/
datacomparator.py
926 lines (776 loc) · 39.9 KB
/
datacomparator.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
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
""" Defines the DataComparator class used to compare multiple DataSets."""
#***************************************************************************************************
# Copyright 2015, 2019 National Technology & Engineering Solutions of Sandia, LLC (NTESS).
# Under the terms of Contract DE-NA0003525 with NTESS, the U.S. Government retains certain rights
# in this software.
# Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except
# in compliance with the License. You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0 or in the LICENSE file in the root pyGSTi directory.
#***************************************************************************************************
import numpy as _np
import scipy as _scipy
import copy as _copy
from scipy import stats as _stats
import collections as _collections
from .multidataset import MultiDataSet as _MultiDataSet
from .hypothesistest import HypothesisTest as _HypothesisTest
def xlogy(x, y):
"""
Returns x*log(y).
"""
if x == 0:
return 0
else:
return x * _np.log(y)
def likelihood(pList, nList):
"""
The likelihood for probabilities `pList` of a die,
given `nList` counts for each outcome.
"""
output = 1.
for i, pVal in enumerate(pList):
output *= pVal**nList[i]
return output
def loglikelihood(pList, nList):
"""
The log of the likelihood for probabilities `pList` of a die,
given `nList` counts for each outcome.
"""
output = 0.
for i, pVal in enumerate(pList):
output += xlogy(nList[i], pVal)
return output
# Only used by the rectify datasets function, which is commented out,
# so this is also commented out.
# def loglikelihoodRatioObj(alpha,nListList,dof):
# return _np.abs(dof - loglikelihoodRatio(alpha*nListList))
def loglikelihoodRatio(nListList):
"""
Calculates the log-likelood ratio between the null hypothesis
that a die has *the same* probabilities in multiple "contexts" and
that it has *different* probabilities in multiple "contexts".
Parameters
----------
nListList : List of lists of ints
A list whereby element i is a list containing observed counts for
all the different possible outcomes of the "die" in context i.
Returns
-------
float
The log-likehood ratio for this model comparison.
"""
nListC = _np.sum(nListList, axis=0)
pListC = nListC / _np.float(_np.sum(nListC))
lC = loglikelihood(pListC, nListC)
li_list = []
for nList in nListList:
pList = _np.array(nList) / _np.float(_np.sum(nList))
li_list.append(loglikelihood(pList, nList))
lS = _np.sum(li_list)
return -2 * (lC - lS)
def JensenShannonDivergence(nListList):
"""
Calculates the Jensen-Shannon divergence (JSD) between between different
observed frequencies, obtained in different "contexts", for the different
outcomes of a "die" (i.e., coin with more than two outcomes).
Parameters
----------
nListList : List of lists of ints
A list whereby element i is a list containing observed counts for
all the different possible outcomes of the "die" in context i.
Returns
-------
float
The observed JSD for this data.
"""
total_counts = _np.sum(_np.array(nListList))
return loglikelihoodRatio(nListList) / (2 * total_counts)
def pval(llrval, dof):
"""
The p-value of a log-likelihood ratio (LLR), comparing a
nested null hypothsis and a larger alternative hypothesis.
Parameters
----------
llrval : float
The log-likehood ratio
dof : int
The number of degrees of freedom associated with
the LLR, given by the number of degrees of freedom
of the full model space (the alternative hypothesis)
minus the number of degrees of freedom of the restricted
model space (the null hypothesis space).
Returns
-------
float
An approximation of the p-value for this LLR. This is
calculated as 1 - F(llrval,dof) where F(x,k) is the
cumulative distribution function, evaluated at x, for
the chi^2_k distribution. The validity of this approximation
is due to Wilks' theorem.
"""
return 1 - _stats.chi2.cdf(llrval, dof)
def llr_to_signed_nsigma(llrval, dof):
"""
Finds the signed number of standard deviations for the input
log-likelihood ratio (LLR). This is given by
(llrval - dof) / (sqrt(2*dof)).
This is the number of standard deviations above the mean
that `llrval` is for a chi^2_(dof) distribution.
Parameters
----------
llrval : float
The log-likehood ratio
dof : int
The number of degrees of freedom associated with
the LLR, given by the number of degrees of freedom
of the full model space (the alternative hypothesis)
minus the number of degrees of freedom of the restricted
model space (the null hypothesis space), in the hypothesis
test.
