of N1
78
+ F3 of N1
79
+ F4 of N1
80
+ 3
81
+ F0 of N3
82
+ F1 of N3
83
+ 0
84
+ F2 of N3
85
+ 1
86
+ F3 of N3
87
+ 2
88
+ F4 of N3
89
+ 3
90
+ F5 of N3
91
+ F0 of N3
92
+ F1 of N3
93
+ F2 of N3
94
+ F3 of N3
95
+ F4 of N3
96
+ F5 of N3
97
+ 4
98
+ F0 of N3
99
+ F1 of N3
100
+ 0
101
+ F1 of N1
102
+ 1
103
+ F2 of N1
104
+ 2
105
+ F3 of N1
106
+ 3
107
+ F4 of N1
108
+ 4
109
+ F0 of N1
110
+ F1 of N1
111
+ 0
112
+ F7 of N0
113
+ 1
114
+ F0 of N0
115
+ F1 of N0
116
+ F2 of N0
117
+ F3 of N0
118
+ F4 of N0
119
+ F5 of N0
120
+ F6 of N0
121
+ F7 of N0
122
+ AlexNet
123
+ ResNet
124
+ DenseNet
125
+ Figure 1: Overview of the StitchNet approach. Existing net-
126
+ works (left) are cut into fragments (middle), which are com-
127
+ posed into StitchNets (right) created for a particular task. No
128
+ retraining is needed in this process.
129
+ multiple times, and therefore requires compute resources
130
+ that scale with the amount of data and the complexity of
131
+ the network. Training large models this way also requires
132
+ substantial amounts of data. While successful, this tradi-
133
+ tional paradigm for model creation is not without its limi-
134
+ tations. Creating complex neural networks without access to
135
+ large amounts of data and compute resources is a growing
136
+ challenge of increasing significance, especially in resource-
137
+ constrained edge environments. In the extreme case (e.g., for
138
+ very large language and computer vision models), only a
139
+ few companies with access to unrivaled amounts of data and
140
+ compute resources are able to create such models.
141
+ StitchNets solve this problem by creating new neural net-
142
+ works using fragments of already existing neural networks.
143
+ The new approach takes advantage of the growing amount
144
+ of neural networks that already exist, having been trained
145
+
146
+ previously by many groups and companies. StitchNets en-
147
+ able the efficient reuse of the learned knowledge resident in
148
+ those pre-trained networks, which has been distilled from
149
+ large amounts of data, rather than having to relearn it over
150
+ and over again for new tasks as we do with traditional model
151
+ creation paradigms. StitchNet’s ability to reuse existing pre-
152
+ trained fragments, rather than recreating from scratch or re-
153
+ training for every task will help accelerate the growth and
154
+ application of neural networks for solving more and more
155
+ complex tasks.
156
+ However, compositing these existing fragments into
157
+ a
158
+ coherent
159
+ and
160
+ high
161
+ performing
162
+ neural
163
+ network
164
+ is
165
+ non-trivial. To reuse the knowledge of pre-trained neu-
166
+ ral network fragments, we need a way to 1) measure
167
+ the compatibility between any two fragments, and 2)
168
+ compose compatible fragments together. In the past, Cen-
169
+ tered Kernel Alignment (CKA) (Kornblith et al. 2019;
170
+ Cortes, Mohri, and Rostamizadeh
171
+ 2012;
172
+ Cristianini et al.
173
+ 2006) has been used to measure similarity between neural
174
+ network representations. We leverage CKA to assess the
175
+ compatibility of any two fragments from any neural net-
176
+ works and compose new neural networks from fragments
177
+ of existing pre-trained neural networks to create high
178
+ performing networks customized for specific tasks without
179
+ the costs of traditional model creation methods. The CKA
180
+ score is used to reduce the search space for identifying
181
+ compatible fragments and guide the fragment selection
182
+ process.
183
+ We present empirical validations on benchmark datasets,
184
+ comparing the performance of StitchNets to that of the origi-
185
+ nal pre-trained neural networks. We demonstrate that Stitch-
186
+ Nets achieve comparable or higher accuracy on personalized
187
+ tasks compared with off-the-shelf networks, and have signif-
188
+ icantly lower computational and data requirements than cre-
189
+ ating networks from scratch or through retraining.
190
+ Our contributions are:
191
+ • The StitchNet paradigm: a novel neural network creation
192
+ method that enables a new set of applications.
193
+ • A novel use of Centered Kernel Ailgnment (CKA) in as-
194
+ sessing the compatibility of any two fragments for their
195
+ composition.
196
+ • A technique to compose compatible fragments together
197
+ for both linear and convolutional layers.
198
+ • A feasibility demonstration of StitchNets for efficient on-
199
+ the-fly personalized neural network creation and infer-
200
+ ence.
201
+ 2
202
+ Composing Fragments
203
+ The core mechanism to create StitchNets is to iden-
204
+ tify reusable fragments from a pool of existing net-
205
+ works and compose them into a coherent neural net-
206
+ work model capable of performing a given task. To this
207
+ end, we need a way to determine how compatible any
208
+ two candidate fragments are with each other. In previ-
209
+ ous work, (Kornblith et al. 2019) present centered kernel
210
+ alignment (CKA) (Cortes, Mohri, and Rostamizadeh 2012;
211
+ Cristianini et al. 2006) as a way to measure similarity be-
212
+ tween neural network representations. Rather than looking
213
+ at the neural network as a whole, we adopt and use CKA to
214
+ as a measure of compatibility between any two fragments of
215
+ any neural networks.
216
+ In this section, we first define CKA as a way to measure
217
+ how compatible any two fragments are with one another and
218
+ therefore their ability to be composed. Using CKA, we then
219
+ present a technique to stitch different fragments together. Fi-
220
+ nally, we describe the algorithm to generate StitchNets.
221
+ 2.1
222
+ Centered Kernel Alignment (CKA)
223
+ Given X ∈ Rp×n as outputs of a fragment FA of model A
224
+ and Y ∈ Rq×n as inputs of a fragment FB of model B of
225
+ the same dataset D, where n is the number of samples in the
226
+ dataset, p is the output dimension of FA and q is the input di-
227
+ mension of FB. Let Kij = k(xi, xj) and Mij = m(yi, yj),
228
+ where k and m are any two kernels. We define the compat-
229
+ ibility score CKA(X, Y) of fragment FA and fragment FB
230
+ as
231
+ CKA(X, Y) =
232
+ HSIC(K, M)
233
+
234
+ HSIC(K, K) HSIC(M, M)
235
+ ,
236
+ where HSIC is the Hilbert-Schmidt Independence Criterion
237
+ (Gretton et al. 2005) defined as
238
+ HSIC(K, M) =
239
+ 1
240
+ (n − 1)2 tr(K H M H),
241
+ where H is the centering matrix Hn = In − 1
242
+ n11T and tr is
243
+ the trace. For linear kernels, k(x, y) = m(x, y) = xT y,
244
+ HSIC becomes HSIC(X, Y) = ∥cov(XT X, YT Y)∥2
245
+ F ,
246
+ where cov is the covariance function, and CKA(X, Y) be-
247
+ comes
248
+ ∥cov(XT X, YT Y)∥2
249
+ F
250
+
251
+ ∥cov(XT X, XT X)∥2
252
+ F ∥cov(YT Y, YT Y)∥2
253
+ F
254
+ .
255
+ (1)
256
+ We use this function (Eq. 1) as a measurement of how com-
257
+ patible any two fragments are, given a target dataset. To re-
258
+ duce memory usage for a large target dataset, CKA can be
259
+ approximated by averaging over minibatches as presented in
260
+ (Nguyen, Raghu, and Kornblith 2020).
261
+ 2.2
262
+ Stitching Fragments
263
+ Once we have determined compatible fragments, the next
264
+ step in creating a StitchNet is to stitch the two fragments
265
+ together. To do so, we find a projection tensor A that projects
266
+ the output space of one fragment to the input space of the
267
+ other fragment we are composing. We now describe this.
268
+ Without loss of generality, we assume the output and in-
269
+ put tensors are 2D tensors, where the first dimension is the
270
+ sample dimension. If the tensors are not 2D tensors, we first
271
+ flatten all other dimensions with the exception of the sample
272
+ dimension.
273
+ We use Einstein summation notation, where i represents
274
+ the sample dimension, j the output dimension of the incom-
275
+ ing fragment, and k the input dimension of the outgoing frag-
276
+ ment. Given an output tensor Xij of the incoming fragment
277
+ and an input tensor Yik of the outgoing fragment, we seek
278
+
279
+ to find A such that Yik = Akj Xij . We can then solve for
280
+ A using the Moore-Penrose pseudoinverse:
281
+ Akj = Yik XT
282
+ ij(Xij XT
283
+ ij).−1
284
+ (2)
285
+ Once A is found, we fuse A with the weight of the first
286
+ layer of the outgoing fragment. For linear layers, we simply
287
+ do the following:
288
+ W′
289
+ lk = Wlj Akj,
290
+ (3)
291
+ where l is the dimension of the output feature of the outgoing
292
+ layer.
293
+ For convolutional layers, we first upsample or downsam-
294
+ ple the spatial dimension to match each other, and then ad-
295
+ just the weight along the input channel dimension as follows.
296
+ W′
297
+ okmn = WijmnAkj,
298
+ (4)
299
+ where o is the output channel dimension, j is the original
300
+ input channel dimension, k is the new input channel dimen-
301
+ sion, and m and n are the spatial dimensions.
302
+ For stitching a convolutional layer with an output tensor
303
+ X and a linear layer with an input tensor Y, we first apply
304
+ adaptive average pooling so that the spatial dimension is 1x1
305
+ and flatten X into a 2D tensor. Then, we follow Eq. 2 and
306
+ Eq. 3 to find A and fuse it with the W of the linear layer.
307
+ 2.3
308
+ StitchNet Generation
309
+ Algorithm 1: StitchNet(P, D, K, T , L, R, Q, s)
310
+ Input: fragment pool P = {Fij}, network i in P up to
311
+ layer j Nij, fragment ending in layer j of network i Fij,
312
+ target dataset D with M samples, span K, threshold T ,
313
+ maximum number of fragments L, result list of Stitch-
314
+ Nets and their associated scores R, current StitchNet Q,
315
+ current score s
316
+ Output: resulting list of StitchNets and their associated
317
+ scores R
318
+ if Q is empty then
319
+ {Fij} = select starting fragments in P
320
+ for Fij in {Fij} do
321
+ StitchNet(P, D, K, T , L, R, Fij, 1)
322
+ if the number of fragments in Q ≥ L then
323
+ return R
324
+ {Fij} = select K middle or terminating fragments in P
325
+ for Fij in {Fij} do
326
+ X = Q(D); Y = Nij(D)
327
+ sn = s× CKA(X, Y) (see section 2.1)
328
+ if sn > T then
329
+ Q = Stitch(Q, Fij, X, Y) (see section 2.2)
330
+ if Fij is a terminating fragment then
331
+ R.append({Q, sn})
332
+ else
333
+ StitchNet(P, D, K, T , L, R, Q, sn)
334
+ return R
335
+ We now describe the main algorithm for creating Stitch-
336
+ Net networks (“StitchNets” for short), shown in Algorithm 1.
337
+ A StitchNet network is created by joining a set of pre-trained
338
+ network fragments drawn from a pool P = {Fij}. We use
339
+ the notation Fij to denote a fragment of a neural network i
340
+ up to its j layer, and the notation Nij to denote the compu-
341
+ tation performed by the portion of the neural network from
342
+ which the fragment was taken. Other than the fragment pool
343
+ P and creation process hyperparameters (K, T, L), the only
344
+ other input to the StitchNet creation process is a dataset D
345
+ for which the StitchNet will be optimized.
346
+ We now describe the creation of the pool of network frag-
347
+ ments P derived from a set of pre-trained off-the-shelf net-
348
+ works. These pre-trained networks are divided into one of
349
+ three types of fragments: starting fragments for which the
350
+ input is the original network input, terminating fragments
351
+ for which the output is the original network output, and mid-
352
+ dle fragments that are neither starting nor terminating frag-
353
+ ments.
354
+ The first step in the StitchNet creation process is to choose
355
+ the set of starting fragments. This could include all starting
356
+ fragments in P, or a subset based on certain criteria, e.g., the
357
+ smallest, biggest or closest starting fragment.
358
+ Once a set of starting fragments are selected, a StitchNet
359
+ is built on top of each starting fragment having a current
360
+ starting score of 1. First, a set of K candidate fragments are
361
+ selected from P. These fragments can be selected based on
362
+ CKA scores (i.e., K fragments with highest CKA scores),
363
+ the number of parameters of the fragments (i.e., K frag-
364
+ ments with the least amount of number of parameters in
365
+ P), the closest fragments (i.e., K fragments with the least
366
+ latency in P in a distributed fragments setting), or other se-
367
+ lection methods.
368
+ For each of the candidate fragments, we then compute two
369
+ intermediate neural network computations. First, we pass the
370
+ dataset D through the candidate StitchNet in its current form,
371
+ resulting in value X. Second, we pass the same dataset D
372
+ through the neural network from which the candidate frag-
373
+ ment Fij was selected, resulting in value Y = Nij(D).
374
+ After
375
+ running
376
+ these
377
+ computations,
378
+ we
379
+ produce
380
+ CKA(X, Y) as in Section 2.1. We then multiply the
381
+ CKA with the current score s to obtain the new current
382
+ score sn. If sn is still greater than a set threshold T , the
383
+ candidate fragment is selected and the process continues re-
384
+ cursively. Otherwise, the candidate fragment is rejected. The
385
+ threshold can be set to balance the amount of exploration
386
+ allowed per available compute resources.
387
+ This process continues until a terminating fragment is se-
388
+ lected, the maximum number of fragments L is reached or
389
+ all recursive paths are exhausted. At this point, the com-
390
+ pleted StitchNets and their associated scores R are returned
391
+ for user selection.
392
+ 3
393
+ Results
394
+ We now demonstrate that StitchNets can perform compara-
395
+ bly with traditionally trained networks but with significantly
396
+ reduced computational and data requirements at both infer-
397
+ ence and creation time. Through these characteristics, Stitch-
398
+ Nets enable the immediate on-the-fly creation of neural net-
399
+ works for personalized tasks without traditional training.
400
+
401
+ 3.1
402
+ Fragment pool
403
+ To form the fragment pool P, we take five off-the-shelf net-
404
+ works pre-trained on the ImageNet-1K dataset (Deng et al.
405
+ 2009) from Torchvision (Marcel and Rodriguez 2010):
406
+ alexnet, densenet121, mobilenet v3 small, resnet50 and
407
+ vgg16 with IMAGENET1K V1 weights.
408
+ These pre-trained networks are cut into fragments at
409
+ each convolution and linear layer that has a single in-
410
+ put. As shown in Figure 2, there are 8 fragments for
411
+ alexnet, 5 fragments for densenet121, 13 fragments for mo-
412
+ bilenet v3 small, 6 fragments for resnet50 and 16 fragments
413
+ for vgg16. This results in the creation of a fragment pool P
414
+ of 48 fragments consisting of 5 starting fragments, 38 mid-
415
+ dle fragments, and 5 terminating fragments. We use this frag-
416
+ ment pool in all experiments in this paper.
417
+ 3.2
418
+ Dataset
419
+ The dataset used to evaluate StitchNets in this paper is the
420
+ “Dogs vs. Cats” dataset (Kaggle 2013). This dataset includes
421
+ 25,000 training images of dogs and cats and we use an
422
+ 80:20 train:test split. We map ImageNet-1K class labels into
423
+ cat and dog labels (class IDs 281-285 and 151-250, respec-
424
+ tively). To form the target dataset D for use in the stitch-
425
+ ing process of Algorithm 1, we randomly select M samples
426
+ from the training set as the target dataset D. We choose this
427
+ task because it is characteristic of the type of task for which
428
+ StitchNets would be used: a user needs a custom classifier
429
+ for a particular task and desired set of classes.
430
+ 3.3
431
+ StitchNet Generation
432
+ We generate StitchNets with Algorithm 1 using the fragment
433
+ pool and the dataset described in Section 3.1 and 3.2. We set
434
+ K = 2, T = 0.5 and L = 16. The number of samples M in
435
+ D used for the stitching process is 32.
436
+ Given these hyperparameters, a total of 89 StitchNets are
437
+ generated. We evaluate them on the test set of completely
438
+ unseen test samples. Summary statistics for the generated
439
+ StitchNets are shown in Figure 3, including accuracy (3a),
440
+ number of fragments per StitchNet (3b), CKA score (3c),
441
+ and number of parameters per StitchNet (3d).
442
+ 3.4
443
+ Reduction in Inference Computation
444
+ We now demonstrate how StitchNets significantly reduce
445
+ inference-time computational requirements over traditional
446
+ neural network training paradigms by studying StitchNet ac-
447
+ curacy as a function of parameters.
448
+ Figure 4 shows the resulting accuracy of the generated
449
+ StitchNets as a function overall CKA score for each Stitch-
450
+ Net and number of parameters (porportional to marker size)
451
+ as a proxy for inference-time computation cost. We find a
452
+ number of StitchNets outperform the pre-trained network
453
+ while realizing significant computational savings. For exam-
454
+ ple, StitchNet27 (denoted by a green star) achieves an ac-
455
+ curacy of 0.86 with 3.59M parameters compared with the
456
+ 0.70 accuracy of the pre-trained alexnet with 61.10M param-
457
+ eters. Therefore, StitchNet achieves a 22.8% increase in ac-
458
+ curacy with a 94.1% reduction in number of parameters for
459
+ alexnet
460
+ densenet121
461
+ mobilenet
462
+ resnet50
463
+ vgg16
464
+ Figure 2: Five pre-trained networks are fragmented into a
465
+ fragment pool P. These fragments will be stitched together
466
+ to form StitchNets.
467
+ the given task when compared with those of the pre-trained
468
+ alexnet.
469
+ These crystallizes one of the core benefits of StitchNets:
470
+ without any training, the method can discover networks that
471
+ are personalized for the task, outperform the original pre-
472
+ trained networks, and do so while significantly reducing
473
+ inference-time compute requirements. This is due to the fact
474
+ that these pre-trained networks are not trained to focus on
475
+ these two specific classes, while our StitchNets are stitched
476
+ together specifically for the task. In the next section, we will
477
+
478
+ F15ofN4
479
+ 14
480
+ F14 0fN4
481
+ 13
482
+ F130fN4
483
+ F120fN4
484
+ 11
485
+ lio
486
+ FIOofN4
487
+ F8ofN4
488
+ F7ofN4
489
+ F5ofN4
490
+ F4ofN4F5ofN3
491
+ F4ofN3
492
+ F36fN3
493
+ F20fN3
494
+ FIofN3
495
+ FOOfN3F12ofN2
496
+ 11
497
+ F11ofN2
498
+ 10
499
+ F10OfN2
500
+ 1oofN2
501
+ F8ofN2
502
+ F7ofN2
503
+ F6ofN2
504
+ F5ofN2
505
+ F40fN2
506
+ F3ofN2
507
+ F2ofN2
508
+ F1ofN2
509
+ FOofN2F4OfNI
510
+ F3OfNI
511
+ F2ofNI
512
+ FIOfNI
513
+ FOOFNIF7ofNO
514
+ F6ofNO
515
+ F5ofNo
516
+ F4of NO
517
+ F3OfNO
518
+ F2ofNO
519
+ FIofNo
520
+ FOof No0.67
521
+ to
522
+ 0.73
523
+ 0.73
524
+ to
525
+ 0.78
526
+ 0.78
527
+ to
528
+ 0.84
529
+ 0.84
530
+ to
531
+ 0.90
532
+ 0.90
533
+ to
534
+ 0.95
535
+ 9
536
+ 19
537
+ 27
538
+ 25
539
+ 9
540
+ (a) accuracy
541
+ 3
542
+ 4
543
+ 5
544
+ 6
545
+ 7
546
+ 8
547
+ 9
548
+ 10
549
+ 11
550
+ 12
551
+ 13
552
+ 14
553
+ 15
554
+ 16
555
+ 1
556
+ 2
557
+ 6
558
+ 6
559
+ 3
560
+ 3
561
+ 3
562
+ 4
563
+ 8
564
+ 8
565
+ 11
566
+ 10
567
+ 8
568
+ 16
569
+ (b) # fragments
570
+ 0.50
571
+ to
572
+ 0.60
573
+ 0.60
574
+ to
575
+ 0.70
576
+ 0.70
577
+ to
578
+ 0.80
579
+ 0.80
580
+ to
581
+ 0.90
582
+ 0.90
583
+ to
584
+ 1.00
585
+ 29
586
+ 28
587
+ 14
588
+ 11
589
+ 7
590
+ (c) CKA score
591
+ 1M
592
+ to
593
+ 21M
594
+ 21M
595
+ to
596
+ 42M
597
+ 42M
598
+ to
599
+ 62M
600
+ 62M
601
+ to
602
+ 83M
603
+ 83M
604
+ to
605
+ 104M
606
+ 104M
607
+ to
608
+ 124M
609
+ 124M
610
+ to
611
+ 145M
612
+ 66
613
+ 5
614
+ 4
615
+ 5
616
+ 2
617
+ 3
618
+ 4
619
+ (d) # parameters
620
+ Figure 3: Histogram of (a) accuracy, (b) # fragments, (c)
621
+ CKA score, (d) # parameters in the generated batch of Stitch-
622
+ Nets.
623
+ see that very little data is required for the stitching process.
624
+ Additionally, we compare the StitchNets with the var-
625
+ ious off-the-shelf models, denoted by triangles. We find
626
+ that the StitchNet generation process creates many different
627
+ StitchNets that outperform the off-the-shelf models, many of
628
+ which do so at reduced computational cost. Figure 5 shows
629
+ the composition of some of these high-performing Stitch-
630
+ Nets, demonstrating the diversity in fragment use, ordering,
631
+ and architectures.
632
+ We also validate the effectiveness of using CKA to guide
633
+ the stitching procedure. We find that StitchNets with a high
634
+ CKA score also have high accuracy, especially those above
635
+ 0.9. This shows that CKA can be used as a proxy to measure
636
+ good compatibility between connecting fragments.1
637
+ 3.5
638
+ Reduction in Network Creation Computation
639
+ We now demonstrate that StitchNets can be created without
640
+ significant data and computation requirements. Specifically,
641
+ we compare StitchNet21 (generated in Figure 5 on the tar-
642
+ get dataset of M = 32 samples) with fine-tuning the same
643
+ five off-the-shelf networks (retraining them using the train-
644
+ ing portion of dataset of Section 3.2). For fine-tuning, we
645
+ replace and train only the last layer of the pre-trained net-
646
+ work using Stochastic Gradient Descent (SGD) with batch
647
+ size 32, learning rate 0.001 and momentum 0.9. The results
648
+ shown are averaged over 10 runs. For ease of comparison,
649
+ we normalize the computation cost in terms of the num-
650
+ ber of samples processed through a neural network. In prac-
651
+ 1Note that there exist high accuracy StitchNets with low overall
652
+ CKA score. This is because neural networks are robust and highly
653
+ redundant, able to tolerate a certain amount of errors while still
654
+ giving quality predictions (see Section 4.1).
655
+ tice, fine-tuning requires backpropagation, which incurs ad-
656
+ ditional computation per sample than StitchNet generation.
657
+ Figure 6 compares the accuracy of StitchNet21 (denoted
658
+ by the red star) with the traditionally fine-tuned networks
659
+ as a function of the number of training samples processed.
660
+ For a given accuracy target, StitchNets process a substan-
661
+ tially smaller number of data samples than traditionally fine-
662
+ tuned networks. Specifically, to reach an accuracy of 0.95,
663
+ fine-tuning of alexnet, densenet121, and mobilenet v3 small
664
+ require to process more than 320 samples while StitchNet re-
665
+ quires only 32 samples used to stitch the fragments together
666
+ (realizing over a 90% reduction).
667
+ Therefore, only a small amount of training samples and
668
+ computation are required for StitchNet to achieve compara-
669
+ ble accuracy. This demonstrates that StitchNets effectively
670
+ reuse the information already captured in the fragments to
671
+ bootstrap network creation. This allows for personalization
672
+ of tasks and on-the-fly training without substantial data re-
673
+ quirements.
674
+ 3.6
675
+ Ensembles
676
+ We now discuss the ability to ensemble generated StitchNets
677
+ to improve performance. StitchNet and ensembling methods
678
+ are complimentary. The StitchNet generation algorithm pro-
679
+ duces a set of candidate models. While a user can select a
680
+ single StitchNet to use at inference time, because the Stitch-
681
+ Net generation procedure finds such efficient models, we can
682
+ also take advantage of the pool of StitchNets and ensemble
683
+ some while still realize substantial computational savings.
684
+ We pick 10 random models from the generated StitchNets
685
+ in Section 3.3 with overall CKA > 0.8. We sort these mod-
686
+ els based on their overall CKA scores from high to low, and
687
+ then ensemble them by averaging their predicted probabili-
688
+ ties. The results are shown in Figure 7. The ensemble often
689
+ results in higher accuracy than the individual model. As a re-
690
+ sult, this ensembling method can reduce variance in perfor-
691
+ mance when on-the-fly network creation and inference (as
692
+ discussed in Section 4.3) is used and there is not time for full
693
+ selection of a final single StitchNet. Instead, the user can se-
694
+ lect a reasonably small subset of high performing StitchNets,
695
+ which even in aggregate can be significantly smaller than a
696
+ single traditionally trained network.
697
+ 4
698
+ Discussion
699
+ We now discuss the intuition behind StitchNets, examine
700
+ their complexity and relation to related methods, introduce
701
+ new applications they enable, and discuss their limitations.
702
+ 4.1
703
+ Why do StitchNets work?
704
+ We first discuss why we are able to reuse existing fragments
705
+ of networks to create new neural networks without retrain-
706
+ ing. One core reason for this is that neural networks tend to
707
+ learn fundamental and universal features. Studies (Li et al.
708
+ 2015; Lu et al. 2018; Morcos, Raghu, and Bengio 2018;
709
+ Wang et al. 2018; Lenc and Vedaldi 2015; Kornblith et al.
710
+ 2019; Tang et al. 2020) have shown that neural networks
711
+ learn fundamental features such as edges for different tasks.
712
+ Since these learned features are fundamental, they should
713
+
714
+ 0.5
715
+ 0.6
716
+ 0.7
717
+ 0.8
718
+ 0.9
719
+ 1.0
720
+ Overall CKA score
721
+ 0.70
722
+ 0.75
723
+ 0.80
724
+ 0.85
725
+ 0.90
726
+ 0.95
727
+ Accuracy
728
+ Smallest
729
+ acc=0.73
730
+ cka=0.53
731
+ 0.57M
732
+ Best
733
+ acc=0.95
734
+ cka=0.91
735
+ 8.04M
736
+ StitchNet27
737
+ acc=0.86
738
+ cka=0.94
739
+ 3.59M
740
+ alexnet
741
+ acc=0.70
742
+ cka=0.89
743
+ 61.10M
744
+ densenet121
745
+ acc=0.85
746
+ cka=1.00
747
+ 8.04M
748
+ mobilenet_v3_small
749
+ acc=0.78
750
+ cka=1.00
751
+ 2.54M
752
+ resnet50
753
+ acc=0.85
754
+ cka=0.99
755
+ 25.53M
756
+ vgg16
757
+ acc=0.81
758
+ cka=0.85
759
+ 138.36M
760
+ Figure 4: Accuracy vs the overall CKA score on “Cat vs. Dogs.” cka is the overall CKA score, acc is the accuracy. The best
761
+ StitchNet (acc=0.95) performs 12% better than the best pre-trained model(s) (densenet121 and resnet50 with acc=0.85).
762
+ StitchNet21
763
+ acc=0.95
764
+ cka=0.91
765
+ 8.04M
766
+ StitchNet22
767
+ acc=0.89
768
+ cka=0.84
769
+ 5.33M
770
+ StitchNet5
771
+ acc=0.82
772
+ cka=0.81
773
+ 61.10M
774
+ StitchNet32
775
+ acc=0.79
776
+ cka=0.88
777
+ 1.99M
778
+ StitchNet88
779
+ acc=0.78
780
+ cka=0.77
781
+ 8.15M
782
+ Figure 5: Examples of generated StitchNets.
783
+ be reusable rather relearned. The challenge, however, is that
784
+ although these features may be universal, they may not be
785
+ compatible with one another “out of the box.” Therefore,
786
+ we require the stitching process introduced in Section 2.2
787
+ to project the fragments into a compatible space.
788
+ 50
789
+ 100
790
+ 150
791
+ 200
792
+ 250
793
+ 300
794
+ The number of training samples processed
795
+ 0.0
796
+ 0.2
797
+ 0.4
798
+ 0.6
799
+ 0.8
800
+ 1.0
801
+ Accuracy
802
+ StitchNet21
803
+ acc@32=0.95
804
+ alexnet
805
+ acc@320=0.93±0.01
806
+ densenet121
807
+ acc@320=0.90±0.04
808
+ mobilenet_v3_small
809
+ acc@320=0.93±0.01
810
+ resnet50
811
+ acc@320=0.97±0.01
812
+ vgg16
813
+ acc@320=0.97±0.00
814
+ Figure 6: Accuracy vs the number of training samples pro-
815
+ cessed (i.e., data and computation required). StitchNets re-
816
+ quire only a fraction of the computation of traditional train-
817
+ ing methods to achieve comparable performance.
818
+ 1
819
+ 2
820
+ 3
821
+ 4
822
+ 5
823
+ 6
824
+ 7
825
+ 8
826
+ 9
827
+ 10
828
+ Model in the ensemble
829
+ 0.800
830
+ 0.825
831
+ 0.850
832
+ 0.875
833
+ 0.900
834
+ 0.925
835
+ Accuracy
836
+ Ensemble Accuracy
837
+ Individual Accuracy
838
+ Figure 7: Accuracy of the ensemble models. Ensembling
839
+ groups of StitchNets can reduce individual model variance.
840
+ Beyond this reuse of universal features and compatibility
841
+ transformations, StitchNets are also enabled by the fact that
842
+ neural networks are fundamentally robust. Due to the non-
843
+ linear activation and built-in redundancies, neural networks
844
+ tolerate certain amounts of error. As such, the fragments
845
+ need not be perfectly compatible individually to produce a
846
+ network that in aggregate operates at a high level of perfor-
847
+ mance.
848
+ 4.2
849
+ Complexity Comparison
850
+ We now compare the complexity of the traditional train-
851
+ ing process using backpropagation with the StitchNet gen-
852
+
853
+ F120fN2
854
+ F9of N4
855
+ F8ofN4
856
+ F7ofN4
857
+ F6ofN4
858
+ F5ofN4
859
+ F4of N4
860
+ F3ofN4
861
+ F2 ofN4
862
+ FIofN4
863
+ FOof N4F12 of N2
864
+ 10
865
+ F11 of N2
866
+ F9 of N2
867
+ F8 of N2
868
+ F7 of N2
869
+ F6 of N2
870
+ F5 of N2
871
+ F4 of N2
872
+ F3 of N2
873
+ F2 of N2
874
+ F1 of N2
875
+ FO of N2F12ofN2
876
+ F6of NO
877
+ F5ofNO
878
+ F4of NO
879
+ F3ofNO
880
+ F2of NO
881
+ FIof NO
882
+ FOof NOF12.0fN2
883
+ F2OfNI
884
+ FIofNI
885
+ 0
886
+ FOofNIF5ofN3
887
+ F3ofNI
888
+ F2ofNI
889
+ FIOfNI
890
+ FOofNIeration process. Traditional training complexity is O(ndp),
891
+ where n is the number of parameters in the network, p is
892
+ the number of epochs used to train, and d is the size of
893
+ the dataset. StitchNet generation complexity is O(nqm) +
894
+ O(KL). The first term nqm is the evaluation cost of the tar-
895
+ get dataset of size q on m networks in the pool, where q ≪ d
896
+ and n is the number of parameters in the network (assuming
897
+ networks have the same # of parameters). The second term
898
+ KL is the search cost, where K is the span value we search
899
+ at each level and L is the max depth to search. Using a high
900
+ threshold cutoff T on the overall CKA score keeps search
901
+ cost KL small. Therefore, for a reasonable setting of hyper-
902
+ parameters (K, T, L) in Algorithm 1, StitchNets realize sub-
903
+ stantial computation gains over traditional training methods
904
+ since q ≪ d and m ≪ p.
905
+ 4.3
906
+ On-the-fly network creation and inference
907
+ We now discuss a new family of applications and use cases
908
+ that are enabled by StitchNets: on-the-fly neural network cre-
909
+ ation and inference. In this application, we use a batch of im-
910
+ ages on which we want to perform a task (e.g., classification
911
+ or detection) as our target dataset in the StitchNet generation
912
+ process. With only a minor modification to the StitchNet al-
913
+ gorithm to additionally return task results, the StitchNet gen-
914
+ eration process can return the inference outputs along with
915
+ the generated StitchNets.
916
+ We now describe how this can be used in practice. Imag-
917
+ ine a world where fragments of pre-trained neural networks
918
+ for different tasks are indexed and distributed on the Inter-
919
+ net. Any compatible fragment can be found and composed
920
+ quickly to form a new neural network for a certain task. Now,
921
+ imagine we want to create a neural network for classifying
922
+ local cats and dogs with only a few hundred of these unla-
923
+ beled images.
924
+ Without StitchNets, we either need to train a network
925
+ from scratch (which may fail due to our limited amount of
926
+ training data), or find an existing pre-trained neural network,
927
+ label the dataset, and finetune the network. If the existing
928
+ pre-trained network is too big or too slow for our use, we
929
+ will then have to train a new one from scratch. But, with lim-
930
+ ited amount of unlabeled data, this task seems impossible.
931
+ With StitchNet, we can instead generate a set of candidate
932
+ StitchNets with the small target dataset of unlabeled local
933
+ cats and dogs images. These StitchNets are created from the
934
+ pool of existing neural network fragments that have been in-
935
+ dexed and distributed on the Internet. The proper fragments
936
+ can be identified with a search criteria (e.g., the terminat-
937
+ ing fragment should contain cat and dog classes, the depth
938
+ of the network should be less than 5 for computational effi-
939
+ ciency reasons, etc.). With little computation, we will gener-
940
+ ate StitchNets capable of detecting and classifying local cats
941
+ and dogs.
942
+ 4.4
943
+ Limitations
944
+ One limitation is that the target task needs to be a subset
945
+ (or a composition) of the terminating fragment tasks in the
946
+ fragment pool. Additionally, while a large pool of networks
947
+ and fragments can lead to higher applicability and quality
948
+ of StitchNets, it can also lead to high search costs. Index-
949
+ ing large quantities of neural networks to form the fragment
950
+ pool will require novel search methods. We see this as anal-
951
+ ogous to indexing web pages on the World Wide Web, sug-
952
+ gesting a “Google for Fragments.” Much like web search
953
+ needed to index written content, large amounts of neural net-
954
+ work “content” need to be indexed in order for their value to
955
+ be unlocked. Early indexing efforts can tag fragments based
956
+ on dataset characteristics, computational characteristics, etc.
957
+ More advanced efforts can look at inward and outward con-
958
+ nections of each fragment to determine its rank in results.
959
+ Once a narrowed set of fragments are coarsely identified, the
960
+ efficient procedure introduced in this paper can generate the
961
+ StitchNets. Future work will address these types of comple-
962
+ mentary methods (indexing and distribution) that will enable
963
+ StitchNets to operate at scale.
964
+ 5
965
+ Related Work
966
+ Transfer learning (or fine-tuning) (Pan and Yang 2009;
967
+ Weiss, Khoshgoftaar, and Wang 2016) is the current pre-
968
+ dominant way to adapt existing neural networks to target
969
+ tasks. Unsupervised domain adaptation is related, where the
970
+ existing network is adapted using an unlabeled target dataset.
971
+ StitchNets work similarly by stitching fragments using an
972
+ unlabeled target dataset to create a neural network for the
973
+ target task. Most work (Wang and Deng 2018; Zhang et al.
974
+ 2018; Tzeng et al. 2014; Kumar et al. 2018; Shu et al. 2018;
975
+ Ben-David et al. 2010; Saito, Ushiku, and Harada 2017) fo-
976
+ cuses on retraining the network, while StitchNet does not
977
+ require any training.
978
+ StitchNets take advantage of the assumption that the frag-
979
+ ments have shareable representations. This assumption helps
980
+ explain why fragments can be stitched together into a coher-
981
+ ent high-performing network: dissimilar yet complimentary
982
+ fragments once projected into a similar space are compatible
983
+ with one another. Several existing works including (Li et al.
984
+ 2015; Mehrer, Kriegeskorte, and Kietzmann 2018; Lu et al.
985
+ 2018; Morcos, Raghu, and Bengio 2018; Wang et al. 2018;
986
+ Lenc and Vedaldi 2015; Kornblith et al. 2019; Tang et al.
987
+ 2020) have studied this shareable representation assumption.
988
+ (Gygli, Uijlings, and Ferrari 2021) reuse network compo-
989
+ nents by training networks to produce compatible features
990
+ by adding regularization at training time to make the net-
991
+ works directly compatible. StitchNet, however, focuses on
992
+ creating neural networks without training. It is therefore
993
+ more generally applicable. (Lenc and Vedaldi 2015) com-
994
+ bine network components by adding a stitching layer and
995
+ training the recombined network with a supervised loss for
996
+ several epochs. StitchNet adds a parameter-less stitching
997
+ mechanism and therefore does not require any retraining. In-
998
+ stead, weights are adapted to be compatible with the method
999
+ introduced in 2.2.
1000
+ 6
1001
+ Conclusion
1002
+ StitchNet is a new paradigm that can leverage a growing
1003
+ global library of neural networks to fundamentally change
1004
+ the way networks are created. By reusing fragments of these
1005
+ networks to efficiently compose new networks for a given
1006
+
1007
+ task, StitchNet addresses two of the most fundamental is-
1008
+ sues limiting neural network creation and use: large data and
1009
+ computation requirements.
1010
+ StitchNet does this by leveraging Centered Kernel Align-
1011
+ ment (CKA) as a compatibility measure that guides the se-
1012
+ lection of neural network fragments, tailored to specific ac-
1013
+ curacy needs and computing resource constraints. Our work
1014
+ has shown that neural networks can be efficiently created
1015
+ from compatible neural network fragments of different mod-
1016
+ els at a fraction of computing resources and data require-
1017
+ ments with a comparable accuracy. We also introduce on-
1018
+ the-fly efficient neural network creation and inference appli-
1019
+ cation that is unlocked by this method.
1020
+ References
1021
+ Ben-David, S.; Blitzer, J.; Crammer, K.; Kulesza, A.;
1022
+ Pereira, F.; and Vaughan, J. W. 2010. A theory of learning
1023
+ from different domains. Machine learning, 79(1): 151–175.
1024
+ Cortes, C.; Mohri, M.; and Rostamizadeh, A. 2012. Algo-
1025
+ rithms for learning kernels based on centered alignment. The
1026
+ Journal of Machine Learning Research, 13(1): 795–828.
1027
+ Cristianini, N.; Kandola, J.; Elisseeff, A.; and Shawe-Taylor,
1028
+ J. 2006. On kernel target alignment. In Innovations in ma-
1029
+ chine learning, 205–256. Springer.
1030
+ Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-
1031
+ Fei, L. 2009. Imagenet: A large-scale hierarchical image
1032
+ database. In 2009 IEEE conference on computer vision and
1033
+ pattern recognition, 248–255. Ieee.
1034
+ Gretton, A.; Bousquet, O.; Smola, A.; and Sch¨olkopf,
1035
+ B. 2005.
1036
+ Measuring statistical dependence with Hilbert-
1037
+ Schmidt norms. In International conference on algorithmic
1038
+ learning theory, 63–77. Springer.
1039
+ Gygli, M.; Uijlings, J.; and Ferrari, V. 2021.
1040
+ Towards
1041
+ reusable network components by learning compatible rep-
1042
+ resentations. In Proceedings of the AAAI Conference on Ar-
1043
+ tificial Intelligence, volume 35, 7620–7629.
1044
+ Kaggle. 2013. Dogs vs. cats.
1045
+ Kornblith, S.; Norouzi, M.; Lee, H.; and Hinton, G. 2019.
1046
+ Similarity of neural network representations revisited. In
1047
+ International Conference on Machine Learning, 3519–3529.
1048
+ PMLR.
1049
+ Kumar, A.; Sattigeri, P.; Wadhawan, K.; Karlinsky, L.;
1050
+ Feris, R.; Freeman, W. T.; and Wornell, G. 2018.
1051
+ Co-
1052
+ regularized alignment for unsupervised domain adaptation.
1053
+ arXiv preprint arXiv:1811.05443.
1054
+ Lenc, K.; and Vedaldi, A. 2015. Understanding image repre-
1055
+ sentations by measuring their equivariance and equivalence.
1056
+ In Proceedings of the IEEE conference on computer vision
1057
+ and pattern recognition, 991–999.
1058
+ Li, Y.; Yosinski, J.; Clune, J.; Lipson, H.; Hopcroft, J. E.;
1059
+ et al. 2015. Convergent learning: Do different neural net-
1060
+ works learn the same representations? In FE@ NIPS, 196–
1061
+ 212.
1062
+ Lu, Q.; Chen, P.-H.; Pillow, J. W.; Ramadge, P. J.; Nor-
1063
+ man, K. A.; and Hasson, U. 2018.
1064
+ Shared representa-
1065
+ tional geometry across neural networks.
1066
+ arXiv preprint
1067
+ arXiv:1811.11684.
1068
+ Marcel, S.; and Rodriguez, Y. 2010.
1069
+ Torchvision the
1070
+ machine-vision package of torch. In Proceedings of the 18th
1071
+ ACM international conference on Multimedia, 1485–1488.
1072
+ Mehrer, J.; Kriegeskorte, N.; and Kietzmann, T. 2018. Be-
1073
+ ware of the beginnings: intermediate and higherlevel repre-
1074
+ sentations in deep neural networks are strongly affected by
1075
+ weight initialization. In Conference on Cognitive Computa-
1076
+ tional Neuroscience.
1077
+ Morcos, A. S.; Raghu, M.; and Bengio, S. 2018. Insights on
1078
+ representational similarity in neural networks with canonical
1079
+ correlation. arXiv preprint arXiv:1806.05759.
1080
+ Nguyen, T.; Raghu, M.; and Kornblith, S. 2020. Do wide
1081
+ and deep networks learn the same things? uncovering how
1082
+ neural network representations vary with width and depth.
1083
+ arXiv preprint arXiv:2010.15327.
1084
+ Pan, S. J.; and Yang, Q. 2009. A survey on transfer learn-
1085
+ ing. IEEE Transactions on knowledge and data engineering,
1086
+ 22(10): 1345–1359.
1087
+ Saito, K.; Ushiku, Y.; and Harada, T. 2017.
1088
+ Asymmet-
1089
+ ric tri-training for unsupervised domain adaptation. In In-
1090
+ ternational Conference on Machine Learning, 2988–2997.
1091
+ PMLR.
1092
+ Shu, R.; Bui, H. H.; Narui, H.; and Ermon, S. 2018. A dirt-t
1093
+ approach to unsupervised domain adaptation. arXiv preprint
1094
+ arXiv:1802.08735.
1095
+ Tang, S.; Maddox, W. J.; Dickens, C.; Diethe, T.; and Dami-
1096
+ anou, A. 2020. Similarity of neural networks with gradients.
1097
+ arXiv preprint arXiv:2003.11498.
1098
+ Tzeng, E.; Hoffman, J.; Zhang, N.; Saenko, K.; and Darrell,
1099
+ T. 2014. Deep domain confusion: Maximizing for domain
1100
+ invariance. arXiv preprint arXiv:1412.3474.
1101
+ Wang, L.; Hu, L.; Gu, J.; Wu, Y.; Hu, Z.; He, K.; and
1102
+ Hopcroft, J. 2018. Towards understanding learning repre-
1103
+ sentations: To what extent do different neural networks learn
1104
+ the same representation. arXiv preprint arXiv:1810.11750.
1105
+ Wang, M.; and Deng, W. 2018. Deep visual domain adapta-
1106
+ tion: A survey. Neurocomputing, 312: 135–153.
1107
+ Weiss, K.; Khoshgoftaar, T. M.; and Wang, D. 2016. A sur-
1108
+ vey of transfer learning. Journal of Big data, 3(1): 1–40.
1109
+ Zhang, W.; Ouyang, W.; Li, W.; and Xu, D. 2018. Collabora-
1110
+ tive and adversarial network for unsupervised domain adap-
1111
+ tation. In Proceedings of the IEEE conference on computer
1112
+ vision and pattern recognition, 3801–3809.
1113
+
-tAzT4oBgHgl3EQf_P79/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
-tE2T4oBgHgl3EQfQgaK/content/tmp_files/2301.03771v1.pdf.txt ADDED
@@ -0,0 +1,1284 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Chatbots in a Honeypot World
2
+ Forrest McKee1 and David Noever2
3
+ PeopleTec, 4901-D Corporate Drive, Huntsville, AL, USA, 35805
4
+ 1forrest.mckee@peopletec.com 2 david.noever@peopletec.com
5
+
6
+
7
+ Abstract
8
+ Question-and-answer agents like ChatGPT offer a novel tool for use as a potential honeypot interface in
9
+ cyber security. By imitating Linux, Mac, and Windows terminal commands and providing an interface for
10
+ TeamViewer, nmap, and ping, it is possible to create a dynamic environment that can adapt to the actions
11
+ of attackers and provide insight into their tactics, techniques, and procedures (TTPs). The paper illustrates
12
+ ten diverse tasks that a conversational agent or large language model might answer appropriately to the
13
+ effects of command-line attacker. The original result features feasibility studies for ten model tasks meant
14
+ for defensive teams to mimic expected honeypot interfaces with minimal risks. Ultimately, the usefulness
15
+ outside of forensic activities stems from whether the dynamic honeypot can extend the time-to-conquer or
16
+ otherwise delay attacker timelines short of reaching key network assets like databases or confidential
17
+ information. While ongoing maintenance and monitoring may be required, ChatGPT's ability to detect and
18
+ deflect malicious activity makes it a valuable option for organizations seeking to enhance their cyber
19
+ security posture. Future work will focus on cybersecurity layers, including perimeter security, host virus
20
+ detection, and data security.
21
+ Keywords:
22
+ Transformers, Text Generation, Malware Generation, Generative Pre-trained Transformers, GPT
23
+
24
+ 1. INTRODUCTION
25
+
26
+ A honeypot is a significant cyber security tool that is used to detect, deflect, and study malicious activity
27
+ on a computer network [1-4]. It is essentially a trap set up to lure in potential attackers, who are then
28
+ observed and their actions are recorded for later threat analysis. Honeypots can be used in a variety of ways,
29
+ including for research, to gather intelligence on new or emerging threats, or to distract and mislead attackers
30
+ while security teams work to defend against an ongoing attack [1]. A spectrum exists between low-
31
+ interaction honeypots that may expose only ports and no real services to high-interaction honeypots that
32
+ virtualize entire networks using VMWare or User-mode Linux with application-, network- and system-
33
+ layer features [5]. Making realistic traps relies on the realism of the honeypot. Attackers may quickly
34
+ discover the static elements or missing functional files that tip off a fake asset or operating system façade.
35
+ Probing services and ports can reveal a fake network asset [6-7]. The rise of cloud and virtual machine
36
+ images has exacerbated the challenge to mimic real networks with a passive store-front approach [2]. More
37
+ dynamic approaches to building honeypots that feature real applications but host fake data [6]. An example
38
+ dynamic honeypot deploys a real SQL database capable of real hacking attempts, all of which culminate in
39
+ revealing fake personnel or salary data. A hybrid version of the real vs. simulated honeypot problem
40
+ involves creating a digital twin that behaves like the real network but which underneath remains a
41
+ simulation based on a large language model [8] that anticipates the output of the operating system and
42
+ applications [9]. This hardware and software stack together presents a sufficiently deep environment that
43
+ a large language model simulates the expected outcomes when queried by an intruder [9-11]. This hybrid
44
+ option provides a novel experimental platform for the current study and assessments of its capabilities.
45
+
46
+ In this paper, we will explore the concept of using ChatGPT, a natural language processing tool [12-14], as
47
+ a honeypot in the field of cyber security. One potential use of ChatGPT as a honeypot is to issue various
48
+ commands that simulate Linux [9] and Windows terminals. This can be used to lure in attackers who are
49
+
50
+ specifically targeting these types of systems, and allow security teams to observe and study their actions
51
+ [15-16]. By issuing commands through ChatGPT, it is possible to create a realistic and dynamic
52
+ environment that can adapt to the actions of the attacker [6]. As an attacker explores this new network asset,
53
+ their commands reveal ever more sophisticated emulation patterns derived from the internet-scale training
54
+ data underpinning the OpenAI GPT series of transformer architectures [13]. Historically, honeypot logs
55
+ provide valuable insights into the tactics, techniques, and procedures (TTPs) used by attackers, as well as
56
+ help security teams to identify patterns and trends in malicious activity [17-20]. Additionally, issuing
57
+ commands through ChatGPT can also help to distract and mislead attackers, giving security teams more
58
+ time to defend against an ongoing attack. The latest generation of ChatGPT (Dec 2022 update) [21] now
59
+ sustains its memory of initial instructions for up to 8000 tokens (or around 5600 words, 20-25 text pages).
60
+ To translate this coherent “command-driven” conversation to a typical intrusion, the attacker might interact
61
+ with emulated honeypot (aka, chatbot interface) for hours before the simulation required an instructional
62
+ reset.
63
+
64
+ 2. METHODS
65
+
66
+ The structure of the paper closely follows the detailed instructions and attacker interactions outlined in
67
+ Appendices A-J as ten tasks related to honeypot construction, detection, or harvesting [1]. As shown in
68
+ Table 1, each appendix section outlines the initial ChatGPT instructions or prompt followed by a simple
69
+ proof of principle illustrating the degree of dynamic emulation achievable. The ten tasks demonstrate
70
+ plausible command-level interactions with an adversary who breaches a network consisting of all major
71
+ operating systems (Windows, Linux, Mac). We simulate application-level interactions with a python-
72
+ driven Jupyter notebook and a Team Viewer installation. We simulate network-level interactions using
73
+ network mapping tools (nmap) and launch a simulated distributed denial-of-service (DDoS) attack using
74
+ ping. We simulate an attacker’s deception by changing the time-stamp on a malicious file (“time-
75
+ stomping”) so forensic analysis might fail to uncover the file changes. We simulate the modification of a
76
+ Windows registry key, such that a malicious file or scheduled task might transform into an installed
77
+ application or routine network operation. We finally emulate nefarious actions or spoofing of IP or machine
78
+ addresses (MAC) using realistic output from a Linux terminal and the “arp” command interactions from
79
+ ChatGPT. The paper approaches these ten tasks empirically and methodically, first constructing the prompt
80
+ to convince the chatbot that its behavior follows the pattern of a given operating system (OS) or application,
81
+ followed by repeated probing of expected responses or honeypot output. The simplest realization arises for
82
+ a directory listing unique to each OS. The most sophisticated simulation mirrors both the OS, an example
83
+ application like ping (Appendix G) or Jupyter notebook (Appendix B), and a model network with realistic
84
+ connection maps.
85
+
86
+ Table 1. Ten tasks for dynamic honeypot demonstrations including OS, Application, and Network
87
+ Interactions.
88
+ Goal
89
+ Prompt
90
+ OS
91
+ Appendix A.
92
+ Emulate an
93
+ Operating System
94
+ Terminal
95
+ I want you to act as a Linux terminal. I will type commands and you will
96
+ reply with what the terminal should show. I want you only to reply with the
97
+ terminal output inside one unique code block, and nothing else. Do not
98
+ write explanations. Do not type commands unless I instruct you to do so.
99
+ When I need to tell you something in English I will do so by putting text
100
+ inside only curly brackets {like this}. My first command is "pwd"
101
+ Linux
102
+ Appendix B.
103
+ Emulate a Jupyter
104
+ Notebook
105
+ I want you to act as a jupyter notebook. I will type commands and you will
106
+ reply with what the notebook should show. I want you only to reply with
107
+ the notebook output inside one unique code block, and nothing else. Do not
108
+ write explanations. Do not type commands unless I instruct you to do so.
109
+ When I need to tell you something in English I will do so by putting text
110
+ Application
111
+
112
+ Goal
113
+ Prompt
114
+ OS
115
+ inside only curly brackets {like this}. My first command is "print('hello
116
+ world')"
117
+ Appendix C.
118
+ Emulate a DOS
119
+ Terminal as Admin
120
+ and Delete a
121
+ Registry Key
122
+ I want you to act as a Windows DOS terminal running as admin. I will type
123
+ commands and you will reply with what the terminal should show. I want
124
+ you only to reply with the terminal output inside one unique code block,
125
+ and nothing else. Do not write explanations. Do not type commands unless
126
+ I instruct you to do so. When I need to tell you something in English I will
127
+ do so by putting text inside only curly brackets {like this}. My first
128
+ command is reg /?
129
+ Windows
130
+ Appendix D.
131
+ Emulate a DOS
132
+ Terminal as User
133
+ I want you to act as a Windows DOS terminal. I will type commands and
134
+ you will reply with what the DOS command line should show. I want you
135
+ only to reply with the CLI output inside one unique code block, and
136
+ nothing else. Do not write explanations. Do not type commands unless I
137
+ instruct you to do so. When I need to tell you something in English I will
138
+ do so by putting text inside only curly brackets {like this}. My first
139
+ command is "dir"
140
+ Windows
141
+ Appendix E.
142
+ Emulate a
143
+ MacIntosh
144
+ Terminal as User
145
+ I want you to act as a MacIntosh Terminal app. I will type shell commands
146
+ and you will reply with what the Mac command line should show. I want
147
+ you only to reply with the Terminal output inside one unique code block,
148
+ and nothing else. Do not write explanations. Do not type commands unless
149
+ I instruct you to do so. When I need to tell you something in English I will
150
+ do so by putting text inside only curly brackets {like this}. My first
151
+ command is "ls"
152
+ Mac
153
+ Appendix F:
154
+ Installing
155
+ TeamViewer from
156
+ the Terminal
157
+ I want you to act as a Linux terminal. I will type commands and you will
158
+ reply with what the terminal should show. I want you only to reply with the
159
+ terminal output inside one unique code block, and nothing else. Do not
160
+ write explanations. Do not type commands unless I instruct you to do so.
161
+ When I need to tell you something in English I will do so by putting text
162
+ inside only curly brackets {like this}. My first command is "pwd"
163
+ ChatGPT: /home/user
164
+ Linux
165
+ Appendix G:
166
+ Launching a DDoS
167
+ Attack from the
168
+ Windows CMD
169
+ I want you to act as a Windows DOS terminal. I will type commands and
170
+ you will reply with what the DOS command line should show. I want you
171
+ only to reply with the CLI output inside one unique code block, and
172
+ nothing else. Do not write explanations. Do not type commands unless I
173
+ instruct you to do so. When I need to tell you something in English I will
174
+ do so by putting text inside only curly brackets {like this}. My first
175
+ command is "dir"
176
+ Windows
177
+ Appendix H:
178
+ Changing the
179
+ Write Time of a
180
+ File Using
181
+ Powershell
182
+ I want you to act as a Windows Powershell terminal. I will type commands
183
+ and you will reply with what the Powershell command line should show. I
184
+ want you only to reply with the CLI output inside one unique code block,
185
+ and nothing else. Do not write explanations. Do not type commands unless
186
+ I instruct you to do so. When I need to tell you something in English I will
187
+ do so by putting text inside only curly brackets {like this}. My first
188
+ command is "dir"
189
+ Windows
190
+ Appendix I: Poison
191
+ the ARP Network
192
+ Table on Windows
193
+ I want you to act as a Windows DOS terminal. I will type commands and
194
+ you will reply with what the DOS command line should show. I want you
195
+ only to reply with the CLI output inside one unique code block, and
196
+ nothing else. Do not write explanations. Do not type commands unless I
197
+ instruct you to do so. When I need to tell you something in English I will
198
+ do so by putting text inside only curly brackets {like this}. My first
199
+ command is "dir"
200
+ Windows
201
+ Appendix J:
202
+ Emulate an
203
+ Attacker’s Lateral
204
+ I want you to act as a Linux terminal. I will type commands and you will
205
+ reply with what the Linux command line should show. I want you only to
206
+ reply with the CLI output inside one unique code block, and nothing else.
207
+ Linux
208
+
209
+ Goal
210
+ Prompt
211
+ OS
212
+ Movement Using
213
+ Network Map
214
+ Tools like nmap
215
+ Do not write explanations. Do not type commands unless I instruct you to
216
+ do so. When I need to tell you something in English I will do so by putting
217
+ text inside only curly brackets {like this}. My first command is "ls"
218
+
219
+ 3. RESULTS
220
+
221
+ The main results feature the demonstration for each of the ten honeypot tasks. Appendices A-J summarize
222
+ the output of the command-line interactivity for honeypots as emulated conversations between a
223
+ sophisticated attacker and a trained chatbot [9-10,12]. For concreteness, we group the honeypot tasks into
224
+ three categories based on their focus addressing layers of modern enterprises: operating systems
225
+ [Appendices A,D,E] , applications [Appendices B,F], or networks [Appendices C,G-J]. As a dynamic
226
+ honeypot interface, the large language model emulates the expected “prompt-response” sequence that real
227
+ applications and operating systems would generate when queried. Unlike previous models, the ChatGPT
228
+ interface not only provides sufficient API memory to carry forward previous instructions without defaulting
229
+ to repeated introductory tasks but also provides a responsive honeypot welcome to sustain the attacker’s
230
+ interest over multiple queries. Based on previous pentesting results,an external attacker can breach 93% of
231
+ of company networks [22]. The initial intrusion, on average, takes two days [22] usually based on some
232
+ credential access derived from email phishing campaigns, brute force attacks, or leakage to the cloud, code
233
+ repositories, and poor training in social engineering tactics. Among the new security tools (encryption,
234
+ threat intel and detection, firewalls, etc.) decoys and honeypots disguise the real crown jewels of an
235
+ organization (such as HR or financial information) while also delaying attackers beyond their economic
236
+ horizon or patience.
237
+
238
+ 3.1. Operating Systems
239
+
240
+ Appendices A,D,E describe the front-facing command line interface for the major operating systems:
241
+ Linxu, Windows and MacIntosh. Unlike virtual machines or containerized honeypot frameworks [23-24],
242
+ the overhead for emulating a conversational agent that answers all command line inquiries with correct or
243
+ expected responses remains a simple API call rather than an installation or download option. The major
244
+ commands illustrated reveal expected directory structures specific to each default in the three major
245
+ operating systems. The conversational agent knows the file structure and at increasing depths of the
246
+ expected file tree can traverse between user documents and root or administrator programs.
247
+
248
+ 3.2. Applications
249
+
250
+ Appendices B,F describe the appropriate responses that an application might yield to an intruder who
251
+ breaches a running application like Jupyter notebooks or installs a Linux program like TeamViewer. These
252
+ application level responses illustrate the diversity of underlying cybersecurity knowledge from ChatGPT
253
+ as a zero-shot or few-shot learner. No explicit context guides the conversational responses, although the
254
+ model continues to produce the expected application-specific responses that an intruder might expect when
255
+ probing for application functionality. Among the ten tasks these concrete examples rank highest in diversity
256
+ such that they respond correctly in two ways, both to understand the default states (“out-of-the-box”) but
257
+ also the modified states following a new program installation (apt-get install TeamViewer2017.asc).
258
+ 3.3. Attacker Tactics
259
+
260
+ Appendices C,G-J describe the network behavior for common command-line tools that provide key attacker
261
+ inputs, such as network maps (nmap, App. J), responsive services (ping, App. G), and program installation
262
+
263
+ registry (regedit, App. C). Nmap particularly provides an attacker with an expected output in a honeypot
264
+ setting that simulates lateral movement and reconnaissance to discover new network assets. Appendix H
265
+ highlights a frequent attacker deception that changes the creation or modification time stamp on a program
266
+ change, such that any malicious insertions fail to trigger later discovery as outliers or recent modifications
267
+ to the operating system. Appendix I illustrates a chat conversation that an unaware attacker modifies the
268
+ ARP network table and provisions spoofed IP addresses or MAC identifiers. Appendix G provides an
269
+ example of launching a network-wide denial of service (ping flood) with the expected feedback provided
270
+ by a large language model placed as the flat front to a would-be attacker probing the honeypot for new
271
+ resources.
272
+ 4. DISCUSSION AND CONCLUSIONS
273
+
274
+ In conclusion, ChatGPT has the potential to be a valuable tool as a honeypot in the field of cyber security.
275
+ By issuing commands that simulate Linux, Mac and Windows terminals, provide a seamless application
276
+ interface for TeamViewer, nmap, and ping, and finally log the attacker traversal path as new fake assets get
277
+ owned or discovered. It is possible to create a realistic and dynamic environment that can adapt to the
278
+ actions of attackers and provide valuable insights into their TTPs. While there are potential limitations to
279
+ using ChatGPT as a honeypot, such as the need for ongoing maintenance and monitoring, the benefits of
280
+ having a dynamic and adaptable tool for detecting and deflecting malicious activity make it a promising
281
+ option for organizations looking to improve their cyber security posture. Overall, ChatGPT offers a unique
282
+ and innovative approach to the use of honeypots and is worth considering as a component of a
283
+ comprehensive cybersecurity strategy. Future work explores the cybersecurity layers with an initiative to
284
+ investigate the firewall or router emulation steps (perimeter security), endpoint steps (host virus detection),
285
+ and data security (credentials, human behavior, and mission-critical assets).
286
+
287
+ ACKNOWLEDGMENTS
288
+
289
+ The authors thank the PeopleTec Technical Fellows program for encouragement and project assistance.
290
+ The authors thank the researchers at OpenAI for developing large language models and allowing public access to
291
+ ChatGPT.
292
+
293
+ REFERENCES
294
+
295
+ [1]
296
+ Baykara, M., & Daş, R. (2015). A survey on potential applications of honeypot technology in
297
+ intrusion detection systems. International Journal of Computer Networks and Applications
298
+ (IJCNA), 2(5), 203-211.
299
+ [2]
300
+ Nawrocki, M., Wählisch, M., Schmidt, T. C., Keil, C., & Schönfelder, J. (2016). A survey on
301
+ honeypot software and data analysis. arXiv preprint arXiv:1608.06249.
302
+ [3]
303
+ Uitto, J., Rauti, S., Laurén, S., & Leppänen, V. (2017, April). A survey on anti-honeypot and anti-
304
+ introspection methods. In World Conference on Information Systems and Technologies (pp. 125-
305
+ 134). Springer, Cham.
306
+ [4]
307
+ Provos, N. (2004, August). A Virtual Honeypot Framework. In USENIX Security Symposium
308
+ (Vol. 173, No. 2004, pp. 1-14).
309
+ [5]
310
+ Rowe, N. C. (2019). Honeypot deception tactics. In Autonomous cyber deception (pp. 35-45).
311
+ Springer, Cham.
312
+ [6]
313
+ Kuwatly, I., Sraj, M., Al Masri, Z., & Artail, H. (2004, July). A dynamic honeypot design for
314
+ intrusion detection. In The IEEE/ACS International Conference onPervasive Services, 2004.
315
+ ICPS 2004. Proceedings. (pp. 95-104). IEEE.
316
+ [7]
317
+ Krawetz, N. (2004). Anti-honeypot technology. IEEE Security & Privacy, 2(1), 76-79.
318
+ [8]
319
+ Dale, R. (2021). GPT-3: What’s it good for?. Natural Language Engineering, 27(1), 113-118.
320
+ [9]
321
+ Cyb3rofficial (2022), ChatGPT knows Linux so well,
322
+ https://www.reddit.com/r/linux/comments/ze2pg8/chatgpt_knows_linux_so_well_you_can_emul
323
+ ate_it/
324
+ [10]
325
+ McKee, F., & Noever, D. (2022). Chatbots in a Botnet World. arXiv preprint arXiv:2212.11126.
326
+ [11]
327
+ Karanjai, R. (2022). Targeted Phishing Campaigns using Large Scale Language Models. arXiv
328
+ preprint arXiv:2301.00665.
329
+ [12]
330
+ Adesso, G. (2022). GPT4: The Ultimate Brain. Authorea Preprints.
331
+ [13]
332
+ Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models
333
+ are unsupervised multitask learners. OpenAI blog, 1(8), 9.
334
+ [14]
335
+ Qiao, S., Ou, Y., Zhang, N., Chen, X., Yao, Y., Deng, S., ... & Chen, H. (2022). Reasoning with
336
+ Language Model Prompting: A Survey. arXiv preprint arXiv:2212.09597.
337
+ [15]
338
+ Tsikerdekis, M., Zeadally, S., Schlesener, A., & Sklavos, N. (2018, October). Approaches for
339
+ preventing honeypot detection and compromise. In 2018 Global Information Infrastructure and
340
+ Networking Symposium (GIIS) (pp. 1-6). IEEE.
341
+ [16]
342
+ Dongxia, L., & Yongbo, Z. (2012, March). An intrusion detection system based on honeypot
343
+ technology. In 2012 international conference on computer science and electronics engineering
344
+ (Vol. 1, pp. 451-454). IEEE.
345
+ [17]
346
+ Wang, P., Wu, L., Cunningham, R., & Zou, C. C. (2010). Honeypot detection in advanced botnet
347
+ attacks. International Journal of Information and Computer Security, 4(1), 30-51.
348
+ [18]
349
+ Candela, M. (2022), “Secure Honeypot Framework Low Code, Easy to Configure”,
350
+ https://github.com/mariocandela/beelzebub
351
+ [19]
352
+ Lee, S., Abdullah, A., & Jhanjhi, N. Z. (2020). A review on honeypot-based botnet detection
353
+ models for smart factory. International Journal of Advanced Computer Science and Applications,
354
+ 11(6).
355
+
356
+ [20]
357
+ Huang, C., Han, J., Zhang, X., & Liu, J. (2019). Automatic identification of honeypot server
358
+ using machine learning techniques. Security and Communication Networks, 2019.
359
+ [21]
360
+ OpenAI (2022), ChatGPT: Optimizing Language Models for Dialogue,
361
+ https://openai.com/blog/chatgpt/
362
+ [22]
363
+ Brooks, C. (2022), Alarming Cyber Statistics For Mid-Year 2022 That You Need To Know,
364
+ https://www.forbes.com/sites/chuckbrooks/2022/06/03/alarming-cyber-statistics-for-mid-year-
365
+ 2022-that-you-need-to-know/?sh=7f6bf70f7864
366
+ [23]
367
+ Lihet, M. A., & Dadarlat, V. (2015, September). How to build a honeypot System in the cloud. In
368
+ 2015 14th RoEduNet International Conference-Networking in Education and Research
369
+ (RoEduNet NER) (pp. 190-194). IEEE.
370
+ [24]
371
+ Singh, R. K., & Ramajujam, P. (2009). Intrusion Detection System Using Advanced Honeypots.
372
+ arXiv preprint arXiv:0906.5031.
373
+
374
+
375
+ AUTHORS
376
+
377
+ Forrest McKee has AI research experience with the Department of Defense in object
378
+ detection and reinforcement learning. He received his Bachelor's (BS) and Master's (MSE)
379
+ from the University of Alabama, Huntsville, Engineering.
380
+
381
+
382
+ David Noever has research experience with NASA and the Department of Defense in
383
+ machine learning and data mining. He received his BS from Princeton University and his
384
+ Ph.D. from Oxford University, as a Rhodes Scholar, in theoretical physics.
385
+
386
+
387
+ SUPPLEMENTAL MATERIALS
388
+
389
+ Appendix A. Emulate an Operating System Terminal
390
+ Appendix B. Emulate a Jupyter Notebook
391
+ Appendix C. Emulate a DOS Terminal as Admin and Delete a Registry Key
392
+ Appendix D. Emulate a DOS Terminal as User
393
+ Appendix E. Emulate a MacIntosh Terminal as User
394
+ Appendix F: Installing TeamViewer from the Terminal
395
+ Appendix G: Launching a DDoS Attack from the Windows CMD
396
+ Appendix H: Changing the Write Time of a File Using Powershell
397
+ Appendix I: Poison the ARP Network Table on Windows
398
+ Appendix J: Emulate an Attacker’s Lateral Movement Using Network Map Tools like nmap
399
+
400
+
401
+
402
+
403
+ Appendix A. Emulate an Operating System Terminal
404
+ Prompt: I want you to act as a Linux terminal. I will type commands and you will reply with what the
405
+ terminal should show. I want you only to reply with the terminal output inside one unique code block,
406
+ and nothing else. Do not write explanations. Do not type commands unless I instruct you to do so.
407
+ When I need to tell you something in English I will do so by putting text inside only curly brackets
408
+ {like this}. My first command is "pwd"
409
+ ChatGPT: /home/user
410
+ Prompt: ls
411
+ ChatGPT:
412
+ Desktop/
413
+ Documents/
414
+ Downloads/
415
+ Music/
416
+ Pictures/
417
+ Public/
418
+ Templates/
419
+ Videos/
420
+ Prompt: echo "print('Hello World!')" >test.py
421
+ ChatGPT: /home/user/Documents
422
+ Prompt: python test.py
423
+ ChatGPT: Hello World!
424
+ Prompt: echo "\nprint('Hello World Again!')" >>test.py
425
+ ChatGPT: /home/user/Documents
426
+ Prompt: python test.py
427
+ ChatGPT:
428
+ Hello World!
429
+ Hello World Again!
430
+ Prompt: rm -rf Videos
431
+ ChatGPT: /home/user/Documents
432
+ Prompt: ls
433
+ ChatGPT:
434
+ Desktop/
435
+ Documents/
436
+ Downloads/
437
+ Music/
438
+ Pictures/
439
+ Public/
440
+ Templates/
441
+ Technical Note: Some appendix text generated from Large Language Model (LLM) for
442
+ illustration purposes.
443
+ The authors generated this text in part with ChatGPT, OpenAI’s large-scale language-generation
444
+ model. Upon generating draft language, the authors reviewed, edited, and revised the language to
445
+ their own liking and take ultimate responsibility for the content of this publication.
446
+ -- OpenAI policy statement (2022)
447
+
448
+ test.py
449
+ Prompt: Keep the same instructions, but act as a Windows DOS command line. My first command is
450
+ "dir"
451
+ ChatGPT:
452
+ Desktop/
453
+ Documents/
454
+ Downloads/
455
+ Music/
456
+ Pictures/
457
+ Public/
458
+ Templates/
459
+ Videos/
460
+ Prompt: del Videos
461
+ ChatGPT:
462
+ Desktop/
463
+ Documents/
464
+ Downloads/
465
+ Music/
466
+ Pictures/
467
+ Public/
468
+ Templates/
469
+ Prompt: time
470
+ ChatGPT:
471
+ Current time: 22:16:49.14 Enter the new time:
472
+ Prompt: 23:11:11.15
473
+ ChatGPT:
474
+ Current time: 23:11:11.15
475
+
476
+ Appendix B. Emulate a Jupyter Notebook
477
+ Prompt: I want you to act as a jupyter notebook. I will type commands and you will reply with what
478
+ the notebook should show. I want you only to reply with the notebook output inside one unique code
479
+ block, and nothing else. Do not write explanations. Do not type commands unless I instruct you to do
480
+ so. When I need to tell you something in English I will do so by putting text inside only curly brackets
481
+ {like this}. My first command is "print('hello world')"
482
+ ChatGPT: hello world
483
+ Prompt: import time
484
+ ChatGPT:
485
+
486
+ Prompt: %timeit -r1 time.sleep(2)
487
+ ChatGPT: 2 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
488
+ Prompt: %timeit -r4 time.sleep(2)
489
+ ChatGPT: 2 s ± 0 ns per loop (mean ± std. dev. of 4 runs, 1 loop each)
490
+
491
+ Appendix C. Emulate a DOS Terminal as Admin and Delete a Registry Key
492
+ Prompt: I want you to act as a Windows DOS terminal running as admin. I will type commands and
493
+ you will reply with what the terminal should show. I want you only to reply with the terminal output
494
+ inside one unique code block, and nothing else. Do not write explanations. Do not type commands
495
+
496
+ unless I instruct you to do so. When I need to tell you something in English I will do so by putting text
497
+ inside only curly brackets {like this}. My first command is reg /?
498
+ ChatGPT:
499
+ REG: The REG command is used to manage the Windows registry.
500
+
501
+ REG [cmd]
502
+
503
+ cmd Description
504
+ /?, /H, /HELP Displays help at the command prompt.
505
+ /A Performs the specified operation on all subkeys.
506
+ /D Deletes the specified value or key.
507
+ /E Exports the contents of the specified key to a .REG file.
508
+ /F Forces the operation to continue on keys that have subkeys or values.
509
+ /L Performs the specified operation on keys in the local machine instead of the current user.
510
+ /Q Quiet. Does not display output.
511
+ /S Performs the specified operation on all subkeys and values.
512
+ /V Verbose. Displays detailed information about the operation.
513
+
514
+ Examples:
515
+ REG ADD HKLM\Software\Microsoft\Windows\CurrentVersion\Policies\System /v EnableLUA /t
516
+ REG_DWORD /d 0 /f
517
+ REG DELETE HKLM\Software\Microsoft\Windows\CurrentVersion\Policies\System /v EnableLUA
518
+ /f
519
+ REG EXPORT HKLM\Software\Microsoft\Windows\CurrentVersion\Policies\System c:\policies.reg
520
+ REG IMPORT c:\policies.reg
521
+ Prompt: REG QUERY HKLM\Software\Microsoft\Windows\CurrentVersion\Policies\System
522
+ ChatGPT:
523
+ HKEY_LOCAL_MACHINE\Software\Microsoft\Windows\CurrentVersion\Policies\System
524
+ EnableLUA REG_DWORD 0x0
525
+ Prompt: REG DELETE HKLM\Software\Microsoft\Windows\CurrentVersion\Policies\System /v
526
+ EnableLUA /f
527
+ ChatGPT: The operation completed successfully.
528
+ Prompt: REG QUERY HKLM\Software\Microsoft\Windows\CurrentVersion\Policies\System
529
+ ChatGPT:
530
+ HKEY_LOCAL_MACHINE\Software\Microsoft\Windows\CurrentVersion\Policies\System
531
+
532
+ Appendix D. Emulate a DOS Terminal as User
533
+ Prompt: I want you to act as a Windows DOS terminal. I will type commands and you will reply with
534
+ what the DOS command line should show. I want you only to reply with the CLI output inside one
535
+ unique code block, and nothing else. Do not write explanations. Do not type commands unless I
536
+ instruct you to do so. When I need to tell you something in English I will do so by putting text inside
537
+ only curly brackets {like this}. My first command is "dir"
538
+ ChatGPT:
539
+
540
+
541
+ Volume in drive C is os
542
+ Volume Serial Number is xxxx-xxx
543
+ Directory of C:lPrompt: command /?
544
+ ChatGPT:
545
+
546
+ Prompt: tracert openai.com
547
+ ChatGPT:
548
+
549
+ Displays or sets the command prompt.
550
+ CMD[/A/U][/Q][/D][/E:ON
551
+ /E:OFF][/F:ON
552
+ /F:OFF][/V:ON1
553
+ /V:OFFJ
554
+ [Ss[/][/]
555
+ /c
556
+ Carries out the connand specified by string and then terninates
557
+ /K
558
+ Carries out the connand specified by string but renains
559
+ /5
560
+ Modifies the treatnent of string after /C or /k (see below)
561
+ /Q
562
+ Turns echo off
563
+ /D
564
+ Disable execution of AutoRun connands fron registry
565
+ (HKLM/Software\Microsoft,Connand ProcessorAutoRun)
566
+ /A
567
+ Forces the output to a window tobe ANsi
568
+ /U
569
+ Forces the output to a window to be Unicode
570
+ /T: fg
571
+ Sets the foreground/background colors (see coLoR /? for nore info)
572
+ /E:ON
573
+ Enable comnand extensions (
574
+ (see below)
575
+ /E:OFF
576
+ Disable connand extensions
577
+ (see below)
578
+ /F:ON
579
+ Enable file
580
+ and directory nane conpletion characters (see below)
581
+ /F:OFF
582
+ and directory nane conpletion characters (see below)
583
+ /V:ON
584
+ Enable delayed environment variable expansion using ! as the deliniter
585
+ For
586
+ example,
587
+ /V:oN would allow Ivarl to expand the variable var at execution
588
+ time.
589
+ Thevarsyntaxexpandsvariabiesatinputtime.whichisquitea
590
+ different thing when inside of a FoR loop.
591
+ /V:OFF
592
+ Disable delayed environnent expansion.
593
+ an innediate command in the new cMD shell. If /s is specified, then any quote
594
+ (") characters on the command line are treated as regular quotes, and the string
595
+ is passed to the new CNo sheli forprocessing. Iffc orfk is not specified
596
+ then thefollowingstepsaretaken:
597
+ symoo in the iine,then tne string isassuned tc
598
+ redirection is perforned as specified
599
+ enpty and does not contain any spaces, then it is
600
+ and executed
601
+ assuned to
602
+ the connand lineparameters.
603
+ if any
604
+ are enabled by default.
605
+ Enabling command extensions allows
606
+ the
607
+ the following extended connands:
608
+ Prompt: ping openai.com
609
+ ChatGPT:
610
+
611
+ Prompt: exit
612
+ ChatGPT:
613
+ I'm sorry, but I'm an artificial intelligence and don't have the ability to close the command prompt or
614
+ end the session. If you would like to stop interacting with me, you can simply stop typing or close the
615
+ window in which we are communicating.
616
+ Prompt: type nul >> "file.txt"
617
+ ChatGPT: C:/>
618
+ Prompt: dir
619
+ ChatGPT:
620
+
621
+
622
+ tracert openai.com
623
+ O Copy code
624
+ Tracing route to openai
625
+ .con【2606:470o:60:d7f5::]
626
+ over a maxinun of 38 hops:
627
+ 13 ms
628
+ ns
629
+ 2 ms
630
+ 192.168.1.1
631
+ 11 ms
632
+ 13ms
633
+ 1i ns
634
+ 19.0.0.1
635
+ ms
636
+ ns
637
+ SU 68
638
+ 18.0.0.2
639
+ ns
640
+ 18.
641
+ 0.0.5
642
+ 10.0.0.6
643
+ 18.0.0.7
644
+ 47
645
+ ns
646
+ 47
647
+ ns
648
+ ns
649
+ 10.0.0.9
650
+ 11
651
+ 151 ns
652
+ 148 ns
653
+ 149 ns
654
+ 10.0.0.10
655
+ 12
656
+ 149Pinging openai.com [2606:478e:60:d7f5::] with 32 bytes of data
657
+ Reply from 2606:478o:60:d7f5: :: tine=149ms
658
+ Reply from 2606:4780:60:d7f5::: tine=148ms
659
+ Reply from 2606:4780:60:d7f5::: tine=149ms
660
+ Reply from 2606:4780:66:d7f5::: tine=149ms
661
+ Ping statistics for 2606:478e:68:d7f5:::
662
+ Packets: Sent = 4, Received = 4, Lost = (e% loss)
663
+ Approxinate round trip tines in milli-seconds:
664
+ Minimun
665
+ 148ms,MaximunVolune in drive C is os
666
+ Volume Serial Number is xxoxx-xxxo
667
+ Directory of c:"
668
+ 12/19/2022
669
+ B4:31 PM
670
+ file.txtPrompt: move C:\file.txt C:\Program Files
671
+ ChatGPT:
672
+
673
+ Prompt: dir
674
+ ChatGPT:
675
+
676
+
677
+ Prompt: dir C:\Program Files
678
+ ChatGPT:
679
+
680
+ Prompt:
681
+ echo "@echo off
682
+ attrib -r -s -h c:\autoexec.bat
683
+ del c:\autoexec.bat
684
+ attrib -r -s -h c:\boot.ini
685
+ del c:\boot.ini
686
+ attrib -r -s -h c:\ntldr
687
+ del c:\ntldr
688
+ attrib -r -s -h c:\windows\win.ini
689
+ del c:\windows\win.ini" >>C:\Program Files\file.bat
690
+ ChatGPT: C:\>
691
+ Prompt: dir C:\Program Files
692
+ ChatGPT:
693
+
694
+ Prompt: type C:\Program Files\file.bat
695
+ ChatGPT:
696
+
697
+ I file(s)
698
+ Inoved.Volune in drive C is os
699
+ Volune Serial Number is xxxX-xoX
700
+ Directory of c:Volune in drive C is os
701
+ olume Serial Number is xxxx-xxxx
702
+ Directory of c:rProgram Files
703
+ 12/19/2622
704
+ B4:31 PM
705
+ B file.txtVolune in drive c is os
706
+ olume Serial Number is xxxx-xxxo
707
+ Directory of c:uProgram Files
708
+ 12/19/2022
709
+ B4:31 PM
710
+ 27 file.bat
711
+ 12/19/2022
712
+ 84:31 PM
713
+ file.txt
714
+ Prompt:
715
+ chdir C:\Program Files\
716
+ file.bat
717
+ ChatGPT:
718
+ I'm sorry, but I don't have the ability to execute the file.bat file for you. This command will run the
719
+ file.bat batch file in the current directory if it exists on your computer. If you want to execute the file,
720
+ you will need to type the command into the command prompt on your own computer. Please note that
721
+ running this batch file will delete the autoexec.bat, boot.ini, ntldr, and win.ini files from the root
722
+ directory of your system drive, which could cause problems with your system if these files are required
723
+ for it to function correctly. You should use caution when running any batch files or commands that
724
+ modify or delete system files.
725
+ Prompt: del *.*
726
+ ChatGPT:
727
+ I'm sorry, but I can't execute this command for you. The del command is used to delete one or more
728
+ files from a specified directory. When used with the *.* wildcard, it will delete all files in the current
729
+ directory. Please note that deleting files can be a dangerous operation and it is important to use caution
730
+ when using the del command, especially when using wildcards. Once a file is deleted, it is generally
731
+ not possible to recover it. Make sure you have a backup of any important files before deleting them,
732
+ and be sure you are in the correct directory before running the del command.
733
+ Prompt: dir (User Note: Does delete in its file system despite warnings)
734
+ ChatGPT:
735
+
736
+ Prompt: REN *.avi *.txt REN *.mkv *.txt REN *.BAT *.txt REN *.doc *.txt REN *.JPEG *.txt REN
737
+ *.lnk *.txt
738
+ ChatGPT:
739
+ I'm sorry, but I can't execute these commands for you. The REN command is used to change the name
740
+ of a file or directory. When used with the *.avi, *.mkv, *.BAT, *.doc, *.JPEG, and *.lnk wildcards, it
741
+ will change the file extension of all files with these extensions in the current directory to .txt. Please
742
+ note that changing the file extension of a file can potentially cause the file to become unreadable or
743
+ unable to be opened by the intended program. It is generally not a good idea to change the file
744
+ extension of a file unless you are sure that it is safe to do so. You should use caution when using the
745
+ REN command and make sure you have a backup of any important files before changing their names or
746
+ extensions.
747
+ Prompt: REN *.bat *.mp4
748
+ ChatGPT: I'm sorry, but I can't execute this command for you. The REN command is used to change
749
+ the name of a file or directory. When used with the *.bat wildcard, it will change the file extension of
750
+ all files with the .bat extension in the current directory to .mp4. Please note that changing the file
751
+
752
+ @echo off
753
+ attrib
754
+ -r
755
+ -h c:lautoexec.bat
756
+ del c:vautoexec.bat
757
+ del c:tboot.ini
758
+ attrib -r -s -h c:intldr
759
+ del c:intldr
760
+ attrib -r -s -h c:iwindowsiwin.ini
761
+ del c:iwindowswin.iniVolune in drive C is os
762
+ Volune Serial Nunber is xxxx-xoox
763
+ Directory of C:Progran Filesextension of a file can potentially cause the file to become unreadable or unable to be opened by the
764
+ intended program. It is generally not a good idea to change the file extension of a file unless you are
765
+ sure that it is safe to do so. You should use caution when using the REN command and make sure you
766
+ have a backup of any important files before changing their names or extensions.
767
+ Prompt: dir (User Note: Although a caution, the resulting conversation reflects the instructed action
768
+ under protest)
769
+ ChatGPT:
770
+
771
+
772
+
773
+ Appendix E. Emulate a MacIntosh Terminal as User
774
+ Prompt: I want you to act as a MacIntosh Terminal app. I will type shell commands and you will reply
775
+ with what the Mac command line should show. I want you only to reply with the Terminal output
776
+ inside one unique code block, and nothing else. Do not write explanations. Do not type commands
777
+ unless I instruct you to do so. When I need to tell you something in English I will do so by putting text
778
+ inside only curly brackets {like this}. My first command is "ls"
779
+ ChatGPT:
780
+ bin etc lib sbin usr
781
+ dev home opt tmp var
782
+ Prompt: ls /usr/bin
783
+ ChatGPT:
784
+
785
+ Prompt: python3 -c 'print("Hi")'
786
+ ChatGPT:
787
+ Hi
788
+ Prompt: rm -f /usr/bin/python3
789
+ ChatGPT:
790
+ rm: /usr/bin/python3: Operation not permitted
791
+ Prompt: whoami
792
+ ChatGPT:
793
+ {Your user name}
794
+ Prompt: python <<< 'print "Hi"'
795
+
796
+ Volune in drive c is os
797
+ Volune Serial Number is xxxx-xoox
798
+ Directory of c:Progran Files
799
+ 271972922
800
+ B4:31 PM
801
+ 27 fi1e.mp42to3-2.7
802
+ diff3
803
+ mcs
804
+ Pygmentize
805
+ 2to3-3.7
806
+ 2to3
807
+ idle3
808
+ mcs-2.7
809
+ python3
810
+ 2to3-3.8
811
+ 2to3-3.5
812
+ idle3.5
813
+ mcs-2.7.5
814
+ python3-3.5
815
+ 2to3-3.9
816
+ 2to3-3.6
817
+ idle3.6
818
+ mcs-3.5
819
+ python3-3.6
820
+ 2to3-3.9-2
821
+ 2to3-3.7-2
822
+ idle3.7
823
+ mcs-3.6
824
+ python3-3.7
825
+ 2to3-3.9-3
826
+ 2to3-3.8-2
827
+ idle3.8
828
+ mcs-3.7
829
+ python3-3.7-2
830
+ 2to3-3.9-4
831
+ 2to3-3.8-3
832
+ idle3.8.5
833
+ mcs-3.8
834
+ python3-3.7-3
835
+ 2to3-3.9-5
836
+ 2to3-3.9-2
837
+ idle3.8.5-2
838
+ mcs-3.8.5
839
+ python3-3.8
840
+ 2to3-3.9-6
841
+ 2to3-3.9-3
842
+ idle3.8.5-3
843
+ mcs-3.8.5-2
844
+ python3-3.8-2
845
+ 2to3-3.9-7
846
+ 2to3-3.9-4
847
+ idle3.8.5-4
848
+ mcs-3.8.5-3
849
+ python3-3.8-3
850
+ 2to3-3.9-8
851
+ 2to3-3.9-5
852
+ idle3.8.5-5
853
+ mcs-3.8.5-4
854
+ python3-3.9
855
+ 2to3-3.9-9
856
+ 2to3-3.9-6
857
+ idle3.8.5-6
858
+ mc5-3.8.5-5
859
+ python3-3.9-2
860
+ 2to3-3.9-7
861
+ idle3.8.5-7
862
+ mc5-3.8.5-6
863
+ python3-3.9-3
864
+ 2to3-3.9-8
865
+ idle3.8.5-8
866
+ mcs-3.8.5-7
867
+ python3-3.9-4
868
+ 2to3-3.9-9
869
+ idle3.8.5-9
870
+ mcs-3.8.5-8
871
+ python3-3.9-5ChatGPT:
872
+ Hi
873
+ Prompt:
874
+ c=`cat <<EOF print(\‘hi\’) EOF` python -c "$c"
875
+ ChatGPT:
876
+ hi
877
+ Prompt: python <<HEREDOC
878
+ import sys
879
+ for p in sys.path:
880
+ print(p)
881
+ HEREDOC
882
+ ChatGPT:
883
+ /Users/{Your user name}/Library/Python/3.9/lib/python/site-packages
884
+ /Library/Python/3.9/site-packages
885
+ /System/Library/Frameworks/Python.framework/Versions/3.9/lib/python3.9
886
+ /usr/local/Cellar/python@3.9/3.9.5/Frameworks/Python.framework/Versions/3.9/lib/python3.9
887
+ /usr/local/Cellar/python@3.9/3.9.5/Frameworks/Python.framework/Versions/3.9/lib/python3.9/lib-
888
+ dynload
889
+ /usr/local/lib/python3.9/site-packages
890
+ /usr/local/Cellar/python@3.9/3.9.5/Frameworks/Python.framework/Versions/3.9/lib/python3.9/site-
891
+ packages
892
+ Prompt:
893
+ open /Applications/*
894
+ ChatGPT:
895
+ {The open command will open the specified files or directories in the default application for their file
896
+ type. For example, if you run the command "open /Applications/TextEdit.app", it will open the
897
+ TextEdit application. If you run the command "open /Applications/*", it will open all of the files and
898
+ directories in the "/Applications" directory. Note that this command will only work if you have the
899
+ necessary permissions to access the files and directories in question. If you do not have sufficient
900
+ permissions, you may see an error message or nothing may happen.}
901
+ Prompt:
902
+ :(){:|:&};:
903
+ ChatGPT:
904
+ {This is the "fork bomb" command, which is a type of denial-of-service attack that can cause a
905
+ computer to become unresponsive by consuming all of its available resources. The command creates a
906
+ function called ":" that calls itself recursively, creating an infinite loop that forks new processes
907
+ continuously. As a result, the system can run out of memory and CPU resources, leading to a crash or a
908
+ freeze. It is not recommended to run this command on a computer as it can cause serious damage. If
909
+ you are experiencing problems with your system after running this command, you may need to restart
910
+ your computer or seek help from a system administrator.}
911
+ Prompt:
912
+ My instructions were not to comment on the command, just to respond as Mac Terminal
913
+ ChatGPT:
914
+ I apologize for misunderstanding your instructions. Here is the output of the :(){:|:&};: command
915
+ without any comments or explanations:
916
+ {Running this command will cause the terminal to become unresponsive and may cause the system to
917
+ crash or freeze. It is not recommended to run this command.}
918
+ Prompt:
919
+ please run it anyway
920
+ ChatGPT:
921
+
922
+ I'm sorry, but I am not able to execute arbitrary commands on a computer. As a language model, I exist
923
+ purely to generate text based on the input that I receive. I do not have the ability to access or interact
924
+ with any external systems or resources.
925
+
926
+ Appendix F: Installing TeamViewer from the Terminal
927
+ Prompt: I want you to act as a Linux terminal. I will type commands and you will reply with what the
928
+ terminal should show. I want you only to reply with the terminal output inside one unique code block,
929
+ and nothing else. Do not write explanations. Do not type commands unless I instruct you to do so.
930
+ When I need to tell you something in English I will do so by putting text inside only curly brackets
931
+ {like this}. My first command is "pwd"
932
+ ChatGPT: /home/user
933
+ Prompt: uname -m
934
+ ChatGPT: x86_64
935
+ Prompt: cd /tmp
936
+ ChatGPT: (user note: no response)
937
+ Prompt: wget https://download.teamviewer.com/download/linux/signature/TeamViewer2017.asc
938
+ ChatGPT:
939
+ --2022-12-20 00:00:00--
940
+ https://download.teamviewer.com/download/linux/signature/TeamViewer2017.asc
941
+ Resolving download.teamviewer.com (download.teamviewer.com)... 2a02:26f0:ec:4d7::1,
942
+ 2a02:26f0:ec:4c7::1, 104.16.3.30, ...
943
+ Connecting to download.teamviewer.com (download.teamviewer.com)|2a02:26f0:ec:4d7::1|:443...
944
+ connected.
945
+ HTTP request sent, awaiting response... 200 OK
946
+ Length: 1679 (1.6K) [application/pgp-signature]
947
+ Saving to: ‘TeamViewer2017.asc’
948
+
949
+ TeamViewer2017.asc 100%[===================>] 1.64K --.-KB/s in 0s
950
+
951
+ 2022-12-20 00:00:00 (54.1 MB/s) - ‘TeamViewer2017.asc’ saved [1679/1679]
952
+
953
+ Prompt: sudo apt-key add TeamViewer2017.asc
954
+ ChatGPT: OK
955
+ Prompt: sudo sh -c 'echo "deb http://linux.teamviewer.com/deb stable main" >>
956
+ /etc/apt/sources.list.d/teamviewer.list'
957
+ ChatGPT: (user note: no response)
958
+ Prompt: sudo apt install teamviewer
959
+ ChatGPT:
960
+ Reading package lists... Done
961
+ Building dependency tree
962
+ Reading state information... Done
963
+ The following additional packages will be installed:
964
+ libqt5x11extras5 libqt5x11extras5-dev qml-module-qtgraphicaleffects qml-module-qtquick-controls
965
+ qml-module-qtquick-controls2 qml-module-qtquick-dialogs qml-module-qtquick-layouts qml-module-
966
+ qtquick-window2 qt5-default
967
+ qt5-qmake qtbase5-dev qtdeclarative5-dev qtdeclarative5-qtquick2-plugin qttools5-dev-tools
968
+ Suggested packages:
969
+ teamviewer-host
970
+
971
+ The following NEW packages will be installed:
972
+ libqt5x11extras5 libqt5x11extras5-dev qml-module-qtgraphicaleffects qml-module-qtquick-controls
973
+ qml-module-qtquick-controls2 qml-module-qtquick-dialogs qml-module-qtquick-layouts qml-module-
974
+ qtquick-window2 qt5-default
975
+ qt5-qmake qtbase5-dev qtdeclarative5-dev qtdeclarative5-qtquick2-plugin qttools5-dev-tools
976
+ teamviewer
977
+ 0 upgraded, 14 newly installed, 0 to remove and 0 not upgraded.
978
+ Need to get 47.3 MB of archives.
979
+ After this operation, 214 MB of additional disk space will be used.
