-
Notifications
You must be signed in to change notification settings - Fork 1
/
GILT.py
388 lines (330 loc) · 13.5 KB
/
GILT.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
# Scaling dimensions from linearized tensor renormalization group transformations
# https://arxiv.org/pdf/2102.08136.pdf
# Renormalization of tensor networks using graph independent local truncations
# https://arxiv.org/pdf/1709.07460.pdf
# https://github.com/Gilt-TNR/Gilt-TNR
# https://github.com/Gilt-TNR/Gilt-TNR/blob/master/GiltTNR3D.py
import torch
from tqdm.auto import tqdm
import numpy as np
from opt_einsum import contract
import itertools as itt
from dataclasses import dataclass
def _toN(t):
return t.detach().cpu().tolist() if isinstance(t,torch.Tensor) else t
#from safe_svd import svd,sqrt # TODO is it necessary???
from torch.linalg import svd
from torch import sqrt
# Basic idea:
# Given a subgraph and a specific leg in that subgraph, the environment tensor E = break the leg, the mapping from the two ends of the broken leg to the external legs of the subgraph
# We can insert arbitrary projector R to that leg if that's in the kernel of E
# We chose R_ab=t_i' U_abi, where E_ab,external=U_ab,i t_i Vh_i,external, and only change t_i which are small: t_i'=t_i**2/(t_i**2+gilt_eps**2), we also do gilt_nIter iterations
# To see how it works
# E factorizes as E(UV and IR entanglement to outside) \otimes I(UV entanglement inside)
# TODO
#svd,sqrt=torch.linalg.svd,torch.sqrt
if torch.get_default_dtype() not in {torch.float64}:
print('[GILT] Warning! float32 is not precise enough, leads to bad RG behavior')
@dataclass
class GILT_options:
enabled:bool=True
eps:float=8e-7 #too los like 1e-8 will result false fixed topological point
nIter:int=1 #enough
split_insertion:bool=True
TRG_method:str='A'
HOTRG_3D_method:str='square_only'
fix_gauge:bool=True
record_S:bool=False
make_isometric:bool=False
recorded_S=[]
def GILT_getuvh(EEh,options:GILT_options=GILT_options()):
d=EEh.shape[0]
uu,vvh=torch.eye(d),torch.eye(d)
for _iter in range(options.nIter):
if _iter==0:
U,S,_=svd(EEh.reshape(d**2,d**2))
else:
uvUS=contract('aA,Bb,abc,c->ABc',u,vh,U,S).reshape(d**2,d**2)
U,S,_=svd(uvUS)
U=U.reshape(d,d,d**2)
t=contract('aac->c',U)
Sn=S/torch.max(S)
if options.record_S:
recorded_S.append(_toN(Sn))
t=t*(Sn**2/(Sn**2+options.eps**2))
Q=contract('abc,c->ab',U,t)
if options.split_insertion:
u,s,vh=svd(Q) # is it necessary to split?
s=sqrt(s).diag()
u,vh=u@s,s@vh
else:
# not make sense, introduces numerical error!
u,vh=Q,torch.eye(d)
uu,vvh=uu@u,vh@vvh
return uu,vvh
def GILT_getEEh(As,Ais:"list[list[str]]"):
def process(edgeid,legid,tensorid,replicaid):
if edgeid is None:
#contract between corresponding replicas
return 'T'+str(tensorid)+'L'+str(legid)
else:
#internal legs
return 'R'+str(replicaid)+'E'+str(edgeid)
R1Ais=[[process(edgeid,legid,tensorid,0) for legid,edgeid in enumerate(Ai)]for tensorid,Ai in enumerate(Ais)]
R2Ais=[[process(edgeid,legid,tensorid,1) for legid,edgeid in enumerate(Ai)]for tensorid,Ai in enumerate(Ais)]
AAis=[list(filter(lambda x:x[0]=='R',R1Ai+R2Ai)) for R1Ai,R2Ai in zip(R1Ais,R2Ais)]
Ti=['R0Eu','R0Ev','R1Eu','R1Ev']
AAs=[contract(A,R1Ai,A,R2Ai,AAi) for A,R1Ai,R2Ai,AAi in zip(As,R1Ais,R2Ais,AAis)]
#print(AAis)
T=contract(*itt.chain(*zip(AAs,AAis)),Ti)
#print(R1Ais);print(R2Ais);print(AAis);print(Ti)
#for AA,AAi in zip(AAs,AAis):
# print(AA.shape)
# print(AAi)
#print(T.shape)
#print(Ti)
return T
#======================================================
#def GILT_Square_one(As,options:GILT_options=GILT_options()):
# # A1- -A2 A1u vA2 0
# # | O | --> | U | 2A3
# # A3---A4 A3---A4 1
# A1i=[None,'13',None,'u']
# A2i=[None,'24','v',None]
# A3i=['13',None,None,'34']
# A4i=['24',None,'34',None]
# EEh=GILT_getEEh(As,[A1i,A2i,A3i,A4i])
# u,vh=GILT_getuvh(EEh,options=options)
# assert not u.isnan().any() and not vh.isnan().any()
# return u,vh
def replace_leg_with_u_and_v(Ais,leg):
flag=False
for Ai in Ais:
if leg in Ai:
if not flag:
Ai[Ai.index(leg)]='u'
flag=True
else:
Ai[Ai.index(leg)]='v'
return Ais
def GILT_Square_one(As,leg,options:GILT_options=GILT_options()):
# leg: 12 for example
# A1- -A2 A1u vA2 0
# | O | --> | U | 2A3
# A3---A4 A3---A4 1
Ais=[
[None,'13',None,'12'],
[None,'24','12',None],
['13',None,None,'34'],
['24',None,'34',None],
]
if(len(As[0].shape)==5): #it might also works for PEPS I hope
Ais=[Ai+[None] for Ai in Ais]
assert leg in {'12','34','13','24'}
Ais=replace_leg_with_u_and_v(Ais,leg)
EEh=GILT_getEEh(As,Ais)
u,vh=GILT_getuvh(EEh,options=options)
assert not u.isnan().any() and not vh.isnan().any()
return u,vh
def GILT_Cube_one(As,leg,options:GILT_options=GILT_options()):
# leg: 12 for example
# A5+------+A6
# |`. |`. 0
# | A1+-u v-+A2 5`|
# | | | | 2-o-3
# A7+---|--+A8 | |`4
# `. | `. | 1
# A3+------+A4
Ais=[
[None,'13',None,'12',None,'15'],
[None,'24','12',None,None,'26'],
['13',None,None,'34',None,'37'],
['24',None,'34',None,None,'48'],
[None,'57',None,'56','15',None],
[None,'68','56',None,'26',None],
['57',None,None,'78','37',None],
['68',None,'78',None,'48',None],
]
if(len(As[0].shape)==7): #it might also works for PEPS I hope
Ais=[Ai+[None] for Ai in Ais]
assert leg in {'12','34','56','78','13','24','57','68','15','26','37','48'}
Ais=replace_leg_with_u_and_v(Ais,leg)
EEh=GILT_getEEh(As,Ais)
u,vh=GILT_getuvh(EEh,options=options)
assert not u.isnan().any() and not vh.isnan().any()
return u,vh
def GILT_HOTRG2D(T1,T2,options:GILT_options=GILT_options()):
# O | O
# /v1-T1-u1\ 0 2
# -w |\ w- 2T3 -> 0T'1
# \v2-T2-u2/ 14 34
# O |\O
#Y1,Y2=T1,T2
contract_path={4:'ijkl,Kk,Ll->ijKL',5:'ijkla,Kk,Ll->ijKLa'}[len(T1.shape)]
u1,vh1=GILT_Square_one([T2,T2,T1,T1],leg='34',options=options)
T1=contract(contract_path,T1,vh1,u1.T)
u2,vh2=GILT_Square_one([T2,T2,T1,T1],leg='12',options=options)
T2=contract(contract_path,T2,vh2,u2.T)
I=torch.eye(T1.shape[0])
gg=[[I,I,vh1,u1.T],[I,I,vh2,u2.T]]
#Y1=contract('ijkl,Ii,Jj,Kk,Ll->IJKL',Y1,*gg[0])
#Y2=contract('ijkl,Ii,Jj,Kk,Ll->IJKL',Y2,*gg[1])
#print((T1-Y1).norm(),(T2-Y2).norm())
return T1,T2,gg
def GILT_HOTRG3D_square_only(T1,T2,options:GILT_options=GILT_options()):
# g4| 5--6
# /g1-T1-g2\ 50 34 |1--2
# -w g8|g3 w- 2T3 -> 0T'1 7| 8|
# \g5-T2-g6/ 14 52 3--4
# |g7
raise NotImplementedError
def GILT_HOTRG3D(T1,T2,options:GILT_options=GILT_options()):
print('not tested!')