Returns
-------
float
The signed standard deviations.
"""
return (llrval - dof) / _np.sqrt(2 * dof)
def is_circuit_allowed_by_exclusion(op_exclusions, circuit):
"""
Returns True if `circuit` does not contain any gates from `op_exclusions`.
Otherwise, returns False.
"""
for gate in op_exclusions:
if gate in circuit:
return False
return True
def is_circuit_allowed_by_inclusion(op_inclusions, circuit):
"""
Returns True if `circuit` contains *any* of the gates from `op_inclusions`.
Otherwise, returns False. The exception is the empty circuit, which always
returns True.
"""
if len(circuit) == 0: return True # always include the empty string
for gate in op_inclusions:
if gate in circuit:
return True
return False
def compute_llr_threshold(significance, dof):
"""
Given a p-value threshold, *below* which a pvalue
is considered statistically significant, it returns
the corresponding log-likelihood ratio threshold, *above*
which a LLR is considered statically significant. For a single
hypothesis test, the input pvalue should be the desired "significance"
level of the test (as a value between 0 and 1). For multiple hypothesis
tests, this will normally be smaller than the desired global significance.
Parameters
----------
pVal : float
The p-value
dof : int
The number of degrees of freedom associated with
the LLR , given by the number of degrees of freedom
of the full model space (the alternative hypothesis)
minus the number of degrees of freedom of the restricted
model space (the null hypothesis space), in the hypothesis
test.
Returns
-------
float
The significance threshold for the LLR, given by
1 - F^{-1}(pVal,dof) where F(x,k) is the cumulative distribution
function, evaluated at x, for the chi^2_k distribution. This
formula is based on Wilks' theorem.
"""
return _scipy.stats.chi2.isf(significance, dof)
def tvd(nListList):
"""
Calculates the total variation distance (TVD) between between different
observed frequencies, obtained in different "contexts", for the *two* set of
outcomes for roles of a "die".
Parameters
----------
nListList : List of lists of ints
A list whereby element i is a list counting counts for the
different outcomes of the "die" in context i, for *two* contexts.
Returns
-------
float
The observed TVD between the two contexts
"""
assert(len(nListList) == 2), "Can only compute the TVD between two sets of outcomes!"
num_outcomes = len(nListList[0])
assert(num_outcomes == len(nListList[1])), "The number of outcomes must be the same in both contexts!"
N0 = _np.sum(nListList[0])
N1 = _np.sum(nListList[1])
return 0.5 * _np.sum(_np.abs(nListList[0][i] / N0 - nListList[1][i] / N1) for i in range(num_outcomes))
class DataComparator():
"""
This object can be used to implement all of the "context dependence detection" methods described
in "Probing context-dependent errors in quantum processors", by Rudinger et al.
(See that paper's supplemental material for explicit demonstrations of this object.)
This object stores the p-values and log-likelihood ratio values from a consistency comparison between
two or more datasets, and provides methods to:
- Perform a hypothesis test to decide which sequences contain statistically significant variation.
- Plot p-value histograms and log-likelihood ratio box plots.
- Extract (1) the "statistically significant total variation distance" for a circuit,
(2) various other quantifications of the "amount" of context dependence, and (3)
the level of statistical significance at which any context dependence is detected.
"""
def __init__(self, dataset_list_or_multidataset, circuits='all',
op_exclusions=None, op_inclusions=None, DS_names=None,
allow_bad_circuits=False):
"""
Initializes a DataComparator object.
Parameters
----------
dataset_list_multidataset : List of DataSets or MultiDataSet
Either a list of DataSets, containing two or more sets of data to compare,
or a MultiDataSet object, containing two or more sets of data to compare. Note
that these DataSets should contain data for the same set of Circuits (although
if there are additional Circuits these can be ignored using the parameters below).