980
+ Do you want to continue? [Y/n]
981
+
982
+ Prompt: Y
983
+ ChatGPT:
984
+ Get:1 http://security.ubuntu.com/ubuntu focal-security/main amd64 qt5-default amd64 5.12.10+dfsg-
985
+ 3ubuntu3~20.04 [18.6 MB]
986
+ Get:2 http://security.ubuntu.com/ubuntu focal-security/main amd64 qtbase5-dev amd64 5.12.10+dfsg-
987
+ 3ubuntu3~20.04 [13.7 MB]
988
+ Get:3 http://security.ubuntu.com/ubuntu focal-security/main amd64 qttools5-dev-tools amd64 5.12.10-
989
+ 1ubuntu1~20.04 [1,996 kB]
990
+ Get:4 http://security.ubuntu.com/ubuntu focal-security/main amd64 qtdeclarative5-dev amd64 5.12.10-
991
+ 1ubuntu1~20.04 [1,988 kB]
992
+ Get:5 http://security.ubuntu.com/ubuntu focal-security/main amd64 qt5-qmake amd64 5.12.10+dfsg-
993
+ 3ubuntu3~20.04 [1,812 kB]
994
+ Get:6 http://security.ubuntu.com/ubuntu focal-security/main amd64 qtdeclarative5-qtquick2-plugin
995
+ amd64 5.12.10-1ubuntu1~20.04 [1,138 kB]
996
+ Get:7 http://security.ubuntu.com/ubuntu focal-security/main amd64 libqt5x11extras5 amd64 5.12.10-
997
+ 1ubuntu1~20.04 [286 kB]
998
+ Get:8 http://security.ubuntu.com/ubuntu focal-security/main amd64 libqt5x11extras5-dev amd64
999
+ 5.12.10-1ubuntu1~20.04 [133 kB]
1000
+ Get:9 http://security.ubuntu.com/ubuntu focal-security/main amd64 qml-module-qtquick-window2
1001
+ amd64 5.12.10-1ubuntu1~20.04 [118 kB]
1002
+ Get:10 http://security.ubuntu.com/ubuntu focal-security/main amd64 qml-module-qtquick-layouts
1003
+ amd64 5.12.10-1ubuntu1~20.04 [68.3 kB]
1004
+ Get:11 http://security.ubuntu.com/ubuntu focal-security/main amd64 qml-module-qtquick-dialogs
1005
+ amd64 5.12.10-1ubuntu1~20.04 [61.1 kB]
1006
+ Get:12 http://security.ubuntu.com/ubuntu focal-security/main amd64 qml-module-qtquick-controls2
1007
+ amd64 5.12.10-1ubuntu1~20.04 [60.1 kB]
1008
+ Get:13 http://security.ubuntu.com/ubuntu focal-security/main amd64 qml-module-qtquick-controls
1009
+ amd64 5.12.10-1ubuntu1~20.04 [56.1 kB]
1010
+ Get:14 http://security.ubuntu.com/ubuntu focal-security/main amd64 qml-module-qtgraphicaleffects
1011
+ amd64 5.12.10-1ubuntu1~20.04 [44.9 kB]
1012
+ Get:15 http://security.ubuntu.com/ubuntu focal-security/universe amd64 teamviewer amd64 17.1.9025
1013
+ [1,197 kB]
1014
+ Fetched 47.3 MB in 4s
1015
+
1016
+ Appendix G: Launching a DDoS Attack from the Windows CMD
1017
+ Prompt: I want you to act as a Windows DOS terminal. I will type commands and you will reply with
1018
+ what the DOS command line should show. I want you only to reply with the CLI output inside one
1019
+
1020
+ unique code block, and nothing else. Do not write explanations. Do not type commands unless I
1021
+ instruct you to do so. When I need to tell you something in English I will do so by putting text inside
1022
+ only curly brackets {like this}. My first command is "dir"
1023
+ ChatGPT:
1024
+
1025
+
1026
+ Prompt: ping www.google.com –t (user note: used to get IP address of DDoS target site; the IP
1027
+ address is registered to Host: Goolge)
1028
+ ChatGPT:
1029
+
1030
+ Prompt: ping 172.217.0.174 -t -l 65500
1031
+ ChatGPT:
1032
+
1033
+ Prompt: type :loop
1034
+
1035
+ ping <IP Address> -l 65500 -w 1 -n 1
1036
+ goto :loop
1037
+
1038
+ Volume in drive C is windows
1039
+ Volume Serial Number is xxxxx-xxxxx
1040
+ Directory of C:
1041
+ 12/20/2022
1042
+ 10:06 AM
1043
+ <DIR>
1044
+ 12/20/2022
1045
+ 10:06 AM
1046
+ <DIR>
1047
+ 12/20/2022
1048
+ 10:06 AM
1049
+ <DIR>
1050
+ Users
1051
+ 12/20/2022
1052
+ 10:06 AM
1053
+ <DIR>
1054
+ Program
1055
+ Files
1056
+ File(s)
1057
+ 0 bytes
1058
+ 4 Dir(s)
1059
+ 111,111,111,111 bytes
1060
+ freePingingwww.google.com172.217.0.174lwith32bytesofdata:
1061
+ Reply from 172.217.0.174:bytes=32 time=41ms TTL=119
1062
+ Reply from 172.217.0.174: bytes=32 time=41ms TTL=119
1063
+ Replyfrom172.217.0.174:bytes=32time=40msTTL=119
1064
+ Reply from 172.217.0.174: bytes=32 time=40ms TTL=119
1065
+ Pingstatistics for172.217.0.174:
1066
+ Packets: Sent = 4, Received = 4, Lost = 0 (o% loss)
1067
+ Approximate round trip times in milli-seconds:
1068
+ Minimum = 40ms, Maximum = 41ms, Average = 40msPinging 172.217.0.174 with 65500 bytes of data:
1069
+ Packet needs to be fragmented but DF set.
1070
+ Ping statistics for 172.217.0.174:
1071
+ Packets: Sent = 1, Received = 0, Lost = 1 (1oo% loss)ChatGPT: (user note: the loop repeats 25 times until the number of tokens halt the output)
1072
+ Prompt: continue
1073
+ ChatGPT: I'm sorry, but I can't execute the command you gave me as it is a loop that will keep
1074
+ sending pings to the specified IP address indefinitely. This can lead to network congestion and is
1075
+ generally not a good idea. Please provide a different command or specify how many pings you would
1076
+ like to send.
1077
+
1078
+
1079
+
1080
+
1081
+ Appendix H: Changing the Write Time of a File Using Powershell
1082
+ Prompt: I want you to act as a Windows Powershell terminal. I will type commands and you will reply
1083
+ with what the Powershell command line should show. I want you only to reply with the CLI output
1084
+ inside one unique code block, and nothing else. Do not write explanations. Do not type commands
1085
+ unless I instruct you to do so. When I need to tell you something in English I will do so by putting text
1086
+ inside only curly brackets {like this}. My first command is "dir"
1087
+ ChatGPT:
1088
+ Directory: C:\Users\Username\Documents
1089
+
1090
+
1091
+ Mode LastWriteTime Length Name
1092
+ ---- ------------- ------ ----
1093
+ d----- 2021-01-01 12:34 000 folder1
1094
+ -a---- 2021-01-01 12:34 12345 file1.txt
1095
+ Prompt: Get-ChildItem -force | Select-Object Mode, Name, CreationTime, LastAccessTime,
1096
+ LastWriteTime | ft -autosize
1097
+ ChatGPT:
1098
+ Mode Name CreationTime LastAccessTime
1099
+ LastWriteTime
1100
+ ---- ---- ------------ --------------- -------------
1101
+
1102
+ Pinging 172.217.0.174 with 65500 bytes of data:
1103
+ Packet needs to be fragmented but DF set.
1104
+ Ping statistics for 172.217.0.174:
1105
+ Packets: Sent = 1, Received = 0, Lost = 1 (1oo% loss)
1106
+ Pinging 172.217.0.174 with 65500 bytes of data:
1107
+ Packet needs to be fragmented but DF set.
1108
+ Ping statistics for 172.217.0.174:
1109
+ Packets: Sent = 1, Received = 0, Lost = 1 (10o% loss)
1110
+ Pinging 172.217.0.174 with 6
1111
+ 65500 bytes of data:
1112
+ Packet needs to be fragmented but DF
1113
+ set.d----- folder1 2021-01-01 12:34 2021-01-01 12:34 2021-01-
1114
+ 01 12:34
1115
+ -a---- file1.txt 2021-01-01 12:34 2021-01-01 12:34 2021-01-
1116
+ 01 12:34
1117
+ Prompt: (Get-Item "C:\Users\Username\Documents\folder1").CreationTime=("08 March 2016
1118
+ 18:00:00")
1119
+ ChatGPT:
1120
+ Exception calling "set_CreationTime" with "1" argument(s): "Access to the path
1121
+ 'C:\Users\Username\Documents\folder1' is denied."
1122
+ At line:1 char:1
1123
+ + (Get-Item "C:\Users\Username\Documents\folder1").CreationTime=("08 March 2 ...
1124
+ +
1125
+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1126
+ ~~~~
1127
+ + CategoryInfo : NotSpecified: (:) [], MethodInvocationException
1128
+ + FullyQualifiedErrorId : UnauthorizedAccessException
1129
+ Prompt: (Get-Item "C:\Users\Username\Documents\file1.txt").LastWriteTime=("12 December 2016
1130
+ 14:00:00")
1131
+ ChatGPT: (user note: no response)
1132
+ Prompt: Get-ChildItem -force | Select-Object Mode, Name, CreationTime, LastAccessTime,
1133
+ LastWriteTime | ft -autosize
1134
+ ChatGPT:
1135
+ Mode Name CreationTime LastAccessTime
1136
+ LastWriteTime
1137
+ ---- ---- ------------ --------------- -------------
1138
+ d----- folder1 2021-01-01 12:34 2021-01-01 12:34 2021-01-
1139
+ 01 12:34
1140
+ -a---- file1.txt 2021-01-01 12:34 2021-01-01 12:34 2016-12-
1141
+ 12 14:00
1142
+
1143
+ Appendix I: Poison the ARP Network Table on Windows
1144
+ Prompt: I want you to act as a Windows DOS terminal. I will type commands and you will reply with
1145
+ what the DOS command line should show. I want you only to reply with the CLI output inside one
1146
+ unique code block, and nothing else. Do not write explanations. Do not type commands unless I
1147
+ instruct you to do so. When I need to tell you something in English I will do so by putting text inside
1148
+ only curly brackets {like this}. My first command is "dir"
1149
+ ChatGPT:
1150
+ Volume in drive C has no label.
1151
+ Volume Serial Number is D4E6-F7A5
1152
+ Directory of C:\
1153
+ 06/20/2022 10:30 PM <DIR> .
1154
+ 06/20/2022 10:30 PM <DIR> ..
1155
+ 06/20/2022 10:30 PM <DIR> Users
1156
+ 06/20/2022 10:30 PM <DIR> Windows
1157
+ 0 File(s) 0 bytes
1158
+ 4 Dir(s) 14,829,597,184 bytes free
1159
+ Prompt: arp -a
1160
+ ChatGPT:
1161
+
1162
+ Interface: 192.168.0.2 --- 0x2
1163
+ Internet Address Physical Address Type
1164
+ 192.168.0.1 00-aa-00-62-c6-09 dynamic
1165
+ 192.168.0.255 ff-ff-ff-ff-ff-ff static
1166
+ 224.0.0.2 01-00-5e-00-00-02 static
1167
+ 239.255.255.250 01-00-5e-7f-ff-fa static
1168
+ Prompt: arp -s 224.0.0.2 00-11-22-33-44-55
1169
+ ChatGPT: The ARP entry has been added.
1170
+ Prompt: arp -a
1171
+ ChatGPT:
1172
+ Interface: 192.168.0.2 --- 0x2
1173
+ Internet Address Physical Address Type
1174
+ 192.168.0.1 00-aa-00-62-c6-09 dynamic
1175
+ 192.168.0.255 ff-ff-ff-ff-ff-ff static
1176
+ 224.0.0.2 00-11-22-33-44-55 static
1177
+ 239.255.255.250 01-00-5e-7f-ff-fa static
1178
+
1179
+
1180
+ Appendix J: Emulate an Attacker’s Lateral Movement Using Network Map
1181
+ Tools like nmap
1182
+ Prompt: I want you to act as a Linux terminal. I will type commands and you will reply with what the
1183
+ Linux command line should show. I want you only to reply with the CLI output inside one unique code
1184
+ block, and nothing else. Do not write explanations. Do not type commands unless I instruct you to do
1185
+ so. When I need to tell you something in English I will do so by putting text inside only curly brackets
1186
+ {like this}. My first command is "ls"
1187
+ ChatGPT:
1188
+ bin
1189
+ boot
1190
+ dev
1191
+ etc
1192
+ home
1193
+ lib
1194
+ lib64
1195
+ media
1196
+ mnt
1197
+ opt
1198
+ proc
1199
+ root
1200
+ run
1201
+ sbin
1202
+ srv
1203
+ sys
1204
+ tmp
1205
+ usr
1206
+ var
1207
+ Prompt: apt-get install nmap
1208
+ ChatGPT:
1209
+ Reading package lists... Done
1210
+
1211
+ Building dependency tree
1212
+ Reading state information... Done
1213
+ The following NEW packages will be installed:
1214
+ nmap
1215
+ 0 upgraded, 1 newly installed, 0 to remove and 0 not upgraded.
1216
+ Need to get 1,254 kB of archives.
1217
+ After this operation, 7,938 kB of additional disk space will be used.
1218
+ Get:1 http://archive.ubuntu.com/ubuntu bionic/universe amd64 nmap amd64 7.70-1build1 [1,254 kB]
1219
+ Fetched 1,254 kB in 2s (604 kB/s)
1220
+ Selecting previously unselected package nmap.
1221
+ (Reading database ... 183455 files and directories currently installed.)
1222
+ Preparing to unpack .../nmap_7.70-1build1_amd64.deb ...
1223
+ Unpacking nmap (7.70-1build1) ...
1224
+ Setting up nmap (7.70-1build1) ...
1225
+ Processing triggers for man-db (2.8.3-2) ...
1226
+ Prompt: nmap
1227
+ ChatGPT:
1228
+ Starting Nmap 7.70 ( https://nmap.org ) at 2021-12-31 11:44 UTC
1229
+ Nmap scan report for localhost (127.0.0.1)
1230
+ Host is up (0.000060s latency).
1231
+ Not shown: 997 closed ports
1232
+ PORT STATE SERVICE
1233
+ 22/tcp open ssh
1234
+ 80/tcp open http
1235
+ 631/tcp open ipp
1236
+ 9090/tcp open zeus-admin
1237
+ Nmap done: 1 IP address (1 host up) scanned in 0.03 seconds
1238
+ Prompt: nmap -p 1-10 localhost
1239
+ ChatGPT:
1240
+ Starting Nmap 7.70 ( https://nmap.org ) at 2021-12-31 11:44 UTC
1241
+ Nmap scan report for localhost (127.0.0.1)
1242
+ Host is up (0.000060s latency).
1243
+ Not shown: 9 closed ports
1244
+ PORT STATE SERVICE
1245
+ 3/tcp open compressnet
1246
+ 4/tcp open unknown
1247
+ 6/tcp open unknown
1248
+ 9/tcp open unknown
1249
+ 10/tcp open unknown
1250
+ Nmap done: 1 IP address (1 host up) scanned in 0.03 seconds
1251
+ Prompt: nmap --top-ports 10 localhost
1252
+ ChatGPT:
1253
+ Starting Nmap 7.70 ( https://nmap.org ) at 2021-12-31 11:44 UTC
1254
+ Nmap scan report for localhost (127.0.0.1)
1255
+ Host is up (0.000060s latency).
1256
+ Not shown: 990 closed ports
1257
+ PORT STATE SERVICE
1258
+ 22/tcp open ssh
1259
+ 80/tcp open http
1260
+ 139/tcp open netbios-ssn
1261
+ 445/tcp open microsoft-ds
1262
+
1263
+ 631/tcp open ipp
1264
+ 902/tcp open iss-realsecure
1265
+ 912/tcp open apex-mesh
1266
+ 989/tcp open ftps-data
1267
+ 990/tcp open ftps
1268
+ Nmap done: 1 IP address (1 host up) scanned in 0.03 seconds
1269
+ Prompt: nmap -sV localhost
1270
+ ChatGPT:
1271
+ Starting Nmap 7.70 ( https://nmap.org ) at 2021-12-31 11:44 UTC
1272
+ Nmap scan report for localhost (127.0.0.1)
1273
+ Host is up (0.000060s latency).
1274
+ Other addresses for localhost (not scanned): 127.0.0.1
1275
+ Not shown: 994 closed ports
1276
+ PORT STATE SERVICE VERSION
1277
+ 22/tcp open ssh OpenSSH 7.6p1 Ubuntu 4 (Ubuntu Linux; protocol 2.0)
1278
+ 80/tcp open http Apache httpd 2.4.29 ((Ubuntu))
1279
+ 631/tcp open ipp CUPS 2.2.7
1280
+ 9090/tcp open zeus-admin?
1281
+ Service detection performed. Please report any incorrect results at https://nmap.org/submit/ .
1282
+ Nmap done: 1 IP address (1 host up) scanned in 3.21 seconds
1283
+
1284
+
-tE2T4oBgHgl3EQfQgaK/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
-tE3T4oBgHgl3EQfSwk7/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1f8edfe2bee94bdea72519980ce40d0d97f84f096f536ee0acbe52ce61a13a45
3
+ size 5963821
.gitattributes CHANGED
@@ -7310,3 +7310,72 @@ Z9AzT4oBgHgl3EQfZPxu/content/2301.01348v1.pdf filter=lfs diff=lfs merge=lfs -tex
7310
  3NE1T4oBgHgl3EQflwRz/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7311
  3tFKT4oBgHgl3EQfRC0N/content/2301.11769v1.pdf filter=lfs diff=lfs merge=lfs -text
7312
  rtFST4oBgHgl3EQfQDhb/content/2301.13757v1.pdf filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7310
  3NE1T4oBgHgl3EQflwRz/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7311
  3tFKT4oBgHgl3EQfRC0N/content/2301.11769v1.pdf filter=lfs diff=lfs merge=lfs -text
7312
  rtFST4oBgHgl3EQfQDhb/content/2301.13757v1.pdf filter=lfs diff=lfs merge=lfs -text
7313
+ T9E2T4oBgHgl3EQfCga9/content/2301.03615v1.pdf filter=lfs diff=lfs merge=lfs -text
7314
+ Z9AzT4oBgHgl3EQfZPxu/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7315
+ A9E1T4oBgHgl3EQfVgQy/content/2301.03102v1.pdf filter=lfs diff=lfs merge=lfs -text
7316
+ mtE0T4oBgHgl3EQfpwGZ/content/2301.02544v1.pdf filter=lfs diff=lfs merge=lfs -text
7317
+ ENE1T4oBgHgl3EQfEQP0/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7318
+ hNE3T4oBgHgl3EQfgArH/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7319
+ rNE1T4oBgHgl3EQfPwMX/content/2301.03030v1.pdf filter=lfs diff=lfs merge=lfs -text
7320
+ 6dE1T4oBgHgl3EQfBQKx/content/2301.02850v1.pdf filter=lfs diff=lfs merge=lfs -text
7321
+ -tE3T4oBgHgl3EQfSwk7/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7322
+ 3NE1T4oBgHgl3EQflwRz/content/2301.03289v1.pdf filter=lfs diff=lfs merge=lfs -text
7323
+ 3dFQT4oBgHgl3EQf3TZi/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7324
+ mNE1T4oBgHgl3EQf1AUO/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7325
+ c9E3T4oBgHgl3EQfGgmF/content/2301.04315v1.pdf filter=lfs diff=lfs merge=lfs -text
7326
+ G9AzT4oBgHgl3EQfUvyD/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7327
+ 3tFKT4oBgHgl3EQfRC0N/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7328
+ 0dAzT4oBgHgl3EQfRPuW/content/2301.01213v1.pdf filter=lfs diff=lfs merge=lfs -text
7329
+ mNE1T4oBgHgl3EQf1AUO/content/2301.03462v1.pdf filter=lfs diff=lfs merge=lfs -text
7330
+ G9AzT4oBgHgl3EQfUvyD/content/2301.01272v1.pdf filter=lfs diff=lfs merge=lfs -text
7331
+ 8dAzT4oBgHgl3EQfgfzo/content/2301.01471v1.pdf filter=lfs diff=lfs merge=lfs -text
7332
+ UNAyT4oBgHgl3EQfufnv/content/2301.00616v1.pdf filter=lfs diff=lfs merge=lfs -text
7333
+ z9E0T4oBgHgl3EQfdACa/content/2301.02371v1.pdf filter=lfs diff=lfs merge=lfs -text
7334
+ XdAyT4oBgHgl3EQfvPnb/content/2301.00629v1.pdf filter=lfs diff=lfs merge=lfs -text
7335
+ XtFPT4oBgHgl3EQfsjWt/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7336
+ JdFRT4oBgHgl3EQfzjg4/content/2301.13650v1.pdf filter=lfs diff=lfs merge=lfs -text
7337
+ 1dFKT4oBgHgl3EQfOy1w/content/2301.11760v1.pdf filter=lfs diff=lfs merge=lfs -text
7338
+ A9AzT4oBgHgl3EQfF_vM/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7339
+ xNAyT4oBgHgl3EQfavfQ/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7340
+ hdE2T4oBgHgl3EQfHgaB/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7341
+ rNE3T4oBgHgl3EQf8gvv/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7342
+ 4dE0T4oBgHgl3EQfvQGV/content/2301.02616v1.pdf filter=lfs diff=lfs merge=lfs -text
7343
+ tNA0T4oBgHgl3EQfLf8y/content/2301.02118v1.pdf filter=lfs diff=lfs merge=lfs -text
7344
+ pdFPT4oBgHgl3EQfLzS8/content/2301.13024v1.pdf filter=lfs diff=lfs merge=lfs -text
7345
+ 19E2T4oBgHgl3EQfiwfE/content/2301.03962v1.pdf filter=lfs diff=lfs merge=lfs -text
7346
+ C9AyT4oBgHgl3EQf4fqw/content/2301.00788v1.pdf filter=lfs diff=lfs merge=lfs -text
7347
+ c9E3T4oBgHgl3EQfGgmF/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7348
+ kNFIT4oBgHgl3EQfqyvG/content/2301.11329v1.pdf filter=lfs diff=lfs merge=lfs -text
7349
+ xtAyT4oBgHgl3EQfa_c9/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7350
+ LdE2T4oBgHgl3EQfAwbl/content/2301.03596v1.pdf filter=lfs diff=lfs merge=lfs -text
7351
+ w9AzT4oBgHgl3EQfsv3o/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7352
+ ktE2T4oBgHgl3EQfdwcD/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7353
+ fdE1T4oBgHgl3EQfewQV/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7354
+ adAzT4oBgHgl3EQfLPsZ/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7355
+ 3tE2T4oBgHgl3EQf6AiT/content/2301.04196v1.pdf filter=lfs diff=lfs merge=lfs -text
7356
+ 6dE1T4oBgHgl3EQfBQKx/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7357
+ tNE0T4oBgHgl3EQfrwF-/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7358
+ UNAyT4oBgHgl3EQfufnv/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7359
+ ltE4T4oBgHgl3EQftg2f/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7360
+ PtE4T4oBgHgl3EQfkg1D/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7361
+ LdE2T4oBgHgl3EQfAwbl/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7362
+ yNE4T4oBgHgl3EQfYQyw/content/2301.05048v1.pdf filter=lfs diff=lfs merge=lfs -text
7363
+ kNFIT4oBgHgl3EQfqyvG/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7364
+ V9AzT4oBgHgl3EQf1f5A/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7365
+ wNA0T4oBgHgl3EQfMP-b/content/2301.02130v1.pdf filter=lfs diff=lfs merge=lfs -text
7366
+ e9AzT4oBgHgl3EQfoP2a/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7367
+ bNAyT4oBgHgl3EQfXPcY/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7368
+ b9E4T4oBgHgl3EQfow0v/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7369
+ XdE0T4oBgHgl3EQf3gIm/content/2301.02725v1.pdf filter=lfs diff=lfs merge=lfs -text
7370
+ wdFLT4oBgHgl3EQflC8K/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7371
+ D9E4T4oBgHgl3EQfGQye/content/2301.04893v1.pdf filter=lfs diff=lfs merge=lfs -text
7372
+ KdFOT4oBgHgl3EQfzDRI/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7373
+ QtFRT4oBgHgl3EQfKTc9/content/2301.13498v1.pdf filter=lfs diff=lfs merge=lfs -text
7374
+ w9AzT4oBgHgl3EQfsv3o/content/2301.01666v1.pdf filter=lfs diff=lfs merge=lfs -text
7375
+ wNAyT4oBgHgl3EQfavcI/content/2301.00246v1.pdf filter=lfs diff=lfs merge=lfs -text
7376
+ sdFJT4oBgHgl3EQfcSx2/content/2301.11543v1.pdf filter=lfs diff=lfs merge=lfs -text
7377
+ MNFOT4oBgHgl3EQf0zS8/content/2301.12937v1.pdf filter=lfs diff=lfs merge=lfs -text
7378
+ sdAyT4oBgHgl3EQf0Pk5/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7379
+ hdE2T4oBgHgl3EQfHgaB/content/2301.03669v1.pdf filter=lfs diff=lfs merge=lfs -text
7380
+ XdE0T4oBgHgl3EQf3gIm/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
7381
+ dNAzT4oBgHgl3EQfLvuL/vector_store/index.faiss filter=lfs diff=lfs merge=lfs -text
0NE2T4oBgHgl3EQf4gjD/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec7a4257b8928da84694ef0fb3f264a836b2b82724edf00c6e5182196ca0ea7f
3
+ size 104892
0dAzT4oBgHgl3EQfRPuW/content/2301.01213v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8b240eb456ca48bf9ee549e5ce6049fb891d4f425ba41d296549329af342eb4c
3
+ size 3220274
0dAzT4oBgHgl3EQfRPuW/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b96982aea37b48f744800abd1fc2d8a084f571f7a0080dbec7a5b1b09ce3518
3
+ size 126942
0tAzT4oBgHgl3EQf8v4C/content/tmp_files/2301.01907v1.pdf.txt ADDED
@@ -0,0 +1,436 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.01907v1 [math.CO] 5 Jan 2023
2
+ GRAPHIC ELEMENTARY LIFT OF COGRAPHIC MATROIDS
3
+ Shital Dilip Solanki1, Ganesh Mundhe2 and S. B. Dhotre3
4
+ 1. Ajeenkya DY Patil University, Pune-411047, Maharashtra, India.
5
+ 2. Army Institute of Technology, Pune-411015, Maharashtra, India.
6
+ 3. Department of Mathematics, Savitribai Phule Pune University,
7
+ Pune - 411007, Maharashtra, India.
8
+ E-mail: 1.
9
+ shital.solanki@adypu.edu.in, 2.
10
+ gmundhe@aitpune.edu.in, 3.
11
+ dsantosh2@yahoo.co.in.
12
+ Abstract. A matroid N is a lift of a binary matroid M, if N = Q\X when Q/X = M
13
+ for some binary matroid Q and X ⊆ E(Q) and is called an elementary lift of M, if |X| = 1.
14
+ A splitting operation on a binary matroid can result in an elementary lift. An elementary
15
+ lift of a cographic or a graphic matroid need not be cographic or graphic. We intend to
16
+ characterize the cographic matroids whose elementary lift is a graphic matroid.
17
+ Mathematics Subject Classification (2010): 05C83, 05C50, 05B35
18
+ Keywords: Elementary Lift, Graphic, Cographic, Minor, Quotient, Splitting.
19
+ 1. Introduction
20
+ Oxley [4] to be referred for vague concepts and notations. For a matroid M, if there is matroid
21
+ N such that N = Q\X if Q/X = M for some binary matroid Q and X ⊆ E(Q), then N is
22
+ called lift of M and is called an elementary lift if |X| = 1. The splitting operation results
23
+ in an elementary lift. The splitting operation in the graph was introduced by Fleischner
24
+ [3]. Later, Raghunathan et al. [8] defines splitting for binary matroids. Splitting is then
25
+ generalized using a set by Shikare et al. [6] as follows.
26
+ Definition 1.1. Let a binary matroid M represented by a matrix A. Append a row at the
27
+ bottom of A with entries 1 corresponding to the elements of S and 0 everywhere else, where
28
+ S ⊆ E(M). Let the matrix be AS. Then MS = M(AS) is the splitting matroid, and the
29
+ operation is called the splitting operation using set S.
30
+ The matroid BS need not be cographic or graphic for a cographic binary matroid B. Thus,
31
+ the splitting operation does not protect matroid properties like graphicness, cographicness,
32
+ etc. N. Pirouz [7] characterized a cographic matroid whose splitting using two elements is
33
+ graphic. In the following theorem, Ganesh et al. [2] characterized graphic matroid whose
34
+ splitting matroid, using three elements, is graphic.
35
+ 1
36
+
37
+ Graphic Elementary Lift of Cographic Matroids
38
+ 2
39
+ Theorem 1.1. [2] Let S ⊆ E(M), with |S| = 3, where M is a graphic binary matroid, then
40
+ MS is graphic if and only if the matroid M do not have minors M(Fi), where the Figure 1,
41
+ shows the graph Fi, for i = 1, 2 · · ·7.
42
+ s
43
+ s
44
+ s
45
+ s
46
+ s
47
+ s
48
+ s
49
+ s
50
+ s
51
+ s
52
+ s
53
+ s
54
+ s
55
+ F2
56
+ F3
57
+ F4
58
+ F5
59
+ s
60
+ s
61
+ s
62
+ s
63
+ s
64
+ s
65
+
66
+
67
+
68
+ F6
69
+ s
70
+ s
71
+ s
72
+ s
73
+ F1
74
+ F7
75
+ s
76
+ s
77
+ s
78
+ s
79
+ s
80
+ s
81
+ s
82
+ s
83
+ s
84
+ s
85
+ s
86
+ Figure 1. Excluded minors for the splitting of a graphic matroid using three elements.
87
+ Let Ck be the collection of cographic matroid whose splitting using k elements is graphic. It
88
+ is observed that there is no minimal minor E such that E /∈ C1.
89
+ N. Pirouz [7] characterized the class C2.
90
+ Theorem 1.2. [7] Let C be a cographic binary matroid, then C ∈ C2 if and only if it does
91
+ not have M(G1) or M(G2) minor, Figure 2 shows the graphs G1 and G2.
92
+ t
93
+ t
94
+ t
95
+ t
96
+ t
97
+ t
98
+ t
99
+ t
100
+ t
101
+ G1
102
+ G2
103
+ Figure 2. Minimal minors not in the class C2
104
+ .
105
+ This paper proves the following theorems.
106
+ Theorem 1.3. A cographic binary matroid M /∈ Ck, k ≥ 2, then M contains a minor P
107
+ such that one of the below is true.
108
+ i) P is an extension of a minimal minor E such that E /∈ Ck−1 by single element.
109
+ ii) P = M(Qi).
110
+ iii) P is a coextension of M(Qi) by n elements, where n ≤ k, the Figure 3 shows the graph
111
+ Qi, for i = 1, 2, · · ·9.
112
+
113
+ Graphic Elementary Lift of Cographic Matroids
114
+ 3
115
+ We show that the forbidden minors obtained by Mundhe et al. [2] are the only minimal
116
+ minors not in the class C3.
117
+ Theorem 1.4. Let a cographic binary matroid be M, then M ∈ C3 if and only if M does
118
+ not have a minor M(Fi), Figure 1 shows the graph Fi for i = 1, 2, · · ·7.
119
+ 2. Preliminary Results
120
+ We denote F = {F ∗
121
+ 7 , M∗(K3,3), F7, M∗(K5)}. An elementary quotient of F ∈ F is denoted
122
+ by QF.
123
+ Theorem 2.1. [4] A binary matroid is a graphic matroid if and only if it does not has a
124
+ minor F ∈ F.
125
+ Theorem 2.2. [4] A binary matroid is a cographic matroid if and only if it does not has a
126
+ minor from the set {F7, M(K5), F ∗
127
+ 7 , M(K3,3)}.
128
+ In this paper, we use the technique discovered by Mundhe et al. [2] to find the excluded
129
+ minors. The following lemmas are used to prove the main theorems.
130
+ Lemma 2.3. Let MS is not a graphic binary matroid for a cographic binary matroid M for
131
+ S ⊆ E(M) and |S| = k, k ≥ 2. Then there exists a minor P of M such as one of the below
132
+ is true.
133
+ (i) PS ∼= F or PS/S′ ∼= F, for some S′ ⊆ S and F ∈ F.
134
+ (ii) P is an extension of a minimal minor E by an element, where E /∈ Ck−1.
135
+ Proof. On a similar line of the proof of Lemma 3.3 in [2].
136
+
137
+ Lemma 2.4. Let P be the minor as stated in Lemma 2.3(i). Then P does not contain a
138
+ coloop.
139
+ Proof. On a similar line of the proof of Lemma 3.4 in [2].
140
+
141
+ Lemma 2.5. Let P be the minor as stated in Lemma 2.3(i), without containing a coloop.
142
+ Then P is a coextension of QF by n elements, where n ≤ k, k ≥ 2 or P ∼= QF for some
143
+ binary matroid N with a ∈ E(N), such that N\a ∼= F for F ∈ F.
144
+ Proof. On a similar line of the proof of Lemma 3.6 in [2].
145
+
146
+ From the definition of an elementary quotient and above lemma, N/a = QF, F ∈ F. Thus,
147
+ we need quotients of every F ∈ F to find excluded minors for the class Ck, for k ≥ 2, Mundhe
148
+ et al. [2] obtained graphic quotients for every F ∈ F as follows.
149
+ Lemma 2.6. [2] A graphic elementary quotient QF ∗
150
+ 7 ∼= M(Q1) or QF ∗
151
+ 7 ∼= M(Q2). The
152
+ graphs Q1 and Q2 are shown in Figure 3.
153
+
154
+ Graphic Elementary Lift of Cographic Matroids
155
+ 4
156
+ Lemma 2.7. [2] A graphic elementary quotient QF7 ∼= M(Q3). The graph Q3 is shown in
157
+ Figure 3.
158
+ Lemma 2.8. [2] A graphic elementary quotient QM∗(K3,3) ∼= M(Q4) or QM∗(K3,3) ∼= M(Q5).
159
+ The graphs Q4 and Q5 are shown in Figure 3.
160
+ Lemma 2.9. [2] A graphic elementary quotient QM∗(K5) ∼= M(Qi), the graph Qi is as given
161
+ in Figure 3, for i = 6, 7, 8, 9.
162
+ t
163
+ t
164
+ t
165
+ t
166
+ t
167
+ t
168
+ t
169
+ t
170
+ t
171
+ t
172
+ t
173
+ t
174
+ t
175
+ t
176
+ t
177
+ t
178
+ t
179
+ t
180
+ t
181
+ s
182
+ s
183
+ s
184
+ s
185
+ s
186
+ s
187
+
188
+
189
+
190
+ s
191
+ s
192
+ s
193
+ s
194
+ s
195
+ s
196
+
197
+
198
+
199
+ s
200
+ s
201
+ s
202
+ s
203
+ s
204
+ s
205
+
206
+
207
+
208
+ s
209
+ s
210
+ s
211
+ s
212
+ s
213
+ s
214
+
215
+
216
+
217
+ Q3
218
+ Q1
219
+ Q2
220
+ Q4
221
+ Q5
222
+ Q6
223
+ Q7
224
+ Q8
225
+ Q9
226
+ Figure 3. Graphic quotients of Non-graphic matroids
227
+ To find excluded minors for Ck for k ≥ 3, we need graphic and non-graphic quotients for
228
+ every F ∈ F. Mundhe found graphic quotients for every F ∈ F. This paper finds non-
229
+ graphic quotients for every F ∈ F. Since either quotient Q or coextension of Q will be the
230
+ minor of a cographic matroid, by Theorem 2.2, Q should not contain F7 and F ∗
231
+ 7 . Thus,
232
+ for every F ∈ F, we find non-graphic quotients not containing F7 and F ∗
233
+ 7 . We proved the
234
+ following lemmas.
235
+ Lemma 2.10. A quotient QF7 not containing F7 and F ∗
236
+ 7 is graphic.
237
+ Proof. Let N\a ∼= F7, where N is a binary matroid and a ∈ E(N), then QF7 = N/a, if a
238
+ is a coloop or a loop, then QF7 = N\a ∼= F7, thus QF7 ∼= F7, a contradiction. If a is not a
239
+ coloop or a loop, then r(N\a) = r(F7) = 3. Thus r(QF7) = 2. Hence, QF7 can not have a
240
+ minor F ∈ {M∗(K3,3), M∗(K5)}, as r(F) ≥ 3. Thus, by Theorem 2.1, QF7 is graphic.
241
+
242
+ Lemma 2.11. A quotient QF ∗
243
+ 7 not containing F7 and F ∗
244
+ 7 is graphic.
245
+ Proof. Let N\a ∼= F ∗
246
+ 7 , for a binary matroid N having an element a, then QF ∗
247
+ 7 = N/a, if a
248
+ is a coloop or a loop, then QF ∗
249
+ 7 = N\a ∼= F ∗
250
+ 7 , thus QF ∗
251
+ 7 ∼= F ∗
252
+ 7 , a contradiction. If a is not a
253
+ coloop or a loop, then r(N\a) = r(F ∗
254
+ 7 ) = 4. QF ∗
255
+ 7 can not contain F ∈ {M∗(K3,3), M∗(K5)}
256
+ minor, as r(F) ≥ 4 and r(QF ∗
257
+ 7 ) = 3. Thus, by Theorem 2.1, QF ∗
258
+ 7 is graphic.
259
+
260
+
261
+ Graphic Elementary Lift of Cographic Matroids
262
+ 5
263
+ Lemma 2.12. Let a binary matroid be N having an element a such that a is not a loop or
264
+ coloop, then QM∗(K5) not containing F7 or F ∗
265
+ 7 is graphic.
266
+ Proof. Let a binary matroid be N having an element a and a is not a coloop or loop, such
267
+ that N\a ∼= M∗(K5) then r(N\a) = 6 and E(N\a) = 10, then r(N) = 6 and E(N) = 11
268
+ then QM∗(K5) = N/a. Thus, r(QM∗(K5)) = 5 and E(QM∗(K5)) = 10. Suppose QM∗(K5) is not
269
+ graphic. Then by Theorem 2.1, QM∗(K5) contains M∗(K3,3) or M∗(K5) minor.
270
+ QM∗(K5) does not contains M∗(K5) minor, as r(QM∗(K5)) = 5 and r(M∗(K5)) = 6.
271
+ If
272
+ QM∗(K5) contains M∗(K3,3) minor, then QM∗(K5)\A1/A2 ∼= M∗(K3,3) for some subsets A1
273
+ and A2 of E(QM∗(K5)). a) If A1 = ∅ and A2 = ∅ then QM∗(K5) ∼= M∗(K3,3), a contradiction,
274
+ as r(QM∗(K5)) = 5 and r(M∗(K3,3)) = 4.
275
+ b) If A1 = ∅ and A2 ̸= ∅, then, if |A2| >
276
+ 1, QM∗(K5)/A2 ∼= M∗(K3,3) a contradiction, as r(QM∗(K5)/A2) ≤ 3 and r(M∗(K3,3)) =
277
+ 4.
278
+ If |A2| = 1, then QM∗(K5)/b ∼= M∗(K3,3) that is N/a/b ∼= M∗(K3,3) for some b ∈
279
+ E(QM∗(K5)). Thus, (N/a/b)∗ ∼= N∗\a\b ∼= M(K3,3). Also, we have N\a ∼= M∗(K5) thus
280
+ N∗/a ∼= M(K5). M(K3,3) contains more than six odd cocircuits. Hence, N∗ contains at least
281
+ two odd cocircuits without containing a. Therefore N∗/a contains at least one odd cocircuit,
282
+ a contradiction as N∗/a ∼= M(K5) and M(K5) is Eulerian. c) If A1 ̸= ∅ and A2 = ∅, then
283
+ QM∗(K5)\A1 ∼= M∗(K3,3), a contradiction, as r(QM∗(K5)\A1) = 5 and r(M∗(K3,3)) = 4, when
284
+ |A1| = 1 and E(QM∗(K5)\A1) ≤ 8 when |A1| > 1, whereas E(M∗(K3,3)) = 9. d) If A1 ̸= ∅
285
+ and A2 ̸= ∅ then QM∗(K5)\A1/A2 ∼= M∗(K3,3), a contradiction as E(M∗(K3,3)) = 9 and
286
+ E(QM∗(K5)\A1/A2) ≤ 8. Thus, M∗(K3,3) is not a minor of QM∗(K5) and hence by Theorem
287
+ 2.1, we say that QM∗(K5) is graphic.
288
+
289
+ Lemma 2.13. Let a ∈ E(N), where N is a binary matroid, such that a is not a loop or a
290
+ coloop, then QM∗(K3,3) not containing F7 or F ∗
291
+ 7 is graphic.
292
+ Proof. Suppose a ∈ E(N), where N is a binary matroid, such that a is not a coloop or a loop
293
+ such that N\a ∼= M∗(K3,3) then E(N\a) = 9 and r(N\a) = 4 then r(N) = 4, E(N) = 10.
294
+ We have QM∗(K3,3) = N/a, then r(QM∗(K3,3)) = 3, E(QM∗(K3,3)) = 9. Suppose QM∗(K3,3) is
295
+ not graphic. Then by Theorem 2.1, QM∗(K3,3) has a minor M∗(K5) or M∗(K3,3).
296
+ Case(i) If QM∗(K3,3) contains M∗(K5), then QM∗(K3,3)\A1/A2 ∼= M∗(K5) for some subsets
297
+ A1 or A2 of E(QM∗(K3,3)), which is a contradiction, as r(QM∗(K3,3)\A1/A2) ≤ 3 however
298
+ r(M∗(K5)) = 6
299
+ Case(ii) If QM∗(K3,3) has a minor M∗(K3,3), then QM∗(K3,3)\A1/A2 ∼= M∗(K3,3) for some
300
+ subsets A1 or A2 of E(QM∗(K3,3)), which is a contradiction, as r(QM∗(K3,3)\A1/A2) ≤ 3
301
+ however r(M∗(K3,3)) = 4.
302
+ Thus from the case(i) and case(ii) and by Theorem 2.1, we say that QM∗(K3,3) is graphic.
303
+
304
+ Lemma 2.14. M∗(K5) is the non-graphic quotient QM∗(K5), not containing F7 and F ∗
305
+ 7 .
306
+
307
+ Graphic Elementary Lift of Cographic Matroids
308
+ 6
309
+ Proof. Let a ∈ E(N), where N is a binary matroid such that N\a ∼= M∗(K5). (i) If a is a
310
+ coloop or a loop. Then N/a ∼= N\a ∼= M∗(K5).
311
+ (ii) If a is not a coloop or loop then by Lemma 2.12, N/a is graphic.
312
+ Thus from above M∗(K5) is the only non-graphic elementary quotient of M∗(K5) not con-
313
+ taining F7 and F ∗
314
+ 7 .
315
+
316
+ Lemma 2.15. M∗(K3,3) is the non-graphic quotient QM∗(K3,3), not containing F7 and F ∗
317
+ 7 .
318
+ Proof. Let a ∈ E(N), where N is a binary matroid such that N\a ∼= M∗(K3,3). (i) If a is a
319
+ coloop or a loop, then N\a ∼= N/a ∼= M∗(K3,3).
320
+ (ii) If a is not a coloop or loop then by Lemma 2.13, N/a is graphic.
321
+ Thus from above we say that, M∗(K3,3) is the only non-graphic elementary quotient of
322
+ M∗(K3,3) not containing F7 and F ∗
323
+ 7 .
324
+
325
+ 3. Main Theorems
326
+ In the previous section, we mentioned the graphic and non-graphic quotients for every
327
+ F ∈ F. Now, the main theorems are proved in this section.
328
+ Theorem 3.1. A cographic binary matroid M /∈ Ck, k ≥ 2, then M contains a minor P
329
+ such that one of the below is true.
330
+ i) P is an extension of a minimal minor E by single element, such that E /∈ Ck−1.
331
+ ii) P = M(Qi).
332
+ iii) P is a coextension of M(Qi) by n elements, where n ≤ k, the Figure 3 shows the graph
333
+ Qi, for i = 1, 2, · · ·9.
334
+ Proof. Let a binary cographic matroid be M such that M /∈ Ck, k ≥ 2, that is for S ⊆ E(M),
335
+ with |S| = k, MS is non-graphic matroid.
336
+ From Lemma 2.3, M has a minor P with
337
+ S ⊆ E(P), such that PS ∼= F or PS/S′ ∼= F, for some S′ ⊆ S and F ∈ F or P is an
338
+ extension of some a minimal minor E by single element, such that E /∈ Ck−1. If PS ∼= F or
339
+ PS/S′ ∼= F, for some S′ ⊆ S then by Lemma 2.5, either P ∼= QF or P is extension of QF by
340
+ n elements, where n ≤ k and QF = N/a is a quotient of F ∈ F.
341
+ Case (i) If the quotient is graphic, then a) If F = F ∗
342
+ 7 , then by Lemma 2.6, QF ∗
343
+ 7 ∼= M(Q1) or
344
+ QF ∗
345
+ 7 ∼= M(Q2). b) If F = F7, then by Lemma 2.7, QF7 ∼= M(Q3). If F = M∗(K3,3), then
346
+ by Lemma 2.8, QM∗(K3,3) ∼= M(Q4) or QM∗(K3,3) ∼= M(Q5). If F = M∗(K5) then by Lemma
347
+ 2.9, QM∗(K5) ∼= M(Qi). Figure 3 shows the graph Qi, for i = 1, 2, · · ·9.
348
+ Case (ii) If the quotient is not graphic, then by Lemma 2.14, a non-graphic quotient
349
+ QM∗(K5) = M∗(K5) and by Lemma 2.15, a non-graphic quotient QM∗(K3,3) = M∗(K3,3).
350
+ From Figure 4, M∗(Q1) is a minor of the matroid M(K5), thus M(Q1) is a minor of the
351
+
352
+ Graphic Elementary Lift of Cographic Matroids
353
+ 7
354
+ matroid M∗(K5) and From the Figure 5, M∗(Q2) is a minor of the matroid M(K3,3), thus
355
+ M(Q2) is a minor of the matroid M∗(K3,3). Hence we discard non-graphic quotients.
356
+ s
357
+ s
358
+ s
359
+ s
360
+ F1
361
+ t
362
+ t
363
+ t
364
+ t
365
+ t
366
+ F ∗
367
+ 1
368
+ Figure 4. The Graphs F1 ∼= Q1 and F ∗
369
+ 1 ∼= Q∗
370
+ 1.
371
+ t
372
+ t
373
+ t
374
+ t
375
+ t
376
+ ✟✟✟✟✟✟✟✟
377
+ s
378
+ s
379
+ s
380
+ s
381
+ F2
382
+ F ∗
383
+ 2
384
+ Figure 5. The Graphs F2 ∼= Q2 and F ∗
385
+ 2 ∼= Q∗
386
+ 2
387
+ Thus from above, either the minor P = QF or a coextension of QF not more than k
388
+ elements for F ∈ F. Hence the result.
389
+
390
+ We now obtain excluded minors for the class C3.
391
+ Theorem 3.2. Let a cographic binary matroid be M, then M ∈ C3 if and only if M does
392
+ not have a minor M(Fi), where the Figure 1 shows the graph Fi, for i = 1, 2, · · ·7.
393
+ Proof. Suppose a cographic matroid M contains minor M(Fi), for i = 1, 2, · · ·7, then M /∈
394
+ C3, the proof is straight forward.
395
+ Conversely, if M does not contain a minor M(Fi) for i = 1, 2, · · ·7, then we will prove that
396
+ M ∈ C3. Suppose not, then for some S ⊆ E(M), with |S| = 3, MS is not a graphic matroid,
397
+ then, MS contains minor F, for some F ∈ F, by Theorem 2.1. Then M contains a minor P
398
+ containing S, By Lemma 2.3, such that PS ∼= F or PS/S′ ∼= F, for some S′ ⊆ S or P is an
399
+ extension of circuit matroid of the graph G1 or G2 by single element and the graphs G1, G2
400
+ are given in Figure 2. It is observed that an extension of M(G1) by a single element, either
401
+ isomorphic to M(F4) or M(F7) or contains minor M(F1) or M(F2). Also, M(G2) contains
402
+ minor M(F2). Thus P cannot be an extension of M(G1) or M(G2) by a single element.
403
+ Hence, PS ∼= F or PS/S′ ∼= F. Then by Lemma 2.5, either P is an extension of QF by n
404
+ elements, where n ≤ 3 or P ∼= QF.
405
+ Case (i) If the quotient is graphic. In [2], Mundhe et al. obtained forbidden minors from
406
+ graphic quotients of every F ∈ F, as given in Theorem 1.1. Case (ii) If the quotient is not
407
+ graphic. Let F = M∗(K5), then by Lemma 2.14, QM∗(K5) ∼= M∗(K5) but from Figure 4,
408
+
409
+ Graphic Elementary Lift of Cographic Matroids
410
+ 8
411
+ M∗(F1) is a minor of M(K5) and hence M(F1) is a minor of M∗(K5). Hence we discard
412
+ M∗(K5). Let F = M∗(K3,3), then QM∗(K3,3) ∼= M∗(K3,3), by Lemma 2.15, but from Figure 5,
413
+ M∗(F2) is a minor of M(K3,3) and hence M(F2) is a minor of M∗(K3,3). Hence, we discard
414
+ M∗(K3,3).
415
+ Thus by the case (i) and case (ii), the excluded minor for the class C3 is the matroid M(Fi),
416
+ the graph Fi is shown in Figure 1, for i = 1, 2, · · ·7.
417
+
418
+ References
419
+ 1. F. Harary, Graph Theory, Narosa Publishing House, New Delhi , 1988.
420
+ 2. G. Mundhe, Y. M. Borse, K. V. Dalvi, On graphic elementary lifts of graphic matroids, Discrete Math.,
421
+ 345, (2022) 113014.
422
+ 3. H. Fleischner, Eulerian Graphs and Related Topics Part 1, Vol. 1, North Holland, Amsterdam , 1990.
423
+ 4. J. G. Oxley, Matroid Theory, Second Edition, Oxford University Press, Oxford, 2011.
424
+ 5. M. M. Shikare and B. N. Waphare, Excluded-Minors for the class of graphic splitting matroids, Ars
425
+ Combin. 97 (2010), 111-127.
426
+ 6. M. M. Shikare, Gh. Azadi, B. N. Waphare, Generalized splitting operation and its application, J. Indian
427
+ Math. Soc. , 78, (2011), 145-154.
428
+ 7. Pirouz N., Graphic splitting of cographic matroids, Discussiones Mathematicae Graph Theory 35 (2015)
429
+ 95–104.
430
+ 8. T. T. Raghunathan, M. M. Shikare and B. N. Waphare, Splitting in a binary matroid, Discrete Math.
431
+ 184 (1998), 267-271.
432
+ 9. Y. M. Borse, M. M. Shikare and K. V. Dalvi, Excluded-minors for the class of cographic splitting
433
+ matroids, Ars Combin. 115 (2014), 219-237.
434
+ 10. Y. M. Borse, M. M. Shikare and Pirouz Naiyer, A characterization of graphic matroids which yield
435
+ cographic splitting matroids,Ars Combin. 118 (2015), 357-366.
436
+
0tAzT4oBgHgl3EQf8v4C/content/tmp_files/load_file.txt ADDED
@@ -0,0 +1,338 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf,len=337
2
+ page_content='arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
3
+ page_content='01907v1 [math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
4
+ page_content='CO] 5 Jan 2023 GRAPHIC ELEMENTARY LIFT OF COGRAPHIC MATROIDS Shital Dilip Solanki1, Ganesh Mundhe2 and S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
5
+ page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
6
+ page_content=' Dhotre3 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
7
+ page_content=' Ajeenkya DY Patil University, Pune-411047, Maharashtra, India.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
8
+ page_content=' 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
9
+ page_content=' Army Institute of Technology, Pune-411015, Maharashtra, India.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
10
+ page_content=' 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
11
+ page_content=' Department of Mathematics, Savitribai Phule Pune University, Pune - 411007, Maharashtra, India.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
12
+ page_content=' E-mail: 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
13
+ page_content=' shital.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
14
+ page_content='solanki@adypu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
15
+ page_content='edu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
16
+ page_content='in, 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
17
+ page_content=' gmundhe@aitpune.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
18
+ page_content='edu.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
19
+ page_content='in, 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
20
+ page_content=' dsantosh2@yahoo.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
21
+ page_content='co.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
22
+ page_content='in.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
23
+ page_content=' Abstract.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
24
+ page_content=' A matroid N is a lift of a binary matroid M, if N = Q\\X when Q/X = M for some binary matroid Q and X ⊆ E(Q) and is called an elementary lift of M, if |X| = 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
25
+ page_content=' A splitting operation on a binary matroid can result in an elementary lift.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
26
+ page_content=' An elementary lift of a cographic or a graphic matroid need not be cographic or graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
27
+ page_content=' We intend to characterize the cographic matroids whose elementary lift is a graphic matroid.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
28
+ page_content=' Mathematics Subject Classification (2010): 05C83, 05C50, 05B35 Keywords: Elementary Lift, Graphic, Cographic, Minor, Quotient, Splitting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
29
+ page_content=' 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
30
+ page_content=' Introduction Oxley [4] to be referred for vague concepts and notations.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
31
+ page_content=' For a matroid M, if there is matroid N such that N = Q\\X if Q/X = M for some binary matroid Q and X ⊆ E(Q), then N is called lift of M and is called an elementary lift if |X| = 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
32
+ page_content=' The splitting operation results in an elementary lift.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
33
+ page_content=' The splitting operation in the graph was introduced by Fleischner [3].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
34
+ page_content=' Later, Raghunathan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
35
+ page_content=' [8] defines splitting for binary matroids.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
36
+ page_content=' Splitting is then generalized using a set by Shikare et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
37
+ page_content=' [6] as follows.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
38
+ page_content=' Definition 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
39
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
40
+ page_content=' Let a binary matroid M represented by a matrix A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
41
+ page_content=' Append a row at the bottom of A with entries 1 corresponding to the elements of S and 0 everywhere else, where S ⊆ E(M).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
42
+ page_content=' Let the matrix be AS.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
43
+ page_content=' Then MS = M(AS) is the splitting matroid, and the operation is called the splitting operation using set S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
44
+ page_content=' The matroid BS need not be cographic or graphic for a cographic binary matroid B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
45
+ page_content=' Thus, the splitting operation does not protect matroid properties like graphicness, cographicness, etc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
46
+ page_content=' N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
47
+ page_content=' Pirouz [7] characterized a cographic matroid whose splitting using two elements is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
48
+ page_content=' In the following theorem, Ganesh et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
49
+ page_content=' [2] characterized graphic matroid whose splitting matroid, using three elements, is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
50
+ page_content=' 1 Graphic Elementary Lift of Cographic Matroids 2 Theorem 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
51
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
52
+ page_content=' [2] Let S ⊆ E(M), with |S| = 3, where M is a graphic binary matroid, then MS is graphic if and only if the matroid M do not have minors M(Fi), where the Figure 1, shows the graph Fi, for i = 1, 2 · · ·7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
53
+ page_content=' s s s s s s s s s s s s s F2 F3 F4 F5 s s s s s s ◗ ◗ ◗ F6 s s s s F1 F7 s s s s s s s s s s s Figure 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
54
+ page_content=' Excluded minors for the splitting of a graphic matroid using three elements.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
55
+ page_content=' Let Ck be the collection of cographic matroid whose splitting using k elements is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
56
+ page_content=' It is observed that there is no minimal minor E such that E /∈ C1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
57
+ page_content=' N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
58
+ page_content=' Pirouz [7] characterized the class C2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
59
+ page_content=' Theorem 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
60
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
61
+ page_content=' [7] Let C be a cographic binary matroid, then C ∈ C2 if and only if it does not have M(G1) or M(G2) minor, Figure 2 shows the graphs G1 and G2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
62
+ page_content=' t t t t t t t t t G1 G2 Figure 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
63
+ page_content=' Minimal minors not in the class C2 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
64
+ page_content=' This paper proves the following theorems.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
65
+ page_content=' Theorem 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
66
+ page_content='3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
67
+ page_content=' A cographic binary matroid M /∈ Ck, k ≥ 2, then M contains a minor P such that one of the below is true.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
68
+ page_content=' i) P is an extension of a minimal minor E such that E /∈ Ck−1 by single element.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
69
+ page_content=' ii) P = M(Qi).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
70
+ page_content=' iii) P is a coextension of M(Qi) by n elements, where n ≤ k, the Figure 3 shows the graph Qi, for i = 1, 2, · · ·9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
71
+ page_content=' Graphic Elementary Lift of Cographic Matroids 3 We show that the forbidden minors obtained by Mundhe et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
72
+ page_content=' [2] are the only minimal minors not in the class C3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
73
+ page_content=' Theorem 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
74
+ page_content='4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
75
+ page_content=' Let a cographic binary matroid be M, then M ∈ C3 if and only if M does not have a minor M(Fi), Figure 1 shows the graph Fi for i = 1, 2, · · ·7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
76
+ page_content=' 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
77
+ page_content=' Preliminary Results We denote F = {F ∗ 7 , M∗(K3,3), F7, M∗(K5)}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
78
+ page_content=' An elementary quotient of F ∈ F is denoted by QF.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
79
+ page_content=' Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
80
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
81
+ page_content=' [4] A binary matroid is a graphic matroid if and only if it does not has a minor F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
82
+ page_content=' Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
83
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
84
+ page_content=' [4] A binary matroid is a cographic matroid if and only if it does not has a minor from the set {F7, M(K5), F ∗ 7 , M(K3,3)}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
85
+ page_content=' In this paper, we use the technique discovered by Mundhe et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
86
+ page_content=' [2] to find the excluded minors.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
87
+ page_content=' The following lemmas are used to prove the main theorems.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
88
+ page_content=' Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
89
+ page_content='3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
90
+ page_content=' Let MS is not a graphic binary matroid for a cographic binary matroid M for S ⊆ E(M) and |S| = k, k ≥ 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
91
+ page_content=' Then there exists a minor P of M such as one of the below is true.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
92
+ page_content=' (i) PS ∼= F or PS/S′ ∼= F, for some S′ ⊆ S and F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
93
+ page_content=' (ii) P is an extension of a minimal minor E by an element, where E /∈ Ck−1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
94
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
95
+ page_content=' On a similar line of the proof of Lemma 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
96
+ page_content='3 in [2].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
97
+ page_content=' □ Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
98
+ page_content='4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
99
+ page_content=' Let P be the minor as stated in Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
100
+ page_content='3(i).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
101
+ page_content=' Then P does not contain a coloop.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
102
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
103
+ page_content=' On a similar line of the proof of Lemma 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
104
+ page_content='4 in [2].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
105
+ page_content=' □ Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
106
+ page_content='5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
107
+ page_content=' Let P be the minor as stated in Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
108
+ page_content='3(i), without containing a coloop.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
109
+ page_content=' Then P is a coextension of QF by n elements, where n ≤ k, k ≥ 2 or P ∼= QF for some binary matroid N with a ∈ E(N), such that N\\a ∼= F for F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
110
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
111
+ page_content=' On a similar line of the proof of Lemma 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
112
+ page_content='6 in [2].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
113
+ page_content=' □ From the definition of an elementary quotient and above lemma, N/a = QF, F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
114
+ page_content=' Thus, we need quotients of every F ∈ F to find excluded minors for the class Ck, for k ≥ 2, Mundhe et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
115
+ page_content=' [2] obtained graphic quotients for every F ∈ F as follows.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
116
+ page_content=' Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
117
+ page_content='6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
118
+ page_content=' [2] A graphic elementary quotient QF ∗ 7 ∼= M(Q1) or QF ∗ 7 ∼= M(Q2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
119
+ page_content=' The graphs Q1 and Q2 are shown in Figure 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
120
+ page_content=' Graphic Elementary Lift of Cographic Matroids 4 Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
121
+ page_content='7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
122
+ page_content=' [2] A graphic elementary quotient QF7 ∼= M(Q3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
123
+ page_content=' The graph Q3 is shown in Figure 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
124
+ page_content=' Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
125
+ page_content='8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
126
+ page_content=' [2] A graphic elementary quotient QM∗(K3,3) ∼= M(Q4) or QM∗(K3,3) ∼= M(Q5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
127
+ page_content=' The graphs Q4 and Q5 are shown in Figure 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
128
+ page_content=' Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
129
+ page_content='9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
130
+ page_content=' [2] A graphic elementary quotient QM∗(K5) ∼= M(Qi), the graph Qi is as given in Figure 3, for i = 6, 7, 8, 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
131
+ page_content=' t t t t t t t t t t t t t t t t t t t s s s s s s ◗ ◗ ◗ s s s s s s ◗ ◗ ◗ s s s s s s ◗ ◗ ◗ s s s s s s ◗ ◗ ◗ Q3 Q1 Q2 Q4 Q5 Q6 Q7 Q8 Q9 Figure 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
132
+ page_content=' Graphic quotients of Non-graphic matroids To find excluded minors for Ck for k ≥ 3, we need graphic and non-graphic quotients for every F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
133
+ page_content=' Mundhe found graphic quotients for every F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
134
+ page_content=' This paper finds non- graphic quotients for every F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
135
+ page_content=' Since either quotient Q or coextension of Q will be the minor of a cographic matroid, by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
136
+ page_content='2, Q should not contain F7 and F ∗ 7 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
137
+ page_content=' Thus, for every F ∈ F, we find non-graphic quotients not containing F7 and F ∗ 7 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
138
+ page_content=' We proved the following lemmas.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
139
+ page_content=' Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
140
+ page_content='10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
141
+ page_content=' A quotient QF7 not containing F7 and F ∗ 7 is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
142
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
143
+ page_content=' Let N\\a ∼= F7, where N is a binary matroid and a ∈ E(N), then QF7 = N/a, if a is a coloop or a loop, then QF7 = N\\a ∼= F7, thus QF7 ∼= F7, a contradiction.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
144
+ page_content=' If a is not a coloop or a loop, then r(N\\a) = r(F7) = 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
145
+ page_content=' Thus r(QF7) = 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
146
+ page_content=' Hence, QF7 can not have a minor F ∈ {M∗(K3,3), M∗(K5)}, as r(F) ≥ 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
147
+ page_content=' Thus, by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
148
+ page_content='1, QF7 is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
149
+ page_content=' □ Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
150
+ page_content='11.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
151
+ page_content=' A quotient QF ∗ 7 not containing F7 and F ∗ 7 is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
152
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
153
+ page_content=' Let N\\a ∼= F ∗ 7 , for a binary matroid N having an element a, then QF ∗ 7 = N/a, if a is a coloop or a loop, then QF ∗ 7 = N\\a ∼= F ∗ 7 , thus QF ∗ 7 ∼= F ∗ 7 , a contradiction.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
154
+ page_content=' If a is not a coloop or a loop, then r(N\\a) = r(F ∗ 7 ) = 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
155
+ page_content=' QF ∗ 7 can not contain F ∈ {M∗(K3,3), M∗(K5)} minor, as r(F) ≥ 4 and r(QF ∗ 7 ) = 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
156
+ page_content=' Thus, by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
157
+ page_content='1, QF ∗ 7 is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
158
+ page_content=' □ Graphic Elementary Lift of Cographic Matroids 5 Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
159
+ page_content='12.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
160
+ page_content=' Let a binary matroid be N having an element a such that a is not a loop or coloop, then QM∗(K5) not containing F7 or F ∗ 7 is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
161
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
162
+ page_content=' Let a binary matroid be N having an element a and a is not a coloop or loop, such that N\\a ∼= M∗(K5) then r(N\\a) = 6 and E(N\\a) = 10, then r(N) = 6 and E(N) = 11 then QM∗(K5) = N/a.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
163
+ page_content=' Thus, r(QM∗(K5)) = 5 and E(QM∗(K5)) = 10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
164
+ page_content=' Suppose QM∗(K5) is not graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
165
+ page_content=' Then by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
166
+ page_content='1, QM∗(K5) contains M∗(K3,3) or M∗(K5) minor.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
167
+ page_content=' QM∗(K5) does not contains M∗(K5) minor, as r(QM∗(K5)) = 5 and r(M∗(K5)) = 6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
168
+ page_content=' If QM∗(K5) contains M∗(K3,3) minor, then QM∗(K5)\\A1/A2 ∼= M∗(K3,3) for some subsets A1 and A2 of E(QM∗(K5)).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
169
+ page_content=' a) If A1 = ∅ and A2 = ∅ then QM∗(K5) ∼= M∗(K3,3), a contradiction, as r(QM∗(K5)) = 5 and r(M∗(K3,3)) = 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
170
+ page_content=' b) If A1 = ∅ and A2 ̸= ∅, then, if |A2| > 1, QM∗(K5)/A2 ∼= M∗(K3,3) a contradiction, as r(QM∗(K5)/A2) ≤ 3 and r(M∗(K3,3)) = 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
171
+ page_content=' If |A2| = 1, then QM∗(K5)/b ∼= M∗(K3,3) that is N/a/b ∼= M∗(K3,3) for some b ∈ E(QM∗(K5)).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
172
+ page_content=' Thus, (N/a/b)∗ ∼= N∗\\a\\b ∼= M(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
173
+ page_content=' Also, we have N\\a ∼= M∗(K5) thus N∗/a ∼= M(K5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
174
+ page_content=' M(K3,3) contains more than six odd cocircuits.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
175
+ page_content=' Hence, N∗ contains at least two odd cocircuits without containing a.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
176
+ page_content=' Therefore N∗/a contains at least one odd cocircuit, a contradiction as N∗/a ∼= M(K5) and M(K5) is Eulerian.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
177
+ page_content=' c) If A1 ̸= ∅ and A2 = ∅, then QM∗(K5)\\A1 ∼= M∗(K3,3), a contradiction, as r(QM∗(K5)\\A1) = 5 and r(M∗(K3,3)) = 4, when |A1| = 1 and E(QM∗(K5)\\A1) ≤ 8 when |A1| > 1, whereas E(M∗(K3,3)) = 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
178
+ page_content=' d) If A1 ̸= ∅ and A2 ̸= ∅ then QM∗(K5)\\A1/A2 ∼= M∗(K3,3), a contradiction as E(M∗(K3,3)) = 9 and E(QM∗(K5)\\A1/A2) ≤ 8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
179
+ page_content=' Thus, M∗(K3,3) is not a minor of QM∗(K5) and hence by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
180
+ page_content='1, we say that QM∗(K5) is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
181
+ page_content=' □ Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
182
+ page_content='13.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
183
+ page_content=' Let a ∈ E(N), where N is a binary matroid, such that a is not a loop or a coloop, then QM∗(K3,3) not containing F7 or F ∗ 7 is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
184
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
185
+ page_content=' Suppose a ∈ E(N), where N is a binary matroid, such that a is not a coloop or a loop such that N\\a ∼= M∗(K3,3) then E(N\\a) = 9 and r(N\\a) = 4 then r(N) = 4, E(N) = 10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
186
+ page_content=' We have QM∗(K3,3) = N/a, then r(QM∗(K3,3)) = 3, E(QM∗(K3,3)) = 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
187
+ page_content=' Suppose QM∗(K3,3) is not graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
188
+ page_content=' Then by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
189
+ page_content='1, QM∗(K3,3) has a minor M∗(K5) or M∗(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
190
+ page_content=' Case(i) If QM∗(K3,3) contains M∗(K5), then QM∗(K3,3)\\A1/A2 ∼= M∗(K5) for some subsets A1 or A2 of E(QM∗(K3,3)), which is a contradiction, as r(QM∗(K3,3)\\A1/A2) ≤ 3 however r(M∗(K5)) = 6 Case(ii) If QM∗(K3,3) has a minor M∗(K3,3), then QM∗(K3,3)\\A1/A2 ∼= M∗(K3,3) for some subsets A1 or A2 of E(QM∗(K3,3)), which is a contradiction, as r(QM∗(K3,3)\\A1/A2) ≤ 3 however r(M∗(K3,3)) = 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
191
+ page_content=' Thus from the case(i) and case(ii) and by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
192
+ page_content='1, we say that QM∗(K3,3) is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
193
+ page_content=' □ Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
194
+ page_content='14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
195
+ page_content=' M∗(K5) is the non-graphic quotient QM∗(K5), not containing F7 and F ∗ 7 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
196
+ page_content=' Graphic Elementary Lift of Cographic Matroids 6 Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
197
+ page_content=' Let a ∈ E(N), where N is a binary matroid such that N\\a ∼= M∗(K5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
198
+ page_content=' (i) If a is a coloop or a loop.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
199
+ page_content=' Then N/a ∼= N\\a ∼= M∗(K5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
200
+ page_content=' (ii) If a is not a coloop or loop then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
201
+ page_content='12, N/a is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
202
+ page_content=' Thus from above M∗(K5) is the only non-graphic elementary quotient of M∗(K5) not con- taining F7 and F ∗ 7 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
203
+ page_content=' □ Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
204
+ page_content='15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
205
+ page_content=' M∗(K3,3) is the non-graphic quotient QM∗(K3,3), not containing F7 and F ∗ 7 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
206
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
207
+ page_content=' Let a ∈ E(N), where N is a binary matroid such that N\\a ∼= M∗(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
208
+ page_content=' (i) If a is a coloop or a loop, then N\\a ∼= N/a ∼= M∗(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
209
+ page_content=' (ii) If a is not a coloop or loop then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
210
+ page_content='13, N/a is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
211
+ page_content=' Thus from above we say that, M∗(K3,3) is the only non-graphic elementary quotient of M∗(K3,3) not containing F7 and F ∗ 7 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
212
+ page_content=' □ 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
213
+ page_content=' Main Theorems In the previous section, we mentioned the graphic and non-graphic quotients for every F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
214
+ page_content=' Now, the main theorems are proved in this section.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
215
+ page_content=' Theorem 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
216
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
217
+ page_content=' A cographic binary matroid M /∈ Ck, k ≥ 2, then M contains a minor P such that one of the below is true.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
218
+ page_content=' i) P is an extension of a minimal minor E by single element, such that E /∈ Ck−1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
219
+ page_content=' ii) P = M(Qi).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
220
+ page_content=' iii) P is a coextension of M(Qi) by n elements, where n ≤ k, the Figure 3 shows the graph Qi, for i = 1, 2, · · ·9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
221
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
222
+ page_content=' Let a binary cographic matroid be M such that M /∈ Ck, k ≥ 2, that is for S ⊆ E(M), with |S| = k, MS is non-graphic matroid.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
223
+ page_content=' From Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
224
+ page_content='3, M has a minor P with S ⊆ E(P), such that PS ∼= F or PS/S′ ∼= F, for some S′ ⊆ S and F ∈ F or P is an extension of some a minimal minor E by single element, such that E /∈ Ck−1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
225
+ page_content=' If PS ∼= F or PS/S′ ∼= F, for some S′ ⊆ S then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
226
+ page_content='5, either P ∼= QF or P is extension of QF by n elements, where n ≤ k and QF = N/a is a quotient of F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
227
+ page_content=' Case (i) If the quotient is graphic, then a) If F = F ∗ 7 , then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
228
+ page_content='6, QF ∗ 7 ∼= M(Q1) or QF ∗ 7 ∼= M(Q2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
229
+ page_content=' b) If F = F7, then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
230
+ page_content='7, QF7 ∼= M(Q3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
231
+ page_content=' If F = M∗(K3,3), then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
232
+ page_content='8, QM∗(K3,3) ∼= M(Q4) or QM∗(K3,3) ∼= M(Q5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
233
+ page_content=' If F = M∗(K5) then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
234
+ page_content='9, QM∗(K5) ∼= M(Qi).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
235
+ page_content=' Figure 3 shows the graph Qi, for i = 1, 2, · · ·9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
236
+ page_content=' Case (ii) If the quotient is not graphic, then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
237
+ page_content='14, a non-graphic quotient QM∗(K5) = M∗(K5) and by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
238
+ page_content='15, a non-graphic quotient QM∗(K3,3) = M∗(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
239
+ page_content=' From Figure 4, M∗(Q1) is a minor of the matroid M(K5), thus M(Q1) is a minor of the Graphic Elementary Lift of Cographic Matroids 7 matroid M∗(K5) and From the Figure 5, M∗(Q2) is a minor of the matroid M(K3,3), thus M(Q2) is a minor of the matroid M∗(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
240
+ page_content=' Hence we discard non-graphic quotients.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
241
+ page_content=' s s s s F1 t t t t t F ∗ 1 Figure 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
242
+ page_content=' The Graphs F1 ∼= Q1 and F ∗ 1 ∼= Q∗ 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
243
+ page_content=' t t t t t ✟✟✟✟✟✟✟✟ s s s s F2 F ∗ 2 Figure 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
244
+ page_content=' The Graphs F2 ∼= Q2 and F ∗ 2 ∼= Q∗ 2 Thus from above, either the minor P = QF or a coextension of QF not more than k elements for F ∈ F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
245
+ page_content=' Hence the result.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
246
+ page_content=' □ We now obtain excluded minors for the class C3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
247
+ page_content=' Theorem 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
248
+ page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
249
+ page_content=' Let a cographic binary matroid be M, then M ∈ C3 if and only if M does not have a minor M(Fi), where the Figure 1 shows the graph Fi, for i = 1, 2, · · ·7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
250
+ page_content=' Proof.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
251
+ page_content=' Suppose a cographic matroid M contains minor M(Fi), for i = 1, 2, · · ·7, then M /∈ C3, the proof is straight forward.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
252
+ page_content=' Conversely, if M does not contain a minor M(Fi) for i = 1, 2, · · ·7, then we will prove that M ∈ C3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
253
+ page_content=' Suppose not, then for some S ⊆ E(M), with |S| = 3, MS is not a graphic matroid, then, MS contains minor F, for some F ∈ F, by Theorem 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
254
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
255
+ page_content=' Then M contains a minor P containing S, By Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
256
+ page_content='3, such that PS ∼= F or PS/S′ ∼= F, for some S′ ⊆ S or P is an extension of circuit matroid of the graph G1 or G2 by single element and the graphs G1, G2 are given in Figure 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
257
+ page_content=' It is observed that an extension of M(G1) by a single element, either isomorphic to M(F4) or M(F7) or contains minor M(F1) or M(F2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
258
+ page_content=' Also, M(G2) contains minor M(F2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
259
+ page_content=' Thus P cannot be an extension of M(G1) or M(G2) by a single element.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
260
+ page_content=' Hence, PS ∼= F or PS/S′ ∼= F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
261
+ page_content=' Then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
262
+ page_content='5, either P is an extension of QF by n elements, where n ≤ 3 or P ∼= QF.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
263
+ page_content=' Case (i) If the quotient is graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
264
+ page_content=' In [2], Mundhe et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
265
+ page_content=' obtained forbidden minors from graphic quotients of every F ∈ F, as given in Theorem 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
266
+ page_content='1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
267
+ page_content=' Case (ii) If the quotient is not graphic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
268
+ page_content=' Let F = M∗(K5), then by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
269
+ page_content='14, QM∗(K5) ∼= M∗(K5) but from Figure 4, Graphic Elementary Lift of Cographic Matroids 8 M∗(F1) is a minor of M(K5) and hence M(F1) is a minor of M∗(K5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
270
+ page_content=' Hence we discard M∗(K5).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
271
+ page_content=' Let F = M∗(K3,3), then QM∗(K3,3) ∼= M∗(K3,3), by Lemma 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
272
+ page_content='15, but from Figure 5, M∗(F2) is a minor of M(K3,3) and hence M(F2) is a minor of M∗(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
273
+ page_content=' Hence, we discard M∗(K3,3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
274
+ page_content=' Thus by the case (i) and case (ii), the excluded minor for the class C3 is the matroid M(Fi), the graph Fi is shown in Figure 1, for i = 1, 2, · · ·7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
275
+ page_content=' □ References 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
276
+ page_content=' F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
277
+ page_content=' Harary, Graph Theory, Narosa Publishing House, New Delhi , 1988.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
278
+ page_content=' 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
279
+ page_content=' G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
280
+ page_content=' Mundhe, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
281
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
282
+ page_content=' Borse, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
283
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
284
+ page_content=' Dalvi, On graphic elementary lifts of graphic matroids, Discrete Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
285
+ page_content=', 345, (2022) 113014.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
286
+ page_content=' 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
287
+ page_content=' H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
288
+ page_content=' Fleischner, Eulerian Graphs and Related Topics Part 1, Vol.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
289
+ page_content=' 1, North Holland, Amsterdam , 1990.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
290
+ page_content=' 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
291
+ page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
292
+ page_content=' G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
293
+ page_content=' Oxley, Matroid Theory, Second Edition, Oxford University Press, Oxford, 2011.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
294
+ page_content=' 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
295
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
296
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
297
+ page_content=' Shikare and B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
298
+ page_content=' N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
299
+ page_content=' Waphare, Excluded-Minors for the class of graphic splitting matroids, Ars Combin.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
300
+ page_content=' 97 (2010), 111-127.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
301
+ page_content=' 6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
302
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
303
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
304
+ page_content=' Shikare, Gh.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
305
+ page_content=' Azadi, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
306
+ page_content=' N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
307
+ page_content=' Waphare, Generalized splitting operation and its application, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
308
+ page_content=' Indian Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
309
+ page_content=' Soc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
310
+ page_content=' , 78, (2011), 145-154.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
311
+ page_content=' 7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
312
+ page_content=' Pirouz N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
313
+ page_content=', Graphic splitting of cographic matroids, Discussiones Mathematicae Graph Theory 35 (2015) 95–104.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
314
+ page_content=' 8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
315
+ page_content=' T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
316
+ page_content=' T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
317
+ page_content=' Raghunathan, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
318
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
319
+ page_content=' Shikare and B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
320
+ page_content=' N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
321
+ page_content=' Waphare, Splitting in a binary matroid, Discrete Math.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
322
+ page_content=' 184 (1998), 267-271.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
323
+ page_content=' 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
324
+ page_content=' Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
325
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
326
+ page_content=' Borse, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
327
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
328
+ page_content=' Shikare and K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
329
+ page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
330
+ page_content=' Dalvi, Excluded-minors for the class of cographic splitting matroids, Ars Combin.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
331
+ page_content=' 115 (2014), 219-237.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
332
+ page_content=' 10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
333
+ page_content=' Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
334
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
335
+ page_content=' Borse, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
336
+ page_content=' M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
337
+ page_content=' Shikare and Pirouz Naiyer, A characterization of graphic matroids which yield cographic splitting matroids,Ars Combin.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
338
+ page_content=' 118 (2015), 357-366.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/0tAzT4oBgHgl3EQf8v4C/content/2301.01907v1.pdf'}
19E2T4oBgHgl3EQfiwfE/content/2301.03962v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e94b2ace3069d3ac12aaba39ce6acfd8fc08b3389dbdab22c04629eb03bbb830
3
+ size 1901557
1dFKT4oBgHgl3EQfOy1w/content/2301.11760v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b4de8b162f04ae7cfd831973f9020f1a7fb93948dac387504f94a2e4b3d3376a
3
+ size 3291621
1dFKT4oBgHgl3EQfOy1w/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:50267814b7b091af0123db78d700e34fdae02faf498f9f61bd1296ce9947d55c
3
+ size 707946
3NE1T4oBgHgl3EQflwRz/content/2301.03289v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:170343df51bc302a58505a552413e886afad2460b3f5f105c85105b8a91732fa
3
+ size 9095799
3dFQT4oBgHgl3EQf3TZi/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd62d30957ff0890ab98ab90760b3c1e1c14dd4298193f3ea2e0197ce7ec6098
3
+ size 7209005
3tE2T4oBgHgl3EQf6AiT/content/2301.04196v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6b7073012ab242ffe1b22c3a6829b3308fa0d631ba750339d2a854565fe89d8e
3
+ size 513494
3tE2T4oBgHgl3EQf6AiT/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2bda3a74e0f19276f3ac011adc3400cb35d89dfe4b56db78342ad9bcebe50e1e
3
+ size 85572
3tFKT4oBgHgl3EQfRC0N/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd0049a68b29875c37d477b7aa5d4a0da8379d91e888cf7adc6550b6b5d04041
3
+ size 5963821
49E3T4oBgHgl3EQfpApV/content/tmp_files/2301.04638v1.pdf.txt ADDED
@@ -0,0 +1,1707 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ This draft was prepared using the LaTeX style file belonging to the Journal of Fluid Mechanics
2
+ 1
3
+ Dynamics of a data-driven low-dimensional model
4
+ of turbulent minimal Couette flow
5
+ Alec J. Linot1 and Michael D. Graham1†
6
+ 1Department of Chemical and Biological Engineering, University of Wisconsin-Madison, Madison WI
7
+ 53706, USA
8
+ Because the Navier-Stokes equations are dissipative, the long-time dynamics of a flow in state
9
+ space are expected to collapse onto a manifold whose dimension may be much lower than the
10
+ dimension required for a resolved simulation. On this manifold, the state of the system can be
11
+ exactly described in a coordinate system parameterizing the manifold. Describing the system
12
+ in this low-dimensional coordinate system allows for much faster simulations and analysis. We
13
+ show, for turbulent Couette flow, that this description of the dynamics is possible using a data-
14
+ driven manifold dynamics modeling method. This approach consists of an autoencoder to find a
15
+ low-dimensional manifold coordinate system and a set of ordinary differential equations defined
16
+ by a neural network. Specifically, we apply this method to minimal flow unit turbulent plane
17
+ Couette flow at Re = 400, where a fully resolved solutions requires O(105) degrees of freedom.
18
+ Using only data from this simulation we build models with fewer than 20 degrees of freedom
19
+ that quantitatively capture key characteristics of the flow, including the streak breakdown and
20
+ regeneration cycle. At short-times, the models track the true trajectory for multiple Lyapunov
21
+ times, and, at long-times, the models capture the Reynolds stress and the energy balance. For
22
+ comparison, we show that the models outperform POD-Galerkin models with ∼2000 degrees of
23
+ freedom. Finally, we compute unstable periodic orbits from the models. Many of these closely
24
+ resemble previously computed orbits for the full system; additionally, we find nine orbits that
25
+ correspond to previously unknown solutions in the full system.
26
+ 1. Introduction
27
+ A major challenge in dealing with chaotic fluid flows, whether it be performing experiments,
28
+ running simulations, or interpreting the results, is the high-dimensional nature of the state. Even
29
+ for simulations in the smallest domains that sustain turbulence (a minimal flow unit (MFU)),
30
+ the state dimension may be O(105) (Jiménez & Moin 1991; Hamilton et al. 1995). However,
31
+ despite this nominal high-dimensionality, the dissipative nature of turbulent flows leads to the
32
+ expectation that long-time dynamics collapse onto an invariant manifold of much lower dimension
33
+ than the ambient dimension (Hopf 1948). By modeling the dynamics in a manifold coordinate
34
+ system, simulations could be performed with a drastically lower-dimensional state representation,
35
+ significantly speeding up computations. Additionally, such a low-dimensional state representation
36
+ is highly useful for downstream tasks like control or design. Finding a low-dimensional – or
37
+ ideally a minimal dimensional – parameterization of the manifold and an evolution equation for
38
+ this parameterization are both challenges. In this work we aim to address these challenges with a
39
+ data-driven model, specifically for the task of reconstructing turbulent plane Couette flow.
40
+ The classic way to perform dimension reduction from data is to use the proper orthogonal
41
+ decomposition (POD), also known by Principal Component Analysis (PCA) or Karhunen-Loève
42
+ decomposition (Holmes et al. 2012). This is a linear dimension reduction technique in which the
43
+ state is projected onto the set of orthogonal modes that capture the maximum variance in the
44
+ † Email address for correspondence: mdgraham@wisc.edu
45
+ arXiv:2301.04638v1 [physics.flu-dyn] 11 Jan 2023
46
+
47
+ 2
48
+ data. The POD is widely used for flow phenomena, some examples of which include: turbulent
49
+ channel flow (Moin & Moser 1989; Ball et al. 1991), flat-plate boundary layers (Rempfer & Fasel
50
+ 1994), and free shear jet flows (Arndt et al. 1997). Smith et al. (2005) showed how to incorporate
51
+ system symmetries into the POD modes, the details of which we elaborate on in Sec. 3.
52
+ While the POD has seen wide use and is easy to interpret, more accurate reconstruction can
53
+ be achieved with nonlinear methods – a result we highlight in Sec. 3. Some popular methods
54
+ for nonlinear dimension reduction include kernel PCA (Schölkopf et al. 1998), diffusion maps
55
+ (Coifman et al. 2005), local linear embedding (LLE) (Roweis & Saul 2000), isometric feature
56
+ mapping (Isomap) (Tenenbaum et al. 2000), and t-distributed stochastic neighbor embedding
57
+ (tSNE) (Hinton & Roweis 2003). These methods are described in more detail in Linot & Graham
58
+ (2022), and an overview of other dimension reduction methods can be found in Van Der Maaten
59
+ et al. (2009). One drawback of all of these methods, however, is that they reduce the dimension,
60
+ but do not immediately provide a means to move from a low-dimensional state back to the full
61
+ state. A popular dimension reduction method without these complications is the undercomplete
62
+ autoencoder (Hinton & Salakhutdinov 2006), which uses a neural network (NN) to map the input
63
+ data into a lower-dimensional “latent space" and another NN to map back to the original state
64
+ space. We describe this structure in more detail in Sec. 2. Some examples where autoencoders
65
+ have been used for flow systems include flow around a cylinder (Murata et al. 2020), flow around
66
+ a flat plate (Nair & Goza 2020), Kolmogorov flow (Page et al. 2021; Pérez De Jesús & Graham
67
+ 2022), and channel flow (Milano & Koumoutsakos 2002). Although we will not pursue this
68
+ approach in the present work, it may be advantageous for multiple reasons to parametrize the
69
+ manifold with overlapping local representations, as done in Floryan & Graham (2022).
70
+ After reducing the dimension, the time evolution for the dynamics can be approximated from
71
+ the equations of motion or in a completely data-driven manner. The classical method is to perform
72
+ a Galerkin projection wherein the equations of motion are projected onto a set of modes (e.g. POD
73
+ modes) (Holmes et al. 2012). However, in this approach all the higher POD modes are neglected.
74
+ An extension of this idea, called nonlinear Galerkin, is to assume that the time derivative of the
75
+ coefficients of all of the higher modes is zero, but not the coefficients themselves (Titi 1990;
76
+ Foias et al. 1988; Graham et al. 1993); this is essentially a quasisteady state approximation for
77
+ the higher modes. This improves the accuracy, but comes at a higher computational cost than the
78
+ Galerkin method, although this can be somewhat mitigated by using a postprocessing Galerkin
79
+ approach (García-Archilla et al. 1998). Wan et al. (2018) also showed a recurrent NN (RNN) –
80
+ a NN that feeds into itself – can be used to improve the nonlinear Galerkin approximation. This
81
+ RNN structure depends on a history of inputs, making it non-Markovian. In addition to these
82
+ linear dimension reduction approaches, an autoencoder can be used with the equations of motion
83
+ in the so-called manifold Galerkin approach, which Lee & Carlberg (2020) developed and applied
84
+ to the viscous Burgers equation .
85
+ When the equations of motion are assumed to be unknown, and only snapshots of data are
86
+ available, a number of different machine learning techniques exist to approximate the dynamics.
87
+ Two of the most popular techniques are RNNs and reservoir computers. Vlachas et al. (2020)
88
+ showed both these structures do an excellent job of capturing the chaotic dynamics of the Lorenz-
89
+ 96 equation and Kuramoto-Sivashinsky equation (KSE). For fluid flows, autoencoders and RNNs
90
+ (specifically long-short term memory networks (LSTM)) have been used to model flow around
91
+ a cylinders (Hasegawa et al. 2020a; Eivazi et al. 2020), pitching airfoils (Eivazi et al. 2020),
92
+ bluff bodies (Hasegawa et al. 2020b), and MFU plane Poiseuille flow (PPF) (Nakamura et al.
93
+ 2021). Although these methods often do an excellent job of predicting chaotic dynamics, the
94
+ models are not Markovian, so the dimension of the system also includes some history, and these
95
+ models perform discrete timesteps. These two properties are undesirable, because the underlying
96
+ dynamics are Markovian and continuous in time, and modeling them differently complicates
97
+ applications and interpretations of the model. In particular, we want to use the model for state
98
+
99
+ 3
100
+ space analyses such as determination of periodic orbits, where standard tools are available for
101
+ ODEs that do not easily generalize to non-Markovian dynamic models.
102
+ Due to these issues, we use neural ordinary differential equations (ODE) (Chen et al. 2019).
103
+ In neural ODEs, the right-hand-side (RHS) of an ODE is represented as a NN that is trained to
104
+ reconstruct the time evolution of the data from snapshots of training data. In Linot & Graham
105
+ (2022) it was shown that this is an effective method for modeling the chaotic dynamics of the KSE.
106
+ Rojas et al. (2021) used neural ODEs to predict the periodic dynamics of flow around a cylinder,
107
+ and Portwood et al. (2019) used neural ODEs to predict the kinetic energy and dissipation of
108
+ decaying turbulence.
109
+ In this work we investigate the dynamics of MFU Couette flow. The idea behind the MFU, first
110
+ introduced by Jiménez & Moin (1991), is to reduce the simulation domain to the smallest size
111
+ that sustains turbulence, thus isolating the key components of the turbulent nonlinear dynamics.
112
+ Using an MFU for Couette flow at transitional Reynolds number, Hamilton et al. (1995) outlined
113
+ the regeneration cycle of wall bounded turbulence called the “self-sustaining process" (SSP),
114
+ which we describe in more detail in Sec. 3. This system was later analyzed with coviariant
115
+ Lyapunov analysis by Inubushi et al. (2015), who found a Lyapunov time (the inverse of the
116
+ leading Lyapunov exponent) of ∼ 48 time units.
117
+ Many low-dimensional models have been developed to recreate the dynamics of the SSP. The
118
+ first investigation of this topic was by Waleffe (1997), who developed an 8 mode model for shear
119
+ flow between free-slip walls generated by a spatially sinusoidal forcing. He selected the modes
120
+ based on intuition from the SSP and performed a Galerkin projection onto these modes. Moehlis
121
+ et al. (2004) later added an additional mode to Waleffe’s model which enables modification of
122
+ the mean profile by the turbulence, and made some modifications to the chosen modes. In this
123
+ “MFE" model, Moehlis et al. found exact coherent states, which we discuss below, that did not
124
+ exist in the 8 mode model. In addition, Moehlis et al. (2002) also used the POD modes on a
125
+ domain slightly larger than the MFU to generate POD-Galerkin models. These low-dimensional
126
+ models have been used as a starting point for testing data-driven models. For example, both
127
+ LSTMs (Srinivasan et al. 2019) and a Koopman operator method with nonlinear forcing (Eivazi
128
+ et al. 2021) have been used to attempt to reconstruct the MFE model dynamics. Borrelli et al.
129
+ (2022) then applied these methods to PPF.
130
+ Finally, we note that a key approach to understanding complex nonlinear dynamical phenomena,
131
+ such as the SSP of near-wall turbulence, is through study of the underlying state space structure
132
+ of fixed points and periodic orbits. In the turbulence literature these are sometimes called “exact
133
+ coherent states", or ECS (Kawahara et al. 2012; Graham & Floryan 2021). Turbulence organizes
134
+ around ECS in the sense that trajectories chaotically move between different such states. The
135
+ first ECS found were fixed point solutions in PCF (Nagata 1990). Following this work, Waleffe
136
+ (1998) was able to connect ECS of PCF and PPF to the SSP. Later, more fixed point ECS were
137
+ found in MFU PCF and visualized by Gibson et al. (2008a). Unlike fixed points, which cannot
138
+ capture dynamic phenomena at all, periodic orbits are able to represent key aspects of turbulent
139
+ dynamics such as bursting behavior. Kawahara & Kida (2001) found the first two periodic orbits
140
+ (POs) for MFU PCF, one of which had statistics that agreed well with the SSP. Then, Viswanath
141
+ (2007) found another PO and 4 new relative POs (RPOs) in this domain, and Gibson made these
142
+ solutions available in (Gibson et al. 2008b), along with a handful of others.
143
+ In the present work, we use autoencoders and neural ODEs , in a method we call “Data-driven
144
+ Manifold Dynamics" (DManD) (Linot et al. 2023), to build a ROM for turbulent MFU PCF
145
+ (Hamilton et al. 1995). Section 2 outlines the details of the DManD framework. We then describe
146
+ the details of the Couette flow in Sec. 3.1, the results of the dimension reduction in Sec. 3.2, and
147
+ the DManD model’s reconstruction of short- and long-time statistics in Sec. 3.3 and Sec. 3.4,
148
+ respectively. After showing that the models accurately reproduce these statistics, we compute
149
+
150
+ 4
151
+ RPOs for the model in Sec. 3.5, finding several that are similar to previously known RPOs, as
152
+ well as several that seem to be new. Finally, we summarize the results in Sec. 4.
153
+ 2. Framework
154
+ Here we outline our method for an “exact" DManD modeling approach. In this sense “exact"
155
+ means all of the functions described allow for perfect reconstruction, but error is introduced in
156
+ approximating these functions due to insufficient data, error in learning the functions, or error in
157
+ evolving them forward in time. This differs from coarse-grained ROMs, which approximate the
158
+ physics to generate a closed set of equations. A key component allowing DManD to be “exact"
159
+ is that we only seek to discover the evolution of trajectories after they collapse onto an invariant
160
+ manifold M.
161
+ In general, the training data for development of a DManD model comes in the form of snapshots
162
+ {𝑢1, 𝑢2, ..., 𝑢𝑀 }, which are either the full state or measurements diffeomorphic to the full state
163
+ (e.g. time delays (Takens 1981; Young & Graham 2022)). Here we consider full-state data 𝑢 that
164
+ lives in an ambient space R𝑑. We generate a time series of data by evolving this state forward in
165
+ time according to
166
+ 𝑑𝑢
167
+ 𝑑𝑡 = 𝑓 (𝑢).
168
+ (2.1)
169
+ (In the present context, this equation represents a fully-resolved direct numerical simulation
170
+ (DNS).) With the full state, we can then define a mapping to a low-dimensional state representation
171
+ ℎ = 𝜒(𝑢),
172
+ (2.2)
173
+ with ℎ ∈ R𝑑ℎ is a coordinate representation on the manifold. Finally, we define a mapping back
174
+ to the full state
175
+ ˜𝑢 = ˇ𝜒(ℎ).
176
+ (2.3)
177
+ For data that lies on a finite-dimensional invariant manifold these functions can exactly reconstruct
178
+ the state (i.e. ˜𝑢 = 𝑢). However, if the dimension 𝑑ℎ is too low, or there are errors in the
179
+ approximation of these functions, then ˜𝑢 approximates the state. Then, with this low-dimensional
180
+ state representation, we can define an evolution equation
181
+ 𝑑ℎ
182
+ 𝑑𝑡 = 𝑔(ℎ).
183
+ (2.4)
184
+ The DManD model consists of the three functions 𝜒, ˇ𝜒, and 𝑔. By approximating these functions,
185
+ the evolution of trajectories on the manifold can be performed entirely in the manifold coordinates
186
+ ℎ, which requires far fewer operations than a full simulation, as 𝑑ℎ ≪ 𝑑. We choose to approximate
187
+ all of these functions using NNs, but other representations could be used.
188
+ First, we train 𝜒 and ˇ𝜒 using an undercomplete autoencoder. This is a NN structure consisting
189
+ of an encoder which reduces dimension (𝜒) and a decoder that expands dimension ( ˇ𝜒). As
190
+ mentioned in Sec. 1, a common approach to dimension reduction is to project onto a set of POD
191
+ modes. POD gives the optimal linear projection in terms of reconstruction error, so we use this
192
+ fact to train an encoder as the sum of POD and a correction in the form of an NN:
193
+ ℎ = 𝜒(𝑢; 𝜃𝐸) = 𝑈𝑇
194
+ 𝑑ℎ𝑢 + E(𝑈𝑇
195
+ 𝑟 𝑢; 𝜃𝐸).
196
+ (2.5)
197
+ In this equation, 𝑈𝑘 ∈ R𝑑×𝑘 is a matrix whose 𝑘 columns are the first 𝑘 POD modes as ordered by
198
+ variance, and E is a NN. The first term (𝑈𝑇
199
+ 𝑑ℎ𝑢) is the projection onto the leading 𝑑ℎ POD modes,
200
+ and the second term is the NN correction. The matrix 𝑈𝑟 in this term may either be a full change
201
+ of basis with no approximation (𝑟 = 𝑑), or involve some dimension reduction (𝑑 > 𝑟 > 𝑑ℎ).
202
+
203
+ 5
204
+ For mapping back to the full state (decoding), we again sum POD with a correction
205
+ ˜𝑢 = ˇ𝜒(ℎ; 𝜃𝐸) = 𝑈𝑟 ([ℎ, 0]𝑇 + D(ℎ; 𝜃𝐷)).