# g4| 5--6
# /g1-T1-g2\ 50 34 |1--2
# -w g8|g3 w- 2T3 -> 0T'1 7| 8|
# \g5-T2-g6/ 14 52 3--4
# |g7
Y1,Y2=T1,T2
T21s=[T2,T2,T2,T2,T1,T1,T1,T1]
T12s=[T1,T1,T1,T1,T2,T2,T2,T2]
contract23='ijklmn,Kk,Ll->ijKLmn'
contract45='ijklmn,Mm,Nn->ijklMN'
cube_apply_inner=True
u,vh=GILT_Cube_one(T21s,leg='34',options=options)
T1,g1,g2=contract(contract23,T1,vh,u.T),vh,u.T
u,vh=GILT_Cube_one(T21s,leg='78',options=options)
T1,g1,g2=contract(contract23,T1,vh,u.T),vh@g1,u.T@g2
if cube_apply_inner:
u,vh=GILT_Cube_one(T12s,leg='12',options=options)
T1,g1,g2=contract(contract23,T1,vh,u.T),vh@g1,u.T@g2
u,vh=GILT_Cube_one(T12s,leg='56',options=options)
T1,g1,g2=contract(contract23,T1,vh,u.T),vh@g1,u.T@g2
u,vh=GILT_Cube_one(T21s,leg='37',options=options)
T1,g3,g4=contract(contract45,T1,vh,u.T),vh,u.T
u,vh=GILT_Cube_one(T21s,leg='48',options=options)
T1,g3,g4=contract(contract45,T1,vh,u.T),vh@g3,u.T@g4
if cube_apply_inner:
u,vh=GILT_Cube_one(T12s,leg='15',options=options)
T1,g3,g4=contract(contract45,T1,vh,u.T),vh@g3,u.T@g4
u,vh=GILT_Cube_one(T12s,leg='26',options=options)
T1,g3,g4=contract(contract45,T1,vh,u.T),vh@g3,u.T@g4
u,vh=GILT_Cube_one(T21s,leg='12',options=options)
T2,g5,g6=contract(contract23,T2,vh,u.T),vh,u.T
u,vh=GILT_Cube_one(T21s,leg='56',options=options)
T2,g5,g6=contract(contract23,T2,vh,u.T),vh@g5,u.T@g6
if cube_apply_inner:
u,vh=GILT_Cube_one(T12s,leg='34',options=options)
T2,g5,g6=contract(contract23,T2,vh,u.T),vh@g5,u.T@g6
u,vh=GILT_Cube_one(T12s,leg='78',options=options)
T2,g5,g6=contract(contract23,T2,vh,u.T),vh@g5,u.T@g6
u,vh=GILT_Cube_one(T21s,leg='15',options=options)
T2,g7,g8=contract(contract45,T2,vh,u.T),vh,u.T
u,vh=GILT_Cube_one(T21s,leg='26',options=options)
T2,g7,g8=contract(contract45,T2,vh,u.T),vh@g7,u.T@g8
if cube_apply_inner:
u,vh=GILT_Cube_one(T12s,leg='37',options=options)
T2,g7,g8=contract(contract45,T2,vh,u.T),vh@g7,u.T@g8
u,vh=GILT_Cube_one(T12s,leg='48',options=options)
T2,g7,g8=contract(contract45,T2,vh,u.T),vh@g7,u.T@g8
I=torch.eye(T1.shape[0])
gg=[[I,I,g1,g2,g3,g4],[I,I,g5,g6,g7,g8]]
Y1=contract('ijklmn,Ii,Jj,Kk,Ll,Mm,Nn->IJKLMN',Y1,*gg[0])
Y2=contract('ijklmn,Ii,Jj,Kk,Ll,Mm,Nn->IJKLMN',Y2,*gg[1])
print((Y1-T1).norm(),(Y2-T2).norm())
return T1,T2,gg
def GILT_HOTRG(T1,T2,options:GILT_options=GILT_options()):
_GILT_HOTRG={4:GILT_HOTRG2D,5:GILT_HOTRG2D,6:GILT_HOTRG3D}[len(T1.shape)]
T1,T2,gg=_GILT_HOTRG(T1,T2,options=options)
if options.record_S:
import matplotlib.pyplot as plt
plt.hist(recorded_S[0],bins=np.logspace(-9,0,50),log=True)
plt.xscale('log')
plt.show()
recorded_S.clear()
return T1,T2,gg
'''
from HOSVD import HOSVD_layer
from fix_gauge import fix_gauge_2D,minimal_canonical_form,fix_phase
def GILT_HOSVD_layer(T1,T2,max_dim,dimR=None,options:GILT_options=GILT_options()):
if options.enabled:
assert not dimR
T1,T2,gg=GILT_HOTRG(T1,T2,options=options)
Tn,layer=HOSVD_layer(T1,T2,max_dim=max_dim,dimR=dimR)
layer.gg=gg
return Tn,layer
else:
return HOSVD_layer(T1,T2,max_dim=max_dim,dimR=dimR)
'''
#============= GILT On TRG=============
from TRG import TRG_AB
from HOTRGZ2 import gauge_invariant_norm
def GILT_SquareA(A,options:GILT_options=GILT_options()):