This object is then intended to be used test to see if the results are indicative
that the outcome probabilities for these Circuits has changed between the "contexts" that
the data was obtained in.
circuits : 'all' or list of Circuits, optional (default is 'all')
If 'all' the comparison is implemented for all Circuits in the DataSets. Otherwise,
this should be a list containing all the Circuits to implement the comparison for (although
note that some of these Circuits may be ignored with non-default options for the next two
inputs).
op_exclusions : None or list of gates, optional (default is None)
If not None, all Circuits containing *any* of the gates in this list are discarded,
and no comparison will be made for those strings.
op_exclusions : None or list of gates, optional (default is None)
If not None, a Circuit will be dropped from the list to implement the comparisons for
if it doesn't include *some* gate from this list (or is the empty circuit).
DS_names : None or list, optional (default is None)
If `dataset_list_multidataset` is a list of DataSets, this can be used to specify names
for the DataSets in the list. E.g., ["Time 0", "Time 1", "Time 3"] or ["Driving","NoDriving"].
allow_bad_circuits : bool, optional
Whether or not the data is allowed to have zero total counts for any circuits in any of the
passes. If false, then an error will be raise when there are such unimplemented circuits. If
true, then the data from those circuits that weren't run in one or more of the passes will
be discarded before any analysis is performed (equivalent to excluding them explicitly in with
the `circuits` input.
Returns
-------
A DataComparator object.
"""
if DS_names is not None:
if len(DS_names) != len(dataset_list_or_multidataset):
raise ValueError('Length of provided DS_names list must equal length of dataset_list_or_multidataset.')
if isinstance(circuits, str):
assert(circuits == 'all'), "If circuits is a string it must be 'all'!"
if isinstance(dataset_list_or_multidataset, list):
dsList = dataset_list_or_multidataset
olIndex = dsList[0].olIndex
olIndexListBool = [ds.olIndex == (olIndex) for ds in dsList]
DS_names = list(range(len(dataset_list_or_multidataset)))
if not _np.all(olIndexListBool):
raise ValueError('Outcomes labels and order must be the same across datasets.')
if circuits == 'all':
circuitList = dsList[0].keys()
circuitsListBool = [ds.keys() == circuitList for ds in dsList]
if not _np.all(circuitsListBool):
raise ValueError(
'If circuits="all" is used, then datasets must contain identical circuits. (They do not.)')
circuits = circuitList
elif isinstance(dataset_list_or_multidataset, _MultiDataSet):
dsList = [dataset_list_or_multidataset[key] for key in dataset_list_or_multidataset.keys()]
if circuits == 'all':
circuits = dsList[0].keys()
if DS_names is None:
DS_names = list(dataset_list_or_multidataset.keys())
else:
raise ValueError("The `dataset_list_or_multidataset` must be a list of DataSets of a MultiDataSet!")
if allow_bad_circuits:
trimmedcircuits = []
for circuit in circuits:
if min([ds[circuit].total for ds in dsList]) > 0:
trimmedcircuits.append(circuit)
circuits = trimmedcircuits
if op_exclusions is not None:
circuits_exc_temp = []
for circuit in circuits:
if is_circuit_allowed_by_exclusion(op_exclusions, circuit):
circuits_exc_temp.append(circuit)
circuits = list(circuits_exc_temp)
if op_inclusions is not None:
circuits_inc_temp = []
for circuit in circuits:
if is_circuit_allowed_by_inclusion(op_inclusions, circuit):
circuits_inc_temp.append(circuit)
circuits = list(circuits_inc_temp)
llrs = {}
pVals = {}
jsds = {}
dof = (len(dsList) - 1) * (len(dsList[0].olIndex) - 1)
total_counts = []
if len(dataset_list_or_multidataset) == 2:
tvds = {}
for circuit in circuits:
datalineList = [ds[circuit] for ds in dsList]
nListList = _np.array([list(dataline.allcounts.values()) for dataline in datalineList])
total_counts.append(_np.sum(nListList))
llrs[circuit] = loglikelihoodRatio(nListList)
jsds[circuit] = JensenShannonDivergence(nListList)
pVals[circuit] = pval(llrs[circuit], dof)
if len(dataset_list_or_multidataset) == 2:
tvds[circuit] = tvd(nListList)
self.dataset_list_or_multidataset = dataset_list_or_multidataset
self.pVals = pVals
self.pVals_pseudothreshold = None
self.llrs = llrs
self.llrs_pseudothreshold = None
self.jsds = jsds
if len(dataset_list_or_multidataset) == 2:
self.tvds = tvds
self.op_exclusions = op_exclusions
self.op_inclusions = op_inclusions
self.pVals0 = str(len(self.pVals) - _np.count_nonzero(list(self.pVals.values())))
self.dof = dof
self.num_strs = len(self.pVals)
self.DS_names = DS_names
if _np.std(_np.array(total_counts)) > 10e-10:
self.fixed_totalcount_data = False
self.counts_per_sequence = None
else:
self.fixed_totalcount_data = True
self.counts_per_sequence = int(total_counts[0])
self.aggregate_llr = _np.sum(list(self.llrs.values()))
self.aggregate_llr_threshold = None
self.aggregate_pVal = pval(self.aggregate_llr, self.num_strs * self.dof)