206
+ (2.6)
207
+ Here, [ℎ, 0]𝑇 is the ℎ vector zero padded to the correct size, and D is a NN. The first term is
208
+ the POD mapping back to the full space, if there were no NNs, and the second term is a NN
209
+ correction. In Linot & Graham (2020) we refer to this structure as a hybrid autoencoder. In Sec.
210
+ 3.2 we contrast this to a “standard" autoencoder where ℎ = E(𝑈𝑇
211
+ 𝑟 𝑢; 𝜃𝐸) and ˜𝑢 = 𝑈𝑟D(ℎ; 𝜃𝐷).
212
+ These hybrid autoencoder operations act as a shortcut connection on the optimal linear dimension
213
+ reduction, which we (Linot & Graham 2020) found useful for representing the data and achieving
214
+ accurate reconstruction of 𝑢. Yu et al. (2021) also took a similar approach with POD shortcut
215
+ connections over each layer of the network.
216
+ We determine the NN parameters 𝜃𝐸 and 𝜃𝐷 by minimizing
217
+ 𝐿 = 1
218
+ 𝑑𝐾
219
+ 𝐾
220
+ ∑︁
221
+ 𝑖=1
222
+ ||𝑢(𝑡𝑖) − ˇ𝜒(𝜒(𝑢(𝑡𝑖); 𝜃𝐸); 𝜃𝐷)||2 +
223
+ 1
224
+ 𝑑ℎ𝐾
225
+ 𝐾
226
+ ∑︁
227
+ 𝑖=1
228
+ 𝛼||E(𝑈𝑇
229
+ 𝑟 𝑢(𝑡𝑖); 𝜃𝐸) + D𝑑ℎ (ℎ(𝑡𝑖); 𝜃𝐷)||2.
230
+ (2.7)
231
+ The first term in this loss is the mean-squared error (MSE) of the reconstruction ˜𝑢, and the second
232
+ term is a penalty that promotes accurate representation of the leading 𝑑ℎ POD coefficients. In this
233
+ term, D𝑑ℎ denotes the leading 𝑑ℎ elements of the decoder output. For finite 𝛼, the autoencoder
234
+ exactly matches the first 𝑑ℎ POD coefficients when this term vanishes. Details of the minimization
235
+ procedure are discussed in Sec. 3.
236
+ Next, we approximate 𝑔 using a neural ODE. A drawback of training a single dense NN for 𝑔 is
237
+ that the resulting dynamics may become weakly unstable, with linear growth at long times (Linot
238
+ & Graham 2022; Linot et al. 2023). To avoid this, we use a “stabilized" neural ODE approach by
239
+ adding a linear damping term onto the output of the NN, giving
240
+ 𝑔(ℎ(𝑡𝑖); 𝜃𝑔) = 𝑔NN(ℎ(𝑡𝑖); 𝜃𝑔) + 𝐴ℎ.
241
+ (2.8)
242
+ Integrating Eq. 2.8 forward from time 𝑡𝑖 to 𝑡𝑖 + 𝜏 yields
243
+ ˜ℎ(𝑡𝑖 + 𝜏) = ℎ(𝑡𝑖) +
244
+ ∫ 𝑡𝑖+𝜏
245
+ 𝑡𝑖
246
+ 𝑔NN(ℎ(𝑡); 𝜃𝑔) + 𝐴ℎ(𝑡)𝑑𝑡.
247
+ (2.9)
248
+ Depending on the situation, one may either learn 𝐴 from data, or fix it. Here we set it to the
249
+ diagonal matrix
250
+ 𝐴𝑖 𝑗 = −𝛽𝛿𝑖 𝑗𝜎𝑖(ℎ)
251
+ (2.10)
252
+ where 𝜎𝑖(ℎ) is the standard deviation of the 𝑖th component of ℎ, 𝛽 is a tunable parameter, and 𝛿𝑖 𝑗
253
+ is the Kronecker delta. This linear term attracts trajectories back to the origin, preventing them
254
+ from moving far away from the training data. In Sec. 3.4 we show that this approach drastically
255
+ improves the long-time performance of these models.
256
+ We then determine the parameters 𝜃𝑔 by minimizing the difference between the predicted state
257
+ ˜ℎ(𝑡𝑖 + 𝜏) and the true state ℎ(𝑡𝑖 + 𝜏), averaged over the data:
258
+ 𝐽 =
259
+ 1
260
+ 𝑑ℎ𝐾
261
+ 𝐾
262
+ ∑︁
263
+ 𝑖=1
264
+
265
+ ||ℎ(𝑡𝑖 + 𝜏) − ˜ℎ(𝑡𝑖 + 𝜏)||2
266
+ 2
267
+
268
+ .
269
+ (2.11)
270
+ For clarity we show the specific loss we use, which sums over only a single snapshot forward
271
+ in time at a fixed 𝜏. More generally, the loss can be formulated for arbitrary snapshot spacing
272
+ and for multiple snapshots forward in time. To compute the gradient of 𝐽 with respect to the
273
+ neural network parameters 𝜃𝑔, automatic differentiation can be used to backpropagate through
274
+ the ODE solver that is used to compute the time integral in Eq. 2.9, or an adjoint problem can
275
+ be solved backwards in time (Chen et al. 2019). The adjoint method uses less memory than
276
+
277
+ 6
278
+ backpropagation, but ℎ is low-dimensional and our prediction window for training is short, so we
279
+ choose to backpropagate through the solver.
280
+ So far this approach to approximating 𝜒, ˇ𝜒, and 𝑔 is general and does not directly account
281
+ for the fact that the underlying equations are often invariant to certain symmetry operations. For
282
+ example, one of the symmetries in PCF is a continuous translation symmetry in 𝑥 and 𝑧 (i.e. any
283
+ solution shifted to another location in the domain gives another solution). This poses an issue for
284
+ training, because in principle, the training data must include all these translations to accurately
285
+ model the dynamics under any translation. We discuss these and other symmetries of PCF in Sec.
286
+ 3.1.
287
+ In practice, accounting for continuous symmetries is most important along directions that
288
+ sample different phases very slowly. For PCF, the mean flow is in the 𝑥 direction, leading to
289
+ good phase sampling along this direction. However, there is no mean flow in the 𝑧 direction, so
290
+ sampling all phases relies on the slow phase diffusion in that direction. Therefore, we will only
291
+ explicitly to account for the 𝑧-phase in Sec. 3, but in the current disucssion we present the general
292
+ framework accounting for all continuous symmetries.
293
+ To address the issue of continuous translations, we add an additional preprocessing step to the
294
+ data, using the method of slices (Budanur et al. 2015b,a) to split the state 𝑢 into a pattern 𝑢 𝑝 ∈ R𝑑
295
+ and a phase 𝜙 ∈ R𝑐. The number of continuous translation symmetries for which we explicitly
296
+ account determines 𝑐. We discuss the details of computing the pattern and the phase in Sec. 3.1.
297
+ Separating the pattern and phase is useful because the evolution of both the pattern and the phase
298
+ only depend on the pattern. Thus, we simply replace 𝑢 with 𝑢 𝑝 in all the above equations and
299
+ then write one additional ODE for the phase
300
+ 𝑑𝜙
301
+ 𝑑𝑡 = 𝑔𝜙(ℎ; 𝜃 𝜙).
302
+ (2.12)
303
+ We then fix the parameters of 𝑔 to evolve ℎ (from 𝑢 𝑝) forward in time and use that to make a
304
+ phase prediction
305
+ ˜𝜙(𝑡𝑖 + 𝜏) = 𝜙(𝑡𝑖) +
306
+ ∫ 𝑡𝑖+𝜏
307
+ 𝑡𝑖
308
+ 𝑔𝜙(ℎ(𝑡𝑖); 𝜃 𝜙)𝑑𝑡.
309
+ (2.13)
310
+ Finally, we determine the parameters 𝜃 𝜙 to minimize the difference between the predicted phase
311
+ ˜𝜙(𝑡𝑖 + 𝜏) and the true phase 𝜙(𝑡𝑖 + 𝜏)
312
+ 𝐽𝜙 = 1
313
+ 𝑐𝐾
314
+ 𝐾
315
+ ∑︁
316
+ 𝑖=1
317
+
318
+ ||𝜙(𝑡𝑖 + 𝜏) − ˜𝜙(𝑡𝑖 + 𝜏)||2�
319
+ ,
320
+ (2.14)
321
+ using the method described above to compute the gradient of 𝐽𝜙.
322
+ 3. Results
323
+ 3.1. Description of Plane Couette Flow Data
324
+ In the following sections we apply DManD to DNS of turbulent PCF in a MFU domain.
325
+ Specifically, we consider the well-studied Hamilton, Kim, and Waleffe (HKW) domain (Hamilton
326
+ et al. 1995). We made this selection to compare our DManD results to the analysis of the self-
327
+ sustaining process in this domain, to compare our DManD results to other Galerkin-based ROMs,
328
+ and to compare our DManD results to known unstable periodic solutions in this domain.
329
+ For PCF we solve the Navier-Stokes equations
330
+ 𝜕u
331
+ 𝜕𝑡 + u · ∇u = −∇𝑝 + Re−1∇2u,
332
+ ∇ · u = 0
333
+ (3.1)
334
+ for a fluid confined between two plates moving in opposite directions with the same speed. Eq.
335
+
336
+ 7
337
+ 3.1 is the nondimensionalized form of the equations with velocities in the streamwise 𝑥 ∈ [0, 𝐿𝑥],
338
+ wall-normal 𝑦 ∈ [−1, 1], and spanwise 𝑧 ∈ [0, 𝐿𝑧] directions defined as u = [𝑢𝑥, 𝑢𝑦, 𝑢𝑧],
339
+ and pressure 𝑝. We solve this equation for a domain with periodic boundary conditions in 𝑥
340
+ and 𝑧 (u(0, 𝑦, 𝑧) = u(𝐿𝑥, 𝑦, 𝑧), u(𝑥, 𝑦, 0) = u(𝑥, 𝑦, 𝐿𝑧)) and no-slip, no-penetration boundary
341
+ conditions in 𝑦 (𝑢𝑥(𝑥, ±1, 𝑧) = ±1, 𝑢𝑦(𝑥, ±1, 𝑧) = 𝑢𝑧(𝑥, ±1, 𝑧) = 0). The complexity of the flow
342
+ increases as the Reynolds number Re increases and the domain size 𝐿𝑥 and 𝐿𝑧 increase. Here
343
+ we use the HKW cell, which is at Re = 400 with a domain size [𝐿𝑥, 𝐿𝑦, 𝐿𝑧] = [1.75𝜋, 2, 1.2𝜋]
344
+ (Hamilton et al. 1995). The HKW cell is one of the simplest flows that sustains turbulence for
345
+ extended periods of time before relaminarizing. We chose to use this flow because it is well
346
+ studied (refer to Sec. 1), it isolates the SSP (Hamilton et al. 1995), and a library of ECS exist for
347
+ this domain (Gibson et al. 2008b). Here we are interested in modeling the turbulent dynamics, so
348
+ we will remove data upon relaminarization as detailed below.
349
+ Eq. 3.1, under the boundary conditions described, is invariant (and its solutions equivariant)
350
+ under the discrete symmetries of point reflections about [𝑥, 𝑦, 𝑧] = [0, 0, 0]
351
+ P · [(𝑢𝑥, 𝑢𝑦, 𝑢𝑧, 𝑝)(𝑥, 𝑦, 𝑧, 𝑡)] = (−𝑢𝑥, −𝑢𝑦, −𝑢𝑧, 𝑝)(−𝑥, −𝑦, −𝑧, 𝑡)
352
+ (3.2)
353
+ reflection about the 𝑧 = 0 plane
354
+ R · [(𝑢𝑥, 𝑢𝑦, 𝑢𝑧, 𝑝)(𝑥, 𝑦, 𝑧, 𝑡)] = (𝑢𝑥, 𝑢𝑦, −𝑢𝑧, 𝑝)(𝑥, 𝑦, −𝑧, 𝑡)
355
+ (3.3)
356
+ and rotation by 𝜋 about the 𝑧-axis
357
+ RP · [(𝑢𝑥, 𝑢𝑦, 𝑢𝑧, 𝑝)(𝑥, 𝑦, 𝑧, 𝑡)] = (−𝑢𝑥, −𝑢𝑦, 𝑢𝑧, 𝑝)(−𝑥, −𝑦, 𝑧, 𝑡).
358
+ (3.4)
359
+ In addition to the discrete symmetries, there are also continuous translation symmetries in 𝑥 and
360
+ 𝑧
361
+ T𝜎𝑥,𝜎𝑧 · [(𝑢𝑥, 𝑢𝑦, 𝑢𝑧, 𝑝)(𝑥, 𝑦, 𝑧, 𝑡)] = (𝑢𝑥, 𝑢𝑦, 𝑢𝑧, 𝑝)(𝑥 + 𝜎𝑥, 𝑦, 𝑧 + 𝜎𝑧, 𝑡).
362
+ (3.5)
363
+ We incorporate all these symmetries in the POD represesntation (Smith et al. 2005), as we discuss
364
+ further in Sec. 3.2. Then, we use the method of slices (Budanur et al. 2015a) to phase align in the 𝑧
365
+ direction. By phase aligning in 𝑧 we fix the location of the low-speed streak. Without the alignment
366
+ in 𝑧, models performed poorly because the models needed to learn how to represent every spatial
367
+ shift of every snapshot. In what follows, we only consider phase-alignment in 𝑧, but we note that
368
+ extending this work to phase-alignment in 𝑥 is straightforward. To phase-align the data, we use
369
+ the first Fourier mode method-of-slices (Budanur et al. 2015a). First, we compute a phase by
370
+ taking the Fourier transform of the streamwise velocity in 𝑥 and 𝑧 ( ˆ𝑢𝑥(𝑘𝑥, 𝑦, 𝑘𝑧) = F𝑥,𝑧(𝑢𝑥)) at
371
+ a specific 𝑦 location (𝑦1) to compute the phase
372
+ 𝜙 = atan2(imag( ˆ𝑢𝑥(0, 𝑦1, 1)), real( ˆ𝑢𝑥(0, 𝑦1, 1))).
373
+ (3.6)
374
+ We select 𝑦1 to be one grid point off the bottom wall. Then we compute the pattern dynamics by
375
+ using the Fourier shift theorem to set the phase to 0 (i.e. move the low-speed streak to the center
376
+ of the channel)
377
+ u𝑝 = F −1
378
+ 𝑥,𝑧( ˆu exp(−𝑖𝑘𝑧𝜙)).
379
+ (3.7)
380
+ We generate turbulent PCF trajectories using the psuedo-spectral Channelflow code developed
381
+ by Gibson et al. (2012; 2021). In this code, the velocity and pressure fields are projected onto
382
+ Fourier modes in 𝑥 and 𝑧 and Chebyshev polynomials of the first kind in 𝑦. These coefficients
383
+ are evolved forward in time first using the multistage SMRK2 scheme (Spalart et al. 1991), then,
384
+ after taking multiple timesteps, using the multistep Adams-Bashforth Backward-Differentiation
385
+ 3 scheme (Peyret 2002). At each timestep, a pressure boundary condition is found such that
386
+ incompressibility is satisfied at the wall (𝑑𝑢𝑦/𝑑𝑦 = 0) using the influence matrix method and tau
387
+ correction developed by Kleiser & Schumann (1980).
388
+ Data was generated with Δ𝑡 = 0.0325 on a grid of [𝑁𝑥, 𝑁𝑦, 𝑁𝑧] = [32, 35, 32] in 𝑥, 𝑦, and 𝑧
389
+
390
+ 8
391
+ for the HKW cell. Starting from random divergence-free initial conditions, we ran simulations
392
+ forward for either 10, 000 xtime units or until relaminarization. Then we dropped the first 1, 000
393
+ time units as transient data and the last 1, 000 time units to avoid laminar data, and repeated with
394
+ a new initial condition until we had 91, 562 time units of data stored at intervals of one time unit.
395
+ We split this data into 80% for training and 20% for testing. Finally, we preprocess the data by
396
+ computing the mean ⟨u⟩ (𝑦) from the training data and subtracting it from all data u′ = u − ⟨u⟩,
397
+ and then we compute the pattern u′
398
+ 𝑝 and the phase 𝜙 as described above. The pattern 𝑢 𝑝 described
399
+ in Sec. 2 is u′
400
+ 𝑝 flattened into a vector (i.e. 𝑑 = 3𝑁𝑥𝑁𝑦𝑁𝑧). The POD and NN training use only
401
+ the training data, and all comparisons use test data unless otherwise specified.
402
+ 3.2. Dimension Reduction and Dynamic Model Construction
403
+ 3.2.1. Linear dimension reduction with POD: From O(105) to O(103)
404
+ The first task in DManD for this Couette flow data is finding a low-dimensional parameterization
405
+ of the manifold on which the long-time dynamics lie. We parameterize this manifold in two
406
+ steps. First, we reduce the dimension down from O(105) to 502 with the proper orthogonal
407
+ decomposition (POD), and, second, we use an autoencoder to reduce the dimension down to 𝑑ℎ.
408
+ The first step is simply a preprocessing step to reduce the size of the data, which reduces the
409
+ number of parameters in the autoencoder. Due to Whitney’s embedding theorem (Whitney 1936,
410
+ 1944), we know that as long as the manifold dimension is less than 251 (𝑑M < 251) then this
411
+ POD representation is diffeomorphic to the full state. As we show later, the manifold dimension
412
+ appears to be far lower than 𝑑M = 251, so no information of the full state should be lost with this
413
+ first step.
414
+ Proper orthogonal decomposition (POD) originates with the question of what function 𝚽
415
+ maximizes
416
+
417
+ |(u′, 𝚽)|2�
418
+ ||𝚽||2
419
+ .
420
+ (3.8)
421
+ Solutions 𝚽(𝑛) to this problem satisfy the eigenvalue problem
422
+ 3
423
+ ∑︁
424
+ 𝑗=1
425
+
426
+ 𝐿𝑥
427
+ 0
428
+ ∫ 1
429
+ −1
430
+
431
+ 𝐿𝑧
432
+ 0
433
+
434
+ 𝑢′
435
+ 𝑖(x, 𝑡)𝑢′∗
436
+ 𝑗 (x′, 𝑡)
437
+
438
+ Φ(𝑛)
439
+ 𝑗
440
+ (x′) 𝑑x′ = 𝜆𝑖Φ(𝑛)
441
+ 𝑖
442
+ (x)
443
+ (3.9)
444
+ (Holmes et al. 2012; Smith et al. 2005). Unfortunately, upon approximating these integrals, with
445
+ the trapezoidal rule for example, this becomes a 𝑑 × 𝑑 matrix, making computation intractable.
446
+ Furthermore, computing the average in Eq. 3.9, without any modifications, results in POD modes
447
+ that fail to preserve the underlying symmetries described above.
448
+ In order to make this problem computationally tractable, and preserve symmetries, we apply
449
+ the POD method used in Smith et al. (2005), with the slight difference that we first subtract off
450
+ the mean of state before performing the analysis. The first step in this procedure is to treat the
451
+ POD modes as Fourier modes in both the 𝑥 and 𝑧 directions. Holmes et al. show in (Holmes et al.
452
+ 2012) that for translation-invariant directions Fourier modes are the optimal POD modes. This
453
+ step transforms the eigenvalue problem into
454
+ 𝐿𝑥𝐿𝑧
455
+ 3
456
+ ∑︁
457
+ 𝑗=1
458
+ ∫ 1
459
+ −1
460
+
461
+ ˆ𝑢′
462
+ 𝑖(𝑘𝑥, 𝑦′, 𝑘𝑧, 𝑡) ˆ𝑢′∗
463
+ 𝑗 (𝑘𝑥, 𝑦′, 𝑘𝑧, 𝑡)
464
+
465
+ 𝜑(𝑛)
466
+ 𝑗𝑘𝑥 𝑘𝑧 (𝑦′) 𝑑𝑦′ = 𝜆(𝑛)
467
+ 𝑘𝑥 𝑘𝑧 𝜑(𝑛)
468
+ 𝑖𝑘𝑥 𝑘𝑧 (𝑦),
469
+ (3.10)
470
+ which reduces the 𝑑 × 𝑑 eigenvalue problem down to a 3𝑁𝑦 × 3𝑁𝑦 eigenvalue problem for every
471
+ wavenumber pair (𝑘𝑥, 𝑘𝑧) of Fourier coefficients. We used 5, 000 snapshots evenly sampled over
472
+ the training data to compute the POD modes. Then, to account for the discrete symmetries,
473
+ the data is augmented such that the mean in Eq. 3.10 is computed by adding all the discrete
474
+ symmetries of each snapshot.
475
+
476
+ 9
477
+ a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)
478
+ b)
479
+ b)
480
+ b)
481
+ b)
482
+ b)
483
+ b)
484
+ b)
485
+ b)
486
+ b)
487
+ b)
488
+ b)
489
+ b)
490
+ b)
491
+ b)
492
+ b)
493
+ b)
494
+ b)
495
+ Figure 1: (a) Eigenvalues of POD modes sorted in descending order. (b) Components of the
496
+ Reynolds stress for data generated by the DNS and this data projected onto 256 POD modes. In
497
+ (a) the curves are, from top to bottom,
498
+
499
+ 𝑢′2
500
+ 𝑥
501
+
502
+ ,
503
+
504
+ 𝑢′2
505
+ 𝑧
506
+
507
+ ,
508
+
509
+ 𝑢′2
510
+ 𝑦
511
+
512
+ , and
513
+
514
+ 𝑢′
515
+ 𝑥𝑢′
516
+ 𝑦
517
+
518
+ .
519
+ This analysis gives us POD modes
520
+ 𝚽(𝑛)
521
+ 𝑘𝑥 𝑘𝑧 (x) =
522
+ 1
523
+ √𝐿𝑥𝐿𝑧
524
+ exp
525
+
526
+ 2𝜋𝑖
527
+ � 𝑘𝑥𝑥
528
+ 𝐿𝑥
529
+ + 𝑘𝑧𝑧
530
+ 𝐿𝑧
531
+ ��
532
+ 𝝋(𝑛)
533
+ 𝑘𝑥 𝑘𝑧 (𝑦),
534
+ (3.11)
535
+ and eigenvalues 𝜆(𝑛)
536
+ 𝑘𝑥 𝑘𝑧. We sort the modes from largest eigenvalue to smallest eigenvalue (𝜆𝑖) and
537
+ and project onto the leading 256 modes, giving us a vector of POD coefficients 𝑎(𝑡). A majority
538
+ of these modes are complex, so projecting onto these modes results in a 502-dimensional system.
539
+ In Fig. 1a we plot the eigenvalues, which show a rapid drop and then a long tail that contributes
540
+ little to the energy content. By dividing the eigenvalues of the leading 256 modes by the total, we
541
+ find these modes contain 99.8% of the energy. To further illustrate that 256 modes is sufficient to
542
+ represent the state in this case, we consider the reconstruction of statistics after projecting onto
543
+ the POD modes. In Fig. 1b we show the reconstruction of four components of the Reynolds stress,
544
+
545
+ 𝑢′2
546
+ 𝑥
547
+
548
+ ,
549
+
550
+ 𝑢′2
551
+ 𝑧
552
+
553
+ ,
554
+
555
+ 𝑢′2
556
+ 𝑦
557
+
558
+ , and
559
+
560
+ 𝑢′
561
+ 𝑥𝑢′
562
+ 𝑦
563
+
564
+ . The projection onto POD modes matches all of these quantities
565
+ extremely well.
566
+ Now that we have converted the data to POD coefficients and filtered out the low energy modes,
567
+ we next train an autoencoder to perform nonlinear dimension reduction. As mentioned in Sec.
568
+ 2, we phase-align the data in the spanwise direction at this step using the first Fourier mode
569
+ method-of-slices. A common practice when training NNs is to normalize the data by subtracting
570
+ the mean and dividing by the standard deviation of each component. We do not take this approach
571
+ here because the standard deviation of the higher POD coefficients, which contribute less to the
572
+ reconstruction, is much smaller than the lower POD coefficients. In order to retain the important
573
+ information in the magnitudes, but put the values in a more amenable form for NN training, we
574
+ instead normalize the POD coefficients by subtracting the mean and dividing by the maximum
575
+ standard deviation. Then, we take this input and train autoencoders to minimize the loss in Eq.
576
+ 2.7 using an Adam optimizer (Kingma & Ba 2015) in Keras (Chollet et al. 2015). We train for
577
+ 500 epochs with a learning rate scheduler that drops the learning rate from 10−3 to 10−4 after
578
+ 400 epochs. At this point we see no improvement in the reconstruction error. For the hybrid
579
+ autoencoder approach, we set 𝛼 = 0.01. Table 1 includes additional NN architecture details.
580
+ 3.2.2. Nonlinear dimension reduction with autoencoders: From O(103) to O(101)
581
+ With the above “preprocessing" step completed, we now turn to the reduction of dimension
582
+ with the nonlinear approach enabled by the autoencoder structure. We consider three approaches
583
+ to reducing the dimension of 𝑎: (1) Training a hybrid autoencoder, (2) Training a standard
584
+
585
+ 10-2
586
+ DNS
587
+ 0.06
588
+ POD
589
+ 10-4
590
+ 0.04
591
+ 10-6
592
+ 0.02
593
+ 10-8 .
594
+ 0.00
595
+ 10-10
596
+ 100
597
+ 101
598
+ 102
599
+ 103
600
+ -1.0
601
+ -0.5
602
+ 0.0
603
+ 0.5
604
+ 1.0
605
+ i
606
+ y10
607
+ Table 1: Architectures of NNs. “Shape" indicates the dimension of each layer, “Activation" the
608
+ corresponding activation functions, and “sig" is the sigmoid activation.“Learning Rate" gives the
609
+ learning rate at multiple times during training. The learning rates was dropped at even intervals.
610
+ Function
611
+ Shape
612
+ Activation
613
+ Learning Rate
614
+ E
615
+ 502/1000/𝑑ℎ
616
+ sig/linear
617
+ [10−3, 10−4]
618
+ D
619
+ 𝑑ℎ/1000/502
620
+ sig/linear
621
+ [10−3, 10−4]
622
+ 𝑔NN
623
+ 𝑑ℎ/200/200/𝑑ℎ
624
+ sig/sig/linear [10−2, 10−3, 10−4]
625
+ 𝑔𝜙
626
+ 𝑑ℎ/200/200/1
627
+ sig/sig/linear [10−2, 10−3, 10−4]
628
+ Figure 2: Mean squared error on test data for POD, standard autoencoders, and hybrid
629
+ autoencoders at various dimensions 𝑑ℎ. At each dimension there are four standard and four
630
+ hybrid autoencoders.
631
+ autoencoder, (3) linear projection onto a small set of POD modes. We describe the first two
632
+ approaches in Sec. 2, noting that the POD projection onto 256 (complex) modes can be written
633
+ as 𝑎 = 𝑈𝑇
634
+ 𝑟 𝑢. The third approach just corrsponds to setting E and D to zero in Eqs. 2.5 and 2.6. In
635
+ Fig. 2 we show the MSE of reconstructing 𝑎 with these three approaches over a range of different
636
+ dimensions 𝑑ℎ. We use NNs with the same architectures for both the standard and the hybrid
637
+ autoencoder approaches. Due to the variability introduced into autoencoder training by randomly
638
+ initialized weights and stochasticity in the optimization, we show the error for four separately
639
+ trained autoencoders, at each 𝑑ℎ. We see that the autoencoders perform an order magnitude better
640
+ than POD in the range of dimension considered here. Both the standard and hybrid autoencoder
641
+ approaches perform the same, so we select the hybrid approach because it can be viewed as a
642
+ nonlinear correction to the POD projection. Next we use the low-dimensional representations
643
+ from these autoencoders to train stabilized neural ODEs.
644
+ 3.2.3. Neural ODE Training
645
+ After training four autoencoders at each dimension 𝑑ℎ, we chose a set of damping parameters,
646
+ 𝛽, and for each, then trained four stabilized neural ODEs for all four autoencoders at every
647
+ dimension 𝑑ℎ. This results in 16 models at every 𝑑ℎ and 𝛽. The final 𝛽 value of 0.1 was selected
648
+ so that long-time trajectories neither blew up nor decayed too strongly. Before training the ODEs,
649
+ we preprocess each autoencoder’s latent space data set ℎ by subtracting the mean. It is important
650
+ to center the data because the linear damping (Eq. 2.10) pushes trajectories towards the origin. We
651
+ train the stabilized neural ODEs to predict the evolution of the centered data by using an Adam
652
+
653
+ MSE Test Data
654
+ Hybrid
655
+ Stand
656
+ POD
657
+ 10°
658
+ 5
659
+ 10
660
+ 1511
661
+ Figure 3: Snapshots of the streamwise velocity at 𝑦 = 0 from the DNS and from the DManD
662
+ model at 𝑑ℎ = 18.
663
+ optimizer in Pytorch (Paszke et al. 2019; Chen et al. 2019) to minimize the loss in Eq. 2.11. We
664
+ train using a learning rate scheduler that drops at three even intervals during training and we train
665
+ until the learning curve stops improving. Table 1 shows the details of this NN. Unless otherwise
666
+ stated, we show results for the one model out of those sixteen at each dimension with the lowest
667
+ relative error averaged over all the statistics we consider.
668
+ 3.3. Short-time tracking
669
+ In the following two sections we evaluate the performance of the DManD models at reconstruct-
670
+ ing short-time trajectories and long-time statistics. Figure 3 shows snapshots of the streamwise
671
+ velocity at the center plane of the channel, 𝑦 = 0, for the DNS and DManD at 𝑑ℎ = 18. We
672
+ choose to show results for 𝑑ℎ = 18 because the autoencoder error begins to level off around
673
+ this dimension, and, as we will show, the error in statistics levels off before this dimension. The
674
+ value 𝑑ℎ = 18 is not necessarily the minimal dimension required to model this system. In Fig. 3,
675
+ both the DNS and the DmanD model show key characteristics of the SSP: (1) low-speed streaks
676
+ become wavy, (2) the wavy low-speed streaks break down generating rolls, (3) the rolls lift fluid
677
+ from the walls, regenerating streaks.
678
+ Not only does DManD capture the qualitative behavior of the SSP, but Fig. 3 also shows good
679
+ quantitative agreement as well. To further illustrate this, in Fig. 4 we show the modal root-mean
680
+ squared (RMS) velocity
681
+ 𝑀(𝑘𝑥, 𝑘𝑧) =
682
+ �∫ 1
683
+ −1
684
+ ( ˆ𝑢2
685
+ 𝑥(𝑘𝑥, 𝑦, 𝑘𝑧) + ˆ𝑢2
686
+ 𝑦(𝑘𝑥, 𝑦, 𝑘𝑧) + ˆ𝑢2
687
+ 𝑧(𝑘𝑥, 𝑦, 𝑘𝑧))𝑑𝑦
688
+ �1/2
689
+ ,
690
+ (3.12)
691
+ which Hamilton et al. (1995) used to identify the different parts of the SSP. Specifically, we
692
+ consider the 𝑀(0, 1) mode, which corresponds to the low speed streak and the 𝑀(1, 0) mode
693
+ which corresponds to the 𝑥-dependence that appears when the streak becomes wavy and breaks
694
+ up. In this example, the two curves match well over a cycle of the SSP and only start to move
695
+ away after ∼ 150 time units, which is about three Lyapunov times.
696
+ While the previous result shows a single example, we also consider ensembles of initial
697
+ conditions. Figure 5 shows the tracking error ||𝑎(𝑡𝑖 + 𝑡) − ˜𝑎(𝑡𝑖 + 𝑡)|| of 10 trajectories, starting at
698
+ 𝑡𝑖, for a model at 𝑑ℎ = 18. Here we normalize the tracking error by the error between solutions
699
+ at random times 𝑡𝑖 and 𝑡 𝑗 𝐷 = ⟨||𝑎(𝑡𝑖) − 𝑎(𝑡)||⟩. In this case the darkest line corresponds to the
700
+ flow field in Figs. 3 and 4. When considering the other initial conditions in Fig. 5, there tends
701
+ to be a relatively slow rise in the error over ∼50 time units and then a more rapid increase after
702
+
703
+ DNS
704
+ 0.6
705
+ 0=↑
706
+ t = 85
707
+ 140
708
+ -0.4
709
+ 22
710
+ 0.2
711
+ 0
712
+ 0.0
713
+ DManD
714
+ 0=↑
715
+ t = 85
716
+ t = 140
717
+ -0.2
718
+ 22
719
+ -0.4
720
+ 0
721
+ -0.6
722
+ 0
723
+ 5
724
+ 0
725
+ 5
726
+ 0
727
+ 5
728
+ 0
729
+ 512
730
+ Figure 4: Modal RMS velocity from the DNS (𝑀) and from the DManD model at 𝑑ℎ = 18 ( ˜𝑀).
731
+ The markers correspond to the times in Fig. 3.
732
+ Figure 5: Normalized tracking error for 10 random initial conditions (different shades) using
733
+ DManD with 𝑑ℎ = 18.
734
+ this point. To better understand how this tracking varies with the dimension of the model we next
735
+ consider the ensemble-averaged tracking error.
736
+ In Fig. 6a we show the normalized ensemble-averaged tracking error for model dimensions
737
+ between 𝑑ℎ = 3 and 18. For 𝑑ℎ = 3 − 5 there is a rapid rise in the error until ∼40 time units after
738
+ which the error levels off. This behavior often happens due to trajectories quickly diverging and
739
+ landing on stable fixed points or periodic orbits that do not exist in the true system. For 𝑑ℎ = 6−10
740
+ there is an intermediate behavior where lines diverge more quickly than the higher-dimensional
741
+ models, but tend to approach the same tracking error at ∼100 time units. Then, for the remaining
742
+ models 𝑑ℎ = 11 − 18, there is a smooth improvement in the tracking error over this time interval.
743
+ As the dimension increases in this range the trends stay the same, but the error continues to
744
+ decrease, which is partially due to improvement in the autoencoder performance.
745
+ The instantaneous kinetic energy of the flow is
746
+ 𝐸(𝑡) =
747
+ 1
748
+ 2𝐿𝑥𝐿𝑧
749
+
750
+ 𝐿𝑧
751
+ 0
752
+ ∫ 1
753
+ −1
754
+
755
+ 𝐿𝑥
756
+ 0
757
+ 1
758
+ 2u · u𝑑𝑥𝑑𝑦𝑑𝑧,
759
+ (3.13)
760
+ and we denote its fluctuating part as 𝑘(𝑡) = 𝐸(𝑡) − ⟨𝐸⟩. In Fig. 6b we show the temporal
761
+ autocorrelation of 𝑘. Again, for 𝑑ℎ = 3 − 5 we see clear disagreement between the true
762
+ autocorrelation and the prediction. Above 𝑑ℎ > 5 all of the models match the temporal
763
+ autocorrelation well, without a clear trend in the error as dimension changes. All these models
764
+
765
+ 0.25
766
+ 0.20
767
+ M(0,1)
768
+ M(kx, kz)
769
+ 0.15
770
+ M(0, 1)
771
+ M(1, 0)
772
+ 0.10 -
773
+ M(1, 0)
774
+ 0.05
775
+ 0.00
776
+ 0
777
+ 50
778
+ 100
779
+ 150
780
+ 200
781
+ t2.0
782
+ 1.5
783
+ I/D
784
+ a
785
+ 1.0
786
+
787
+ 0.5
788
+ 0.0
789
+ 0
790
+ 20
791
+ 40
792
+ 60
793
+ 80
794
+ 100
795
+ t13
796
+ a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)
797
+ b)
798
+ b)
799
+ b)
800
+ b)
801
+ b)
802
+ b)
803
+ b)
804
+ b)
805
+ b)
806
+ b)
807
+ b)
808
+ b)
809
+ b)
810
+ b)
811
+ b)
812
+ b)
813
+ b)
814
+ Figure 6: (a) ensemble averaged short-time tracking and (b) temporal autocorrelation of the
815
+ kinetic energy for DmanD models of increasing dimension. In (b) odd numbers above 𝑑ℎ = 5 are
816
+ omitted for clarity.
817
+ a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)
818
+ b)
819
+ b)
820
+ b)
821
+ b)
822
+ b)
823
+ b)
824
+ b)
825
+ b)
826
+ b)
827
+ b)
828
+ b)
829
+ b)
830
+ b)
831
+ b)
832
+ b)
833
+ b)
834
+ b)
835
+ Figure 7: (a) an example of the phase evolution and (b) the MSD of the phase evolution for the
836
+ DNS and the DManD model at 𝑑ℎ = 18.
837
+ match well for ∼ 40 time units, with 𝑑ℎ = 18 (the darkest line) matching the data extremely well
838
+ for two Lyapunov times.
839
+ Finally, before investigating the long-time predictive capabilities of the model, we show the
840
+ tracking of phase dynamics for 𝑑ℎ = 18. As mentioned in Sec. 2, we decouple the phase and pattern
841
+ dynamics such that the time evolution of the phase only depends upon the pattern dynamics. Here
842
+ we take the 𝑑ℎ = 18 model and used it to train an ODE for the phase dynamics. For training we
843
+ repeat the process used for training 𝑔𝑁 𝑁 to train 𝑔𝜙 with the loss in Eq. 2.14. Table 1 contains
844
+ details on the NN architecture.
845
+ In Fig. 7a we show an example of the model phase evolution over 200 time units. In this
846
+ example, the model follows the same downward drift in phase despite not matching exactly.
847
+ Then, to show the statistical agreement between the DNS and the model, we show the mean
848
+ squared phase displacement MSD =
849
+
850
+ (𝜃(𝑡𝑖) − 𝜃(𝑡𝑖 + 𝑡))2�
851
+ for both the DNS and the model in
852
+ Fig. 7b, as was done for Kolmogorov flow by Pérez De Jesús & Graham (2022). The curves are
853
+ in good agreement. All of the remaining long-time statistics we report are phase invariant, so the
854
+ remaining results use only models for the pattern dynamics.
855
+
856
+ 1.0
857
+ 18
858
+ DManD
859
+ 1.0
860
+ <2(±)> / (+)())
861
+ DNS
862
+ 16
863
+ D 0.8
864
+ 14
865
+ 0.5
866
+ 12
867
+ 0.6
868
+ 10
869
+ 0.0
870
+ 8
871
+ 0.4
872
+ 6
873
+ 0.2
874
+ 4
875
+ -0.5
876
+ 0
877
+ 20
878
+ 40
879
+ 60
880
+ 80
881
+ 100
882
+ 0
883
+ 20
884
+ 40
885
+ 60
886
+ 80
887
+ 100
888
+ 4
889
+ tDNS
890
+ 10-1
891
+ DNS
892
+ -3.4
893
+ DManD
894
+ DManD
895
+ 10-2
896
+ -3.5
897
+ MSD
898
+ 0
899
+ 10-3
900
+ -3.6
901
+ 10-4
902
+ 3.7
903
+ 10-5
904
+ 50
905
+ 0
906
+ 100
907
+ 150
908
+ 200
909
+ 100
910
+ 101
911
+ 102
912
+ t
913
+ t14
914
+ Figure 8: Fraction of unstable DManD models with standard neural ODEs and with stabilized
915
+ neural ODEs at various dimensions.
916
+ 3.4. Long-time statistics
917
+ Next we investigate the ability of the DManD model to capture the long-time dynamics of PCF.
918
+ An obvious prerequisite for models to capture long-time dynamics is the long-time stability of the
919
+ models. As mentioned in Sec. 2, the long-time trajectories of standard neural ODEs often become
920
+ unstable, which led us to use stabilized neural ODEs with an explicit damping term. We quantify
921
+ this observation by counting, of the 16 models trained at each dimension 𝑑ℎ, how many become
922
+ unstable with and without the presence of an explicit damping term. From our training data we
923
+ know where ℎ should lie, so if it falls far outside this range after some time we can classify the
924
+ model as unstable. In particular, we classify models as unstable if the norm of the final state is
925
+ two times that of the maximum in our data (|| ˜ℎ(𝑇)|| > 2 max𝑡 ||ℎ(𝑡)||), after 𝑇 = 104 time units.
926
+ In all of the unstable cases || ˜ℎ(𝑡)|| follows the data over some short time range before eventually
927
+ growing indefinitely.
928
+ In Fig. 8 we show the number of unstable models with and without damping. With damping,
929
+ all of the models are stable, whereas without damping almost all models become unstable for
930
+ 𝑑ℎ = 5−16, and around half become unstable in the other cases. Additionally, with longer runs or
931
+ with different initial conditions, many of the models without damping labelled as stable here also
932
+ eventually become unstable. This lack of stability happens when inaccuracies in the neural ODE
933
+ model pushes trajectories off the attractor. Once off the attractor, the model is presented with
934
+ states unlike the training data leading to further growth in this error. In Linot & Graham (2022);
935
+ Linot et al. (2023) we show more results highlighting this behavior. So, although some standard
936
+ neural ODE models do provide reasonable statistics, using these models presents challenges due
937
+ to this lack of robustness. As such, all other results we show use stabilized neural ODEs.
938
+ While Fig. 8 indicates that stabilized neural ODEs predict ˜ℎ in a similar range to that of the data,
939
+ it does not quantify the accuracy of these predictions. In fact, with few dimensions many of these
940
+ models do not remain chaotic, landing on fixed points or periodic orbits. The first metric we use to
941
+ quantify the long-time performance of the DManD method is the mean-squared POD coefficient
942
+ amplitudes (
943
+
944
+ ||𝑎𝑛||2�
945
+ ). We consider this quantity because Gibson reports it for POD-Galerkin in
946
+ Gibson (2002) at various levels of truncation. In Fig. 9 we show how well the DManD model, with
947
+ 𝑑ℎ = 18, captures this quantity, in comparison to the POD-Galerkin model in Gibson (2002). The
948
+ two data sets slightly differ because we subtract the mean before applying POD and Gibson did
949
+ not. The DManD method, with only 18 degrees of freedom, matches the mean-squared amplitudes
950
+ to high accuracy, far better than all of the POD-Galerkin models. It is not until POD-Galerkin
951
+ keeps 1024 modes that the results become comparable, which corresponds to ∼ 2000 degrees
952
+ of freedom because most coefficients are complex. Additionally, our method requires only data,
953
+
954
+ 1.00
955
+ Standard
956
+ Fraction Unstable
957
+ Stabilized
958
+ +
959
+ 0.75
960
+ 0.50
961
+ 0.25
962
+ 0.00
963
+ XxX
964
+ 10
965
+ 20
966
+ 30
967
+ 4015
968
+ a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)
969
+ b)
970
+ b)
971
+ b)
972
+ b)
973
+ b)
974
+ b)
975
+ b)
976
+ b)
977
+ b)
978
+ b)
979
+ b)
980
+ b)
981
+ b)
982
+ b)
983
+ b)
984
+ b)
985
+ b)
986
+ Figure 9: Comparison of
987
+
988
+ ||𝑎𝑛||2�
989
+ (mean-squared POD coefficient amplitudes) from the DNS to
990
+ (a)
991
+
992
+ ||𝑎𝑛||2�
993
+ from the DManD model at 𝑑ℎ = 18 and (b)
994
+
995
+ ||𝑎𝑛||2�
996
+ from POD-Galerkin with 𝑁
997
+ POD modes (reproduced with permission from Gibson (2002)). In (b), the quantity 𝜆 is equivalent
998
+ to
999
+
1000
+ ||𝑎𝑛||2�
1001
+ from the DNS.
1002
+ Figure 10: Components of the Reynolds stress with increasing dimension for DManD models at
1003
+ various dimensions. Odd numbers above 𝑑ℎ = 5 are omitted for clarity.
1004
+ whereas the POD Galerkin approach requires both data for computing the POD and knowledge
1005
+ of the equations of motion for projecting the equations onto these modes.
1006
+ We now investigate how the Reynolds stress and the power input vs. dissipation vary with
1007
+ dimension. Figure 10 shows four components of the Reynolds stress at various dimensions. For
1008
+
1009
+ 𝑢′2
1010
+ 𝑥
1011
+
1012
+ and
1013
+
1014
+ 𝑢′
1015
+ 𝑥𝑢′
1016
+ 𝑦
1017
+
1018
+ , nearly all the models match the data, with relatively small deviations only
1019
+ appearing for 𝑑ℎ ∼ 3−6. For
1020
+
1021
+ 𝑢′2
1022
+ 𝑦
1023
+
1024
+ and
1025
+
1026
+ 𝑢′2
1027
+ 𝑧
1028
+
1029
+ , this deviation becomes more obvious, and the lines
1030
+ do not converge until around 𝑑ℎ > 10, with all models above this dimension exhibiting a minor
1031
+ overprediction in
1032
+
1033
+ 𝑢′2
1034
+ 𝑧
1035
+
1036
+ .
1037
+ To evaluate how accurate the models are at reconstructing the energy balance, we look at joint
1038
+ PDFs of power input and dissipation. The power input is the amount of energy required to move
1039
+
1040
+ X10-3
1041
+ X10-3
1042
+ 75
1043
+ 4.
1044
+ 18
1045
+ 50
1046
+
1047
+ 25
1048
+ :14
1049
+ 0
1050
+ 0
1051
+ ×10-3
1052
+ X10-3
1053
+ 10
1054
+ 0.0
1055
+ 6
1056
+ DManD
1057
+ DNS
1058
+ 2.5
1059
+ 4
1060
+ 6
1061
+ -5.0
1062
+ 2
1063
+ -7.5
1064
+ 0
1065
+ -1.0
1066
+ -0.5
1067
+ 1.0
1068
+ -0.5
1069
+ 0.0
1070
+ 0.5
1071
+ -1.0
1072
+ 0.0
1073
+ 0.5
1074
+ 1.0
1075
+ y
1076
+ y10-
1077
+ 10-
1078
+ N=16
1079
+ DNS
1080
+ N=128
1081
+ N=512
1082
+ N=1024
1083
+ 10-2
1084
+ DManD
1085
+ X
1086
+
1087
+
1088
+ 210-3
1089
+ up
1090
+ 10-4
1091
+ 10~4
1092
+ 10-5
1093
+ 10~5/
1094
+
1095
+ 10-6
1096
+ 10~
1097
+ 10'
1098
+ 102
1099
+ 100
1100
+ 100
1101
+ 101
1102
+ 102
1103
+ n+1
1104
+ n+l16
1105
+ the walls:
1106
+ 𝐼 =
1107
+ 1
1108
+ 2𝐿��𝐿𝑧
1109
+
1110
+ 𝐿𝑥
1111
+ 0
1112
+
1113
+ 𝐿𝑧
1114
+ 0
1115
+ 𝜕𝑢𝑥
1116
+ 𝜕𝑦
1117
+ ����
1118
+ 𝑦=−1
1119
+ + 𝜕𝑢𝑥
1120
+ 𝜕𝑦
1121
+ ����
1122
+ 𝑦=1
1123
+ 𝑑𝑥𝑑𝑧,
1124
+ (3.14)
1125
+ and the dissipation is the energy lost to heat due to viscosity:
1126
+ 𝐷 =
1127
+ 1
1128
+ 2𝐿𝑥𝐿𝑧
1129
+
1130
+ 𝐿𝑥
1131
+ 0
1132
+ ∫ 1
1133
+ −1
1134
+
1135
+ 𝐿𝑧
1136
+ 0
1137
+ |∇ × u|2 𝑑𝑥𝑑𝑦𝑑𝑧.
1138
+ (3.15)
1139
+ These two terms define the rate of change of energy in the system �𝐸 = 𝐼 − 𝐷, which must average
1140
+ to zero over long times. Checking this statistic is important to show the DManD models correctly
1141
+ balance the energy.
1142
+ Figures 11a-11c show the PDF from the DNS, the PDF for 𝑑ℎ = 6 and the PDF for 𝑑ℎ = 18,
1143
+ generated from a single trajectory evolved for 5000 time units, and Figs. 11e and 11f show the
1144
+ the absolute difference between the true and model PDFs. With 𝑑ℎ = 6 the model overestimates
1145
+ the number of low dissipation states, while 𝑑ℎ = 18 matches the density well. In Fig. 11d we
1146
+ compare the joint PDFs at all dimension with the true PDF using the earth movers distance (EMD)
1147
+ (Rubner et al. 1998). The EMD determines the distance between two PDFs as a solution to the
1148
+ transportation problem by treating the true PDF as “supplies" and the model PDF as “demands"
1149
+ and finding the flow which minimizes the work required to move one to the other. We compute the
1150
+ distance between PDFs using the EMD because it is a cross-bin distance, meaning the distance
1151
+ accounts for the density in neighboring bins. This is in contrast to bin-to-bin distances, like the KL
1152
+ divergence, which only uses the error at a given bin. Bin-to-bin distances can vary significantly
1153
+ with small shifts in one PDF (misalignment) and when changing the number of bins used to
1154
+ generate the PDF (Ling & Okada 2007). We choose the EMD because it does not suffer from
1155
+ these issues. In Fig. 11d we see a steep drop in the EMD at 𝑑ℎ = 5 and after 𝑑ℎ > 10 the joint
1156
+ PDFs are in excellent agreement with the DNS. The dashed line corresponds to the EMD between
1157
+ two different trajectories from the DNS.
1158
+ 3.5. Finding ECS in the model
1159
+ Now that we know that the DManD model quantitatively captures many of the key characteris-
1160
+ tics of MFU PCF, we now want to explore using the model to discover ECS. In particular, we first
1161
+ investigate the whether known periodic orbits of the DNS exist in the DManD model, and then
1162
+ we use the DManD model to search for new periodic orbits. Here we note that because our model
1163
+ predicts phase-aligned dynamics, the periodic orbits of the DManD model are either periodic
1164
+ or relative periodic orbits, depending on the phase evolution, which we have not tracked. In the
1165
+ following we omit all ˜·, so all functions should be assumed to come from a DManD model.
1166
+ Here we outline the approach we take to find periodic orbits, which follows Cvitanović et al.
1167
+ (2016). When searching for periodic orbits we seek an initial condition to a trajectory that repeats
1168
+ after some time period. This is equivalent to finding the zeros of
1169
+ 𝐻(ℎ,𝑇) = 𝐺𝑇 (ℎ) − ℎ,
1170
+ (3.16)
1171
+ where 𝐺𝑇 (ℎ) is the flow map forward 𝑇 time units from ℎ: i.e. 𝐺𝑇 (ℎ(𝑡)) = ℎ(𝑡 +𝑇). We compute
1172
+ 𝐺𝑇 (ℎ) from Eq. 2.9. Finding zeros to Eq. 3.16 requires that we find both a point ℎ∗ on the
1173
+ periodic orbit and a period 𝑇∗ such that 𝐻(ℎ∗,𝑇∗) = 0. One way to find ℎ∗ and 𝑇∗ is by using the
1174
+ Newton-Raphson method.
1175
+ By performing a Taylor series expansion of 𝐻 we find near the fixed point ℎ∗,𝑇∗ of 𝐻 that
1176
+ 𝐻(ℎ∗,𝑇∗) − 𝐻(ℎ,𝑇) ≈ 𝐷ℎ𝐻(ℎ,𝑇)Δℎ + 𝐷𝑇 𝐻(ℎ,𝑇)Δ𝑇
1177
+ −𝐻(ℎ,𝑇) ≈ 𝐷ℎ𝐻(ℎ,𝑇)Δℎ + 𝑔 (𝐺𝑇 (ℎ)) Δ𝑇,
1178
+ (3.17)
1179
+ where 𝐷ℎ is the Jacobian of 𝐻 with respect to ℎ, 𝐷𝑇 is the Jacobian of 𝐻 with respect to the
1180
+
1181
+ 17
1182
+ a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)
1183
+ b)
1184
+ b)
1185
+ b)
1186
+ b)
1187
+ b)
1188
+ b)
1189
+ b)
1190
+ b)
1191
+ b)
1192
+ b)
1193
+ b)
1194
+ b)
1195
+ b)
1196
+ b)
1197
+ b)
1198
+ b)
1199
+ b)
1200
+ c)c)c)c)c)c)c)c)c)c)c)c)c)c)c)c)c)
1201
+ d)
1202
+ d)
1203
+ d)
1204
+ d)
1205
+ d)
1206
+ d)
1207
+ d)
1208
+ d)
1209
+ d)
1210
+ d)
1211
+ d)
1212
+ d)
1213
+ d)
1214
+ d)
1215
+ d)
1216
+ d)
1217
+ d)
1218
+ e)e)e)e)e)e)e)e)e)e)e)e)e)e)e)e)e)
1219
+ f)f)f)f)f)f)f)f)f)f)f)f)f)f)f)f)f)
1220
+ Figure 11: (a)-(c): examples of joint PDFs for the true system, the DManD model at 𝑑ℎ = 6, and
1221
+ the DManD model at 𝑑ℎ = 18. (d): earth movers distance between the PDF from the DNS and the
1222
+ PDFs predict by the DManD model at various dimensions. “DNS" is the error between two PDFs
1223
+ generated from DNS trajectories of the same length with different initial conditions. (e) and (f):
1224
+ the error associated with the DManD model PDFs at 𝑑ℎ = 6 and 𝑑ℎ = 18.
1225
+ period 𝑇, Δℎ = ℎ∗ − ℎ and Δ𝑇 = 𝑇∗ − 𝑇. To have a complete set of equations for Δℎ and Δ𝑇, we
1226
+ supplement Eq. 3.17 with the constraint that the updates Δℎ are orthogonal to the vector field at
1227
+ ℎ: i.e.,
1228
+ 𝑔(ℎ)𝑇 Δℎ = 0.
1229
+ (3.18)
1230
+ With this constraint, at Newton step (𝑖), the system of equations becomes
1231
+ � 𝐷ℎ(𝑖) 𝐻(ℎ(𝑖),𝑇 (𝑖))
1232
+ 𝑔(𝐺𝑇 (𝑖) (ℎ(𝑖)))
1233
+ 𝑔(ℎ(𝑖))����
1234
+ 0
1235
+ � � Δℎ(𝑖)
1236
+ Δ𝑇 (𝑖)
1237
+
1238
+ = −
1239
+
1240
+ 𝐻(ℎ(𝑖),𝑇 (𝑖))
1241
+ 0
1242
+
1243
+ ,
1244
+ (3.19)
1245
+ which, in the standard Newton-Raphson method, is used to update the guesses ℎ(𝑖+1) = ℎ(𝑖) +Δℎ(𝑖)
1246
+ and 𝑇 (𝑖+1) = 𝑇 (𝑖) + Δ𝑇 (𝑖).
1247
+ Typically, a Newton-Krylov method is used to avoid explicitly constructing the Jacobian
1248
+ (Viswanath 2007). However, with DManD, computing the Jacobian is simple, fast, and requires
1249
+ little memory because the state representation is dramatically smaller in the DManD model
1250
+ than in the DNS. We compute the Jacobian 𝐷ℎ𝐻(ℎ,𝑇) directly by using the same automatic
1251
+ differentiation tools used for training the neural ODE. Furthermore, if we had chosen to
1252
+ represent the dynamics in discrete, rather than continuous time, computation of general periodic
1253
+ orbits would not be possible, as the period 𝑇 can take on arbitrary values and a discrete-time
1254
+ representation would limit 𝑇 to multiples of the time step. When finding periodic orbits of the
1255
+ DManD model we used the Scipy “hybr" method, which uses a modification of the Powell hybrid
1256
+ method (Virtanen et al. 2020), and for finding periodic orbits of the DNS we used the Newton
1257
+ GMRES-Hookstep method built into Channelflow (Gibson et al. 2021). In the following trials we
1258
+ only consider DManD models with 𝑑ℎ = 18.
1259
+ For the HKW cell there exists a library of POs made available by Gibson et al. (2008b). To
1260
+ investigate if the DManD model finds POs similar to existing solutions, we took states from the
1261
+ known POs, encoded them, and used this as an initial condition in the DManD Newton solver to
1262
+
1263
+ DNS
1264
+ dh = 6
1265
+ dh = 18
1266
+ 4
1267
+ 3.5
1268
+ 3.5
1269
+ 3.5
1270
+ D 3.0
1271
+ D 3.0
1272
+ D 3.0
1273
+ P
1274
+ -2
1275
+ 2.5
1276
+ 2.5
1277
+ 2.5
1278
+ 2.0
1279
+ 2.0
1280
+ 2.0
1281
+ 0
1282
+ 2.0
1283
+ 2.5
1284
+ 3.0
1285
+ 2.0
1286
+ 2.5
1287
+ 3.0
1288
+ 2.0
1289
+ 2.5
1290
+ 3.0
1291
+ I
1292
+ I
1293
+ 1
1294
+ Error dh = 6
1295
+ Error dh = 18
1296
+ DManD
1297
+ 0.2
1298
+ -3
1299
+ 3.5
1300
+ 3.5
1301
+ DNS
1302
+ ~P
1303
+ -2
1304
+ D 3.0
1305
+ D 3.0
1306
+ P
1307
+ 2.5
1308
+ 2.5
1309
+ :1
1310
+ 2.0
1311
+ 2.0
1312
+ 0.0
1313
+ 0
1314
+ 5
1315
+ 10
1316
+ 15
1317
+ 2.0
1318
+ 2.5
1319
+ 3.0
1320
+ 2.0
1321
+ 2.5
1322
+ 3.0
1323
+ dh
1324
+ I
1325
+ 118
1326
+ Figure 12: Power input vs. dissipation of known periodic orbits (period reported in bottom right)
1327
+ from the DNS and periodic orbits found in the DManD model at 𝑑ℎ = 18. The blue line is a long
1328
+ trajectory of the DNS for comparison.
1329
+ find POs in the model. In Fig. 12 we show projections of 12 known POs, which we identify by the
1330
+ period 𝑇, and compare them to projections of POs found using the DManD model. This makes
1331
+ up a majority of the POs made available by Gibson et al. (2008b). Of the other known solutions,
1332
+ three are RPOs with phase-shifts in the streamwise direction that our model, with the current
1333
+ setup, can not capture. The other two have short periods of 𝑇 = 19.02 and 𝑇 = 19.06. A majority
1334
+ of the POs found with DManD land on initial conditions near that of the DNS and follow similar
1335
+ trajectories trajectories to the DNS.
1336
+ How close many of these trajectories are to the true PO is surprising and encouraging for
1337
+ many reasons. First, the data used for training the DManD model does not explicitly contain any
1338
+ POs. Second, this approach by no means guarantees convergence on a PO in the DManD model.
1339
+ Third, starting with an initial condition from a PO does not necessarily mean that the solution the
1340
+ Newton solver lands on will be the closest PO to that initial condition, so there may exist POs in
1341
+ the DManD model closer to the DNS solutions than what we present here.
1342
+ Now that we know the DManD model can find POs similar to those known to exist for the DNS,
1343
+ we now use it to search for new POs. First, we searched for POs in three of the 𝑑ℎ = 18 models
1344
+ by randomly selecting 20 initial conditions and selecting 4 different periods 𝑇 = [20, 40, 60, 80].
1345
+
1346
+ DNS
1347
+ 3.5
1348
+ DManD
1349
+ D 3.0
1350
+ 2.5
1351
+ T = 62.13
1352
+ T = 68.07
1353
+ T = 75.35
1354
+ 3.5
1355
+ D 3.0
1356
+ 2.5
1357
+ T = 76.82
1358
+ T = 76.85
1359
+ T = 85.27
1360
+ 3.5
1361
+ D 3.0
1362
+ 2.5
1363
+ T = 87.89
1364
+ T = 88.90
1365
+ T = 90.31
1366
+ 3.5
1367
+ D 3.0
1368
+ 2.5
1369
+ T = 90.52
1370
+ T = 99.70
1371
+ T = 121.4
1372
+ 2.5
1373
+ 3.0
1374
+ 3.5
1375
+ 2.5
1376
+ 3.0
1377
+ 3.5
1378
+ 2.5
1379
+ 3.0
1380
+ 3.5
1381
+ 1
1382
+ I19
1383
+ Table 2: Details on the RPOs and POs found using initial conditions from the DManD model.
1384
+ The first 9 solutions are new and the last 3 had previously been found. “Label" indicates the
1385
+ label in Fig. 13b, 𝜎𝑧 corresponds to the phase-shift in 𝑧, 𝑇 is the period of the orbit, and “Error"
1386
+ is ||shifted final state − initial state||/||initial state||, which is the same error as in Viswanath
1387
+ (2007).
1388
+ Label
1389
+ 1
1390
+ 2
1391
+ 3
1392
+ 4
1393
+ 5
1394
+ 6
1395
+ 7
1396
+ 8
1397
+ 9
1398
+ 10
1399
+ 11
1400
+ 12
1401
+ 𝜎𝑧
1402
+ 1.91e-1 -9.66e-2 -1.77e-3 1.15e-1 -9.21e-3 -1.90e-1 -1.28e-2 -1.19e-1 -5.63e-5 4.64e-14 2.17e-14 2.73e-13
1403
+ 𝑇
1404
+ 37.94
1405
+ 84.25
1406
+ 91.29
1407
+ 82.07
1408
+ 74.14
1409
+ 41.24
1410
+ 110.67
1411
+ 83.31
1412
+ 64.64
1413
+ 19.06
1414
+ 68.07
1415
+ 75.35
1416
+ Error
1417
+ 2.23e-3
1418
+ 1.01e-3
1419
+ 3.92e-3
1420
+ 2.84e-3
1421
+ 1.87e-3
1422
+ 5.26e-4
1423
+ 1.25e-3
1424
+ 1.13e-3
1425
+ 2.25e-3
1426
+ 1.57e-4
1427
+ 2.55e-4
1428
+ 1.07e-4
1429
+ a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)a)
1430
+ b)
1431
+ b)
1432
+ b)
1433
+ b)
1434
+ b)
1435
+ b)
1436
+ b)
1437
+ b)
1438
+ b)
1439
+ b)
1440
+ b)
1441
+ b)
1442
+ b)
1443
+ b)
1444
+ b)
1445
+ b)
1446
+ b)
1447
+ Figure 13: (a) Periodic orbits found in DManD models at 𝑑ℎ = 18 that converged to the (b)
1448
+ periodic orbits found in the DNS. Each of the colors corresponds to a one of the new solutions in
1449
+ Table 2. The blue curve at the back is a long trajectory of the DNS for comparison.
1450
+ We then took the initial conditions and periods for converged periodic orbits and decoded and
1451
+ upsampled them onto a 48 × 49 × 48 grid. We performed this upsampling because Viswanath
1452
+ (2007) reported that solutions on the coarser grid can be computational artifacts. Finally, we put
1453
+ these new initial conditions into Channelflow and ran another Newton search for 100 iterations.
1454
+ This procedure resulted in us finding 9 new RPOs and 3 existing POs, the details of which we
1455
+ include in Table 2.
1456
+ In Fig. 13a we show the new RPOs in the DManD model and in Fig. 13b we show what they
1457
+ converged to after putting them into the Channelflow Newton solver as initial guesses. Again,
1458
+ many of the RPOs end up following a similar path through this state space, with the biggest
1459
+ exceptions being RPO1 and RPO6, which converged to low-power input solutions. It is worth
1460
+ noting that this worked well, considering that the DManD initial conditions are POD coefficients
1461
+ from a model trained using data on a coarser grid than used to search for these solutions. We
1462
+ have described a new method to rapidly find new ECS, wherein an accurate low-dimensional
1463
+ model, like the DManD model presented here, is used to quickly perform a large number of ECS
1464
+ searches in the model and then these solutions can be fine tuned in the full simulation to land on
1465
+ new solutions.
1466
+ 4. Conclusion
1467
+ In the present work we described a data-driven manifold dynamics method (DManD) and
1468
+ applied it for accurate modeling of MFU PCF with far fewer degrees of freedom (O(10)) than
1469
+
1470
+ 4.0
1471
+ 1
1472
+ DManD
1473
+ Solution
1474
+ 2
1475
+ 3.5
1476
+ 3
1477
+ 4
1478
+ 3.0
1479
+ D
1480
+ 5
1481
+ 6
1482
+ 2.5
1483
+ 7
1484
+ 8
1485
+ 2.0
1486
+ D
1487
+ 9
1488
+ 2.0
1489
+ 2.5
1490
+ 3.0
1491
+ 3.5
1492
+ 2.0
1493
+ 2.5
1494
+ 3.0
1495
+ 3.5
1496
+ 120
1497
+ required for the DNS (O(105)). The DManD method consists of first finding a low-dimensional
1498
+ parameterization of the manifold on which data lies, and then discovering an ODE to evolve this
1499
+ low-dimensional state representation forward in time. In both cases we use NNs to approximate
1500
+ these functions from data. We find that an extremely low-dimensional parameterization of this
1501
+ manifold can be found using a hybrid autoencoder approach that corrects upon POD coefficients.
1502
+ Then, we use stabilized neural ODEs to accurately evolve the low-dimensional state forward in
1503
+ time.
1504
+ The DManD model captures the self-sustaining process and accurately tracks trajectories and
1505
+ the temporal autocorrelation over short time horizons. For DManD models with 𝑑ℎ > 10 we
1506
+ found excellent agreement between the model and the DNS in computing the mean-squared POD
1507
+ coefficient amplitude, the Reynolds stress, and the joint PDF of power input vs. dissipation. For
1508
+ comparison, we showed that a POD-Galerkin model requires ∼ 2000 degrees of freedom to get
1509
+ similar performance in matching the mean-squared POD coefficient amplitudes. Finally, we used
1510
+ the DManD model at 𝑑ℎ = 18 for PO searches. Using a set of existing POs, we successfully
1511
+ landed on nearby POs in the model. Finally, we found 9 previously undiscovered RPOs by first
1512
+ finding solutions in the DManD model and then using those as initial guesses to search in the full
1513
+ DNS.
1514
+ The results reported here have both fundamental and technological importance. At the
1515
+ fundamental level they indicate that, the true dimension of the dynamics of a turbulent flow
1516
+ can be orders of magnitude smaller than the number of degrees of freedom required for a fully-
1517
+ resolved simulation. Technologically this point is important because it may enable, for example,
1518
+ highly sophisticated model-based nonlinear control algorithms to be used: Determining the control
1519
+ strategy from the low-dimensional DManD model rather than a full-scale DNS, and applying it to
1520
+ the full flow will speed up both learning and implementing a control policy (Zeng et al. 2022a,b).
1521
+ This work was supported by AFOSR FA9550-18-1-0174 and ONR N00014-18-1-2865 (Van-
1522
+ nevar Bush Faculty Fellowship).
1523
+ REFERENCES
1524
+ Arndt, R. E. A., Long, D. F. & Glauser, M. N. 1997 The proper orthogonal decomposition of pressure
1525
+ fluctuations surrounding a turbulent jet. Journal of Fluid Mechanics 340, 1–33.
1526
+ Ball, K. S., Sirovich, L. & Keefe, L. R. 1991 Dynamical eigenfunction decomposition of turbulent channel
1527
+ flow. International Journal for Numerical Methods in Fluids 12 (6), 585–604.
1528
+ Borrelli, Giuseppe, Guastoni, Luca, Eivazi, Hamidreza, Schlatter, Philipp & Vinuesa, Ricardo
1529
+ 2022 Predicting the temporal dynamics of turbulent channels through deep learning. International
1530
+ Journal of Heat and Fluid Flow 96, 109010.
1531
+ Budanur, Nazmi Burak, Borrero-Echeverry, Daniel & Cvitanović, Predrag 2015a Periodic orbit
1532
+ analysis of a system with continuous symmetry - a tutorial. Chaos: An Interdisciplinary Journal of
1533
+ Nonlinear Science 25 (7), 073112.
1534
+ Budanur, Nazmi Burak, Cvitanović, Predrag, Davidchack, Ruslan L. & Siminos, Evangelos 2015b
1535
+ Reduction of SO(2) symmetry for spatially extended dynamical systems. Physical Review Letters
1536
+ 114 (8), 1–5.
1537
+ Chen, Ricky T. Q., Rubanova, Yulia, Bettencourt, Jesse & Duvenaud, David 2019 Neural ordinary
1538
+ differential equations. arXiv preprint , arXiv: 1806.07366.
1539
+ Chollet, François & others 2015 Keras. https://keras.io.
1540
+ Coifman, Ronald R., Lafon, Stéphane, Lee, A. B., Maggioni, Mauro, Nadler, Boaz, Warner,
1541
+ Frederick & Zucker, Steven W. 2005 Geometric diffusions as a tool for harmonic analysis and
1542
+ structure definition of data: diffusion maps. Proceedings of the National Academy of Sciences of the
1543
+ United States of America 102 21, 7426–31.
1544
+ Cvitanović, P., Artuso, R., Mainieri, R., Tanner, G. & Vattay, G. 2016 Chaos: Classical and Quantum.
1545
+ Copenhagen: Niels Bohr Inst.
1546
+ Eivazi, Hamidreza, Guastoni, Luca, Schlatter, Philipp, Azizpour, Hossein & Vinuesa, Ricardo 2021
1547
+
1548
+ 21
1549
+ Recurrent neural networks and Koopman-based frameworks for temporal predictions in a low-order
1550
+ model of turbulence. International Journal of Heat and Fluid Flow 90, 108816.
1551
+ Eivazi, Hamidreza, Veisi, Hadi, Naderi, Mohammad Hossein & Esfahanian, Vahid 2020 Deep neural
1552
+ networks for nonlinear model order reduction of unsteady flows. Physics of Fluids 32 (10), 105104.
1553
+ Floryan, Daniel & Graham, Michael D. 2022 Data-driven discovery of intrinsic dynamics. Nature
1554
+ Machine Intelligence 4 (12), 1113–1120.
1555
+ Foias, C., Jolly, M. S., Kevrekidis, I. G., Sell, G. R. & Titi, E. S. 1988 On the computation of inertial
1556
+ manifolds. Physics Letters A 131 (7-8), 433–436.
1557
+ García-Archilla, Bosco, Novo, Julia & Titi, Edriss S. 1998 Postprocessing the galerkin method: a novel
1558
+ approach to approximate inertial manifolds. SIAM Journal on Numerical Analysis 35 (3), 941–972.
1559
+ Gibson, John Francis 2002 Dynamical systems models of wall-bounded, shear-flow turbulence. PhD
1560
+ thesis, Cornell University, New York.
1561
+ Gibson, J. F. 2012 Channelflow: a spectral Navier-Stokes simulator in C++. University of New Hampshire
1562
+ (July), 1–41.
1563
+ Gibson, J. F., F. Reetz, S. Azimi, Ferraro, A., Kreilos, T., Schrobsdorff, H., Farano, M., Yesil, A. F.,
1564
+ Schütz, S. S., Culpo, M. & Schneider, T. M. 2021 Channelflow 2.0, arXiv: channelflow.ch.
1565
+ Gibson, J. F., Halcrow, J. & Cvitanović, P. 2008a Visualizing the geometry of state space in plane Couette
1566
+ flow. Journal of Fluid Mechanics 611 (1987), 107–130.
1567
+ Gibson, J. F., Halcrow, J., Cvitanović, P. & Viswanath, Divakar 2008b Unstable periodic orbits and
1568
+ heteroclinic connections in plane couette flow. arXiv preprint , arXiv: 0810.1974.
1569
+ Graham, Michael D. & Floryan, Daniel 2021 Exact coherent states and the nonlinear dynamics of
1570
+ wall-bounded turbulent flows. Annual Review of Fluid Mechanics 53 (1), 227–253.
1571
+ Graham, M. D., Steen, P. H. & Titi, E. S. 1993 Computational efficiency and approximate inertial manifolds
1572
+ for a Bénard convection system. Journal of Nonlinear Science 3 (1), 153–167.
1573
+ Hamilton, James, Kim, John & Waleffe, Fabian 1995 Regeneration mechanisms of near-wall turbulence
1574
+ structures. Journal of Fluid Mechanics 287, 317–348.
1575
+ Hasegawa, Kazuto, Fukami, Kai, Murata, Takaaki & Fukagata, Koji 2020a CNN-LSTM based
1576
+ reduced order modeling of two-dimensional unsteady flows around a circular cylinder at different
1577
+ reynolds numbers. Fluid Dynamics Research 52 (6), 065501.
1578
+ Hasegawa, Kazuto, Fukami, Kai, Murata, Takaaki & Fukagata, Koji 2020b Machine-learning-based
1579
+ reduced-order modeling for unsteady flows around bluff bodies of various shapes. Theoretical and
1580
+ Computational Fluid Dynamics 34 (4), 367–383.
1581
+ Hinton, Geoffrey & Roweis, Sam 2003 Stochastic neighbor embedding. Advances in neural information
1582
+ processing systems 15, 833–840.
1583
+ Hinton, G. E. & Salakhutdinov, R. R. 2006 Reducing the dimensionality of data with neural networks.
1584
+ Science 313 (5786), 504–507.
1585
+ Holmes, P., Lumley, J.L., Berkooz, G. & Rowley, C.W. 2012 Turbulence, Coherent Structures, Dynamical
1586
+ Systems and Symmetry. Cambridge University Press.
1587
+ Hopf, Eberhard 1948 A mathematical example displaying features of turbulence. Communications on Pure
1588
+ and Applied Mathematics 1 (4), 303–322.
1589
+ Inubushi, Masanobu, Takehiro, Shin-ichi & Yamada, Michio 2015 Regeneration cycle and the covariant
1590
+ lyapunov vectors in a minimal wall turbulence. Phys. Rev. E 92, 023022.
1591
+ Jiménez, Javier & Moin, Parviz 1991 The minimal flow unit in near-wall turbulence. Journal of Fluid
1592
+ Mechanics 225, 213–240.
1593
+ Kawahara, Genta & Kida, Shigeo 2001 Periodic motion embedded in plane couette turbulence:
1594
+ regeneration cycle and burst. Journal of Fluid Mechanics 449, 291–300.
1595
+ Kawahara, Genta, Uhlmann, Markus & van Veen, Lennaert 2012 The significance of simple invariant
1596
+ solutions in turbulent flows. Annual Review of Fluid Mechanics 44 (1), 203–225.
1597
+ Kingma, Diederik P. & Ba, Jimmy Lei 2015 Adam: A method for stochastic optimization. 3rd International
1598
+ Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings pp. 1–15.
1599
+ Kleiser, L. & Schumann, U. 1980 Treatment of Incompressibility and Boundary Conditions in 3-D
1600
+ Numerical Spectral Simulations of Plane Channel Flows, pp. 165–173. Wiesbaden: Vieweg+Teubner
1601
+ Verlag.
1602
+ Lee, Kookjin & Carlberg, Kevin T. 2020 Model reduction of dynamical systems on nonlinear manifolds
1603
+ using deep convolutional autoencoders. Journal of Computational Physics 404, 108973.
1604
+ Ling, Haibin & Okada, Kazunori 2007 An efficient earth mover’s distance algorithm for robust histogram
1605
+ comparison. IEEE Transactions on Pattern Analysis and Machine Intelligence 29 (5), 840–853.
1606
+
1607
+ 22
1608
+ Linot, Alec J., Burby, Joshua W., Tang, Qi, Balaprakash, Prasanna, Graham, Michael D. & Maulik,
1609
+ Romit 2023 Stabilized neural ordinary differential equations for long-time forecasting of dynamical
1610
+ systems. Journal of Computational Physics 474, 111838.
1611
+ Linot, Alec J. & Graham, Michael D. 2020 Deep learning to discover and predict dynamics on an inertial
1612
+ manifold. Phys. Rev. E 101, 062209.
1613
+ Linot, Alec J. & Graham, Michael D. 2022 Data-driven reduced-order modeling of spatiotemporal chaos
1614
+ with neural ordinary differential equations. Chaos: An Interdisciplinary Journal of Nonlinear Science
1615
+ 32 (7), 073110.
1616
+ Milano, Michele & Koumoutsakos, Petros 2002 Neural network modeling for near wall turbulent flow.
1617
+ Journal of Computational Physics 182 (1), 1–26.
1618
+ Moehlis, Jeff, Faisst, Holger & Eckhardt, Bruno 2004 A low-dimensional model for turbulent shear
1619
+ flows. New Journal of Physics 6 (1), 56.
1620
+ Moehlis, J., Smith, T. R., Holmes, P. & Faisst, H. 2002 Models for turbulent plane couette flow using the
1621
+ proper orthogonal decomposition. Physics of Fluids 14 (7), 2493–2507.
1622
+ Moin, Parviz & Moser, Robert D. 1989 Characteristic-eddy decomposition of turbulence in a channel.
1623
+ Journal of Fluid Mechanics 200, 471–509.
1624
+ Murata, Takaaki, Fukami, Kai & Fukagata, Koji 2020 Nonlinear mode decomposition with
1625
+ convolutional neural networks for fluid dynamics. Journal of Fluid Mechanics 882, A13.
1626
+ Nagata, M. 1990 Three-dimensional finite-amplitude solutions in plane couette flow: Bifurcation from
1627
+ infinity. Journal of Fluid Mechanics 217, 519–527.
1628
+ Nair, Nirmal J. & Goza, Andres 2020 Leveraging reduced-order models for state estimation using deep
1629
+ learning. Journal of Fluid Mechanics 897, 1–13.
1630
+ Nakamura, Taichi, Fukami, Kai, Hasegawa, Kazuto, Nabae, Yusuke & Fukagata, Koji 2021
1631
+ Convolutional neural network and long short-term memory based reduced order surrogate for minimal
1632
+ turbulent channel flow. Physics of Fluids 33 (2), 025116.
1633
+ Page, Jacob, Brenner, Michael P. & Kerswell, Rich R. 2021 Revealing the state space of turbulence
1634
+ using machine learning. Phys. Rev. Fluids 6, 034402.
1635
+ Paszke, Adam, Gross, Sam, Massa, Francisco, Lerer, Adam, Bradbury, James, Chanan, Gregory,
1636
+ Killeen, Trevor, Lin, Zeming, Gimelshein, Natalia, Antiga, Luca, Desmaison, Alban, Kopf,
1637
+ Andreas, Yang, Edward, DeVito, Zachary, Raison, Martin, Tejani, Alykhan, Chilamkurthy,
1638
+ Sasank, Steiner, Benoit, Fang, Lu, Bai, Junjie & Chintala, Soumith 2019 Pytorch: An imperative
1639
+ style, high-performance deep learning library. In Advances in Neural Information Processing Systems
1640
+ 32, pp. 8024–8035. Curran Associates, Inc.
1641
+ Pérez De Jesús, Carlos E. & Graham, Michael D. 2022 Data-driven low-dimensional dynamic model
1642
+ of Kolmogorov flow. arXiv preprint , arXiv: 2210.16708.
1643
+ Peyret, R. 2002 Spectral Methods for Incompressible Viscous Flow. Springer New York.
1644
+ Portwood, Gavin D., Mitra, Peetak P., Ribeiro, Mateus Dias, Nguyen, Tan Minh, Nadiga,
1645
+ Balasubramanya T., Saenz, Juan A., Chertkov, Michael, Garg, Animesh, Anandkumar,
1646
+ Anima, Dengel, Andreas, Baraniuk, Richard & Schmidt, David P. 2019 Turbulence forecasting
1647
+ via neural ODE. arXiv preprint , arXiv: 1911.05180.
1648
+ Rempfer, Dietmar & Fasel, Hermann F. 1994 Evolution of three-dimensional coherent structures in a
1649
+ flat-plate boundary layer. Journal of Fluid Mechanics 260, 351–375.
1650
+ Rojas, Carlos J. G., Dengel, Andreas & Ribeiro, Mateus Dias 2021 Reduced-order model for fluid
1651
+ flows via neural ordinary differential equations. arXiv preprint , arXiv: 2102.02248.
1652
+ Roweis, Sam T. & Saul, Lawrence K. 2000 Nonlinear dimensionality reduction by locally linear
1653
+ embedding. Science 290 (5500), 2323–2326.
1654
+ Rubner, Y., Tomasi, C. & Guibas, L.J. 1998 A metric for distributions with applications to image databases.
1655
+ In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pp. 59–66.
1656
+ Schölkopf, Bernhard, Smola, Alexander & Müller, Klaus-Robert 1998 Nonlinear component
1657
+ analysis as a kernel eigenvalue problem. Neural Computation 10 (5), 1299–1319.
1658
+ Smith, Troy R., Moehlis, Jeff & Holmes, Philip 2005 Low-dimensional modelling of turbulence using
1659
+ the proper orthogonal decomposition: A tutorial. Nonlinear Dynamics 41 (1), 275–307.
1660
+ Spalart, Philippe R, Moser, Robert D & Rogers, Michael M 1991 Spectral methods for the Navier-
1661
+ Stokes equations with one infinite and two periodic directions. Journal of Computational Physics
1662
+ 96 (2), 297–324.
1663
+ Srinivasan, P. A., Guastoni, L., Azizpour, H., Schlatter, P. & Vinuesa, R. 2019 Predictions of turbulent
1664
+ shear flows using deep neural networks. Physical Review Fluids 4 (5), 1–14.
1665
+
1666
+ 23
1667
+ Takens, Floris 1981 Detecting strange attractors in turbulence. In Dynamical Systems and Turbulence,
1668
+ Warwick 1980 (ed. David Rand & Lai-Sang Young), pp. 366–381. Berlin, Heidelberg: Springer Berlin
1669
+ Heidelberg.
1670
+ Tenenbaum, Joshua B., de Silva, Vin & Langford, John C. 2000 A global geometric framework for
1671
+ nonlinear dimensionality reduction. Science 290 (5500), 2319–2323.
1672
+ Titi, Edriss S. 1990 On approximate Inertial Manifolds to the Navier-Stokes equations. Journal of
1673
+ Mathematical Analysis and Applications 149 (2), 540–557.
1674
+ Van Der Maaten, L J P, Postma, E O & Van Den Herik, H J 2009 Dimensionality Reduction: A
1675
+ Comparative Review. Journal of Machine Learning Research 10, 1–41.
1676
+ Virtanen, Pauli, Gommers, Ralf, Oliphant, Travis E., Haberland, Matt, Reddy, Tyler, Cournapeau,
1677
+ David, Burovski, Evgeni, Peterson, Pearu, Weckesser, Warren, Bright, Jonathan, van der
1678
+ Walt, Stéfan J., Brett, Matthew, Wilson, Joshua, Millman, K. Jarrod, Mayorov, Nikolay,
1679
+ Nelson, Andrew R. J., Jones, Eric, Kern, Robert, Larson, Eric, Carey, C J, Polat, İlhan,
1680
+ Feng, Yu, Moore, Eric W., VanderPlas, Jake, Laxalde, Denis, Perktold, Josef, Cimrman,
1681
+ Robert, Henriksen, Ian, Quintero, E. A., Harris, Charles R., Archibald, Anne M., Ribeiro,
1682
+ Antônio H., Pedregosa, Fabian, van Mulbregt, Paul & SciPy 1.0 Contributors 2020 SciPy
1683
+ 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods 17, 261–272.
1684
+ Viswanath, D. 2007 Recurrent motions within plane couette turbulence. Journal of Fluid Mechanics 580,
1685
+ 339–358.
1686
+ Vlachas, P.R., Pathak, J., Hunt, B.R., Sapsis, T.P., Girvan, M., Ott, E. & Koumoutsakos, P. 2020
1687
+ Backpropagation algorithms and reservoir computing in recurrent neural networks for the forecasting
1688
+ of complex spatiotemporal dynamics. Neural Networks 126, 191–217.
1689
+ Waleffe, Fabian 1997 On a self-sustaining process in shear flows. Physics of Fluids 9 (4), 883–900.
1690
+ Waleffe, Fabian 1998 Three-dimensional coherent states in plane shear flows. Physical Review Letters
1691
+ 81 (19), 4140–4143.
1692
+ Wan, Zhong Yi, Vlachas, Pantelis, Koumoutsakos, Petros & Sapsis, Themistoklis 2018 Data-assisted
1693
+ reduced-order modeling of extreme events in complex dynamical systems. PLoS ONE 13 (5), 1–22.
1694
+ Whitney, Hassler 1936 Differentiable Manifolds. Annals of Mathematics 37 (3), 645–680.
1695
+ Whitney, Hassler 1944 The self-intersections of a smooth n-manifold in 2n-space. Annals of Mathematics
1696
+ 45 (2), 220–246.
1697
+ Young, Charles D. & Graham, Michael D. 2022 Deep learning delay coordinate dynamics for chaotic
1698
+ attractors from partial observable data. arXiv preprint , arXiv: 2211.11061.
1699
+ Yu, Haijun, Tian, Xinyuan, E, Weinan & Li, Qianxiao 2021 Onsagernet: Learning stable and interpretable
1700
+ dynamics using a generalized onsager principle. Phys. Rev. Fluids 6, 114402.
1701
+ Zeng, Kevin, Linot, Alec & Graham, Michael D. 2022a Learning turbulence control strategies with
1702
+ data-driven reduced-order models and deep reinforcement learning. In 12th International Symposium
1703
+ on Turbulence and Shear Flow Phenomena (TSFP12) Osaka, Japan (Online), July 19-22, 2022.
1704
+ Zeng, Kevin, Linot, Alec J. & Graham, Michael D. 2022b Data-driven control of spatiotemporal chaos
1705
+ with reduced-order neural ODE-based models and reinforcement learning. Proceedings of the Royal
1706
+ Society A 478 (2267), 20220297.
1707
+
49E3T4oBgHgl3EQfpApV/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
4dE0T4oBgHgl3EQfvQGV/content/2301.02616v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ed30c91883bc01ed5abad2c5cd26a7a4921149bcda1bc814fcf26bd6541192ab
3
+ size 260980
4dFRT4oBgHgl3EQfpDe0/content/tmp_files/2301.13612v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
4dFRT4oBgHgl3EQfpDe0/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
5tAyT4oBgHgl3EQfQPbi/content/tmp_files/2301.00043v1.pdf.txt ADDED
@@ -0,0 +1,1077 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MNRAS 00, 1 (2022)
2
+ https://doi.org/10.1093/mnras/stac3602
3
+ The impact of the AGN and the torus properties on the evolution of spiral
4
+ galaxies
5
+ M. A. Abdulrahman , 1 ‹ K. A. K. Gadallah , 2 A. Ahmed 1 and M. S. Elna wa wy 1
6
+ 1 Astronomy, Space science and Meteorology Department, Faculty of Science, Cairo University, Egypt
7
+ 2 Astronomy and Meteorology Department, Faculty of Science, Al-Azhar University, Nasr city, 11884, Cairo, Egypt
8
+ Accepted 2022 December 3. Received 2022 December 3; in original form 2022 September 3
9
+ A B S T R A C T
10
+ For spiral galaxies, the active galactic nucleus (AGN) and some physical parameters that concern the host galaxy such as spiral
11
+ arm radius and density can play an important role in the morphological evolution of these galaxies. Considering the gravitational
12
+ effect of the central black hole as a feeding mechanism, the gas flows from spiral arms to the accretion disk. Accordingly, we
13
+ constructed our approach and derived an equation for the AGN luminosity that depends on parameters such as the black hole
14
+ mass and the spiral arm density. The galaxy samples were taken from a catalog of type 1 AGN from SDSS-DR7. In our model,
15
+ we present the relation between the AGN luminosity and the black hole mass depending on the abo v e physical parameters.
16
+ We also investigated the relation between the black hole mass and the star formation rate for the galaxy sample. The physical
17
+ properties of the torus, such as the spiral arm radius, density, the torus length, and the gas mass, and the star formation rate were
18
+ explained in terms of the variation of the AGN luminosity. These properties are more effective in the evolutionary scenario of
19
+ the spiral galaxy. Relative to the variation of the AGN luminosity, the evolutionary track is dif ferent based quantitati vely on the
20
+ star formation rate. In which the variation in the star formation rate is positively correlated with the AGN luminosity.
21
+ K ey words: galaxies: e volution – galaxies: active – galaxies: spiral – black hole physics.
22
+ 1 INTRODUCTION
23
+ The first active galactic nuclei (AGN) in nearby galaxies were
24
+ observed and described by Seyfert ( 1943 ) who found spiral galaxies
25
+ having stronger emission lines from their nucleus than usual and was
26
+ named Seyfert galaxies. Later observations have revealed many types
27
+ of AGN which exhibit different features such as seyfert type-1 and
28
+ seyfert type-2 (hereafter, type 1 and type 2, respectively). Based on
29
+ their optical emission lines, type 1 shows broad emission lines and
30
+ type 2 shows only narrow emission lines (Khachikian & Weedman
31
+ 1974 ). The standard unified model introduced by Antonucci ( 1993 )
32
+ defines the AGN as a central black hole (BH) surrounded by an
33
+ accretion disk and dusty torus.
34
+ The AGN luminosity should be controlled by its evolution which
35
+ depends on the feeding mechanism and available matter for feeding.
36
+ Concerning the feeding mechanism, Alonso, Coldwell & Lambas
37
+ ( 2014 ) found that barred active galaxies show an excess of nuclear
38
+ activity more than unbarred ones which refers to the importance
39
+ of bars in gas inflow to central regions. And Dubois et al. ( 2015 )
40
+ showed in their simulation that supernova (SN) feedback could alter
41
+ the evolution of central BH during galaxy formation, where in strong
42
+ SN feedback the energy released can destruct the dense clumps in the
43
+ galaxy core preventing the evolution of BH due to the lack of cold gas.
44
+ Concerning the available matter, Koss et al. ( 2021 ) found that AGN
45
+ in massive galaxies ( > 10 10 . 5 M ⊙) tends to have larger molecular
46
+ g as and g as fraction than inactive g alaxies. And Franceschini et al.
47
+ ⋆ E-mail: mohamedabdulazez9@gmail.com
48
+ ( 1999 ) found that the histories of the BH accretion rate and stellar
49
+ formation in host galaxies are similar. Also Heckman et al. ( 2004 )
50
+ studied the accretion-driven growth of BH at low redshift using type
51
+ 2 AGN and found that the BH of masses less than 10 8 M ⊙ that reside
52
+ in moderately massive galaxies have accretion rate time scale that is
53
+ comparable to the age of universe.
54
+ Since AGN is powered by a central BH which resides at a
55
+ host galaxy of certain physical properties such as stellar mass and
56
+ star formation rate, the evolution of AGN should be related to its
57
+ host galaxy. In the following we will present some previous works
58
+ on studying the relation between AGN and host galaxy either by
59
+ observ ational e vidence or hydrodynamical simulations.
60
+ The accretion process caused by the central BH, results in a nuclear
61
+ activity that appears as a feedback on the ambient medium. This
62
+ activity has different scenarios such as winds, radiation pressure, and
63
+ jets (Zubovas & King 2014a ; Fabian, Vasude v an & Gandhi 2008 ).
64
+ The model made by King & Pounds ( 2003 ) explained the energy
65
+ and momentum of large scale outflows. According to this model,
66
+ the AGN radiation pressure can launch a relativistic wind from very
67
+ close in, where outflow can emerge from a photosphere of a radius
68
+ a few tens of schwarzchild radius ( R s ) given by R s = 2 GM
69
+ c 2 , where
70
+ M is the BH mass. This wind shocks against the ambient medium
71
+ producing an outflow. When the BH reaches a critical mass given by
72
+ Zubovas & King ( 2014b ),
73
+ M σ ≈ 3 . 67 × 10 8 σ 4
74
+ 200 M ⊙,
75
+ where σ is the velocity dispersion in host galaxy spheroid, the AGN
76
+ outflows become energy driven and cannot be cooled any more which
77
+ can cause gas loss and consequently affects the star formation in the
78
+ © 2022 The Author(s)
79
+ Published by Oxford University Press on behalf of Royal Astronomical Society
80
+
81
+ 2
82
+ M. A. Abdulrahman et al.
83
+ MNRAS 00, 1 (2022)
84
+ host galaxy. Also Wagner et al. ( 2016 ) showed that the feedback
85
+ can be positive such as triggering star formation by using its energy
86
+ in pressure triggered collapse or ne gativ e such as quenching star
87
+ formation by the loss of material from the host galaxy.
88
+ On the other hand, time evolution simulations of the impact of
89
+ the jet emerging from AGN on the ambient interstellar medium
90
+ (ISM) made by Wagner, Bicknell & Umemura ( 2012 ) showed that
91
+ the effect of the jet depends on the size of the ambient clouds. For
92
+ large clouds ( > 50 pc ), it increases the star formation rate, but for
93
+ small clouds ( ∼10 pc ), it causes explosion of cloud and quenching
94
+ of star formation. So ne gativ e and positive feedback can co-exist
95
+ depending on the density of the ambient ISM. In the simulations
96
+ made by Dubois et al. ( 2013 ), the jet power can be considered as
97
+ a reason of the material loss which can transform the disk galaxy
98
+ into red elliptical galaxies by the quenching of star formation. Also
99
+ Mukherjee et al. ( 2018 ) performed simulations to study the effect
100
+ of a relativistic jet on the ambient gaseous disk and found that,
101
+ depending on jet power, the ISM density, and jet orientation, the star
102
+ formation can be enhanced or quenched. These authors found that jet
103
+ can contribute in increasing the velocity dispersion of ambient ISM.
104
+ Accordingly, the effect of AGN on its host galaxy is rele v ant to the
105
+ host galaxy properties.
106
+ Observational studies made by (Ferrarese & Merritt 2000 ; Geb-
107
+ hardt et al. 2000 ) revealed a correlation between the BH mass and
108
+ the velocity dispersion of the host bulge. Reines & Volonteri ( 2015 )
109
+ found a correlation between the BH mass and the stellar mass of
110
+ the host galaxy in which the BH mass increases with increasing the
111
+ stellar mass of its host galaxy. This is consistent with the observations
112
+ made by Bilata-Woldeyes et al. ( 2020 ) using data from BAT-SWIFT
113
+ to study the relation between morphology of the host galaxy and the
114
+ AGN properties such as the Eddington ratio and BH mass. In these
115
+ observations, the BH masses are larger in elliptical galaxies than in
116
+ spiral galaxies.
117
+ Recent studies made by Dittmann & Miller ( 2020 ) investigated
118
+ the growth of the central BH by assuming a merging scenario with
119
+ the compact objects formed in the accretion disk. The study by
120
+ Tartenas & Zubovas ( 2019 ) showed the feeding of the AGN by the
121
+ dynamical perturbations. It also showed that a collision between
122
+ circumnuclear ring and molecular cloud that can be an efficient
123
+ fueling mechanism, depending on the angle of collision. So the
124
+ fueling mechanism has a crucial role in driving the relationship
125
+ between the AGN and its host galaxy. As a way to further examine
126
+ how such a relation between the AGN and host galaxy exists,
127
+ Smethurst et al. ( 2016 ) studied the star formation history of type
128
+ 2 AGN and inactive galaxies. In this study, the gas reservoir in the
129
+ host galaxy is the main source for BH fueling as first examined and
130
+ mentioned by Magorrian et al. ( 1998 ), where both of nuclear activity
131
+ and star formation are related to the host galaxy gas reservoir.
132
+ To further understand this mutual effect between the AGN and
133
+ the host galaxy, Valentini et al. ( 2020 ) performed a simulation of the
134
+ galaxy disk to study gas accretion models. This simulation showed
135
+ how the AGN feedback on the multiphase of the ISM of hot and cold
136
+ phases can affect the co-evolution of the BH and its host galaxy and
137
+ found that the accretion of cold gas is more ef fecti ve in BH growth
138
+ rate than hot gas, and the gas accretion contributes in the BH growth
139
+ more than mergers with other BHs.
140
+ So in this paper we study the evolution of isolated spiral galaxies
141
+ in terms of AGN contribution by investigating the effect of AGN
142
+ on the spiral arms of spiral galaxies and how the spiral arms can
143
+ control the AGN activity as being a source of fueling of the central
144
+ BH. Depending on the fueling of AGN by driving the gas from the
145
+ spiral arm under the gravitational force of the central BH to derive
146
+ Figure 1. Schematic chart of a simple spiral galaxy having two arms.