# Not good precision. Why?
# A- -A vAu vAu
# | O | --> | O |
# A---A vAu-vAu
for i in range(4):
u,vh=GILT_Square_one([A,A,A,A],leg='12',options=options)
A=contract('abcd,Cc,dD->abCD',A,vh,u)
A=contract('abcd->dcab',A)
return A
def GILT_SquareABCD(A,B,C,D,options:GILT_options=GILT_options()):
# A- -B Au vB 0
# | O | --> | O | 2A3
# C---D C---D 1
for i in range(2):
for i in range(4):
u,vh=GILT_Square_one([A,B,C,D],leg='12',options=options)
A,B=contract('abcd,dD->abcD',A,u),contract('abcd,Cc->abCd',B,vh)
CCW='abcd->dcab'
A,B,C,D=contract(CCW,B),contract(CCW,D),contract(CCW,A),contract(CCW,C)#rotate CCW
A,B,C,D=B,A,D,C
return A,B,C,D
def GILT_SquareAB(A,B,options:GILT_options=GILT_options()):
# A---B Au vB 0
# | O | --> | O | 2A3
# B---A vB---Au 1
for i in range(4):
u,vh=GILT_Square_one([A,B,B,A],leg='12',options=options)
A,B=contract('abcd,dD->abcD',A,u),contract('abcd,Cc->abCd',B,vh)
A,B=contract('abcd->dcab',B),contract('abcd->dcab',A)#rotate
return A,B
# def evolve_TRG_GILT_2D(T0,nLayers,max_dim,return_layers=False,options:GILT_options=GILT_options()):
# T,logTotal=T0,0
# if return_layers:
# Ts,logTotals=[T],[logTotal]
# for i in tqdm(range(nLayers),leave=False):
# norm=gauge_invariant_norm(T)
# T=T/norm
# logTotal=2*(logTotal+norm.log())
# if options.TRG_method=='A':
# A=GILT_SquareA(T,options=options)
# T=TRG_AB(A,A,max_dim)
# elif options.TRG_method=='AB':
# A,B=GILT_SquareAB(T,T,options=options)
# T=TRG_AB(A,B,max_dim)
# elif options.TRG_method=='BA':
# A,B=GILT_SquareAB(T,T,options=options)
# T=TRG_AB(B,A,max_dim)
# elif options.TRG_method=='BAAB':
# A,B=GILT_SquareAB(T,T,options=options)
# A,B=GILT_SquareAB(B,A,options=options)
# T=TRG_AB(A,B,max_dim)
# elif options.TRG_method=='BABA':
# A,B=GILT_SquareAB(T,T,options=options)
# A,B=GILT_SquareAB(B,A,options=options)
# T=TRG_AB(B,A,max_dim)
# elif options.TRG_method=='ABCD':
# assert False
# else:
# assert False
# if options.fix_gauge and T.shape==Ts[-1].shape:
# #T=fix_gauge_2D(T,Ts[-1])
# T,_=minimal_canonical_form(T)
# T,_=fix_phase(T,Ts[-1])
# if return_layers:
# Ts.append(T);logTotals.append(logTotal)
# if return_layers:
# return Ts,logTotals
# return T,logTotal