self.aggregate_pVal_threshold = None
# Convert the aggregate LLR to a signed standard deviations.
self.aggregate_nsigma = llr_to_signed_nsigma(self.aggregate_llr, self.num_strs * self.dof)
self.aggregate_nsigma_threshold = None
# All attributes to be populated in methods that can be called from .get methods, so
# we can raise a meaningful warning if they haven't been calculated yet.
self.sstvds = None
self.pVal_pseudothreshold = None
self.llr_pseudothreshold = None
self.pVal_pseudothreshold = None
self.jsd_pseudothreshold = None
def implement(self, significance=0.05, per_sequence_correction='Hochberg',
aggregate_test_weighting=0.5, pass_alpha=True, verbosity=2):
"""
Implements statistical hypothesis testing, to detect whether there is statistically
significant variation between the DateSets in this DataComparator. This performs
hypothesis tests on the data from individual circuits, and a joint hypothesis test
on all of the data. With the default settings, this is the method described and implemented
in "Probing context-dependent errors in quantum processors", by Rudinger et al. With
non-default settings, this is some minor variation on that method.
Note that the default values of all the parameters are likely sufficient for most
purposes.
Parameters
----------
significance : float in (0,1), optional (default is 0.05)
The "global" statistical significance to implement the tests at. I.e, with
the standard `per_sequence_correction` value (and some other values for this parameter)
the probability that a sequence that has been flagged up as context dependent
is actually from a context-independent circuit is no more than `significance`.
Precisely, `significance` is what the "family-wise error rate" (FWER) of the full set
of hypothesis tests (1 "aggregate test", and 1 test per sequence) is controlled to,
as long as `per_sequence_correction` is set to the default value, or another option
that controls the FWER of the per-sequence comparion (see below).
per_sequence_correction : string, optional (default is 'Hochberg')
The multi-hypothesis test correction used for the per-circuit/sequence comparisons.
(See "Probing context-dependent errors in quantum processors", by Rudinger et al. for
the details of what the per-circuit comparison is). This can be any string that is an allowed
value for the `localcorrections` input parameter of the HypothesisTest object. This includes:
- 'Hochberg'. This implements the Hochberg multi-test compensation technique. This
is strictly the best method available in the code, if you wish to control the FWER,
and it is the method described in "Probing context-dependent errors in quantum processors",
by Rudinger et al.
- 'Holms'. This implements the Holms multi-test compensation technique. This
controls the FWER, and it results in a strictly less powerful test than the Hochberg
correction.
- 'Bonferroni'. This implements the well-known Bonferroni multi-test compensation
technique. This controls the FWER, and it results in a strictly less powerful test than
the Hochberg correction.
- 'none'. This implements no multi-test compensation for the per-sequence comparsions,
so they are all implemented at a "local" signifincance level that is altered from `significance`
only by the (inbuilt) Bonferroni-like correction between the "aggregate" test and the per-sequence
tests. This option does *not* control the FWER, and many sequences may be flagged up as context
dependent even if none are.
-'Benjamini-Hochberg'. This implements the Benjamini-Hockberg multi-test compensation
technique. This does *not* control the FWER, and instead controls the "False Detection Rate"
(FDR); see, for example, https://en.wikipedia.org/wiki/False_discovery_rate. That means that
the global significance is maintained for the test of "Is there any context dependence?". I.e.,
one or more tests will trigger when there is no context
dependence with at most a probability of `significance`. But, if one or more per-sequence tests
trigger then we are only guaranteed that (in expectation) no more than a fraction of
"local-signifiance" of the circuits that have been flagged up as context dependent actually aren't.