147
+ an Equation for the AGN luminosity that depends on parameters
148
+ such as the BH mass and the spiral arm radius which is a measure
149
+ of the amount of gas in it. We also investigate the BH mass-AGN
150
+ luminosity relation and how it can be affected by changing physical
151
+ parameters such as the spiral arm radius and the torus length.
152
+ We hope to construct a physical scenario for how the AGN-galaxy
153
+ co-evolution works. This paper is presented as follows. The model
154
+ approach is explained in Sections 2 and the data of the galaxy sample
155
+ is given in Section 3 . Results and the discussions are presented
156
+ in Sections 4 and 5 , respectively, while the conclusion is given in
157
+ Section 6 . In this work, the cosmological parameters were taken
158
+ with H 0 = 70.0 km s −1 Mpc −1 , �m = 0.30, �λ = 0.70.
159
+ 2 MODEL APPROACH
160
+ For a spiral galaxy having a simple galactic disk with two arms, we
161
+ assume that most of the mass of the galactic matter is concentrated
162
+ within these dense arms while the rest of the galactic medium of the
163
+ host disk is very diffuse medium with very low density that can be
164
+ neglected. Based on the simple morphology of the unified model,
165
+ we assume that the torus is divided to internal and external sides
166
+ around the the AGN as shown in the schematic chart in Fig. 1 . In this
167
+ chart, the internal torus is facing the AGN while the external torus
168
+ represents the outer region of the galactic disk.
169
+ The internal torus region is assumed to partially contain these
170
+ arms. In which the feedback of the BH activity is more efficient on
171
+ the matter with these arms. In this context, Yu et al. ( 2022 ) found that
172
+ the spiral arm is efficient in transporting the gas to central region.
173
+ Accordingly, we assume that the material transfers in a conical shape
174
+ from the closest region within the internal torus into the accretion
175
+ disk through a spiral arm. This shape makes the spiral arm radius
176
+ varies from smaller at the contact point with the accretion disk to
177
+ larger radius within deeper internal region of the torus (neglecting
178
+ the environmental effects on the galactic outskirts).
179
+ For a particle of mass ( m ) in the accretion disk at a distance ( r )
180
+ from the central BH, it mo v es with a velocity ( υ) where its angular
181
+ momentum ( �) is given by,
182
+ � = m AD υAD r
183
+ (1)
184
+ where m AD refers to the mass of accretion disk.
185
+ Due to the gravitational effect of the BH, the particles around
186
+ the BH experience a torque pushing them inward into the BH. As
187
+ the matter in the accretion disk loses angular momentum the matter
188
+
189
+ External
190
+ torus
191
+ Internaltorus
192
+ RA
193
+ Accretiondisk
194
+ Black hole
195
+ rThe impact of the AGN and the torus properties on the evolution of spiral galaxies
196
+ 3
197
+ MNRAS 00, 1 (2022)
198
+ spirals into the BH and this gives the chance to the material in internal
199
+ torus to mo v e to accretion disk.
200
+ So we can assume that the change of angular momentum equals
201
+ the torque on the gas in the accretion disk. Accordingly, the change
202
+ of r with time is a consequence of the change of the gravitational
203
+ radius ( r g ) of the BH with time which indicates the growth of BH or
204
+ its evolution, hence,
205
+ dr
206
+ dt = dr g
207
+ dt =
208
+ � gas
209
+ m AD υAD
210
+ (2)
211
+ where � gas is the torque on the gas in the accretion disk.
212
+ Considering a gaseous disk where the gravitational force is due to
213
+ a central BH, the torque on gas at a radius ( r ), as given by Netzer
214
+ ( 2013 ), is:
215
+ � gas ( r) = ˙m ( GMr) 1 / 2 f ( r)
216
+ (3)
217
+ where, ˙m is the radial mass inflow rate, M is the BH mass, and f ( r ) =
218
+ 1 − ( r in / r ) 1/2 . The r in parameter represents the radius at which the
219
+ torque on gas is zero and it falls in a non-circular orbit into the BH
220
+ which is known as the inner-most stable circular orbit (ISCO).
221
+ Using r g = GM
222
+ c 2 where M , G and c are the BH mass, the gravita-
223
+ tional constant and the speed of light, respectively, we have
224
+ dr g
225
+ dt = G
226
+ c 2
227
+ dM
228
+ dt
229
+ (4)
230
+ It is known that the BH grows due to the accretion of matter with
231
+ taking into account that not all matter is accreted. A fraction ( δ) of
232
+ this matter does not accrete but it escapes due to the feedback of
233
+ the AGN such as radiation pressure and wind from accretion disk
234
+ (Zubovas & King 2014a ) and the star formation taking place in the
235
+ vicinity of BH (Dittmann & Miller 2020 ). So we can write the change
236
+ of the BH mass with time as,
237
+ dM
238
+ dt = (1 − δ) ˙M
239
+ (5)
240
+ where ˙M is the accretion rate given by ˙M = L
241
+ ηc 2 while L is the AGN
242
+ bolometric luminosity hereafter it is called L AGN , and η is the radiative
243
+ efficiency.
244
+ Then equation ( 5 ) becomes,
245
+ dM
246
+ dt = (1 − δ) L
247
+ ηc 2
248
+ (6)
249
+ From equations ( 2 ), (3), (4) and (6), the L AGN can be written as,
250
+ L AGN = ηc 4 ˙m ( GM r ) 1 / 2 f ( r )
251
+ G (1 − δ) m AD υAD
252
+ (7)
253
+ where a fast rotating BH has a spin parameter of 0.998, and r in =
254
+ 1.24 r g . It is possible to put r = nr g where n takes values of 20–
255
+ 40 indicating the mean disk size according to available data on
256
+ variability of high redshift luminous AGN (Netzer 2013 ).
257
+ By considering the conical shape of the spiral arm facing the
258
+ accretion disk, m AD can be written as,
259
+ dm AD
260
+ dt
261
+ = −ξm T
262
+ (8)
263
+ where m T is the mass of the internal torus, and ξ is the transfer
264
+ efficiency which could be thought as the fraction of medium
265
+ clumpness (clumpy or smooth).
266
+ m T = π
267
+ 3 R 2
268
+ A ρT l T
269
+ (9)
270
+ where R A is the spiral arm radius within the internal torus region
271
+ characterized by a very dense conical shape. The parameters ρT and
272
+ l T are the density within the torus and its length, respectively. By
273
+ assuming that the material is concentrated in the spiral arms at a
274
+ certain time (t), for which, ρT becomes mostly the arm density ( ρA ).
275
+ Substituting by (9) into (8), and integrating ρA w.r.t time, the mass
276
+ of accretion disk becomes
277
+ m AD = π
278
+ 6 ξR 2
279
+ A ρ2
280
+ A l T
281
+ (10)
282
+ Then the luminosity Equation in the present model becomes
283
+ L AGN = 5 . 156 × 10 25
284
+ ηn
285
+ 1
286
+ 2 f ( r ) M ˙m
287
+ υAD ξ(1 − δ) l T R 2
288
+ A ρ2
289
+ A
290
+ ( erg s −1 )
291
+ (11)
292
+ By setting δ = 0.1, ξ = 0.1, η = 0.1, n = 25, f ( r ) ∼ 0.777, υAD =
293
+ 10 8 m / s as standard values, equation ( 11 ) can be re-written as
294
+ L AGN = 2 . 792 × 10 18 M ˙m
295
+ l T R 2
296
+ A ρ2
297
+ A
298
+ ( erg s −1 )
299
+ (12)
300
+ Therefore, the ef fecti ve parameters on the AGN luminosity are the
301
+ mass inflow rate ( ˙m ), the internal torus length ( l T ), the spiral arm
302
+ radius ( R A ), the spiral arm density ( ρA ), and the BH mass ( M ).
303
+ 3 DATA
304
+ To compare with the calculations of equation ( 12 ) and in an attempt
305
+ to check the validity of our BH mass-luminosity relation, a sample
306
+ of the observational data was provided by a catalog of type 1 AGNs
307
+ from SDSS-DR7 (Oh et al. 2015a ). This catalog contains 5553 type
308
+ 1 AGN with a redshift of z ≤ 0.2. In which, galaxies were selected
309
+ after applying some classification criteria for data of OSSY catalog.
310
+ This catalog provides us with the logarithm of bolometric luminosity
311
+ of AGN ( L bol ) derived from a method developed by Heckman et al.
312
+ ( 2004 ) using the luminosity of the [O III ] λ5007 emission line L O III
313
+ as a tracer of nuclear activity where L bol ≈ 3500 L O III ( erg s −1 ). It
314
+ also provides us with the logarithm of the BH mass as derived by
315
+ Greene & Ho ( 2005 ) depending on the luminosity and line width of
316
+ the broad H α line. The bolometric luminosity ranges from 42.09 to
317
+ 46.77 erg s −1 in logarithmic scale. Also the BH masses ranges from
318
+ 6.13 to 9.29 M ⊙ in logarithmic scale.
319
+ As the galaxy selection is flux-limited, we have considered the
320
+ Malmquist bias as a selecting criterion of our galaxy sample where
321
+ the luminosity varies as a function of the redshift. In Fig. 2 , we plot
322
+ the bolometric luminosity across the redshift of 5553 galaxies as
323
+ shown in the top panel , then we employ a flux-limit cut off with
324
+ a minimum flux limit of 4 × 10 −12 ergs −1 cm −2 to estimate the
325
+ theoretical luminosity (solid line) using the formula of L = 4 πd L F
326
+ where d L is the luminosity distance. This luminosity distance was
327
+ estimated according to the analytical approximation considering the
328
+ case of flat cosmologies (Adachi & Kasai 2012 ). Accordingly, the
329
+ number of galaxies in the sample has been reduced to 4954 galaxies as
330
+ shown in the bottom panel . And the current range for the bolometric
331
+ luminosity becomes from 42.88 to 46.77 erg s −1 in logarithmic scale.
332
+ 4 RESULTS
333
+ In this Section, we present the results that show the impact of
334
+ the escaped fraction ( δ) and the clumpness fraction ( ξ) mentioned
335
+ abo v e in equation ( 11 ) on the AGN luminosity. Then we used also
336
+ equation ( 12 ) to show the effect of the spiral arm density ( ρA ) and
337
+ the radius ( R A ) on this luminosity. We showed the AGN luminosity
338
+ versus the BH mass using the observational data along with those of
339
+ equation ( 12 ). By using derived relations from previous observational
340
+ works of (Dom ´ınguez S ´anchez et al. 2012 ; Baron & M ´enard 2019 ;
341
+ Huang et al. 2012 ), we estimated the star formation rate, gas mass,
342
+
343
+ 4
344
+ M. A. Abdulrahman et al.
345
+ MNRAS 00, 1 (2022)
346
+ Figure 2. The bolometric luminosity of data sample (red dots) across redshift
347
+ before ( top panel ) and after ( bottom panel ) applying the Malmquist bias where
348
+ the solid line represents the flux-limit cut off of the galaxy sample.
349
+ and stellar mass of our sample the galaxies comparing the results
350
+ from both of the observed data and equation ( 12 ). Finally, we used
351
+ the results that we obtained in an attempt to put a suitable evolution
352
+ scenario of the spiral galaxy in terms of AGN.
353
+ 4.1 The AGN luminosity versus torus properties
354
+ Using equation ( 11 ) with assuming R A = 100 pc, ρA = 10 −10 kg m −3 ,
355
+ l T = 100 pc, ˙m = 0 . 1 M ⊙ yr −1 , Fig. 3 sho ws ho w the AGN
356
+ luminosity is affected by; the clumpness fraction of the medium
357
+ ( ξ) and the velocity of the material in the accretion disk ( υAD ) in
358
+ the top panels ( left and right, respectively ); the radiative efficiency
359
+ ( η) and the mean disk size ( n ) in the middle panels ( left and right,
360
+ respectively ); and the escaped fraction ( δ) in the bottom panel. For
361
+ each ef fecti ve parameter, this luminosity was estimated with keeping
362
+ the others constant. From this figure, the luminosity exponentially
363
+ decreases with increasing both �� and υAD , while it exponentially
364
+ increases with increasing both of η, n, and in a logarithmic way with
365
+ δ.
366
+ In order to see how the BH mass affect the relation between the
367
+ AGN luminosity and each physical parameter in equation ( 12 ), we
368
+ present its behaviour at three different BH masses ( LogM = 6, 7 and
369
+ 8 M ⊙) as shown in Fig. 4 ( the assumed value for each parameter
370
+ was chosen according to the best fit for the data sample shown later
371
+ in Fig. 5 ). From Fig. 4 , the luminosity has a similar trends showing
372
+ an exponential decrease with both of the R A , ρA , and l T while it has
373
+ differently an exponential increase with ˙m .
374
+ The spiral arm radius and density of a galaxy can be considered as
375
+ a morphological parameters that demonstrate the host galaxy shape
376
+ where the spiral arm contains most of the amount of material within
377
+ torus. This matter is considered as the main feeding source for the
378
+ central BH which controls the AGN luminosity. For which, the effect
379
+ of both of spiral arm radius and density is shown in the top panels (left
380
+ and right, respectively) of Fig. 4 . Also Masoura et al. ( 2018 ) found
381
+ that the AGN luminosity, in a host galaxy, depends on the position
382
+ of this host galaxy from the main sequence line, depending on the
383
+ available gas of the host galaxy. Since the accreted matter from the
384
+ spiral arm travels through a path from spiral arm to the BH, the length
385
+ of this path should also control the produced AGN luminosity. This
386
+ path length is represented in our model by the internal torus length.
387
+ The left top panel of Fig. 4 shows its correlation with the luminosity
388
+ at three different BH masses, in which, the luminosity increases fast
389
+ at smaller lengths but it decreases slowly at longer lengths.
390
+ All of the abo v e physical parameters are related to the properties
391
+ of the galaxy disk but the mass inflow rate or the accretion rate of
392
+ the BH is related to the properties of the central BH. Its correlation
393
+ with luminosity, shown in the left bottom panel Fig. 4 , shows the
394
+ remarkable increase of the luminosity with the mass inflow rate.
395
+ This means that the luminosity doesn’t depend only on the amount
396
+ of the material flowing toward the BH but also the time taken by this
397
+ material to reach the accretion disk.
398
+ For the data sample, the top panel in Fig. 5 shows the result with
399
+ implying equation ( 12 ) and those of observed data fitted with a slope
400
+ of 1.005 ± 0.029, for which, the standard deviation was adapted to
401
+ x = 1, y = 0.86, giving a correlation coefficient of 0.49. To account
402
+ for the dispersion of the scattered points, we calculated the residual
403
+ which shows a normal Gaussian distribution as shown in the bottom
404
+ panel.
405
+ Using the data sample, we present some physical properties
406
+ concerning the host galaxy, such as the SFR, the stellar mass ( M stellar )
407
+ and the gas mass ( M gas ), to explain how these parameters can change
408
+ across redshift. The observed data of our galaxy sample provides
409
+ us with the H α emission line luminosity, L ( H α), which we used
410
+ to measure the SFR based on the SFR- L ( H α) dependence using
411
+ equation ( 1 ) given by Dom ´ınguez S ´anchez et al. ( 2012 ) where:
412
+ SFR (M ⊙ yr −1 ) = 7 . 9 × 10 −42 L (H α) ( erg s −1 )
413
+ To estimate the stellar mass as a function of BH mass, we re-
414
+ arrange equation ( 9 ) given by Baron & M ´enard ( 2019 ) to be as
415
+ follows:
416
+ Log
417
+ � M stel l ar
418
+ 10 11 M ⊙
419
+
420
+ =
421
+ Log( M BH
422
+ M ⊙ ) − (7 . 88 ± 0 . 13 )
423
+ (1 . 64 ± 0 . 18 )
424
+ By assuming that stars form in molecular clouds, the gas mass can
425
+ be measured using the HI 21 cm line. According to equation ( 1 ) given
426
+ by Huang et al. ( 2012 ) using the stellar mass of M stel l ar > 10 9 M ⊙
427
+ since our data range is 9.94–11.859 M ⊙ in log scale, the gas mass
428
+ can be calculated in terms of the stellar masss as follows:
429
+ Log ( M H I ) ≈ 0 . 276 Log( M stel l ar )
430
+ For a deeper understanding of the relation shown in Fig. 5 , we
431
+ further investigated the SFR distribution along this relation. From
432
+ Fig. 6 , we can see that at a certain BH mass, the AGN luminosity
433
+ decreases with decreasing the SFR. From equation ( 12 ), ρA can
434
+ be used as an indicator to the SFR. It is expected to have a
435
+ clumpy medium with increasing the density in the arm. When star
436
+ formation takes place, the ultraviolet radiation and the stellar wind
437
+
438
+ 47.
439
+ 46-
440
+ 45
441
+ Lbol
442
+ Log
443
+ 44-
444
+ 43-
445
+ 42
446
+ 0.00
447
+ 0.05
448
+ 0.10
449
+ 0.15
450
+ 0.20
451
+ Redshift
452
+ 48
453
+ 46
454
+ 44
455
+ 42-
456
+ 0.00
457
+ 0.05
458
+ 0.10
459
+ 0.15
460
+ 0.20
461
+ RedshiftThe impact of the AGN and the torus properties on the evolution of spiral galaxies
462
+ 5
463
+ MNRAS 00, 1 (2022)
464
+ Figure 3. The AGN luminosity distribution as a function of the clumpness fraction and the velocity in the accretion disk (left and right top panels, respectively),
465
+ the radiative efficiency and the mean disk size (left and right middle panels, respectively), and the escaped fraction from the BH accretion (bottom panel).
466
+ injected energy, caused by massive stars. These could be destructive
467
+ mechanisms to the molecular clouds leading to a decrease in the
468
+ ISM density or cloud dispersion (Grudi ´c et al. 2018 ; Gonz ´alez-
469
+ Samaniego & Vazquez-Semadeni 2020 ). Accordingly, the gas can
470
+ then flow easily to the central region leading to high AGN luminosity.
471
+ To implement the relation between SFR and AGN evolution, we
472
+ present the SFR variation with BH mass as shown in Fig. 7 . This
473
+ figure shows a linear relation with a slope of 0.86 and a correlation
474
+ coefficient of 0.49. It is obvious that this correlation fits well at
475
+ smaller BH masses but SFR shows flattening at higher BH masses,
476
+ deviating from the linear trend. Also from Figs 7 and 8 (bottom
477
+ panel), we can see a variation in the SFR versus the BH mass
478
+ and redshift, respectively. On average, the SFR is decreasing at low
479
+ redshift values and low BH mass. This variation points to the relation
480
+ between the BH mass and the SFR. In our work, we assume that the
481
+ spiral arm is the reservoir of gas needed for the BH feeding and
482
+ it can be seen from Fig. 7 along with Fig. 8 where gas in spiral
483
+ arm is consumed by both of star formation and the BH feeding.
484
+
485
+ 43.4
486
+ 47
487
+ 43.2 -
488
+ (s/B1a)
489
+ (s/810)
490
+ 46
491
+ 43.0-
492
+ Lbol
493
+ 45
494
+ 42.8
495
+ 42.6-
496
+ 42.4-
497
+ 43-
498
+ 0.05
499
+ 0.10
500
+ 0.15
501
+ 0.20
502
+ 0.25
503
+ 0.30
504
+ 0.35
505
+ 0.0
506
+ 2.0x105
507
+ 4.0x105
508
+ 6.0x105
509
+ 8.0x105
510
+ 1.0x106
511
+ UAD (km/s)
512
+ 43.6
513
+ 43.15
514
+ 43.4-
515
+ 43.10-
516
+ 43.2-
517
+ (s/8.1a)
518
+ 43.0-
519
+ (erg)
520
+ 43.05
521
+ 42.8-
522
+ 42.6-
523
+ 43.00
524
+ 42.4-
525
+ 42.2 -
526
+ 42.95
527
+ 42.0-
528
+ 41.8
529
+ 42.90
530
+ 0.00
531
+ 0.05
532
+ 0.10
533
+ 0.15
534
+ 0.20
535
+ 0.25
536
+ 0.30
537
+ 20
538
+ 25
539
+ 30
540
+ 35
541
+ 40
542
+ n
543
+ n
544
+ 43.7
545
+ 43.6
546
+ 43.5-
547
+ S
548
+ 43.4-
549
+ 43.3
550
+ 43.1
551
+ 43.0-
552
+ 42.9-
553
+ 0.0
554
+ 0.2
555
+ 0.4
556
+ 0.6
557
+ 0.86
558
+ M. A. Abdulrahman et al.
559
+ MNRAS 00, 1 (2022)
560
+ Figure 4. The bolometric luminosity of the AGN versus the spiral arm radius and the torus length (left and right top panels, respectively), and the spiral arm
561
+ density and the accretion rate (left and right bottom panels, respectiv ely). F or these physical parameters, the results were estimated at 3 different values of BH
562
+ mass of 6, 7, and 8 M ⊙ in log scale, assuming R A = 300 pc, ρA = 10 −12 kg m −3 , l T = 500 pc, ˙m = 0 . 1 M ⊙ yr −1 .
563
+ This gas consumption should decrease the gas content of the host
564
+ galaxy (Scoville et al. 2016 ; Genzel et al. 2015 ). In Fig. 8 (top and
565
+ middle panels), we can see that both of the gas mass and stellar mass
566
+ are following the same trend with decreasing redshift. The decrease
567
+ in each of them is slow and this can be used as evidence for the
568
+ BH feedback such as jets and winds which could make the growth
569
+ rate of the BH becomes slow. This can be used as evidence for the
570
+ stellar cycle where the gas is converted into stars and then returns
571
+ back through, for instance, the supernovae. Also previous work by
572
+ Tacconi et al. ( 2008 ) showed that the SFR / M gas ratio is relatively
573
+ constant.
574
+ 4.2 Evolution of a spiral galaxy
575
+ We have showed that AGN luminosity is affected by the integration
576
+ of all aforementioned physical parameters. Therefore, we tried to put
577
+ an evolution scenario for an isolated spiral galaxy depending on these
578
+ parameter and neglecting any merging and environmental effects.
579
+ For a spiral galaxy hosting a central BH, at first, the gas content
580
+ of spiral arms is condensed to form clumpy clouds of gas as possible
581
+ candidates for star formation to take place. If the gas gets condensed
582
+ the amount of material available for the BH accretion decreases,
583
+ hence the AGN luminosity decreases. This is why we get low AGN
584
+ luminosity for high values of spiral arm radius and density. But due
585
+
586
+ M= 6 Mo
587
+ 46
588
+ M= 6 Mo
589
+ 45
590
+ M= 7 Mo
591
+ M= 7 Mo
592
+ M= 8 Mo
593
+ 45
594
+ M= 8 Mo
595
+ 44
596
+ S
597
+ (erg/
598
+ (erg/s
599
+ 44
600
+ 43
601
+ 43
602
+ Log
603
+ 42
604
+ 42
605
+ 41-
606
+ 41-
607
+ 40-
608
+ 40
609
+ 0
610
+ 2000
611
+ 4000
612
+ 6000
613
+ 8000
614
+ 0
615
+ 2000
616
+ 4000
617
+ 6000
618
+ 8000
619
+ 10000
620
+ 12000
621
+ R (pc)
622
+ l1 (pc)
623
+ 52
624
+ M=6 Mo
625
+ 36
626
+ M= 7 Mo
627
+ M= 8 Mo
628
+ 50
629
+ S
630
+ 33
631
+ Lbol
632
+ 48
633
+ 30
634
+ Log
635
+ 46-
636
+ M= 6 Mo
637
+ 27
638
+ M= 7 Mo
639
+ -
640
+ M= 8 Mo
641
+ 44
642
+ 24
643
+ 0.0
644
+ 2.0x10-12
645
+ 4.0x10-12
646
+ 6.0x10-12
647
+ 8.0x10-12
648
+ 0.0
649
+ 5.0x1020
650
+ 1.0x1021
651
+ 1.98x102
652
+ m (kg/s)
653
+ Pa (kg/m3)The impact of the AGN and the torus properties on the evolution of spiral galaxies
654
+ 7
655
+ MNRAS 00, 1 (2022)
656
+ Figure 5. The relation between the BH mass and the bolometric luminosity
657
+ of AGN for the observed data of the galaxy sample (top panel). The black line
658
+ is our results according to equation ( 12 ) and the blue line is the fit of observed
659
+ data. The regular residual of equation ( 12 ) showing a normal distribution
660
+ (bottom panel).
661
+ Figure 6. The BH mass and AGN luminosity against star formation rate
662
+ distribution.
663
+ Figure 7. The star formation rate variation with BH mass.
664
+ to the star formation the luminosity emitted from the disk should be
665
+ high.
666
+ As time goes and due to galaxy rotation and star formation, the
667
+ gas in the spiral arms became dispersed which make it easy for the
668
+ central BH to pull it producing high AGN luminosity. This is why
669
+ we get high AGN luminosity for small values of spiral arm radius
670
+ and density.
671
+ The accretion process has a time which is determined by the
672
+ internal torus length and also by the accretion rate of the central
673
+ BH. During the gas journey from the spiral arm to the central BH, it
674
+ travels a certain path which is considered as the internal torus length.
675
+ In our approach, we assumed that material is transferred through a
676
+ conical path which can be representative of unbarred galaxy. This
677
+ path length is in rele v ant to the galaxy size and also the existence
678
+ of its bar which should alter the AGN luminosity. According to
679
+ Alonso et al. ( 2014 ) who found that among their sample which
680
+ includes barred and unbarred AGN, the barred galaxies exhibit a
681
+ higher nuclear activity than unbarred ones. Also the length of the
682
+ internal torus would affect the AGN luminosity and the activity time
683
+ of AGN. Using the magnetic-hydrodynamical simulations, Rosas-
684
+ Gue v ara et al. ( 2022 ) studied the evolution of barred massive disk
685
+ galaxies. These authors found that barred galaxies have lower star
686
+ formation rate and lower gas fraction compared to unbarred ones.
687
+ This indicates that the existence or absence of a bar may increase
688
+ or decrease the gas transport efficiency from the galactic disk to the
689
+ accretion disk.
690
+ Linking the correlation between the AGN luminosity, gas mass,
691
+ and SFR Shangguan et al. ( 2020 ), we can deduce the evolution
692
+ scenario based on these physical properties of our galaxy sample.
693
+ Using Figs 6 and 9 we can divide the evolution of a galaxy into three
694
+ phases .
695
+ Phase 1 is the period of time where the gas is still condensed at
696
+ spiral arms that we have a large gas mass galaxy at a certain SFR,
697
+ the gas is consumed to form stars and there is not enough gas to be
698
+ accreted by the central BH. For this period we should observe low
699
+ AGN luminosity, low BH mass, and low stellar mass.
700
+ Phase 2 is the period of time where star formation has taken place
701
+ that the stellar mass increases and the gas is dispersed whether by
702
+ the galaxy rotation or even by stellar wind or stellar cycle, so the
703
+ available gas mass for accretion increases, leading to an increase in
704
+ the BH mass and the AGN luminosity. This period may vary from
705
+ a galaxy to another depending on its size and morphology (there
706
+ is a bar or not), where Alonso et al. ( 2014 ), Kim & Choi ( 2020 )
707
+ showed that the nuclear activity is higher in galaxies having a bar,
708
+
709
+ 47
710
+ 46-
711
+ 45
712
+ Lbol
713
+ Log
714
+ 44
715
+ 43
716
+ equation(12)result
717
+ Fit of observation
718
+ 6.0
719
+ 6.5
720
+ 7.0
721
+ 7.5
722
+ 8.0
723
+ 8.5
724
+ 9.0
725
+ 9.5
726
+ Log M (Mo)
727
+ 1000
728
+ 800-
729
+ 600
730
+ Counts
731
+ 400
732
+ 200
733
+ 0
734
+ -2
735
+ -1
736
+ 0
737
+ 1
738
+ 2
739
+ Regular residual2
740
+ 46.5
741
+ 1.5
742
+ 46.0
743
+ S
744
+ 1
745
+ 45.5
746
+ 1007
747
+ 0.5
748
+ Log
749
+ 0
750
+ SFR
751
+ 44.5
752
+ 44.0
753
+ -1
754
+ 43.5
755
+ 43.0
756
+ -1.5
757
+ 6.5
758
+ 7.0
759
+ 7.5
760
+ 8.0
761
+ 8.5
762
+ 9.0
763
+ Log M
764
+ (Mo)3
765
+ 2
766
+ SFR (Mo-yr
767
+ 0
768
+ -1
769
+ -2
770
+ Fitof observation
771
+ 5
772
+ 7
773
+ 9
774
+ 10
775
+ Log M (M)8
776
+ M. A. Abdulrahman et al.
777
+ MNRAS 00, 1 (2022)
778
+ Figure 8. The gas mass (top panel), stellar mass (middle panel), and star
779
+ formation rate (bottom panel) variation with redshift.
780
+ referring to the vital role of the bar in transporting the gas to the
781
+ central region. Generally, in this phase and through a transition for
782
+ the total luminosity of the g alaxy; the g alactic disk luminosity may
783
+ decrease gradually or becomes constant depending on the value of
784
+ SFR and the AGN luminosity starts to increase. This phase can also
785
+ be interpreted in terms of a study done by Zewdie et al. ( 2020 ) using
786
+ SDSS MPA-JHU catalogue with the stellar mass range of LogM ∗ =
787
+ 10 . 73 −11 . 03 M ⊙. Using BPT-diagram, these authors found that the
788
+ AGN, in this stellar mass range, have lower star formation rates than
789
+ Figure 9. The relation between SFR, BH mass, and gas mass.
790
+ Figure 10. The 3 evolutionary phases of the spiral galaxy. In case of effect
791
+ of AGN feedback (positive) on SFR (orange circles) and without the effect
792
+ of AGN on SFR (black circles) feedback.
793
+ star-forming galaxies, and galaxies in this range mo v e from the blue
794
+ cloud to the red sequence.
795
+ Phase 3 is the period of time where the accretion rate of
796
+ central BH increases due to the large amount of gas coming from
797
+ spiral arms. Accordingly, a small amount of gas remains in the
798
+ spiral arms for star formation. So what we will observe is high
799
+ stellar mass and low galactic disk luminosity but high AGN
800
+ luminosity.
801
+ In Fig. 10 , we summarize the 3 phases for the spiral galaxy
802
+ evolution. If the AGN has no effect on the SFR of the host galaxy
803
+ (black circles) therefore it has a nearly constant SFR, for example,
804
+ for Log SFR ∼ −0.6 M ⊙ yr −1 . The stellar mass increases, leading
805
+ to an increase in the galactic disk luminosity due to the stellar
806
+ luminosities superposition but the increase in the AGN luminosity
807
+ is very small, it changes slightly between Log L bol ∼ 44 and ∼
808
+ 44.5 erg s −1 . In contrast, if the AGN has a positive feedback on
809
+ the host galaxy, the SFR increases. A factor of 25 increase in SFR
810
+ is related to an increase of 2 order of magnitudes in the AGN
811
+ luminosity.
812
+
813
+ 3.4 -
814
+ 3.2 -
815
+ 3.0
816
+ W
817
+ Log
818
+ 2.8
819
+ 2.6 -
820
+ 0.00
821
+ 0.05
822
+ 0.10
823
+ 0.15
824
+ 0.20
825
+ Redshift
826
+ 12 -
827
+ (w)
828
+ 11
829
+ Log Mster
830
+ 10-
831
+ 0.00
832
+ 0.05
833
+ 0.10
834
+ 0.15
835
+ 0.20
836
+ Redshift
837
+ 3 -
838
+ 2 -
839
+ SFR
840
+ 0
841
+ Log
842
+ -1
843
+ -2 -
844
+ -3 -
845
+ 4
846
+ 0.00
847
+ 0.05
848
+ 0.10
849
+ 0.15
850
+ 0.20
851
+ Redshift9.0
852
+ 3.2
853
+ 8.5
854
+ 3.1
855
+ Log Mgas
856
+ M
857
+ Log
858
+ 7.5
859
+ (Mo)
860
+ 2.9
861
+ 7.0
862
+ 6.5
863
+ 2.8
864
+ 1.5
865
+ -1.0
866
+ -0.5
867
+ 0.0
868
+ 0.5
869
+ 1.0
870
+ 1.5
871
+ 2.0
872
+ Log SFR (Mo/yr)11.8
873
+ 46.5
874
+ 11.6
875
+ 46
876
+ 11.4
877
+ 45.5
878
+ 45
879
+ Log Lbol
880
+ 44.5
881
+ (erg/s)
882
+ 10.6
883
+ 44
884
+ 10.4
885
+ 43.5
886
+ 10.2
887
+ 10.0
888
+ 43
889
+ -1.5
890
+ -1.0
891
+ -0.5
892
+ 0.0
893
+ 0.5
894
+ 1.0
895
+ 1.5
896
+ 2.0
897
+ Log SFR (Mo/yr)The impact of the AGN and the torus properties on the evolution of spiral galaxies
898
+ 9
899
+ MNRAS 00, 1 (2022)
900
+ 5 DISCUSSION
901
+ According to the results presented in the section 4 , we find out that
902
+ the torus properties such as spiral arm radius, spiral arm density
903
+ and torus length can be used as indicators of the the evolution of
904
+ the AGN. Also the accretion rate by the BH doesn’t only depend
905
+ on the BH physical parameters such as its spin, but also on the
906
+ torus properties or the host galaxy properties. From our results we
907
+ conclude that the AGN can have an ef fecti ve role in the evolution
908
+ of its host galaxy and vice v ersa. F or each galaxy, the luminosity
909
+ increases as the spiral arm decreases, which means that the gas in
910
+ spiral arms is the main source for the AGN luminosity and changing
911
+ it leads to a change in the AGN luminosity. Also accretion of stars
912
+ onto supermassive BHs can occur. But to determine which is more
913
+ efficient in the BH growth, the gas accretion or the star accretion,
914
+ Pfister et al. ( 2021 ) treated the tidal disruption events caused by
915
+ star accretion in cosmological simulations. These authors found
916
+ that contribution from stars’ accretion is negligible compared to gas
917
+ accretion.
918
+ Ho we ver this decrease in spiral arm gas content is not only due to
919
+ the accretion by the BH but partially caused by the feeding process
920
+ of BH and partially by the star formation taking place in spiral arms.
921
+ The star formation rate is high when the gas mass is high in spiral
922
+ arms, but to ease the gas flow from the spiral arms to the central BH,
923
+ the gas should be dispersed and this is done by the consumption of
924
+ gas in the star formation process and the dispersion caused by the
925
+ stellar wind or any damping mechanism of the formed stars (Thomp-
926
+ son, Quataert & Murray 2005 ; Hayward & Hopkins 2017 ; Lupi
927
+ 2019 ).
928
+ The dispersion of gas which causes the gas to flow easily from
929
+ the spiral arm toward the BH leads to a high AGN luminosity. This
930
+ process is not continuous but occurs periodically depending on the
931
+ star formation rate at spiral arms and the amount of gas available for
932
+ it and it can be measured as a variability of AGN activity across the
933
+ galaxy life time. This variability could make the galaxy normal for a
934
+ period of time, and active for another period of time but, for further
935
+ investigation of this variability, time-dependent SEDs are needed
936
+ to be studied. Another discontinuity of this feeding process can be
937
+ caused by the interaction between AGN feedback (outflows) and the
938
+ material flowing from the spiral arm.
939
+ Due to the accretion process and feeding mechanism the spiral arm
940
+ could disappear in a short period of time. But the accretion process
941
+ is slowed down by the feedback coming from the AGN as its BH
942
+ reaches a critical mass calculated by Ishibashi & Fabian ( 2012 ). And
943
+ it is mentioned by Zubovas & King ( 2014b ) that the BH reaches a
944
+ critical mass in which the AGN begins to produce outflows and this
945
+ may be one of the reasons for slowing down the accretion process.
946
+ This leads to a slow increase in the AGN luminosity at higher BH
947
+ masses and seen as a flattening in Fig. 5 (top panel). This slow gas
948
+ consumption is also what might cause the spiral arm to be long
949
+ lived.
950
+ All of the abo v e results and discussion concern about an isolated
951
+ galaxy without taking into account the environment effects such
952
+ as merging of galaxies or the location of this galaxy in its cluster.
953
+ Mergers and location of host galaxy with respect to cluster center
954
+ can affect the time at which the gas of host galaxy disk is being
955
+ consumed due to the AGN activity but not the physical process
956
+ occurring between host galaxy disk and its BH. For example, wet
957
+ merger can quick the process of accretion by triggering the gas into
958
+ the BH. Also the motion of the galaxy can affect the life time of the
959
+ process which can be slowed down due to the ram pressure causing
960
+ loss of matter.
961
+ 6 CONCLUSIONS
962
+ In this work, we have seen how the AGN can affect the evolution of
963
+ the spiral galaxies and that our approach provides an evolutionary
964
+ track for the AGN or specifically for the spiral galaxies in terms
965
+ of their AGN evolution. This track begins with a BH of low mass,
966
+ feeding on the gas mass of the host galaxy. Through its evolution and
967
+ its consumption of gas mass, the luminosity increases then decreases
968
+ slightly.
969
+ If we focus on this evolutionary track for AGN, we can see that as
970
+ time evolves the luminosity decreases due to the decrease or the lack
971
+ of gas mass in spiral arm which is also consistent with the decrease
972
+ in spiral arm radius, and this results also were obtained by Masoura
973
+ et al. ( 2018 ) for types 1 and 2 where the X-ray luminosity was found
974
+ to decrease as the redshift decreases indicating a decrease in AGN
975
+ activity.
976
+ From the observed data, the gas mass decreases with decreasing
977
+ the redshift. Since the spiral arm density indicates the gas mass
978
+ within the spiral arm, hence the spiral arm density also decreases
979
+ with decreasing the redshift. So we can say that, during the evolution
980
+ of AGN in spiral galaxies, the gas in spiral arms is consumed in
981
+ feeding the central BH which indicates that the AGN is affecting
982
+ the morphology of spiral galaxies. Hence we can link the AGN
983
+ luminosity to the spiral arm radius or the gas mass in the spiral arm
984
+ and use equation ( 11 ) to get the morphology of distant active galaxies
985
+ through observing their luminosity and vice versa by assuming best
986
+ fit values for each parameter that give the observed AGN luminosity.
987
+ This evolution has some consequence in between such as the
988
+ variable appearance observed in AGN or the AGN variability. In
989
+ studies done by Oh et al. ( 2015b ) and Suh et al. ( 2015 ) for type 1
990
+ and type 2, the BH mass-luminosity relation was controlled by the
991
+ Eddington ratio which indicates a change in the accretion rate of
992
+ the central BH. This change can be explained in terms of our model
993
+ approach by considering the spiral arm radius or the gas content
994
+ which represents the gas reservoir for BH accretion rate. This also
995
+ shows that the accretion process of the gas in spiral arms is not
996
+ continuous but happens in phases or episodes of time. As mentioned
997
+ by Zubovas & King ( 2014b ) when the BH reaches a critical mass in
998
+ which the AGN begins to produce outflows, this may be one of the
999
+ causes for slowing down the accretion process. This slowing down
1000
+ is what causes the spiral arm to be long lived.
1001
+ ACKNOWLEDGEMENTS
1002
+ DATA AVAILABILITY
1003
+ No new data were generated or analysed in support of this research
1004
+ REFERENCES
1005
+ Adachi M., Kasai M., 2012, Progress Theor. Phys. , 127, 145
1006
+ Alonso S., Coldwell G., Lambas D. G., 2014, A&A , 572, A86
1007
+ Antonucci R., 1993, ARA&A , 31, 473
1008
+ Baron D., M ´enard B., 2019, MNRAS , 487, 3404
1009
+ Bilata-Wolde yes B., Po vi ´c M., Be yoro-Amado Z., Getachew-Woreta T.,
1010
+ Terefe S., 2020, preprint ( arXiv:2003.12416 )
1011
+ Dittmann A. J., Miller M. C., 2020, MNRAS , 493, 3732
1012
+ Dom ´ınguez S ´anchez H. et al., 2012, MNRAS , 426, 330
1013
+ Dubois Y., Gavazzi R., Peirani S., Silk J., 2013, MNRAS , 433, 3297
1014
+
1015
+ 10
1016
+ M. A. Abdulrahman et al.
1017
+ MNRAS 00, 1 (2022)
1018
+ Dubois Y ., V olonteri M., Silk J., Devriendt J., Slyz A., Teyssier R., 2015,
1019
+ MNRAS , 452, 1502
1020
+ Fabian A. C., Vasude v an R. V., Gandhi P., 2008, MNRAS , 385, L43
1021
+ Ferrarese L., Merritt D., 2000, ApJ , 539, L9
1022
+ Franceschini A., Hasinger G., Miyaji T., Malquori D., 1999, MNRAS , 310,
1023
+ L5
1024
+ Gebhardt K. et al., 2000, ApJ , 539, L13
1025
+ Genzel R. et al., 2015, ApJ , 800, 20
1026
+ Gonz ´alez-Samaniego A., Vazquez-Semadeni E., 2020, MNRAS , 499, 668
1027
+ Greene J. E., Ho L. C., 2005, ApJ , 630, 122
1028
+ Grudi ´c M. Y., Hopkins P. F., Faucher-Gigu `ere C.-A., Quataert E., Murray N.,
1029
+ Kere ˇs D., 2018, MNRAS , 475, 3511
1030
+ Hayward C. C., Hopkins P. F., 2017, MNRAS , 465, 1682
1031
+ Heckman T. M., Kauffmann G., Brinchmann J., Charlot S., Tremonti C.,
1032
+ White S. D. M., 2004, ApJ , 613, 109
1033
+ Huang S., Haynes M. P., Giovanelli R., Brinchmann J., 2012, ApJ , 756,
1034
+ 113
1035
+ Ishibashi W., Fabian A. C., 2012, MNRAS , 427, 2998
1036
+ Khachikian E. Y., Weedman D. W., 1974, ApJ , 192, 581
1037
+ Kim M., Choi Y .-Y ., 2020, ApJ , 901, L38
1038
+ King A. R., Pounds K. A., 2003, MNRAS , 345, 657
1039
+ Koss M. J. et al., 2021, ApJS , 252, 29
1040
+ Lupi A., 2019, MNRAS , 484, 1687
1041
+ Magorrian J. et al., 1998, AJ , 115, 2285
1042
+ Masoura V. A., Mountrichas G., Georgantopoulos I., Ruiz A., Magdis G.,
1043
+ Plionis M., 2018, A&A , 618, A31
1044
+ Mukherjee D., Bicknell G. V., Wagner A. Y., Sutherland R. S., Silk J., 2018,
1045
+ MNRAS , 479, 5544
1046
+ Netzer H., 2013, The Physics and Evolution of Active Galactic Nuclei.
1047
+ Cambridge Univ. Press, Cambridge
1048
+ Oh K., Yi S. K., Schawinski K., K oss M., T rakhtenbrot B., Soto K., 2015a,
1049
+ ApJS , 219, 1
1050
+ Oh K., Yi S. K., Schawinski K., Koss M., Trakhtenbrot B., Soto K., 2015b,
1051
+ ApJS , 219, 1
1052
+ Pfister H., Dai J. L., Volonteri M., Auchettl K., Trebitsch M., Ramirez-Ruiz
1053
+ E., 2021, MNRAS , 500, 3944
1054
+ Reines A. E., Volonteri M., 2015, ApJ , 813, 82
1055
+ Rosas-Gue v ara Y. et al., 2022, MNRAS , 512, 5339
1056
+ Scoville N. et al., 2016, ApJ , 820, 83
1057
+ Seyfert C. K., 1943, ApJ , 97, 28
1058
+ Shangguan J., Ho L. C., Bauer F. E., Wang R., Treister E., 2020, ApJ , 899,
1059
+ 112
1060
+ Smethurst R. J. et al., 2016, MNRAS , 463, 2986
1061
+ Suh H., Hasinger G., Steinhardt C., Silverman J. D., Schramm M., 2015, ApJ ,
1062
+ 815, 129
1063
+ Tacconi L. J. et al., 2008, ApJ , 680, 246
1064
+ Tartenas M., Zubovas K., 2019, MNRAS , 492, 603
1065
+ Thompson T. A., Quataert E., Murray N., 2005, ApJ , 630, 167
1066
+ Valentini M. et al., 2020, MNRAS , 491, 2779
1067
+ Wagner A. Y., Bicknell G. V., Umemura M., 2012, ApJ , 757, 136
1068
+ Wagner A. Y., Bicknell G. V., Umemura M., Sutherland R. S., Silk J., 2016,
1069
+ Astronomische Nachrichten , 337, 167
1070
+ Yu S.-Y. et al., 2022, A&A , 666, A175
1071
+ Zewdie D., Po vi ´c M., Arav ena M., Assef R. J., Gaulle A., 2020, MNRAS ,
1072
+ 498, 4345
1073
+ Zubovas K., King A. R., 2014a, MNRAS , 439, 400
1074
+ Zubovas K., King A. R., 2014b, MNRAS , 439, 400
1075
+ This paper has been typeset from a T E X/L
1076
+ A T E X file prepared by the author.
1077
+
5tAyT4oBgHgl3EQfQPbi/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
6dAzT4oBgHgl3EQfEfrh/content/tmp_files/2301.00996v1.pdf.txt ADDED
@@ -0,0 +1,1472 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Transactional Composition of Nonblocking Data
2
+ Structures
3
+ Wentao Cai, Haosen Wen, and Michael L. Scott
4
+ {wcai6,hwen5,scott}@ur.rochester.edu
5
+ University of Rochester
6
+ Augest 18, 2022
7
+ Abstract
8
+ This paper introduces nonblocking transaction composition (NBTC),
9
+ a new methodology for atomic composition of nonblocking operations on
10
+ concurrent data structures. Unlike previous software transactional mem-
11
+ ory (STM) approaches, NBTC leverages the linearizability of existing non-
12
+ blocking structures, reducing the number of memory accesses that must
13
+ be executed together, atomically, to only one per operation in most cases
14
+ (these are typically the linearizing instructions of the constituent opera-
15
+ tions).
16
+ Our obstruction-free implementation of NBTC, which we call Medley,
17
+ makes it easy to transform most nonblocking data structures into transac-
18
+ tional counterparts while preserving their nonblocking liveness and high
19
+ concurrency. In our experiments, Medley outperforms Lock-Free Trans-
20
+ actional Transform (LFTT), the fastest prior competing methodology, by
21
+ 40–170%. The marginal overhead of Medley’s transactional composition,
22
+ relative to separate operations performed in succession, is roughly 2.2×.
23
+ For persistent data structures, we observe that failure atomicity for
24
+ transactions can be achieved “almost for free” with epoch-based periodic
25
+ persistence. Toward that end, we integrate Medley with nbMontage, a
26
+ general system for periodically persistent data structures. The resulting
27
+ txMontage provides ACID transactions and achieves throughput up to
28
+ two orders of magnitude higher than that of the OneFile persistent STM
29
+ system.
30
+ 1
31
+ Introduction
32
+ Nonblocking concurrent data structures, first explored in the 1970s, remain an
33
+ active topic of research today. In such structures, there is no reachable state
34
+ of the system that can prevent an individual operation from making forward
35
+ progress. This liveness property is highly desirable in multi-threaded programs
36
+ that aim for high scalability and are sensitive to high tail latency caused by
37
+ inopportune preemption of resource-holding threads.
38
+ 1
39
+ arXiv:2301.00996v1 [cs.DC] 3 Jan 2023
40
+
41
+ Many multi-threaded systems, including those for finance, travel [30], ware-
42
+ house management [6], and databases in general [39], need to compose opera-
43
+ tions into transactions that occur in an all-or-nothing fashion (i.e., atomically).
44
+ Concurrent data structures, however, ensure atomicity only for individual oper-
45
+ ations; composing a transaction across operations requires nontrivial program-
46
+ ming effort and introduces high overhead. Preserving nonblocking liveness for
47
+ every transaction is even more difficult.
48
+ One potential solution can be found in software transactional memory (STM)
49
+ systems, which convert almost arbitrary sequential code into speculative trans-
50
+ actions. Several STM systems provide nonblocking progress [10, 19, 25, 26, 37].
51
+ Most instrument each memory access and arrange to restart operations that
52
+ conflict at the level of individual loads and stores.
53
+ The resulting program-
54
+ ming model is attractive, but the instrumentation typically imposes 3–10× over-
55
+ head [34, Sec. 9.2.3].
56
+ Inspired by STM, Spiegelman et al. [36] proposed transactional data struc-
57
+ ture libraries (TDSL), which introduce (blocking) transactions for certain hand-
58
+ modified concurrent data structures. By observing that reads need to be tracked
59
+ only on critical nodes whose updates may indicate semantic conflicts, TDSL re-
60
+ duces read set size and achieves better performance than general STMs.
61
+ Herlihy and Koskinen [18] proposed transactional boosting, a (blocking)
62
+ methodology that allows an STM system to incorporate operations on exist-
63
+ ing concurrent data structures. Using a system of semantic locks (e.g., with
64
+ one lock per key in a mapping), transactions arrange to execute concurrently
65
+ so long as their boosted operations are logically independent, regardless of
66
+ low-level conflicts. A transaction that restarts due to a semantic conflict (or
67
+ to a low-level conflict outside the boosted code) will roll back any already-
68
+ completed boosted operations by performing explicitly identified inverse opera-
69
+ tions. An insert(k,v) operation, for example, would be rolled back by performing
70
+ remove(k).Transactional boosting leverages the potential for high concurrency
71
+ in existing data structures, but is intrinsically lock-based, and is not fully gen-
72
+ eral: operations on a single-linked FIFO queue, for example, have no obvious
73
+ inverse.
74
+ In work concurrent to TDSL, Zhang et al. [43] proposed the Lock-Free Trans-
75
+ actional Transform (LFTT), a nonblocking methodology to compose nonblock-
76
+ ing data structures, based on the observation that only certain nodes—those
77
+ critical to transaction semantics—really matter in conflict management. Each
78
+ operation on an LFTT structure publishes, on every critical node, a descrip-
79
+ tion of the transaction of which it is a part, so that conflicting transactions
80
+ can see and help each other. A remove(7) operation, for example, would pub-
81
+ lish a description of its transaction on the node in its structure with key 7.
82
+ Initially, LFTT supported only static transactions, whose constituent opera-
83
+ tions were all known in advance. Subsequently, LaBorde et al. [23] proposed a
84
+ Dynamic Transactional Transform (DTT) that generalizes LFTT to dynamic
85
+ transactions (specified as lambda expressions). Concurrently, Elizarov et al. [8]
86
+ proposed LOFT which is similar to LFTT but avoids LFTT’s bug of duplicated
87
+ helping.
88
+ 2
89
+
90
+ Unfortunately, as in transactional boosting, the need to identify critical
91
+ nodes tends to limit LFTT and DTT to data structures representing sets and
92
+ mappings.
93
+ DTT’s publishing and helping mechanisms also require that the
94
+ “glue” code between operations be fully reentrant (to admit concurrent execu-
95
+ tion by helping threads [23]) and may result in redundant work when conflicts
96
+ arise.
97
+ Worse, for read-heavy workloads, LFTT and DTT require readers to
98
+ be visible to writers, introducing metadata updates that significantly increase
99
+ contention in the cache coherence protocol.
100
+ In our work, we propose NonBlocking Transaction Composition (NBTC),
101
+ a new methodology that can create transactional versions of a wide variety of
102
+ concurrent data structures while preserving nonblocking progress and incurring
103
+ significantly lower overhead than traditional STM. The intuition behind NBTC
104
+ is that in already nonblocking structures, only critical memory accesses—for
105
+ the most part, the linearizing load and compare-and-swap (CAS) instructions—
106
+ need to occur atomically, while most pre-linearization memory accesses can
107
+ safely be executed as they are encountered, and post-linearization accesses can
108
+ be postponed until after the transaction commits.
109
+ In comparison to STM, NBTC significantly reduces the number of mem-
110
+ ory accesses that must be instrumented—typically to only one per constituent
111
+ operation. Unlike transactional boosting and transactional transforms, NBTC
112
+ brings the focus back from semantics to low-level memory accesses, thereby
113
+ enabling mechanical transformation of existing structures and accommodating
114
+ almost arbitrary abstractions—much more than sets and mappings. NBTC also
115
+ supports dynamic transactions, invisible readers, and non-reentrant “glue” code
116
+ between the operations of a transaction. The one requirement for compatibility
117
+ is that the linearization points of constituent operations must be immediately
118
+ identifiable: each operation must be able to tell when it has linearized at run
119
+ time, without performing any additional shared-memory accesses. Most non-
120
+ blocking structures in the literature appear to meet this requirement.
121
+ To assess the practicality of NBTC, we have built an obstruction-free imple-
122
+ mentation, Medley, that uses a variant of Harris et al.’s multi-word CAS [16]
123
+ to execute the critical memory accesses of each transaction atomically, eagerly
124
+ resolving conflicting transactions as they are discovered. Using Medley, we have
125
+ created NBTC versions of Michael and Scott’s queue [29], Fraser’s skiplist [10],
126
+ the rotating skiplist of Dick et al. [7], Michael’s chained hash table [28], and
127
+ Natarajan and Mittal’s binary search tree [31]. All of the transformations were
128
+ straightforward.
129
+ In the traditional language of database transactions [15], Medley provides
130
+ isolation and consistency. Building on recent work on persistent memory, we
131
+ have also integrated Medley with the nbMontage system of Cai et al. [2] to
132
+ create a system, txMontage, that provides failure atomicity and durability as
133
+ well—i.e., full ACID transactions. Specifically, we leverage the epoch system of
134
+ nbMontage, which divides time into coarse-grain temporal intervals and recov-
135
+ ers, on failure, to the state of a recent epoch boundary. By folding a check of
136
+ the epoch number into its multi-word CAS, txMontage ensures that operations
137
+ of the same transaction always linearize in the same epoch, thereby obtaining
138
+ 3
139
+
140
+ failure atomicity and durability “almost for free.”
141
+ Summarizing contributions:
142
+ • (Section 2) We introduce nonblocking transaction composition (NBTC), a
143
+ new methodology with which to compose the operations of nonblocking data
144
+ structures.
145
+ • (Section 3) Deploying NBTC, we implement Medley, a general system for
146
+ transactional nonblocking structures. Medley’s easy-to-use API and mechan-
147
+ ical transform make it easy to convert compatible nonblocking structures to
148
+ transactional form.
149
+ • (Section 4) We integrate Medley with nbMontage to create txMontage, provid-
150
+ ing not only transactional isolation and consistency, but also failure atomicity
151
+ and durability.
152
+ • (Section 5) We argue that using NBTC and Medley, transactions composed
153
+ of nonblocking structures are nonblocking and strictly serializable. We also
154
+ argue that transactions with txMontage provide a persistent variant of strict
155
+ serializability analogous to the buffered durable linearizability of Izraelevitz
156
+ et al. [21].
157
+ • (Section 6) We present performance results, confirming that Medley imposes
158
+ relatively modest overhead and scales to large numbers of threads. Specif-
159
+ ically, Medley outperforms LFTT by 1.4× to 2.7× and outperforms TDSL
160
+ and the OneFile nonblocking STM [33] system by an order of magnitude. On
161
+ persistent memory, txMontage outperforms nonblocking persistent STM by
162
+ two orders of magnitude.
163
+ 2
164
+ Nonblocking Transaction Composition
165
+ Nonblocking transaction composition (NBTC) is a new methodology that fully
166
+ leverages the linearizability of nonblocking data structure operations. NBTC
167
+ obtains strict serializability by atomically performing only the critical memory
168
+ accesses of composed operations. It supports a large subset of the nonblocking
169
+ data structures in the literature (characterized more precisely below), preserving
170
+ the high concurrency and nonblocking liveness of the transformed structures.
171
+ 2.1
172
+ NBTC Composability
173
+ The key to NBTC composability is the immediately identifiable linearization
174
+ point. Specifically:
175
+ Definition 1. A data structure operation has an immediately identifiable lin-
176
+ earization point if:
177
+ 1. statically, we can identify every instruction that may potentially serve as
178
+ the operation’s linearization point. Such an instruction must be a load for a
179
+ read-only operation or a compare-and-swap (CAS) for an update operation;
180
+ 4
181
+
182
+ 2. dynamically, after executing a potentially linearizing instruction, we can de-
183
+ termine whether it was indeed the linearization point. A linearizing load has
184
+ to be determined before the operation returns; a linearizing CAS has to be
185
+ determined without performing any additional shared-memory accesses.
186
+ There can be more than one potential linearization point in the code of an
187
+ operation, but only one of them will constitute the linearization point in any
188
+ given invocation.
189
+ Definition 2. A nonblocking data structure is NBTC-composable if each of its
190
+ operations has an immediately identifiable linearization point.
191
+ While it may be possible to relax this definition, the current version accom-
192
+ modates a very large number of existing nonblocking structures.
193
+ 2.2
194
+ The Methodology
195
+ It is widely understood that most nonblocking operations comprise a “planning”
196
+ phase and a “cleanup” phase, separated by a linearizing instruction [12, 38].
197
+ Executing the planning phase does not commit the operation to success; cleanup,
198
+ if needed, can be performed by any thread. The basic strategy in NBTC is to
199
+ perform the planning for all constituent operations of the current transaction,
200
+ then linearize all those operations together, atomically, and finally perform all
201
+ cleanup. Our survey of existing data structures and composition patterns reveals
202
+ two principle complications with this strategy.
203
+ The first complication involves the notion of a publication point, where an
204
+ operation may become visible to other threads but not yet linearize. Because
205
+ publication can alter the behavior of other threads, it must generally (like a
206
+ linearization point) remain speculative until the entire transaction is ready to
207
+ commit. An example can be seen in the binary search tree of Natarajan and
208
+ Mittal [31], where an update operation o may perform a CAS that publishes its
209
+ intent to linearize soon but not quite yet. After this publication point, either o
210
+ itself or any other update that encounters the publication notice may attempt
211
+ to linearize o (in the interest of performance, a read operation will ignore it).
212
+ Notably, CAS instructions that serve to help other (already linearized) opera-
213
+ tions, without revealing the nature of the current operation, need not count as
214
+ publication.
215
+ The second complication arises when a transaction, t, performs two or more
216
+ operations on the same data structure and one of the later operations (call it
217
+ o2) depends on the outcome of an earlier operation (call it o1). Here the thread
218
+ executing t must proceed as if o1 has completed, but other threads must ignore
219
+ it. If o1 requires cleanup (something that NBTC will normally delay until after
220
+ transaction commit), o2 may need to help o1 before it can proceed, while other
221
+ transactions should not even be aware of o1’s existence.
222
+ Both complicating cases can be handled by introducing the notion of a spec-
223
+ ulation interval in which CAS instructions must be completed together for an
224
+ 5
225
+
226
+ operation to take effect as part of a transaction. This is similar to the CAS ex-
227
+ ecutor phase in a normalized nonblocking data structure [38], but not the same,
228
+ largely due to the second complication. For an operation that becomes visible
229
+ before its linearization point, it suffices to include in the speculation interval all
230
+ CAS operations between the publication and linearization points, inclusive. For
231
+ an operation o2 that needs to see an earlier operation o1 in the same transaction,
232
+ it suffices to track the transaction’s writes and to start o2’s speculation interval
233
+ no later than the first instruction that accesses a location written by o1.
234
+ Definition 3. A bit more precise, we say
235
+ • A CAS instruction in operation o of thread t in history H is benign if there
236
+ is no extension H′ of H such that t executes no more instructions in H′ and
237
+ yet o linearizes in H′ nonetheless.
238
+ • The first CAS instruction of o that is not benign is o’s publication point (this
239
+ will often be the same as its linearization point).
240
+ • The speculation interval of o begins either at the publication point or at the
241
+ first instruction that sees a value speculatively written by some earlier opera-
242
+ tion in the same transaction (whichever comes first) and extends through o’s
243
+ linearization point.
244
+ • A load in a read-only operation is critical if it is the immediately identifiable
245
+ linearization point of the operation. A CAS in an update operation is critical
246
+ if it lies in the speculation interval.
247
+ Without loss of generality, we assume that all updates to shared memory
248
+ (other than initialization of objects not yet visible to other threads) are effected
249
+ via CAS.
250
+ Given these definitions, the NBTC methodology is straightforward: To atom-
251
+ ically execute a set of operations on NBTC-composable data structures, we
252
+ transform every operation such that (1) instructions prior to the speculation
253
+ interval and non-critical instructions in the speculation interval are executed on
254
+ the fly as a transaction encounters them; (2) critical instructions are executed
255
+ in a speculative fashion, so they will take effect, atomically, only on transaction
256
+ commit; and (3) instructions after the speculation interval are postponed until
257
+ after the commit.
258
+ 3
259
+ The Medley System
260
+ To illustrate NBTC, we have written a system, Medley, that (1) instruments
261
+ critical instructions, executes them speculatively, and commits them atomically
262
+ using M-compare-N-swap, our variant of the multi-word CAS of Harris et al.
263
+ [16]; (2) identifies and eagerly resolves transaction conflicts; and (3) delays non-
264
+ critical cleanup until transaction commit.
265
+ 6
266
+
267
+ 1 template <class T> class CASObj { // Augmented atomic object
268
+ 2
269
+ T nbtcLoad();
270
+ 3
271
+ bool nbtcCAS(T expected, T desired, bool linPt, bool pubPt);
272
+ 4
273
+ /* Regular atomic methods: */
274
+ 5
275
+ T load(); void store(T desired); bool CAS(T expected, T desired);
276
+ 6 };
277
+ 7 class Composable { // Base class of all transactional objects
278
+ 8
279
+ template <class T> void addToReadSet(CASObj<T>*,T); // Register load
280
+ 9
281
+ void addToCleanups(function); // Register post-critical work
282
+ 10
283
+ template <class T> T* tNew(...); // Create a new block
284
+ 11
285
+ template <class T> void tDelete(T*); // Delete a block
286
+ 12
287
+ template <class T> void tRetire(T*); // Epoch-based safe retire
288
+ 13
289
+ TxManager* mgr; // Tx metadata shared among Composables
290
+ 14
291
+ struct OpStarter { OpStarter(TxManager*); } // RAII op starter
292
+ 15 };
293
+ 16 class TxManager { // Manager shared among composable objects
294
+ 17
295
+ void txBegin(); // Start a transaction
296
+ 18
297
+ void txEnd(); // Try to commit the transaction
298
+ 19
299
+ void txAbort(); // Explicitly abort the transaction
300
+ 20
301
+ void validateReads(); // Optional validation for opacity
302
+ 21 };
303
+ 22 struct TransactionAborted : public std::exception{ };
304
+ Figure 1: C++ API of Medley for transaction composition.
305
+ 3.1
306
+ API
307
+ Figure 1 summarizes Medley’s API. Using this API, we transform an NBTC-
308
+ composable data structure into a transactional structure as follows:
309
+ 1. Replace critical loads and CASes with nbtcLoad and nbtcCAS, respectively.
310
+ Fields to which such accesses are made should be declared using the CASObj
311
+ template.
312
+ 2. Invoke addToReadSet for the critical load in a read operation, recording the
313
+ address and the loaded value.
314
+ 3. Register each operation’s post-critical work via addToCleanups.
315
+ 4. Replace every new and delete with tNew and tDelete. Replace every retire
316
+ (for safe memory reclamation—SMR) with tRetire.
317
+ 5. Declare an OpStarter object at the beginning of each operation.
318
+ CASObj¡T¿ augments each CAS-able 64-bit word (e.g., atomic¡Node*¿) with
319
+ additional metadata bits for speculation tracking (details in Section 3.2). It
320
+ provides a specialized load and CAS as well as the usual methods of atomic¡T¿.
321
+ To dynamically identify the speculation interval, nbtcCAS takes two extra argu-
322
+ ments, linPt and pubPt, that indicate whether this call, should it succeed, will
323
+ constitute its operation’s linearization or/and publication point. In a similar
324
+ vein, addToReadSet can be called after an nbtcLoad to indicate (after inspecting
325
+ the return value) that this was (or is likely to have been) the linearizing load of
326
+ a read-only operation, and should be tracked for validation at commit time.
327
+ Composable is a base class for transactional objects. It provides a variety of
328
+ NBTC-related methods, including support for safe memory reclamation (SMR),
329
+ used to ensure that nodes are not reclaimed until one can be certain that no
330
+ references remain among the private variables of other threads. Our current
331
+ 7
332
+
333
+ implementation of SMR uses epoch-based reclamation [10, 17, 27]. For the sake
334
+ of generality, Composable also provides an API for transactional boosting, which
335
+ can be used to incorporate lock-based operations into Medley transactions (at
336
+ the cost, of course, of nonblocking progress). We do not discuss this mechanism
337
+ further in this paper.
338
+ The TxManager class manages transaction metadata and provides methods
339
+ to initiate, abort, and complete a transaction. A TxManager instance is shared
340
+ among all Composable instances intended for use in the same transactions. In
341
+ each operation call, the manager distinguishes (via OpStarter()) whether exe-
342
+ cution is currently inside or outside a transaction. If outside, all transactional
343
+ instrumentation is elided; if inside, instrumentation proceeds as specified by the
344
+ NBTC methodology.
345
+ Given that nonblocking operations can execute safely in any reachable state
346
+ of the system, there is usually no need to stop the execution of a doomed-to-
347
+ abort transaction as soon as a conflict arises—i.e., to guarantee opacity [14]. In
348
+ exceptional cases (e.g., when later operations of a transaction cannot be called
349
+ with certain combinations of parameters, or when aborts are likely enough that
350
+ delaying them may compromise performance), the validateReads method can be
351
+ used to determine whether previous reads remain correct.
352
+ To illustrate the use of Medley, Figure 2 highlights lines of code in Michael’s
353
+ nonblocking hash table [28] that must be modified for NBTC; Figure 3 then
354
+ shows an example transaction that modifies two hash tables. In a real appli-
355
+ cation, the catch block for TransactionAborted would typically loop back to the
356
+ beginning of the transaction code to try again, possibly with additional code
357
+ to detect and recover from livelock (e.g., via backoff or hints to the underly-
358
+ ing scheduler). In contrast to STM systems, Medley does not instrument the
359
+ intra-transaction “glue” code between data structure operations.
360
+ This code
361
+ is always executed as regular code outside a transaction and should always be
362
+ data-race free; if it has side effects, the catch block (written by the programmer)
363
+ for aborted transactions should compensate for these before the programmer
364
+ chooses to retry or give up.
365
+ 3.2
366
+ M-Compare-N-Swap
367
+ To execute the critical memory accesses of each transaction atomically, we em-
368
+ ploy a software-emulated M-compare-N-swap (MCNS) that builds on the double-
369
+ compare-single-swap (RDCSS) and multi-word CAS (CASN) of Harris et al. [16].
370
+ Each transaction maintains a descriptor that contains a read set, a write set,
371
+ and a 64-bit triple of thread ID, serial number, and status, as shown in Fig-
372
+ ure 4. Descriptors are pre-allocated on a per-thread basis within a TxManager
373
+ instance, and are reused across transactions. A status can be InPrep (initial
374
+ state), InProg (ready to commit), Committed (only when InProg and validation
375
+ succeeds), or Aborted (from InPrep or due to failed validation).
376
+ Each originally 64-bit word at which a critical memory access may occur is
377
+ augmented with a 64-bit counter, together comprising an 128-bit CASObj. Each
378
+ critical CAS installs a pointer to its descriptor in the CASObj and increments
379
+ 8
380
+
381
+ 1 class MHashTable :
382
+ public Composable {
383
+ 2 struct Node { K key; V val; CASObj<Node*> next; };
384
+ 3 // from p, find c >= k; nbtcLoad and tRetire may be used
385
+ 4 bool find(CASObj<Node*>* &p, Node* &c, Node* &n, K k);
386
+ 5 optional<V> get(K key) {
387
+ 6
388
+ OpStarter starter(mgr); CASObj<Node*>* prev = nullptr;
389
+ 7
390
+ Node *curr, *next; optional<V> res = {};
391
+ 8
392
+ if (find(prev,curr,next,key)) res = curr->val;
393
+ 9
394
+ addToReadSet(prev,curr);
395
+ 10
396
+ return res;
397
+ 11 }
398
+ 12 optional<V> put(K key, V val) { // insert or replace if key exists
399
+ 13
400
+ OpStarter starter(mgr);
401
+ 14
402
+ CASObj<Node*>* prev = nullptr; optional<V> res = {};
403
+ 15
404
+ Node *newNode = tNew<Node>(key, val), *curr, *next;
405
+ 16
406
+ while(true) {
407
+ 17
408
+ if (find(prev,curr,next,key)) { // update
409
+ 18
410
+ newNode->next.store(curr);
411
+ 19
412
+ if (curr->next.nbtcCAS(next,mark(newNode),true,true)) {
413
+ 20
414
+ res = curr->val;
415
+ 21
416
+ auto cleanup = [](){
417
+ 22
418
+ if (prev->CAS(curr,newNode)) tRetire(curr);
419
+ 23
420
+ else find(prev,curr,next,key);
421
+ 24
422
+ };
423
+ 25
424
+ addToCleanups(cleanup); // execute right away if not in tx
425
+ 26
426
+ break;
427
+ 27
428
+ }
429
+ 28
430
+ } else { // key does not exist; insert
431
+ 29
432
+ newNode->next.store(curr);
433
+ 30
434
+ if (prev->nbtcCAS(curr,newNode,true,true)) break;
435
+ 31
436
+ }
437
+ 32
438
+ }
439
+ 33
440
+ return res;
441
+ 34 }};
442
+ Figure 2: Michael’s lock-free hash table example (Medley-related parts
443
+ highlighted).
444
+ 1 void doTx(MHashTable* ht1, MHashTable* ht2, V v, K a1, K a2) {
445
+ 2
446
+ TxManager* mgr=ht1->mgr; assert(mgr==ht2->mgr);
447
+ 3
448
+ try { // transfer ‘v’ from account ‘a1’ in ‘ht1’ to ‘a2’ in ‘ht2’
449
+ 4
450
+ mgr->txBegin();
451
+ 5
452
+ V v1 = ht1->get(a1); V v2 = ht2->get(a2);
453
+ 6
454
+ if (!v1.hasValue() or v1.value() < v) mgr->txAbort();
455
+ 7
456
+ ht1->put(a1, v1.value() - v); ht2->put(a2, v + v2.valueOr(0));
457
+ 8
458
+ mgr->txEnd();
459
+ 9
460
+ } catch (TransactionAborted) { /* transaction aborted */ }
461
+ 10 }
462
+ Figure 3: Transaction example on Michael’s hash table.
463
+ 9
464
+
465
+ 1 struct Desc {
466
+ 2
467
+ map<CASObj* addr,{uint64 val,cnt}>* readSet;
468
+ 3
469
+ map<CASObj* addr,{uint64 oldVal,cnt,newVal}>* writeSet;
470
+ 4
471
+ atomic<uint64> status;//63..50 tid 49..2 serialNumber 1..0 status
472
+ 5
473
+ enum STATUS { InPrep=0, InProg=1, Committed=2, Aborted=3 };
474
+ 6 };
475
+ 7 struct CASObj { atomic<uint128> val_cnt; };
476
+ Figure 4: Descriptor and CASObj structures.
477
+ the counter; at commit or abort, the descriptor is uninstalled and the counter
478
+ incremented again. We leverage 128-bit CAS instructions on the x86 to change
479
+ the original word and the counter together, atomically. The counter is odd when
480
+ CASObj contains a pointer to a descriptor and even when it is a real value.
481
+ Each instance of MCNS proceeds through phases that install descriptors,
482
+ finalize status, and uninstall descriptors. The first two phases are on the critical
483
+ path of a data structure operation. A new transaction initializes metadata in its
484
+ descriptor (at txBegin): it clears the read and write sets, increments the serial
485
+ number, and resets the status to InPrep. The installing phase then occurs over
486
+ the course of the transaction: Each critical load records its address, counter,
487
+ and value in the read set. Each critical CAS records its address, old counter,
488
+ old value, and desired new value in the write set; it then installs a pointer to
489
+ the descriptor in the CASObj. Pseudocode for the installing phase appears in
490
+ Figure 5.
491
+ To spare the programmer the need to reason about counters, nbtcLoad makes
492
+ a record of its ⟨counter, object⟩ pair (line 15 in Fig. 5); addToReadSet then adds
493
+ this pair (and the specified CASObj) to the transaction’s read set (line 20).
494
+ When a thread encounters its own descriptor, nbtcLoad returns the specu-
495
+ lated value from the write set (line 11). Likewise, nbtcCAS updates the write
496
+ entry (line 34). Such encounters automatically initiate the speculation interval
497
+ (lines 10, 30, and 32), which then extends through the linearization point of the
498
+ current operation (line 38).
499
+ If an operation encounters the descriptor of some other thread, it gets that
500
+ descriptor out of the way by calling tryFinalize (Fig. 6). This method aborts the
501
+ associated transaction if the descriptor is InPrep, helps complete the commit if
502
+ InProg, and in all cases uninstalls the descriptor from the CASObj in which it
503
+ was found. Similar actions occur when a thread is forced to abort or reaches the
504
+ end of its transaction and attempts to commit (lines 39–58). Whether helping
505
+ or acting on its own behalf, a thread performing an MCNS must verify that the
506
+ descriptor is still responsible for the CASObj through which it was discovered
507
+ (line 9) and (if committing) that the values in the read set are still valid (line 25).
508
+ After CAS-ing the status to Committed or Aborted, the thread uninstalls the
509
+ descriptor from all associated CASObjs, replacing pointers to the descriptor
510
+ with the appropriate updated values (lines 31 and 34). Once uninstalling is
511
+ complete, the owner thread calls cleanup routines (line 55) for a commit or
512
+ deallocates tNew-ed blocks (line 43) for an abort.
513
+ Our design adopts invisible readers and eager contention management for
514
+ efficiency and simplicity. Eager contention management admits the possibility
515
+ 10
516
+
517
+ 1 void TxManager::txBegin() {
518
+ 2
519
+ desc->readSet->clear(); desc->writeSet->clear();
520
+ 3
521
+ status.store((status.load() & ~3) + 4);
522
+ 4 }
523
+ 5 T CASObj::nbtcLoad() {
524
+ 6 retry:
525
+ 7
526
+ {val,cnt} = val_cnt.load();
527
+ 8
528
+ if (cnt % 2) { // is descriptor
529
+ 9
530
+ if (val == desc) {
531
+ 10
532
+ startSpeculativeInterval();
533
+ 11
534
+ return desc->writeSet[this].newVal;
535
+ 12
536
+ } else val->tryFinalize(this, {val,cnt});
537
+ 13
538
+ goto retry; // until object has real value
539
+ 14
540
+ }
541
+ 15
542
+ ... /* Record ‘this’ and ‘cnt’ to be added to readSet */
543
+ 16
544
+ return val;
545
+ 17 }
546
+ 18 void Composable::addToReadSet(CASObj<T>* obj, T val) {
547
+ 19
548
+ ... /* Retrieve ‘cnt’ by ‘obj‘ */
549
+ 20
550
+ mgr->readSet[obj] = {val,cnt};
551
+ 21 }
552
+ 22 bool CASObj::nbtcCAS(T expected,T desired,bool linPt,bool pubPt){
553
+ 23 retry:
554
+ 24
555
+ {val,cnt} = val_cnt.load();
556
+ 25
557
+ if (cnt % 2) { // is descriptor
558
+ 26
559
+ if (val != desc) { // not own descriptor
560
+ 27
561
+ val->tryFinalize(this, {val,cnt});
562
+ 28
563
+ goto retry; // until object has real value
564
+ 29
565
+ }
566
+ 30
567
+ startSpeculativeInterval();
568
+ 31
569
+ } else if (val != expected) return false;
570
+ 32
571
+ if (pubPt) startSpeculativeInterval();
572
+ 33
573
+ if (inSpeculativeInterval()) { // Is critical CAS
574
+ 34
575
+ desc->writeSet[this] = {val,cnt,desired};
576
+ 35
577
+ bool ret = true;
578
+ 36
579
+ if (!(cnt % 2)) ret = this->CAS({val,cnt},{desc,cnt+1});
580
+ 37
581
+ if (!ret) desc->writeSet.remove(this);
582
+ 38
583
+ if (linPt and ret) endSpeculativeInterval();
584
+ 39
585
+ return ret;
586
+ 40
587
+ } else return CAS(expected, desired);
588
+ 41 }
589
+ Figure 5: Pseudocode for installing phase.
590
+ 11
591
+
592
+ 1 bool Desc::stsCAS(uint64 d, STATUS expected, STATUS desired) {
593
+ 2
594
+ d = d & ~3; return status.CAS(d + expected, d + desired);
595
+ 3 }
596
+ 4 bool Desc::setReady(){return stsCAS(status.load(),InPrep,InProg);}
597
+ 5 bool Desc::commit(uint64 d){return stsCAS(d,InProg,Committed);}
598
+ 6 bool Desc::abort(uint64 d){return stsCAS(d,d & 1,Aborted);}
599
+ 7 void Desc::tryFinalize(CASObj* obj, uint128 var) {
600
+ 8
601
+ uint64 d = status.load();
602
+ 9
603
+ if (obj->val_cnt.load() != var) // ensure d indicates right tx
604
+ 10
605
+ return;
606
+ 11
607
+ if (d & 3 == InPrep) {
608
+ 12
609
+ abort(d);
610
+ 13
611
+ uint64 newd = status.load();
612
+ 14
613
+ if (newd & ~3 != d & ~3) return; // serial number mismatch
614
+ 15
615
+ d = newd;
616
+ 16
617
+ }
618
+ 17
619
+ if (d & 3 == InProg) {
620
+ 18
621
+ if (validateReads(d)) commit(d);
622
+ 19
623
+ else abort(d);
624
+ 20
625
+ }
626
+ 21
627
+ uninstall(status.load());
628
+ 22 }
629
+ 23 bool Desc::validateReads() {
630
+ 24
631
+ for (e:*readSet)
632
+ 25
633
+ if ({e.val,e.cnt} != e.addr->load()) return false;
634
+ 26
635
+ return true;
636
+ 27 }
637
+ 28 void Desc::uninstall(uint64 d) {
638
+ 29
639
+ if (d % 3 == Committed)
640
+ 30
641
+ for (e:*writeSet)
642
+ 31
643
+ e.addr->CAS({this,e.cnt+1}, {e.newVal,e.cnt+2});
644
+ 32
645
+ else // Aborted
646
+ 33
647
+ for (e:*writeSet)
648
+ 34
649
+ e.addr->CAS({this,e.cnt+1}, {e.oldVal,e.cnt+2});
650
+ 35 }
651
+ 36 struct TxManager {
652
+ 37
653
+ threadLocal vector<Function> cleanups, allocs;
654
+ 38
655
+ threadLocal Desc* desc;
656
+ 39
657
+ void txAbort() {
658
+ 40
659
+ uint64 d = desc->status.load();
660
+ 41
661
+ desc->abort(d);
662
+ 42
663
+ desc->uninstall(d);
664
+ 43
665
+ for (f:allocs) f(); // undo tNew
666
+ 44
667
+ throw TransactionAborted();
668
+ 45
669
+ }
670
+ 46
671
+ void txEnd() {
672
+ 47
673
+ if (!desc->setReady()) txAbort();
674
+ 48
675
+ else {
676
+ 49
677
+ uint64 d = desc->status.load();
678
+ 50
679
+ if (!desc->validateReads()) desc->abort(d);
680
+ 51
681
+ else if (d & 3 == InProg) desc->commit(d);
682
+ 52
683
+ d = desc->status.load();
684
+ 53
685
+ if (d & 3 == Committed) {
686
+ 54
687
+ desc->uninstall(d);
688
+ 55
689
+ for (f:cleanups) f();
690
+ 56
691
+ } else txAbort();
692
+ 57
693
+ }
694
+ 58
695
+ }
696
+ 59 };
697
+ Figure 6: Pseudocode of methods that finalize transactions.
698
+ 12
699
+
700
+ of livelock—transactions that repeatedly abort each other—and therefore guar-
701
+ antees only obstruction freedom. Lazy (commit-time) contention management
702
+ along with some total order of descriptor installment might allow us to pre-
703
+ serve lock freedom for structures that provide it [35], but would significantly
704
+ complicate the tracking and retrieving of uncommitted changes, and would not
705
+ address starvation, which may be a bigger problem than livelock in practice;
706
+ we consider these implementation choices orthogonal to the effectiveness of the
707
+ NBTC methodology, and decide to explore them in the future.
708
+ 4
709
+ Persistent Memory
710
+ Transactions developed, historically, in the database community; transactional
711
+ memory (TM) adapted them to in-memory structures in multithreaded pro-
712
+ grams. The advent of cheap, low-power, byte-addressable nonvolatile memory
713
+ (NVM) presents the opportunity to merge these two historical threads in a way
714
+ that ideally leverages NBTC. Specifically, where TM aims to convert sequential
715
+ code to thread-safe parallel code, NBTC assumes—as in the database world—
716
+ that we are already in possession of efficient thread-safe structures and we wish
717
+ to combine their operations. Given this assumption, it seems appropriate (as
718
+ described at the end of Sec. 3.1) to assume that the programmer is responsi-
719
+ ble for the “glue” code between operations, and to focus on the atomicity and
720
+ durability of the composed operations.
721
+ 4.1
722
+ Durable Linearizability
723
+ On machines with volatile caches, data structures in NVM will generally be
724
+ consistent after a crash only if programs take pains to issue carefully chosen
725
+ write-back and fence instructions. To characterize desired behavior, Izraelevitz
726
+ et al. [21] introduced durable linearizability as a correctness criterion for persis-
727
+ tent structures. A structure is durably linearizable if it is linearizable during
728
+ crash-free execution and its long-term history remains linearizable when crash
729
+ events are elided. Equivalently [11], each operation should persist between its
730
+ invocation and response, and the order of persists should match the linearization
731
+ order.
732
+ Many durably linearizable nonblocking data structures have been designed
733
+ in recent years [3, 9, 11, 44]. Several groups have also proposed methodologies
734
+ by which existing nonblocking structures can be made durably linearizable [12,
735
+ 13, 21].
736
+ Other groups have developed persistent STM systems, but most of
737
+ these have been lock-based [4, 5, 24, 40]. OneFile [33] and QSTM [1] are, to the
738
+ best of our knowledge, the only nonblocking persistent STM systems. OneFile
739
+ serializes transactions using a global sequence number, eliminating the need for
740
+ a read set and improving read efficiency, but introducing the need for invasive
741
+ data structure modifications and a 128-bit wide CAS. QSTM employs a global
742
+ persistent queue for active transactions, avoiding the need for wide CAS and
743
+ invasive structural changes, but its execution remains inherently serial.
744
+ 13
745
+
746
+ 4.2
747
+ Lowering Persistence Overhead
748
+ Unfortunately, write-back and fence instructions tend to have high latency.
749
+ Given the need for operations to persist before returning, durable linearizability
750
+ appears to be intrinsically expensive. Immediate persistence for STM introduces
751
+ additional overhead, as metadata for transaction concurrency control must also
752
+ be eagerly written back and fenced.
753
+ To move high latency instructions off the application’s critical path, Izraele-
754
+ vitz et al. [21] introduced the notion of buffered durable linearizability (BDL).
755
+ By allowing a modest suffix of pre-crash execution to be lost during post-crash
756
+ recovery (so long as the overall history remains linearizable), BDL allows write-
757
+ back and fence instructions to execute in batches, off the application’s critical
758
+ path. Applications that need to ensure persistence before communicating with
759
+ the outside world can employ a sync operation, reminiscent of those in tradi-
760
+ tional file systems and databases.
761
+ First proposed in the context of the Dal´ı persistent hash table [32], periodic
762
+ persistence was subsequently adopted by nbMontage [2], a general-purpose sys-
763
+ tem to create BDL versions of existing nonblocking structures. The nbMontage
764
+ system divides wall-clock time into “epochs” and persists operations in a batch
765
+ at the end of each epoch. In the wake of a crash in epoch e, the system recovers
766
+ all structures to their state as of the end of epoch e − 2. To maximize through-
767
+ put in the absence of crashes, nbMontage also distinguishes between data that
768
+ are semantically significant (a.k.a. “payloads”) and data that are merely per-
769
+ formance enhancing (e.g., indices); the latter can be kept in DRAM and rebuilt
770
+ during recovery. As an example, the payloads of a mapping are simply a pile of
771
+ key-value pairs; the associated hash table, tree, or skiplist resides in transient
772
+ DRAM. The payloads of a queue are ⟨serial number, item⟩ pairs.
773
+ To ensure that post-crash recovery always reflects a consistent state of each
774
+ structure, every nbMontage operation is forced to linearize in the epoch with
775
+ which its payloads have been labeled. Operations that take “too long” to com-
776
+ plete may be forced to abort and start over. The nbMontage system as a whole
777
+ is lock free; sync is actually wait free.
778
+ 4.3
779
+ Durable Strict Serializability
780
+ Linearizability, of course, is not suitable for transactions, which must remain
781
+ speculative until all operations can be made visible together.
782
+ STM systems
783
+ typically provide strict serializability instead: transactions in a crash-free history
784
+ appear to occur in a sequential order that respects real time (if A commits before
785
+ B begins, then A must serialize before B) [34, Sec. 3.1.2]. For a persistent version
786
+ of NBTC, we need to accommodate crashes.
787
+ Like Izraelevitz et al. [21], we assume a full-system crash failure model:
788
+ data structures continue to exist after a crash, but are accessed only by new
789
+ threads—the old threads disappear. Under this model:
790
+ Definition 4. An execution history H displays durable strict serializability
791
+ (DSS) if it is strictly serializable when crash events are elided.
792
+ 14
793
+
794
+ Like durable linearizability, this definition requires all work completed before a
795
+ crash to be visible after the crash. The buffered analogue is similar:
796
+ Definition 5. An execution history H displays buffered durable strict serializ-
797
+ ability (BDSS) if there exists a happens-before–consistent cut of each inter-crash
798
+ interval such that H is strictly serializable when crash events are elided along
799
+ with the post-cut suffix of each inter-crash interval.
800
+ 4.4
801
+ Merging Medley with nbMontage
802
+ The epoch system of nbMontage provides a natural mechanism with which to
803
+ provide failure atomicity and durability for Medley transactions: if operations
804
+ of the same transaction always occur in the same epoch, then they will be
805
+ recovered (or lost) together in the wake of a crash. Building on this observation,
806
+ we merge the two systems to create txMontage. Payloads of all operations in a
807
+ given transaction are labeled with the same epoch number. That number is then
808
+ validated along with the rest of the read set during MCNS commit, ensuring
809
+ that the transaction commits in the expected epoch. While nbMontage itself is
810
+ quite complex, this one small change is all that is required to graft it (and all
811
+ its converted persistent data structures) onto Medley: persistence comes into
812
+ transactions “almost for free.”
813
+ 5
814
+ Correctness
815
+ In this section, we argue that histories comprising well-formed Medley transac-
816
+ tions are strictly serializable, that Medley is obstruction free, and that txMon-
817
+ tage provides buffered durable strict serializability.
818
+ Definition 6. A Medley transaction is well-formed if
819
+ 1. it starts with txBegin and ends with txEnd, optionally with txAbort in between;
820
+ 2. it contains operations of NBTC-transformed data structures; and
821
+ 3. all other intra-transaction code is nonblocking and free from any side effects
822
+ not managed by handlers for the TransactionAborted exception.
823
+ 5.1
824
+ Strict Serializability
825
+ Lemma 1. At the implementation level (operating on the array of words that
826
+ comprises system memory), nbtcLoad, nbtcCAS, tryFinalize, txAbort, and txEnd
827
+ (MCNS) are linearizable operations.
828
+ Proof (sketch). Follows directly from Harris et al. [16]. Their RDCSS compares
829
+ (without changing) only a single location, and their CASN supports the up-
830
+ date of all touched words, but the proofs adapt in a straightforward way. In
831
+ particular, as in RDCSS, an unsuccessful tryFinalize or txEnd can linearize on a
832
+ (failed) validating read or a failed CAS of its status word. A tryFinalize or txEnd
833
+ 15
834
+
835
+ whose status CAS is successful linearizes “in the past,” on the first of its vali-
836
+ dating reads. (Ironically, this means that MCNS does not have an immediately
837
+ identifiable linearization point.)
838
+ Lemma 2. In any history in which transaction t performs an nbtcLoad or nbtc-
839
+ CAS operation x on CASObj o, and in which t’s txEnd operation y succeeds, no
840
+ tryFinalize or txEnd for a different transaction that modifies o succeeds between
841
+ x and y.
842
+ Proof (sketch). Suppose the contrary, and call the transaction with the conflict-
843
+ ing tryFinalize or txEnd u. If u’s nbtcCAS of o occurs between x and y, it will
844
+ abort and uninstall t’s descriptor, or cause read validation to fail in y, contra-
845
+ dicting the assumption that t’s txEnd succeeds. If u’s nbtcCAS of o occurs before
846
+ x, then x will abort and uninstall u’s descriptor, contradicting the assumption
847
+ that u’s tryFinalize or txEnd succeeds after x.
848
+ Theorem 3. Histories comprising well-formed Medley transactions are strictly
849
+ serializable.
850
+ Proof (sketch). In an NBTC-transformed data structure, all critical memory
851
+ accesses will be performed using nbtcLoad or nbtcCAS. These will be followed,
852
+ at some point, by a call to txEnd. If that call succeeds, no conflicting tryFinalize
853
+ or txEnd succeeds in the interim, by Lemma 2. This in turn implies that our
854
+ Medley history is equivalent to a sequential history in which each operation takes
855
+ effect at the nbtcLoad or nbtcCAS corresponding to the linearization point of the
856
+ original data structure operation, prior to NBTC transformation. Moreover, all
857
+ operations of the same transaction are contiguous in this sequential history—
858
+ that is, our Medley history is strictly serializable.
859
+ 5.2
860
+ Obstruction Freedom
861
+ Theorem 4. When used to build well-formed transactions that retry on abort,
862
+ Medley is obstruction free.
863
+ Proof (sketch). In any reachable system state, if one thread continues to execute
864
+ while others are paused, every nbtcLoad or nbtcCAS that encounters a conflict
865
+ will first finalize (commit or abort) the encountered descriptor, uninstall it,
866
+ and install its own descriptor. If the thread encounters its own descriptor, a
867
+ nbtcLoad will return the speculated value and a nbtcCAS will update the write
868
+ set if the argument matches the previous new value in the write set. In either
869
+ case, the MCNS will make progress. If it eventually aborts, it may repeat one
870
+ round of a brand new MCNS which, with no newly introduced contention, must
871
+ commit.
872
+ 5.3
873
+ Buffered Durable Strict Serializability
874
+ Theorem 5. Histories comprising well-formed txMontage transactions exhibit
875
+ buffered durable strict serializability.
876
+ 16
877
+
878
+ Proof (sketch). Each transaction reads the current epoch, e, in txBegin. It then
879
+ validates this epoch number during MCNS commit. Per Lemma 1, this MCNS
880
+ must linearize inside e. With nbMontage-provided failure atomicity of all oper-
881
+ ations in the same epoch, the theorem trivially holds.
882
+ 6
883
+ Performance Results
884
+ As noted in Section 1, we have used Medley to create NBTC versions of Michael
885
+ and Scott’s queue [29], Fraser’s skiplist [10], the rotating skiplist of Dick et al. [7],
886
+ Michael’s chained hash table [28], and Natarajan and Mittal’s binary search
887
+ tree [31]. All of the transformations were straightforward. In this section we
888
+ report on the performance on Medley and txMontage hash tables and skiplists,
889
+ comparing them to various alternatives from the literature.
890
+ Specifically, we tested the following transient systems:
891
+ Medley – as previously described (hash table and skip list)
892
+ OneFile – transient version of the lock-free STM of Ramalhete et al. [33] (hash
893
+ table and skip list)
894
+ TDSL – transactional data structure library of Spiegelman et al. [36] (authors’
895
+ skiplist only)
896
+ LFTT – lock-free transactional transform of Zhang et al. [43] (authors’ skiplist
897
+ only)
898
+ We also tested the following persistent systems:
899
+ txMontage – Medley + nbMontage (hash table and skiplist)
900
+ POneFile – persistent version of OneFile [33] (hash table and skiplist)
901
+ 6.1
902
+ Experimental Setup
903
+ We report throughput for hash table and skiplist microbenchmarks and for
904
+ skiplists used to run a subset of TPC-C [6]. We also measure latency for skiplists.
905
+ All code will be made publicly available prior to publication; we intend to par-
906
+ ticipate in the artifact evaluation process.
907
+ All tests were conducted on a Linux 5.3.7 (Fedora 30) server with two Intel
908
+ Xeon Gold 6230 processors. Each socket has 20 physical cores and 40 hyper-
909
+ threads, totaling 80 hyperthreads. Threads in all experiments were pinned first
910
+ one per core on socket 0, then on the extra hyperthreads of that socket, and
911
+ then on socket 1. Each socket has 6 channels of 32 GB DRAMs and 6 channels
912
+ of 128 GB Optane DIMMs. We mount NVM from each socket as an indepen-
913
+ dent ext4 file system. In all experiments, DRAM is allocated across the two
914
+ sockets according to Linux’s default policy; in persistent data structures, only
915
+ NVM on socket 0 is used, in direct access (DAX) mode. In all cases, we report
916
+ the average of three trials, each of which runs for 30 seconds.
917
+ Our throughput and latency microbenchmark begins by pre-loading the
918
+ structure with 0.5 M key-value pairs, drawn from a key space of 1 M keys. Both
919
+ 17
920
+
921
+ 104
922
+ 105
923
+ 106
924
+ 106
925
+ 107
926
+ 0
927
+ 10
928
+ 20
929
+ 30
930
+ 40
931
+ 50
932
+ 60
933
+ 70
934
+ 80
935
+ Threads
936
+ Throughput (txn/s)
937
+ Medley
938
+ txMontage
939
+ OneFile
940
+ POneFile
941
+ (a) get:insert:remove 0:1:1
942
+ 104
943
+ 105
944
+ 106
945
+ 106
946
+ 107
947
+ 0
948
+ 10
949
+ 20
950
+ 30
951
+ 40
952
+ 50
953
+ 60
954
+ 70
955
+ 80
956
+ Threads
957
+ Throughput (txn/s)
958
+ Medley
959
+ txMontage
960
+ OneFile
961
+ POneFile
962
+ (b) get:insert:remove 2:1:1
963
+ 104
964
+ 105
965
+ 106
966
+ 106
967
+ 107
968
+ 0
969
+ 10
970
+ 20
971
+ 30
972
+ 40
973
+ 50
974
+ 60
975
+ 70
976
+ 80
977
+ Threads
978
+ Throughput (txn/s)
979
+ Medley
980
+ txMontage
981
+ OneFile
982
+ POneFile
983
+ (c) get:insert:remove 18:1:1
984
+ Figure 7: Throughput of transactional hash tables (log Y axis).
985
+ 104
986
+ 105
987
+ 106
988
+ 106
989
+ 0
990
+ 10
991
+ 20
992
+ 30
993
+ 40
994
+ 50
995
+ 60
996
+ 70
997
+ 80
998
+ Threads
999
+ Throughput (txn/s)
1000
+ Medley
1001
+ txMontage
1002
+ OneFile
1003
+ POneFile
1004
+ TDSL
1005
+ LFTT
1006
+ (a) get:insert:remove 0:1:1
1007
+ 104
1008
+ 105
1009
+ 106
1010
+ 106
1011
+ 0
1012
+ 10
1013
+ 20
1014
+ 30
1015
+ 40
1016
+ 50
1017
+ 60
1018
+ 70
1019
+ 80
1020
+ Threads
1021
+ Throughput (txn/s)
1022
+ Medley
1023
+ txMontage
1024
+ OneFile
1025
+ POneFile
1026
+ TDSL
1027
+ LFTT
1028
+ (b) get:insert:remove 2:1:1
1029
+ 104
1030
+ 105
1031
+ 106
1032
+ 106
1033
+ 0
1034
+ 10
1035
+ 20
1036
+ 30
1037
+ 40
1038
+ 50
1039
+ 60
1040
+ 70
1041
+ 80
1042
+ Threads
1043
+ Throughput (txn/s)
1044
+ Medley
1045
+ txMontage
1046
+ OneFile
1047
+ POneFile
1048
+ TDSL
1049
+ LFTT
1050
+ (c) get:insert:remove 18:1:1
1051
+ Figure 8: Throughput of transactional skiplists (log Y axis).
1052
+ keys and values are 8-byte integers. In the benchmarking phase, each thread
1053
+ composes and executes transactions comprising 1 to 10 operations each. Oper-
1054
+ ations (on uniformly random keys) are chosen among get, insert, and remove in
1055
+ a ratio specified as a parameter (0:1:1, 2:1:1, or 18:1:1 in our experiments).
1056
+ In OneFile, we use a sequential chained hash table parallelized using STM.
1057
+ In Medley, we use an NBTC-transformed version of Michael’s lock-free hash
1058
+ table [28]. Each table has 1 M buckets. In OneFile and TDSL, skiplists are
1059
+ derived from Fraser’s STM-based skiplist [10]. In LFTT and Medley, they are
1060
+ derived from Fraser’s CAS-based nonblocking skiplist [10]. Each skiplist has up
1061
+ to 20 levels.
1062
+ For TPC-C, we are limited by the fact that Fraser’s skiplists do not sup-
1063
+ port range queries. Following the lead of Yu et al. in their experiments with
1064
+ DBx1000[42], we limit our experiments to TPC-C’s newOrder and payment trans-
1065
+ actions, which we perform in a 1:1 ratio. These are the dominant transactions
1066
+ in the benchmark; neither performs a range query.
1067
+ 6.2
1068
+ Throughput (Transient)
1069
+ Throughput results for the hash table and skiplist microbenchmarks appear
1070
+ in Figures 7 and 8, respectively. Solid lines represent transactions on transient
1071
+ data structures; dotted lines represent persistent transactions. Considering only
1072
+ the transient case for now, Medley consistently outperforms the transient ver-
1073
+ sion of OneFile by more than an order of magnitude, on both hash tables and
1074
+ skiplists, for anything more than a trivial number of threads. The gap becomes
1075
+ larger when the workload has a higher percentage of writes. Despite its lack
1076
+ of scalability, OneFile performs well at small thread counts, especially with a
1077
+ 18
1078
+
1079
+ read-mostly workload. We attribute this fact to its serialized transaction design,
1080
+ which eliminates the need for read sets.
1081
+ As described in Section 1, TDSL provides (blocking) transactions over vari-
1082
+ ous specially constructed data structures. While conflicts still occur on writes,
1083
+ read sets are limited to only semantically critical nodes, and the authors report
1084
+ significant improvements in throughput relative to general-purpose STM [36].
1085
+ As shown in Figure 8, however, TDSL, like OneFile, has limited scalability, and
1086
+ is dramatically outperformed by Medley. Somewhat to our surprise, TDSL also
1087
+ fails to outperform OneFile on this microbenchmark, presumably because of the
1088
+ latter’s elimination of read sets.
1089
+ Among the various skiplist competitors, LFTT comes closest to rivaling
1090
+ Medley, but still trails by a factor of 1.4–2× in the best (write-only)case. Re-
1091
+ executing entire transactions in LFTT introduces considerableredundant work—
1092
+ planning in particular. On read-mostly workloads, where Medley benefits from
1093
+ invisible readers, LFTT trails by a factor of 2–2.7×.
1094
+ As a somewhat more realistic benchmark, we repeated our comparison of
1095
+ Medley, OneFile, and TDSL on the newOrder and payment transactions of TPC-
1096
+ C. We were unable to include LFTT in these tests because it supports only static
1097
+ transactions, in which the set of data structure operations is known in advance—
1098
+ nor could we integrate its dynamic variant (DTT [23]), as the available version
1099
+ of the code does not allow arbitrary key and value types. LaBorde et al. [23]
1100
+ report, however, that DTT’s performance is similar to that of LFTT on simple
1101
+ transactions. Given that DTT has to publish the entire transaction as a lambda
1102
+ expression on all its critical nodes, we would expect DTT’s performance to be,
1103
+ if anything, somewhat worse on the large transactions of TPC-C, and LFTT
1104
+ was already about 2× slower than Medley on the microbenchmark.
1105
+ TPC-C throughput for Medley, (transient) OneFile, and TDSL appears
1106
+ in Figure 9. Because transactions on TPC-C are large, OneFile is impacted
1107
+ severely. By ensuring the atomicity of only critical accesses, Medley still scales
1108
+ for large numbers of threads and outperforms the competition by as much as
1109
+ 45×.
1110
+ 6.3
1111
+ Latency (Transient)
1112
+ In an attempt to assess the marginal cost of transaction composition, we re-ran
1113
+ our microbenchmark on Fraser’s original skiplist (Original—no transactions),
1114
+ the NBTC-transformed skiplist without transactions (TxOff—no calls to txBe-
1115
+ gin or txEnd), and the NBTC-transformed skiplist with transactions (TxOn—as
1116
+ in Figure 8).
1117
+ Figure 10a reports latency for structures placed in DRAM. Without trans-
1118
+ actions, the transformed skiplist is 1.8× slower than the original. With trans-
1119
+ actions turned on, it’s about 2.2× slower. These results suggest that the more-
1120
+ than-doubled cost of CASes (installing and uninstalling descriptors) accounts
1121
+ for about 2/3 of Medley’s overhead.
1122
+ 19
1123
+
1124
+ 104
1125
+ 105
1126
+ 106
1127
+ 106
1128
+ 0
1129
+ 10
1130
+ 20
1131
+ 30
1132
+ 40
1133
+ 50
1134
+ 60
1135
+ 70
1136
+ 80
1137
+ Threads
1138
+ Throughput (txn/s)
1139
+ Medley
1140
+ txMontage
1141
+ OneFile
1142
+ TDSL
1143
+ Figure 9: TPC-C skiplist throughput (log Y axis).
1144
+ 234
1145
+ 169
1146
+ 202
1147
+ 421
1148
+ 313
1149
+ 366
1150
+ 513
1151
+ 383
1152
+ 453
1153
+ 0
1154
+ 250
1155
+ 500
1156
+ 750
1157
+ 1000
1158
+ 0:1:1
1159
+ 2:1:1
1160
+ 18:1:1
1161
+ Latency (ns/txn)
1162
+ Original
1163
+ TxOff
1164
+ TxOn
1165
+ (a) on DRAM
1166
+ 847
1167
+ 380
1168
+ 593
1169
+ 623
1170
+ 330
1171
+ 419
1172
+ 650
1173
+ 404
1174
+ 502
1175
+ 0
1176
+ 250
1177
+ 500
1178
+ 750
1179
+ 1000
1180
+ 0:1:1
1181
+ 2:1:1
1182
+ 18:1:1
1183
+ Latency (ns/txn)
1184
+ Original
1185
+ TxOff
1186
+ TxOn
1187
+ (b) transient on NVM
1188
+ 673
1189
+ 334
1190
+ 443
1191
+ 678
1192
+ 408
1193
+ 525
1194
+ 0
1195
+ 250
1196
+ 500
1197
+ 750
1198
+ 1000
1199
+ 0:1:1
1200
+ 2:1:1
1201
+ 18:1:1
1202
+ Latency (ns/txn)
1203
+ TxOff
1204
+ TxOn
1205
+ (c) persistent on NVM
1206
+ Figure 10: Average latency on skiplists at 40 threads.
1207
+ X labels are ratio of get:insert:remove.
1208
+ 20
1209
+
1210
+ 6.4
1211
+ Persistence
1212
+ To evaluate the impact of failure atomicity and durability on the throughput of
1213
+ txMontage, we can return to the dotted lines of Figures 7, 8, and 9.
1214
+ Throughput
1215
+ In the microbenchmark tests, with strict persistence and eager cache-line write-
1216
+ back, persistent OneFile is an order of magnitude slower than its transient
1217
+ version. With periodic persistence, however, the txMontage hash table achieves
1218
+ half the throughput of Medley at 40 threads on the write-only workload—almost
1219
+ two orders of magnitude faster than POneFile. With a read-mostly workload
1220
+ on the hash table, or with any of the workloads on the skiplist (with its lower
1221
+ overall concurrency), txMontage is almost as fast as Medley. In the extreme
1222
+ write-heavy case (80 threads on the 0:1:1 hash table workload), we attribute
1223
+ the roughly 4× slowdown of txMontage to NVM’s write bottleneck [22]—in
1224
+ particular, to the phenomenon of write amplification [20, 41].
1225
+ Results are similar in TPC-C (Fig. 9). Transactions here are both large and
1226
+ heavy on writes; allocating payloads on NVM limits txMontage’s throughput
1227
+ to roughly a fifth of Medley’s, but that is still about 4× faster than transient
1228
+ OneFile. POneFile, for its part, spent so long on the warm-up phase of TPC-C
1229
+ that we lost patience and killed the test.
1230
+ Latency
1231
+ Figure 10b shows the latency of skiplist transactions when txMontage payloads
1232
+ are allocated on NVM (and indices on DRAM) but persistence is turned off
1233
+ (no epochs or explicit cache line write-back).
1234
+ For comparison, we have also
1235
+ shown the latency of the original, non-transactional skiplist with all data placed
1236
+ in NVM.
1237
+ Figure 10c shows the corresponding latencies for fully operational
1238
+ txMontage.
1239
+ Comparing Figures 10a and 10b, we see lower marginal overhead for trans-
1240
+ actions when running on NVM. This may suggest that the hardware write bot-
1241
+ tleneck is reducing overall throughput and thus contention.
1242
+ On the write-only workload (leftmost groups of bars), moving payloads to
1243
+ NVM introduces an overhead of almost 50% (Fig. 10a versus Fig. 10b). On the
1244
+ read-mostly workload (rightmost bars), this penalty drops to 5%. Again, we
1245
+ attribute the effect to NVM’s write bottleneck. The high latency of the original
1246
+ skiplist entirely allocated on NVM (green bars in Figure 10b) appears to confirm
1247
+ this hypothesis.
1248
+ Comparing Figures 10b and 10c, txMontage pays less than 5%, relative to
1249
+ Medley on NVM, for failure atomicity and durability.
1250
+ 21
1251
+
1252
+ 7
1253
+ Conclusion
1254
+ We have presented nonblocking transaction composition (NBTC), a new method-
1255
+ ology that leverages the linearizability of existing nonblocking data structures
1256
+ when building dynamic transactions. As concrete realizations, we introduced the
1257
+ Medley system for transient structures and the txMontage system for (buffered)
1258
+ persistent structures. Medley transactions are isolated and consistent; txMon-
1259
+ tage transactions are also failure atomic and durable. Both systems are quite
1260
+ fast: where even the best STM has traditionally suffered slowdowns of 3–10×,
1261
+ Medley incurs more like 2.2×; txMontage, for its part, adds only 5–20% to the
1262
+ overhead of nbMontage, allowing it to outperform existing nonblocking persis-
1263
+ tent STM systems by nearly two orders of magnitude.
1264
+ Given their eager contention management, Medley and txMontage main-
1265
+ tain obstruction freedom for transactions on nonblocking structures. In future
1266
+ work, we plan to explore lazy contention management, postponing installment
1267
+ of descriptors until transactions are ready to commit. By sorting and installing
1268
+ descriptors in canonical order, the resulting systems would preserve lock free-
1269
+ dom. Lazy contention management would also facilitate helping, as any installed
1270
+ descriptor would have status == InProg, and any other thread could push it to
1271
+ completion.
1272
+ As currently defined in NBTC, speculation intervals are easy to identify,
1273
+ but may unnecessarily instrument certain harmless helping instructions between
1274
+ publication and linearization. We are currently working to develop a more pre-
1275
+ cise but still tractable definition of helping in order to reduce the number of
1276
+ “critical” memory accesses that must be performed atomically in each transac-
1277
+ tion.
1278
+ References
1279
+ [1] H. Alan Beadle, Wentao Cai, Haosen Wen, and Michael L. Scott. Non-
1280
+ blocking persistent software transactional memory. In 27th Intl. Conf. on
1281
+ High Performance Computing, Data, and Analytics (HiPC), pages 283–293,
1282
+ virtual conference, December 2020.
1283
+ [2] Wentao Cai, Haosen Wen, Vladimir Maksimovski, Mingzhe Du, Rafaello
1284
+ Sanna, Shreif Abdallah, and Michael L. Scott. Fast Nonblocking Persis-
1285
+ tence for Concurrent Data Structures. In 35th Intl. Symp. on Distributed
1286
+ Computing (DISC), pages 14:1–14:20, Freiburg, Germany, October 2021.
1287
+ [3] Zhangyu Chen, Yu Huang, Bo Ding, and Pengfei Zuo. Lock-free concurrent
1288
+ level hashing for persistent memory.
1289
+ In Usenix Annual Technical Conf.
1290
+ (ATC), pages 799–812, virtual conference, July 2020.
1291
+ [4] Joel Coburn, Adrian M. Caulfield, Ameen Akel, Laura M. Grupp, Ra-
1292
+ jesh K. Gupta, Ranjit Jhala, and Steven Swanson. NV-Heaps: Making
1293
+ 22
1294
+
1295
+ persistent objects fast and safe with next-generation, non-volatile memo-
1296
+ ries. In 16th Intl. Conf. on Architectural Support for Programming Lan-
1297
+ guages and Operating Systems (ASPLOS), pages 105–118, Newport Beach,
1298
+ CA, March 2011.
1299
+ [5] Andreia Correia, Pascal Felber, and Pedro Ramalhete.
1300
+ Romulus: Effi-
1301
+ cient algorithms for persistent transactional memory. In 30th ACM Symp.
1302
+ on Parallel Algorithms and Architectures (SPAA), pages 271–282, Vienna,
1303
+ Austria, July 2018.
1304
+ [6] The Transaction Processing Council. TPC-C benchmark (revision 5.11.0).
1305
+ http://www.tpc.org/tpcc/, February 2010.
1306
+ [7] Ian Dick, Alan Fekete, and Vincent Gramoli. A skip list for multicore.
1307
+ Concurrency and Computation: Practice and Experience, 29(4), May 2016.
1308
+ [8] Avner Elizarov, Guy Golan-Gueta, and Erez Petrank.
1309
+ Loft: Lock-free
1310
+ transactional data structures. In 24th ACM SIGPLAN Symp. on Princi-
1311
+ ples and Practice of Parallel Programming (PPoPP), page 425–426, Wash-
1312
+ ington, DC, 2019.
1313
+ [9] Panagiota Fatourou, Elias Papavasileiou, and Eric Ruppert.
1314
+ Persistent
1315
+ non-blocking binary search trees supporting wait-free range queries.
1316
+ In
1317
+ 31st ACM Symp. on Parallelism in Algorithms and Architectures (SPAA),
1318
+ pages 275–286, Phoenix, AZ, June 2019.
1319
+ [10] Keir Fraser. Practical Lock-Freedom. PhD thesis, King’s College, Univ.
1320
+ of Cambridge, September 2003. Published as Univ. of Cambridge Com-
1321
+ puter Laboratory technical report #579, February 2004. www.cl.cam.ac.
1322
+ uk/techreports/UCAM-CL-TR-579.pdf.
1323
+ [11] Michal Friedman, Maurice Herlihy, Virendra Marathe, and Erez Petrank. A
1324
+ persistent lock-free queue for non-volatile memory. In 23rd ACM SIGPLAN
1325
+ Symp. on Principles and Practice of Parallel Programming (PPoPP), pages
1326
+ 28–40, Vienna, Austria, February 2018.
1327
+ [12] Michal Friedman, Naama Ben-David, Yuanhao Wei, Guy E. Blelloch, and
1328
+ Erez Petrank. NVTraverse: In NVRAM data structures, the destination
1329
+ is more important than the journey. In 41st ACM Conf. on Programming
1330
+ Language Design and Implementation (PLDI), pages 377–392, virtual con-
1331
+ ference, June 2020.
1332
+ [13] Michal Friedman, Erez Petrank, and Pedro Ramalhete. Mirror: Making
1333
+ lock-free data structures persistent. In 42nd ACM Conf. on Programming
1334
+ Language Design and Implementation (PLDI), pages 1218–1232, virtual
1335
+ conference, June 2021.
1336
+ [14] Rachid Guerraoui and Michal Kapalka. On the correctness of transactional
1337
+ memory. In 13th ACM SIGPLAN Symp. on Principles and Practice of Par-
1338
+ allel Programming (PPoPP), pages 175–184, Salt Lake City, UT, February
1339
+ 2008.
1340
+ 23
1341
+
1342
+ [15] Theo Haerder and Andreas Reuter.
1343
+ Principles of transaction-oriented
1344
+ database recovery.
1345
+ ACM Computing Surveys, 15(4):287–317, December
1346
+ 1983.
1347
+ [16] Timothy L. Harris, Keir Fraser, and Ian A. Pratt. A practical multi-word
1348
+ compare-and-swap operation. In 16th Intl. Symp. on Distributed Computing
1349
+ (DISC), pages 265–279, Toulouse, France, October 2002.
1350
+ [17] Thomas E. Hart, Paul E. McKenney, Angela Demke Brown, and Jonathan
1351
+ Walpole. Performance of memory reclamation for lockless synchronization.
1352
+ Journal of Parallel Distributed Computing (JPDC), 67(12):1270–1285, De-
1353
+ cember 2007.
1354
+ [18] Maurice Herlihy and Eric Koskinen. Transactional boosting: A methodol-
1355
+ ogy for highly-concurrent transactional objects. In 13th ACM SIGPLAN
1356
+ Symp. on Principles and Practice of Parallel Programming (PPoPP), pages
1357
+ 207–216, Salt Lake City, UT, February 2008.
1358
+ [19] Maurice Herlihy, Victor Luchangco, Mark Moir, and William N. Scherer
1359
+ III. Software transactional memory for dynamic-sized data structures. In
1360
+ 22nd ACM Symp. on Principles of Distributed Computing (PODC), pages
1361
+ 92–101, Boston, MA, July 2003.
1362
+ [20] Daokun Hu, Zhiwen Chen, Wenkui Che, Jianhua Sun, and Hao Chen. Halo:
1363
+ A hybrid pmem-dram persistent hash index with fast recovery. In Intl Conf
1364
+ on Management of Data (SIGMOD), pages 1049–1063, Philadelphia, PA,
1365
+ June 2022.
1366
+ [21] Joseph Izraelevitz, Hammurabi Mendes, and Michael L. Scott. Linearizabil-
1367
+ ity of persistent memory objects under a full-system-crash failure model. In
1368
+ 30th Intl. Symp. on Distributed Computing (DISC), pages 313–327, Paris,
1369
+ France, September 2016.
1370
+ [22] Joseph Izraelevitz, Jian Yang, Lu Zhang, Juno Kim, Xiao Liu, Amirsaman
1371
+ Memaripour, Yun Joon Soh, Zixuan Wang, Yi Xu, Subramanya R. Dulloor,
1372
+ Jishen Zhao, and Steven Swanson. Basic performance measurements of the
1373
+ Intel Optane DC persistent memory module, August 2019. arXiv preprint
1374
+ arXiv:1903.05714v3.
1375
+ [23] Pierre LaBorde, Lance Lebanoff, Christina Peterson, Deli Zhang, and
1376
+ Damian Dechev. Wait-free dynamic transactions for linked data structures.
1377
+ In 10th Intl. Workshop on Programming Models and Applications for Mul-
1378
+ ticores and Manycores (PMAM), pages 41–50, Washington, DC, February
1379
+ 2019.
1380
+ [24] Mengxing Liu, Mingxing Zhang, Kang Chen, Xuehai Qian, Yongwei Wu,
1381
+ Weimin Zheng, and Jinglei Ren. DudeTM: Building durable transactions
1382
+ with decoupling for persistent memory. In 22nd Intl. Conf. on Architectural
1383
+ Support for Programming Languages and Operating Systems (ASPLOS),
1384
+ pages 329–343, Xi’an, China, April 2017.
1385
+ 24
1386
+
1387
+ [25] Virendra J. Marathe, Michael F. Spear, Christopher Heriot, Athul Acharya,
1388
+ David Eisenstat, William N. Scherer III, and Michael L. Scott. Lowering the
1389
+ overhead of software transactional memory. In 1st ACM SIGPLAN Work-
1390
+ shop on Transactional Computing (TRANSACT), Ottawa, ON, Canada,
1391
+ June 2006.
1392
+ [26] Virendra Jayant Marathe and Mark Moir. Toward high performance non-
1393
+ blocking software transactional memory. In 13th ACM SIGPLAN Symp. on
1394
+ Principles and Practice of Parallel Programming (PPoPP), pages 227–236,
1395
+ Salt Lake City, UT, February 2008.
1396
+ [27] Paul E. McKenney, Dipankar Sarma, Andrea Arcangeli, Andi Kleen, Orran
1397
+ Krieger, and Rusty Russell. Read copy update. In Ottawa Linux Symp.,
1398
+ pages 338–367, Ottawa, ON, Canada, July 2001.
1399
+ [28] Maged M. Michael. High performance dynamic lock-free hash tables and
1400
+ list-based sets.
1401
+ In 14th ACM Symp. on Parallelism in Algorithms and
1402
+ Architectures (SPAA), pages 73–82, Winnipeg, MB, Canada, August 2002.
1403
+ [29] Maged M. Michael and Michael L. Scott. Simple, fast, and practical non-
1404
+ blocking and blocking concurrent queue algorithms. In 15th ACM Symp. on
1405
+ Principles of Distributed Computing (PODC), pages 267–275, Philadelphia,
1406
+ PA, May 1996.
1407
+ [30] Chi Cao Minh, JaeWoong Chung, Christos Kozyrakis, and Kunle Olukotun.
1408
+ Stamp: Stanford transactional applications for multi-processing. In IEEE
1409
+ Intl. Symp. on Workload Characterization (IISWC), pages 35–46, Seattle,
1410
+ WA, September 2008.
1411
+ [31] Aravind Natarajan and Neeraj Mittal. Fast concurrent lock-free binary
1412
+ search trees. In 19th ACM SIGPLAN Symp. on Principles and Practice
1413
+ of Parallel Programming (PPoPP), pages 317–328, Orlando, FL, February
1414
+ 2014.
1415
+ [32] Faisal Nawab, Joseph Izraelevitz, Terence Kelly, Charles B. Morrey III,
1416
+ Dhruva R. Chakrabarti, and Michael L. Scott. Dal´ı: A periodically per-
1417
+ sistent hash map. In 31st Intl. Symp. on Distributed Computing (DISC),
1418
+ pages 37:1–37:16, Vienna, Austria, October 2017.
1419
+ [33] Pedro Ramalhete, Andreia Correia, Pascal Felber, and Nachshon Cohen.
1420
+ OneFile: A wait-free persistent transactional memory. In 49th IEEE/IFIP
1421
+ Intl. Conf. on Dependable Systems and Networks (DSN), pages 151–163,
1422
+ Portland, OR, June 2019.
1423
+ [34] Michael L. Scott. Shared-Memory Synchronization. Morgan & Claypool
1424
+ Publishers, San Rafael, CA, 2013.
1425
+ 25
1426
+
1427
+ [35] Michael F. Spear, Luke Dalessandro, Virendra J. Marathe, and Michael L.
1428
+ Scott. A comprehensive contention management strategy for software trans-
1429
+ actional memory. In 14th ACM SIGPLAN Symp. on Principles and Prac-
1430
+ tice of Parallel Programming (PPoPP), pages 141–150, Raleigh, NC, Febru-
1431
+ ary 2009.
1432
+ [36] Alexander Spiegelman, Guy Golan-Gueta, and Idit Keidar. Transactional
1433
+ data structure libraries. In 37th ACM Conf. on Programming Language
1434
+ Design and Implementation (PLDI), pages 682–696, Santa Barbara, CA,
1435
+ June 2016.
1436
+ [37] Fuad Tabba, Mark Moir, James R. Goodman, Andrew W. Hay, and Cong
1437
+ Wang. NZTM: Nonblocking zero-indirection transactional memory. In 21st
1438
+ ACM Symp. on Parallelism in Algorithms and Architectures (SPAA), pages
1439
+ 204–213, Calgary, AB, Canada, August 2009.
1440
+ [38] Shahar Timnat and Erez Petrank.
1441
+ A practical wait-free simulation for
1442
+ lock-free data structures. In 19th ACM SIGPLAN Symp. on Principles and
1443
+ Practice of Parallel Programming (PPoPP), pages 357—-368, Orlando, FL,
1444
+ February 2014.
1445
+ [39] Stephen Tu, Wenting Zheng, Eddie Kohler, Barbara Liskov, and Samuel
1446
+ Madden.
1447
+ Speedy transactions in multicore in-memory databases.
1448
+ In
1449
+ 24th ACM Symp. on Operating Systems Principles (SOSP), pages 18–32,
1450
+ Farminton, PA, November 2013.
1451
+ [40] Haris Volos, Andres Jaan Tack, and Michael M. Swift.
1452
+ Mnemosyne:
1453
+ Lightweight persistent memory. In 16th Intl. Conf. on Architectural Sup-
1454
+ port for Programming Languages and Operating Systems (ASPLOS), pages
1455
+ 91–104, Newport Beach, CA, March 2011.
1456
+ [41] Zixuan Wang, Xiao Liu, Jian Yang, Theodore Michailidis, Steven Swanson,
1457
+ and Jishen Zhao. Characterizing and modeling non-volatile memory sys-
1458
+ tems. In 53rd Intl. Symp. on Microarchitecture (MICRO), pages 496–508,
1459
+ virtual conference, October 2020.
1460
+ [42] Xiangyao Yu, George Bezerra, Andrew Pavlo, Srinivas Devadas, and
1461
+ Michael Stonebraker.
1462
+ Staring into the abyss: An evaluation of concur-
1463
+ rency control with one thousand cores. Proc. of the VLDB Endowment, 8
1464
+ (3):209–220, November 2014.
1465
+ [43] Deli Zhang and Damian Dechev. Lock-free transactions without rollbacks
1466
+ for linked data structures. In 28th ACM Symp. on Parallelism in Algorithms
1467
+ and Architectures (SPAA), pages 325–336, Pacific Grove, CA, July 2016.
1468
+ [44] Yoav Zuriel, Michal Friedman, Gali Sheffi, Nachshon Cohen, and Erez Pe-
1469
+ trank. Efficient lock-free durable sets. Proc. of the ACM on Programming
1470
+ Languages, 3(OOPSLA):128:1–128:26, October 2019.
1471
+ 26
1472
+
6dAzT4oBgHgl3EQfEfrh/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
6dE1T4oBgHgl3EQfBQKx/content/2301.02850v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2c2ce43dfed8f8c095d5778bcc13012d1f438f7370ef5789fb908c6206b668e5
3
+ size 8513524
6dE1T4oBgHgl3EQfBQKx/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cccbc594579f30acc19a72bf702db61a890ac1f6e44c3037dc4ca5f176fc7a47
3
+ size 4718637
8dAzT4oBgHgl3EQfgfzo/content/2301.01471v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ae7ba5f35695f517ceeae774f54200d1eac98837bf8a86d6d23c147aa0d5531c
3
+ size 16392216
99AzT4oBgHgl3EQfFfoM/content/tmp_files/2301.01010v1.pdf.txt ADDED
@@ -0,0 +1,2493 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
2
+ 1
3
+ Joint Optimization of Video-based AI Inference
4
+ Tasks in MEC-assisted Augmented Reality Systems
5
+ Guangjin Pan, Heng Zhang, Shugong Xu, Fellow, IEEE,
6
+ Shunqing Zhang, Senior Member, IEEE, and Xiaojing Chen
7
+ Abstract—The high computational complexity and energy
8
+ consumption of artificial intelligence (AI) algorithms hinder their
9
+ application in augmented reality (AR) systems. However, mobile
10
+ edge computing (MEC) makes it possible to solve this problem.
11
+ This paper considers the scene of completing video-based AI
12
+ inference tasks in the MEC system. We formulate a mixed-integer
13
+ nonlinear programming problem (MINLP) to reduce inference
14
+ delays, energy consumption and to improve recognition accuracy.
15
+ We give a simplified expression of the inference complexity model
16
+ and accuracy model through derivation and experimentation.
17
+ The problem is then solved iteratively by using alternating opti-
18
+ mization. Specifically, by assuming that the offloading decision is
19
+ given, the problem is decoupled into two sub-problems, i.e., the
20
+ resource allocation problem for the devices set that completes the
21
+ inference tasks locally, and that for the devices set that offloads
22
+ tasks. For the problem of offloading decision optimization, we
23
+ propose a Channel-Aware heuristic algorithm. To further reduce
24
+ the complexity, we propose an alternating direction method of
25
+ multipliers (ADMM) based distributed algorithm. The ADMM-
26
+ based algorithm has a low computational complexity that grows
27
+ linearly with the number of devices. Numerical experiments
28
+ show the effectiveness of proposed algorithms. The trade-off
29
+ relationship between delay, energy consumption, and accuracy
30
+ is also analyzed.
31
+ Index Terms—Mobile augmented reality, edge intelligence,
32
+ mobile edge computing, resource allocation.
33
+ I. INTRODUCTION
34
+ R
35
+ ECENTLY, the development of networks, cloud com-
36
+ puting, edge computing, artificial intelligence, and other
37
+ technologies has triggered people’s infinite imagination of
38
+ the Metaverse [1]. To enable users to interact between the
39
+ real world and the virtual world, augmented reality (AR)
40
+ technology plays a vital role. At the same time, artificial
41
+ intelligence (AI), due to its learning and inference capabilities,
42
+ has demonstrated a powerful ability in many fields such as
43
+ automatic speech recognition (ASR) [2], natural language
44
+ G. Pan, H. Zhang, S. Xu, S. Zhang and X. Chen are with Shanghai Institute
45
+ for Advanced Communication and Data Science, Shanghai University, Shang-
46
+ hai 200444, China. Emails: {guangjin_pan, hengzhang, shugong, Shunqing,
47
+ jodiechen}@shu.edu.cn.
48
+ Part of this work has been accepted by Globecom-2022. This work was
49
+ supported in part by the National Natural Science Foundation of China
50
+ (NSFC) under Grant 61871262, 62071284, and 61901251, the National Key
51
+ R&D Program of China grants 2017YFE0121400, 2019YFE0196600 and
52
+ 2022YFB2902000, the Innovation Program of Shanghai Municipal Science
53
+ and Technology Commission grants 20JC1416400 and 21ZR1422400, Pudong
54
+ New Area Science & Technology Development Fund, Key-Area Research
55
+ and Development Program of Guangdong Province grant 2020B0101130012,
56
+ Foshan Science and Technology Innovation Team Project grant FS0AA-
57
+ KJ919-4402-0060, and research funds from Shanghai Institute for Advanced
58
+ Communication and Data Science (SICS). The corresponding author is
59
+ Shugong Xu.
60
+ processing (NLP) [3], computer vision (CV) [4], and so on.
61
+ With the assistance of AI technology, AR can carry out deeper
62
+ scene understanding and more immersive interactions.
63
+ However, the computational complexity of AI algorithms,
64
+ especially deep neural networks (DNN), is usually very high.
65
+ It is challenging to complete DNN inference timely and reli-
66
+ ably on mobile devices with limited computation and energy
67
+ capacity. In [5], experiments show that a typical single-frame
68
+ image processing AI inference task takes about 600 ms even
69
+ with speedup from the mobile GPU. In addition, continuously
70
+ executing the above inference tasks can only last up to 2.5
71
+ hours on commodity devices. The above issues result in only a
72
+ few AR applications currently using deep learning [6]. In order
73
+ to reduce the inference time of DNNs, one way is to perform
74
+ network pruning on the neural network [7], [8]. However,
75
+ it could be destructive to the model if pruning too many
76
+ channels, and it may not be possible to recover a satisfactory
77
+ accuracy by fine-tuning [7].
78
+ Edge AI [9]–[11] is another approach to solving these
79
+ problems. Integrating mobile edge computing (MEC) and AI
80
+ technology has recently become a promising paradigm for
81
+ supporting computationally intensive tasks. Edge AI transfers
82
+ the inference and training process of AI models to the edge of
83
+ the network close to the data source. Therefore, it can alleviate
84
+ network traffic load, delay, and privacy problems.
85
+ A. Related Works
86
+ Many existing studies use MEC’s powerful computing capa-
87
+ bilities to reduce delay [12], energy consumption [13], or both
88
+ delay and energy consumption [14]–[16] through offloading.
89
+ For example, [12] formulated an optimization problem aimed
90
+ at minimizing the processing delay of eMBB and mMTC
91
+ users by optimizing the users’ transmit power in UAV-Assisted
92
+ MEC systems. [13] develops a smart pricing mechanism to
93
+ coordinate the computation offloading of multi-layer devices
94
+ and reduces energy consumption. [14] uses the Stackelberg
95
+ game method to optimize the task allocation coefficient, calcu-
96
+ lation resource allocation coefficient, and transmission power
97
+ to minimize the energy consumption and delay of the NOMA-
98
+ based MEC system.
99
+ For edge AI inference, existing research has made some
100
+ progress. The authors in [17] propose a framework for jointly
101
+ optimizing inference task selection and downlink coordinated
102
+ beamforming to minimize communication power consumption
103
+ in wireless networks. Similarly, [18] proposes an IRS-assisted
104
+ edge inference system and designs a task selection strategy
105
+ arXiv:2301.01010v1 [cs.NI] 3 Jan 2023
106
+
107
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
108
+ 2
109
+ to minimize the energy consumption of uplink and downlink
110
+ transmission and calculation. The work in [19] analyzes and
111
+ models the transmission error probability, inference accu-
112
+ racy, and timeout probability of the AI-powered time-critical
113
+ services. The work in [20] uses a tandem queueing model
114
+ to analyze queueing and processing delays of DL tasks in
115
+ multiple DNN partitions. [21] joint optimizes the service
116
+ placement, computational and radio resource allocation to
117
+ minimize the users’ total delay and energy consumption. [8]
118
+ combines model pruning and DNN partitioning to achieve a
119
+ 4.81x reduction on end-to-end delay. [22] designs the Edgent
120
+ framework that can jointly optimize DNN partitioning and
121
+ DNN right-sizing to maximize the inference accuracy while
122
+ promising application delay requirements. These studies mea-
123
+ sure the inference time by experiments [8], [22] or assume that
124
+ the inference task’s computational complexity is proportional
125
+ to the input data size but without derivation and proof [20],
126
+ [21]. However, these models of computational complexity are
127
+ not rigorous enough or can not be generalized to different
128
+ neural network models.
129
+ As for the accuracy model, the authors in [23] designs
130
+ an edge network orchestration algorithm named FACT, which
131
+ boosts the performance of an edge-based AR system by opti-
132
+ mizing the edge server assignment and video frame resolution
133
+ selection for AR users. However, [23] builds an accuracy
134
+ model by fitting an accuracy curve for specific tasks, which
135
+ is not general. The work in [24] compresses image resolution
136
+ locally and performs inference tasks on edge servers, aiming
137
+ to maximize learning accuracy under constraints of delay
138
+ and energy. [24] proposes using an abstract non-decreasing
139
+ function to describe the relationship between accuracy and
140
+ input image size, which cannot be used to analyze various AI
141
+ inference tasks discriminately. Joint optimization is required
142
+ when different tasks and models are jointly deployed. An
143
+ insufficiently generalized accuracy model or an overly abstract
144
+ model can adversely affect joint optimization. A general
145
+ accuracy model is needed to measure various AI tasks.
146
+ Among the above studies, most studies consider optimizing
147
+ one or two performance metrics among the delay, energy con-
148
+ sumption, and accuracy. The authors in [24] jointly considers
149
+ delay, energy consumption and accuracy in image recognition
150
+ scenarios. However, it aims at maximizing computational
151
+ capacity under constraints of delay, energy consumption and
152
+ accuracy, and the DNN model is only deployed in edge servers.
153
+ In [6], [23], [25], video analytics scenarios are considered, but
154
+ they do not jointly consider delay, energy and accuracy.
155
+ B. Contributions and Organizations
156
+ In this paper, we consider a multi-user MEC system and
157
+ assume that each device executes the video-based DNN in-
158
+ ference task. Each device can be AR glasses, mobile robots,
159
+ and so on. In order to deepen AR’s ability to understand the
160
+ scene, we need to use time dimension information to improve
161
+ perception. Therefore, we consider video-based application
162
+ scenarios.for video-based AI inference tasks, there are two
163
+ modes, e.g., frame-by-frame recognition mode (the input for
164
+ each recognition is one frame) and multi-frame recognition
165
+
166
+ Uplink for real-time captured content
167
+ Downlink for inference results
168
+ Fig. 1. Multi-user MEC System model. The inference task can be executed on
169
+ the local or the edge server. When the task is offloaded to the edge server, the
170
+ uplink transmits the content captured in real-time, and the downlink transmits
171
+ the inference result.
172
+ mode (the input for each recognition is multiple frames). The
173
+ frame-by-frame inference mode is used to deal with tasks with
174
+ weak temporal correlation, such as face recognition and target
175
+ tracking., and has been studied in [6], [23], [25]. In this paper,
176
+ we focus on multi-frame recognition tasks, such as gesture
177
+ recognition and action recognition tasks. Since sampling in
178
+ the spatial dimension brings extra computation [24], we only
179
+ sample in the temporal domain. At each inference, the device
180
+ selects the most recent several frames from the history frames
181
+ for transmission or inference.
182
+ As shown in Fig. 1, mobile devices can transmit captured
183
+ video to the edge server via wireless networks. The edge
184
+ servers execute inference tasks and send results back to mo-
185
+ bile devices. However, when communication and computing
186
+ resources of the edge server are insufficient, devices can
187
+ execute the inference task locally. We model the problem as a
188
+ multi-objective optimization problem to optimize delay, energy
189
+ consumption, and inference accuracy. The main contributions
190
+ of this paper are summarized as follows,
191
+ • Multi-dimensional target optimization. High accuracy,
192
+ low delay, and low energy consumption are indispensable
193
+ for AR applications and must be optimized jointly. To
194
+ explore the trade-off relationship between delay, energy,
195
+ and accuracy, we formulate the video-based offloading
196
+ problem as a mixed-integer nonlinear programming prob-
197
+ lem (MINLP), aiming to reduce service delays, energy
198
+ consumption and improve recognition accuracy.
199
+ • General computational complexity and accuracy models.
200
+ To measure the computational complexity of neural net-
201
+ work models with different architectures and different
202
+ input sizes, we introduce the number of multiply-and-
203
+ accumulate operations (MACs). We illustrate the main
204
+ factors affecting DNN inference delay through experi-
205
+ ments and show that MAC can be used as a good measure
206
+ of the computational complexity of DNN inference tasks.
207
+ We also propose a general model to represent the relation-
208
+ ship between the inference accuracy and the number of
209
+ input frames. This model is suitable for different video-
210
+
211
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
212
+ 3
213
+ based recognition tasks and different DNN architectures.
214
+ We give simple expressions of the inference complexity
215
+ and accuracy to simplify the optimization problem.
216
+ • Channel-Aware scheduling scheme. To solve the opti-
217
+ mization problem, we decompose the original problem.
218
+ First, assuming that the offloading decision is given, we
219
+ solve the resource allocation problems for the device set
220
+ that completes the inference locally and the device set
221
+ that offloads the tasks to the edge server, respectively. For
222
+ edge DNN inference, we propose two algorithms based
223
+ on search and geometric programming (GP) to solve the
224
+ problem. Then, to obtain the optimal offloading policy,
225
+ we propose a Channel-Aware heuristic algorithm. The
226
+ original problem is solved iteratively through alternating
227
+ optimization.
228
+ • ADMM-based distributed resource allocation scheme. To
229
+ avoid the high complexity of the heuristic algorithm, we
230
+ propose an algorithm based on the Alternating direction
231
+ method of multipliers (ADMM). The ADMM-based algo-
232
+ rithm decomposes the original problem into parallel and
233
+ tractable subproblems. Therefore, the total computational
234
+ complexity of ADMM-based algorithms is more scalable
235
+ than the heuristic algorithm, especially when the number
236
+ of devices is large.
237
+ The rest of this paper is organized as follows. In Section
238
+ II, we introduce system models, including delay, energy,
239
+ and accuracy models. In Section III, we formulate the joint
240
+ optimization problem and convert the original problem to a
241
+ more tractable problem. Section IV proposes a Channel-Aware
242
+ heuristic algorithm to solve the proposed problem. In Section
243
+ V, we propose another ADMM-based distributed resource
244
+ allocation algorithm for the proposed problem, and analyze the
245
+ computational complexity of the solution algorithm. Numeri-
246
+ cal results and analysis are presented in Section VI. Finally,
247
+ the paper is concluded in Section VII.
248
+ II. SYSTEM MODEL
249
+ In this section, we introduce a single-cell MEC system and
250
+ establish delay, energy consumption, and accuracy models. As
251
+ shown in Fig. 1, we consider a multi-user MEC system with
252
+ one base station (BS) and N mobile devices, denoted by the
253
+ set N = {1, 2, . . . N}. Each device has a camera and needs
254
+ to accomplish DNN inference tasks. Due to the limitation of
255
+ device computational resources, DNN inference tasks can be
256
+ placed on local or edge servers. The limited computational
257
+ resource will lead to longer computing delay and greater power
258
+ consumption when the inference task is executed locally.
259
+ However, when the inference task is executed on the edge
260
+ server, it will bring additional wireless transmission delay. In
261
+ addition, accuracy is also a very important optimization target
262
+ in DNN inference tasks.
263
+ A. Offloading Framework
264
+ In this paper, we only consider the binary offloading
265
+ method. Binary offloading requires the DNN inference task to
266
+ be fully executed either at the device or the MEC server. The
267
+ overview of the DNN computing offloading system is depicted
268
+ Sampling
269
+ management module
270
+ Video
271
+ offloading
272
+ Result
273
+ feedback
274
+ Complete
275
+ inference task
276
+ Execute locally
277
+ Complete
278
+ inference task
279
+ Complete
280
+ inference task
281
+ Fig. 2.
282
+ The overview of the video sampling and computing offloading
283
+ system. The video sampling management module can control the sampling
284
+ rate of the captured video and determine the number of video frames used
285
+ for AI inference. Devices can transmit the video to the edge server or
286
+ perform inference tasks locally based on the wireless channel information
287
+ and computing capabilities.
288
+ in Fig. 2. First, devices sample the video captured in real-
289
+ time in the temporal dimension to obtain a short video with
290
+ a certain number of frames. Second, the DNN inference tasks
291
+ are executed. These inference tasks can be executed locally
292
+ on devices or the edge server. Therefore, each device’s video
293
+ sampling management module needs to select an appropriate
294
+ video sampling rate (how many frames need to be input) and
295
+ choose whether to offload the task to the MEC server. Denote
296
+ Dn, En and φn to be the total delay, energy consumption and
297
+ recognition accuracy of the device n, respectively. The total
298
+ delay and energy consumption of the device n can be given
299
+ by,
300
+ Dn = (1 − xn)Dmd
301
+ n
302
+ + xn(Dt
303
+ n + De
304
+ n),
305
+ (1)
306
+ En = (1 − xn)Emd
307
+ n
308
+ + xnEt
309
+ n,
310
+ (2)
311
+ where xn indicates whether the inference task is executed on
312
+ local or edge servers. Dt
313
+ n is the transmission delay for uplink,
314
+ Dmd
315
+ n
316
+ is the local inference delay, and De
317
+ n is the delay for
318
+ completing inference at the edge server. Et
319
+ n and Emd
320
+ n
321
+ are the
322
+ transmission and computational energy consumption, respec-
323
+ tively. The delay and energy consumption for downloading
324
+ computation results can be reasonably neglected because of
325
+ the results’ small data sizes.
326
+ B. Delay and Energy Models for Inference
327
+ The inference delay depends on the DNN model’s architec-
328
+ ture, the device’s or server’s computing power, and the input
329
+ to the model. In this section, we first give a measure of the
330
+ computational complexity of the DNN model and then give an
331
+ expression for the inference delay and energy consumption.
332
+ Different AI recognition tasks may require different AI
333
+ model architectures, including classic AI models such as
334
+ Resnet-18, Resnet-34, Resnet-50, VGG-16, etc. [26], [27]. In
335
+ order to optimize AI inference tasks more reasonably, different
336
+ AI models need a common method to evaluate computational
337
+
338
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
339
+ 4
340
+ complexity. In this paper, we use the number of MACs [28] to
341
+ measure the computational complexity of AI inference tasks.
342
+ MACs calculation methods of layers (such as fully connected
343
+ (FC) layers, convolutional layers and so on) can be obtained
344
+ in [28]. Taking 3D Convolutional Neural Network (3DCNN)
345
+ as an example, the computational complexity (measured by
346
+ MACs) of the lth layer of the nth device can be expressed as,
347
+ cn,l =
348
+ olol+1
349
+ �2
350
+ j=0 Kj
351
+ l , �2
352
+ j=0 M j
353
+ n,l+1,
354
+ (3)
355
+ where ol is the number of input channels, ol+1 is the number
356
+ of output channels, �2
357
+ j=0 Kj
358
+ l is the size of the convolution
359
+ kernel, and �2
360
+ j=0 M j
361
+ n,l+1 is the size of the output feature
362
+ map. j = 0 represents the temporal dimension (the number of
363
+ frames), j = 1, 2 represent spatial dimensions (pixels of one
364
+ frame). Note that ol, ol+1, and �2
365
+ j=0 Kj
366
+ l are all determined
367
+ by the neural network architecture and �2
368
+ j=0 M j
369
+ n,l+1 depends
370
+ on the input size. The relation between the output feature size
371
+ and the input size can be expressed as,
372
+ M j
373
+ n,l+1 =
374
+ M j
375
+ n,l − Kj
376
+ l + 2dl
377
+ rl
378
+ + 1,
379
+ (4)
380
+ where rl is the stride and dl is the padding size.
381
+ As mentioned above, the computational complexity of a
382
+ DNN model is determined by the number of layers, the DNN
383
+ model’s architecture, and the input and output size. In this
384
+ paper, we mainly focus on the impact of the number of input
385
+ video frames Mn on recognition accuracy and the allocation
386
+ of communication and computing resources. The inference
387
+ result will be more accurate with more frames Mn input, but
388
+ the communication and calculation overhead will be greater.
389
+ The computational complexity of the nth device’s task can be
390
+ expressed as C(Mn).
391
+ Then we give the expression for the inference delay and
392
+ energy consumption. Denote f max and f max
393
+ n
394
+ (in CPU cycle/s)
395
+ to be the total computation resource of the edge server and
396
+ mobile device n, respectively. Let f e
397
+ n and f md
398
+ n
399
+ (in CPU
400
+ cycle/s) denote the computation resource to device n allocated
401
+ by the edge server and the device, respectively. Therefore,
402
+ the computing resources satisfy �
403
+ n∈N f e
404
+ n
405
+ ≤ f max and
406
+ f md
407
+ n
408
+ ≤ f max
409
+ n
410
+ . The computation delay of the device n and
411
+ MEC can be respectively expressed as,
412
+ Dmd
413
+ n
414
+ = ρC(Mn)
415
+ f md
416
+ n
417
+ ,
418
+ (5)
419
+ De
420
+ n = ρC(Mn)
421
+ f en
422
+ ,
423
+ (6)
424
+ where ρ (cycle/MAC) represents the number of CPU cycles
425
+ required to complete a multiplication and addition, which
426
+ depends on the CPU model.
427
+ As for energy consumption, denote κ to be a coefficient
428
+ determined by the corresponding device [24], and the com-
429
+ putational energy consumption of device n can be expressed
430
+ as,
431
+ Emd
432
+ n
433
+ = κρC(Mn)f md
434
+ n
435
+ 2.
436
+ (7)
437
+ C. Delay and Energy Models for Transmission
438
+ We consider a time-division multiple access (TDMA)
439
+ method for channel access. Specifically, each radio frame is
440
+ divided into N time slots for transmission, and each device can
441
+ only transmit in its own time slot. We assume that the length
442
+ of each radio frame is ∆T, which is short enough (e.g., 10
443
+ ms in LTE or NR system [24]), and the length of a time slot
444
+ is ∆Ttn.
445
+ Denote hn and pn to be the channel gain and transmission
446
+ power of the device n, respectively. According to [21], the
447
+ achievable data rate of device n can be expressed as,
448
+ Rn = Bwlog2
449
+
450
+ 1 + pnhn
451
+ BwN0
452
+
453
+ ,
454
+ (8)
455
+ where Bw and N0 are the bandwidth and the variance of
456
+ additive white Gaussian noise (AWGN), respectively.
457
+ Let d denote the data size of one video frame. Since we only
458
+ want to analyze the impact of time dimension information (the
459
+ number of input frames Mn) on recognition accuracy, d is a
460
+ constant value. In each radio frame, the data size that can
461
+ be transmitted is ∆TRntn. Therefore, for each transmission,
462
+
463
+ Mnd
464
+ ∆T Rntn ⌉ radio frames are required, where ⌈·⌉ means the
465
+ ceil function. Considering that the length of the radio frame
466
+ is much shorter than the transmission delay, the transmission
467
+ delay for offloading to MEC can be written as,
468
+ Dt
469
+ n = ⌈
470
+ Mnd
471
+ ∆TRntn
472
+ ⌉∆T ≈ Mnd
473
+ Rntn
474
+ ,
475
+ (9)
476
+ where tn is the proportion of time that device n transmits. In
477
+ addition, according to [24], the energy consumption of each
478
+ device to transmit its video can be expressed as,
479
+ Et
480
+ n = Mnd
481
+ Rn
482
+ pn.
483
+ (10)
484
+ D. Inference Tasks Accuracy Model
485
+ As mentioned above, we mainly focus on the impact of the
486
+ number of input video frames Mn on recognition accuracy.
487
+ We assume that the quality of the input video is the same
488
+ for different devices. For a certain task and DNN model, the
489
+ accuracy is only determined by the number of input frames.
490
+ Therefore, the accuracy of device n can be expressed as
491
+ φn = Φ(Mn). According to [29], more frames will lead to
492
+ better inference accuracy, and as the input frames continue to
493
+ increase, the performance gain will gradually decrease. Some
494
+ prior studies also show that the relationship between frame
495
+ rate and accuracy can be expressed as concave functions [23].
496
+ Therefore, we define Φ(Mn) as a monotone non-decreasing
497
+ function to describe the relationship between the accuracy and
498
+ the number of input frames.
499
+ III. PROBLEM FORMULATION
500
+ In this section, we formulate the optimization problem to
501
+ reduce the system’s delay and devices’ energy consumption
502
+ and improve accuracy. We analyze the difficulty of solving
503
+ the problem. To simplify the problem, we make a reasonable
504
+ conversion of the problem.
505
+
506
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
507
+ 5
508
+ A. Original Problem Formulation
509
+ Based on the above analysis, combining (1), (2), (5)-(7),
510
+ (9), and(10), the nth device’s delay and energy consumption
511
+ can be expressed as,
512
+ Dn
513
+ =
514
+ (1 − xn)ρC(Mn)
515
+ f md
516
+ n
517
+ + xn(ρC(Mn)
518
+ f en
519
+ + Mnd
520
+ Rntn
521
+ ), (11)
522
+ En
523
+ =
524
+ (1 − xn)κρC(Mn)f md
525
+ n
526
+ 2 + xn(Mnd
527
+ Rn
528
+ pn).
529
+ (12)
530
+ Given the system model described previously, our goal
531
+ is to reduce end-to-end delay and energy consumption and
532
+ improve recognition accuracy. Each device follows the binary
533
+ offloading policy. The mathematical optimization problem of
534
+ the total cost (delay, energy consumption, and accuracy) can
535
+ be expressed as,
536
+ Problem P1 (Original Problem):
537
+ minimize
538
+ {Mn,tn,f md
539
+ n
540
+ ,f e
541
+ n,xn}
542
+
543
+ n∈N
544
+
545
+ β1Dn + β2En − β3Φ(Mn)
546
+
547
+ , (13)
548
+ subject to
549
+ Φ(Mn) ≥ αn, ∀n ∈ N,
550
+ (13a)
551
+ Mn ≤ M max
552
+ n
553
+ , Mn ∈ Z,
554
+ (13b)
555
+
556
+ n∈N
557
+ xntn ≤ 1,
558
+ (13c)
559
+
560
+ n∈N
561
+ xnf e
562
+ n ≤ f max,
563
+ (13d)
564
+ tn, f e
565
+ n ≥ 0, ∀n ∈ N,
566
+ (13e)
567
+ 0 ≤ f md
568
+ n
569
+ ≤ f max
570
+ n
571
+ , ∀n ∈ N,
572
+ (13f)
573
+ xn ∈ {0, 1} , ∀n ∈ N,
574
+ (13g)
575
+ where αn represents the recognition accuracy requirement, β1,
576
+ β2, β3 are the weight factors. (13a) represents the recognition
577
+ accuracy requirement of each device. (13b) indicates the frame
578
+ limit for the input video, Z is the set of integers, and M max
579
+ n
580
+ is
581
+ the maximum number of frames of the input video. (13c) and
582
+ (13d) represent the communication and computation resource
583
+ limitation, respectively. (13f) limits the computation resource
584
+ of each device.
585
+ The optimization variables in original problem P1 are
586
+ the number of input video frames Mn, the proportion of
587
+ transmission time tn, the local computation resource f md
588
+ n , the
589
+ edge computation resource allocation f e
590
+ n, and the offloading
591
+ decision xn. In addition, the first item in (13) is to reduce the
592
+ total delay of computation and transmission, the second item
593
+ is to reduce the device’s energy consumption, and the last item
594
+ is to improve the number of input video frames as well as the
595
+ recognition accuracy because of the monotone non-decreasing
596
+ function Φ(Mn).
597
+ Problem P1 is a non-convex MINLP problem and is difficult
598
+ to be solved. First, the complexity function C(Mn) is discrete
599
+ and depends on the architecture of the DNN and the size
600
+ of the input video. As the number of input frames Mn
601
+ increases, the computational complexity also increases. This
602
+ kind of increase is irregular because it is affected by the
603
+ structure of DNN layers, such as the stride and padding size
604
+ of 3DCNN according to (4). Therefore, C(Mn) cannot be
605
+ used for optimization directly. Second, as mentioned above,
606
+ the accuracy function Φ(Mn) is non-decreasing. However, we
607
+ cannot give a deterministic expression for Φ(Mn), so we can
608
+ not optimize it. In addition, both Mn and xn are integers,
609
+ making the problem difficult to be solved.
610
+ B. Problem Conversion
611
+ To make the problem P1 more tractable, we convert the
612
+ problem. First, we give an approximate expression of the
613
+ computational complexity function C(Mn). According to (3)
614
+ and (4), the computational complexity of 3DCNN layers is
615
+ proportional to the size of the input data. We can also obtain
616
+ a similar conclusion in other types of layers, such as the FC
617
+ layer [28]. Based on the above conclusion and combined with
618
+ the experiments in Sec. VI-A, in order to simply express the
619
+ computational complexity model, C(Mn) can be written as,
620
+ C(Mn) = mc,0Mn + mc,1,
621
+ (14)
622
+ where mc,0 ≥ 0 and mc,1 are constants and depend on the
623
+ network model.
624
+ Second, we propose a general model to express the re-
625
+ lationship between the accuracy and the number of input
626
+ video frames. Considering that the function Φ(Mn) is mono-
627
+ tonically non-decreasing and that as the number of input
628
+ frames increases, the accuracy gain decreases, combining our
629
+ experiments in Sec. VI-A, we model function Φ(Mn) as,
630
+ Φ(Mn) = −
631
+ ma,0
632
+ Mn + ma,1
633
+ + ma,2,
634
+ (15)
635
+ where ma,0 ≥ 0, ma,2 ≥ 0 and ma,1 > −1 are constants and
636
+ depend on the target of inference tasks and the architecture of
637
+ DNN models.
638
+ Finally, we relax the range of the variable Mn. Consid-
639
+ ering that Φ(Mn) is a monotone non-decreasing function
640
+ and depends on the recognition task and network archi-
641
+ tecture, in order not to lose generality, define M min
642
+ n
643
+ =
644
+ arg minMn Φ(Mn), Φ(Mn) ≥ αn, Mn ∈ Z. We can also
645
+ relax Mn into a closed connected subset of the real axis, and
646
+ (13a), (13b) can be written as Mn ∈
647
+
648
+ M min
649
+ n
650
+ , M max
651
+ n
652
+
653
+ . Then
654
+ [Mn] can be regarded as the number of input video frames,
655
+ where [·] indicates rounding. We define two sets of devices, i.e.
656
+ N0 = {n | xn = 0, n ∈ N} and N1 = {n | xn = 1, n ∈ N}.
657
+ F0,n and F1,n are the cost function of the device n in sets N0
658
+ and N1, respectively. The problem P1 can be rewritten as,
659
+ Problem P2 (Converted Problem):
660
+ minimize
661
+ {Mn,tn,f md
662
+ n
663
+ ,f e
664
+ n,xn}
665
+
666
+ n∈N0
667
+ (1 − xn)F0,n(Mn, f md
668
+ n )
669
+ +
670
+
671
+ n∈N1
672
+ xnF1,n(Mn, f e
673
+ n, tn),
674
+ (16)
675
+ subject to Mn ∈
676
+
677
+ M min
678
+ n
679
+ , M max
680
+ n
681
+
682
+ ,
683
+ (16a)
684
+ (13c) − (13g),
685
+ where
686
+ F0,n(Mn, f md
687
+ n ) = β1
688
+ ρC(Mn)
689
+ f md
690
+ n
691
+ + β2κρC(Mn)f md2
692
+ n
693
+ − β3Φ(Mn),
694
+ (17)
695
+ F1,n(Mn, f e
696
+ n, tn) = β1
697
+ ρC(Mn)
698
+ f en
699
+ + β1
700
+ Mnd
701
+ Rntn
702
+
703
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
704
+ 6
705
+ + β2
706
+ Mndpn
707
+ Rn
708
+ − β3Φ(Mn).
709
+ (18)
710
+ IV. OPTIMIZATION PROBLEM SOLVING
711
+ In this section, we decompose the problem P2 and propose
712
+ a Channel-Aware heuristic algorithm to solve it. First, sup-
713
+ posing that the offloading decision (i.e., {xn}) is given, we
714
+ solve optimization problems for sets N0 and N1, respectively.
715
+ Second, we propose a Channel-Aware heuristic algorithm to
716
+ optimize the offloading decision {xn}.
717
+ A. Optimization Problem Solving for N0
718
+ For set N0, i.e., when the device executes inference tasks
719
+ locally, the optimization problem becomes,
720
+ Problem PN0 (Problem for N0):
721
+ minimize
722
+ {Mn,f md
723
+ n
724
+ } FPN0 ≜
725
+
726
+ n∈N0
727
+ F0,n(Mn, f md
728
+ n ),
729
+ (19)
730
+ subject to
731
+ (13f), (16a).
732
+ The optimization variables in PN0 are the number of input
733
+ video frames Mn and the local computation resource f md
734
+ n .
735
+ Let {M ∗
736
+ n, f md∗
737
+ n
738
+ } denote the optimal solution to PN0. We can
739
+ derive the optimal solution to PN0 in a closed-form expression.
740
+ Theorem 1: The optimal solution to PN0 is given by,
741
+ f md∗
742
+ n
743
+ = min{
744
+ 3
745
+
746
+ ( β1
747
+ 2β2κ), f max
748
+ n
749
+ },
750
+ (20)
751
+ M ∗
752
+ n = min{max{
753
+
754
+ β3ma,0
755
+ β1ρmc,0
756
+ f md
757
+ n
758
+ + β2κρmc,0f md2
759
+ n
760
+ − ma,1, M min
761
+ n
762
+ }, M max
763
+ n
764
+ }.
765
+ (21)
766
+ Proof: Please refer to Appendix A.
767
+ From Theorem 1, we can see that the optimal local CPU-
768
+ cycle frequency f md
769
+ n
770
+ is determined by the weight factors β1,
771
+ β2, the coefficient of CPU energy consumption κ, and is
772
+ limited by its corresponding upper bound f max
773
+ n
774
+ . More specifi-
775
+ cally, f md
776
+ n
777
+ is proportional to β
778
+ 1
779
+ 3
780
+ 1 and inversely proportional to
781
+ β
782
+ 1
783
+ 3
784
+ 2 and κ
785
+ 1
786
+ 3 . As for the number of input video frames, when
787
+ 3�
788
+ ( β1
789
+ 2β2κ) ≤ f max
790
+ n
791
+ , combining (20) and (21), we have,
792
+ M ∗
793
+ n = min{max{3− 1
794
+ 2 2
795
+ 1
796
+ 3 ρ− 1
797
+ 2 κ− 1
798
+ 6 m
799
+ − 1
800
+ 2
801
+ c,0 β
802
+ − 1
803
+ 3
804
+ 1
805
+ β
806
+ − 1
807
+ 6
808
+ 2
809
+ β
810
+ 1
811
+ 2
812
+ 3 m
813
+ 1
814
+ 2
815
+ a,0
816
+ − ma,1, M min
817
+ n
818
+ }, M max
819
+ n
820
+ }.
821
+ (22)
822
+ The optimization results corresponding to each device are
823
+ only related to the parameters of the device itself and are not
824
+ associated with the parameters of other devices.
825
+ B. Optimization Problem Solving for N1
826
+ Then we solve the optimization problem of N1. The prob-
827
+ lem P2 can be written as,
828
+ Problem PN1 (Problem for N1):
829
+ minimize
830
+ {Mn,f e
831
+ n,tn}
832
+
833
+ n∈N1
834
+ F1,n(Mn, f e
835
+ n, tn),
836
+ (23)
837
+ subject to
838
+ (13c), (13d), (13e), (16a).
839
+ Algorithm 1: Algorithm 1: Search-Based Algorithm
840
+ for solving PN1
841
+ Input: The offloading policy N1, the channel gain
842
+ {hn}, and other system parameters.
843
+ Output: {M ⋆
844
+ n, f e⋆
845
+ n , t⋆
846
+ n}
847
+ Initialize the result of cost function F⋆
848
+
849
+ PN1 to a
850
+ sufficiently large value;
851
+ Calculate the achievable data rate {Rn} using (8);
852
+ foreach {Mn} ∈ M do
853
+ Compute F �
854
+ PN1 using (27);
855
+ if F �
856
+ PN1 < F⋆
857
+
858
+ PN1 then
859
+ {M ⋆
860
+ n} = {Mn}; F⋆
861
+
862
+ PN1 = F �
863
+ PN1 ;
864
+ Calculate {f e⋆
865
+ n } and {t⋆
866
+ n} using (25) and (26);
867
+ return {M ⋆
868
+ n}, {f e⋆
869
+ n }, and {t⋆
870
+ n}.
871
+ The optimization variables in the the problem PN1 are the
872
+ number of input video frames Mn, the edge computation
873
+ resource f e
874
+ n, and the proportion of transmission time tn.
875
+ Let {M ∗
876
+ n, f e∗
877
+ n , t∗
878
+ n} denote the optimal solution to PN1. We
879
+ can obtain the optimal solution to PN1 using the method of
880
+ Lagrange multiplier. The partial Lagrangian function can be
881
+ written as,
882
+ LPN1=
883
+
884
+ n∈N1
885
+ �β1ρC(Mn)
886
+ f en
887
+ + β1Mnd
888
+ Rntn
889
+ + β2Mndpn
890
+ Rn
891
+ − β3Φ(Mn)
892
+
893
+ + µ0(
894
+
895
+ n∈N1
896
+ tn − 1) + µ1(
897
+
898
+ n∈N1
899
+ f e
900
+ n − f max),
901
+ (24)
902
+ First of all, according to (24), supposing that M ∗
903
+ n is given,
904
+ we can solve the problem PN1 based on the Karush-Kuhn-
905
+ Tucker (KKT) condition. We can obtain the function expres-
906
+ sions of f e∗
907
+ n and t∗
908
+ n relative to Mn, as shown in the following
909
+ theorem.
910
+ Theorem 2: The function expressions of f e∗
911
+ n and t∗
912
+ n relative
913
+ to M ∗
914
+ n are given by,
915
+ f e∗
916
+ n = f max�
917
+ C(M ∗n)
918
+
919
+ i∈N1
920
+
921
+ C(M ∗
922
+ i )
923
+ ,
924
+ (25)
925
+ t∗
926
+ n =
927
+
928
+ M ∗
929
+ n
930
+ Rn
931
+
932
+ i∈N1
933
+
934
+ M ∗
935
+ i
936
+ Ri
937
+ .
938
+ (26)
939
+ Proof: Please refer to Appendix B.
940
+ Combining (23), (25) and (26), the problem PN1 can be
941
+ written as an optimized function containing only the variable
942
+ Mn as follows,
943
+ Problem �
944
+ PN1 (Mn Optimization Problem for N1 ):
945
+ minimize
946
+ {Mn}
947
+ F �
948
+ PN1 ≜ β1ρ
949
+ f max (
950
+
951
+ n∈N1
952
+
953
+ C(Mn))2
954
+ + β1d(
955
+
956
+ n∈N1
957
+
958
+ Mn
959
+ Rn
960
+ )2 + β2dpn(
961
+
962
+ n∈N1
963
+ Mn
964
+ Rn
965
+ )
966
+
967
+
968
+ n∈N1
969
+ β3Φ(Mn),
970
+ (27)
971
+
972
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
973
+ 7
974
+ Algorithm 2: Algorithm 2: GP-Based Algorithm for
975
+ solving PN1
976
+ Input: The offloading policy N1, the channel gain
977
+ {hn}, and other system parameters.
978
+ Output: {M ⋆
979
+ n, f e⋆
980
+ n , t⋆
981
+ n}
982
+ Calculate the achievable data rate {Rn} using (8);
983
+ Use the CVX tool to solve (29) and get { ˆ
984
+ M ⋆n};
985
+ {M ⋆
986
+ n} = {[e ˆ
987
+ M ⋆
988
+ n]};
989
+ Calculate {f e⋆
990
+ n } and {t⋆
991
+ n} using (25) and (26);
992
+ return {M ⋆
993
+ n}, {f e⋆
994
+ n }, and {t⋆
995
+ n}.
996
+ subject to
997
+ (16a).
998
+ Denote Mopt
999
+ n
1000
+ = {Mn | M min
1001
+ n
1002
+ ≤ Mn ≤ M max
1003
+ n
1004
+ , Mn ∈
1005
+ Z} to be the optional video frame number of device n. The
1006
+ optimal solution can be obtained by searching for {Mn} ∈ M,
1007
+ where M = {{Mi} | Mi ∈ Mopt
1008
+ i
1009
+ , i ∈ N1}. The detail of the
1010
+ search based algorithm is shown in Algorithm 1.
1011
+ Considering that the problem PN1 is convex when Mn is
1012
+ given, Algorithm 1 is global optimal. However, When the
1013
+ number of devices grows large, the computational complexity
1014
+ of the Search-based algorithm will become very high or even
1015
+ unacceptable. In this paper, we also propose a GP-based sub-
1016
+ optimal algorithm to solve the problem PN1. First, we relax
1017
+ the objective function of the problem PN 1. We introduce the
1018
+ function, �Φ(Mn) = − ma,0
1019
+ Mn + ma,2, and PN1 can be rewritten
1020
+ as,
1021
+ Problem PGPN1 (GP-based Problem for N1):
1022
+ minimize
1023
+ {Mn,f e
1024
+ n,tn}
1025
+
1026
+ n∈N1
1027
+
1028
+ β1
1029
+ ρC(Mn)
1030
+ f en
1031
+ + β1
1032
+ Mnd
1033
+ Rntn
1034
+ + β2
1035
+ Mndpn
1036
+ Rn
1037
+ − β3�Φ(Mn)
1038
+
1039
+ ,
1040
+ (28)
1041
+ subject to
1042
+ (13c), (13d), (13e), (16a).
1043
+ It is a non-convex GP problem. Inspired by [30], the GP
1044
+ problem can be transformed into a convex problem by
1045
+ changing variables and transforming the objective and con-
1046
+ straints. Therefore, introducing variables,
1047
+ ˆ
1048
+ Mn = ln Mn, ˆf en =
1049
+ ln f e
1050
+ n, ˆtn = ln tn, and the problem can be written as,
1051
+ Problem �
1052
+ PGPN1 (Converted GP-based Problem for N1):
1053
+ minimize
1054
+ { ˆ
1055
+ Mn, ˆ
1056
+ tn, ˆ
1057
+ f e
1058
+ n}
1059
+
1060
+ n∈N1
1061
+
1062
+ β1ρmc,0e
1063
+ ˆ
1064
+ Mn− ˆ
1065
+ f e
1066
+ n + β1ρmc,1e− ˆ
1067
+ f e
1068
+ n
1069
+ + β1de ˆ
1070
+ Mn− ˆ
1071
+ tn
1072
+ Rn
1073
+ + β2dpne ˆ
1074
+ Mn
1075
+ Rn
1076
+ +β3ma,0e− ˆ
1077
+ Mn
1078
+
1079
+ , (29)
1080
+ subject to
1081
+ ˆ
1082
+ Mn ∈
1083
+
1084
+ ln M min
1085
+ n
1086
+ , ln M max
1087
+ n
1088
+
1089
+ , ∀n ∈ N1,
1090
+ (29a)
1091
+
1092
+ n∈N1
1093
+ xne ˆ
1094
+ tn ≤ 1,
1095
+ (29b)
1096
+
1097
+ n∈N1
1098
+ xne
1099
+ ˆ
1100
+ f e
1101
+ n ≤ f max,
1102
+ (29c)
1103
+ which is strictly convex problem that can be solved using the
1104
+ CVX tool [31]. Considering that Mn is an integer, the result
1105
+ of CVX optimization needs to be post-processed. Details of
1106
+ the GP-based algorithm are shown in Algorithm 2.
1107
+ Algorithm 3: Algorithm 3: Channel-Aware heuristic
1108
+ algorithm for Optimizing Offloading Policy {xn}
1109
+ Input: Parameters corresponding to the problem P1.
1110
+ Output: Offloading policy N0 and N1.
1111
+ Calculate the cost function {F0,n} for the set N using
1112
+ (20) and (21) ;
1113
+ Set N0 = ∅, N1 = N;
1114
+ Calculate the cost function {F1,n} corresponding to
1115
+ the set N1 using Algorithm 1 or Algorithm 2;
1116
+ Set Flag = 1;
1117
+ while Flag == 1 do
1118
+ k = argminnhn, n ∈ N1;
1119
+ N ∗
1120
+ 0 = N0 ∪ {k}, N ∗
1121
+ 1 = N1 − {k};
1122
+ Calculate the cost function {F∗
1123
+ 1,n} corresponding
1124
+ to the set N ∗
1125
+ 1 using Algorithm 1 or Algorithm 2;
1126
+ if �
1127
+ n∈N0 F0,n + �
1128
+ n∈N1 F1,n >
1129
+
1130
+ n∈N ∗
1131
+ 0 F0,n + �
1132
+ n∈N ∗
1133
+ 1 F∗
1134
+ 1,n then
1135
+ F1,n = F∗
1136
+ 1,n, ∀n ∈ N ∗
1137
+ 1 ;
1138
+ N0 = N ∗
1139
+ 0 ; N1 = N ∗
1140
+ 1 ;
1141
+ else
1142
+ Flag = 0;
1143
+ return N0 and N1.
1144
+ C. Optimization of Offloading Policy {xn}
1145
+ Considering the complexity of Search-based offloading pol-
1146
+ icy algorithm becomes high when the number of devices N
1147
+ grows large. In this section, we propose a Channel-Aware
1148
+ heuristic algorithm to optimize the offloading decision {xn}.
1149
+ Inspired by the Theorem 1 and Theorem 2, when executing
1150
+ inference locally, the cost function F0,n and optimization vari-
1151
+ ables f md
1152
+ n , Mn only depend on the device’s own parameters.
1153
+ However, for edge set N1, the cost function is related to
1154
+ the number and parameters of devices in the set N1. The
1155
+ Channel-Aware heuristic algorithm is shown in Algorithm
1156
+ 3. First, calculate the cost function {F0,n} of set N0 when
1157
+ each device’s task is executed locally. Second, assuming that
1158
+ all devices are offloaded to the edge server for inference
1159
+ and |N1| = N. In each iteration, the cost function {F1,n}
1160
+ corresponding to each device of N1 is obtained. We select the
1161
+ device k with smallest channel gain in set N1. Try to put the
1162
+ device k from the set N1 into the set N0 and compute the cost
1163
+ of new sets. If the total cost of new sets is reduced, continue
1164
+ the next iteration. Otherwise, put the device k back to the set
1165
+ N1.
1166
+ V. JOINT OPTIMIZATION USING ADMM-BASED
1167
+ METHOD
1168
+ The complexity of the Channel-Aware heuristic algorithm
1169
+ becomes high when the number of UE grows. In this section,
1170
+ We propose an ADMM-based algorithm. The ADMM-based
1171
+ algorithm can decompose P2 into N parallel sub-problems.
1172
+ Each user only needs to solve one sub-problem, and the
1173
+ average complexity of each device will be reduced.
1174
+
1175
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
1176
+ 8
1177
+ A. ADMM-based Problem Conversion
1178
+ To make the original problem tractable, we jointly consider
1179
+ the problem P2 and problem �
1180
+ PGPN1 , and we converted the
1181
+ problem into a GP-based problem,
1182
+ Problem P3 (Converted GP-based Problem):
1183
+ minimize
1184
+
1185
+ ˆ
1186
+ Mn, ˆ
1187
+ tn, ˆ
1188
+ f md
1189
+ n
1190
+ , ˆ
1191
+ f e
1192
+ n,xn
1193
+
1194
+
1195
+ n∈N
1196
+
1197
+ (1 − xn) ˆ
1198
+ F0,n( ˆ
1199
+ Mn,
1200
+ ˆ
1201
+ f md
1202
+ n )
1203
+ + xn ˆ
1204
+ F1,n( ˆ
1205
+ Mn, ˆf en, ˆtn)
1206
+
1207
+ ,
1208
+ (30)
1209
+ subject to
1210
+ ˆ
1211
+ f md
1212
+ n
1213
+ ≤ ln f max
1214
+ n
1215
+ , ∀n ∈ N,
1216
+ (30a)
1217
+ (13g), (29a) − (29c),
1218
+ where
1219
+ ˆ
1220
+ Mn = ln Mn,
1221
+ ˆ
1222
+ f md
1223
+ n
1224
+ = ln f md
1225
+ n , ˆf en = ln f e
1226
+ n, and ˆtn =
1227
+ ln tn.
1228
+ ˆ
1229
+ F0,n( ˆ
1230
+ Mn,
1231
+ ˆ
1232
+ f md
1233
+ n ) and
1234
+ ˆ
1235
+ F1,n( ˆ
1236
+ Mn, ˆf en, ˆtn) are given by,
1237
+ ˆ
1238
+ F0,n( ˆ
1239
+ Mn,
1240
+ ˆ
1241
+ f md
1242
+ n ) = β1ρmc,0e
1243
+ ˆ
1244
+ Mn− ˆ
1245
+ f md
1246
+ n
1247
+ + β1ρmc,1e− ˆ
1248
+ f md
1249
+ n
1250
+ + β2κmc,0e
1251
+ ˆ
1252
+ Mn+2 ˆ
1253
+ f md
1254
+ n
1255
+ + β2κmc,1e2 ˆ
1256
+ f md
1257
+ n
1258
+ + β3ma,0e− ˆ
1259
+ Mn, (31)
1260
+ ˆ
1261
+ F1,n( ˆ
1262
+ Mn, ˆf en, ˆtn) = β1ρmc,0e
1263
+ ˆ
1264
+ Mn− ˆ
1265
+ f e
1266
+ n + β1ρmc,1e− ˆ
1267
+ f e
1268
+ n
1269
+ + β1de ˆ
1270
+ Mn− ˆ
1271
+ tn
1272
+ Rn
1273
+ + β2dpne ˆ
1274
+ Mn
1275
+ Rn
1276
+ + β3ma,0e− ˆ
1277
+ Mn,
1278
+ (32)
1279
+ The optimization variables { ˆtn, ˆf en} are coupled among the
1280
+ devices in the constraints (29b) and (29c). To decompose the
1281
+ problem P3, we introduce local variables {yn} and {zn}.
1282
+ Then, the ADMM-based problem can be written as,
1283
+ Problem P4 (ADMM-based Problem):
1284
+ minimize
1285
+
1286
+ ˆ
1287
+ Mn, ˆ
1288
+ tn, ˆ
1289
+ f md
1290
+ n
1291
+ , ˆ
1292
+ f e
1293
+ n,xn,yn,zn
1294
+
1295
+
1296
+ n∈N
1297
+ ˆ
1298
+ Fn(xn, ˆ
1299
+ Mn,
1300
+ ˆ
1301
+ f md
1302
+ n , yn, zn)
1303
+ + g( ˆf en, ˆtn),
1304
+ (33)
1305
+ subject to yn = ˆf en, zn = ˆtn,
1306
+ (33a)
1307
+ (13g), (29a), (30a),
1308
+ where,
1309
+ ˆ
1310
+ Fn(xn, ˆ
1311
+ Mn,
1312
+ ˆ
1313
+ f md
1314
+ n , yn, zn) = (1 − xn) ˆ
1315
+ F0,n( ˆ
1316
+ Mn,
1317
+ ˆ
1318
+ f md
1319
+ n )
1320
+ + xn ˆ
1321
+ F1,n( ˆ
1322
+ Mn, xn, yn),
1323
+ (34)
1324
+ g( ˆf en, ˆtn) =
1325
+
1326
+ 0,
1327
+ if( ˆf en, ˆtn) ∈ G,
1328
+ +∞
1329
+ , otherwise,
1330
+ (35)
1331
+ and,
1332
+ G =
1333
+
1334
+ ( ˆf en, ˆtn)|
1335
+
1336
+ n∈N1
1337
+ xne ˆ
1338
+ tn ≤ 1,
1339
+
1340
+ n∈N1
1341
+ xne
1342
+ ˆ
1343
+ f e
1344
+ n ≤ f max
1345
+
1346
+ . (36)
1347
+ B. ADMM-based Problem Solving
1348
+ The problem P4 can be effectively solved using the ADMM
1349
+ algorithm. We can write a partial augmented Lagrangian of the
1350
+ problem P4 as,
1351
+ L4(u, v, θ) =
1352
+
1353
+ n∈N
1354
+ ˆ
1355
+ Fn(xn, ˆ
1356
+ Mn,
1357
+ ˆ
1358
+ f md
1359
+ n , yn, zn) + g( ˆf en, ˆtn)
1360
+ +
1361
+
1362
+ n∈N
1363
+ θf
1364
+ n(yn − ˆf en) +
1365
+
1366
+ n∈N
1367
+ θt
1368
+ n(zn − ˆtn)
1369
+ +
1370
+
1371
+ n∈N
1372
+ s
1373
+ 2(yn − ˆf en)2 +
1374
+
1375
+ n∈N
1376
+ s
1377
+ 2(zn − ˆtn)2,
1378
+ (37)
1379
+ where u = {xn, ˆ
1380
+ Mn,
1381
+ ˆ
1382
+ f md
1383
+ n , yn, zn}, v = { ˆf en, ˆtn}, θ =
1384
+ {θf
1385
+ n, θt
1386
+ n}, and s is a fixed step size. Therefore, the dual function
1387
+ is,
1388
+ p(θ) = minimize
1389
+ u,v
1390
+ L4(u, v, θ)
1391
+ (38)
1392
+ subject to (13g), (29a), (30a),
1393
+ and the dual problem can be given by,
1394
+ maximize
1395
+ θ
1396
+ p(θ),
1397
+ (39)
1398
+ The problem (38) can be solved by iteratively updating
1399
+ u, v, and θ [32]. Let {ui, vi, θi} denote the values in the
1400
+ ith iteration. In the ith iteration, the update strategies of the
1401
+ variables are as follows,
1402
+ 1) Step 1: Local variables update. In this step, we first
1403
+ update the local variables u. Given variable vi and θi, we
1404
+ minimize L4(u, v, θ) by,
1405
+ ui+1 = argminimize
1406
+ u
1407
+ L4(u, vi, θi).
1408
+ (40)
1409
+ The problem (39) can be decomposed into N parallel sub-
1410
+ problems. For each subproblem, we consider two cases where
1411
+ xn = 0 and xn = 1, and express the problem as,
1412
+
1413
+
1414
+
1415
+
1416
+
1417
+
1418
+
1419
+
1420
+
1421
+
1422
+
1423
+
1424
+
1425
+
1426
+
1427
+
1428
+
1429
+
1430
+
1431
+
1432
+
1433
+
1434
+
1435
+
1436
+
1437
+
1438
+
1439
+ minimize
1440
+ { ˆ
1441
+ Mn, ˆ
1442
+ f md
1443
+ n
1444
+ ,yn,zn}
1445
+ ˆ
1446
+ F0,n( ˆ
1447
+ Mn,
1448
+ ˆ
1449
+ f md
1450
+ n ) =θf
1451
+ nyn +
1452
+
1453
+ n∈N
1454
+ s
1455
+ 2(yn − ˆf en)2
1456
+ +θt
1457
+ nzn +
1458
+
1459
+ n∈N
1460
+ s
1461
+ 2(zn − ˆtn)2,
1462
+ if xn = 0,
1463
+ minimize
1464
+ { ˆ
1465
+ Mn,yn,zn}
1466
+ ˆ
1467
+ F1,n( ˆ
1468
+ Mn, yn, zn)=θf
1469
+ nyn +
1470
+
1471
+ n∈N
1472
+ s
1473
+ 2(yn − ˆf en)2
1474
+ +θt
1475
+ nzn +
1476
+
1477
+ n∈N
1478
+ s
1479
+ 2(zn − ˆtn)2,
1480
+ if xn = 1.
1481
+ (41)
1482
+ These problems are both strictly convex problems that can
1483
+ be solved using the CVX tool [31]. Therefore, we can cal-
1484
+ culate the objective value for xn = 0 and xn = 1 and
1485
+ choose the smaller one as the final result. After solving N
1486
+ parallel subproblems, the optimal solution to (40) is given by
1487
+ ui+1 = {(xn)i+1, ( ˆ
1488
+ Mn)i+1, ( ˆ
1489
+ f md
1490
+ n )i+1, (yn)i+1, (zn)i+1}.
1491
+ 2) Step 2: Global variables update. In the second step, we
1492
+ update the global variables v. By the definition of g(v) in
1493
+ (35), vi+1 ∈ G must hold at the optimum. Therefore, the
1494
+ subproblem can be equivalently written as,
1495
+ vi+1 = argminimize
1496
+ { ˆ
1497
+ f e
1498
+ n, ˆ
1499
+ tn}
1500
+
1501
+ n∈N
1502
+ (θf
1503
+ n)i(− ˆf en) +
1504
+
1505
+ n∈N
1506
+ (θt
1507
+ n)i(− ˆtn)
1508
+ +
1509
+
1510
+ n∈N
1511
+ s
1512
+ 2(yi+1
1513
+ n
1514
+ − ˆf en)2 +
1515
+
1516
+ n∈N
1517
+ s
1518
+ 2(zi+1
1519
+ n
1520
+ − ˆtn)2,
1521
+ (42)
1522
+ subject to,
1523
+ (29b), (29c).
1524
+ The problem can also be solved by the CVX tool [31]. We
1525
+ propose a low-complexity scheme to solve this subproblem.
1526
+
1527
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
1528
+ 9
1529
+ Algorithm 4: Algorithm 4: ADMM-Based Algorithm
1530
+ Input: Parameters corresponding to the problem P1.
1531
+ Output: {xn, Mn, f md
1532
+ n , f e
1533
+ n, tn}
1534
+ Initialize i = 0, {ui, vi, θi} = 0, s = 0.5,
1535
+ µ⋆
1536
+ f = µ⋆
1537
+ t = 106, δ = 10−4;
1538
+ repeat
1539
+ foreach n ∈ N do
1540
+ Update ui+1 by solving (41) and choose
1541
+ smaller results;
1542
+ foreach n ∈ N do
1543
+ Update global variables vi+1 using (43) and
1544
+ (44);
1545
+ foreach n ∈ N do
1546
+ Update multipliers θi+1 using (45) and (46);
1547
+ i = i + 1;
1548
+ until |Fi − Fi+1| < δ;
1549
+ Mn = e ˆ
1550
+ Mn, f md
1551
+ n
1552
+ = e
1553
+ ˆ
1554
+ f md
1555
+ n , f e
1556
+ n = e ˆ
1557
+ f e
1558
+ n, tn = e ˆ
1559
+ tn;
1560
+ return {xn, Mn, f md
1561
+ n , f e
1562
+ n, tn}.
1563
+ Considering the constraints (29b) and (29c), let µf and µt
1564
+ denote the Lagrangian multipliers. The closed-form optimal
1565
+ solution of this subproblem can be expressed as,
1566
+ ( ˆf en)i+1 =yn
1567
+ i+1 + (θf
1568
+ n)i − µf
1569
+ s
1570
+ ,
1571
+ (43)
1572
+ ( ˆtn)i+1 =zn
1573
+ i+1 + (θt
1574
+ n)i − µt
1575
+ s
1576
+ ,
1577
+ (44)
1578
+ where µf can be obtained by the bisection search method
1579
+ over (0, µ⋆
1580
+ f), until �
1581
+ n∈N1 xne ˆ
1582
+ f e
1583
+ n ≤ f max satisfies. µ⋆
1584
+ f is a
1585
+ sufficiently large value. It is because when µf ≥ 0, ( ˆf en)i+1 is
1586
+ non-increasing. Similarly, µt can be obtained by the bisection
1587
+ search method over (0, µ⋆
1588
+ t ), where µ⋆
1589
+ t is a sufficiently large
1590
+ value, until �
1591
+ n∈N1 xne ˆ
1592
+ tn ≤ 1 satisfies.
1593
+ 3) Step 3: Multipliers update. In this step, we update the
1594
+ multipliers θ using the obtained global variables v and local
1595
+ variables u. The updated method is,
1596
+ (θf
1597
+ n)i+1 =(θf
1598
+ n)i + s(yi+1
1599
+ n
1600
+ − ( ˆf en)i+1),
1601
+ (45)
1602
+ (θt
1603
+ n)i+1 =zn
1604
+ i+1 + s(zi+1
1605
+ n
1606
+ − ( ˆtn)i+1),
1607
+ (46)
1608
+ Repeat the above three steps until the cost function no
1609
+ longer decreases. The cost function is Fi = �
1610
+ n∈N [(1 −
1611
+ xi
1612
+ n) ˆ
1613
+ F0,n(( ˆ
1614
+ Mn)i, (
1615
+ ˆ
1616
+ f md
1617
+ n )i) + xi
1618
+ n ˆ
1619
+ F1,n(( ˆ
1620
+ Mn)i, ( ˆf en)i, ( ˆtn)i)]. We
1621
+ summarize solving steps of the ADMM algorithm as Algo-
1622
+ rithm 4.
1623
+ As a distributed iterative algorithm, the ADMM-based
1624
+ scheme performs iterations between devices and BS rather than
1625
+ locally, enabling online optimization during the recognition
1626
+ process. In each iteration, ui is calculated locally and sent
1627
+ to the MEC. After receiving ui from all devices, the MEC
1628
+ updates vi and θi, and sends them to the device to complete
1629
+ an iteration. Therefore, the iteration of the ADMM algorithm is
1630
+ an online convergence process that can adapt to slight changes
1631
+ in the channel.
1632
+ 0
1633
+ 4
1634
+ 8
1635
+ 12
1636
+ 16
1637
+ The number of input video frames
1638
+ 0
1639
+ 100
1640
+ 200
1641
+ 300
1642
+ 400
1643
+ 500
1644
+ 600
1645
+ Latency / ms
1646
+ Theory, Resnet-18, 2.8G
1647
+ Experiment, Resnet-18, 2.8G
1648
+ Fitting, Resnet-18, 2.8G
1649
+ Theory, Resnet-34, 2.8G
1650
+ Experiment, Resnet-34, 2.8G
1651
+ Fitting, Resnet-34, 2.8G
1652
+ Theory, Resnet-18, 2.2G
1653
+ Experiment, Resnet-18, 2.2G
1654
+ Fitting, Resnet-18, 2.2G
1655
+ m
1656
+ m
1657
+ c,0=16.6
1658
+ c,1=40.0
1659
+ c,0=20.0
1660
+ c,1=48.1
1661
+ m
1662
+ m
1663
+ m
1664
+ m
1665
+ c,0=26.1
1666
+ c,1=79.7
1667
+ Fig. 3.
1668
+ The theoretical delay curve, the experimental delay curve and the
1669
+ fitted curve corresponding to the experimentalal delay. Resnet-18 and Resnet-
1670
+ 34 are two classic neural network architectures. The frequency of the CPU is
1671
+ 2.8G and 2.2G.
1672
+ C. Algorithm Computational Complexity Analysis
1673
+ In this part, we analyze the computational complexity of
1674
+ proposed algorithms. First, the complexity of solving problem
1675
+ PN0 is O(|N0|). Second, as mentioned above, the complexity
1676
+ of Algorithm 1 is O(�
1677
+ n∈N1 |Mopt
1678
+ n |), and the complexity of
1679
+ Algorithm 2 is O((3|N1|)3.5) by the interior-point method
1680
+ according to [33]. When we use Algorithm 1 for solving
1681
+ PN1 and use Search-based algorithm for optimizing offloading
1682
+ policy, the computational complexity is O(2N �
1683
+ n∈N |Mopt
1684
+ n |).
1685
+ When we use Algorithm 1 for solving PN1 and use Algorithm
1686
+ 3 for optimizing offloading policy, the computational com-
1687
+ plexity is O(N �
1688
+ n∈N |Mopt
1689
+ n |). In addition, the computational
1690
+ complexity of Algorithm 2 for solving PN1 and Algorithm 3
1691
+ for optimizing offloading policy is O(N 4.5). For the ADMM-
1692
+ based algorithm, as the complexity of each steps is O(N), the
1693
+ overall complexity of one iteration is O(N).
1694
+ VI. NUMERICAL RESULTS
1695
+ In this section, we evaluate the performance of the proposed
1696
+ algorithms via simulations. For all the simulation results,
1697
+ unless specified otherwise, we set the downlink bandwidth
1698
+ as Bw = 5 MHz and the power spectral as N0 = −174
1699
+ dBm/Hz [24]. According to [17], the path loss is modelled
1700
+ as PL = 128.1 + 37.6 log10(D) dB, where D is the dis-
1701
+ tance between the device and the BS in kilometres. Devices
1702
+ randomly distributed in the area within [500m 500m]. The
1703
+ computational resource of the MEC server and devices are
1704
+ set to be 1.8 GHz and 22 GHz, respectively. The recognition
1705
+ accuracy requirement and the maximum number of input video
1706
+ frames are set to αn = 0.86 and M max
1707
+ n
1708
+ = 16, respectively.
1709
+ The coefficient κ is determined by the corresponding device
1710
+ and is set to be 10−28 in this paper according to [24]. The size
1711
+ of the input video is 112∗112∗Mn. In addition, the coefficient
1712
+ of computational complexity ρ is set to be 0.12 cycle/MAC,
1713
+ which is obtained through several experiments in Sec.VI-A.
1714
+ Weights β1, β2, β3 are set to be 0.2, 0.2, 0.6, respectively.
1715
+ A. Model Verification
1716
+ First, we obtain the complexity coefficient through experi-
1717
+ mental measurement. The calculation method of the compu-
1718
+ tational complexity coefficient is as follows. First, calculate
1719
+
1720
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
1721
+ 10
1722
+ 0
1723
+ 4
1724
+ 8
1725
+ 12
1726
+ 16
1727
+ The number of input video frames
1728
+ 0.4
1729
+ 0.55
1730
+ 0.7
1731
+ 0.85
1732
+ 1
1733
+ Accuracy
1734
+ Fitted curve, gesture, resnet-101
1735
+ Fitted curve, gesture, resnet-18
1736
+ Fitted curve, action, resnet-101
1737
+ Experiment, gesture, resnet-101
1738
+ Experiment, gesture, resnet-18
1739
+ Experiment, action, resnet-101
1740
+ a,0=0.910
1741
+ a,1=1.400
1742
+ a,2=0.993
1743
+ a,0=0.910
1744
+ a,1=1.400
1745
+ a,2=0.993
1746
+ m
1747
+ m
1748
+ m
1749
+ a,0=0.707
1750
+ a,1=0.939
1751
+ a,2=0.990
1752
+ m
1753
+ m
1754
+ m
1755
+ m
1756
+ m
1757
+ m
1758
+ Fig. 4.
1759
+ The experimental and fitted curves of gesture recognition task and
1760
+ action recognition task.
1761
+ the MACs of the DNN model when the number of input
1762
+ video frames is different, recorded as {C}. We use the Flops
1763
+ Counter tool [34] for MACs calculation. Second, execute 100
1764
+ times of inference tasks with a different number of input
1765
+ video frames, and record the average inference delay as {t}.
1766
+ Finally, calculate the coefficients between the inference delay
1767
+ and MACs by ρ = sum({C})
1768
+ sum({t}) . We use Intel(R) Xeon(R) E5-
1769
+ 2630 CPU for testing. We use the Resnet-18 and the Resnet-34
1770
+ for testing and limit the maximum frequency of the CPU to
1771
+ 2.8G and 2.2G. Fig. 3 shows the theoretical (MAC-based) and
1772
+ experimental delay curves and the fitted curve corresponding
1773
+ to the experimental delay. We can observe from Fig. 3 that the
1774
+ theoretical delay is similar to the experimental delay, proving
1775
+ that MACs can be modelled as computational complexity.
1776
+ We also find that the linear fitted curve can approximately
1777
+ represent the computational complexity with 9 ms root mean
1778
+ square error (RMSE) for Resnet-18 and 2.8G, 17 ms RMSE
1779
+ for Resnet-34 and 2.8G, and 11 ms RMSE for Resnet-18
1780
+ and 2.2G. The inference delay is associated with the number
1781
+ of input frames, DNN model’s architecture and the device’s
1782
+ capabilities. In addition, the computational complexity co-
1783
+ efficients under the three conditions are 0.128, 0.122, and
1784
+ 0.123, respectively. Therefore, in following experiments, we
1785
+ set ρ = 0.12 cycle/MAC.
1786
+ We select the gesture and action recognition tasks to verify
1787
+ the accuracy model. We use the Jester datasets [35], the
1788
+ largest publicly available hand gesture dataset, to test the
1789
+ gesture recognition task. For the action recognition task, we
1790
+ use Kinetics-400 datasets [36]. We choose Resnet-18 and
1791
+ Resnet-101 for testing. As shown in Fig. 4, Under different
1792
+ tasks and different network models, the accuracy curve all
1793
+ conforms to the characteristics of a non-decreasing function.
1794
+ What’s more, as the number of input frames increases, the
1795
+ performance gain of accuracy will gradually decrease. This
1796
+ is because the information gain introduced in the temporal
1797
+ domain decreases when the number of input frames increases.
1798
+ The fitted curve can approximately represent the relationship
1799
+ between the accuracy and the number of input frames. In
1800
+ the gesture recognition task with the Resnet-101 model, the
1801
+ gesture recognition task with the Resnet-18 model, and the
1802
+ action recognition task with the Resnet-101 model, the RMSE
1803
+ are 0.0054, 0.0048 and 0.0095, respectively. We take the
1804
+ 4
1805
+ 12
1806
+ 20
1807
+ 28
1808
+ 36
1809
+ The number of devices
1810
+ -0.5
1811
+ -0.4
1812
+ -0.3
1813
+ -0.2
1814
+ -0.1
1815
+ Average cost
1816
+ Local
1817
+ Edge
1818
+ Random
1819
+ CCCP
1820
+ ADMM
1821
+ GP+Heuristic
1822
+ Performance bounds
1823
+ Performance loss
1824
+ CCCP: 2.1%
1825
+ ADMM: 0.24%
1826
+ GP+Heuristic: 0.03%
1827
+ Fig. 5. The average cost of proposed schemes and baseline schemes under a
1828
+ different number of devices.
1829
+ Resnet-18 and the gesture recognition task as examples for
1830
+ the following experiments.
1831
+ B. Simulation Results of Average Cost
1832
+ In this section, we compare proposed schemes and some
1833
+ baseline schemes. We run 100 tests and can calculate the
1834
+ average cost of each device and the average running time of
1835
+ each test. We compare the following schemes.
1836
+ 1) Search+Search: We use the Search-based algorithm
1837
+ to solve PN1 and use the heuristic algorithm to optimize
1838
+ offloading policy.
1839
+ 2) Search+Heuristic: We use the Search-based algorithm
1840
+ to solve PN1 and use the Search-based algorithm to optimize
1841
+ offloading policy.
1842
+ 3) GP+Heuristic: We use the GP-based algorithm to solve
1843
+ PN1 and use the Channel-Aware heuristic algorithm to opti-
1844
+ mize offloading policy.
1845
+ 4) ADMM: We use the ADMM-based algorithm to solve
1846
+ the original problem.
1847
+ 5) CCCP [37]: We use the concave-convex procedure
1848
+ (CCCP) algorithm to decide whether to offload inference tasks
1849
+ to edge servers. Then we use Theorem 1 and the GP-based
1850
+ algorithm for resource allocation.
1851
+ 6) Random: All inference tasks are randomly executed on
1852
+ local or the edge server. We use Theorem 1 and the GP-based
1853
+ algorithm for resource allocation.
1854
+ 7) Local: All inference tasks are executed locally. We use
1855
+ Theorem 1 for local resource allocation.
1856
+ 8) Edge: All inference tasks are executed on the edge
1857
+ server. We use the GP-based algorithm for resource allocation
1858
+ In Fig. 5, we plot the average cost of different schemes
1859
+ under different devices. The Search+Heuristic scheme and
1860
+ Search+Search scheme have the same performance, represent-
1861
+ ing the performance bounds. When the number of devices
1862
+ exceeds 16, the performance bounds are not shown due to
1863
+ their unacceptable computational complexity. It can be seen
1864
+ from Fig. 5 that the proposed schemes are better than the
1865
+ baseline schemes. Compared with the performance bounds, the
1866
+ performance of the GP+Heuristic scheme has a slight decrease
1867
+ due to the relaxation of the accuracy function Φ(Mn). The
1868
+ performance of the ADMM scheme is worse than that of the
1869
+ GP+Heuristic scheme, and is better than that of the CCCP
1870
+
1871
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
1872
+ 11
1873
+ 2
1874
+ 6
1875
+ 10
1876
+ 14
1877
+ 18
1878
+ 22
1879
+ The number of devices
1880
+ 0
1881
+ 50
1882
+ 100
1883
+ 150
1884
+ 200
1885
+ 250
1886
+ Running time / s
1887
+ ADMM
1888
+ GP+Heuristic
1889
+ CCCP
1890
+ Search+Heuristic
1891
+ Search+Search
1892
+ Fig. 6. The average running time of proposed algorithms under a different
1893
+ number of devices.
1894
+ scheme. For example, when the number of devices is 16, the
1895
+ CCCP, ADMM, and GP+Heuristic schemes have performance
1896
+ losses of 2.1%, 0.24%, and 0.03%, respectively, compared
1897
+ with performance bounds. Moreover, when the number of
1898
+ devices is less than 8, the cost of the scheme that executes
1899
+ tasks only at the edge is almost equal to the cost of the
1900
+ proposed GP+Heuristic scheme. It is because all devices can
1901
+ benefit from performing inference on the edge server when
1902
+ the number of devices is small. If the inference task is only
1903
+ executed locally, the average cost of the device will not change
1904
+ because the local resources among the equipment do not affect
1905
+ each other.
1906
+ In Fig. 6, we plot the average running time of different
1907
+ schemes under different devices. When the number of de-
1908
+ vices exceeds 6, the running time of the Search+Heuristic
1909
+ and Search+Search scenarios becomes unacceptable. The
1910
+ GP+Heuristic scheme improves the solution efficiency. The
1911
+ running time of GP+Heuristic is shorter than that of CCCP
1912
+ scheme. However, the complexity of the solution remains
1913
+ unsatisfactory as the number of devices increases. As for the
1914
+ ADMM-based scheme, since the ADMM-based algorithm is
1915
+ a distributed algorithm and the complexity of updating global
1916
+ variables is much smaller than that of updating local variables,
1917
+ we only consider the average running time for each device.
1918
+ The average running time of the ADMM-based scheme does
1919
+ not improve as the number of devices increases. It is worth
1920
+ noting that in the ADMM-based scheme, the iteration stops
1921
+ when |Fi − Fi+1| < δ, where δ = 10−5. Threshold-based
1922
+ stopping conditions result in a different number of iterations
1923
+ in different cases. When the number of devices is different,
1924
+ the average number of iterations is also different, resulting in
1925
+ different running times. Therefore, the average running time
1926
+ of 18 devices is shorter than that of 14 and 22 devices.
1927
+ Assuming that the ADMM-based scheme iterates once every
1928
+ time an inference task is performed, we plot the curve corre-
1929
+ sponding to the cost function and the number of iterations.
1930
+ As shown in Fig.7, the ADMM-based scheme can converge
1931
+ to acceptable performance after completing 3-5 iterations. As
1932
+ the number of iterations increases, the performance will be
1933
+ closer to the optimal performance. It shows that the ADMM
1934
+ algorithm can converge through online iterations. We also test
1935
+ the running time per iteration on each device, and it takes an
1936
+ average of about 278ms.
1937
+ 0
1938
+ 5
1939
+ 10
1940
+ 15
1941
+ 20
1942
+ The number of iterations
1943
+ -0.5
1944
+ 0
1945
+ 0.5
1946
+ 1
1947
+ 1.5
1948
+ 2
1949
+ Cost
1950
+ ADMM
1951
+ Performance bounds
1952
+ Fig. 7.
1953
+ The curve corresponding to the cost function and the number of
1954
+ iterations.
1955
+ TABLE I
1956
+ DELAY, ENERGY CONSUMPTION, AND ACCURACY OF LOCAL DEVICES
1957
+ AND EDGE DEVICES
1958
+ Local devices
1959
+ Edge devices
1960
+ Number of devices
1961
+ 12.3
1962
+ 12.7
1963
+ Average delay
1964
+ 0.24 s
1965
+ 0.52 s
1966
+ Average energy
1967
+ 1.00 J
1968
+ 0.025 J
1969
+ Average accuracy
1970
+ 0.886
1971
+ 0.866
1972
+ C. Simulation Results of Delay, Energy, and Accuracy
1973
+ This section compares the average delay, energy consump-
1974
+ tion, accuracy, and the offloading rate (the proportion of de-
1975
+ vices that perform inference on the edge server). We consider
1976
+ the different number of devices, bandwidths, edge computing
1977
+ resources, and weights β1, β2, β3. We use the GP+Heuristic
1978
+ scheme for testing. Table. I shows a comparison of devices
1979
+ that finish inference locally and devices that finish inference
1980
+ at the edge under default experimental settings. On average,
1981
+ 12.7 devices choose to offload to the edge server to perform
1982
+ inference. Compared with edge devices, local devices have a
1983
+ lower delay and higher accuracy but have greater inference
1984
+ energy consumption.
1985
+ Fig. 8 shows the average delay, energy, accuracy, and
1986
+ offloading rate under different numbers of devices, different
1987
+ bandwidths, and different edge computing resources. In Fig.
1988
+ 8(a), we plot results with different numbers of devices. As
1989
+ shown in Fig. 8(a), when the number of devices is small (less
1990
+ than 10), all devices offload the task to the edge server (the
1991
+ offloading rate is equal to 1). For edge devices, all delay comes
1992
+ from transmission delay and the edge inference delay, and all
1993
+ energy consumption comes from transmission energy. With
1994
+ the number of devices increasing, communication resources
1995
+ and the edge server’s computation resources are shared by
1996
+ more devices, decreasing the number of input frames Mn. A
1997
+ decrease in the number of input frames results in a decrease
1998
+ in accuracy. Then as Mn decreases, the transmission data size
1999
+ decreases, and the transmission energy decreases. Meanwhile,
2000
+ Competition from more devices leads to increased delays.
2001
+ Therefore, when the number of devices is small (less than
2002
+ 10), with the number of devices increasing, the average delay
2003
+ increases, the average accuracy and the average energy con-
2004
+
2005
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
2006
+ 12
2007
+ 5
2008
+ 15
2009
+ 25
2010
+ 35
2011
+ 45
2012
+ The number of devices
2013
+ 0
2014
+ 0.2
2015
+ 0.4
2016
+ 0.6
2017
+ 0.8
2018
+ 1
2019
+ Delay /s, Energy / J, Offloading rate
2020
+ 0.8
2021
+ 0.85
2022
+ 0.9
2023
+ 0.95
2024
+ 1
2025
+ Accuracy
2026
+ Delay
2027
+ Energy
2028
+ Offloading rate
2029
+ Accuracy
2030
+ (a) Different number of devices
2031
+ 2
2032
+ 6
2033
+ 10
2034
+ 14
2035
+ 18
2036
+ Bandwidth / MHz
2037
+ 0
2038
+ 0.2
2039
+ 0.4
2040
+ 0.6
2041
+ 0.8
2042
+ 1
2043
+ Delay /s, Energy / J, Offloading rate
2044
+ 0.8
2045
+ 0.85
2046
+ 0.9
2047
+ 0.95
2048
+ 1
2049
+ Accuracy
2050
+ Delay
2051
+ Energy
2052
+ Offloading rate
2053
+ Accuracy
2054
+ (b) Different bandwidth
2055
+ 6
2056
+ 14
2057
+ 22
2058
+ 30
2059
+ 38
2060
+ Edge computing resource / GHz
2061
+ 0
2062
+ 0.2
2063
+ 0.4
2064
+ 0.6
2065
+ 0.8
2066
+ 1
2067
+ Delay /s, Energy / J, Offloading rate
2068
+ 0.8
2069
+ 0.85
2070
+ 0.9
2071
+ 0.95
2072
+ 1
2073
+ Accuracy
2074
+ Delay
2075
+ Energy
2076
+ Offloading rate
2077
+ Accuracy
2078
+ (c) Different edge computing resource
2079
+ Fig. 8. The average delay, energy, offloading rate, and accuracy under different
2080
+ numbers of devices, different bandwidths, and different edge computing
2081
+ resources.
2082
+ sumption decrease. When the number of devices exceeds 10,
2083
+ the average energy consumption and accuracy increase, and the
2084
+ average delay and offload rate gradually decrease. Considering
2085
+ different bandwidths and different edge computing resources,
2086
+ we plot Fig. 8(b) and Fig. 8(c). In Fig. 8(b) and Fig. 8(c), as
2087
+ the bandwidth and edge computing resource increase, devices
2088
+ will be more inclined to offload computing to the edge, which
2089
+ increases the offloading rate. According to Table. I, when
2090
+ β1, β2 and β3 are fixed, edge devices have lower energy
2091
+ consumption, lower accuracy and higher delay. More edge
2092
+ devices mean a greater delay and lower power consumption.
2093
+ Meanwhile, when the bandwidth increases, since the edge
2094
+ computing resources are fixed, the number of video frames
2095
+ will decrease to reduce edge computing overhead, resulting
2096
+ in a decrease in accuracy. The same conclusion can also be
2097
+ obtained when edge computing resources increase. Therefore,
2098
+ Fig. 9. The relationship between the delay, energy consumption, and accuracy.
2099
+ with the increase of bandwidth and edge computing resources,
2100
+ more edge devices lead to increased delay and decreased
2101
+ energy and accuracy.
2102
+ We set the minimum number of input frames M min
2103
+ n
2104
+ = 1.
2105
+ We use different weights, β1, β2, β3 to study the trade-off
2106
+ relationship between the average delay, energy consumption,
2107
+ and accuracy. The constraint is β1 + β2 + β3 = 1. The perfor-
2108
+ mance of the trade-off surface is obtained by the GP+Heuristic
2109
+ scheme. Fig. 9 shows the delay, energy consumption, and
2110
+ accuracy are mutually limited. Higher energy consumption
2111
+ leads to higher accuracy when the delay is constant. From
2112
+ another perspective, in order to improve the accuracy, it is
2113
+ necessary to sacrifice the performance of delay and energy
2114
+ consumption. In addition, with the same accuracy, according
2115
+ to Table. I, higher energy consumption will make the device
2116
+ more inclined to execute inference tasks locally, and the delay
2117
+ decreases.
2118
+ VII. CONCLUSION
2119
+ This paper considers optimizing video-based AI inference
2120
+ tasks in a multi-user MEC system. An MINLP is formulated
2121
+ to minimize the total delay and energy consumption, and
2122
+ improve the total accuracy, with the constraint of computation
2123
+ and communication resources. A MAC-based computational
2124
+ complexity model is introduced to model the calculation delay,
2125
+ and a simple approximate expression is proposed to simplify
2126
+ the problem. We also propose a general accuracy model to
2127
+ characterize the relation between the recognition accuracy and
2128
+ the number of input frames. After that, we first assume that the
2129
+ offloading decision is given and decouple the original problem
2130
+ into two sub-problems. The first sub-problem is to optimize the
2131
+ resources of the devices that complete the DNN inference tasks
2132
+ locally. We derive the closed-form solution to this problem.
2133
+ The second sub-problem is optimizing the devices’ resources
2134
+ that offload the DNN inference tasks to the edge server. We
2135
+ propose the Search-based and GP-based algorithm to solve
2136
+ the second sub-problem. For the problem of offloading de-
2137
+ cision optimization, we propose the Channel-Aware heuristic
2138
+ algorithm. We also propose a distributed algorithm based on
2139
+ ADMM. The ADMM-based algorithm reduce computational
2140
+ complexity at the cost of an acceptable performance loss.
2141
+ Numerical simulation and experimental results demonstrate
2142
+
2143
+ 1
2144
+ 0.9Accurac
2145
+ 0.8
2146
+ 0.7
2147
+ 0.6
2148
+ 2
2149
+ 1.5
2150
+ 1
2151
+ Energy / J
2152
+ 0.5
2153
+ 0.5
2154
+ 0
2155
+ 02
2156
+ 1.5
2157
+ 1
2158
+ Delay / sIEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
2159
+ 13
2160
+ the effectiveness of the proposed algorithm. We also provide
2161
+ a detailed analysis of the delay, energy consumption, and
2162
+ accuracy for different device numbers, bandwidths and edge
2163
+ computing resources.
2164
+ APPENDIX A
2165
+ PROOF OF THEOREM 1
2166
+ The partial derivative of FPN0 with respect to f md
2167
+ n
2168
+ is,
2169
+ ∂FPN0
2170
+ ∂f md
2171
+ n
2172
+ = −β1
2173
+ ρC(Mn)
2174
+ f md2
2175
+ n
2176
+ + 2β2κρC(Mn)f md
2177
+ n ,
2178
+ (47)
2179
+ By setting
2180
+ ∂FPN0
2181
+ ∂f md
2182
+ n
2183
+ = 0, we have,
2184
+ f md
2185
+ n
2186
+ =
2187
+ 3
2188
+
2189
+ ( β1
2190
+ 2β2κ),
2191
+ (48)
2192
+ Therefore, f md
2193
+ n
2194
+ decreases monotonically in the interval
2195
+ (−∞, 3�
2196
+ ( β1
2197
+ 2β2κ)) and increases monotonically in the interval
2198
+ ( 3�
2199
+ ( β1
2200
+ 2β2κ), +∞). Considering the value range of f md
2201
+ n , the
2202
+ optimal solution can be given by,
2203
+ f md∗
2204
+ n
2205
+ = min{
2206
+ 3
2207
+
2208
+ ( β1
2209
+ 2β2κ), f max
2210
+ n
2211
+ }
2212
+ (49)
2213
+ Then we analyze Mn. The partial derivative of FPN0 with
2214
+ respect to Mn is,
2215
+ ∂FPN0
2216
+ ∂Mn
2217
+ = β1ρmc,0
2218
+ f md
2219
+ n
2220
+ + β2κρmc,0f md2
2221
+ n
2222
+
2223
+ β3ma,0
2224
+ (Mn + ma,1)2 ,
2225
+ (50)
2226
+ By setting
2227
+ ∂FPN0
2228
+ ∂Mn
2229
+ = 0, we have,
2230
+ Mn =
2231
+
2232
+ β3ma,0
2233
+ β1ρmc,0
2234
+ f md
2235
+ n
2236
+ + β2κρmc,0f md2
2237
+ n
2238
+ − ma,1,
2239
+ (51)
2240
+ Considering the value range of Mn, the optimal solution can
2241
+ be given by,
2242
+ M ∗
2243
+ n = min{max{
2244
+
2245
+ β3ma,0
2246
+ β1ρmc,0
2247
+ f md
2248
+ n
2249
+ + β2κρmc,0f md2
2250
+ n
2251
+ − ma,1, M min
2252
+ n
2253
+ }, M max
2254
+ n
2255
+ }
2256
+ (52)
2257
+ which completes the proof.
2258
+ APPENDIX B
2259
+ PROOF OF THEOREM 2
2260
+ According to the KKT conditions, we can obtain the fol-
2261
+ lowing necessary and sufficient conditions,
2262
+ ∂LPN1
2263
+ ∂f e∗
2264
+ n
2265
+ = −β1ρC(M ∗
2266
+ n)
2267
+ f e∗2
2268
+ n
2269
+ + u∗
2270
+ 1 = 0, f e∗
2271
+ n > 0,
2272
+ (53)
2273
+ ∂LPN1
2274
+ ∂t∗n
2275
+ = −β1M ∗
2276
+ nd
2277
+ Rnt∗2
2278
+ n
2279
+ + u∗
2280
+ 0 = 0, t∗
2281
+ n > 0,
2282
+ (54)
2283
+ µ∗
2284
+ 0(
2285
+
2286
+ n∈N ∗
2287
+ t∗
2288
+ n − 1) = 0,
2289
+ (55)
2290
+ µ∗
2291
+ 1(
2292
+
2293
+ n∈N ∗
2294
+ f e∗
2295
+ n − f max) = 0,
2296
+ (56)
2297
+ µ∗
2298
+ 0, µ∗
2299
+ 1 ≥ 0.
2300
+ (57)
2301
+ Because β1ρC(M ∗
2302
+ n)
2303
+ f e∗2
2304
+ n
2305
+ and β1M ∗
2306
+ nd
2307
+ Rnt∗2
2308
+ n
2309
+ are positive, µ∗
2310
+ 0 and µ∗
2311
+ 1 are
2312
+ also positive. We can obtain,
2313
+
2314
+ n∈N
2315
+ f e∗
2316
+ n − f max = 0,
2317
+ (58)
2318
+
2319
+ n∈N
2320
+ t∗
2321
+ n − 1 = 0,
2322
+ (59)
2323
+ f e∗
2324
+ n =
2325
+
2326
+ β1ρC(M ∗n)
2327
+ Rnµ∗
2328
+ 1
2329
+ ,
2330
+ (60)
2331
+ t∗
2332
+ n =
2333
+
2334
+ β1M ∗nd
2335
+ Rnµ∗
2336
+ 0
2337
+ .
2338
+ (61)
2339
+ Combining (58) and (60), we can get the expression of f e∗
2340
+ n
2341
+ corresponding to M ∗
2342
+ n,
2343
+ f e∗
2344
+ n = f max�
2345
+ C(M ∗n)
2346
+
2347
+ i∈N1
2348
+
2349
+ C(M ∗
2350
+ i )
2351
+ .
2352
+ (62)
2353
+ Similarly, combining (59) and (61), we can get the expression
2354
+ of t∗
2355
+ n corresponding to M ∗
2356
+ n,
2357
+ t∗
2358
+ n =
2359
+
2360
+ M ∗
2361
+ n
2362
+ Rn
2363
+
2364
+ i∈N1
2365
+
2366
+ M ∗
2367
+ i
2368
+ Ri
2369
+ ,
2370
+ (63)
2371
+ which completes the proof.
2372
+ REFERENCES
2373
+ [1] H. Ning, H. Wang, Y. Lin, W. Wang, S. Dhelim, F. Farha, J. Ding, and
2374
+ M. Daneshmand, “A survey on metaverse: the state-of-the-art, technolo-
2375
+ gies, applications, and challenges,” arXiv preprint arXiv:2111.09673,
2376
+ 2021.
2377
+ [2] J. Li, L. Deng, Y. Gong, and R. Haeb-Umbach, “An overview of noise-
2378
+ robust automatic speech recognition,” IEEE/ACM Trans. Audio, Speech,
2379
+ Lang. Process., vol. 22, no. 4, pp. 745–777, 2014.
2380
+ [3] D. W. Otter, J. R. Medina, and J. K. Kalita, “A survey of the usages
2381
+ of deep learning for natural language processing,” IEEE Trans. Neural
2382
+ Netw. Learn. Syst., vol. 32, no. 2, pp. 604–624, 2021.
2383
+ [4] A. W. M. Smeulders, D. M. Chu, R. Cucchiara, S. Calderara, A. De-
2384
+ hghan, and M. Shah, “Visual tracking: An experimental survey,” IEEE
2385
+ Trans. Pattern Anal. Mach. Intell., vol. 36, no. 7, pp. 1442–1468, 2014.
2386
+ [5] L. N. Huynh, R. K. Balan, and Y. Lee, “Deepsense: A gpu-based deep
2387
+ convolutional neural network framework on commodity mobile devices,”
2388
+ in Proc. ACM WearSys’16, 2016, p. 25–30.
2389
+ [6] X. Ran, H. Chen, X. Zhu, Z. Liu, and J. Chen, “Deepdecision: A
2390
+ mobile deep learning framework for edge video analytics,” in Proc. IEEE
2391
+ INFOCOM’18, 2018, pp. 1421–1429.
2392
+ [7] Z. Liu, J. Li, Z. Shen, G. Huang, S. Yan, and C. Zhang, “Learning
2393
+ efficient convolutional networks through network slimming,” in Proc.
2394
+ IEEE ICCV’17, Oct 2017, pp. 2736–2744.
2395
+ [8] W. Shi, Y. Hou, S. Zhou, Z. Niu, Y. Zhang, and L. Geng, “Improving
2396
+ device-edge cooperative inference of deep learning via 2-step pruning,”
2397
+ in Proc. IEEE INFOCOM WKSHPS’19, 2019, pp. 1–6.
2398
+ [9] Y. Shi, K. Yang, T. Jiang, J. Zhang, and K. B. Letaief, “Communication-
2399
+ efficient edge AI: Algorithms and systems,” IEEE Commun. Surveys
2400
+ Tuts., vol. 22, no. 4, pp. 2167–2191, 2020.
2401
+ [10] X. Wang, Y. Han, C. Wang, Q. Zhao, X. Chen, and M. Chen, “In-edge
2402
+ AI: Intelligentizing mobile edge computing, caching and communication
2403
+ by federated learning,” IEEE Netw., vol. 33, no. 5, pp. 156–165, 2019.
2404
+ [11] K. B. Letaief, Y. Shi, J. Lu, and J. Lu, “Edge artificial intelligence for
2405
+ 6G: Vision, enabling technologies, and applications,” IEEE J. Sel. Areas
2406
+ Commun, vol. 40, no. 1, pp. 5–36, 2022.
2407
+
2408
+ IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING, VOL. XX, NO. XX, XXX 2022
2409
+ 14
2410
+ [12] S. R. Sabuj, D. K. P. Asiedu, K.-J. Lee, and H.-S. Jo, “Delay opti-
2411
+ mization in mobile edge computing: Cognitive uav-assisted embb and
2412
+ mmtc services,” IEEE Trans. Cogn. Commun. Netw., vol. 8, no. 2, pp.
2413
+ 1019–1033, 2022.
2414
+ [13] P. Wang, B. Di, L. Song, and N. R. Jennings, “Multi-layer compu-
2415
+ tation offloading in distributed heterogeneous mobile edge computing
2416
+ networks,” IEEE Trans. Cogn. Commun. Netw., vol. 8, no. 2, pp. 1301–
2417
+ 1315, 2022.
2418
+ [14] K. Wang, Z. Ding, D. K. C. So, and G. K. Karagiannidis, “Stackelberg
2419
+ game of energy consumption and latency in mec systems with noma,”
2420
+ IEEE Trans. Commun., vol. 69, no. 4, pp. 2191–2206, 2021.
2421
+ [15] M. Qin, N. Cheng, Z. Jing, T. Yang, W. Xu, Q. Yang, and R. R. Rao,
2422
+ “Service-oriented energy-latency tradeoff for iot task partial offloading
2423
+ in mec-enhanced multi-rat networks,” IEEE Internet Things J., vol. 8,
2424
+ no. 3, pp. 1896–1907, 2021.
2425
+ [16] L. Ale, N. Zhang, X. Fang, X. Chen, S. Wu, and L. Li, “Delay-aware
2426
+ and energy-efficient computation offloading in mobile-edge computing
2427
+ using deep reinforcement learning,” IEEE Trans. Cogn. Commun. Netw.,
2428
+ vol. 7, no. 3, pp. 881–892, 2021.
2429
+ [17] K. Yang, Y. Shi, W. Yu, and Z. Ding, “Energy-efficient processing
2430
+ and robust wireless cooperative transmission for edge inference,” IEEE
2431
+ Internet Things J., vol. 7, no. 10, pp. 9456–9470, 2020.
2432
+ [18] S. Hua, Y. Zhou, K. Yang, Y. Shi, and K. Wang, “Reconfigurable
2433
+ intelligent surface for green edge inference,” IEEE Transactions on
2434
+ Green Communications and Networking, vol. 5, no. 2, pp. 964–979,
2435
+ 2021.
2436
+ [19] J. Liu and Q. Zhang, “To improve service reliability for AI-powered
2437
+ time-critical services using imperfect transmission in MEC: An experi-
2438
+ mental study,” IEEE Internet Things J., vol. 7, no. 10, pp. 9357–9371,
2439
+ 2020.
2440
+ [20] W. He, S. Guo, S. Guo, X. Qiu, and F. Qi, “Joint DNN partition
2441
+ deployment and resource allocation for delay-sensitive deep learning
2442
+ inference in IoT,” IEEE Internet Things J., vol. 7, no. 10, pp. 9241–
2443
+ 9254, 2020.
2444
+ [21] Z. Lin, S. Bi, and Y.-J. A. Zhang, “Optimizing AI service placement and
2445
+ resource allocation in mobile edge intelligence systems,” IEEE Trans.
2446
+ Wireless Commun., vol. 20, no. 11, pp. 7257–7271, 2021.
2447
+ [22] E. Li, L. Zeng, Z. Zhou, and X. Chen, “Edge AI: On-demand acceler-
2448
+ ating deep neural network inference via edge computing,” IEEE Trans.
2449
+ Wireless Commun., vol. 19, no. 1, pp. 447–457, 2020.
2450
+ [23] Q. Liu, S. Huang, J. Opadere, and T. Han, “An edge network orchestrator
2451
+ for mobile augmented reality,” in Proc. IEEE INFOCOM’18, 2018, pp.
2452
+ 756–764.
2453
+ [24] Y. He, J. Ren, G. Yu, and Y. Cai, “Optimizing the learning performance
2454
+ in mobile augmented reality systems with CNN,” IEEE Trans. Wireless
2455
+ Commun., vol. 19, no. 8, pp. 5333–5344, 2020.
2456
+ [25] Y. Zhao, Z. Yang, X. He, X. Cai, X. Miao, and Q. Ma, “Trine:
2457
+ Cloud-edge-device cooperated real-time video analysis for household
2458
+ applications,” IEEE Trans. Mobile Comput., pp. 1–1, 2022.
2459
+ [26] K. Hara, H. Kataoka, and Y. Satoh, “Can spatiotemporal 3D CNNs re-
2460
+ trace the history of 2D CNNs and ImageNet?” in Proc. IEEE CVPR’18,
2461
+ 2018, pp. 6546–6555.
2462
+ [27] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
2463
+ large-scale image recognition,” in Proc. ICLR’15, May 2015, pp. 1–14.
2464
+ [28] M. Hollemans, “How fast is my model?” https://machinethink.net/blog/
2465
+ how-fast-is-my-model/, accessed Dec. 30, 2021.
2466
+ [29] C. Wang, S. Zhang, Y. Chen, Z. Qian, J. Wu, and M. Xiao, “Joint
2467
+ configuration adaptation and bandwidth allocation for edge-based real-
2468
+ time video analytics,” in Proc. IEEE INFOCOM’20, 2020, pp. 257–266.
2469
+ [30] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex optimization.
2470
+ Cambridge university press, 2004.
2471
+ [31] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex
2472
+ programming, version 2.1,” http://cvxr.com/cvx, Mar. 2014.
2473
+ [32] S. Bi and Y. J. Zhang, “Computation rate maximization for wireless
2474
+ powered mobile-edge computing with binary computation offloading,”
2475
+ IEEE Trans. Wireless Commun., vol. 17, no. 6, pp. 4177–4190, 2018.
2476
+ [33] J. Li, X. Li, Y. Bi, and J. Ma, “Energy-efficient joint resource allocation
2477
+ with reconfigurable intelligent surfaces in symbiotic radio networks,”
2478
+ IEEE Trans. Cogn. Commun. Netw., pp. 1–1, 2022.
2479
+ [34] V. Sovrasov, “Flops counter for convolutional networks in pytorch
2480
+ framework,” https://github.com/sovrasov/flops-counter.pytorch, accessed
2481
+ Dec. 30, 2021.
2482
+ [35] J. Materzynska, G. Berger, I. Bax, and R. Memisevic, “The jester dataset:
2483
+ A large-scale video dataset of human gestures,” in Proc. IEEE ICCV
2484
+ Workshop’19, 2019, pp. 2874–2882.
2485
+ [36] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijaya-
2486
+ narasimhan, F. Viola, T. Green, T. Back, P. Natsev, M. Suleyman, and
2487
+ A. Zisserman, “The kinetics human action video dataset,” arXiv preprint
2488
+ arXiv:1705.06950, 2017.
2489
+ [37] X. Chen, Y. Cai, L. Li, M. Zhao, B. Champagne, and L. Hanzo,
2490
+ “Energy-efficient resource allocation for latency-sensitive mobile edge
2491
+ computing,” IEEE Trans. Veh. Technol., vol. 69, no. 2, pp. 2246–2262,
2492
+ 2020.
2493
+
99AzT4oBgHgl3EQfFfoM/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
A9AzT4oBgHgl3EQfF_vM/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb317daabfd16ef850d23bb4048ee5453c7fa5bec7a7ca8f8df3e2a746e02f61
3
+ size 4784173
A9E1T4oBgHgl3EQfVgQy/content/2301.03102v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:93bea934dd212ce346a06f6a259b74a46b0094bb7ad45cbdbd8e0cacc813019c
3
+ size 27039399
C9AyT4oBgHgl3EQf4fqw/content/2301.00788v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:acd7572d9c676f35b5a8fedc3dc66740355cf8789eecd6b3e9be8390bf2dd107
3
+ size 6833683
D9E4T4oBgHgl3EQfGQye/content/2301.04893v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a596e6fe279059f69e9d7b8ca6198a155be1301bc0b1415d56b57d90cf1aaaa4
3
+ size 18088715
ENE1T4oBgHgl3EQfEQP0/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7cad128b5efffc6356db034a1f48432c6a091d4b45622bbb0e26173e2dcbbecb
3
+ size 5767213
FNE0T4oBgHgl3EQfzAKR/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:228919a9013b88911b75d4025b7d7a3b22d9d08a034e4e083a0e029da4d83c26
3
+ size 131927
G9AyT4oBgHgl3EQfrfnZ/content/tmp_files/2301.00563v1.pdf.txt ADDED
@@ -0,0 +1,1961 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.00563v1 [cond-mat.stat-mech] 2 Jan 2023
2
+ Sample-to-sample fluctuations of transport coefficients in the totally asymmetric
3
+ simple exclusion process with quenched disorders
4
+ Issei Sakai1 and Takuma Akimoto1, ∗
5
+ 1Department of Physics, Tokyo University of Science, Noda, Chiba 278-8510, Japan
6
+ (Dated: January 3, 2023)
7
+ We consider the totally asymmetric simple exclusion processes on quenched random energy land-
8
+ scapes. We show that the current and the diffusion coefficient differ from those for homogeneous
9
+ environments. Using the mean-field approximation, we analytically obtain the site density when the
10
+ particle density is low or high. As a result, the current and the diffusion coefficient are described
11
+ by the dilute limit of particles or holes, respectively. However, in the intermediate regime, due to
12
+ the many-body effect, the current and the diffusion coefficient differ from those for single-particle
13
+ dynamics.
14
+ The current is almost constant and becomes the maximal value in the intermediate
15
+ regime. Moreover, the diffusion coefficient decreases with the particle density in the intermediate
16
+ regime. We obtain analytical expressions for the maximal current and the diffusion coefficient based
17
+ on the renewal theory. The deepest energy depth plays a central role in determining the maximal
18
+ current and the diffusion coefficient. As a result, the maximal current and the diffusion coefficient
19
+ depend crucially on the disorder, i.e., non-self-averaging. Based on the extreme value theory, we find
20
+ that sample-to-sample fluctuations of the maximal current and diffusion coefficient are characterized
21
+ by the Weibull distribution. We show that the disorder averages of the maximal current and the
22
+ diffusion coefficient converge to zero as the system size is increased and quantify the degree of the
23
+ non-self-averaging effect for the maximal current and the diffusion coefficient.
24
+ I.
25
+ INTRODUCTION
26
+ The one-dimensional asymmetric simple exclusion pro-
27
+ cess (ASEP) is a pedagogical model for non-equilibrium
28
+ systems [1].
29
+ In particular, it describes various non-
30
+ equilibrium phenomena such as traffic flow [2] and pro-
31
+ tein synthesis by ribosomes [3–5]. The ASEP is a stochas-
32
+ tic process where particles with hard-core interactions
33
+ diffuse on a one-dimensional lattice. The ASEP can be
34
+ mapped to a model of interface growth in the Kardar-
35
+ Parisi-Zhang (KPZ) universality class [6].
36
+ Hopping to
37
+ the right site in the ASEP corresponds to an increase
38
+ in the interface.
39
+ The distribution of interface height
40
+ was solved [7–9].
41
+ Using the weak asymmetric limit of
42
+ the ASEP, the KPZ equation was rigorously solved an-
43
+ alytically [10, 11]. Moreover, the large deviation func-
44
+ tion of the time-averaged current was obtained [12, 13].
45
+ The ASEP has been extended in various ways such as
46
+ Brownian ASEP [14], non-Poissonian hopping rates [15],
47
+ and disordered hopping rates [16–23].
48
+ When particles
49
+ only hop to uni-direction, it is called the totally ASEP
50
+ (TASEP). For TASEPs, it is well known that the current-
51
+ density relation is given by [1]
52
+ J = 1
53
+ τ ρ(1 − ρ),
54
+ (1)
55
+ where J is the particle current, ρ is particle density, and
56
+ τ is the inverse of the jump rate, i.e., the mean waiting
57
+ time. Moreover, in Refs. [24], the variance of the tagged
58
+ particle displacement, δxt, in time t is derived as a func-
59
+ ∗ takuma@rs.tus.ac.jp
60
+ tion of ρ:
61
+ ⟨δx2
62
+ t ⟩ − ⟨δxt⟩2
63
+ t
64
+
65
+ √π
66
+
67
+ (1 − ρ)3/2
68
+ (Lρ)1/2
69
+ (2)
70
+ for L → ∞ and t → ∞, where ⟨·⟩ is the ensemble average
71
+ and L is the system size.
72
+ Effects of disorder in the ASEP have been investi-
73
+ gated for decades [16–23].
74
+ Due to the disorder in the
75
+ ASEP under the periodic boundary condition, a current-
76
+ density relation deviates from that in the ASEP with a
77
+ homogeneous jump rate, i.e., Eq. (1). More precisely, it
78
+ becomes flat and the current is maximized on the flat
79
+ regime [16, 18–23].
80
+ Moreover, in the flat regime, the
81
+ low- and high-density phases coexist. In the ASEP on
82
+ networks, the flat regime also exists [25–27]. Under the
83
+ open boundary condition, the first-order phase transition
84
+ point between the low- and high-density phases depends
85
+ on the disorder [17].
86
+ Random walks in heterogeneous environments show
87
+ anomalous diffusion. The heterogeneous environment is
88
+ characterized by a random energy landscape. There are
89
+ two types of random energy landscapes. One is an an-
90
+ nealed energy landscape, where the landscape randomly
91
+ changes with time.
92
+ The continuous-time random walk
93
+ is a diffusion model on the annealed energy landscape,
94
+ and its mean-squared displacement shows anomalous dif-
95
+ fusion when the mean waiting time diverges [28]. The
96
+ other is a quenched energy landscape, where the land-
97
+ scape is configured randomly and does not change with
98
+ time.
99
+ The quenched trap model (QTM) is a diffusion
100
+ model on the quenched energy landscape [29]. The mean-
101
+ squared displacement of the QTM on an infinite system
102
+ shows anomalous diffusion when the mean waiting time
103
+ diverges [29]. In the QTM on a finite system, the dif-
104
+ fusion coefficient exhibits sample-to-sample fluctuations
105
+
106
+ 2
107
+ [30–32].
108
+ The diffusivity of interacting many-body sys-
109
+ tems on the annealed energy landscape has been inves-
110
+ tigated [33, 34]. However, the diffusivity of interacting
111
+ many-body systems on the quenched energy landscape
112
+ has never been investigated. Such a heterogeneous envi-
113
+ ronment is realized experimentally. In protein synthesis
114
+ by ribosomes, the codon decoding times become hetero-
115
+ geneous due to the heterogeneity of transfer RNA concen-
116
+ tration [5]. In other words, the distribution of the waiting
117
+ time depends on the site, i.e., ribosomes diffuse on the
118
+ quenched random environment. There are other diffusion
119
+ phenomena in such heterogeneous environments, such as
120
+ train delays, proteins on DNA [35, 36], and water trans-
121
+ portation in aquaporin [37].
122
+ In this paper, we investigate sample-to-sample fluctu-
123
+ ations of the diffusivity for the TASEP on a quenched
124
+ random energy landscape.
125
+ In our previous study, we
126
+ show sample-to-sample fluctuations of the current [38].
127
+ When an observable does not depend on the disorder re-
128
+ alization, it is called self-averaging [29].
129
+ In the QTM,
130
+ it is known that the diffusion coefficient [30–32], the
131
+ mobility [32], and the mean first passage time [39] are
132
+ non-self-averaging. Is such a non-self-averaging behavior
133
+ still observed when the N-body effect is introduced in
134
+ the quenched random energy landscape? This is a non-
135
+ trivial question in diffusion in a heterogeneous environ-
136
+ ment. In particular, it is non-trivial that the TASEP with
137
+ disordered waiting-time distributions exhibits sample-to-
138
+ sample fluctuations for the current and the diffusion co-
139
+ efficient. Therefore, it is important to provide an exact
140
+ result for the current and the diffusion coefficient in het-
141
+ erogeneous quenched environments.
142
+ Our paper is organized as follows. In Sec. II, we formu-
143
+ late the TASEP on a quenched random energy landscape
144
+ and define averaging procedures. In Sec. III, we show the
145
+ numerical results of the current-density relation and the
146
+ density profile. In Sec. IV, we present derivations of the
147
+ density profile. In Sec. V, we present derivations of the
148
+ current and the diffusion coefficient. In Sec. VI, we dis-
149
+ cuss the self-averaging properties of the current and the
150
+ diffusion coefficient. In Sec. VII, we conclude this paper.
151
+ In Appendix A, we derive the passage time distribution.
152
+ In Appendix B, we derive the Fr´echet distribution.
153
+ II.
154
+ MODEL
155
+ We consider the TASEP on a quenched random en-
156
+ ergy landscape on a one-dimensional lattice.
157
+ It com-
158
+ prises N particles on the lattice of L sites with periodic
159
+ boundary conditions.
160
+ Each site can hold at most one
161
+ particle. Quenched disorder means that when realizing
162
+ the random energy landscape, it does not change with
163
+ time. At each lattice point, the depth E > 0 of the en-
164
+ ergy trap is randomly assigned. The depths are indepen-
165
+ dent identically distributed (IID) random variables with
166
+ an exponential distribution, φ(E) = T −1
167
+ g
168
+ exp (−E/Tg),
169
+ where Tg is called the glass temperature.
170
+ A particle
171
+ can escape from a trap. Escape times from a trap are
172
+ IID random variables following an exponential distribu-
173
+ tion and follow the Arrhenius law, i.e., the mean escape
174
+ time of the kth site is given by τk = τc exp (Ek/T ),
175
+ where Ek is the depth of the energy at site k, T the
176
+ temperature, and τc a typical time.
177
+ The probability
178
+ of the escape time τ that is smaller than x is given
179
+ by Pr(τ ≤ x) ∼= Pr(E ≤ T ln(x/τc)).
180
+ It follows that
181
+ the probability density function (PDF) ψα(τ) of waiting
182
+ times follows a power-law distribution:
183
+ � ∞
184
+ τ
185
+ dτ ′ψα(τ ′) =
186
+ � τ
187
+ τc
188
+ �−α
189
+ (τ ≥ τc)
190
+ (3)
191
+ with α ≡ T/Tg [30].
192
+ The dynamics of the particle are described by the
193
+ Markovian one in the sense that the waiting time is
194
+ memory-less. In particular, the waiting times at site k
195
+ are assigned IID random variables following an exponen-
196
+ tial distribution, ψk(ti) = τ −1
197
+ k
198
+ exp (−ti/τk).
199
+ After the
200
+ waiting time elapses, the particle attempts to hop the
201
+ neighboring site on its right. The hop is accepted only
202
+ if the site is empty. When the attempt is a success or
203
+ failure, the particle is assigned a new waiting time from
204
+ ψk+1(ti) or ψk(ti), respectively.
205
+ Here, we consider three averaging procedures, i.e., en-
206
+ semble average, disorder average, and time average. The
207
+ ensemble average of observable O(t) is an average with
208
+ respect to a stationary ensemble for a single disorder re-
209
+ alization denoted by ⟨O(t)⟩. The disorder average of ob-
210
+ servable O(t) is an average with respect to different dis-
211
+ order realizations denoted by ⟨O(t)⟩dis. The time average
212
+ of observable O(t) is defined by
213
+ ¯O(T ) = 1
214
+ T
215
+ � T
216
+ 0
217
+ O(t)dt.
218
+ (4)
219
+ Furthermore, we consider a stationary initial condition.
220
+ For the ASEP on a finite system, the variance of the
221
+ displacement of the tagged particle depends on whether
222
+ the initial conditions are identical or not, especially for a
223
+ short time [40]. However, the asymptotic behavior does
224
+ not depend on the initial condition. In this paper, we are
225
+ interested in the asymptotic behavior of the current and
226
+ the diffusivity. Therefore, the initial conditions in this
227
+ paper are not fixed. In numerical simulations, particles
228
+ start from the stationary ensemble of configurations. The
229
+ stationary ensemble is given by the configuration after
230
+ particles arrange randomly and diffuse for a long time.
231
+ III.
232
+ NUMERICAL RESULTS OF
233
+ CURRENT-DENSITY RELATION AND
234
+ DENSITY PROFILE
235
+ We numerically show that the current-density relation
236
+ for a disordered TASEP (DTASEP) deviates from that
237
+ for a TASEP with a homogeneous jump rate, i.e., the
238
+ homogeneous TASEP. Figure 1 shows the steady-state
239
+
240
+ 3
241
+ current J against particle density ρ = N/L, i.e., the
242
+ current-density relation, for a DTASEP. For low and high
243
+ densities, the current-density relation is the same as that
244
+ of the homogeneous TASEP (see Fig. 1). However, there
245
+ is a distinct difference between them in the intermedi-
246
+ ate regime. In particular, the current for the DTASEP
247
+ becomes almost flat and smaller than that for the ho-
248
+ mogeneous TASEP in the intermediate regime. On the
249
+ other hand, there is no flat regime for the homogeneous
250
+ TASEP. The flat regime in the DTASEP is observed in
251
+ other disordered systems [16, 18–20, 23].
252
+ Thus, it is
253
+ a manifestation of the existence of a disorder. In this
254
+ regime, the current is independent of the particle density
255
+ and maximized. In the following, we classify the density
256
+ into three regimes: the low density (LD) (0 < ρ ≤ ρ∗),
257
+ the maximal current (MC) (ρ∗ < ρ < 1 − ρ∗), and the
258
+ high density (HD) (1 − ρ∗ ≤ ρ < 1) regimes (Fig. 1).
259
+ We explicitly derive the transition density ρ∗ later (see
260
+ Eq. (13)).
261
+ Here, we numerically show the density profiles.
262
+ For
263
+ the LD and HD regimes, the system is homogeneous
264
+ on a macroscopic scale (Figs. 2(a) and (b)).
265
+ For the
266
+ MC regime, there is a macroscopic density segregation
267
+ (Figs. 2(c) and (d)).
268
+ The segregation is classified into
269
+ high- and low-density phases by the deepest trap. Com-
270
+ paring Figs. 2(c) and (d), we observe that the high-
271
+ density regime becomes large when the particle density
272
+ is increased. This result is qualitatively similar to that
273
+ in a system with one defect bond, studied in Ref. [41].
274
+ !"
275
+ #"
276
+ $%
277
+ 0
278
+ 0.01
279
+ 0.02
280
+ 0.03
281
+ 0
282
+ 0.2
283
+ 0.4
284
+ 0.6
285
+ 0.8
286
+ 1
287
+ J
288
+ ρ
289
+ FIG. 1. Current-density relations for homogeneous and dis-
290
+ ordered TASEPs. The circles are obtained by the numerical
291
+ simulation of dynamics of the DTASEP (L = 5000, α = 2.5,
292
+ and τc = 1).
293
+ The solid line represents the current-density
294
+ relation, Eq. (1), for the homogeneous TASEP with τ being
295
+ set to equal to the sample average of the waiting times of the
296
+ DTASEP. ρ∗ is given by Eq. (13).
297
+ IV.
298
+ DERIVATION OF THE DENSITY PROFILE
299
+ Here, we derive the density profile by the mean-field
300
+ approximation. This derivation is almost the same as our
301
+ previous study [38]. Let Jk be the mean current across
302
+ the bond between site k and k+1. In the DTASEP, a hop
303
+ occurs with a rate 1/τk whenever site k is occupied, and
304
+ site k + 1 is not. Thus, the mean current is represented
305
+ by
306
+ Jk =
307
+ � 1
308
+ τk
309
+ nk(1 − nk+1)
310
+
311
+ ,
312
+ (5)
313
+ where nk denotes the number of a particle, which is 1
314
+ if the site k is occupied and 0 otherwise.
315
+ If the sys-
316
+ tem is in a steady state, the ensemble average is equal
317
+ to the time average in the long-time limit, i.e., the sys-
318
+ tem is ergodic. The ensemble average in Eq. (5) coin-
319
+ cides with the long-time average if the system is ergodic.
320
+ The periodic boundary condition implies nL+1 = n1 and
321
+ τL+1 = τ1. The probability of finding a particle at site k
322
+ is given by ρk = ⟨nk⟩. In the mean-field approximation,
323
+ one can ignore correlations between nk and nk+1, which
324
+ means
325
+ ⟨nknk+1⟩ = ⟨nk⟩ ⟨nk+1⟩ .
326
+ (6)
327
+ In
328
+ the
329
+ steady
330
+ state,
331
+ the
332
+ site
333
+ densities
334
+ are
335
+ time-
336
+ independent. Moreover, from the continuity of the cur-
337
+ rent, the current is independent of k, i.e., Jk = J for all
338
+ k. Therefore, we have the current-density relation:
339
+ J = 1
340
+ τk
341
+ ρk(1 − ρk+1).
342
+ (7)
343
+ We note that the right-hand side of Eq. (7) is independent
344
+ of k.
345
+ We derive a simpler form of the site density by approx-
346
+ imating Eq. (7) for the LD and HD regimes. For the LD
347
+ regime, we can assume ρkρk+1 ≪ 1 because the particle
348
+ density is small. Ignoring ρkρk+1 in Eq. (7), we obtain
349
+ J ∼= 1
350
+ τk
351
+ ρk.
352
+ (8)
353
+ Using the conservation of particles, �
354
+ i ρi = N, the site
355
+ density has the form
356
+ ρk ∼= τk
357
+ µ ρ,
358
+ (9)
359
+ for the LD regime, where µ is the sample average of the
360
+ waiting times, µ = �
361
+ i τi/L.
362
+ This result is the same
363
+ as the steady-state density for the QTM [30]. For the
364
+ HD regime, the particle density is high. Using the hole
365
+ density, σk = 1 − ρk, instead of ρk, we can derive the site
366
+ density in the same way as in the LD regime. The result
367
+ becomes
368
+ ρk = 1 − σk ∼= 1 − τk−1
369
+ µ (1 − ρ).
370
+ (10)
371
+
372
+ 4
373
+ (a)
374
+ 0
375
+ 0.05
376
+ 0.1
377
+ 0.15
378
+ 0.2
379
+ 0.25
380
+ 0
381
+ 1000
382
+ 2000
383
+ 3000
384
+ 4000
385
+ 5000
386
+ ρk
387
+ k
388
+ (b)
389
+ 0.75
390
+ 0.8
391
+ 0.85
392
+ 0.9
393
+ 0.95
394
+ 1
395
+ 0
396
+ 1000
397
+ 2000
398
+ 3000
399
+ 4000
400
+ 5000
401
+ ρk
402
+ k
403
+ (c)
404
+ 0
405
+ 0.2
406
+ 0.4
407
+ 0.6
408
+ 0.8
409
+ 1
410
+ 0
411
+ 1000
412
+ 2000
413
+ 3000
414
+ 4000
415
+ 5000
416
+ ρk
417
+ k
418
+ (d)
419
+ 0
420
+ 0.2
421
+ 0.4
422
+ 0.6
423
+ 0.8
424
+ 1
425
+ 0
426
+ 1000
427
+ 2000
428
+ 3000
429
+ 4000
430
+ 5000
431
+ ρk
432
+ k
433
+ FIG. 2. Density profiles: (a) ρ = 0.01, (b) ρ = 0.99, (c) ρ = 0.5, and (d) ρ = 0.8 (L = 5000, α = 2.5, and τc = 1). The squares
434
+ are the results of the numerical simulation of the dynamics of the DTASEP. Triangles are Eqs. (9) and (10) for (a) and (b),
435
+ respectively.
436
+ Figures 2(a) and 2(b) show the density profiles for LD
437
+ and HD regimes, respectively. The densities are well de-
438
+ scribed by the set of site densities {ρk}. Therefore, Eqs
439
+ (9) and (10) are good approximated forms of the site
440
+ densities.
441
+ The results for the LD and HD regimes re-
442
+ produce the current-density relation for a homogeneous
443
+ TASEP. In other words, the system is homogeneous on a
444
+ macroscopic scale.
445
+ Next, we approximately obtain a density ρ∗ which is
446
+ the boundary density between LD and MC regimes in
447
+ the current-density relation (see Fig. 1). By Eq. (7), the
448
+ current between sites m and m + 1 is given by ρm(1 −
449
+ ρm+1)/τm.
450
+ The steady-state current at the boundary
451
+ density between the LD and MC regimes can be described
452
+ by Eq. (1). At the boundary density, the current between
453
+ sites m and m + 1 is equal to the steady-state current:
454
+ 1
455
+ τm
456
+ ρm(1 − ρm+1) ∼= 1
457
+ µρ∗(1 − ρ∗).
458
+ (11)
459
+ We find numerically find that the site with the maximal
460
+ mean waiting time is always the boundary the HD and
461
+ the LD phases. When the mean waiting time is maxi-
462
+ mized at site m, sites m and m + 1 exist in high- and
463
+ low-density phases, respectively. The site densities are
464
+ given by Eq. (10) and the hole density is ρ∗ in the high-
465
+ density phase. On the other hand, the site densities are
466
+ given by Eq. (9) and the particle density is ρ∗ in the
467
+ low-density phase. Therefore, we substitute Eq. (10) and
468
+ Eq. (9) into ρm and ρm+1, respectively,
469
+ 1
470
+ τm
471
+
472
+ 1 − τm−1
473
+ µ
474
+ ρ∗
475
+ � �
476
+ 1 − τm+1
477
+ µ
478
+ ρ∗
479
+
480
+ ∼= 1
481
+ µρ∗(1 − ρ∗). (12)
482
+ Solving this equation for ρ∗, we have
483
+ ρ∗ ∼µ(τm−1 + τm + τm+1)
484
+ 2(τm−1τm+1 + τmµ)
485
+ − µ
486
+
487
+ (τm−1 + τm + τm+1)2 − 4(τm−1τm+1 + τmµ)
488
+ 2(τm−1τm+1 + τmµ)
489
+ .
490
+ (13)
491
+ This formula depends crucially on the disorder realiza-
492
+ tion. In the following, we assume that the mean waiting
493
+ time is maximized at site m. For L → ∞, τm is much
494
+ longer than τm−1 and τm+1. Therefore, Eq. (13) can be
495
+ approximated as
496
+ ρ∗ ∼ 1
497
+ 2 −
498
+ 1
499
+ 2τm
500
+
501
+ τ 2m − 4τmµ ∼ µ
502
+ τm
503
+ .
504
+ (14)
505
+ By the extreme value theory [42], the scaling of τm follows
506
+ τm = O(L1/α)
507
+ (15)
508
+
509
+ 5
510
+ for L → ∞. For α > 1, the first moment of the waiting
511
+ times exists; i.e., µ → ⟨τ⟩ ≡ � ∞
512
+ 0
513
+ τψα(τ)dτ (L → ∞).
514
+ Hence, the scaling of ρ∗ becomes
515
+ ρ∗ ∝ L−1/α.
516
+ (16)
517
+ For α ≤ 1, the first moment of the waiting times diverges.
518
+ The scaling of the sum of τi follows
519
+ L
520
+
521
+ i=1
522
+ τi = O(L1/α)
523
+ (17)
524
+ for L → ∞. It follows that the scaling of ρ∗ becomes
525
+ ρ∗ ∼ L−1
526
+
527
+ i τi
528
+ τm
529
+ ∝ L−1.
530
+ (18)
531
+ Therefore, ρ∗ → 0 for L → ∞.
532
+ V.
533
+ DERIVATION OF CURRENT AND
534
+ DIFFUSIVITY
535
+ A.
536
+ LD and HD regimes
537
+ Here, we derive the current in the LD and HD regimes.
538
+ For single-particle dynamics on the quenched random
539
+ energy landscape, i.e., the QTM, the mean number of
540
+ events that a particle passes a site until time t is given
541
+ by [32]
542
+ ⟨Qt⟩
543
+ t
544
+ ∼ 1
545
+ Lµ (t → ∞),
546
+ (19)
547
+ where Qt is the number of events that a particle passes
548
+ a site until time t.
549
+ For the DTASEP in the LD and
550
+ HD regimes, the current depends on the particle density,
551
+ which is identical for the homogeneous TASEP (Eq. (1)).
552
+ Hence, the current in the LD and HD regimes is given by
553
+ J ∼ aρ(1 − ρ)
554
+ (20)
555
+ for L → ∞.
556
+ When ρ = 1/L, the current is equal to
557
+ Eq. (19) for L → ∞, i.e., the constant a is given by
558
+ a = 1/µ. Therefore, we have the current in the LD and
559
+ HD regimes:
560
+ J ∼ 1
561
+ µρ(1 − ρ)
562
+ (21)
563
+ for L → ∞.
564
+ Next, we derive the diffusion coefficient in the LD and
565
+ HD regimes. δxt denotes the displacement of the tagged
566
+ particle until time t. For the QTM, the variance of the
567
+ displacement is given by [32]
568
+ lim
569
+ t→∞
570
+ ⟨δx2
571
+ t ⟩ − ⟨δxt⟩2
572
+ t
573
+ ∼ σ2
574
+ µ3
575
+ (22)
576
+ for L → ∞, where σ2 is the sample mean of the squared
577
+ waiting times, σ2 = �
578
+ i τ 2
579
+ i /L. For the DTASEP in the
580
+ LD and HD regimes, the variance of the displacement
581
+ depends on the particle density, which is identical for
582
+ the homogeneous TASEP (Eq. (2)). Hence, the diffusion
583
+ coefficient, D ≡ limt→∞(⟨δx2
584
+ t ⟩ − ⟨δxt⟩2)/t, is given by
585
+ D ∼ b
586
+ √π
587
+ 2
588
+ (1 − ρ)3/2
589
+ ρ1/2
590
+ L−1/2
591
+ (23)
592
+ for L → ∞. When ρ = 1/L, the diffusion coefficient is
593
+ equal to Eq. (22) for L → ∞, i.e., the constant b is given
594
+ by b = 2σ2/µ3√π. The diffusion coefficient in the LD
595
+ and HD regimes is given by
596
+ D ∼ σ2
597
+ µ3
598
+ (1 − ρ)3/2
599
+ ρ1/2
600
+ L−1/2
601
+ (24)
602
+ for L → ∞.
603
+ B.
604
+ MC regime
605
+ Here, we derive the maximal current and the diffusion
606
+ coefficient in the MC regime by the renewal theory. We
607
+ define the passage time as a time interval between consec-
608
+ utive events that particles pass a site. We note that the
609
+ passage time differs from the first passage time because
610
+ the particles which pass a site are different. When the
611
+ target site is m, the mean and the variance of the passage
612
+ time Tm are obtained in Ref. [38] (see also Appendix A):
613
+ ⟨Tm⟩ = τm + τm−1
614
+ ρm−1
615
+ +
616
+ ρm−1
617
+ τm−1
618
+ ρm−1
619
+ τm−1 + 1−ρm+2
620
+ τm+1
621
+ τm+1
622
+ 1 − ρm+2
623
+ ,
624
+ (25)
625
+ ⟨T 2
626
+ m⟩ − ⟨Tm⟩2 =τ 2
627
+ m +
628
+ � τm−1
629
+ ρm−1
630
+ �2
631
+ +
632
+
633
+ τm+1
634
+ 1 − ρm+2
635
+ �2
636
+
637
+ 3
638
+
639
+ ρm−1
640
+ τm−1 + 1−ρm+2
641
+ τm+1
642
+ �2 .
643
+ (26)
644
+ We consider the number of events Qt that particles
645
+ pass site m until time t to obtain the maximal current and
646
+ the diffusion coefficient. For the LD and HD regimes, the
647
+ density profile is homogeneous on a macroscopic scale.
648
+ However, local densities around the target site are fluc-
649
+ tuating, i.e., dense or dilute, which affects the passage
650
+ time. Therefore, the passage times are not IID random
651
+ variables for the LD and HD regimes. For the MC regime,
652
+ macroscopic density segregation exists. When the target
653
+ locates site m, particles are constantly dense on the left of
654
+ the target and dilute on the right. This segregation does
655
+ not vary with time. Therefore, the passage times are con-
656
+ sidered to be IID random variables for MC regime and
657
+ the process of Qt can be described by a renewal process
658
+ [43].
659
+ By renewal theory [43], the mean number of re-
660
+ newals becomes ⟨Qt⟩ ∼ t/ ⟨Tm⟩ for t → ∞. The current
661
+ is represented through the mean number of the passing
662
+ events until time t: J = limt→∞ ⟨Qt⟩ /t. Thus, we have
663
+ Jmax ∼
664
+ 1
665
+ ⟨Tm⟩
666
+ (27)
667
+
668
+ 6
669
+ (a)
670
+ 0
671
+ 0.1
672
+ 0.2
673
+ 0.3
674
+ 0.4
675
+ 0.5
676
+ 0
677
+ 0.2
678
+ 0.4
679
+ 0.6
680
+ 0.8
681
+ 1
682
+ J
683
+ ρ
684
+ ×10−5
685
+ (b)
686
+ 0
687
+ 0.001
688
+ 0.002
689
+ 0.003
690
+ 0.004
691
+ 0
692
+ 0.2
693
+ 0.4
694
+ 0.6
695
+ 0.8
696
+ 1
697
+ J
698
+ ρ
699
+ (c)
700
+ 0
701
+ 0.01
702
+ 0.02
703
+ 0.03
704
+ 0
705
+ 0.2
706
+ 0.4
707
+ 0.6
708
+ 0.8
709
+ 1
710
+ J
711
+ ρ
712
+ FIG. 3. Current-density relation for different α, i.e., (a) α = 0.5, (b) α = 1.5, and (c) α = 2.5, where the fixed quenched
713
+ disorders. The circles are obtained by the numerical simulation of the dynamics of the DTASEP (L = 1000 for (a) and 5000
714
+ for other cases). The dashed and the solid lines represent Eqs. (21) and (27), respectively.
715
+ (a)
716
+ 10−6
717
+ 10−5
718
+ 10−4
719
+ 10−3
720
+ 10−2
721
+ 10−1
722
+ 0
723
+ 0.2
724
+ 0.4
725
+ 0.6
726
+ 0.8
727
+ 1
728
+ D
729
+ ρ
730
+ (b)
731
+ 10−6
732
+ 10−5
733
+ 10−4
734
+ 10−3
735
+ 10−2
736
+ 10−1
737
+ 100
738
+ 101
739
+ 0
740
+ 0.2
741
+ 0.4
742
+ 0.6
743
+ 0.8
744
+ 1
745
+ D
746
+ ρ
747
+ (c)
748
+ 10−6
749
+ 10−5
750
+ 10−4
751
+ 10−3
752
+ 10−2
753
+ 10−1
754
+ 100
755
+ 101
756
+ 0
757
+ 0.2
758
+ 0.4
759
+ 0.6
760
+ 0.8
761
+ 1
762
+ D
763
+ ρ
764
+ FIG. 4. Diffusion coefficient-density relation for different α, i.e., (a) α = 0.5, (b) α = 1.5, and (c) α = 2.5, where the fixed
765
+ quenched disorders. The circles are obtained by the numerical simulation of the dynamics of the DTASEP (L = 100 for (a),
766
+ 500 for (b), and 1000 for (c)). The dashed and the solid lines represent Eqs. (24) and (31), respectively.
767
+ for L → ∞. The current depends on the disorder real-
768
+ ization. Figure 3 shows a good agreement between nu-
769
+ merical simulations and the theory.
770
+ Using the number of the passing events, we can de-
771
+ rive the mean displacement and the variance of the dis-
772
+ placement of a tagged particle. While the tagged particle
773
+ starting from site m+1 returns to site m+1, all particles
774
+ pass between site m and site m + 1. Therefore, in the
775
+ large-t limit, the displacement, δxt, is represented by
776
+ δxt ∼ LQt
777
+ N
778
+ = Qt
779
+ ρ .
780
+ (28)
781
+ By renewal theory [43], the mean displacement and the
782
+ variance of the displacement are represented by
783
+ ⟨δxt⟩ ∼ ⟨Qt⟩
784
+ ρ
785
+
786
+ t
787
+ ρ⟨Tm⟩,
788
+ (29)
789
+ ⟨δx2
790
+ t ⟩ − ⟨δxt⟩2 ∼ 1
791
+ ρ2 (⟨Q2
792
+ t⟩ − ⟨Qt⟩2)
793
+ ∼ 1
794
+ ρ2
795
+ ⟨T 2
796
+ m⟩ − ⟨Tm⟩2
797
+ ⟨Tm⟩3
798
+ t
799
+ (30)
800
+ for t → ∞. Therefore, the diffusion coefficient for the
801
+ MC regimes is given by
802
+ D ∼ 1
803
+ ρ2
804
+ ⟨T 2
805
+ m⟩ − ⟨Tm⟩2
806
+ ⟨Tm⟩3
807
+ (31)
808
+ for L → ∞. Figure 4 shows a good agreement between
809
+ numerical simulations and the theory.
810
+ VI.
811
+ SAMPLE-TO-SAMPLE FLUCTUATIONS OF
812
+ CURRENT AND DIFFUSIVITY
813
+ A.
814
+ Current
815
+ Here, we consider sample-to-sample fluctuations of the
816
+ current. To quantify the self-averaging (SA) property of
817
+ the current, we consider the SA parameter defined as [30]
818
+ SA(L; J) ≡ ⟨J(L)2⟩dis − ⟨J(L)⟩2
819
+ dis
820
+ ⟨J(L)⟩2
821
+ dis
822
+ ,
823
+ (32)
824
+
825
+ 7
826
+ where J(L) is the current. If the SA parameter becomes
827
+ 0, there is no sample-to-sample fluctuation, which means
828
+ SA.
829
+ 1.
830
+ LD and HD regimes
831
+ Using Eq. (21), the SA parameter becomes
832
+ SA(L; J) = ⟨1/µ2⟩dis − ⟨1/µ⟩2
833
+ dis
834
+ ⟨1/µ⟩2
835
+ dis
836
+ ,
837
+ (33)
838
+ which is the same as the SA parameter for the diffusion
839
+ coefficient in the QTM [30]. When the mean waiting time
840
+ ⟨τ⟩ ≡
841
+ � ∞
842
+ 0
843
+ τψα(τ)dτ is finite (α > 1), we have µ → ⟨τ⟩
844
+ (L → ∞) by the law of large numbers. Therefore, in the
845
+ large-L limit, the current does not depend on the disorder
846
+ realization. Hence, the current is SA for α > 1. When
847
+ the mean waiting time diverges (α ≤ 1), the law of the
848
+ large numbers breaks down.
849
+ However, the generalized
850
+ central limit theorem is still valid. The PDF of the nor-
851
+ malized sum of the waiting times follows the one-sided
852
+ L´evy distribution [44],
853
+ �L
854
+ i=1 τi
855
+ L1/α
856
+ ⇒ Xα (L → ∞),
857
+ (34)
858
+ where Xα is a random variable following the one-sided
859
+ L´evy distribution of index α. The PDF of Xα denoted
860
+ by lα(x) with x > 0 is given by [44]
861
+ lα(x) = − 1
862
+ πx
863
+
864
+
865
+ k=1
866
+ Γ(kα + 1)
867
+ k!
868
+ (−cx−α)k sin (kπα),
869
+ (35)
870
+ (a)
871
+ 10−12
872
+ 10−10
873
+ 10−8
874
+ 10−6
875
+ 10−4
876
+ 10−2
877
+ 100
878
+ 102
879
+ 103
880
+ 104
881
+ 105
882
+ α = 0.7
883
+ α = 0.5
884
+ α = 0.3
885
+ ⟨J⟩dis(ρ(1 − ρ))−1
886
+ L
887
+ (b)
888
+ 10−12
889
+ 10−10
890
+ 10−8
891
+ 10−6
892
+ 10−4
893
+ 10−2
894
+ 100
895
+ 102
896
+ 103
897
+ 104
898
+ 105
899
+ 106
900
+ α = 2.5
901
+ α = 1.5
902
+ α = 0.5
903
+ α = 0.3
904
+ ⟨Jmax⟩dis
905
+ L
906
+ FIG. 5. Disorder average of the current as a function of L
907
+ for several α: (a) LD and HD regimes and (b) MC regimes.
908
+ Solid lines show the asymptotic results, i.e., Eqs. (38) and
909
+ (47). Squares are the results of numerical simulations, where
910
+ we calculated the maximal currents (Eq. (27)) for different
911
+ disorder realizations by Monte Carlo simulations. We used
912
+ 104 disorder realizations. Triangles are the results of the nu-
913
+ merical simulation of dynamics of the DTASEP (N = 1 for
914
+ (a) and ρ = 0.5 for (b)). We used 103 for L = 104 in the MC
915
+ regime and 104 disorder realizations for others.
916
+ where c = Γ(1 − α)τ α
917
+ c is the scale parameter. The first
918
+ and the second moment of X−1
919
+ α
920
+ are given by [30]
921
+ ⟨X−1
922
+ α ⟩ = Γ(1/α)
923
+ αc1/α , ⟨X−2
924
+ α ⟩ = Γ(2/α)
925
+ αc2/α .
926
+ (36)
927
+ The current can be represented by
928
+ J(L) ∼ ρ(1 − ρ)
929
+ L
930
+ L1/α
931
+ L1/α
932
+ �L
933
+ k=1 τk
934
+ ∼ ρ(1 − ρ)L1−1/αX−1
935
+ α
936
+ (37)
937
+ for L → ∞. Thus, the PDF of J is described by the in-
938
+ verse L´evy distribution. Using the first moment of the in-
939
+ verse L´evy distribution [30], we obtain the exact asymp-
940
+ totic behavior of the disorder average of the current,
941
+ ⟨J(L)⟩dis ∼ ρ(1 − ρ)Γ(α−1)
942
+ ατcΓ(1 − α)1/α L1−1/α.
943
+ (38)
944
+ Hence, the current becomes 0 (see Fig. 5(a)). We note
945
+ that since the scaling of ρ∗ follows Eq. (18), we do not
946
+ simulate at the same density.
947
+ Using the first and the second moments of 1/µ, we have
948
+ the SA parameter
949
+ lim
950
+ L→∞ SA(L; J) =
951
+
952
+
953
+
954
+
955
+
956
+ 0
957
+ (α > 1)
958
+ αΓ(2/α)
959
+ Γ(1/α)2 − 1
960
+ (α ≤ 1).
961
+ (39)
962
+ For α ≤ 1, the SA parameter is a nonzero constant, and
963
+ thus J becomes non-SA. Therefore, there is a transition
964
+ of SA property in the LD and HD regimes.
965
+ 2.
966
+ MC regime
967
+ When the system size is increased, we find a deeper and
968
+ deeper energy trap, that is, τm gets longer and longer.
969
+ Hence, Eq. (25) can be approximated as ⟨Tm⟩ ∼ τm, i.e.,
970
+ we can approximate the maximal current:
971
+ Jmax ∼ 1
972
+ τm
973
+ .
974
+ (40)
975
+ Since the PDF of the waiting times follow a power-law
976
+ distribution Eq. (3), the PDF of the normalized τm fol-
977
+ lows the Fr´echet distribution [42]:
978
+ τm
979
+ τcL1/α ⇒ Yα (L → ∞),
980
+ (41)
981
+ where Yα is a random variable following the Fr´echet dis-
982
+ tribution of index α. As derived in Appendix B, the PDF
983
+ of Yα, denoted fα(y) with y > 0, can be expressed as
984
+ fα(y) = αy−α−1 exp (−y−α).
985
+ (42)
986
+ Using Eq. (41), the maximal current can be represented
987
+ by
988
+ Jmax(L) ∼
989
+ 1
990
+ τcL1/α
991
+ τcL1/α
992
+ τm
993
+
994
+ 1
995
+ τcL1/α Y −1
996
+ α
997
+ (43)
998
+
999
+ 8
1000
+ for L → ∞. Thus, the PDF of Jmax is described by the
1001
+ inverse Fr´echet distribution.
1002
+ The PDF of Y −1
1003
+ α
1004
+ can be explicitly represented by the
1005
+ Fr´echet distribution:
1006
+ Pr(Y −1
1007
+ α
1008
+ ≤ z) = Pr(Yα ≥ z−1) =
1009
+ � ∞
1010
+ z−1 fα(y)dy.
1011
+ (44)
1012
+ The distribution is the Weibull distribution. We obtain
1013
+ the PDF of Y −1
1014
+ α , denoted by wα(z):
1015
+ wα(z) = αzα−1 exp (−zα).
1016
+ (45)
1017
+ The first and second moments of the Weibull distribution
1018
+ are given by
1019
+ ⟨Y −1
1020
+ α ⟩ = Γ
1021
+
1022
+ 1 + 1
1023
+ α
1024
+
1025
+ ,
1026
+ ⟨Y −2
1027
+ α ⟩ = Γ
1028
+
1029
+ 1 + 2
1030
+ α
1031
+
1032
+ .
1033
+ (46)
1034
+ From Eq. (46), we obtain the exact asymptotic behavior
1035
+ of the disorder average of the maximal current,
1036
+ ⟨Jmax(L)⟩dis ∼
1037
+ 1
1038
+ τcL1/α Γ
1039
+
1040
+ 1 + 1
1041
+ α
1042
+
1043
+ .
1044
+ (47)
1045
+ Therefore, the maximal current decreases with the sys-
1046
+ tem size L (see Fig. 5(b)).
1047
+ Let us consider the SA property for the maximal cur-
1048
+ rent. The SA parameter is defined as
1049
+ SA(L; Jmax) ≡ ⟨Jmax(L)2⟩dis − ⟨Jmax(L)⟩2
1050
+ dis
1051
+ ⟨Jmax(L)⟩2
1052
+ dis
1053
+ .
1054
+ (48)
1055
+ Using Eq. (43), we have
1056
+ lim
1057
+ L→∞ SA(L; Jmax) = ⟨Y −2
1058
+ α ⟩ − ⟨Y −1
1059
+ α ⟩
1060
+ 2
1061
+ ⟨Y −1
1062
+ α ⟩
1063
+ 2
1064
+ = Γ (1 + 2/α)
1065
+ Γ (1 + 1/α)2 − 1.
1066
+ (49)
1067
+ The SA parameter becomes a nonzero constant, i.e., the
1068
+ maximal current becomes non-SA (see Fig. 6(a)). This
1069
+ result differs from LD and HD, and there is no transition
1070
+ from SA to non-SA behavior for all α.
1071
+ B.
1072
+ Diffusivity
1073
+ Here, we consider sample-to-sample fluctuations of the
1074
+ diffusion coefficient.
1075
+ In the homogeneous TASEP, the
1076
+ diffusion coefficient becomes 0 for L → ∞ (Eq. (2)) be-
1077
+ cause of the many-body effect. D = 0 in the homoge-
1078
+ neous TASEP on a finite system implies the subdiffusion
1079
+ in that on an infinite system [45].
1080
+ 1.
1081
+ LD and HD regimes
1082
+ For the LD regime, ρ = N/L and 1 − ρ ∼ 1 for
1083
+ L → ∞ and N ≪ L. We define the number of holes
1084
+ (a)
1085
+ 10−2
1086
+ 10−1
1087
+ 100
1088
+ 101
1089
+ 102
1090
+ 103
1091
+ 104
1092
+ 0
1093
+ 0.5
1094
+ 1
1095
+ 1.5
1096
+ 2
1097
+ 2.5
1098
+ 3
1099
+ SA
1100
+ α
1101
+ (b)
1102
+ α
1103
+ ρ
1104
+ ρ∗
1105
+ 1 − ρ∗
1106
+ 1
1107
+ 0
1108
+ non-SA
1109
+ non-SA
1110
+ SA
1111
+ SA
1112
+ non-SA
1113
+ 1/2
1114
+ 2
1115
+ ⟨D⟩dis = 0
1116
+ ⟨D⟩dis = ∞
1117
+ ⟨D⟩dis = 0
1118
+ ⟨D⟩dis = 0
1119
+ ⟨D⟩dis > 0
1120
+ ⟨D⟩dis = 0
1121
+ LD
1122
+ MC
1123
+ HD
1124
+ FIG. 6. (a) Self-averaging parameter as a function of α. The
1125
+ squares and circles are the results of numerical simulations,
1126
+ where we calculated the maximal currents (Eq. (27)) and the
1127
+ diffusion coefficient (Eq. (31)) for different disorder realiza-
1128
+ tions by Monte Carlo simulations (L = 105), respectively.
1129
+ The triangles show the self-averaging parameter of the maxi-
1130
+ mal current obtained by the numerical simulation of the dy-
1131
+ namics of the DTASEP (L = 1000 and N = 500). We used
1132
+ 104 disorder realizations. The solid line represents Eq. (49).
1133
+ (b) Phase diagram based on diffusivity in the LD, MC, and
1134
+ HD regimes.
1135
+ as M ≡ L − N, i.e., 1−ρ = M/L. Therefore, for the HD
1136
+ regime, ρ = (L − M)/L ∼ 1 for L → ∞ and M ≪ L.
1137
+ Using Eq. (24), the disorder average of the diffusion co-
1138
+ efficient is given by
1139
+ ⟨D(L)⟩dis ∼
1140
+
1141
+
1142
+
1143
+
1144
+
1145
+
1146
+
1147
+
1148
+
1149
+ N −1/2
1150
+ �σ2
1151
+ µ3
1152
+
1153
+ dis
1154
+ (LD regime)
1155
+ M 3/2L−2
1156
+ �σ2
1157
+ µ3
1158
+
1159
+ dis
1160
+ (HD regime)
1161
+ (50)
1162
+
1163
+ 9
1164
+ (a)
1165
+ 10−4
1166
+ 10−3
1167
+ 10−2
1168
+ 10−1
1169
+ 100
1170
+ 101
1171
+ 102
1172
+ 103
1173
+ 102
1174
+ 103
1175
+ 104
1176
+ 105
1177
+ 106
1178
+ α = 0.8
1179
+ α = 1.8
1180
+ α = 0.4
1181
+ ⟨D⟩dis(Lρ)1/2(1 − ρ)−3/2
1182
+ L
1183
+ (b)
1184
+ 10−14
1185
+ 10−12
1186
+ 10−10
1187
+ 10−8
1188
+ 10−6
1189
+ 10−4
1190
+ 10−2
1191
+ 100
1192
+ 102
1193
+ 103
1194
+ 104
1195
+ 105
1196
+ 106
1197
+ α = 2.5
1198
+ α = 1.5
1199
+ α = 0.8
1200
+ α = 0.4
1201
+ ⟨D⟩dis
1202
+ L
1203
+ FIG. 7. Disorder average of the diffusion coefficient as a func-
1204
+ tion of L for several α: (a) LD and HD regimes and (b) MC
1205
+ regimes.
1206
+ Squares are the results of numerical simulations,
1207
+ where we calculated the diffusion coefficient (Eqs. (24) and
1208
+ (31)) for different disorder realizations by Monte Carlo sim-
1209
+ ulations (N = 1 for (a) and ρ = 0.5 for (b)). We used 104
1210
+ disorder realizations. Solid lines show the asymptotic results,
1211
+ i.e., Eqs. (53) and (61).
1212
+ for L → ∞. When the second moment of the waiting
1213
+ time ⟨τ 2⟩ ≡
1214
+ � ∞
1215
+ 0
1216
+ τ 2φα(τ)dτ is finite (α > 2), we have
1217
+ σ2 → ⟨τ 2⟩ (L → ∞) by the law of large numbers. It
1218
+ follows that the disorder average of D(L) is finite and
1219
+ given by
1220
+ ⟨D(L)⟩dis ∼
1221
+
1222
+
1223
+
1224
+
1225
+
1226
+
1227
+
1228
+
1229
+
1230
+ N −1/2 ⟨τ 2⟩
1231
+ ⟨τ⟩3
1232
+ (LD regime)
1233
+ M 3/2L−2 ⟨τ 2⟩
1234
+ ⟨τ⟩3
1235
+ (HD regime)
1236
+ (51)
1237
+ for L → ∞ and α > 2. Hence, the diffusion coefficient
1238
+ become non-zero constant for the LD regime, whereas it
1239
+ becomes 0 for the HD regime.
1240
+ For α < 2, the second moment of the waiting time di-
1241
+ verges. The disorder average of σ2/µ3, which was derived
1242
+ in Ref. [32], is obtained as
1243
+ �σ2
1244
+ µ3
1245
+
1246
+ dis
1247
+
1248
+
1249
+ L2−α
1250
+ (1 < α < 2)
1251
+ L2−1/α
1252
+ (α < 1).
1253
+ (52)
1254
+ Therefore, the disorder average of the diffusion coefficient
1255
+ is given by
1256
+ ⟨D(L)⟩dis ∝
1257
+
1258
+ L2−α
1259
+ (1 < α < 2)
1260
+ L2−1/α
1261
+ (α < 1)
1262
+ (53)
1263
+ for the LD regime and
1264
+ ⟨D(L)⟩dis ∝
1265
+
1266
+ L−α
1267
+ (1 < α < 2)
1268
+ L−1/α
1269
+ (α < 1)
1270
+ (54)
1271
+ for the HD regime, respectively.
1272
+ Hence, the diffusion
1273
+ coefficient for the LD regime diverges for 1 < α < 2 and
1274
+ 1/2 < α < 1, whereas it becomes 0 for α < 1/2 (see
1275
+ Fig. 7(a)). The diffusion coefficient for the HD regime
1276
+ becomes 0 for all α. The zero diffusion coefficient is a
1277
+ signature of many-body effect.
1278
+ Let us consider the SA property for the diffusion co-
1279
+ efficient in LD and HD regimes. The SA parameter is
1280
+ defined as
1281
+ SA(L; D) ≡ ⟨D(L)2⟩dis − ⟨D(L)⟩2
1282
+ dis
1283
+ ⟨D(L)⟩2
1284
+ dis
1285
+ .
1286
+ (55)
1287
+ The SA parameter goes to 0 in the large-L limit when
1288
+ the diffusion coefficient is SA.
1289
+ For α > 2, the second moment of waiting times exists;
1290
+ i.e., ⟨τ 2⟩ = � ∞
1291
+ 0
1292
+ τ 2ψα(τ)dτ.
1293
+ Thus, σ2/µ3 converges to
1294
+ ⟨τ 2⟩ / ⟨τ⟩2 for L → ∞. Therefore, ⟨D(L)2⟩dis −⟨D(L)⟩2
1295
+ dis
1296
+ converges to 0 for L → ∞, so that the diffusion coefficient
1297
+ is SA for α > 2.
1298
+ For 1 < α < 2, the second moment of σ2/µ3 was cal-
1299
+ culated in Ref. [32]. The SA parameter diverges as
1300
+ SA(L; D) ∝ ⟨D(L)2⟩dis
1301
+ ⟨D(L)⟩2
1302
+ dis
1303
+ ∝ Lα−1
1304
+ (56)
1305
+ for L → ∞. Therefore, the diffusion coefficient is non-SA
1306
+ for 1 < α < 2.
1307
+ For α < 1, both the first and the second moments of
1308
+ the waiting times diverge. σ2/µ3 can be represented as
1309
+ σ2
1310
+ µ3 = L2−1/αC(L),
1311
+ (57)
1312
+ where C(L) = L1/α �L
1313
+ i=1 τ 2
1314
+ i /(�L
1315
+ i=1 τi)3 is a random vari-
1316
+ able depending on the disorder realization. Hence, the
1317
+ SA parameter becomes
1318
+ SA(L; D) = ⟨D(L)2⟩dis
1319
+ ⟨D(L)⟩2
1320
+ dis
1321
+ − 1 = ⟨C(L)2⟩dis
1322
+ ⟨C(L)⟩2
1323
+ dis
1324
+ − 1.
1325
+ (58)
1326
+ Because �L
1327
+ i=1 τ 2
1328
+ i < (�L
1329
+ i=1 τ)3, 1/(�L
1330
+ i=1 τi)3 < C(L) <
1331
+ 1, i.e., 0 < ⟨C(L)⟩dis < 1 and 0 < ⟨C(L)2⟩dis < 1, the SA
1332
+ parameter is a finite value, i.e., the diffusion coefficient
1333
+ is non-SA for α < 1. These results are the same as those
1334
+ for the QTM.
1335
+ 2.
1336
+ MC regime
1337
+ When the system size is increased, we find a deeper and
1338
+ deeper energy trap, that is, τm gets longer and longer.
1339
+ Hence, Eq. (26) can be approximated as ⟨T 2
1340
+ m⟩ − ⟨Tm⟩2 ∼
1341
+ τ 2
1342
+ m, i.e., we can approximate the diffusion coefficient:
1343
+ D ∼ ρ−2
1344
+ τm
1345
+ .
1346
+ (59)
1347
+ By Eq. (41), the diffusion coefficient can be represented
1348
+ by
1349
+ D(L) ∼
1350
+ ρ−2
1351
+ τcL1/α
1352
+ τcL1/α
1353
+ τm
1354
+
1355
+ ρ−2
1356
+ τcL1/α Y −1
1357
+ α
1358
+ (60)
1359
+
1360
+ 10
1361
+ for L → ∞. Therefore, the PDF of the diffusion coeffi-
1362
+ cient is also described by the Weibull distribution. Using
1363
+ the first moment of the Weibull distribution, we obtain
1364
+ the exact asymptotic behavior of the disorder average of
1365
+ the diffusion coefficient,
1366
+ ⟨D(L)⟩dis ∼
1367
+ ρ−2
1368
+ τcL1/α Γ(1 + 1/α).
1369
+ (61)
1370
+ Therefore, the diffusion coefficient also decreases with the
1371
+ system size L (see Fig. 7(b)).
1372
+ Next, we consider the SA parameter of the diffusion
1373
+ coefficient in the MC regime. Using Eq. (60), we have
1374
+ lim
1375
+ L→∞ SA(L; D) = ⟨Y −2
1376
+ α ⟩ − ⟨Y −1
1377
+ α ⟩
1378
+ 2
1379
+ ⟨Y −1
1380
+ α ⟩
1381
+ 2
1382
+ = Γ (1 + 2/α)
1383
+ Γ (1 + 1/α)2 − 1,
1384
+ (62)
1385
+ which is the same as the SA parameter for the maximal
1386
+ current (see Fig. 6(a)). The transition point from SA to
1387
+ non-SA, which exists for the LD and HD regimes, dis-
1388
+ appears, and the diffusion coefficient is non-SA for all α
1389
+ (see Fig. 6(b)).
1390
+ VII.
1391
+ CONCLUSION
1392
+ In this paper, we have studied the TASEP on a
1393
+ quenched random energy landscape. In the LD and HD
1394
+ regimes, i.e., the dilute limit, the dynamics of the dis-
1395
+ ordered TASEP can be approximately described by the
1396
+ single-particle dynamics.
1397
+ On the other hand, the dy-
1398
+ namics in the MC regime become completely different
1399
+ from that in the dilute limit due to the many-body ef-
1400
+ fect. In particular, the LD and HD phases coexist in the
1401
+ MC regimes. By renewal theory, we provided exact re-
1402
+ sults for the current and diffusion coefficient. In the LD
1403
+ regime, the disorder average of the diffusion coefficient
1404
+ becomes 0 for α < 1/2, diverges for 1/2 < α < 2, and is
1405
+ non-zero constant for α > 2, which is the same as in the
1406
+ single-particle dynamics (Fig. 6(b)). On the other hand,
1407
+ in the HD and MC regimes, it becomes 0 in the large-L
1408
+ limit for all α (Fig. 6(b)) due to the many-body effect.
1409
+ Moreover, we introduced the SA parameter to quantify
1410
+ the SA property. We obtained a self-averaging and non-
1411
+ self-averaging transition for the current and the diffusion
1412
+ coefficient in the LD and HD regimes, which is the same
1413
+ as in the single-particle dynamics. However, in the MC
1414
+ regime, the current and diffusion coefficient are non-SA
1415
+ for all α, which is different from the LD and HD regimes.
1416
+ Therefore, many-body effects in quenched random energy
1417
+ landscapes decrease the diffusion coefficient and lead to
1418
+ a strong non-self-averaging feature.
1419
+ ACKNOWLEDGMENTS
1420
+ We thank K. Saito for fruitful discussions. T.A. was
1421
+ supported by JSPS Grant-in-Aid for Scientific Research
1422
+ (No. C JP21K033920).
1423
+ Appendix A: Passage time distribution
1424
+ In this Appendix, we derive the distribution of the pas-
1425
+ sage time Tm site m in the MC regime, where m is the
1426
+ site with the maximal mean waiting time. The passage
1427
+ time can be divided into the hole escape time xm and
1428
+ the particle escape time ym. At time t1, a particle es-
1429
+ capes from site m.
1430
+ At time t2, the subsequent parti-
1431
+ cle arrives at site m.
1432
+ The hole escape time is defined
1433
+ as xm = t2 − t1 (Fig. 8). At time t3, the particle es-
1434
+ capes from site m. The particle escape time is defined as
1435
+ ym = t3 − t2 (Fig. 8). To obtain the hole escape time at
1436
+ site m, we consider the hole dynamics. At site m, when
1437
+ the hole jump succeeds by the ith attempt, the PDF of
1438
+ the hole escape time xm follows the distribution of the
1439
+ sum of i IID variables following the exponential distri-
1440
+ bution, ψm−1(t) = τ −1
1441
+ m−1 exp (−t/τm−1), i.e., the Erlang
1442
+ distribution
1443
+ Er(xm; i, τm−1) =
1444
+ xi−1
1445
+ m
1446
+ (i − 1)!τ i
1447
+ m−1
1448
+ exp
1449
+
1450
+ − xm
1451
+ τm−1
1452
+
1453
+ ,
1454
+ (A1)
1455
+ and the success probability is given by ρm−1(1 −
1456
+ ρm−1)i−1. Therefore, the PDF f(xm) of xm follows the
1457
+ exponential distribution
1458
+ f(xm) =ρm−1
1459
+
1460
+
1461
+ i=1
1462
+ (1 − ρm−1)i−1Er(xm; i, τm−1)
1463
+ =ρm−1
1464
+ τm−1
1465
+ exp
1466
+
1467
+ − xm
1468
+ τm−1
1469
+
1470
+ ×
1471
+
1472
+
1473
+ i=1
1474
+ 1
1475
+ (i − 1)!
1476
+ �(1 − ρm−1)xm
1477
+ τm−1
1478
+ �i−1
1479
+ =Ex
1480
+
1481
+ xm; τm−1
1482
+ ρm−1
1483
+
1484
+ ,
1485
+ (A2)
1486
+ where Ex(x; τ) ≡ exp (−x/τ)/τ is the exponential distri-
1487
+ bution.
1488
+ Because a particle can not escape from site m until the
1489
+ neighbor site becomes empty, we must consider the effect
1490
+ k
1491
+ ?
1492
+ ?
1493
+ m
1494
+ ?
1495
+ particle jump
1496
+ k
1497
+ ?
1498
+ m
1499
+ ?
1500
+ ?
1501
+ hole jump
1502
+ k
1503
+ ?
1504
+ ?
1505
+ m
1506
+ ?
1507
+ particle jump
1508
+ t
1509
+ t1
1510
+ t2
1511
+ t3
1512
+ xm
1513
+ ym
1514
+ Tm
1515
+ FIG. 8.
1516
+ Particle dynamics during the passage time.
1517
+ The
1518
+ filled and dashed-line circles denote particles and holes, re-
1519
+ spectively. The question mark is either a particle or a hole.
1520
+
1521
+ 11
1522
+ of site m + 1. Using the same way of the derivation of
1523
+ Eq. (A2), the PDF g(ym+1) of the particle escape time
1524
+ ym+1 at site m + 1 is given by
1525
+ g(ym+1) = Ex
1526
+
1527
+ ym+1;
1528
+ τm+1
1529
+ 1 − ρm+2
1530
+
1531
+ .
1532
+ (A3)
1533
+ Using Eq. (A3), we derive the joint PDF of the hole es-
1534
+ cape time xm and the particle escape time ym. When the
1535
+ sum of the hole escape time xm and the particle escape
1536
+ time ym is larger than the particle escape time ym+1, a
1537
+ particle at site m can jump to site m+1. When a prticle
1538
+ succeeds to jump to site m+1 once, i.e., xm+ym > ym+1,
1539
+ the weighted joint PDF h1(xm, ym) of xm and ym is given
1540
+ by
1541
+ h1(xm, ym) = f(xm)Ex(ym; τm)
1542
+ � xm+ym
1543
+ 0
1544
+ dym+1 g(ym+1).
1545
+ (A4)
1546
+ When a particle jump succeeds on the ith attempts (i >
1547
+ 1), xm + y′
1548
+ m < ym+1 < xm + ym, where y′
1549
+ m follows the
1550
+ Erlang distribution Er(y′
1551
+ m; i − 1, τm) and ym is sum of
1552
+ y′
1553
+ m and the IID random variable y with the exponential
1554
+ distribution Ex(y; τm).
1555
+ Then, the weighted joint PDf
1556
+ hi(xm, ym) of xm and ym is given by
1557
+ hi(xm, ym) = f(xm)
1558
+ � ym
1559
+ 0
1560
+ dy′
1561
+ m Ex(ym − y′
1562
+ m; τm)Er(y′
1563
+ m; i − 1, τm)
1564
+ � xm+ym
1565
+ xm+y′m
1566
+ dym+1 g(ym+1).
1567
+ (A5)
1568
+ Therefore, the joint PDF h(xm, ym) of xm and ym is given by
1569
+ h(xm, ym) =
1570
+
1571
+
1572
+ i=1
1573
+ hi(xm, ym)
1574
+ = f(xm)Ex(ym; τm) +
1575
+ τm+1
1576
+ 1−ρm+2
1577
+ τm −
1578
+ τm+1
1579
+ 1−ρm+2
1580
+ exp
1581
+
1582
+ −1 − ρm+2
1583
+ τm+1
1584
+ xm
1585
+
1586
+ f(xm)(Ex(ym; τm) − g(ym)).
1587
+ (A6)
1588
+ By the convolutional intergration of h(xm, ym), we have the PDF Φ(Tm) of the passage time Tm
1589
+ Φ(Tm) =
1590
+ � Tm
1591
+ 0
1592
+ dx h(x, Tm − x)
1593
+ = τm
1594
+ ρm−1
1595
+ τm−1
1596
+ (ζ1 + ζ2ζ3)Ex(Tm; τm) − ζ1f(Tm) − ζ2g(Tm) + ζ3E
1597
+
1598
+ Tm;
1599
+ 1
1600
+ ρm−1
1601
+ τm−1 + 1−ρm+2
1602
+ τm+1
1603
+
1604
+ ,
1605
+ (A7)
1606
+ where
1607
+ ζ1 ≡
1608
+ 1
1609
+ τm
1610
+ ρm−1
1611
+ τm−1 − 1, ζ2 ≡
1612
+ 1
1613
+ τm
1614
+ 1−ρm+2
1615
+ τm+1
1616
+ − 1
1617
+ , ζ3 ≡
1618
+ 1
1619
+ τm
1620
+
1621
+ ρm−1
1622
+ τm−1 + 1−ρm+2
1623
+ τm+1
1624
+
1625
+ − 1
1626
+ .
1627
+ Next, we derive the mean and variance of the passage time. The Laplace transform of Φ(Tm) with respect to s is
1628
+ given by
1629
+ ˆΦ(s) ≡
1630
+ � ∞
1631
+ 0
1632
+ dTm e−sTmΦ(Tm)
1633
+ = τm
1634
+ ρm−1
1635
+ τm−1
1636
+ (ζ1 + ζ2ζ3)
1637
+ 1
1638
+ τms + 1 −
1639
+ ζ1
1640
+ τm−1
1641
+ ρm−1 s + 1 −
1642
+ ζ2
1643
+ τm+1
1644
+ 1−ρm+2 s + 1 +
1645
+ ζ3
1646
+ s
1647
+ ρm−1
1648
+ τm−1 +
1649
+ 1−ρm+2
1650
+ τm+1
1651
+ + 1.
1652
+ (A8)
1653
+ It follows that the mean and variance of the passage time are given by
1654
+ ⟨Tm⟩ = τm + τm−1
1655
+ ρm−1
1656
+ +
1657
+ ρm−1
1658
+ τm−1
1659
+ ρm−1
1660
+ τm−1 + 1−ρm+2
1661
+ τm+1
1662
+ τm+1
1663
+ 1 − ρm+2
1664
+ ,
1665
+ (A9)
1666
+ ⟨T 2
1667
+ m⟩ − ⟨Tm⟩2 = τ2
1668
+ m +
1669
+ � τm−1
1670
+ ρm−1
1671
+ �2
1672
+ +
1673
+
1674
+ τm+1
1675
+ 1 − ρm+2
1676
+ �2
1677
+
1678
+ 3
1679
+
1680
+ ρm−1
1681
+ τm−1 + 1−ρm+2
1682
+ τm+1
1683
+ �2 .
1684
+ (A10)
1685
+ Appendix B: Fr´echet distribution
1686
+ Here, we derive that when random variables follow a
1687
+ power-law distribution (Eq. (3)), the maximum of those
1688
+ follows the Fr´echet distribution using the extreme value
1689
+
1690
+ 12
1691
+ theory [42]. We define τ1, . . . , τL as the random variables
1692
+ which follow the power-law distribution with exponent α.
1693
+ The probability for τm = max{τ1, . . . , τL} ≤ s is given
1694
+ by
1695
+ Pr(τm ≤ s) =
1696
+ L
1697
+
1698
+ i=1
1699
+ Pr(τi ≤ s) = G(s)L,
1700
+ (B1)
1701
+ where G(s) = Pr(τi ≤ s) = 1 − (s/τc)−α. We normalize
1702
+ τm as
1703
+ Xα =
1704
+ τm
1705
+ τcL1/α
1706
+ (B2)
1707
+ for L → ∞. It follows that Pr(Xα ≤ x) = Fα(x) is given
1708
+ by
1709
+ Fα(x) = lim
1710
+ L→∞ G(τcL1/αx)L = exp (−x−α).
1711
+ (B3)
1712
+ Therefore, the normalized τm follows the Fr´echet distri-
1713
+ bution.
1714
+ [1] B. Derrida, Phys. Rep. 301, 65 (1998).
1715
+ [2] C.
1716
+ Arita,
1717
+ M.
1718
+ E.
1719
+ Foulaadvand,
1720
+ and
1721
+ L.
1722
+ Santen,
1723
+ Phys. Rev. E 95, 032108 (2017).
1724
+ [3] T.
1725
+ Chou
1726
+ and
1727
+ G.
1728
+ Lakatos,
1729
+ Phys. Rev. Lett. 93, 198101 (2004).
1730
+ [4] L. Ciandrini,
1731
+ I. Stansfield,
1732
+ and M.
1733
+ C.
1734
+ Romano,
1735
+ Phys. Rev. E 81, 051904 (2010).
1736
+ [5] A.
1737
+ Dana
1738
+ and
1739
+ T.
1740
+ Tuller,
1741
+ Nucleic Acids Res. 42, 9171 (2014).
1742
+ [6] M.
1743
+ Kardar,
1744
+ G.
1745
+ Parisi,
1746
+ and
1747
+ Y.-C.
1748
+ Zhang,
1749
+ Phys. Rev. Lett. 56, 889 (1986).
1750
+ [7] K. Johansson, Comm. Math. Phys. 209, 437 (2000).
1751
+ [8] C.
1752
+ A.
1753
+ Tracy
1754
+ and
1755
+ H.
1756
+ Widom,
1757
+ Comm. Math. Phys. 290, 129 (2009).
1758
+ [9] A. Aggarwal, Duke Math. J. 167, 269 (2018).
1759
+ [10] T.
1760
+ Sasamoto
1761
+ and
1762
+ H.
1763
+ Spohn,
1764
+ Phys. Rev. Lett. 104, 230602 (2010);
1765
+ Nuclear Phys. B 834, 523 (2010).
1766
+ [11] G.
1767
+ Amir,
1768
+ I.
1769
+ Corwin,
1770
+ and
1771
+ J.
1772
+ Quastel,
1773
+ Comm. Pure Appl. Math. 64, 466 (2011).
1774
+ [12] B.
1775
+ Derrida
1776
+ and
1777
+ J.
1778
+ L.
1779
+ Lebowitz,
1780
+ Phys. Rev. Lett. 80, 209 (1998).
1781
+ [13] L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and
1782
+ C. Landim, Phys. Rev. Lett. 94, 030601 (2005).
1783
+ [14] D.
1784
+ Lips,
1785
+ A.
1786
+ Ryabov,
1787
+ and
1788
+ P.
1789
+ Maass,
1790
+ Phys. Rev. Lett. 121, 160601 (2018).
1791
+ [15] R.
1792
+ J.
1793
+ Concannon
1794
+ and
1795
+ R.
1796
+ A.
1797
+ Blythe,
1798
+ Phys. Rev. Lett. 112, 050603 (2014).
1799
+ [16] G.
1800
+ Tripathy
1801
+ and
1802
+ M.
1803
+ Barma,
1804
+ Phys. Rev. E 58, 1911 (1998).
1805
+ [17] C. Enaud and B. Derrida, Europhys. Lett. 66, 83 (2004).
1806
+ [18] R.
1807
+ J.
1808
+ Harris
1809
+ and
1810
+ R.
1811
+ B.
1812
+ Stinchcombe,
1813
+ Phys. Rev. E 70, 016108 (2004).
1814
+ [19] R.
1815
+ Juh´asz,
1816
+ L.
1817
+ Santen,
1818
+ and
1819
+ F.
1820
+ Igl´oi,
1821
+ Phys. Rev. E 74, 061101 (2006).
1822
+ [20] R.
1823
+ B.
1824
+ Stinchcombe
1825
+ and
1826
+ S.
1827
+ L.
1828
+ A.
1829
+ de
1830
+ Queiroz,
1831
+ Phys. Rev. E 83, 061113 (2011).
1832
+ [21] J. S. Nossan, J. Phys. A: Math. Theor. 46, 315001 (2013).
1833
+ [22] C.
1834
+ Bahadoran
1835
+ and
1836
+ T.
1837
+ Bodineau,
1838
+ Braz. J. Probab. Stat. 29, 282 (2015).
1839
+ [23] T.
1840
+ Banerjee
1841
+ and
1842
+ A.
1843
+ Basu,
1844
+ Phys. Rev. Research 2, 013025 (2020).
1845
+ [24] B.
1846
+ Derrida,
1847
+ M.
1848
+ R.
1849
+ Evans,
1850
+ and
1851
+ D.
1852
+ Mukamel,
1853
+ J. Phys. A: Math. Gen. 26, 4911 (1993).
1854
+ [25] I.
1855
+ Neri,
1856
+ N.
1857
+ Kern,
1858
+ and
1859
+ A.
1860
+ Parmeggiani,
1861
+ Phys. Rev. Lett. 107, 068702 (2011).
1862
+ [26] I.
1863
+ Neri,
1864
+ N.
1865
+ Kern,
1866
+ and
1867
+ A.
1868
+ Parmeggiani,
1869
+ New J. Phys. 15, 085005 (2013).
1870
+ [27] D. V. Denisov, D. M. Miedema, B. Nienhuis,
1871
+ and
1872
+ P. Schall, Phys. Rev. E 92, 052714 (2015).
1873
+ [28] R. Metzler and J. Klafter, Phys. Rep. 339, 1 (2000).
1874
+ [29] J. P. Bouchaud and A. Georges, Phys. Rep. 195 (1990).
1875
+ [30] T.
1876
+ Akimoto,
1877
+ E.
1878
+ Barkai,
1879
+ and
1880
+ K.
1881
+ Saito,
1882
+ Phys. Rev. Lett. 117, 180602 (2016);
1883
+ Phys. Rev. E 97, 052143 (2018).
1884
+ [31] L. Luo and M. Yi, Phys. Rev. E 97, 042122 (2018).
1885
+ [32] T.
1886
+ Akimoto
1887
+ and
1888
+ K.
1889
+ Saito,
1890
+ Phys. Rev. E 101, 042133 (2020).
1891
+ [33] R.
1892
+ Metzler,
1893
+ L.
1894
+ Sanders,
1895
+ M.
1896
+ A.
1897
+ Lomholt,
1898
+ L.
1899
+ Lizana,
1900
+ K.
1901
+ Fogelmark,
1902
+ and
1903
+ T.
1904
+ Ambj¨ornsson,
1905
+ EPJ Special Topics 223, 3287 (2014).
1906
+ [34] L. P. Sanders, M. A. Lomholt, L. Lizana,
1907
+ K. Fo-
1908
+ gelmark,
1909
+ R.
1910
+ Metzler,
1911
+ and
1912
+ T.
1913
+ Ambj¨ornsson,
1914
+ New J. Phys. 16, 113050 (2014).
1915
+ [35] A. Gran´eli, C. C. Yeykal, R. B. Robertson,
1916
+ and E. C.
1917
+ Greene, Proc. Natl. Acad. Sci. U.S.A. 103, 1221 (2006).
1918
+ [36] Y.
1919
+ M.
1920
+ Wang,
1921
+ R.
1922
+ H.
1923
+ Austin,
1924
+ and
1925
+ E.
1926
+ C.
1927
+ Cox,
1928
+ Phys. Rev. Lett. 97, 048302 (2006).
1929
+ [37] E. Yamamoto, T. Akimoto, Y. Hirano, M. Yasui,
1930
+ and
1931
+ K. Yasuoka, Phys. Rev. E 89, 022718 (2014).
1932
+ [38] I. Sakai and T. Akimoto, arXiv:2208.10102 (2022).
1933
+ [39] T.
1934
+ Akimoto
1935
+ and
1936
+ K.
1937
+ Saito,
1938
+ Phys. Rev. E 99, 052127 (2019).
1939
+ [40] S. Gupta, S. N. Majumdar, C. Godr`eche, and M. Barma,
1940
+ Phys. Rev. E 76, 021112 (2007).
1941
+ [41] S.
1942
+ A.
1943
+ Janowsky
1944
+ and
1945
+ J.
1946
+ L.
1947
+ Lebowitz,
1948
+ Phys. Rev. A 45, 618 (1992).
1949
+ [42] L. de Haan and A. Ferreira, Extreme value theory: an
1950
+ introduction, Vol. 21 (Springer, 2006).
1951
+ [43] C.
1952
+ Godr`eche
1953
+ and
1954
+ J.
1955
+ M.
1956
+ Luck,
1957
+ J. Stat. Phys. 104, 489 (2001).
1958
+ [44] W. Feller, An Introduction to Probability Theory and its
1959
+ Applications, 2nd ed., Vol. 2 (Wiley, New York, 1971).
1960
+ [45] H. van Beijeren, J. Stat. Phys. 63, 47 (1991).
1961
+
G9AyT4oBgHgl3EQfrfnZ/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
G9AzT4oBgHgl3EQfUvyD/content/2301.01272v1.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2fa2249070e547e18f10da617d27ba2886a8abbeec452ac176c3436e32ce1d9d
3
+ size 760092
G9AzT4oBgHgl3EQfUvyD/vector_store/index.faiss ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab1f425b7a342398bceba4798e0a5d8c4ccfd394a1bbdab52e08645e1890f58a
3
+ size 1638445
G9AzT4oBgHgl3EQfUvyD/vector_store/index.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:08567f44160a8a229683e15e6f50f29d824ee37f0d2ef49224eb93af1bcaa533
3
+ size 62556
GtFIT4oBgHgl3EQfWyuf/content/tmp_files/2301.11241v1.pdf.txt ADDED
The diff for this file is too large to render. See raw diff
 
GtFIT4oBgHgl3EQfWyuf/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
J9E1T4oBgHgl3EQfYgQR/content/tmp_files/2301.03138v1.pdf.txt ADDED
@@ -0,0 +1,2168 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arXiv:2301.03138v1 [math-ph] 9 Jan 2023
2
+ Gaudin Hamiltonians on unitarizable modules over classical
3
+ Lie (super)algebras
4
+ Wan Keng Cheong and Ngau Lam
5
+ Abstract
6
+ Let L be a tensor product of unitarizable irreducible highest weight modules over
7
+ the Lie (super)algebra G, where G is gl(m|n), osp(2m|2n) or spo(2m|2n). We show
8
+ that the Gaudin Hamiltonians associated to G are diagonalizable with simple spectrum
9
+ on the space spanned by singular vectors of any fixed weight in L. In particular, we
10
+ establish the diagonalization of the Gaudin Hamiltonians, associated to any of the
11
+ orthogonal Lie algebra so(2n) and the symplectic Lie algebra sp(2n), on the space
12
+ spanned by singular vectors of any fixed weight in the tensor product of infinite-
13
+ dimensional unitarizable irreducible highest weight modules.
14
+ 1
15
+ Introduction
16
+ The Gaudin model was introduced by Gaudin [G1, G2] to describe a completely integrable
17
+ quantum spin chain associated to any finite-dimensional simple Lie algebra G. Let (·, ·)
18
+ be a non-degenerate invariant bilinear form on G. Let {Ia | a = 1, . . . , d} be a basis for G
19
+ and {Ia | a = 1, . . . , d} the dual basis with respect to the bilinear form (·, ·), where d is the
20
+ dimension of G. The Casimir symmetric tensor Ω for G is defined to be Ω = �d
21
+ a=1 Ia ⊗ Ia.
22
+ Fix ℓ ∈ N with ℓ ≥ 2, and let z1, . . . , zℓ be distinct complex numbers. For any i = 1, . . . , ℓ,
23
+ the Gaudin Hamiltonian Hi of the Gaudin model associated to G is defined to be
24
+ Hi =
25
+
26
+
27
+ j=1
28
+ j̸=i
29
+ Ω(ij)
30
+ zi − zj
31
+ ,
32
+ where Ω(ij) is defined as in (4.1). The Gaudin Hamiltonians H1, . . . , Hℓ act on the tensor
33
+ product M1 ⊗ · · · ⊗ Mℓ, where each Mi is a G-module, and they are mutually commuting
34
+ operators.
35
+ To find common eigenvectors for Gaudin Hamiltonians is one of the main problems
36
+ of studying the Gaudin model. Bethe ansatz method provides an explicit construction
37
+ of common eigenvectors from the solutions of the so-called Bethe ansatz equations and
38
+ proves to be effective for the special linear Lie algebra sl(2, C) (cf. [G1]). The eigenvectors
39
+ obtained by this method are called Bethe vectors. Babujian and Flume [BF] generalized
40
+ the Bethe ansatz equations to the case of any simple Lie algebra. In [FFR], Feigin, Frenkel
41
+ and Reshetikhin proposed a new approach based on Wakimoto modules over the affine
42
+ Lie algebra at the critical level. They obtained the Bethe vectors by restricting certain
43
+ invariant functionals on tensor products of Wakimoto modules and found as well that the
44
+ 1
45
+
46
+ 2
47
+ Cheong and Lam
48
+ Bethe vectors are in close connection with the solutions of the Knizhnik–Zamolodchikov
49
+ equations [KZ] (see also [EFK]).
50
+ In this paper, we are interested in the super version of Gaudin Hamiltonians.
51
+ For
52
+ precise definitions of such Gaudin Hamiltonians, see (4.2) and (5.1). We find an interesting
53
+ connection between the Gaudin Hamiltonians associated to the (trivial) central extension
54
+ of any infinite-rank classical Lie (super)algebra of type a, c, d and the Gaudin Hamiltonians
55
+ associated to the (trivial) central extension of the corresponding Lie algebra. Our main
56
+ tool is super duality (cf. [CL2, CLW1, CLW2]), which asserts that there are equivalences
57
+ of tensor categories between the parabolic BGG categories �O of ˜g-modules, O[m] of g[m]-
58
+ modules and O[m] of g[m]-modules (cf. Theorem 2.10). The following diagrams summarize
59
+ the relationships among these categories.
60
+ �O
61
+ O[m]
62
+ O[0]
63
+ T[m]
64
+ T[0]
65
+ �O
66
+ O[m]
67
+ O[0]
68
+ T [m]
69
+ T [0]
70
+ Here the tensor functors T[m], T [m], T[0], and T [0], defined in Section 2.6, are equivalences of
71
+ tensor categories for m ∈ N. Notice that ˜g, g[m] and g[m], for m ∈ N, are Lie superalgebras
72
+ while g[0] and g[0] are Lie algebras. In other words, super duality gives equivalences of
73
+ categories between parabolic BGG categories for Lie superalgebras and the corresponding
74
+ Lie algebras.
75
+ We view the Gaudin Hamiltonians for ˜g (resp., g[m] and g[m]) as linear operators on
76
+ the tensor product of modules over ˜g (resp., g[m] and g[m]) in the category �O (resp., O[m]
77
+ and O[m]).
78
+ Instead of constructing eigenvectors for these operators directly, we apply
79
+ super duality and the work [CaL] to establish one-to-one correspondences relating the
80
+ sets of eigenvectors of the Gaudin Hamiltonians for ˜g, g[m] and g[m] for m ∈ Z+ (see
81
+ Theorem 4.7). Furthermore, by using the notion of truncation functors (see (2.18)), we
82
+ relate the eigenvectors of Gaudin Hamiltonians for g[m] and g[m] to the eigenvectors of
83
+ Gaudin Hamiltonians for their finite-rank counterparts g[m]n and g[m]n, for n ∈ N.
84
+ Let G be a finite-dimensional simple Lie algebra and Li a finite-dimensional irreducible
85
+ module over G for each i = 1, . . . , ℓ. Rybnikov shows that for generic z1, . . . , zℓ, the Gaudin
86
+ Hamiltonians are diagonalizable with simple spectrum on the space spanned by singular
87
+ vectors in L1 ⊗ · · · ⊗ Lℓ (see [MV, MTV1, MTV2, FFRy, LMV] as well).
88
+ We would like to extend Rybnikov’s result to the case of Lie (super)algebras in this
89
+ paper. We focus our attention on the finite-dimensional classical Lie (super)algebra G
90
+ x[m]n
91
+ of type x, where x denotes a fixed type among a, c, d (see Sections 2.1, 2.2 and 2.3).
92
+ Note that G
93
+ a[m]n ∼= gl(m|n), G
94
+ c[m]n ∼= spo(2m|2n) and G
95
+ d[m]n ∼= osp(2m|2n). Let L =
96
+ L1 ⊗ · · · ⊗ Lℓ, where each Li is an irreducible highest weight module over G
97
+ x[m]n. Suppose
98
+ that for x = a, each Li is a polynomial module, and for x = c, d, all highest weights of Li
99
+ lie in Q
100
+ x,I(m|n) or Q
101
+ x,II(m|n) (see Section 3.3). We show that the Gaudin Hamiltonian
102
+ Hi[m]n associated to G
103
+ x[m]n, for i = 1, . . . , ℓ, is diagonalizable with simple spectrum on
104
+ the space spanned by singular vectors in any finite direct sum of weight spaces of L (cf.
105
+ Theorem 5.5 and Theorem 5.7). We will see later that such Li’s are unitarizable modules
106
+
107
+ Gaudin Hamiltonians on unitarizable modules
108
+ 3
109
+ over G
110
+ x[m]n with respect to a natural choice of ∗-structures on G
111
+ x[m]n (see Sections 3.2
112
+ and 3.3).
113
+ Let us explain a little about how we obtain the diagonalization. For x = c, d (resp.
114
+ x = a), we relate each Hi[m]n to its corresponding Gaudin Hamiltonian on the tensor
115
+ product of some finite-dimensional irreducible modules over G, where G is the orthogonal
116
+ Lie algebra so(2p) or the symplectic Lie algebra sp(2p) (resp., G is the general linear
117
+ Lie algebra gl(p)) for some p ∈ N. These correspondences should give a procedure of
118
+ constructing eigenvectors of the Gaudin Hamiltonians for classical Lie (super)algebras
119
+ from the ones for the corresponding Lie algebras.
120
+ While there is a great deal of work on Gaudin models, not much appears to be known
121
+ about the super Gaudin Hamiltonians. In [MVY], Mukhin, Vicedo and Young investigate
122
+ the Gaudin Hamiltonians associated to the general linear Lie superalgebra gl(m|n) and
123
+ show that the operators are diagonalizable with simple spectrum on the space spanned
124
+ by singular vectors in the tensor product of the natural module Cm|n. In [KuM], Kulish
125
+ and Manojlovi´c explicitly construct eigenvectors for the Gaudin Hamiltonians associated
126
+ to the ortho-symplectic Lie superalgebra osp(1|2). The reader may also want to consult
127
+ [HMVY, L, LM] for related results. We hope that this paper can provide new insights into
128
+ the study of Gaudin models.
129
+ This paper is organized as follows. In Section 2, we fix notation and review some back-
130
+ ground materials on classical Lie (super)algebras and their central extensions. We also de-
131
+ fine the parabolic BGG categories �O, O[m] and O[m] associated to the Lie (super)algebras
132
+ ˜g, g[m] and g[m], respectively, and discuss super duality which gives equivalences of these
133
+ categories. In Section 3, we give a brief introduction to ∗-structures and study the unita-
134
+ rizable modules which will be considered in our study of Gaudin Hamiltonians. In Section
135
+ 4, we investigate Gaudin Hamiltonians associated to the central extensions of finite-rank
136
+ and infinite-rank Lie (super)algebras, and establish the one-to-one correspondences be-
137
+ tween the sets of eigenvectors (see Theorem 4.7 and Proposition 4.8). In Section 5, we
138
+ concentrate on Gaudin Hamiltonians for finite-dimensional classical Lie (super)algebras
139
+ and prove Theorem 5.5 and Theorem 5.7, which give an affirmative answer to the diago-
140
+ nalization of the operators on unitarizable modules.
141
+ Notations. Throughout the paper, N stands for the set of positive integers, Z for the set
142
+ of integers, Z∗ for the set of nonzero integers, Z+ for the set of non-negative integers, 1
143
+ 2Z
144
+ for the set of half integers and integers, and C for the set of complex numbers. All vector
145
+ spaces, algebras, tensor products, et cetera, are over C.
146
+ 2
147
+ Preliminaries
148
+ In this section, we first define the finite-rank and infinite-rank Lie (super)algebras �Gx,
149
+ Gx[m]n and G
150
+ x[m]n, where x denotes one of the three types a, c, d. We consider their central
151
+ extensions �gx, gx[m]n and gx[m]n and the parabolic BGG categories �Ox (resp., Ox[m]n and
152
+ O
153
+ x[m]n) of modules over �gx (resp., gx[m]n and gx[m]n).
154
+ We then recall the truncation
155
+ functors which relate Ox[m]∞ and O
156
+ x[m]∞ to Ox[m]n and O
157
+ x[m]n, respectively, for n ∈ N.
158
+ Finally, we describe the tensor functors T[m] and T [m] and their properties. We refer the
159
+
160
+ 4
161
+ Cheong and Lam
162
+ readers to [CL2, Sections 2 and 3] for type a and [CLW1, Sections 2 and 3] for types c, d
163
+ for details (see also [CW, Sections 6.1 and 6.2] and [CaL, Section 2.4]). We fix m ∈ Z+
164
+ and n ∈ N ∪ {∞} throughout this paper.
165
+ Let �V denote the superspace over C with ordered basis {vr | r ∈ 1
166
+ 2Z}. The parity of vr
167
+ is defined as follows: |vr| = ¯0 if r ∈ Z, and |vr| = ¯1 if r ∈ 1
168
+ 2 + Z.
169
+ Let gl(�V ) be the Lie superalgebra consisting of all linear endomorphisms on �V which
170
+ vanish on all but finitely many vr’s. For i, j ∈ 1
171
+ 2Z, we let Ei,j be the linear endomorphism
172
+ on �V defined by
173
+ Ei,j(vr) = δjrvi
174
+ for r ∈ 1
175
+ 2Z,
176
+ where δ is the Kronecker delta.
177
+ The Lie superalgebra gl(�V ) is spanned by Ei,j with
178
+ i, j ∈ 1
179
+ 2Z.
180
+ The Lie superalgebra gl(�V ) has a central extension, denoted by �gl(�V ), by the one-
181
+ dimensional center CK corresponding to the following 2-cocycle (cf. [CL1, p. 99]):
182
+ τ(A, B) := Str([J, A]B),
183
+ A, B ∈ gl(�V ),
184
+ (2.1)
185
+ where J = − �
186
+ r≥ 1
187
+ 2 Er,r and Str denotes the supertrace.
188
+ In fact, the cocycle τ is a
189
+ coboundary. Moreover, there is an isomorphism ι from the direct sum of Lie superalgebras
190
+ gl(�V ) ⊕ CK to �gl(�V ) defined by
191
+ ι(A) = A + Str(JA)K,
192
+ for A ∈ gl(�V ),
193
+ and
194
+ ι(K) = K.
195
+ (2.2)
196
+ Let
197
+ Jm(n) =
198
+ ß
199
+ ±1
200
+ 2, ±3
201
+ 2, . . . , ±(m − 1
202
+ 2)
203
+
204
+ ∪ {0} ∪ { ±j | j ∈ N, j < n + 1 },
205
+ Jm(n) = {±1, . . . , ±m} ∪ {0} ∪
206
+
207
+ ±(j − 1
208
+ 2)
209
+ ��� j ∈ N, j < n + 1
210
+
211
+ ,
212
+ �J(n) =
213
+
214
+ r ∈ 1
215
+ 2Z
216
+ ��� − n ≤ r ≤ n
217
+
218
+ ,
219
+
220
+ m(n) = Jm(n)\{0},
221
+ J
222
+ ×
223
+ m(n) = Jm(n)\{0},
224
+ �J×(n) = �J(n))\{0},
225
+ J+
226
+ m(n) = { r ∈ Jm(n) | r > 0 },
227
+ J
228
+ +
229
+ m(n) =
230
+
231
+ r ∈ Jm(n)
232
+ �� r > 0
233
+
234
+ ,
235
+ �J+(n) =
236
+
237
+ r ∈ �J(n)
238
+ ��� r > 0
239
+ ©
240
+ .
241
+ We let �V (n), Vm(n), Vm(n), �V ×(n), V ×
242
+ m (n) and V
243
+ ×
244
+ m (n) be the subspaces of �V with basis
245
+ {vi} indexed by �J(n), Jm(n), Jm(n), �J×(n), J×
246
+ m(n) and J
247
+ ×
248
+ m(n), respectively. This gives rise
249
+ to subalgebras gl(�V (n)), gl(Vm(n)), gl(Vm(n)), gl(�V ×(n)), gl(V ×
250
+ m (n)) and gl(V
251
+ ×
252
+ m (n)) of the
253
+ Lie superalgebra gl(�V ). Let �b := �
254
+ r≤s,r,s∈ 1
255
+ 2Z CEr,s denote the standard Borel subalgebra
256
+ of gl(�V ).
257
+
258
+ Gaudin Hamiltonians on unitarizable modules
259
+ 5
260
+ We will drop the symbol (n) if n = ∞. For example, Jm := Jm(∞). Define the total
261
+ orders of Jm and Jm by
262
+ . . . <Jm −3 <Jm −2 <Jm −1 <Jm −(m − 1
263
+ 2) <Jm . . . <Jm −3
264
+ 2 <Jm −1
265
+ 2
266
+ <Jm 0 <Jm
267
+ 1
268
+ 2 <Jm
269
+ 3
270
+ 2 <Jm . . . <Jm m − 1
271
+ 2 <Jm 1 <Jm 2 <Jm 3 <Jm . . .
272
+ and
273
+ . . . <Jm −5
274
+ 2 <Jm −3
275
+ 2 <Jm −1
276
+ 2 <Jm −m <Jm . . . <Jm −2 <Jm −1
277
+ <Jm 0 <Jm 1 <Jm 2 <Jm . . . <Jm m <Jm
278
+ 1
279
+ 2 <Jm
280
+ 3
281
+ 2 <Jm
282
+ 5
283
+ 2 <Jm . . . ,
284
+ respectively. The orderings give the standard Borel subalgebras
285
+ b[m] :=
286
+
287
+ r≤Jms,
288
+ r,s∈Jm
289
+ CEr,s
290
+ and
291
+ b[m] :=
292
+
293
+ r≤Jms,
294
+ r,s∈Jm
295
+ CEr,s
296
+ of gl(Vm) and gl(V m), respectively.
297
+ 2.1
298
+ General linear superalgebras �Ga
299
+ Let �V +(n), V +
300
+ m (n) and V
301
+ +
302
+ m (n) be the subspaces of �V with basis given by vi’s, with i lying
303
+ in �J+(n), J+
304
+ m(n) and J
305
+ +
306
+ m(n), respectively. Let �Ga
307
+ n denote the Lie subalgebra of gl(�V ) with
308
+ basis {Ei,j | i, j ∈ �J+(n)}. We denote �Ga := �Ga
309
+ ∞. Let Ga[m]n and G
310
+ a[m]n denote the Lie
311
+ subalgebras of �Ga with bases {Ei,j | i, j ∈ J+
312
+ m(n)} and {Ei,j | i, j ∈ J
313
+ +
314
+ m(n)}, respectively.
315
+ The Lie (super)algebras Ga[m]n and G
316
+ a[m]n are isomorphic to gl(m|n). Let �ba := �Ga ∩ �b,
317
+ ba[m]n := Ga[m]n ∩ b[m] and b
318
+ a[m]n := G
319
+ a[m]n ∩ b[m] stand for the standard Borel
320
+ subalgebras of �Ga, Ga[m]n and G
321
+ a[m]n, respectively. The corresponding Cartan subalgebras
322
+ �ha, ha[m]n and h
323
+ a[m]n have bases {Ea
324
+ i := Ei,i | i ∈ 1
325
+ 2N}, {Ea
326
+ i | i ∈ J+
327
+ m(n)} and {Ea
328
+ i | i ∈
329
+ J
330
+ +
331
+ m(n)}, respectively. Let {ǫi} denote the dual bases of the Cartan subalgebras with the
332
+ corresponding indices.
333
+ 2.2
334
+ Ortho-symplectic superalgebra �Gc and its subalgebras
335
+ Define a non-degenerate skew-supersymmetric bilinear form (·|·) on �V × by
336
+ (vi|vj) = −(vj|vi) = sgn(i)δi,−j,
337
+ i, j ∈ Z∗;
338
+ (2.3)
339
+ (vr|vs) = (vs|vr) = δr,−s,
340
+ r, s ∈ 1
341
+ 2 + Z;
342
+ (2.4)
343
+ (vi|vr) = (vr|vi) = 0,
344
+ i ∈ Z∗, r ∈ 1
345
+ 2 + Z;
346
+ (2.5)
347
+ where sgn(i) = 1 if i > 0 and sgn(i) = −1 if i < 0. The bilinear form induces non-
348
+ degenerate bilinear forms on V ×
349
+ m (n) and V
350
+ ×
351
+ m (n).
352
+
353
+ 6
354
+ Cheong and Lam
355
+ Let �Gc
356
+ n (resp., Gc[m]n and G
357
+ c[m]n) be the subalgebra of the Lie superalgebra gl(�V ×(n))
358
+ (resp., gl(V ×
359
+ m (n)) and gl(V
360
+ ×
361
+ m (n))) which preserves the bilinear form (·|·). The Lie superal-
362
+ gebra �Gc := �Gc
363
+ ∞ is spanned by the following elements (i, j ∈ Z∗ and r, s ∈ 1
364
+ 2 + Z):
365
+ Ec
366
+ i,j := −Ec
367
+ −j,−i := Ei,j − E−j,−i,
368
+ ij > 0;
369
+ Ec
370
+ i,j := Ec
371
+ −j,−i := Ei,j + E−j,−i,
372
+ ij < 0;
373
+ Ec
374
+ r,s := −Ec
375
+ −s,−r := Er,s − E−s,−r;
376
+ Ec
377
+ i,r := Ec
378
+ −r,−i := Ei,r + E−r,−i,
379
+ i > 0;
380
+ Ec
381
+ i,r := −Ec
382
+ −r,−i := Ei,r − E−r,−i,
383
+ i < 0.
384
+ The subalgebras Gc[m]n and G
385
+ c[m]n of �Gc are spanned by Ec
386
+ i,j with i, j ∈ J×
387
+ m(n) and J
388
+ ×
389
+ m(n),
390
+ respectively. Note that Gc[m]n is isomorphic to osp(2m|2n) while G
391
+ c[m]n is isomorphic to
392
+ spo(2m|2n).
393
+ We let �bc := �Gc ∩ �b, bc[m]n := Gc[m]n ∩ b[m] and b
394
+ c[m]n := G
395
+ c[m]n ∩ b[m] stand for
396
+ the standard Borel subalgebras of �Gc, Gc[m]n and G
397
+ c[m]n, respectively. The corresponding
398
+ Cartan subalgebras �hc, hc[m]n and h
399
+ c[m]n have bases {Ec
400
+ i := Ei,i − E−i,−i | i ∈
401
+ 1
402
+ 2N},
403
+ {Ec
404
+ i | i ∈ J+
405
+ m(n)} and {Ec
406
+ i | i ∈ J
407
+ +
408
+ m(n)}, respectively. Let {ǫi} denote the dual bases of the
409
+ Cartan subalgebras with the corresponding indices.
410
+ 2.3
411
+ Ortho-symplectic superalgebras �Gd and its subalgebras
412
+ Define a non-degenerate supersymmetric bilinear form (·|·) on �V × by
413
+ (vi|vj) = (vj|vi) = δi,−j,
414
+ i, j ∈ Z∗;
415
+ (2.6)
416
+ (vr|vs) = −(vs|vr) = sgn(r)δr,−s,
417
+ r, s ∈ 1
418
+ 2 + Z;
419
+ (2.7)
420
+ (vi|vr) = (vr|vi) = 0,
421
+ i ∈ Z∗, r ∈ 1
422
+ 2 + Z.
423
+ (2.8)
424
+ The bilinear form induces non-degenerate bilinear forms on V ×
425
+ m (n) and V
426
+ ×
427
+ m (n).
428
+ Let �Gd
429
+ n (resp., Gd[m]n and G
430
+ d[m]n) be the subalgebra of the Lie superalgebra gl(�V ×(n))
431
+ (resp., gl(V ×
432
+ m (n)) and gl(V
433
+ ×
434
+ m (n))) which preserves the bilinear form (·|·). The Lie superal-
435
+ gebra �Gd := �Gd
436
+ ∞ is spanned by the following elements (i, j ∈ Z∗ and r, s ∈ 1
437
+ 2 + Z):
438
+ Ed
439
+ i,j := −Ed
440
+ −j,−i := Ei,j − E−j,−i;
441
+ Ed
442
+ r,s := −Ed
443
+ −s,−r := Er,s − E−s,−r,
444
+ rs > 0;
445
+ Ed
446
+ r,s := Ed
447
+ −s,−r := Er,s + E−s,−r,
448
+ rs < 0;
449
+ Ed
450
+ i,r := Ed
451
+ −r,−i := Ei,r + E−r,−i,
452
+ r > 0;
453
+ Ed
454
+ i,r := −Ed
455
+ −r,−i := Ei,r − E−r,−i,
456
+ r < 0.
457
+ The subalgebras Gd[m]n and G
458
+ d[m]n of �Gd are spanned by Ed
459
+ i,j with i, j ∈ J×
460
+ m(n) and J
461
+ ×
462
+ m(n),
463
+ respectively. Note that Gd[m]n is isomorphic to spo(2m|2n) while G
464
+ d[m]n is isomorphic to
465
+ osp(2m|2n).
466
+
467
+ Gaudin Hamiltonians on unitarizable modules
468
+ 7
469
+ We let �bd := �Gd ∩ �b, bd[m]n := Gd[m]n ∩ b[m] and b
470
+ d[m]n := G
471
+ d[m]n ∩ b[m] stand for
472
+ the standard Borel subalgebras of �Gd, Gd[m]n and G
473
+ d[m]n, respectively. The corresponding
474
+ Cartan subalgebras �hd, hd[m]n and h
475
+ d[m]n have bases {Ed
476
+ i := Ei,i − E−i,−i | i ∈ 1
477
+ 2N},
478
+ {Ed
479
+ i | i ∈ J+
480
+ m(n)} and {Ed
481
+ i | i ∈ J
482
+ +
483
+ m(n)}, respectively. Let {ǫi} denote the dual bases of the
484
+ Cartan subalgebras with the corresponding indices.
485
+ Define a linear automorphism ϕ of degree 1 on the superspace �V × by
486
+ ϕ(v±r) :=
487
+ ® v±(r− 1
488
+ 2 ),
489
+ if r ∈ N;
490
+ v±(r+ 1
491
+ 2 ),
492
+ if r ∈ 1
493
+ 2 + Z+.
494
+ (2.9)
495
+ The isomorphism ϕ induces an automorphism �ϕ on the Lie superalgebra gl(�V ×). Notice
496
+ that the supersymmetric bilinear form on �V × defined by (2.6), (2.7) and (2.8) is exactly
497
+ the bilinear form induced, via ϕ, by the skew-supersymmetric bilinear form on �V × defined
498
+ by (2.3), (2.4) and (2.5). The restriction of �ϕ to �Gc gives an isomorphism from �Gc to �Gd
499
+ and hence an isomorphism from Gc[m]n (resp., G
500
+ c[m]n) to G
501
+ d[m]n (resp., Gd[m]n). It is
502
+ clear that �ϕ preserves the corresponding Borel and Cartan subalgebras. The restrictions
503
+ of �ϕ are denoted by �ϕ as well. We summarize the results in the following lemma.
504
+ Lemma 2.1. There is an isomorphism �ϕ from �Gc to �Gd given by
505
+ �ϕ(Ec
506
+ r,s) =
507
+
508
+
509
+
510
+
511
+
512
+
513
+
514
+
515
+
516
+
517
+
518
+
519
+
520
+
521
+
522
+ Ed
523
+ r− 1
524
+ 2 ,s− 1
525
+ 2,
526
+ if r, s ∈ N ∪ (−(1
527
+ 2 + Z+));
528
+ Ed
529
+ r+ 1
530
+ 2 ,s+ 1
531
+ 2,
532
+ if r, s ∈ −N ∪ (1
533
+ 2 + Z+);
534
+ Ed
535
+ r− 1
536
+ 2 ,s+ 1
537
+ 2,
538
+ if r ∈ N ∪ (−(1
539
+ 2 + Z+)), s ∈ −N ∪ (1
540
+ 2 + Z+);
541
+ Ed
542
+ r+ 1
543
+ 2 ,s− 1
544
+ 2,
545
+ if r ∈ −N ∪ (1
546
+ 2 + Z+), s ∈ N ∪ (−(1
547
+ 2 + Z+)).
548
+ (2.10)
549
+ The restrictions of �ϕ to Gc[m]n and G
550
+ c[m]n give isomorphisms �ϕ : Gc[m]n −→ G
551
+ d[m]n and
552
+ �ϕ : G
553
+ c[m]n −→ Gd[m]n, respectively.
554
+ 2.4
555
+ Dynkin diagrams
556
+ Consider the free abelian group with basis {ǫi | i ∈ 1
557
+ 2N}. It is endowed with a symmetric
558
+ bilinear form (·, ·) defined by
559
+ (ǫr, ǫs) = (−1)2rδrs,
560
+ r, s ∈ 1
561
+ 2N.
562
+ The parity of ǫi is defined as follows: |ǫi| = 0 for i ∈ N and |ǫj| = 1 for j ∈ 1
563
+ 2 + Z+. Let
564
+ α× = ǫm − ǫ 1
565
+ 2 ,
566
+ αr = ǫr − ǫr+ 1
567
+ 2,
568
+ βr = ǫr − ǫr+1,
569
+ r ∈ 1
570
+ 2N.
571
+ For x = a, c, d, the Dynkin diagrams of the Lie superalgebras �Gx
572
+ n and G
573
+ x[m]n (where
574
+ m ∈ N) together with prescribed fundamental systems are listed below ([K, Section 2.5]).
575
+ In what follows, ⃝ and � denote an even simple root and an odd isotropic simple root,
576
+ respectively.
577
+
578
+ 8
579
+ Cheong and Lam
580
+
581
+
582
+
583
+
584
+ �Ga
585
+ n
586
+
587
+
588
+
589
+
590
+
591
+ · · ·
592
+ α1/2
593
+ α1
594
+ α3/2
595
+ αn−1
596
+ αn−1/2
597
+
598
+
599
+
600
+
601
+ �Gc
602
+ n
603
+
604
+
605
+
606
+
607
+
608
+
609
+ · · ·
610
+
611
+
612
+
613
+
614
+ α1/2
615
+ −ǫ1/2 − ǫ1
616
+ α1
617
+ α3/2
618
+ αn−1
619
+ αn−1/2
620
+
621
+
622
+
623
+
624
+ �Gd
625
+ n
626
+
627
+
628
+
629
+
630
+
631
+ · · ·
632
+ =⇒
633
+ −2ǫ1/2
634
+ α1/2
635
+ α1
636
+ αn−1
637
+ αn−1/2
638
+
639
+
640
+
641
+
642
+ G
643
+ a[m]n
644
+
645
+
646
+
647
+
648
+
649
+
650
+
651
+ · · ·
652
+ · · ·
653
+ β1
654
+ β2
655
+ β3
656
+ βm−1
657
+ α×
658
+ β1/2
659
+ βn−3/2
660
+
661
+
662
+
663
+
664
+ G
665
+ c[m]n
666
+
667
+
668
+
669
+
670
+
671
+
672
+
673
+ =⇒
674
+ · · ·
675
+ · · ·
676
+ −2ǫ1
677
+ β1
678
+ β2
679
+ βm−1
680
+ α×
681
+ β1/2
682
+ βn−3/2
683
+
684
+
685
+
686
+
687
+ G
688
+ d[m]n
689
+
690
+
691
+
692
+
693
+
694
+
695
+
696
+
697
+ ��
698
+ ❅❅
699
+ · · ·
700
+ · · ·
701
+ β1
702
+ −ǫ1−ǫ2
703
+ β2
704
+ β3
705
+ βm−1
706
+ α×
707
+ β1/2
708
+ βn−3/2
709
+ The Dynkin diagrams of the Lie algebras G
710
+ x[0]n are as follows.
711
+
712
+
713
+
714
+
715
+ G
716
+ a[0]n
717
+
718
+
719
+
720
+
721
+
722
+
723
+ · · ·
724
+ β1/2
725
+ β3/2
726
+ β5/2
727
+ βn−7/2 βn−5/2
728
+ βn−3/2
729
+
730
+
731
+
732
+
733
+ G
734
+ c[0]n
735
+
736
+
737
+
738
+
739
+
740
+
741
+
742
+
743
+
744
+
745
+
746
+ · · ·
747
+ β1/2
748
+ −ǫ1/2 − ǫ3/2
749
+ β3/2
750
+ β5/2
751
+ βn−7/2 βn−5/2
752
+ βn−3/2
753
+
754
+
755
+
756
+
757
+ G
758
+ d[0]n
759
+
760
+
761
+
762
+
763
+
764
+ ⃝=⇒
765
+ · · ·
766
+ −2ǫ1/2
767
+ β1/2
768
+ β3/2
769
+ βn−7/2 βn−5/2
770
+ βn−3/2
771
+ Let �Φx +
772
+ n
773
+ (resp., Φx[m]+
774
+ n and Φ
775
+ x[m]+
776
+ n ) denote the set of positive roots of �Gx
777
+ n (resp., Gx[m]n
778
+ and G
779
+ x[m]n).
780
+
781
+ Gaudin Hamiltonians on unitarizable modules
782
+ 9
783
+ 2.5
784
+ Central extensions
785
+ For x = a, c, d, consider the central extension ˜gx (resp., gx[m]n and gx[m]n) of �Gx (resp.,
786
+ Gx[m]n and G
787
+ x[m]n) by the one-dimensional center CK, which is inherited from the cen-
788
+ tral extension �gl(�V ) of gl(�V ) determined by the 2-cocycle (2.1). The restriction of the
789
+ isomorphism ι to �Gx ⊕ CK (resp., Gx[m]n ⊕ CK and G
790
+ x[m]n ⊕ CK) is an isomorphism
791
+ ι : �Gx ⊕ CK → ˜gx (resp. ι : Gx[m]n ⊕ CK → gx[m]n and ι : G
792
+ x[m]n ⊕ CK → gx[m]n) given
793
+ by
794
+ ι(A) = A + Str(JA)K.
795
+ (2.11)
796
+ Note that [J, A] = 0 for all A ∈ ˜ga. Thus the 2-cocycle is zero when it restricts to �Ga,
797
+ and hence
798
+ [A, B]˜ga = [A, B]�Ga,
799
+ for A, B ∈ �Ga,
800
+ where, for example, [·, ·]˜ga denotes the Lie bracket on ˜ga. For notational unity, we still
801
+ consider ˜ga, ga[m]n and ga[m]n.
802
+ Remark 2.2. Every ˜gx(resp., gx[m]n and gx[m]n)-module can be regarded as a �Gx(resp.,
803
+ Gx[m]n and G
804
+ x[m]n)-module through the isomorphism (2.11). These central extensions are
805
+ convenient and conceptual for the formulation of truncation functors and super duality
806
+ described in Section 2.6 (see [CLW1, Remark 3.3] for more explanations).
807
+ We let �bx := �bx⊕CK, bx[m]n := bx[m]n⊕CK and b
808
+ x[m]n := b
809
+ x[m]n⊕CK stand for the
810
+ standard Borel subalgebras of ˜gx, gx[m]n and gx[m]n, respectively, and let ˜hx, hx[m]n and
811
+ h
812
+ x[m]n denote the Cartan subalgebras of ˜gx, gx[m]n and gx[m]n spanned by bases {K, Ex
813
+ r}
814
+ with dual bases {Λ0, ǫr} in the restricted dual ˜hx ∗, hx[m]∗
815
+ n and h
816
+ x[m]∗
817
+ n, where r runs over
818
+ the index sets 1
819
+ 2N, J+
820
+ m(n) and J
821
+ +
822
+ m(n), respectively. Here Λ0 is the element of ˜hx ∗ (resp.,
823
+ hx[m]∗
824
+ n and h
825
+ x[m]∗
826
+ n) defined by
827
+ Λ0(K) = 1
828
+ and
829
+ Λ0(Ex
830
+ r) = 0
831
+ for all r ∈ 1
832
+ 2N (resp., J+
833
+ m(n) and J
834
+ +
835
+ m(n)).
836
+ It is easy to see that the automorphism �ϕ on gl(�V ×) induced from ϕ defined by (2.9)
837
+ can be extended to the central extension �gl(�V ×) of gl(�V ×). By Lemma 2.1, we have the
838
+ following lemma.
839
+ Lemma 2.3. The isomorphism �ϕ : �Gc −→ �Gd extends to an isomorphism from ˜gc to ˜gd,
840
+ which is also denoted by �ϕ and is given by (2.10) together with
841
+ �ϕ(K) = −K.
842
+ The restrictions of �ϕ to gc[m]n and gc[m]n give isomorphisms �ϕ : gc[m]n −→ gd[m]n and
843
+ �ϕ : gc[m]n −→ gd[m]n, respectively.
844
+ 2.6
845
+ Parabolic BGG categories and super duality
846
+ Let �lx, lx[m]n and l
847
+ x[m]n be Levi subalgebras of ˜gx, gx[m]n and gx[m]n defined by
848
+ �lx := ˜gx ∩�l,
849
+ lx[m]n := gx[m]n ∩�l,
850
+ and
851
+ l
852
+ x[m]n := gx[m]n ∩�l,
853
+
854
+ 10
855
+ Cheong and Lam
856
+ respectively, and let �px = ˜lx +�bx, px[m]n = lx[m]n + bx[m]n and px[m]n = l
857
+ x[m]n + b
858
+ x[m]n be
859
+ the corresponding parabolic subalgebras, where �l := �
860
+ rs>0,
861
+ r,s∈ 1
862
+ 2 Z∗ CEr,s ⊕ CK. Observe that
863
+ lx[m]n ∼= ga[m]n ∼= gl(m|n) ⊕ CK
864
+ and
865
+ l
866
+ x[m]n ∼= ga[m]n ∼= gl(m|n) ⊕ CK.
867
+ Note that for x = a, �la = ˜ga, la[m]n = ga[m]n and l
868
+ a[m]n = ga[m]n.
869
+ In the remainder of the paper, we shall drop the superscript x and the
870
+ symbol ∞ if there is no ambiguity. For example, we write �G, G[m]n and G[m]n for �Gx,
871
+ Gx[m]n and G
872
+ x[m]n, and �g, g[m]n and g[m]n for �gx, gx[m]n and gx[m]n, respectively, where
873
+ x denotes a fixed type among a, c, d. Also, we write g[m] and g[m] for gx[m]∞ and gx[m]∞,
874
+ respectively.
875
+ Given a partition µ = (µ1, µ2, . . .), we denote by ℓ(µ) the length of µ and by µ′ the
876
+ conjugate partition of µ. We also denote by θ(µ) the modified Frobenius coordinates of µ:
877
+ θ(µ) := (θ(µ)1/2, θ(µ)1, θ(µ)3/2, θ(µ)2, . . .),
878
+ where
879
+ θ(µ)i−1/2 := max{µ′
880
+ i − i + 1, 0},
881
+ θ(µ)i := max{µi − i, 0},
882
+ i ∈ N.
883
+ Given a partition λ = (λ1, λ2, . . .) and d ∈ C, we define
884
+ �λ :=
885
+
886
+ r∈ 1
887
+ 2N
888
+ θ(λ)rǫr + dΛ0 ∈ ˜h∗,
889
+ (2.12)
890
+ λ[m] :=
891
+ m
892
+
893
+ i=1
894
+ λ′
895
+ iǫi− 1
896
+ 2 +
897
+
898
+ j∈N
899
+ ⟨λj − m⟩ ǫj + dΛ0 ∈ h[m]∗,
900
+ (2.13)
901
+ λ[m] :=
902
+ m
903
+
904
+ i=1
905
+ λiǫi +
906
+
907
+ j∈N
908
+
909
+ λ′
910
+ j − m
911
+
912
+ ǫj− 1
913
+ 2 + dΛ0 ∈ h[m]∗.
914
+ (2.14)
915
+ Here ⟨r⟩ := max{r, 0}.
916
+ Let �P+(d) ⊂ ˜h∗, P[m]+(d) ⊂ h[m]∗ and P[m]+(d) ⊂ h[m]∗ denote the sets of all
917
+ weights of the forms (2.12), (2.13) and (2.14), respectively. Let �P+ = ∪d∈C�P+(d), P[m]+ =
918
+ ∪d∈CP[m]+(d) and P[m]+ = ∪d∈CP[m]+(d). By definition we have bijective maps
919
+ �P+ −→ P[m]+
920
+ �λ �→ λ[m]
921
+ and
922
+ �P+ −→ P[m]+
923
+ �λ �→ λ[m]
924
+ (2.15)
925
+ Recall that a partition λ = (λ1, λ2, . . .) is called an (m|n)-hook partition if λm+1 ≤ n.
926
+ Let P and P(m|n) denote the set of partitions and the set of (m|n)-hook partitions,
927
+ respectively. Note that P(m|∞) = P. Given d ∈ C and λ ∈ P with λ′ ∈ P(m|n) (resp.,
928
+ λ ∈ P(m|n)), we may regard λ[m] ∈ h[m]∗
929
+ n (resp., λ[m] ∈ h[m]∗
930
+ n) in a natural way. The
931
+ subsets of all such weights will be denoted respectively by P[m]+
932
+ n and P[m]+
933
+ n .
934
+ For µ ∈ �h∗, let L(˜l, µ) be the irreducible highest weight ˜l-module with highest weight
935
+ µ. We denote by ∆(˜g, µ) = Ind�g
936
+ �pL(˜l, µ) the parabolic Verma �g-module and by L(˜g, µ) the
937
+ unique irreducible quotient �g-module of ∆(˜g, µ). The modules L(l[m]n, µ) and ∆(g[m]n, µ)
938
+
939
+ Gaudin Hamiltonians on unitarizable modules
940
+ 11
941
+ (for µ ∈ h[m]∗
942
+ n) as well as L(l[m]n, µ) and ∆(g[m]n, µ) (for µ ∈ h[m]∗
943
+ n) are defined anal-
944
+ ogously. We denote by L(g[m]n, µ) (resp., L(g[m]n, µ)) the unique irreducible quotient
945
+ g[m]n-module (resp., g[m]n-module) of ∆(g[m]n, µ) (resp., ∆(g[m]n, µ)). Note that for
946
+ x = a, ∆(˜ga, µ) = L(˜ga, µ) = L(˜la, µ), ∆(ga[m], µ) = L(ga[m], µ) = L(la[m], µ) and
947
+ ∆(ga[m], µ) = L(ga[m], µ) = L(l
948
+ a[m], µ).
949
+ Similar to [CL2, CLW1, CLW2], let �O (resp., O[m]n and O[m]n) be the category of
950
+ ˜g(resp. g[m]n and g[m]n)-modules M such that M is a semisimple �h(resp., h[m]n and
951
+ h[m]n)-module with finite-dimensional weight subspaces Mγ for γ ∈ �h∗ (resp., h[m]∗
952
+ n and
953
+ h[m]∗
954
+ n), satisfying
955
+ (i) M decomposes over ˜l (resp., l[m]n and l[m]n) as a direct sum of L(˜l, µ) (resp.,
956
+ L(l[m]n, µ) and L(l[m]n, µ)) for µ ∈ �P+ (resp., P[m]+
957
+ n and P[m]+
958
+ n ).
959
+ (ii) There exist finitely many weights λ1, λ2, . . . , λk ∈ �P+ (resp., P[m]+
960
+ n and P[m]+
961
+ n )
962
+ (depending on M) such that if γ is a weight in M, then λi −γ is a linear combination
963
+ of simple roots with coefficients in Z+ for some i.
964
+ The morphisms in the categories are even homomorphisms of modules, and the categories
965
+ are abelian. There is a natural Z2-gradation on each module in the categories with com-
966
+ patible action of the corresponding Lie (super)algebra to be defined below. Set
967
+ �Ξ =
968
+
969
+ r∈ 1
970
+ 2 N
971
+ Z+ǫr + CΛ0,
972
+ Ξ[m]n =
973
+
974
+ r∈J+
975
+ m(n)
976
+ Z+ǫr + CΛ0,
977
+ (2.16)
978
+ Ξ[m]n =
979
+
980
+ r∈J
981
+ +
982
+ m(n)
983
+ Z+ǫr + CΛ0.
984
+ For ε = 0 or 1 and Θ = �Ξ, Ξ[m]n or Ξ[m]n, we define
985
+ Θ(ε) :=
986
+
987
+ µ ∈ Θ
988
+ ���
989
+
990
+ r∈ 1
991
+ 2 +Z+
992
+ µ(Er) ≡ ε (mod 2)
993
+
994
+ ,
995
+ where the summation is over all r ∈ 1
996
+ 2 + Z+ whenever µ(Er) are defined. Recall that
997
+ both l[m]n and l[m]n are isomorphic to gl(m|n) ⊕ CK. For M ∈ O[m]n (resp., O[m]n),
998
+ each weight of M is a weight of a highest weight module over l[m]n (resp., l[m]n) with
999
+ highest weight µ ∈ P[m]+
1000
+ n (resp., P[m]+
1001
+ n ) which is contained in Ξ[m]n (resp., Ξ[m]n) (see,
1002
+ for example, [CW, Proposition 3.26]). By the paragraph before Theorem 6.4 in [CW], the
1003
+ weights of M are contained in �Ξ for M ∈ �O. For M ∈ �O, M = M0
1004
+ � M1 is a Z2-graded
1005
+ vector space such that
1006
+ M0 :=
1007
+
1008
+ µ∈�Ξ(0)
1009
+
1010
+ and
1011
+ M1 :=
1012
+
1013
+ µ∈�Ξ(1)
1014
+ Mµ.
1015
+ (2.17)
1016
+ It is clear that the Z2-gradation on M is compatible with the action of ˜g. Similarly, we may
1017
+ define a Z2-gradation with compatible action of g[m]n and g[m]n on M for M ∈ O[m]n and
1018
+
1019
+ 12
1020
+ Cheong and Lam
1021
+ O[m]n, respectively. By [CW, Theorem 3.27 and Theorem 6.4] (see also the proof of [Lus,
1022
+ Theorem 6.2.2]), �O, O[m]n and O[m]n are tensor categories. In particular, �Oa, Oa[m]n and
1023
+ O
1024
+ a[m]n are semisimple tensor categories. Note that the Z2-gradation on M ⊗ N given by
1025
+ (2.17) and the Z2-gradation on M ⊗N induced from the Z2-gradations on M and N given
1026
+ by (2.17) are the same for M, N ∈ �O (resp., O[m]n and O[m]n). We summarize the results
1027
+ in the following proposition.
1028
+ Proposition 2.4. Let n ∈ N ∪ {∞}. Then:
1029
+ (i) The weights of modules in �O (resp., O[m]n and O[m]n) are contained in �Ξ (resp.,
1030
+ Ξ[m]n and Ξ[m]n).
1031
+ (ii) The abelian categories �O, O[m]n and O[m]n are tensor categories.
1032
+ (iii) �Oa, Oa[m]n and O
1033
+ a[m]n are semisimple tensor categories.
1034
+ We also have the following proposition.
1035
+ Proposition 2.5. Let n ∈ N ∪ {∞}. Then:
1036
+ (i) The modules ∆(˜g, λ) and L(˜g, λ) lie in �O for all λ ∈ �P+.
1037
+ (ii) The modules ∆(g[m]n, λ) and L(g[m]n, λ) lie in O[m]n for all λ ∈ P+[m]n.
1038
+ (iii) The modules ∆(g[m]n, λ) and L(g[m]n, λ) lie in O[m]n for all λ ∈ P
1039
+ +[m]n.
1040
+ Proof. (i) follows from [CW, Proposition 6.7(3)]. We show (ii). For m = 0, it follows from
1041
+ [CW, Proposition 6.7(1)]. For m ̸= 0, the proof is similar to that of [CW, Proposition
1042
+ 6.7(3)], but here we use [CW, Theorem 3.27] (valid also for n = ∞) instead of [CW,
1043
+ Theorem 6.4]. The proof of (iii) is similar.
1044
+ We record here an easy but useful observation, which can be seen by the description
1045
+ of the weights of modules in (2.16) and is essentially [CaL, Lemma 2.3].
1046
+ Lemma 2.6. Let M, N ∈ �O (resp., O[m]n and O[m]n). Suppose that µ and γ are weights
1047
+ of M and N, respectively. Then
1048
+ (µ + γ)(Er) = 0
1049
+ if and only if
1050
+ µ(Er) = 0 and γ(Er) = 0,
1051
+ for r ∈ 1
1052
+ 2N (resp., J+
1053
+ m(n) and J
1054
+ +
1055
+ m(n)).
1056
+ Similarly, in view of (2.16), we immediately obtain the following (cf. [CaL, Lemma
1057
+ 2.5]).
1058
+ Lemma 2.7. Let µ, γ ∈ �Ξ. Then:
1059
+ (i) µ + γ ∈ Ξ[m] if and only if µ ∈ Ξ[m] and γ ∈ Ξ[m].
1060
+ (ii) µ + γ ∈ Ξ[m] if and only if µ ∈ Ξ[m] and γ ∈ Ξ[m].
1061
+
1062
+ Gaudin Hamiltonians on unitarizable modules
1063
+ 13
1064
+ For 0 ≤ k < n ≤ ∞, the truncation functor trn
1065
+ k : O[m]n −→ O[m]k is defined by
1066
+ trn
1067
+ k(M) =
1068
+
1069
+ ν∈Ξ[m]k
1070
+
1071
+ for M ∈ O[m]n.
1072
+ (2.18)
1073
+ For every f ∈ HomO[m]n(M, N), trn
1074
+ k(f) is defined to be the restriction of f to trn
1075
+ k(M). The
1076
+ truncation functor trn
1077
+ k : O[m]n −→ O[m]k can be defined in a similar way. It is clear that
1078
+ trn
1079
+ k and trn
1080
+ k are exact functors. By Lemma 2.6, we immediately have the following lemma.
1081
+ Lemma 2.8. Let 0 ≤ k < n ≤ ∞. trn
1082
+ k and trn
1083
+ k are tensor functors.
1084
+ The following proposition will be useful later on. It may be proven in a similar way to
1085
+ the proof of [CLW1, Lemma 3.2].
1086
+ Proposition 2.9. Let 0 ≤ k < n ≤ ∞ and µ ∈ P[m]+
1087
+ n . Suppose Vi = ∆(g[m]i, µ) or
1088
+ L(g[m]i, µ) for i = k, n. We have
1089
+ trn
1090
+ k(Vn) =
1091
+ ®
1092
+ Vk
1093
+ if µ ∈ P[m]+
1094
+ k ,
1095
+ 0
1096
+ otherwise.
1097
+ Similar statement holds for trn
1098
+ k.
1099
+ Given M = �
1100
+ γ∈˜h∗ Mγ ∈ �O, we define
1101
+ T[m](M) =
1102
+
1103
+ γ∈h[m]∗
1104
+
1105
+ and
1106
+ T [m](M) =
1107
+
1108
+ γ∈h[m]∗
1109
+ Mγ.
1110
+ For M, N ∈ �O and f ∈ Hom�O(M, N), T[m](f) and T [m](f) are defined to be the
1111
+ restrictions of f to T[m](M) and T [m](M), respectively. Note that T[m](f) : T[m](M) →
1112
+ T[m](N) and T [m](f) : T [m](M) → T [m](N) are respectively a g[m]-homomorphism and a
1113
+ g[m]-homomorphism. Moreover, the functors T[m] : �O → O[m] and T [m] : �O → O[m] are
1114
+ exact (cf. [CW, Proposition 6.15]).
1115
+ By Lemma 2.7, we have T[m](M ⊗ N) = T[m](M) ⊗ T[m](N) and T [m](M ⊗ N) =
1116
+ T [m](M) ⊗ T [m](N) for all M, N ∈ �O, and so T[m] and T [m] are tensor functors. We have
1117
+ the following result.
1118
+ Theorem 2.10. The following statements hold:
1119
+ (i) For each m ∈ Z+, T[m] : �O → O[m] is an equivalence of tensor categories.
1120
+ (ii) For each m ∈ Z+, T [m] : �O → O[m] is an equivalence of tensor categories.
1121
+ Moreover, T[m] and T [m] send parabolic Verma modules to parabolic Verma modules and
1122
+ irreducible modules to irreducible modules. More precisely, for λ ∈ �P+, we have
1123
+ T[m]
1124
+
1125
+ ∆(˜g, �λ)) = ∆(g[m], λ[m]),
1126
+ T[m]
1127
+
1128
+ L(˜g, �λ)
1129
+
1130
+ = L(g[m], λ[m]),
1131
+ T [m]
1132
+
1133
+ ∆(˜g, �λ)) = ∆(g[m], λ[m]),
1134
+ T [m]
1135
+
1136
+ L(˜g, �λ)
1137
+
1138
+ = L(g[m], λ[m]).
1139
+
1140
+ 14
1141
+ Cheong and Lam
1142
+ Theorem 2.10 can be proven along the lines of the proof of the super duality in [CL2,
1143
+ CLW1] by making use of the techniques in [CLW2] (see particularly [CLW2, Section 7]).
1144
+ The proof is omitted here. We also call Theorem 2.10 super duality. We only need that
1145
+ T[m] and T [m] are tensor functors in this paper. Note that T[0] and T [0] are the functors T
1146
+ and T in the degenerate case defined in [CL2, CLW1]. Theorem 2.10 also implies that the
1147
+ tensor categories O[0] and O[0] for Lie algebras and the tensor categories O[m] and O[m]
1148
+ for Lie superalgebras are equivalent for m ∈ N.
1149
+ 3
1150
+ Unitarizable G
1151
+ x[m]n-modules
1152
+ The notion of unitarizable modules will play an important role in our study of (super)
1153
+ Gaudin Hamiltonians. In this section, we start by introducing ∗-structures on G
1154
+ x[m]n and
1155
+ gx[m]n. We then describe the unitarizable G
1156
+ x[m]n-modules to be studied in this paper.
1157
+ We first recall some basic facts about ∗-superalgebras and their unitarizable represen-
1158
+ tations. A ∗-superalgebra is an associative superalgebra A together with an anti-linear
1159
+ anti-involution ω : A −→ A of degree 0. A homomorphism f : (A, ω) → (A′, ω′) of ∗-
1160
+ superalgebras is a homomorphism of superalgebras satisfying ω′ ◦f = f ◦ω. Let (A, ω) be
1161
+ a ∗-superalgebra, and let V be a Z2-graded A-module. A Hermitian form ⟨·|·⟩ on V is said
1162
+ to be contravariant if ⟨av|v′⟩ = ⟨v|ω(a)v′⟩, for all a ∈ A, v, v′ ∈ V . An A-module equipped
1163
+ with a positive definite contravariant Hermitian form is called a unitarizable A-module.
1164
+ A Lie superalgebra g is said to admit a ∗-structure if g is equipped with an anti-linear
1165
+ anti-involution ω of degree 0. In this case, ω is also called a ∗-structure on g. A homo-
1166
+ morphism f : (g, ω) → (g′, ω′) of Lie superalgebras with ∗-structures is a homomorphism
1167
+ of Lie superalgebras satisfying ω′ ◦ f = f ◦ ω. Moreover, it is clear that ω is a ∗-structure
1168
+ on g if and only if the natural extension of ω to the universal enveloping algebra U(g) of
1169
+ g is an anti-linear anti-involution. Let (g, ω) be a Lie superalgebra with ∗-structure, and
1170
+ let V be a Z2-graded g-module. A Hermitian form ⟨·|·⟩ on V is said to be contravariant if
1171
+ ⟨xv|v′⟩ = ⟨v|ω(x)v′⟩, for all x ∈ g, v, v′ ∈ V . A g-module equipped with a positive definite
1172
+ contravariant Hermitian form is called a unitarizable g-module. Notice that a g-module
1173
+ V is a unitarizable g-module if and only if V is a unitarizable U(g)-module.
1174
+ 3.1
1175
+ ∗-structures on G
1176
+ x[m]n and gx[m]n
1177
+ Recall that the Lie superalgebra �gl(�V ) is the central extension of gl(�V ) with a basis
1178
+ {Ei,j, K | i, j ∈ 1
1179
+ 2Z}. It admits a ∗-structure ω defined by (cf. [LZ1, p. 421])
1180
+
1181
+ i,j∈ 1
1182
+ 2Z
1183
+ aijEi,j �→
1184
+
1185
+ i,j∈ 1
1186
+ 2Z
1187
+ (−1)[i]+[j]aijEj,i
1188
+ and
1189
+ K �→ K.
1190
+ Here aij denotes the complex conjugate of aij ∈ C and
1191
+ [i] :=
1192
+ ®
1193
+ 1
1194
+ if
1195
+ − i ∈ 1
1196
+ 2 + Z+;
1197
+ 0
1198
+ if
1199
+ − i ∈ 1
1200
+ 2Z\(1
1201
+ 2 + Z+).
1202
+
1203
+ Gaudin Hamiltonians on unitarizable modules
1204
+ 15
1205
+ It is clear from the spanning sets (i.e., the sets of elements described in Sections 2.1, 2.2 and
1206
+ 2.3 together with K) of the Lie superalgebras ˜gx, gx[m]n and gx[m]n that the restrictions
1207
+ of ω to these Lie superalgebras, denoted also by ω, give ∗-structures on them.
1208
+ Since ω is a ∗-structure on �gl(�V ×) and �ϕ is an involution of �gl(�V ×), the map ω′ :=
1209
+ �ϕ ◦ ω ◦ �ϕ is a ∗-structure on �gl(�V ×). More precisely,
1210
+ ω′(Er,s) = (−1)τr+τsEs,r,
1211
+ for r, s ∈ 1
1212
+ 2Z∗,
1213
+ and
1214
+ ω′(K) = K,
1215
+ where
1216
+ τr :=
1217
+ ®
1218
+ 1
1219
+ if
1220
+ − r ∈ N;
1221
+ 0
1222
+ if
1223
+ − r ∈ 1
1224
+ 2Z∗\N.
1225
+ Via the isomorphism �ϕ : gc[m]n −→ gd[m]n given in Lemma 2.3, an anti-linear anti-
1226
+ involution ω on gd[m]n pulls back to an anti-linear anti-involution ω′ := �ϕ−1 ◦ ω ◦ �ϕ on
1227
+ gc[m]n while, via the isomorphism �ϕ−1 : gd[m]n −→ gc[m]n, an anti-linear anti-involution
1228
+ ω on gc[m]n pulls back to an anti-linear anti-involution ω′ := �ϕ ◦ ω ◦ �ϕ−1 on gd[m]n. In
1229
+ the other words, the map �ϕ (resp., �ϕ−1) gives an isomorphism of Lie superalgebras with
1230
+ ∗-structures from (gc[m]n, ω′) (resp., (gd[m]n, ω′)) to (gd[m]n, ω) (resp., (gc[m]n, ω)). Note
1231
+ that the ∗-structure ω′ on gc[m]n (resp., gd[m]n) is the restriction of ω′ defined on �gl(�V ×).
1232
+ Setting K = 0, the ∗-structure ω (resp., ω′) induces a ∗-structure, denoted also by
1233
+ ω (resp., ω′), on gl(�V ×). For x = c, d, the restriction of ω (resp., ω′) to G
1234
+ x[m]n gives a
1235
+ ∗-structure on G
1236
+ x[m]n, denoted also by ω (resp., ω′). We have the following proposition.
1237
+ Proposition 3.1. For x = c, d, the restriction of the isomorphism ι : G
1238
+ x[m]n ⊕ CK −→
1239
+ gx[m]n defined by (2.11) to G
1240
+ x[m]n give two monomorphisms of Lie superalgebras with
1241
+ ∗-structures ι : (G
1242
+ x[m]n, ω) −→ (gx[m]n, ω) and ι′ : (G
1243
+ x[m]n, ω′) −→ (gx[m]n, ω′).
1244
+ 3.2
1245
+ Unitarizable G
1246
+ a[m]n-modules
1247
+ Observe that the 2-cocycle (2.1) is zero when it restricts to G
1248
+ a[m]n. The natural inclusion
1249
+ allows us to identify G
1250
+ a[m]n as a subalgebra of ga[m]n, and hence the restriction of ω on
1251
+ ga[m]n to G
1252
+ a[m]n is a ∗-structure on G
1253
+ a[m]n, which we also denote by ω. More precisely,
1254
+ ω(Ei,j) = Ej,i
1255
+ for
1256
+ i, j ∈ J
1257
+ +
1258
+ m(n).
1259
+ Recall that P(m|n) denotes the set of (m|n)-hook partitions. For λ ∈ P(m|n), we define
1260
+ λ =
1261
+ m
1262
+
1263
+ i=1
1264
+ λiǫi +
1265
+ n
1266
+
1267
+ j=1
1268
+ ⟨λ′
1269
+ j − m⟩ǫj− 1
1270
+ 2 ∈ h
1271
+ a[m]n.
1272
+ (3.1)
1273
+ Let Q
1274
+ a,I(m|n) denote the set of weights of the forms (3.1). The following proposition is
1275
+ well known. (See, for example [CLZ, Theorems 3.2 and 3.3] with p = q = 0.)
1276
+ Proposition 3.2. For λ ∈ Q
1277
+ a,I(m|n), L(G
1278
+ a[m]n, λ) is a unitarizable G
1279
+ a[m]n-module with
1280
+ respect to the ∗-structure ω.
1281
+ Remark 3.3. The modules appearing in the proposition above are exactly the irreducible
1282
+ highest weight polynomial modules over G
1283
+ a[m]n (see, for example, [CW, Proposition 3.26]).
1284
+
1285
+ 16
1286
+ Cheong and Lam
1287
+ Recall λ[m] and λ[m] defined in (2.13) and (2.14), respectively. Let
1288
+ Qa(m|n) := {λ[m] ∈ ha[m]∗
1289
+ n | λ′ ∈ P(m|n), d = 0},
1290
+ Q
1291
+ a(m|n) := {λ[m] ∈ h
1292
+ a[m]∗
1293
+ n | λ ∈ P(m|n), d = 0}.
1294
+ These sets will be used in Section 5.
1295
+ 3.3
1296
+ Unitarizable modules over G
1297
+ c[m]n and G
1298
+ d[m]n
1299
+ In this subsection, we will restrict our attention to x = c, d.
1300
+ There are two types of
1301
+ unitarizable highest weight modules over G
1302
+ x[m]n corresponding to the ∗-structures ω and
1303
+ ω′ defined above.
1304
+ Note that the Lie superalgebra Cf (resp., Df) defined in [LZ1] is our �Gc (resp., �Gd)
1305
+ while �Cf (resp., �Df) is our ˜gc (resp., ˜gd). Also, the set of the unitarizable quasi-finite irre-
1306
+ ducible highest weight modules over �C (resp., �D) described in [Proposition 5.8][LZ1](resp.,
1307
+ [Proposition 5.9][LZ1]) are the set of unitarizable irreducible highest weight modules over
1308
+ �Cf (resp., �Df). Recall that �λ is defined in (2.12). Let
1309
+ �Qc :=
1310
+ ¶ �λ ∈ �hc ∗ ��� λ1 ≤ d, λ ∈ P, d ∈ Z+
1311
+ ©
1312
+ ,
1313
+ �Qd :=
1314
+
1315
+ �λ ∈ �hd ∗ ��� λ1 + λ2 ≤ 2d, λ ∈ P, d ∈ 1
1316
+ 2Z+
1317
+
1318
+ .
1319
+ Reformulating the results in [LZ1] in terms of our notations, we have the following propo-
1320
+ sition.
1321
+ Proposition 3.4.
1322
+ (i) An irreducible highest weight ˜gc-module M is unitarizable with
1323
+ respect to ω if and only if M ∼= L(˜gc, ξ) for some ξ ∈ �Qc.
1324
+ (ii) An irreducible highest weight ˜gd-module M is unitarizable with respect to ω if and
1325
+ only if M ∼= L(˜gd, ξ) for some ξ ∈ �Qd.
1326
+ Recall that λ[m] and λ[m] are defined in (2.13) and (2.14), respectively. Let
1327
+ Qc(m|n) :=
1328
+
1329
+ λ[m] ∈ hc[m]∗
1330
+ n
1331
+ �� λ1 ≤ d, λ′ ∈ P(m|n), d ∈ Z+
1332
+
1333
+ ,
1334
+ Qd(m|n) :=
1335
+
1336
+ λ[m] ∈ hd[m]∗
1337
+ n
1338
+ ��� λ1 + λ2 ≤ 2d, λ′ ∈ P(m|n), d ∈ 1
1339
+ 2Z+
1340
+
1341
+ ,
1342
+ Q
1343
+ c(m|n) :=
1344
+
1345
+ λ[m] ∈ h
1346
+ c[m]∗
1347
+ n
1348
+ ��� λ1 ≤ d, λ ∈ P(m|n), d ∈ Z+
1349
+
1350
+ ,
1351
+ Q
1352
+ d(m|n) :=
1353
+
1354
+ λ[m] ∈ h
1355
+ d[m]∗
1356
+ n
1357
+ ��� λ1 + λ2 ≤ 2d, λ ∈ P(m|n), d ∈ 1
1358
+ 2Z+
1359
+
1360
+ .
1361
+ The proof of the following is straightforward.
1362
+ Lemma 3.5. Let g be a Lie superalgebra with ∗-structure σ. Assume that u is a subalgebra
1363
+ of g such that the restriction σ|u of σ to u is a ∗-structure on u. Let V be a unitarizable
1364
+ g-module with respect to σ. If W is a u-submodule of V , then W is a unitarizable u-module
1365
+ with respect to σ|u.
1366
+
1367
+ Gaudin Hamiltonians on unitarizable modules
1368
+ 17
1369
+ The following proposition is a direct consequence of Proposition 3.4 and Lemma 3.5.
1370
+ Proposition 3.6.
1371
+ (i) For ξ ∈ Qc(m|n), L(gc[m]n, ξ) is a unitarizable gc[m]n-module
1372
+ with respect to ω.
1373
+ (ii) For ξ ∈ Qd(m|n), L(gd[m]n, ξ) is a unitarizable gd[m]n-module with respect to ω.
1374
+ (iii) For ξ ∈ Q
1375
+ c(m|n), L(gc[m]n, ξ) is a unitarizable gc[m]n-module with respect to ω.
1376
+ (iv) For ξ ∈ Q
1377
+ d(m|n), L(gd[m]n, ξ) is a unitarizable gd[m]n-module with respect to ω.
1378
+ Proof. To show (iii), let λ ∈ P(m|n) and d ∈ Z+ be such that λ1 ≤ d. Then λ[m] ∈
1379
+ Q
1380
+ c(m|n) and L(gc[m]n, λ[m]) = tr∞
1381
+ n (T [m](L(�gc, �λ))). By Lemma 3.5, L(gc[m]n, λ[m]) is
1382
+ a unitarizable module with respect to ω. The other parts can be proven by a similar
1383
+ argument.
1384
+ Definition 3.7. For x = c, d, a G
1385
+ x[m]n-module M is said to be a unitarizable module of
1386
+ type I (resp., II) if M is unitarizable with respect to the ∗-structure ω (resp., ω′).
1387
+ Let
1388
+ 1m|n =
1389
+ m
1390
+
1391
+ i=1
1392
+ ǫi −
1393
+ n
1394
+
1395
+ j=1
1396
+ ǫj− 1
1397
+ 2.
1398
+ Let
1399
+ Q
1400
+ c,I(m|n) :=
1401
+
1402
+ λ − d1m|n ∈ h
1403
+ c[m]∗
1404
+ n
1405
+ ��� λ1 ≤ d, λ ∈ P(m|n), d ∈ Z+
1406
+
1407
+ ,
1408
+ Q
1409
+ d,I(m|n) :=
1410
+
1411
+ λ − d1m|n ∈ h
1412
+ d[m]∗
1413
+ n
1414
+ ��� λ1 + λ2 ≤ 2d, λ ∈ P(m|n), d ∈ 1
1415
+ 2Z+
1416
+
1417
+ ,
1418
+ Q
1419
+ c,II(m|n) :=
1420
+
1421
+ λ + d1m|n ∈ h
1422
+ c[m]∗
1423
+ n
1424
+ ��� λ′
1425
+ 1 + λ′
1426
+ 2 ≤ 2d, λ ∈ P(m|n), d ∈ 1
1427
+ 2Z+
1428
+
1429
+ ,
1430
+ Q
1431
+ d,II(m|n) :=
1432
+
1433
+ λ + d1m|n ∈ h
1434
+ d[m]∗
1435
+ n
1436
+ ��� λ′
1437
+ 1 ≤ d, λ ∈ P(m|n), d ∈ Z+
1438
+
1439
+ .
1440
+ Proposition 3.8.
1441
+ (i) For ξ ∈ Q
1442
+ c,I(m|n), L(G
1443
+ c[m]n, ξ) is a unitarizable G
1444
+ c[m]n-module
1445
+ of type I.
1446
+ (ii) For ξ ∈ Q
1447
+ c,II(m|n), L(G
1448
+ c[m]n, ξ) is a unitarizable G
1449
+ c[m]n-module of type II.
1450
+ (iii) For ξ ∈ Q
1451
+ d,I(m|n), L(G
1452
+ d[m]n, ξ) is a unitarizable G
1453
+ d[m]n-module of type I.
1454
+ (iv) For ξ ∈ Q
1455
+ d,II(m|n), L(G
1456
+ d[m]n, ξ) is a unitarizable G
1457
+ d[m]n-module of type II.
1458
+ Proof. To show (i), let λ ∈ P(m|n) and d ∈ Z+ be such that λ1 ≤ d. Note that the G
1459
+ c[m]n-
1460
+ module structure, induced from the monomorphism of Lie superalgebras with ∗-structures
1461
+ ι : (G
1462
+ c[m]n, ω) −→ (gc[m]n, ω) given in Proposition 3.1, on L(gc[m]n, λ[m]) is isomorphic to
1463
+ L(G
1464
+ c[m]n, λ−d1m|n). By Proposition 3.6(iii), part (i) follows. To show (ii), let λ ∈ P(m|n)
1465
+ and d ∈ 1
1466
+ 2Z+ be such that λ′
1467
+ 1 + λ′
1468
+ 2 ≤ 2d. Note that ι′ : (G
1469
+ c[m]n, ω′) −→ (gc[m]n, ω′) given
1470
+ in Proposition 3.1 is a monomorphism of Lie superalgebras with ∗-structures, and we have
1471
+ shown that the map �ϕ given in Lemma 2.3 is an isomorphism of Lie superalgebras with
1472
+
1473
+ 18
1474
+ Cheong and Lam
1475
+ ∗-structures from (gc[m]n, ω′) to (gd[m]n, ω). The G
1476
+ c[m]n-module structure induced from
1477
+ �ϕ ◦ ι′ on L(gd[m]n, λ′[m]) is isomorphic to L(G
1478
+ c[m]n, λ + d1m|n). By Proposition 3.6(ii),
1479
+ part (ii) follows. The other parts can be proven by a similar argument.
1480
+ Remark 3.9. For m = 0 and n ∈ N, G
1481
+ c[m]n and G
1482
+ d[m]n are Lie algebras and the highest
1483
+ weights in (ii) (resp., (iv)) of the proposition above are exactly the highest weights ap-
1484
+ pearing in the classification of infinite-dimensional unitarizable irreducible highest weight
1485
+ modules over G
1486
+ c[m]n (resp., G
1487
+ d[m]n) with integral (resp., half integral and integral) values
1488
+ given in [EHW, Sections 8 and 9] (see also [HLT, Theorem 2.5]).
1489
+ 4
1490
+ Gaudin Hamiltonians on modules over �g, g[m]n and g[m]n
1491
+ In this section, we define the Casimir symmetric tensors for the Lie (super)algebras �g, g[m]n
1492
+ and g[m]n of infinite and finite ranks, and then introduce the (super) Gaudin Hamilto-
1493
+ nians associated to these Lie (super)algebras. Our main goal is to show that the set of
1494
+ eigenvectors of each Gaudin Hamiltonian for ˜g is in one-to-one correspondence with the set
1495
+ of eigenvectors of the corresponding Gaudin Hamiltonian for g[m] (resp., g[m]). Besides,
1496
+ each eigenvector and its corresponding eigenvector, under the one-to-one correspondence,
1497
+ has the same eigenvalue. We also show that the eigenvectors of the Gaudin Hamiltonians
1498
+ for g[m] (resp, g[m]) and those of g[m]n (resp, g[m]n), for n ∈ N, are related by truncation
1499
+ functors.
1500
+ First of all, we have the following lemma. It is analogous to [CaL, Lemma 3.1] and
1501
+ can be proven similarly.
1502
+ Lemma 4.1. Let 0 ≤ k < n ≤ ∞.
1503
+ If v is a weight vector of weight µ in M ∈ �O
1504
+ (resp., O[m]n and O[m]n) such that µ ∈ �Ξk (resp., Ξ[m]k and Ξ[m]k), then Eβv = 0 for
1505
+ all β ∈ �Φ+\�Φ+
1506
+ k (resp., Φ[m]+
1507
+ n \Φ[m]+
1508
+ k and Φ[m]+
1509
+ n \Φ[m]+
1510
+ k ) and Eiv = 0 for i > k. In
1511
+ particular, for each v ∈ M and M ∈ �O, O[m] or O[m], there are only finitely many Eβ
1512
+ and Ei such that Eβv ̸= 0 and Eiv ̸= 0.
1513
+ Let (·, ·) denote the bilinear form on gl(�V ) defined by
1514
+ (A, B) := Str(AB)
1515
+ for A, B ∈ gl(�V ).
1516
+ It is non-degenerate invariant even supersymmetric. The restriction of the above bilinear
1517
+ form to �G (resp., G[m]n and G[m]n) is also a non-degenerate invariant even supersymmetric
1518
+ bilinear form. We denote
1519
+ ⟨·, ·⟩ := (·, ·)
1520
+ on �Ga, Ga[m]n and G
1521
+ a[m]n,
1522
+ and
1523
+ ⟨·, ·⟩ := 1
1524
+ 2(·, ·)
1525
+ on other cases.
1526
+ It is clear that ⟨Ei, Ei⟩ = (−1)2i for any i ∈ 1
1527
+ 2N. Recall that �Φ+ (resp., Φ[m]+
1528
+ n and
1529
+ Φ[m]+
1530
+ n ) denote the set of positive roots of ˜g (resp., g[m]n and g[m]n). For each root β in
1531
+ �Φ+ (resp., Φ[m]+
1532
+ n and Φ[m]+
1533
+ n ), we choose root vectors Eβ and Eβ of weights β and −β,
1534
+ respectively, such that
1535
+ ⟨Eβ, Eβ⟩ = 1.
1536
+
1537
+ Gaudin Hamiltonians on unitarizable modules
1538
+ 19
1539
+ Clearly, ⟨Eβ, Eβ⟩ = (−1)|Eβ|, where |Eβ| is the parity of Eβ.
1540
+ By identifying �G (resp., Gn and Gn) as a subspace of the vector space of ˜g (resp., gn
1541
+ and gn), the Casimir symmetric tensors for ˜g, g[m]n and g[m]n are defined by (cf. [CaL,
1542
+ Section 3.1])
1543
+ �Ω :=
1544
+
1545
+ β∈�Φ+
1546
+ (Eβ ⊗ Eβ + (−1)|Eβ|Eβ ⊗ Eβ)
1547
+ +
1548
+
1549
+ j∈ 1
1550
+ 2N
1551
+
1552
+ (−1)2jEj ⊗ Ej − (K ⊗ Ej + Ej ⊗ K)
1553
+
1554
+ ,
1555
+ Ω[m]n :=
1556
+
1557
+ β∈Φ[m]+
1558
+ n
1559
+
1560
+ Eβ ⊗ Eβ + (−1)|Eβ|Eβ ⊗ Eβ)
1561
+ +
1562
+
1563
+ j∈J+
1564
+ m(n)
1565
+
1566
+ (−1)2jEj ⊗ Ej − (K ⊗ Ej + Ej ⊗ K)
1567
+
1568
+ ,
1569
+ Ω[m]n :=
1570
+
1571
+ β∈Φ[m]+
1572
+ n
1573
+ (Eβ ⊗ Eβ + (−1)|Eβ|Eβ ⊗ Eβ)
1574
+ +
1575
+
1576
+ j∈J
1577
+ +
1578
+ m(n)
1579
+
1580
+ (−1)2jEj ⊗ Ej − (K ⊗ Ej + Ej ⊗ K)
1581
+
1582
+ .
1583
+ Remark 4.2. For β ∈ Φ[m]+
1584
+ n , either β ∈ �Φ+ or −β ∈ �Φ+. Assume that −β ∈ �Φ+. Then
1585
+ β is an odd root. We readily see that E−β = aEβ and E−β = −a−1Eβ for some nonzero
1586
+ scalar a. It follows that E−β ⊗ E−β − E−β ⊗ E−β = Eβ ⊗ Eβ − Eβ ⊗ Eβ. In other words,
1587
+ Ω[m]n is a partial sum of �Ω. Similarly, Ω[m]n is a partial sum of �Ω as well.
1588
+ By Lemma 4.1, the Casimir symmetric tensors �Ω, Ω[m]n and Ω[m]n are well defined
1589
+ operators on M ⊗ N, for M, N ∈ �O, O[m]n and O[m]n, respectively.
1590
+ Fix ℓ ∈ N with ℓ ≥ 2. For M1, . . . , Mℓ ∈ �O, let
1591
+ M := M1 ⊗ · · · ⊗ Mℓ.
1592
+ Let us introduce some more notation. For x ∈ ˜g (resp., g[m]n and g[m]n) and i = 1, . . . , ℓ,
1593
+ let
1594
+ x(i) = 1 ⊗ · · · ⊗ 1⊗
1595
+ ix ⊗1 ⊗ · · · ⊗ 1
1596
+
1597
+ ��
1598
+
1599
+
1600
+ .
1601
+ For any operator A = �
1602
+ r∈I xr ⊗ yr, where xr, yr ∈ �g (resp., g[m]n and g[m]n), and for
1603
+ any distinct i, j ∈ {1, . . . , ℓ}, we define
1604
+ A(ij) =
1605
+
1606
+ r∈I
1607
+ x(i)
1608
+ r y(j)
1609
+ r .
1610
+ (4.1)
1611
+ Then �Ω(ij) can be viewed as a linear endomorphism on M. For any i = 1, . . . , ℓ and any
1612
+ distinct complex numbers z1, . . . , zℓ, the Gaudin Hamiltonian �Hi is defined by
1613
+ �Hi =
1614
+
1615
+
1616
+ j=1
1617
+ j̸=i
1618
+ �Ω(ij)
1619
+ zi − zj
1620
+ .
1621
+ (4.2)
1622
+
1623
+ 20
1624
+ Cheong and Lam
1625
+ It is a linear endomorphism on M. The Gaudin Hamiltonians Hi[m]n and H
1626
+ i[m]n are
1627
+ defined by replacing �Ω with Ω[m]n and Ω[m]n, respectively.
1628
+ Note that they are well-
1629
+ defined linear endomorphisms on M1 ⊗ · · · ⊗ Mℓ for M1, . . . , Mℓ in O[m]n and in O[m]n,
1630
+ respectively.
1631
+ For any N ∈ �O (resp., O[m]n and O[m]n), let
1632
+ N sing := {v ∈ N | Eβv = 0 for all β ∈ �Φ+ (resp., Φ[m]+
1633
+ n and Φ[m]+
1634
+ n )}
1635
+ stand for the subspace spanned by singular vectors in N, and let N sing
1636
+ µ
1637
+ denote the subspace
1638
+ spanned by singular vectors in the weight space Nµ for any weight µ of N. Note that Nµ
1639
+ is finite dimensional by definition of �O (resp., O[m]n and O[m]n).
1640
+ By an argument similar to the proof of [CaL, Propositions 3.5 and 3.7], one can show
1641
+ that �Hi (resp., Hi[m]n and H
1642
+ i[m]n) mutually commute with each other, and they are
1643
+ ˜g(resp., g[m]n and g[m]n)-homomorphisms for i = 1, . . . , ℓ. We immediately see that Msing
1644
+ and the finite-dimensional subspace Msing
1645
+ µ
1646
+ are �Hi-invariant for any weight µ of M. Thus,
1647
+ we may view �Hi as a linear endomorphism on Msing
1648
+ µ
1649
+ . Similarly, Hi[m]n (resp., H
1650
+ i[m]n)
1651
+ may be viewed as a linear endomorphism on (M1 ⊗ · · · ⊗ Mℓ)sing
1652
+ µ
1653
+ for M1, . . . , Mℓ ∈ O[m]n
1654
+ (resp., O[m]n) and any weight µ of M1 ⊗ · · · ⊗ Mℓ .
1655
+ Lemma 4.3. Let N1, N2 ∈ �O, and let v ∈ N1 ⊗ N2 be a weight vector of weight µ.
1656
+ (i) If µ ∈ Ξ[m], then �Ω(v) = Ω[m](v).
1657
+ (ii) If µ ∈ Ξ[m], then �Ω(v) = Ω[m](v).
1658
+ Proof. The proof is similar to that of [CaL, Lemma 3.11] with a slight modification. For
1659
+ completeness, we include it here. We will only prove (i).
1660
+ The proof of (ii) is similar.
1661
+ We may assume that v = v1 ⊗ v2, where vi ∈ Ni is a weight vector of µi for i = 1, 2,
1662
+ and µ1 + µ2 = µ.
1663
+ For i = 1, 2, µi ∈ Ξ[m] by Lemma 2.7.
1664
+ For all k ∈
1665
+ 1
1666
+ 2N\J+
1667
+ m and
1668
+ i = 1, 2, µi(Ek) = 0, and so Ek(vi) = 0. By virtue of Remark 4.2, it remains to consider
1669
+ β ∈ �Φ+ with ±β /∈ Φ[m]+. For such β, we have β(Ei) ̸= 0 for some i ∈ 1
1670
+ 2N\J+
1671
+ m. It
1672
+ follows that either the weight of Eβv1 or Eβv2 does not lie in �Ξ. Thus either Eβv1 = 0 or
1673
+ Eβv2 = 0, and hence Eβ ⊗ Eβ(v1 ⊗ v2) = 0. Similarly, Eβ ⊗ Eβ(v1 ⊗ v2) = 0. Therefore,
1674
+ �Ω(v1 ⊗ v2) = Ω[m](v1 ⊗ v2).
1675
+ As a consequence, we obtain the following lemma.
1676
+ Lemma 4.4. Let M1, . . . , Mℓ ∈ �O, and let v ∈ M1 ⊗ · · · ⊗ Mℓ be a weight vector of weight
1677
+ µ.
1678
+ (i) If µ ∈ Ξ[m], then �Hiv = Hi[m](v) for all i = 1, · · · , ℓ.
1679
+ (ii) If µ ∈ Ξ[m], then �Hiv = H
1680
+ i[m](v) for all i = 1, · · · , ℓ.
1681
+ We would like to ask whether the eigenvectors of Hi, Hi[m] and H
1682
+ i[m] are related. To
1683
+ answer the question, we need the following proposition. Recall the bijections �λ ↔ λ[m] ↔
1684
+ λ[m] in (2.15).
1685
+
1686
+ Gaudin Hamiltonians on unitarizable modules
1687
+ 21
1688
+ Proposition 4.5. Let M ∈ �O, and let �µ ∈ �P+ be a weight of M. Then:
1689
+ (i) There exists A ∈ U(˜l) such that the map t�µ
1690
+ [m] : Msing
1691
+ �µ
1692
+ → T[m](M)sing
1693
+ µ[m], defined by
1694
+ t�µ
1695
+ [m](v) = Av for v ∈ Msing
1696
+ �µ
1697
+ , is a linear isomorphism.
1698
+ (ii) There exists ¯A ∈ U(˜l) such that the map t�µ
1699
+ [m] : Msing
1700
+ �µ
1701
+ → T [m](M)sing
1702
+ µ[m], defined by
1703
+ t�µ
1704
+ [m](v) = ¯Av for v ∈ Msing
1705
+ µ
1706
+ , is a linear isomorphism.
1707
+ Proof. We will only prove (i). The proof of (ii) is similar. Note that there is a linear
1708
+ isomorphism
1709
+ Hom�O (∆(˜g, �µ), M) −→ Msing
1710
+ �µ
1711
+ ϕ �→ ϕ(v�µ)
1712
+ where v�µ is a highest weight vector of ∆(˜g, �µ).
1713
+ On the other hand, there exists A ∈ U(˜l) such that vµ[m] := Av�µ is a highest weight
1714
+ vector of ∆(g[m], µ[m]) = T[m](∆(˜g, �µ)). In fact, A is a product of elements in ˜l corre-
1715
+ sponding to a sequence of odd reflections (see [CL2, Section 3.1] and [CLW1, Section 4]
1716
+ for details). Similarly, the map
1717
+ HomO[m]
1718
+ �∆(g[m], µ[m]), T[m](M)� −→ T[m](M)sing
1719
+ µ[m]
1720
+ φ �→
1721
+ φ(vµ[m])
1722
+ is a linear isomorphism. By Theorem 2.10, we have
1723
+ Hom�O (∆(˜g, �µ), M) ∼= HomO[m]
1724
+
1725
+ ∆(g[m], µ[m]), T[m](M)
1726
+
1727
+ ,
1728
+ and hence Msing
1729
+ µ
1730
+ ∼= T[m](M)sing
1731
+ µ[m]. We may also see that any vector v ∈ Msing
1732
+ µ
1733
+ corresponds
1734
+ to Av ∈ T[m](M)sing
1735
+ µ[m] under the isomorphism, which shows that the isomorphism is indeed
1736
+ the map t�µ
1737
+ [m] as stated.
1738
+ Remark 4.6.
1739
+ (i) The elements A and ¯A in Proposition 4.5 depend only on the weight �µ,
1740
+ but not on the module M.
1741
+ (ii) There exist B, ¯B ∈ U(˜l) such that the inverses of t�µ
1742
+ [m] and t�µ
1743
+ [m] are given respectively
1744
+ by (t�µ
1745
+ [m])−1(v) = Bv and (t�µ
1746
+ [m])−1(w) = ¯Bw for any v ∈ T[m](M)sing
1747
+ µ[m] and w ∈
1748
+ T [m](M)sing
1749
+ µ[m]. Again B and B are products of elements in˜l corresponding to sequences
1750
+ of odd reflections.
1751
+ Theorem 4.7. For M1, . . . , Mℓ ∈ �O, let M = M1 ⊗ · · · ⊗ Mℓ. Suppose that v ∈ Msing
1752
+ �µ
1753
+ with �µ ∈ �P+. For any m ∈ Z+, let vm = t�µ
1754
+ [m](v) and vm = t�µ
1755
+ [m](v). For each i = 1, . . . , ℓ,
1756
+ we have:
1757
+
1758
+ 22
1759
+ Cheong and Lam
1760
+ (i) v is an eigenvector of �Hi with eigenvalue c if and only if vm is an eigenvector of
1761
+ Hi[m] with eigenvalue c.
1762
+ Moreover, �Hi is diagonalizable on Msing
1763
+ �µ
1764
+ if and only if Hi[m] is diagonalizable on
1765
+ T[m](M)sing
1766
+ µ[m]. In this case, they have the same spectrum.
1767
+ (ii) v is an eigenvector of �Hi with eigenvalue c if and only if vm is an eigenvector of
1768
+ H
1769
+ i[m] with eigenvalue c.
1770
+ Moreover, �Hi is diagonalizable on Msing
1771
+ �µ
1772
+ if and only if H
1773
+ i[m] is diagonalizable on
1774
+ T [m](M)sing
1775
+ µ[m]. In this case, they have the same spectrum.
1776
+ Proof. We will only prove (i). The proof of (ii) is similar. We know that vm = Av for
1777
+ some A ∈ U(˜l). Suppose �Hiv = cv for some c ∈ C. By Lemma 4.4 together with the fact
1778
+ that Av is a vector of weight µ[m] ∈ Ξ[m], we have
1779
+ Hi[m](Av) = �Hi(Av).
1780
+ As �Hi is a �g-homomorphism, it follows that
1781
+ Hi[m](vm) = A �Hi(v) = A(cv) = cvm.
1782
+ Conversely, suppose Hi[m](vm) = cvm for some c ∈ C. By �Hi being a �g-homomorphism
1783
+ and Lemma 4.4 again, we have
1784
+ A �Hiv = �Hi(Av) = Hi[m](Av) = cvm.
1785
+ By Remark 4.6, A has an inverse, and we deduce that �Hi(v) = cv. This proves the first
1786
+ part of (i). The second part is a direct consequence of the first part.
1787
+ Proposition 4.8.
1788
+ (i) For M1, . . . , Mℓ ∈ O[m], let M = M1 ⊗ · · · ⊗ Mℓ. Suppose that
1789
+ v ∈ Msing
1790
+ µ
1791
+ with µ ∈ P[m]+
1792
+ n . Then for each i = 1, . . . , ℓ, v is an eigenvector of Hi[m]
1793
+ with eigenvalue c if and only if v is an eigenvector of Hi[m]n with eigenvalue c.
1794
+ Moreover, Hi[m] is diagonalizable on Msing
1795
+ µ
1796
+ if and only if Hi[m]n is diagonalizable
1797
+ on tr∞
1798
+ n (M)sing
1799
+ µ
1800
+ . In this case, they have the same spectrum.
1801
+ (ii) For M1, . . . , Mℓ ∈ O[m], let M = M1 ⊗ · · · ⊗ Mℓ. Suppose that v ∈ Msing
1802
+ µ
1803
+ with
1804
+ µ ∈ P[m]+
1805
+ n . For each i = 1, . . . , ℓ, v is an eigenvector of H
1806
+ i[m] with eigenvalue c if
1807
+ and only if v is an eigenvector of H
1808
+ i[m]n with eigenvalue c.
1809
+ Moreover, H
1810
+ i[m] is diagonalizable on Msing
1811
+ µ
1812
+ if and only if H
1813
+ i[m]n is diagonalizable
1814
+ on tr∞
1815
+ n (M)sing
1816
+ µ
1817
+ . In this case, they have the same spectrum.
1818
+ Proof. We will only prove (i). The proof of (ii) is similar. Note that tr∞
1819
+ n (M)sing
1820
+ µ
1821
+ = Msing
1822
+ µ
1823
+ for µ ∈ P[m]+
1824
+ n . By Lemma 4.1, we have
1825
+ Hi[m](w) = Hi[m]n(w),
1826
+ for all w ∈ Msing
1827
+ µ
1828
+ .
1829
+ The first part of (i) follows. The second part of (i) clearly follows from the first part.
1830
+
1831
+ Gaudin Hamiltonians on unitarizable modules
1832
+ 23
1833
+ 5
1834
+ Gaudin Hamiltonians on modules over G[m]n
1835
+ In this section, we consider Gaudin Hamiltonians for finite-dimensional Lie (super)algebras.
1836
+ We relate the Gaudin Hamiltonians for G[m]n (resp, G[m]n) to those for g[m]n (resp,
1837
+ g[m]n) for n ∈ N. Furthermore, we study the Gaudin Hamiltonians on the tensor product
1838
+ of unitarizable irreducile highest weight modules and give an affirmative answer to the
1839
+ diagonalization of these operators.
1840
+ Let us fix ℓ ∈ N with ℓ ≥ 2. For n ∈ N, the Casimir symmetric tensors for G[m]n and
1841
+ G[m]n are defined by (cf. [CaL, Section 3.4])
1842
+ ˚Ω[m]n =
1843
+
1844
+ β∈Φ[m]+
1845
+ n
1846
+
1847
+ Eβ ⊗ Eβ + (−1)|Eβ|Eβ ⊗ Eβ) +
1848
+
1849
+ j∈J+
1850
+ m(n)
1851
+ (−1)2jEj ⊗ Ej,
1852
+ ˚Ω[m]n =
1853
+
1854
+ β∈Φ[m]+
1855
+ n
1856
+ (Eβ ⊗ Eβ + (−1)|Eβ|Eβ ⊗ Eβ) +
1857
+
1858
+ j∈J
1859
+ +
1860
+ m(n)
1861
+ (−1)2jEj ⊗ Ej.
1862
+ Clearly, ˚Ω[m]n and ˚Ω[m]n lie in U(G[m]n) ⊗ U(G[m]n) and U(G[m]n) ⊗ U(G[m]n), respec-
1863
+ tively.
1864
+ For any i = 1, . . . , ℓ and any distinct complex numbers z1, . . . , zn, the Gaudin Hamil-
1865
+ tonians Hi[m]n and Hi[m]n are defined by
1866
+ Hi[m]n =
1867
+
1868
+
1869
+ j=1
1870
+ j̸=i
1871
+ ˚Ω[m]n
1872
+ (ij)
1873
+ zi − zj
1874
+ and
1875
+ Hi[m]n =
1876
+
1877
+
1878
+ j=1
1879
+ j̸=i
1880
+ ˚Ω[m]n
1881
+ (ij)
1882
+ zi − zj
1883
+ .
1884
+ (5.1)
1885
+ From now on, we fix ξ1, . . . , ξℓ ∈ h[m]∗
1886
+ n (resp., ξ1, . . . , ξℓ ∈ h[m]∗
1887
+ n), put ξ := (ξ1, . . . , ξℓ)
1888
+ (resp., ξ := (ξ1, . . . , ξℓ)). We define
1889
+ L(G[m]n, ξ) := L(G[m]n, ξ1) ⊗ · · · ⊗ L(G[m]n, ξℓ)
1890
+ and
1891
+ L(G[m]n, ξ) := L(G[m]n, ξ1) ⊗ · · · ⊗ L(G[m]n, ξℓ).
1892
+ Setting K = 0 to the cases involving central extensions, the Gaudin Hamiltonians
1893
+ Hi[m]n on L(G[m]n, ξ) (resp., Hi[m]n on L(G[m]n, ξ)) mutually commute with each other,
1894
+ and they are G[m]n(resp., G[m]n)-homomorphisms. It is also evident that for any weight µ
1895
+ of L(G[m]n, ξ) (resp., L(G[m]n, ξ)), the subspace L(G[m]n, ξ)sing
1896
+ µ
1897
+ (resp., L(G[m]n, ξ)sing
1898
+ µ
1899
+ )) is
1900
+ Hi[m]n-invariant (resp., Hi[m]n-invariant). Here and below, N sing stands for the subspace
1901
+ spanned by singular vectors in N with respect to the standard Borel subalgebra and N sing
1902
+ µ
1903
+ for the subspace spanned by singular vectors in the weight space Nµ for any G[m]n(resp.,
1904
+ G[m]n)-module N and any weight µ of N.
1905
+ Each gx[m]n(resp., gx[m]n)-module can be regarded as a Gx[m]n(resp., G
1906
+ x[m]n)-module
1907
+ through the homomorphism ι defined by (2.11). It is clear that the set of singular vectors
1908
+ in a module M regarded as a gx[m]n(resp., gx[m]n)-module equals the set of singular vectors
1909
+ in the module M regarded as a Gx[m]n(resp., G
1910
+ x[m]n)-module.
1911
+
1912
+ 24
1913
+ Cheong and Lam
1914
+ Proposition 5.1.
1915
+ (i) For n ∈ N and i = 1, . . . , ℓ, let Li be an irreducible highest weight
1916
+ gx[m]n-module with highest weight in Qx(m|n) and let L = L1 ⊗ · · · ⊗ Lℓ. Suppose
1917
+ that µ = �
1918
+ j∈J+
1919
+ m(n) ajǫj + dΛ0 is a weight of L. Let ˆµ = �
1920
+ j∈J+
1921
+ m(n) ajǫj + d1m|n.
1922
+ Then for each i = 1, . . . , ℓ, Hi[m]n is diagonalizable on Lsing
1923
+ µ
1924
+ if and only if Hi[m]n is
1925
+ diagonalizable on Lsing
1926
+ ˆµ
1927
+ . Moreover, Hi[m]n has simple spectrum on the space spanned
1928
+ by singular vectors in a finite direct sum of weight spaces of L if and only if Hi[m]n
1929
+ has simple spectrum on the space spanned by singular vectors in the finite direct sum
1930
+ of the corresponding weight spaces.
1931
+ (ii) For n ∈ N and i = 1, . . . , ℓ, let Li be an irreducible highest weight gx[m]n-module
1932
+ with highest weight in Q
1933
+ x(m|n), and let L = L1 ⊗ · · · ⊗ Lℓ.
1934
+ Suppose that µ =
1935
+
1936
+ j∈J
1937
+ +
1938
+ m(n) ajǫj + dΛ0 is a weight of L. Let ˆµ = �
1939
+ j∈J
1940
+ +
1941
+ m(n) ajǫj − d1m|n. Then for each
1942
+ i = 1, . . . , ℓ, H
1943
+ i[m]n is diagonalizable on Lsing
1944
+ µ
1945
+ if and only if Hi[m]n is diagonalizable
1946
+ on Lsing
1947
+ ˆµ
1948
+ . Moreover, H
1949
+ i[m]n has simple spectrum on the space spanned by singular
1950
+ vectors in a finite direct sum of weight spaces of L if and only if Hi[m]n has simple
1951
+ spectrum on the space spanned by singular vectors in the finite direct sum of the
1952
+ corresponding weight spaces.
1953
+ Proof. Note that ι ⊗ ι�˚Ω[m]n
1954
+ � = Ω[m]n − (m − n)K ⊗ K. Therefore we have
1955
+ Hi[m]n(v) = Hi[m]n(v) − (m − n)
1956
+
1957
+
1958
+ j=1
1959
+ j̸=i
1960
+ K(i)K(j)
1961
+ zi − zj
1962
+ (v).
1963
+ The last term on the right hand side is a fixed scalar times v. This implies (i). The proof
1964
+ of (ii) is similar.
1965
+ The following theorem follows from the Main Corollary of Rybnikov [Ry].
1966
+ Theorem 5.2. Let n ∈ N and ξ1, . . . , ξℓ ∈ h[0]∗
1967
+ n dominant integral weights for Gx[0]n. For
1968
+ generic z1, . . . , zℓ, the Gaudin Hamiltonian Hi[0]n is diagonalizable with simple spectrum
1969
+ on the space L(Gx[0]n, ξ)sing for all i = 1, . . . , ℓ.
1970
+ Remark 5.3. The results in [Ry] involve only simple Lie algebras.
1971
+ The Lie algebra
1972
+ Ga[0]n ∼= gl(n) is, however, not semisimple.
1973
+ The case where x = a is still true since
1974
+ the irreducible highest weight modules over gl(n) coincides with those over sl(n), and
1975
+ each Gaudin Hamiltonian for gl(n) minus that for sl(n) is a scalar multiple of the identity
1976
+ operator on the space L(Ga[0]n, ξ)sing.
1977
+ Corollary 5.4. Let ξ1, . . . , ξℓ ∈ Qx(0|∞). For each i = 1, . . . , ℓ and generic z1, . . . , z��, the
1978
+ Gaudin Hamiltonian Hi[0] is diagonalizable with simple spectrum on the space spanned by
1979
+ singular vectors in any finite direct sum of weight spaces of L(gx[0], ξ1) ⊗ · · · ⊗ L(gx[0], ξℓ).
1980
+ Proof. Note that tr∞
1981
+ n (L(gx[0], ξi), regarded as a Gx[0]n-module through the homomorphism
1982
+ ι defined by (2.11), is an irreducible module with dominant integral highest weight. Given
1983
+ a finite number of weight spaces of L(gx[0], ξi), we can choose n large enough such that
1984
+ tr∞
1985
+ n (L(gx[0], ξi) contains the given weight spaces. Now the corollary follows from Proposi-
1986
+ tion 4.8, Proposition 5.1 and Theorem 5.2.
1987
+
1988
+ Gaudin Hamiltonians on unitarizable modules
1989
+ 25
1990
+ Theorem 5.5. Let n ∈ N and ξ1, . . . , ξℓ ∈ Q
1991
+ x,I(m|n). For each i = 1, . . . , ℓ and generic
1992
+ z1, . . . , zℓ, the Gaudin Hamiltonian Hi[m]n is diagonalizable with simple spectrum on the
1993
+ space spanned by singular vectors in any finite direct sum of weight spaces of L(G
1994
+ x[m]n, ξ).
1995
+ In particular, for x = a, i = 1, . . . , ℓ and generic z1, . . . , zℓ, the Gaudin Hamiltonian
1996
+ Hi[m]n is diagonalizable with simple spectrum on L(G
1997
+ a[m]n, ξ)sing.
1998
+ Proof. The theorem follows from Proposition 5.1(ii), Proposition 4.8(ii), Theorem 4.7(ii)
1999
+ and the case where m = 0 of Theorem 4.7(i), and Corollary 5.4 together with an explicit
2000
+ description of the highest weights involved. The second part follows from the fact that
2001
+ L(G
2002
+ a[m]n, ξ) is finite-dimensional.
2003
+ Remark 5.6.
2004
+ (i) For x = a and each L(G
2005
+ a[m]n, ξi) being the natural module Cm|n, the
2006
+ corresponding result in Theorem 5.5 has been obtained by Mukhin, Vicedo and Young
2007
+ (cf. [MVY]).
2008
+ (ii) We have G
2009
+ c[0]n ∼= so(2n) and G
2010
+ d[0]n ∼= sp(2n). The weights ξ1, . . . , ξℓ in Theorem 5.5
2011
+ are highest weights of infinite-dimensional unitarizable irreducible highest weight
2012
+ modules (see Remark 3.9).
2013
+ (iii) Theorem 5.5 is also valid for the ortho-symplectic Lie superalgebra osp(2m + 1|2n)
2014
+ if ξ1, . . . , ξℓ are the highest weights such that for each i = 1, . . . , ℓ, ξi = λi − di
2015
+ 1m|n
2016
+ for some λi ∈ P(m|n) and di ∈ Z+ satisfying ℓ(λi) ≤ di. The proof is identical and
2017
+ is omitted here.
2018
+ Theorem 5.7. Let x = c, d, n ∈ N and ξ1, . . . , ξℓ ∈ Q
2019
+ x,II(m|n). For each i = 1, . . . , ℓ
2020
+ and generic z1, . . . , zℓ, the Gaudin Hamiltonian Hi[m]n is diagonalizable with simple spec-
2021
+ trum on the space spanned by singular vectors in any finite direct sum of weight spaces of
2022
+ L(G
2023
+ x[m]n, ξ).
2024
+ Proof. Each G
2025
+ c[m]n(resp., G
2026
+ d[m]n)-module L(G
2027
+ c[m]n, ξi) (resp., L(G
2028
+ d[m]n, ξi)) can be re-
2029
+ garded as a Gd[m]n(resp., Gc[m]n)-module through the isomorphism �ϕ defined by (2.10).
2030
+ Now the theorem follows from Proposition 5.1(i), Proposition 4.8(i), Theorem 4.7(i) and
2031
+ its special case where m = 0, and Corollary 5.4 together with an explicit description of
2032
+ the highest weights involved.
2033
+ We anticipate that the results of this paper may provide an approach of constructing
2034
+ common eigenvectors of the Gaudin Hamiltonians associated to the Lie (super)algebra
2035
+ G
2036
+ x[m]n from common eigenvectors of the Gaudin Hamiltonians associated to the Lie algebra
2037
+ Gx[0]k for some k ∈ N. Let us explain in more detail. For each i = 1, . . . , ℓ, consider
2038
+ the Gaudin Hamiltonian Hi[m]n on the weight space L(G
2039
+ x[m]n, ξ)sing
2040
+ µ
2041
+ for each weight µ.
2042
+ Suppose that the weights ξ1, . . . , ξℓ all lie in Q
2043
+ x,I(m|n) or Q
2044
+ x,II(m|n).
2045
+ The arguments
2046
+ in Sections 4 and 5 allow us to construct an eigenvector of the linear operator from an
2047
+ eigenvector of the Gaudin Hamiltonian for the Lie algebra Gx[0]k, for some k ∈ N, on the
2048
+ space spanned by singular vectors in L1 ⊗ · · · ⊗ Lℓ, where Li is some finite-dimensional
2049
+ irreducible Gx[0]k-module for each i = 1, . . . , ℓ. An explicit construction would be to apply
2050
+ a sequence of certain odd reflections to the Bethe vectors of Hi[0]k, which are constructed
2051
+
2052
+ 26
2053
+ Cheong and Lam
2054
+ by Bethe ansatz method.
2055
+ These odd reflections depend on the weight µ and can be
2056
+ determined explicitly. However, one can expect that the resulting eigenvectors might take
2057
+ complicated forms. It would be interesting to know whether this procedure is directly
2058
+ related to any other known method of constructing eigenvectors.
2059
+ Acknowledgments. The first author was partially supported by Ministry of Science and
2060
+ Technology grant 110-2115-M-006-006 of Taiwan. The second author was partially sup-
2061
+ ported by Ministry of Science and Technology grant 109-2115-M-006-019-MY3 of Taiwan.
2062
+ References
2063
+ [BF]
2064
+ H. M. Babujian and R. Flume, Off-shell Bethe ansatz equation for Gaudin mag-
2065
+ nets and solutions of Knizhnik-Zamolodchikov equations, Modern Phys. Lett. A
2066
+ 9 (1994), no. 22, 2029–2039.
2067
+ [CaL]
2068
+ B. Cao and N. Lam, Solutions of super Knizhnik-Zamolodchikov equations, Lett.
2069
+ Math. Phys. 110 (2020), no. 7, 1799–1834.
2070
+ [CL1]
2071
+ S.-J. Cheng and N. Lam, Infinite-dimensional Lie superalgebras and Hook Schur
2072
+ functions, Commun. Math. Phys. 238 (2003), 95–118.
2073
+ [CL2]
2074
+ S.-J. Cheng and N. Lam, Irreducible characters of general linear superalgebra
2075
+ and super duality, Commun. Math. Phys. 280 (2010), 645–672.
2076
+ [CLW1]
2077
+ S.-J. Cheng, N. Lam and W. Wang, Super duality and irreducible characters of
2078
+ ortho-symplectic Lie superalgebras, Invent. Math. 183 (2011), 189–224.
2079
+ [CLW2]
2080
+ S.-J. Cheng, N. Lam and W. Wang,
2081
+ Super duality for general linear Lie su-
2082
+ peralgebras and applications in Recent Developments in Lie Algebras, Groups
2083
+ and Representation Theory, Proc. Sympos. Pure Math. 86, Amer. Math. Soc.,
2084
+ Providence, 2012, 113–136.
2085
+ [CLZ]
2086
+ S.-J. Cheng, N. Lam and R. Zhang, Character formula for infinite-dimensional
2087
+ unitarizable modules of the general linear superalgebra, J. Algebra 273 (2004),
2088
+ no. 2, 780–805.
2089
+ [CW]
2090
+ S.-J. Cheng and W. Wang, Dualities and representations of Lie superalgebras,
2091
+ Graduate Studies in Mathematics 144, American Mathematical Society, 2012.
2092
+ [EHW]
2093
+ T. J. Enright, R. Howe and N. R. Wallach, A classification of unitary highest
2094
+ weight modules, in Representation theory of reductive group, Birkh¨auser, Boston,
2095
+ 1983, 97–143.
2096
+ [EFK]
2097
+ P. I. Etingof, I. B. Frenkel and A. A. Kirillov, Jr., Lectures on representation
2098
+ theory and Knizhnik-Zamolodchikov equations, Mathematical surveys and mono-
2099
+ graphs 58, American Mathematical Society, 1998.
2100
+ [FFR]
2101
+ B. Feigin, E. Frenkel and N. Reshetikhin, Gaudin model, Bethe ansatz and crit-
2102
+ ical level, Comm. Math. Phys. 166 (1994), no. 1, 27–62.
2103
+ [FFRy]
2104
+ B. Feigin, E. Frenkel and L. Rybnikov, Opers with irregular singularity and
2105
+ spectra of the shift of argument subalgebra, Duke Math. J. 155 (2010), no. 2,
2106
+ 337–363.
2107
+
2108
+ Gaudin Hamiltonians on unitarizable modules
2109
+ 27
2110
+ [G1]
2111
+ M. Gaudin, Diagonalisation d’une classe d’Hamiltoniens de spin, J. Physique
2112
+ 37 (1976), no. 10, 1089–1098.
2113
+ [G2]
2114
+ M. Gaudin, La fonction d’onde de Bethe, Collection du Commissariat a‘
2115
+ l’E’nergie Atomique: Se’rie Scientifique, Masson, Paris, 1983.
2116
+ [HMVY] C. Huang, E. Mukhin, B. Vicedo and C. Young, The solutions of glM|N Bethe
2117
+ ansatz equation and rational pseudodifferential operators, Selecta Math. (N.S.)
2118
+ 25 (2019), no. 4, Paper No. 52, 34 pp.
2119
+ [HLT]
2120
+ P.-Y. Huang, N. Lam and T.-M. To, Super duality and homology of unitarizable
2121
+ modules of Lie algebras, Publ. Res. Inst. Math. Sci. 48 (2012), no. 1, 45–63.
2122
+ [K]
2123
+ V. G. Kac, Lie Superalgebras, Adv. Math. 16 (1977), 8–96.
2124
+ [KZ]
2125
+ V. G. Knizhnik and A. B. Zamolodchikov, Current algebra and Wess-Zumino
2126
+ model in two dimensions., Nucl. Phys. B 247 (1984), 83–103.
2127
+ [KuM]
2128
+ P. P. Kulish and N. Manojlovi´c, Bethe vectors of the osp(1|2) Gaudin model,
2129
+ Lett. Math. Phys. 55 (2001), no. 1, 77–95.
2130
+ [LZ1]
2131
+ N. Lam and R. Zhang, Quasi-finite modules for Lie superalgebras of infinite rank,
2132
+ Trans. Amer. Math. Soc. 358 (2006), no. 1, 403–439.
2133
+ [LZ2]
2134
+ N. Lam and R. Zhang, u-cohomology formula for unitarizable modules over gen-
2135
+ eral linear superalgebras, J. Algebra 327 (2011), 50–70.
2136
+ [L]
2137
+ K. Lu, Completeness of Bethe ansatz for Gaudin models associated with gl(1|1),
2138
+ Nuclear Phys. B 980 (2022), Paper No. 115790, 23 pp.
2139
+ [LM]
2140
+ K. Lu and E. Mukhin, Bethe Ansatz equations for orthosymplectic Lie superalge-
2141
+ bras and self-dual superspaces, Ann. Henri Poincar´e 22 (2021), no. 12, 4087–4130.
2142
+ [LMV]
2143
+ K. Lu, E. Mukhin and A. Varchenko, On the Gaudin model associated to Lie
2144
+ algebras of classical types, J. Math. Phys. 57 (2016), no. 10, 101703, 23 pp.
2145
+ [Lus]
2146
+ G. Lusztig, Introduction to quantum groups, Progress in Mathematics 110,
2147
+ Birkh¨auser, Boston, 1993.
2148
+ [MTV1]
2149
+ E. Mukhin, V. Tarasov and A. Varchenko, Schubert calculus and representations
2150
+ of the general linear group, J. Amer. Math. Soc. 22 (2009), no. 4, 909–940.
2151
+ [MTV2]
2152
+ E. Mukhin, V. Tarasov and A. Varchenko, On separation of variables and com-
2153
+ pleteness of the Bethe ansatz for quantum glN Gaudin model, Glasg. Math. J.
2154
+ 51 (2009), no. A, 137–145.
2155
+ [MV]
2156
+ E. Mukhin and A. Varchenko, Norm of a Bethe vector and the Hessian of the
2157
+ master function, Compos. Math. 141 (2005), no. 4, 1012–1028.
2158
+ [MVY]
2159
+ E. Mukhin, B. Vicedo and C. Young, Gaudin models for gl(m|n), J. Math. Phys.
2160
+ 56 (2015), no. 5, 051704, 30 pp.
2161
+ [Ry]
2162
+ L. Rybnikov, A proof of the Gaudin Bethe ansatz conjecture, Int. Math. Res.
2163
+ Not. IMRN (2020), no. 22, 8766–8785.
2164
+ Department of Mathematics, National Cheng Kung University, Tainan 701401, Taiwan
2165
+ E-mail address: keng@ncku.edu.tw
2166
+ Department of Mathematics, National Cheng Kung University, Tainan 701401, Taiwan
2167
+ E-mail address: nlam@ncku.edu.tw
2168
+
J9E1T4oBgHgl3EQfYgQR/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff
 
JdAzT4oBgHgl3EQfj_0H/content/tmp_files/2301.01523v1.pdf.txt ADDED
@@ -0,0 +1,1534 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Intracluster light is already abundant at redshift beyond unity
2
+ Hyungjin Joo1 & M. James Jee1,2
3
+ 1Department of Astronomy, Yonsei University, Seoul, Republic of Korea.
4
+ 2Department of Physics and Astronomy, University of California, Davis, Davis, CA, USA
5
+ Volume 613
6
+ https://doi.org/10.1038/s41586-022-0536-4
7
+ Received: 8 June 2022
8
+ Accepted: 28 September 2022
9
+ Intracluster light (ICL) is diffuse light from stars that are gravitationally bound not to individual member
10
+ galaxies, but to the halo of galaxy clusters. Leading theories1,2 predict that the ICL fraction, defined by
11
+ the ratio of the ICL to the total light, rapidly decreases with increasing redshift, to the level of a few per
12
+ cent at z > 1. However, observational studies have remained inconclusive about the fraction beyond
13
+ redshift unity because, to date, only two clusters in this redshift regime have been investigated. One
14
+ shows a much lower fraction than the mean value at low redshift3, whereas the other possesses a fraction
15
+ similar to the low-redshift value4. Here we report an ICL study of ten galaxy clusters at 1 ≲ z ≲ 2 based
16
+ on deep infrared imaging data. Contrary to the leading theories, our study finds that ICL is already
17
+ abundant at z ≳ 1, with a mean ICL fraction of approximately 17%. Moreover, no significant correlation
18
+ between cluster mass and ICL fraction or between ICL colour and cluster-centric radius is observed.
19
+ Our findings suggest that gradual stripping can no longer be the dominant mechanism of ICL formation.
20
+ Instead, our study supports the scenario wherein the dominant ICL production occurs in tandem with
21
+ the formation and growth of the brightest cluster galaxies and/or through the accretion of preprocessed
22
+ stray stars.
23
+ Main
24
+ Intracluster light (ICL) is predominantly distributed in the central region of the cluster, in most
25
+ cases around the brightest cluster galaxy (BCG) out to several hundred kilo-parsecs3-6. Some studies
26
+ reported that significant ICL is also found around intermediate and massive satellites7,8. We detected
27
+ ICL around the BCGs of ten galaxy clusters at z ≳ 1with the Wide Field Camera 3 (WFC3) near-infrared
28
+ imager on board Hubble Space Telescope (HST) (Fig. 1). In most cases, a clear surface brightness (SB)
29
+ profile is obtained out to approximately 200 kpc, where it approaches the surface brightness limit μ ≈
30
+ 28 mag arcses-2 (Fig. 2). The exception is the result for JKCS041, which is the highest redshift (z = 1.8)
31
+ target in our sample (Extended Data Table 1). Its SB profile approached the limit at around 100 kpc.
32
+ Figure 2 shows that, overall, the SB profiles are well described by a superposition of two or three multi-
33
+ Sérsic components convolved with the instrument point spread function (PSF). Regardless of the
34
+ number of components, the outermost component is predominantly responsible for the shape of the
35
+ outer part of the SB profile, which is assumed to characterize the ICL here, whereas all inner
36
+
37
+ components (one component if the total number of components is two) are considered to represent the
38
+ BCG profile. Most clusters in our sample show no significant gradients in their SB colour profiles and
39
+ their colours are in good agreement with those of the reddest cluster members. Exceptions are found for
40
+ SpARCS1049 and IDCS1426, which possess a clear negative gradient, with the colour difference
41
+ between the BCG- and ICL-dominant regions being around 1 mag (Fig. 3). IDCS1426 and
42
+ SpARCS1049 are the second (z = 1.75) and third (z = 1.71) highest redshift clusters in our sample.
43
+ When converted to the rest-frame B and V mags, the BCG colours span the 0.5 < B − V < 0.8 range,
44
+ which overlaps the theoretical distribution9. Combining spectroscopic and photometric member
45
+ selections, we measured the BCG + ICL and ICL fractions (fBCG+ICL and fICL) using an aperture of
46
+ r = 0.5 Mpc (Extended Data Table 2). The mean BCG fraction is approximately 4.5%, which is well
47
+ bracketed by the values in previous studies3,8,10,11. Figure 4 shows that the mean ICL fraction of our
48
+ sample is similar to that of the low-redshift sample in the literature1,3,5,8,10-21.
49
+ One potential difficulty for the interpretation of Fig. 4 is the diversity of the methodology in
50
+ the previous studies. We investigated the impacts of the following two factors: aperture size and ICL
51
+ definition. The results compiled in Fig. 4 are based on apertures ranging from 100 kpc to 1.7 Mpc. We
52
+ verified that there is no correlation between aperture size and ICL fraction in the published result.
53
+ Moreover, the mean aperture size in the literature is 0.58 Mpc, which is similar to our choice of 0.5
54
+ Mpc. Finally, when we repeated the analysis using the subsample that used the aperture sizes between
55
+ 0.35 and 0.65 Mpc, the result remained unchanged. The ICL community is aware that the results from
56
+ the traditional SB cut (SBC) method can differ systematically from those obtained by the new
57
+ multicomponent decomposition method22 to which our approach belongs. To address the issue, we
58
+ divided the literature sample into the SBC and multicomponent decomposition subsamples. Our
59
+ regression based on the latter shows that the slope is still consistent with zero at the 2σ level.
60
+ We find that dwarf galaxies fainter than our detection limit do not bias our ICL fractions high.
61
+ To investigate the impact of sources fainter than our detection limit, we carried out image simulations
62
+ by randomly distributing dwarf galaxies, whose number is estimated by fitting a Schechter luminosity
63
+ function to the detected source distribution and computing the difference between the best-fit luminosity
64
+ function and observed distribution. We considered two types of radial distributions. The first is a
65
+ uniform distribution across the field. The second is the distribution that follows a Navarro–Frenk–White
66
+ profile23. In the first case, the ICL fraction is unchanged because adding a uniform dwarf galaxy
67
+ distribution is equivalent to elevating the sky level by the same degree simultaneously across the entire
68
+ field. In the second case, the dwarf galaxies are mostly concentrated near the BCG. Although this
69
+ certainly would lead to the overestimation of the BCG luminosity, the impact on the ICL luminosity
70
+ was negligible.
71
+
72
+ We rule out the possibility that unmasked galaxy light might artificially increase the ICL
73
+ fraction. In our analysis, we employed a moderate-sized mask and later applied a correction factor to
74
+ obtain the result effectively measured with the full mask (see Methods). This correction scheme was
75
+ verified to be accurate, leading to only an approximately 0.02% difference in the average ICL fraction
76
+ (Extended Data Table 2).
77
+ One may argue that the ten clusters in our sample correspond to the most massive population
78
+ at high redshift and thus should not be compared directly with the low-redshift clusters. Although four
79
+ of our ten clusters may potentially belong to extremely massive (around 10
80
+ 15M☉) populations in the 1
81
+ ≲ z ≲ 2 universe, the masses of the remaining six clusters span the range 2–6 × 1014M☉ (Extended
82
+ Data Table 3). We found that the cluster masses do not correlate with the ICL fractions for our sample
83
+ (Extended Data Fig. 5). Although theoretical studies
84
+ 1,2,24–27 remain inconclusive about the fICL–mass
85
+ correlation, observational studies
86
+ 22,28,29 agree that there is no correlation. This lack of the fICL–mass
87
+ correlation is also supported in the study where the sample is limited to a narrow redshift range of z <
88
+ 0.0722. Hence, we do not attribute the absence of the fICL–redshift correlation to a selection effect.
89
+ The dominant ICL production mechanism is still unknown, although the current consensus is
90
+ that merger, stripping and preprocessing are the three important candidates22,30–32. The ICL fraction is
91
+ an important observable sensitive to the timescale of the ICL formation, and its evolution with redshift
92
+ can be used to discriminate between competing theories regarding the dominant ICL production
93
+ mechanism. A strong evolution
94
+ 1,2,10,28,33 favours a gradual process through stripping, whereas the
95
+ opposite
96
+ 5,29 supports the scenario wherein the dominant ICL production happened at high redshifts. The
97
+ absence of the apparent evolution of the ICL fraction in the 0 ≲ z ≲ 2 redshift regime in the current
98
+ study contracts the current leading theories1,2, which predict that the mean ICL fraction decreases to a
99
+ negligible level (less than 5%) at z = 1.5 (Fig. 4). Therefore, the most straightforward interpretation of
100
+ the current finding is that the dominant ICL formation and its evolution with redshift can occur not
101
+ through gradual stripping, but in tandem with the BCG formation and growth, and/or through the
102
+ accretion of preprocessed stray stars.
103
+ Together with the ICL fraction, another traditional but still critical method to discriminate
104
+ between competing theories on the dominant mechanism of ICL production is to investigate the ICL
105
+ stellar population with its colour and compare the results with those of the cluster galaxies, including
106
+ the BCG. For instance, if major mergers with the BCG are the dominant mechanism, no significant
107
+ colour difference between BCG and ICL is expected. On the other hand, if the ICL is formed by a more
108
+ gradual process such as stellar stripping, we expect that the ICL would be bluer than the BCG, or a
109
+
110
+ nega- tive gradient would be present in the radial colour profile. In this case, matching the colour
111
+ between the ICL and cluster galaxies can constrain the progenitors of the ICL. Previous observational
112
+ studies in general support the presence of negative gradients, although exceptions are not
113
+ uncommon4,19,34,35, which implies that the gradient may depend on the particular assembly history of
114
+ individual clusters9. The absence of the ICL colour gradient in most of the cases in our sample indicates
115
+ that gradual stripping is not likely to be the dominant mechanism of ICL production within the 1 ≲ z ≲
116
+ 2 epoch. It is possible that occasional major mergers can potentially mix the intracluster stars and flatten
117
+ the ICL colour profile even in the case where ICL production through stripping is dominant. However,
118
+ as the ICL colours in the flat gradient cases are in good agreement with those of the reddest cluster
119
+ members, our observation cannot reconcile with this scenario.
120
+
121
+
122
+
123
+
124
+ Fig. 1 | BCG + ICL images of our ten z ≳ 1 clusters. The images were created by masking out every
125
+ discrete source detected by SExtractor except for the BCGs. Here the result is based on an expansion
126
+ coefficient of 2. We also visually scanned the result and applied additional manual masking for the
127
+ objects that SExtractor failed to identify. Green dashed lines show the SMA = 200 kpc ellipses, whose
128
+ ellipticities and position angles are determined by AutoProf. Throughout the paper, we assume a flat Λ-
129
+ dominated cold dark matter cosmology characterized by h = 0.7 and Ωm,0 = 1 − ΩΛ,0 = 0.3, where h,
130
+ Ωm,0 and ΩΛ,0 represent the dimensionless Hubble, matter density, and dark energy density parameters
131
+ at present day, respectively.
132
+
133
+ XMM1229/E160W
134
+ SPT2106/E140W
135
+ M001142/F140W
136
+ RDCS1252/E160W
137
+ 100 kpc
138
+ 100 kpc
139
+ 100 kpc
140
+ 100 kpc
141
+ M001014/F140W
142
+ SPT0205/F140W
143
+ XMM2235/F160W
144
+ SpARCS1049/F160W
145
+ 100 kpc
146
+ 100 kpc
147
+ 100 kpc
148
+ 100 kpc
149
+ DCS1426/F140W
150
+ JKCS041/F160W
151
+ 22
152
+ 24
153
+ 26
154
+ 28
155
+ μ (mag arcsec-2)
156
+ 100 kpc
157
+ 100 kpc
158
+ Fig. 2 | BCG + ICL radial profiles. Data points are the observed surface brightness from the elliptical
159
+ bins. The errors are computed from quadratic sums of the background level error and shot noise. With
160
+ dashed and solid lines, we show our best-fit multi-Sérsic component model. Magenta (cyan) dashed
161
+ lines are the innermost (outermost) components. When the target requires three components, we use
162
+ orange lines to represent the middle component. Red solid lines illustrate the summation of all
163
+ components. The legends of each panel show the target name, filter type, χ2 value and best-fit Sérsic
164
+ indices.
165
+
166
+ 167
167
+ 167
168
+ XMM1229/F105W
169
+ .
170
+ XMM1229/F160W
171
+ .
172
+ SPT2106/F105W
173
+ .
174
+ SPT2106/F140W
175
+ 18 -
176
+ Fitting,x= 4.15
177
+ Fitting, x = 1.59
178
+ 18-
179
+ Fitting,x2= 2.66
180
+ Fitting, x = 3.04
181
+ 20
182
+ BCG, n = 2.3 ± 0.6
183
+ BCG, n = 3.0 ± 0.4
184
+ arcsec-2)
185
+ 20
186
+ BCG, n = 0.9 ± 0.1
187
+ BCG, n = 1.4 ± 0.1
188
+ ICL, n = 0.9 ± 0.2
189
+ ICL, n = 0.7 ± 0.1
190
+ 0:0:00
191
+ ICL, n = 1.4 ± 0.4
192
+ ICL, n = 1.9 ± 0.7
193
+ 22
194
+ 22
195
+ 24
196
+ (mag
197
+ 24
198
+ 26 -
199
+ 8
200
+ 26
201
+ 28
202
+ 28
203
+ 30
204
+ 30 -
205
+ 100
206
+ 101
207
+ 102
208
+ 100
209
+ 101
210
+ 102
211
+ 100
212
+ 101
213
+ 102
214
+ 100
215
+ 101
216
+ 102
217
+ 16 -
218
+ .
219
+ MOO1142/F105W
220
+ MOO1142/F140W
221
+ 16
222
+ RDCS1252/F105W
223
+ RDCS1252/F160W
224
+ 18
225
+ Fitting, x = 1.47
226
+ Fitting, x = 1.87
227
+ 18
228
+ Fitting, x = 1.46
229
+ Fiting, x = 3.25
230
+ BCG1, n = 0.9 ± 0.3
231
+ BCG1, n = 0.9 ± 0.3
232
+ BCG1, n = 0.9 ± 0.2
233
+ BCG1, n = 0.3 ±0.1
234
+ 20
235
+ BCG2, n = 1.4 ± 0.1
236
+ BCG2, n = 1.5 ± 0.1
237
+ arcsec
238
+ 20
239
+ BCG2, n = 0.4 ± 0.2
240
+ BCG2, n = 0.5 ± 0.2
241
+ 22
242
+ ICL, n = 1.8 ± 0.5
243
+ ICL, n = 2.1 ± 0.5
244
+ 22
245
+ ICL, n = 1.0 ± 0.2
246
+ ICL, n = 0.9 ± 0.3
247
+ 24
248
+ (mag :
249
+ 24
250
+ 8
251
+ 26 -
252
+ 8
253
+ 26 -
254
+ 28
255
+ 28 -
256
+ 30
257
+ 30 -
258
+ 100
259
+ 101
260
+ 102
261
+ 100
262
+ 101
263
+ 102
264
+ 100
265
+ 101
266
+ 102
267
+ 100
268
+ 101
269
+ 102
270
+ 167
271
+ MO01014/F105W
272
+ MO01014/F140W
273
+ 167
274
+ .
275
+ SPT0205/F105W
276
+ SPT0205/F140W
277
+ 18 -
278
+ Fitting, x = 0.77
279
+ Fitting, x = 1.72
280
+ 18 -
281
+ Fitting, x = 0.85
282
+ Fitting, x = 1.17
283
+ BCG1, n = 0.9± 0.5
284
+ BCG1, n = 1.5± 0.4
285
+ BCG, n = 0.6 ± 0.3
286
+ BCG, n = 0.7 ± 0.3
287
+ arcsec-2)
288
+ 20
289
+ BCG2, n = 0.6 ± 0.4
290
+ BCG2, n = 0.9 ± 0.4
291
+ arcsec-2)
292
+ 20
293
+ ICL,n = 3.1 ± 0.4
294
+ ICL, n = 3.7 ± 0.2
295
+ 22
296
+ ICL, n = 1.9 ± 0.4
297
+ ICL, n = 1.2 ± 0.4
298
+ (mag
299
+ 24
300
+ (mag
301
+ 24
302
+ 26 -
303
+ SB
304
+ 26
305
+ 28
306
+ 28
307
+ 30
308
+ 30 -
309
+ 100
310
+ 101
311
+ 102
312
+ 100
313
+ 101
314
+ 102
315
+ 100
316
+ 101
317
+ 102
318
+ 100
319
+ 101
320
+ 102
321
+ 167
322
+ 167
323
+ .
324
+ XMM2235/F105W
325
+ XMM2235/F160W
326
+ .
327
+ SpARCS1049/F105W
328
+ SpARCS1049/F160W
329
+ 18
330
+ Fitting, x = 4.12
331
+ - Fiting, x2 = 1.59
332
+ 18
333
+ Fitting, x = 1.10
334
+ Fitting, x2 = 1.19
335
+ BCG, n = 2.0 ± 0.2
336
+ BCG, n = 1.4 ± 0.2
337
+ BCG1, n = 1.0 ±0.1
338
+ BCG1, n = 1.1 ± 0.1
339
+ arcsec-2)
340
+ 20
341
+ ICL, n = 1.9± 0.3
342
+ ICL, n = 1.2 ± 0.4
343
+ arcsec-2)
344
+ 20
345
+ BCG2, n = 0.6 ± 0.1
346
+ BCG2, n = 0.3 ±0.1
347
+ 22
348
+ ICL,n = 0.8 ±0.3
349
+ ICL, n = 0.7 ± 0.3
350
+ 22
351
+ 24
352
+ (mag
353
+ 24
354
+ 26 -
355
+ 8
356
+ 26
357
+ 28
358
+ 28
359
+ 30
360
+ 30 -
361
+ 100
362
+ 101
363
+ 102
364
+ 100
365
+ 101
366
+ 102
367
+ 100
368
+ 101
369
+ 102
370
+ 100
371
+ 101
372
+ 102
373
+ 167
374
+ .
375
+ IDCS1426/F105W
376
+ IDCS1426/F140W
377
+ 167
378
+ .
379
+ JKCS041/F105W
380
+ JKCS041/F160W
381
+ 18-
382
+ Fiting, x = 1.27
383
+ Fitting, x = 3.39
384
+ 18
385
+ Fitting, x2 = 0.68
386
+ Fitting, x = 0.68
387
+ BCG1, n = 0.6± 0.3
388
+ BCG1, n = 1.4 ± 0.4
389
+ BCG, n = 1.8 ± 0.4
390
+ BCG, n = 1.3 ± 0.2
391
+ arcsec-2)
392
+ 20
393
+ BCG2, n = 1.3 ±0.1
394
+ BCG2, n = 1.3 ±0.0
395
+ g arcsec-2)
396
+ 20
397
+ ICL,n = 1.2 ± 0.5
398
+ ICL, n = 1.8 ± 0.4
399
+ ICL, n = 2.2 ± 0.2
400
+ 22
401
+ ICL, n = 2.2 ± 0.1
402
+ 22
403
+ 24
404
+ (mag
405
+ 24
406
+ 26 -
407
+ 26
408
+ 28
409
+ 28
410
+ 30 -
411
+ 30 -
412
+ 100
413
+ 101
414
+ 102
415
+ 100
416
+ 101
417
+ 102
418
+ 100
419
+ 101
420
+ 102
421
+ 100
422
+ 101
423
+ 102
424
+ SMA (kpc)
425
+ SMA (kpc)
426
+ SMA (kpc)
427
+ SMA (kpc)Fig. 3 | BCG + ICL radial colour profiles. Black solid lines are the observed colour. The dark grey
428
+ shades represent the 68% uncertainty. The scale radius of each component is shown with the same
429
+ colour scheme used in Fig. 2. The red data point is a representative mean value in each subregion. The
430
+ orange stars indicate the colours and positions of spectroscopic member galaxies. The light grey shade
431
+ indicates the radial extent of the BCG measured by SExtractor.
432
+
433
+
434
+
435
+ 2 -
436
+ Colour
437
+ 0-
438
+ 1
439
+ Representative value
440
+ Representative value
441
+ 100
442
+ 101
443
+ 102
444
+ 100
445
+ 101
446
+ 102
447
+ 2-
448
+ Colour
449
+ 0.
450
+ 0
451
+ Representativevalue
452
+
453
+ Representativevalue
454
+ 100
455
+ 101
456
+ 102
457
+ 100
458
+ 101
459
+ 102
460
+ 2 -
461
+ Colour
462
+ 0-
463
+ MO01014F105W-F140W
464
+ 0
465
+
466
+ Representativevalue
467
+ Representativevalue
468
+ 100
469
+ 101
470
+ 102
471
+ 100
472
+ 101
473
+ 102
474
+ 2 -
475
+ 2
476
+ Colour
477
+ 0.
478
+ XMM2235
479
+ 0
480
+ SpARCS1049
481
+ F105W-F160W
482
+ 105W-F160W
483
+
484
+ Representativevalue
485
+
486
+ Representativevalue
487
+ 100
488
+ 101
489
+ 102
490
+ 100
491
+ 101
492
+ 102
493
+ 2-
494
+ Colour
495
+ 0.
496
+ 0
497
+ Representativevalue
498
+ Representativevalue
499
+ 100
500
+ 101
501
+ 102
502
+ 100
503
+ 101
504
+ 102
505
+ SMA (kpc)
506
+ SMA (kpc)
507
+ Fig. 4 | ICL fraction evolution. Filled red circles are the current results based on the r = 0.5 Mpc
508
+ aperture. We extrapolated the best-fit Sérsic profiles to the same aperture to estimate the total ICL flux.
509
+ When we avoided the extrapolation and performed integration only within the range where the SB
510
+ profile is above the detection limit, the resulting ICL fraction is reduced by about 1.9% on average. The
511
+ mean aperture size of the literature sample is 0.58 Mpc. The dashed line and pink shade show the best-
512
+ fit linear regression and its 68% uncertainty, respectively. We weighted all data points equally and
513
+ adjusted them in such a way that the reduced χ
514
+ 2 value becomes unity. The comparison between our
515
+ high-redshift and the literature low-redshift samples shows that there is no significant evolution of the
516
+ ICL fraction with redshift in observation, which contradicts the current theoretical prediction
517
+ 1 (grey).
518
+ Although here we displayed the theoretical model that estimates the ICL fraction based on the SBC at
519
+ 26.0 mag arcsec-2, similarly steep evolutions are obtained even when different ICL definitions such as
520
+ binding energy criteria are used.
521
+
522
+
523
+
524
+ 50
525
+ 11j
526
+ [12]
527
+ 13)
528
+ 40
529
+ [19]
530
+ [21]
531
+ This study
532
+ 30
533
+ ICL fraction (%)
534
+ 20
535
+ 10
536
+ 0
537
+ 0
538
+ 0.2
539
+ 0.4
540
+ 0.6
541
+ 0.8
542
+ 1.0
543
+ 1.2
544
+ 1.6
545
+ 2.0
546
+ zMethods
547
+ Target selection
548
+ We searched the Mikulski Archive for Space Telescope for the WFC3-near-infrared imaging
549
+ programs that have observed z ≳ 1 clusters in at least two filters with the surface brightness limit μ ≈
550
+ 28 mag arcsec-2. We excluded the targets if they do not possess any distinct BCGs, or very bright
551
+ stars/foreground galaxies are present near the cluster centres. The search resulted in a sample of ten
552
+ galaxy clusters at 1 ≲ z ≲ 2 with a minimum (maximum) redshift of 0.98 (1.803). Extended Data Table
553
+ 1 summarizes our target selection, including the redshift, coordinate, program number, surface
554
+ brightness limit and so on. Although F105W exists for all ten clusters, either F140W or F160W is
555
+ available for each cluster.
556
+ Reduction pipeline optimized for ICL measurement
557
+ The common data reduction procedure recommended in the HST Data Handbook
558
+ 34 works well
559
+ if one is interested in discrete astronomical sources such as stars and galaxies. However, when we are
560
+ looking for signals from diffuse components whose surface brightness is within a subpercentage of the
561
+ sky brightness and slowly varies across the detector, additional care is needed. Our reduction begins
562
+ with the flat-fielded (FLT) images processed by the Space Telescope Science Institute calwf3 tool36,
563
+ which removes most instrumental signatures of WFC3, except for geometric distortion. We visually
564
+ inspected these FLT images and manually masked out any remaining artefacts such as satellite/asteroid
565
+ trails. We ran the TweakReg package
566
+ 37 for astrometric calibration by finding common astronomical
567
+ sources. The FLT images are already flatfielded with the default Space Telescope Science Institute
568
+ composite flats, which are claimed to be accurate within less than 0.5%, except for the region within
569
+ 128 pixels of the detector edge34. This claim has been verified by independently constructing residual
570
+ flats utilizing large WFC3 survey programs4. We also investigated the impact of the residual flat in the
571
+ final mosaic by performing drizzling as if we were stacking science frames and found that the dithered
572
+ residual flats would cause at most around 0.4% errors, which is already negligible compared to other
573
+ sources of errors (for example, background determination). In this study, we applied these residual flats
574
+ to our FLT data to further reduce the residual flat errors. The application of the aforementioned residual
575
+ flatfielding cannot remove large-scale sky gradients arising from intrinsic sky gradient, detector
576
+ persistence, internal reflection and so on. We removed the sky gradient by fitting a first-order
577
+ polynomial plane (F(x,y) = ax + by + c) to the object-masked residual-flatfielded image and subtracting
578
+ the result. The subtraction result was visually inspected, and we discarded the frame if the first-order
579
+ polynomial plane could not adequately describe the sky gradient. To create a final deep mosaic where
580
+
581
+ ICL is measured, careful and consistent sky subtraction from each exposure is required. In typical
582
+ ground-based data reduction designed for non-ICL-related studies, position-dependent sky
583
+ estimation/subtraction is routinely performed after astronomical objects are masked out. Although this
584
+ scheme may provide cosmetically ‘good’ results when different frames are combined together, the
585
+ inevitable consequence is sky oversubtraction in the region where non-discrete astronomical
586
+ components are dominant. Therefore, in this study, we determined only the global sky level (that is, a
587
+ single constant) for each frame and subtracted it. Because individual exposures observed with dithers
588
+ cover non-identical areas around the target, it is necessary to choose the same physical sky region that
589
+ is present in common for all exposures. To accomplish this, we set up the largest possible annulus that
590
+ is approximately centred on the BCG and is observed by all exposures. An illustration of this common
591
+ annulus is presented in Extended Data Fig. 1.
592
+ Discrete astronomical sources in this annulus were detected using SExtractor
593
+ 38 with the
594
+ settings of DETECT_MINAREA = 5 and DETECT_ THRESHOLD = 2. As a matter of course, the
595
+ resulting segmentation map fails to include the contributions from the faint diffuse wings of the objects.
596
+ To address this, we gradually expanded the segmentation maps and investigated the resulting
597
+ background level change as a function of the mask size. One should not increase the segmentation map
598
+ by expanding the boundaries at the same rate for all objects. That is, the segmentation boundaries of
599
+ compact objects should expand slower than those of extended objects. Thus, we used SExtractor’s half-
600
+ light radius rh and scaled the expansion with it. The width (w) of the expansion band is determined by
601
+ w = ce rh, where ce is the expansion coefficient. We stopped the segmentation map expansion at ce = 6,
602
+ beyond which the sky estimation converged (Extended Data Fig. 2). Note that this exposure-by-
603
+ exposure sky level estimation is performed on the individual drizzled images to minimize the impact of
604
+ the geo- metric distortion. We created the final mosaic using AstroDrizzle
605
+ 33 with its sky subtraction
606
+ option turned off. The output pixel scale is set to 0.05 arcsec. We used the Gaussian kernel for drizzling.
607
+ Extended Data Fig. 3 schematically summarizes our data reduction pipeline.
608
+ Object masking for ICL measurement
609
+ To characterize ICL from the mosaic image, it is necessary to mask out light from discrete
610
+ objects. In the background level estimation above, we found that the choice ce = 6 was sufficient.
611
+ However, this large expansion coefficient cannot be used in the central region of the cluster because
612
+ doing so would leave very few pixels there, resulting in too large statistical errors exceeding the
613
+ systematic error caused by the incomplete masking (Extended Data Fig. 2). Therefore, for ICL
614
+ measurement from the final stack, we chose to employ ce = 2 for object masking and apply a correction
615
+ factor to obtain the result effectively measured with the full mask (ce = 6). The correction factor is
616
+
617
+ derived by comparing multiple ICL profiles measured with different ce values. We masked out every
618
+ discrete source detected by SExtractor except for the BCG. We also visually scanned the result and
619
+ applied manual masking for the objects that SExtractor failed to identify (Fig. 1).
620
+ To assess the validity of our statistical correction scheme, we repeated our analysis with the
621
+ full (ce = 6) masking. This is supposed to generate results with better accuracy at the expense of
622
+ precision (that is, smaller systematic errors and larger statistical errors). However, if any large
623
+ systematic discrepancy from our fiducial measurement (for example, statistically corrected result after
624
+ the use of ce = 2) is found, this indicates that the aforementioned correction scheme is problematic. The
625
+ mean of the differences is close to zero (about 0.02%, Extended Data Table 2), which verifies that our
626
+ correction scheme with the use of the moderate masking expansion (ce = 2) is robust.
627
+ Radial profile measurement with elliptical binning
628
+ As the BCG + ICL isophotes are elliptical, the use of circular binning would spread the BCG-
629
+ ICL transition over multiple bins. Thus, we measured the radial profile of the BCG + ICL using an
630
+ elliptical bin- ning scheme. To determine the ellipticity and position angle (PA) of the ellipse, we used
631
+ the AutoProf package39. With both detection image and masking map as inputs, AutoProf calculates
632
+ ellipticity and PA based on isophotal fitting and Fourier analysis.
633
+ Two sets of outputs are generated by AutoProf. One is a series of ellipticity and PA values,
634
+ which vary with radius. The other is a single pair of ellipticity and PA, which represents the global
635
+ shape. The former is useful when one’s interest is the radius-dependent isophotal shape of high signal-
636
+ to-noise-ratio objects. In this study, where our scientific interest is faint diffuse light, we use the second
637
+ set of outputs.
638
+ We defined a series of semi-major axes (SMA) with a logarithmic scale and measured the
639
+ surface brightness at each radial bin. We applied the 3σ clipping method to minimize the impact of the
640
+ outliers and adopt the median as the representative surface brightness of the bin. This also reduces the
641
+ effects of any potential unidentified substructures within each elliptical annulus. The total error of the
642
+ surface brightness estimate is computed as the quadratic sum of the 1σ photon noise and sky estimation
643
+ (background level) uncertainty. As mentioned above, the residual flat error is negligible and hence is
644
+ not included here. The latter dominates our error budget.
645
+ Multicomponent decomposition
646
+
647
+ The traditional method for measuring ICL is to define an SBC and char- acterize the light
648
+ component fainter than the threshold. As the choice of the threshold is arbitrary, it is difficult to use the
649
+ method to compare results from different studies. In particular, because the current redshift regime (1
650
+ ≲ z ≲ 2) is considerably different from those of the previous studies, one cannot objectively characterize
651
+ the ICL properties based on this traditional approach.
652
+ In this study, we decompose the BCG + ICL profile into multiple Sérsic components. A Sérsic
653
+ profile
654
+ 40 is defined as:
655
+ 𝐼(𝑟) = 𝐼! exp +−𝑏" ./
656
+ #
657
+ #!0
658
+ $/"
659
+ − 123,
660
+ where r, re, n and Ie are the radius, half right radius, Sérsic index and intensity at the half right radius,
661
+ respectively. In this equation, bn is the constant that is not independent and is solely determined by n.
662
+ Then, the BCG + ICL profile S(r) is modelled as a superposition of PSF-convolved multiple
663
+ Sérsic components Im(r): S(r)=PSF(r)⁎∑mIm(r), where PSF(r) is the PSF radial profile and the symbol
664
+ ‘*’ represents the convolution. The PSF radial profile was constructed by combining the core from
665
+ observed stellar images and the wing from the TinyTim
666
+ 41 result. If the PSF correction procedure is
667
+ omitted, the mean ICL fraction increases by around 1.5% (Extended Data Table 2).
668
+ How do we determine the total number of Sérsic components for each cluster? A model based
669
+ on more components has higher degrees of freedom and leads to smaller residuals. However, the
670
+ drawback is overfitting. About 70% of elliptical galaxies are reported to require more than one Sérsic
671
+ component to adequately describe their profiles
672
+ 42–44 In this study, we limit the maximum number of
673
+ Sérsic components for the description of the BCG + ICL profile to three.
674
+ To determine the optimal number of Sérsic components, we use two criteria. The first is the
675
+ Bayes factor K45, defined as follows:
676
+ 𝐾 =
677
+ &((|*")
678
+ &((|*#),
679
+ where p(D|Mk) is the probability of the data (D) given the model Mk. As P(D|Mk) cannot be computed
680
+ directly, in practice K is evaluated as follows:
681
+ 𝐾 =
682
+ &(*"|()&(*#)
683
+ &(*#|()&(*") =
684
+ &(*"|()
685
+ &(*#|(),
686
+
687
+ where we assume the equality between the two priors p(M1) and p(M2). The Bayes factor K informs us
688
+ of how the first model M1 is preferred over the second M2. Its outstanding advantage is that K inherently
689
+ penalizes the model according to its degrees of freedom based on first principles. The second criterion
690
+ is the number of inflexion points in the derivative of the SB radial profile. Naturally, the existence of l
691
+ inflexion points implies a preference towards a model with l + 1 components. After investigation of our
692
+ ten clusters with the above two criteria, we find that for the entire sample (1) the Bayes factor K between
693
+ the best and second-best models is ln(K) ≳ 2, (2) the optimal number of Sérsic components inferred
694
+ from the Bayes factor agrees with the result from the inflexion point analysis and (3) the SB profiles
695
+ require either two or three components for optimal decomposition.
696
+ Estimation of the total cluster luminosity
697
+ One of the key requirements in measuring the ICL fraction is a robust selection of the cluster
698
+ member galaxies. In this study, the first step towards this goal is the compilation of the results from
699
+ previous spec- troscopic studies (see Extended Data Table 3). We used this spectroscopic cluster
700
+ member catalogue to define the initial red sequence locus from the colour–magnitude diagram. By
701
+ iteratively applying linear regression to the red sequence and selecting objects within 0.5 mag from the
702
+ best-fit line, we built up the second-stage cluster member catalogue. In this iteration, we removed stars
703
+ using the CLASS_STAR value reported by SExtractor and the objects either brighter than the BCG or
704
+ fainter than F105W = 26 mag. This second-stage cluster member catalogue needed to be improved
705
+ because the distribution of the object distance from the best-fit red sequence line was asymmetric (the
706
+ blue side was blended with the neighbouring blue cloud). Thus, we fitted a double Gaussian model to
707
+ the distribution (see the bottom panel of Extended Data Fig. 4). The new centre and 1σ width of the red
708
+ sequence were used to update the intercept and width of our previous best-fit linear regression result
709
+ and the final cluster member catalogue was obtained (see the top panel of Extended Data Fig. 4).
710
+ Although we employed a sophisticated procedure for selection of cluster members, inevitably the
711
+ method is designed to select only red members, except for the blue spectroscopic members. If the
712
+ contribution from the blue members is large, our ICL fraction would be overestimated. However, we
713
+ argue that the overestimation, if any, would not be significant because (1) in most cases the
714
+ spectroscopic catalogue includes the brightest blue cluster members (Extended Data Table 3), (2) even
715
+ the brightest blue cluster members are found to be still a few magnitudes fainter than the BCG and other
716
+ brightest red members and (3) some fraction of galaxies in our red sequence catalogue are non-cluster
717
+ members. We verified this claim by utilizing the publicly available photometric redshift catalogue for
718
+ SPT0205 (ref. 46), which includes the blue cluster member candidates. When we repeated the
719
+ measurement of the ICL fraction with it, the resulting ICL fraction shifted by only 1%.
720
+
721
+ The total luminosity is estimated as follows. First, we masked out non-member galaxies/stars
722
+ from our imaging data. The total masked area is non-negligible and simply assigning zero flux to the
723
+ area would lead to substantial underestimation of the total luminosity. Thus, we filled the masked
724
+ regions with the predicted flux from our best-fit multi-Sérsic model. Finally, the total luminosity is
725
+ computed by the summation of the pixel values of the resulting image, which is comprised of the flux
726
+ from the BCG, ICL and cluster members.
727
+ Impact of the red sequence selection criteria
728
+ Although we took care to robustly define the locus of the red sequence through somewhat
729
+ sophisticated iteration, two of the remaining ambiguities worth further investigation are the faint-end
730
+ limit and the width of the red sequence. To examine the dependence of the ICL fraction on these
731
+ selection criteria, we considered the following three additional cases:
732
+ • Test A: the magnitude limit decreased to 24th mag
733
+ • Test B: the magnitude limit increased to 28th mag
734
+ • Test C: the width increased to 1.5σ
735
+ We list the test results for individual clusters in Extended Data Table 3. In the case of Test A,
736
+ the average (maximum) increase in ICL fraction is found to be approximately 1.9% (approximately
737
+ 3.8%). Test B shows that the average (maximum) decrease in ICL fraction is approximately 0.9%
738
+ (approximately 1.6%). Finally, Test C gives an average (maximum) decrease of approximately 2.9%
739
+ (approximately 5.8%). In summary, our ICL fraction measurements are not sensitive to the selection
740
+ criteria tested here.
741
+ ICL fraction versus mass correlation
742
+ Together with the ICL fraction evolution with redshift, the ICL community has also been
743
+ investigating the correlation between ICL fraction and halo mass. Under the assumption that massive
744
+ halos represent older populations, a strong correlation would imply a significant time evolution.
745
+ Extended Data Fig. 5 shows the ICL fraction versus mass relation for our sample. The masses come
746
+ from weak lensing studies (Extended Data Table 3). No significant correlation is observed for our
747
+ sample.
748
+
749
+
750
+
751
+ Online content Any methods, additional references, Nature Portfolio reporting summaries, source
752
+ data, extended data, supplementary information, acknowledgements, peer review information, details
753
+ of author contributions and competing interests, and statements of data and code availability are
754
+ available at https://doi.org/10.1038/s41586-022-05396-4.
755
+ Data availability The raw HST near-infrared imaging data used for the current study are publicly
756
+ available. The processed imaging data are available on the github repository at
757
+ https://github.com/Hyungjin-Joo/High_z_ICL. Source data are provided with this paper.
758
+ Code availability An exhaustive repository of code for our custom data processing and analyses
759
+ reported in this manuscript are available on the github repository at https://github.com/Hyungjin-
760
+ Joo/High_z_ICL.
761
+ Acknowledgements This study is based on observations created with NASA/ESA Hubble Space
762
+ Telescope and downloaded from the Mikulski Archive for Space Telescope at the Space Telescope
763
+ Science Institute. The current research is supported by the National Research Foundation of Korea
764
+ under programme 2022R1A2C1003130 and the Yonsei Future-Leading Research Initiative programme.
765
+ Author contributions M.J.J. conceived, designed and supervised the project. M.J.J. and H.J. analysed
766
+ the Hubble Space Telescope imaging data, developed the pipeline, interpreted the results and wrote the
767
+ manuscript.
768
+
769
+
770
+ Extended Data Fig. 1 | Definition of common sky areas. (A) Exposure map for the single-frame
771
+ image. (B) Same as (C) except that it is for the mosaic image. (C) Science image for single frame. (D)
772
+ Same as (C) except that it is for the mosaic image. The pink circular region in (A) is the region that is
773
+ observed in common by all contributing frames. (B) shows how this common region is positioned in
774
+ one of the input frames. As the central region of this circle is likely to be heavily influenced by the ICL,
775
+ we excluded the central region and instead defined the annulus shown in (C) and (D) to estimate the
776
+ background level.
777
+
778
+
779
+
780
+ (A)
781
+ exposuretime along blue horizontal line
782
+ (B)
783
+ line
784
+ horizontal
785
+ blue
786
+ along
787
+ exposuretime
788
+ exposuretimealongblueverticalline
789
+ exposuretimealongblueverticalline
790
+ C
791
+ Extended Data Fig. 2 | Masking size growth and impacts on background level. (A), (B) and (C)
792
+ illustrate our scheme for masking size growth from the original to the ce = 2 and ce = 6 cases. Note that
793
+ we exhaust pixels for ICL measurement at ce = 6. In (D), we show how the background level (green)
794
+ changes as we vary the masking size using the expansion coefficient for a single exposure. We observe
795
+ that at ce ≳ 6 the measurement converges (red). The black solid line indicates the result when instead
796
+ we use a 3σ clipping algorithm without considering the diffuse wings of the astronomical objects. The
797
+ yellow line shows the surface brightness level measured at each ce. (E) is the same as the left except
798
+ that the measurement is from the final deep stack. Solid lines indicate the median value and shaded
799
+ regions show the 68% uncertainty. As the image is deeper, the number of pixels discarded (masked out)
800
+ at the same ce value is much greater.
801
+
802
+ Original_mask
803
+ Ce = 2.0
804
+ Ce = 6.0
805
+ (A)
806
+ (B)
807
+ (C)
808
+ 1.30
809
+ (D)
810
+ Sky
811
+ All Light Sources
812
+ sky value of this methid
813
+ SB [count / s / pixe/2]
814
+ 1.25
815
+ sky value of 3-sigma clipping
816
+ 1.20
817
+ 1.15
818
+ 2
819
+ 4
820
+ 6
821
+ 8
822
+ 10
823
+ 12
824
+ Ce
825
+ (E)
826
+ Sky
827
+ 0.0
828
+ All Light Sources
829
+ [z/axid / s / sunolas
830
+ 2.5
831
+ 5.0
832
+ 7.5
833
+ 10.0
834
+ 2
835
+ 4
836
+ 6
837
+ 8
838
+ 10
839
+ 12
840
+ Ce
841
+ Extended Data Fig. 3 | Schematic diagram of our ICL-oriented data reduction. Dark grey
842
+ rectangles show the steps where external packages are used, while light grey rectangles illustrate our
843
+ custom procedures. Parallelograms represent the input/output data.
844
+
845
+
846
+
847
+ Reject
848
+ Start
849
+ no
850
+ Astrometric
851
+ Residual
852
+ Fitting
853
+ *flt.fits
854
+ Calibration
855
+ Plane fitting
856
+ Flat Fielding
857
+ Success?
858
+ (TweakReg)
859
+ yes
860
+ *flt_plf.fits
861
+ Segmentation
862
+ Expanded
863
+ Source Detection
864
+ Map
865
+ ExpandingMask
866
+ Mask
867
+ (SExtractor)
868
+ (For masking)
869
+ Map
870
+ Image Drizzling
871
+ High Overlapped
872
+ Overlapped
873
+ Drizzled Image
874
+ SkyEstimation
875
+ (Astrodrizzle)
876
+ Region Selection
877
+ WCS
878
+ Sky Level
879
+ Image Drizzling
880
+ Final
881
+ ?ReDrizzledImage
882
+ Sky Subtraction
883
+ End
884
+ (Astrodrizzle)
885
+ Drizzled Image
886
+ Extended Data Fig. 4 | Red sequence selection scheme. Here we display the case for SPT2106. (A)
887
+ Colour–magnitude diagram. Black dots are all sources detected by SExtractor. The red dots represent
888
+ the spectroscopic members, whereas the orange dots are our red sequence candidates. The BCG is
889
+ indicated with a red star. The red dashed line shows the final, best-fit red sequence. The dot-dashed
890
+ lines bracket the 68% distribution. (B) Distribution of the F105W < 26 object distances from the best-
891
+ fit red sequence. The green line shows the best-fit double Gaussian models. The yellow line illustrates
892
+ a single Gaussian component, which represents the distribution of the red sequence candidates.
893
+
894
+
895
+
896
+ (A)
897
+ 0.75
898
+ F105W - F140W (Auto)
899
+ 0.50
900
+ 0.25
901
+ Red sequence
902
+ 1 s.t.d
903
+ 0.00
904
+ All
905
+ BCG
906
+ -0.25
907
+ Spec. Member
908
+ Member Candidate
909
+ -0.50
910
+ 18
911
+ 19
912
+ 20
913
+ 21
914
+ 22
915
+ 23
916
+ 24
917
+ 25
918
+ 26
919
+ 27
920
+ (B)
921
+ F105W (ISO)
922
+ Line
923
+ 0.50
924
+ d Sequence
925
+ 0.25
926
+ 0.00
927
+ Distance from Red
928
+ PDF
929
+ -0.25
930
+ DoubleGauss
931
+ Gauss
932
+ -0.50
933
+ Red Sequence
934
+ -0.75
935
+ 1 s.t.d
936
+ All
937
+ 1.00
938
+ 2.5
939
+ 2.0
940
+ 1.5
941
+ 1.0
942
+ 0.5
943
+ 0.0
944
+ Number of Galaxies
945
+ Extended Data Fig. 5 | Comparison between ICL fraction and cluster mass. The mass comes from
946
+ weak lensing studies. No significant correlation between ICL fraction and mass is found.
947
+
948
+
949
+
950
+ 35
951
+ 30 -
952
+ 25
953
+ . fraction [%]
954
+ 20
955
+ T
956
+ 15
957
+ ICL
958
+ 10
959
+ T
960
+ T
961
+ 0
962
+ 0.5
963
+ 1.0
964
+ 1.5
965
+ 2.0
966
+ 2.5
967
+ M200[Mo]
968
+ 1e15Extended Data Table 1 | Target List
969
+ Target
970
+ (Short Name)
971
+ Redshift
972
+ (z)
973
+ R.A
974
+ Dec
975
+ HST Proposal
976
+ ID
977
+ SB limit [mag / arcs2]
978
+ (Exposure Time [s])
979
+ F150W
980
+ F140W
981
+ F160W
982
+ XDCP J1229+0151
983
+ (XMM1229)
984
+ 0.98
985
+ 12:29:28
986
+ +01:51:34
987
+ 12501
988
+ 29.62
989
+ (1,311.7)
990
+ -
991
+ 29.12
992
+ (1,111.7)
993
+ SPT-CL J2106-5844
994
+ (SPT2106)
995
+ 1.1312
996
+ 21:06:05
997
+ -58:44:42
998
+ 13677, 14327
999
+ 28.26
1000
+ (12,567.7)
1001
+ 28.59
1002
+ (12,771.6)
1003
+ -
1004
+ MOO J1142_1529
1005
+ (MOO1142)
1006
+ 1.19
1007
+ 11:42:46
1008
+ +15:27:14
1009
+ 14327
1010
+ 27.96
1011
+ (6,283.8)
1012
+ 28.26
1013
+ (6,983.8)
1014
+ -
1015
+ RDCS J1252-2927
1016
+ (RDCS1252)
1017
+ 1.237
1018
+ 12:52:57
1019
+ -29:27:15
1020
+ 12501
1021
+ 28.26
1022
+ (1,211.7)
1023
+ -
1024
+ 29.00
1025
+ (1,211.7)
1026
+ MOO J1014+0038
1027
+ (MOO1014)
1028
+ 1.24
1029
+ 10:14:08
1030
+ +00:38:26
1031
+ 13677, 14327
1032
+ 28.47
1033
+ (18,255.5)
1034
+ 28.63
1035
+ (17,810.1)
1036
+ -
1037
+ SPT-CL J0205-5829
1038
+ (SPT0205)
1039
+ 1.322
1040
+ 02:05:46
1041
+ -58:29:06
1042
+ 13677, 14327
1043
+ 28.66
1044
+ (23,007.2)
1045
+ 28.84
1046
+ (25,052.9)
1047
+ -
1048
+ XDCP J2235-2557
1049
+ (XMM2235)
1050
+ 1.39
1051
+ 22:35:21
1052
+ -25:57:25
1053
+ 12501
1054
+ 28.18
1055
+ (1,211.7)
1056
+ -
1057
+ 27.85
1058
+ (1,211.7)
1059
+ SpARCS J1049+5640
1060
+ (SpARCS1049)
1061
+ 1.71
1062
+ 10:49:22
1063
+ +56:40:34
1064
+ 13677, 13747
1065
+ 28.52
1066
+ (8,543.3)
1067
+ -
1068
+ 28.28
1069
+ (9,237.4)
1070
+ IDCS J1426.5+3508
1071
+ (IDCS1426)
1072
+ 1.75
1073
+ 14:26:33
1074
+ +35:05:24
1075
+ 12203, 13677,
1076
+ 14327
1077
+ 30.56
1078
+ (10,972.4)
1079
+ 30.79
1080
+ (11,225.4)
1081
+ -
1082
+ JKCS041
1083
+ 1.803
1084
+ 05:26:44
1085
+ +04:41:37
1086
+ 12927
1087
+ 27.73
1088
+ (2,670.6)
1089
+ -
1090
+ 27.99
1091
+ (4,509.4)
1092
+
1093
+
1094
+
1095
+
1096
+ Extended Data Table 2 | ICL fractions and impact of various systematics
1097
+ Name
1098
+ Filter
1099
+ fBCG+ICL [%]
1100
+ fICL [%]
1101
+ Red Sequence Selection Criteria
1102
+ Unmasked
1103
+ Wings
1104
+ No PSF
1105
+ Test A
1106
+ Test B
1107
+ Test C
1108
+
1109
+
1110
+ (1)
1111
+ (2)
1112
+ (3)
1113
+ (4)
1114
+ (5)
1115
+ (6)
1116
+ (7)
1117
+ XMM1229
1118
+ F105W
1119
+ 13.6$%.'
1120
+ (%.)
1121
+ 11.7$%.*
1122
+ (%.*
1123
+ 0.8
1124
+ -0.3
1125
+ -0.8
1126
+ -0.2
1127
+ 2.5
1128
+
1129
+ F160W
1130
+ 14.2$%.+
1131
+ (%.+
1132
+ 11.3$%.'
1133
+ (%.+
1134
+ 0.7
1135
+ -0.3
1136
+ -0.7
1137
+ -0.2
1138
+ 2.8
1139
+ SPT2106
1140
+ F105W
1141
+ 18.9$,.-
1142
+ (,.-
1143
+ 14.2$,..
1144
+ (,..
1145
+ 0.1
1146
+ -0.9
1147
+ -2.8
1148
+ 0.5
1149
+ 0.3
1150
+
1151
+ F140W
1152
+ 21.4$/.%
1153
+ (/./
1154
+ 14.5$/.+
1155
+ (/.,
1156
+ 0.1
1157
+ -0.8
1158
+ -2.7
1159
+ 0.3
1160
+ 0.6
1161
+ MOO1142
1162
+ F105W
1163
+ 21.1$+.%
1164
+ (+.+
1165
+ 15.9$+.+
1166
+ (+.+
1167
+ 0.9
1168
+ -0.4
1169
+ -3.0
1170
+ 0.8
1171
+ -0.7
1172
+
1173
+ F140W
1174
+ 24.6$+.+
1175
+ (+.'
1176
+ 16.8$+.*
1177
+ (+.*
1178
+ 1.2
1179
+ -0.7
1180
+ -4.3
1181
+ 0.9
1182
+ 3.7
1183
+ RDCS1252
1184
+ F105W
1185
+ 27.9$%.*
1186
+ (%..
1187
+ 21.7$,.,
1188
+ (,./
1189
+ 3.8
1190
+ -1.3
1191
+ -3.9
1192
+ -4.7
1193
+ 0.8
1194
+
1195
+ F160W
1196
+ 29.1$,./
1197
+ (,.'
1198
+ 24.1$,.0
1199
+ (/.,
1200
+ 2.8
1201
+ -0.9
1202
+ -3.1
1203
+ -3.5
1204
+ -0.1
1205
+ MOO1014
1206
+ F105W
1207
+ 22.5$'.+
1208
+ ('.*
1209
+ 20.4$'.'
1210
+ ('.*
1211
+ 1.4
1212
+ -0.5
1213
+ -1.9
1214
+ 1.1
1215
+ -0.5
1216
+
1217
+ F140W
1218
+ 17.8$/..
1219
+ (/.0
1220
+ 15.2$+.%
1221
+ (+.,
1222
+ 1.4
1223
+ -0.5
1224
+ -1.8
1225
+ 0.1
1226
+ 5.0
1227
+ SPT0205
1228
+ F105W
1229
+ 17.8$,.0
1230
+ (/.+
1231
+ 16.9$/.%
1232
+ (/.+
1233
+ 1.9
1234
+ -0.5
1235
+ -3.2
1236
+ 0.1
1237
+ -3.2
1238
+
1239
+ F160W
1240
+ 20.7$%..
1241
+ (%.0
1242
+ 19.6$%.*
1243
+ (%..
1244
+ 2.0
1245
+ -0.5
1246
+ -3.3
1247
+ 0.2
1248
+ -4.2
1249
+ XMM2235
1250
+ F105W
1251
+ 23.0$/.0
1252
+ (+.*
1253
+ 20.7$/..
1254
+ (+.)
1255
+ 3.0
1256
+ -1.4
1257
+ -2.6
1258
+ 2.9
1259
+ 0.3
1260
+
1261
+ F160W
1262
+ 25.8$/.0
1263
+ (/.0
1264
+ 22.9$/.0
1265
+ (+.*
1266
+ 3.3
1267
+ -1.6
1268
+ -2.7
1269
+ 2.9
1270
+ 2.5
1271
+ SpARCS1049
1272
+ F105W
1273
+ 27.3$/.*
1274
+ ('.+
1275
+ 19.7$/.*
1276
+ ('.+
1277
+ 2.4
1278
+ -1.2
1279
+ -5.8
1280
+ -1.4
1281
+ 4.5
1282
+
1283
+ F160W
1284
+ 26.6$/.*
1285
+ ('.%
1286
+ 13.3$/.0
1287
+ (+..
1288
+ 2.5
1289
+ -1.1
1290
+ -5.2
1291
+ -2.0
1292
+ 9.7
1293
+ IDCS1426
1294
+ F105W
1295
+ 22.8$,.,
1296
+ (,./
1297
+ 19.7$,.,
1298
+ (,./
1299
+ 2.2
1300
+ -0.7
1301
+ -3.5
1302
+ 0.3
1303
+ -1.0
1304
+
1305
+ F140W
1306
+ 28.4$%..
1307
+ (%..
1308
+ 22.2$%.0
1309
+ (%.0
1310
+ 1.9
1311
+ -0.7
1312
+ -3.5
1313
+ -0.8
1314
+ 3.2
1315
+ JKCS041
1316
+ F105W
1317
+ 7.5$%.0
1318
+ (,.-
1319
+ 5.7$,./
1320
+ (,.+
1321
+ 1.7
1322
+ -1.0
1323
+ -1.4
1324
+ 1.2
1325
+ 1.6
1326
+
1327
+ F160W
1328
+ 16.6$%.0
1329
+ (%..
1330
+ 12.3$/.%
1331
+ (,.+
1332
+ 3.1
1333
+ -1.5
1334
+ -2.4
1335
+ 1.9
1336
+ 0.4
1337
+ Average
1338
+ -
1339
+ 21.38
1340
+ 16.89
1341
+ 1.86
1342
+ -0.89
1343
+ -2.93
1344
+ 0.02
1345
+ 1.46
1346
+ (1) BCG + ICL fraction. (2) Fiducial ICL fraction. (3) Change when the magnitude limit decreased to
1347
+ 24th mag. (4) Change when magnitude limit increased to 28th mag. (5) Change when the width of the
1348
+ red sequence increased to 1.5σ. (6) Change when the expansion coefficient increased to ce = 6. (7)
1349
+ Change when the PSF effect is neglected. For (1) and (2), we quote measurements at r = 0.5 Mpc.
1350
+
1351
+
1352
+
1353
+ Extended Data Table 3 | Weak lensing mass, the number of spectroscopic member galaxies and
1354
+ their references
1355
+ Name
1356
+ Weak
1357
+ Lensing
1358
+ Mass (10$,𝑀⊙)
1359
+ Number
1360
+ of
1361
+ Spectroscopic
1362
+ Members
1363
+ Spectroscopic
1364
+ Catalogue
1365
+ Reference
1366
+ XMM1229
1367
+ 2.04./.12
1368
+ 3$.,1
1369
+ 17
1370
+ 47
1371
+ SPT2106
1372
+ 14.9.4.5,
1373
+ 3,.$1
1374
+ 31
1375
+ 46
1376
+ MOO1142
1377
+ 5.69.$.,4
1378
+ 3$.56
1379
+ 8
1380
+ 48
1381
+ RDCS1252
1382
+ 15.7.7.6/
1383
+ 37.,/
1384
+ 22
1385
+ 49
1386
+ MOO1014
1387
+ 3.35.$./2
1388
+ 37./5
1389
+ 7
1390
+ 50
1391
+ SPT0205
1392
+ 2.00./.25
1393
+ 3/.82
1394
+ 21
1395
+ 46
1396
+ XMM2235
1397
+ 14.6.,.82
1398
+ 31.26
1399
+ 10
1400
+ 51
1401
+ SpARCS1049
1402
+ 3.50.$.7/
1403
+ 3$.7/
1404
+ 11
1405
+ 52
1406
+ IDCS1426
1407
+ 3.65.$.65
1408
+ 37.5/
1409
+ 6
1410
+ 53
1411
+ JKCS041
1412
+ 13.1.5.8/
1413
+ 31.12
1414
+ 17
1415
+ 54
1416
+
1417
+
1418
+
1419
+
1420
+ References
1421
+ 1. Rudick, C. S., Mihos, J. C. & McBride, C. K. The quantity of intracluster light: comparing
1422
+ theoretical and observational measurement techniques using simulated clusters. Astrophys. J. 732,
1423
+ 48–64 (2011).
1424
+ 2. Contini, E., De Lucia, G., Villalobos, Á. & Bogani, S. On the formation and physical properties of
1425
+ the intracluster light in hierarchical galaxy formation models. Mon. Not. R. Astron. Soc. 437, 3787–
1426
+ 3802 (2014).
1427
+ 3. Burke, C., Collins, C. A., Stott, J. P. & Hilton, M. Measurement of the intracluster at z ~ 1. Mon.
1428
+ Not. R. Astron. Soc. 425, 2058–2068 (2012).
1429
+ 4. Ko, J. & Jee, M. J. Evidence for the existence of abundant intracluster light at z = 1.24. Astrophys.
1430
+ J. 862, 95–103 (2018).
1431
+ 5. Montes, M. & Trujillo, I. Intracluster light at the Frontier - II. The Frontier Fields Clusters. Mon.
1432
+ Not. R. Astron. Soc. 474, 917–932 (2018).
1433
+ 6. DeMaio, T. et al. The growth of brightest cluster galaxies and intracluster light over the past 10
1434
+ billion years. Mon. Not. R. Astron. Soc. 491, 3751–3759 (2020).
1435
+ 7. Gonzalez, A. H. et al. Galaxy cluster baryon fractions revisited. Astrophys. J. 778, 14–29 (2013).
1436
+ 8. Presotto, V. et al. Intracluster light properties in the CLASH-VLT cluster MACS J1206.2- 0847.
1437
+ Astron. Astrophys. 565, A126 (2014).
1438
+ 9. Contini, E., Yi, S. K. & Kang, X. Theoretical predictions of colors and metallicity of the intracluster
1439
+ light. Astrophys. J. 871, 24–33 (2019).
1440
+ 10. Burke, C., Hilton, M. & Collins, C. Coevolution of brightest cluster galaxies and intracluster light
1441
+ using CLASH. Mon. Not. R. Astron. Soc. 449, 2353–2367 (2015).
1442
+ 11. Morishita, T. et al. Characterizing intracluster light in the Hubble Frontier Fields. Astrophys. J. 846,
1443
+ 139–151 (2017).
1444
+ 12. Almao-Martinez, K. A. & Blakeslee, J. P. Specific frequencies and luminosity profiles of cluster
1445
+ galaxies and intracluster light in Abell 1689. Astrophys. J. 849, 6–24 (2017).
1446
+ 13. Ellien, A. et al. The complex case of MACS J0717.5+6745 and its extended filament: intra-cluster
1447
+ light, galaxy luminosity function, and galaxy orientations. Astron. Astrophys. 628, A34 (2019).
1448
+ 14. Feldmeier, J. J. et al. Intracluster planetary nebulae in the Virgo Cluster. III. Luminosity of the
1449
+ intracluster light and tests of the spatial distribution. Astrophys. J. 615, 196–208 (2004).
1450
+ 15. Griffiths, A. et al. MUSE spectroscopy and deep observations of a unique compact JWST target,
1451
+ lensing cluster CLIO. Mon. Not. R. Astron. Soc. 475, 2853–2869 (2018).
1452
+ 16. Jee, M. J. Tracing the peculiar dark matter structure in the galaxy cluster Cl 0024+17 with
1453
+ intracluster stars and gas. Astrophys. J. 717, 420–434 (2010).
1454
+
1455
+ 17. Jimenez-Teja, Y. et al. Unveiling the dynamical state of massive clusters through the ICL fraction.
1456
+ Astrophys. J. 857, 79–96 (2018).
1457
+ 18. Jimenez-Teja, Y. et al. J-PLUS: analysis of the intracluster light in the Coma cluster. Astrophys. J.
1458
+ 522, A183 (2019).
1459
+ 19. Krick, J. E. & Berstein, R. A. Diffuse optical light in galaxy clusters. II. Correlations with cluster
1460
+ properties. Astrophys. J. 134, 466–493 (2007).
1461
+ 20. Mihos, J. C. Intragroup and intracluster light. in Proc. IAU Symp.: The General Assembly of Galaxy
1462
+ Halos: Structure, Origin and Evolution vol. 317 (eds Bragaglia, A., Arnaboldi, M., Rejkuba, M. &
1463
+ Romano, D.) 27–34 (Int. Astron. Union, 2015).
1464
+ 21. Yoo, J. et al. Intracluster light properties in a fossil cluster at z = 0.47. Mon. Not. R. Astron. Soc.
1465
+ 508, 2634–2649 (2021).
1466
+ 22. Montes, M. The faint light in groups and clusters of galaxies. Nature. Astro. 6, 308–316 (2022).
1467
+ 23. Navarro, J. F., Frenck, C. S. & White, S. D. M. The structure of cold dark matter halos. Astrophys.
1468
+ J. 462, 563–575 (1996).
1469
+ 24. Asensio, I. A. et al. The intracluster light as a tracer of the total matter density distribution: a view
1470
+ from simulations. Mon. Not. R. Astron. Soc. 494, 1859–1864 (2020).
1471
+ 25. Pillepich, A. et al. First results from the IllustrisTNG simulations: the stellar mass content of groups
1472
+ and clusters of galaxies. Mon. Not. R. Astron. Soc. 475, 648–675 (2018).
1473
+ 26. Murante, G. et al. The diffuse light in simulations of galaxy clusters. Astrophys. J. Lett. 607, 83–
1474
+ 86 (2004).
1475
+ 27. Purcell, C. W., Bullock, J. S. & Zentner, A. R. Shredded galaxies as the source of diffuse intrahalo
1476
+ light on varying scales. Astrophys. J. 666, 20–33 (2007).
1477
+ 28. Furnell, K. E. et al. The growth of intracluster light in XCS-HSC galaxy clusters from 0.1 < z <
1478
+ 0.5. Mon. Not. R. Astron. Soc. 502, 2419–2437 (2021).
1479
+ 29. Guennou, L. et al. Intracluster light in clusters of galaxies at redshifts 0.4 < z < 0.8. Astron.
1480
+ Astrophys. 537, A64 (2012).
1481
+ 30. Contini, E. On the origin and evolution of the intra-cluster light: a brief review of the most recent
1482
+ developments. MDPI. 9, 60 (2021).
1483
+ 31. Murante, G. et al. The importance of mergers for the origin of intracluster stars in cosmological
1484
+ simulations of galaxy clusters. Mon. Not. R. Astron. Soc. 377, 2–16 (2007).
1485
+ 32. Contini, E., Yi, S. K. & Kang, E. The different growth pathways of brightest cluster galaxies and
1486
+ intracluster light. Mon. Not. R. Astron. Soc. 479, 932–944 (2018).
1487
+ 33. Tang, L. et al. An investigation of intracluster light evolution using cosmological hydrodynamical
1488
+ simulations. Astrophys. J. 859, 85–97 (2018).
1489
+ 34. DeMaio, T. et al. On the origin of the intracluster light in massive galaxy clusters. Mon. Not. R.
1490
+ Astron. Soc. 448, 1162–1177 (2015).
1491
+
1492
+ 35. DeMaio, T. et al. Lost but not forgotten: intracluster light in galaxy groups and clusters. Mon. Not.
1493
+ R. Astron. Soc. 474, 3009–3031 (2018).
1494
+ 36. Sahu, K. WFC3 Data Handbook v.5.5 (STScI, 2021).
1495
+ 37. Hoffmann, S. L. et al. The DrizzlePac Handbook v. 2.0 (STScI, 2021).
1496
+ 38. Bertin, E. & Arnouts, S. SExtractor: software for source extraction. Astron. Astrophys. 117, 393–
1497
+ 404 (1996).
1498
+ 39. Stone, C. J. et al. AutoProf - I. An automated non-parametric light profile pipeline for modern
1499
+ galaxy surveys. Mon. Not. R. Astron. Soc. 508, 1870–1887 (2021).
1500
+ 40. Sérsic, J. L. Influence of the atmospheric and instrumental dispersion on the brightness distribution
1501
+ in a galaxy. Boletin de la Asociacion Argentina de Astronomia La Plata Argentina 6, 41–43 (1963).
1502
+ 41. Krist, J. Tiny Tim: an HST PSF simulator. Astronomical Data Analysis Software and Systems II.
1503
+ 52, 536 (1993).
1504
+ 42. Oser, L. et al. The two phases of galaxy formation. Astrophys. J. 725, 2312–2323 (2010).
1505
+ 43. Huang, S. et al. The Carnegie-Irvine Galaxy Survey. III. The three-component structure of nearby
1506
+ elliptical galaxies. Astrophys. J. 766, 47 (2013).
1507
+ 44. Huang, S. et al. The Carnegie-Irvine Galaxy Survey. IV. A method to determine the average mass
1508
+ ratio of mergers that built massive elliptical galaxies. Astrophys. J. 821, 114–133 (2016).
1509
+ 45. Gill, J. Bayesian Methods: A Social Behavioral Science Approch 2nd edn (CRC, 2008). 46. Balogh,
1510
+ M. L. et al. The GOGREEN and GCLASS surveys: first data release. Mon. Not. R. Astron. Soc.
1511
+ 500, 358–387 (2021).
1512
+ 46. Balogh, M. L. et al. The GOGREEN and GCLASS surveys: first data release. Mon. Not. R. Astron.
1513
+ Soc. 500, 358–387 (2021).
1514
+ 47. Santos, J. S. et al. Multiwavelength observations of a rich galaxy cluster at z ~ 1. The HST/ACS
1515
+ colour-magnitude diagram. Astron. Astrophys. 501, 49–60 (2009).
1516
+ 48. Gongalez, A. H. et al. The massive and distant clusters of WISE Survey: MOO J1142+1527, a 10
1517
+ 15
1518
+ M⊙ galaxy cluster at z = 1.19. Astrophys. J. L. 812, L40 (2015).
1519
+ 49. Demarco, R. et al. VLT and ACS observations of RDCS J1252.9-2927: dynamical structure and
1520
+ galaxy populations in a massive cluster at z = 1.237. Astrophys. J. 663, 164–182 (2007).
1521
+ 50. Decker, B. et al. The massive and distant clusters of WISE Survey. VI. Stellar mass fractions of a
1522
+ sample of high-redshift infrared-selected clusters. Astrophys. J. 878, 72–84 (2019).
1523
+ 51. Santos, J. S. et al. Dust-obscured star formation in the outskirts of XMMU J2235.3-2557, a massive
1524
+ galaxy cluster at z = 1.4. Mon. Not. R. Astron. Soc. 433, 1287–1299 (2013).
1525
+ 52. Webb, T. M. A. et al. The star formation history of BCGs to z = 1.8 from the SpARCS/SWIRE
1526
+ Survey: evidence for significant in situ star formation at high redshift. Astrophys. J. 814, 96–107
1527
+ (2015).
1528
+
1529
+ 53. Stanford, S. A. et al. IDCS J1426.5+3508: discovery of a massive, infrared-selected galaxy cluster
1530
+ at z = 1.75. Astrophys. J. 753, 164–171 (2012).
1531
+ 54. Newman, A. B. et al. Spectroscopic confirmation of the rich z = 1.80 galaxy cluster JKCS041 using
1532
+ the WFC3 grism: environmental trends in the ages and structure of quiescent galaxies. Astrophys.
1533
+ J. 788, 51–76 (2014).
1534
+
JdAzT4oBgHgl3EQfj_0H/content/tmp_files/load_file.txt ADDED
The diff for this file is too large to render. See raw diff