Here, "local-significance" is the significance at which the per-sequence tests are, together,
implemented, which is `significance`*(1 - `aggregate_test_weighting`) if the aggregate test doesn't
detect context dependence and `significance` if it does (as long as `pass_alpha` is True). This
method is strictly more powerful than the Hochberg correction, but it controls a different, weaker
quantity.
aggregate_test_weighting : float in [0,1], optional (default is 0.5)
The weighting, in a generalized Bonferroni correction, to put on the "aggregate test", that jointly
tests all of the data for context dependence (in contrast to the per-sequence tests). If this is 0 then
the aggreate test is not implemented, and if it is 1 only the aggregate test is implemented (unless it
triggers and `pass_alpha` is True).
pass_alpha : Bool, optional (default is True)
The aggregate test is implemented first, at the "local" significance defined by `aggregate_test_weighting`
and `significance` (see above). If `pass_alpha` is True, then when the aggregate test triggers all the local
significance for this test is passed on to the per-sequence tests (which are then jointly implemented with
significance `significance`, that is then locally corrected for the multi-test correction as specified
above), and when the aggregate test doesn't trigger this local significance isn't passed on. If `pass_alpha`
is False then local significance of the aggregate test is never passed on from the aggregate test. See
"Probing context-dependent errors in quantum processors", by Rudinger et al. (or hypothesis testing
literature) for discussions of why this "significance passing" still maintains a (global) FWER of
`significance`. Note that The default value of True always results in a strictly more powerful test.
verbosity : int, optional (default is 1)
If > 0 then a summary of the results of the tests is printed to screen. Otherwise, the
various .get_...() methods need to be queried to obtain the results of the
hypothesis tests.
Returns
-------
None
"""
self.significance = significance
assert(aggregate_test_weighting <= 1. or aggregate_test_weighting >= 0.), \
"The weighting on the aggregate test must be between 0 and 1!"
if verbosity >= 3:
print("Implementing {0:.2f}% significance statistical hypothesis testing...".format(
self.significance * 100), end='')
circuits = tuple(self.pVals.keys())
hypotheses = ('aggregate', circuits)
weighting = {}
weighting['aggregate'] = aggregate_test_weighting
weighting[circuits] = 1 - aggregate_test_weighting
if pass_alpha: passing_graph = 'Holms'
else: passing_graph = 'none'
hypotest = _HypothesisTest(hypotheses, significance=significance, weighting=weighting,
passing_graph=passing_graph, local_corrections=per_sequence_correction)
extended_pVals_dict = _copy.copy(self.pVals)
extended_pVals_dict['aggregate'] = self.aggregate_pVal
hypotest.add_pvalues(extended_pVals_dict)
hypotest.implement()
self.results = hypotest
if aggregate_test_weighting == 0:
self.aggregate_llr_threshold = _np.inf
self.aggregate_nsigma_threshold = _np.inf
self.aggregate_pVal_threshold = 0.
else:
self.aggregate_llr_threshold = compute_llr_threshold(
aggregate_test_weighting * significance, self.num_strs * self.dof)
self.aggregate_nsigma_threshold = llr_to_signed_nsigma(
self.aggregate_llr_threshold, self.num_strs * self.dof)
self.aggregate_pVal_threshold = aggregate_test_weighting * significance
self.pVal_pseudothreshold = hypotest.pvalue_pseudothreshold[circuits]
self.llr_pseudothreshold = compute_llr_threshold(self.pVal_pseudothreshold, self.dof)
if self.fixed_totalcount_data:
self.jsd_pseudothreshold = self.llr_pseudothreshold / self.counts_per_sequence
temp_hypothesis_rejected_dict = _copy.copy(hypotest.hypothesis_rejected)
self.inconsistent_datasets_detected = any(list(temp_hypothesis_rejected_dict.values()))
del temp_hypothesis_rejected_dict['aggregate']
self.number_of_significant_sequences = _np.sum(list(temp_hypothesis_rejected_dict.values()))
if len(self.dataset_list_or_multidataset) == 2:
sstvds = {}
for opstr in list(self.llrs.keys()):
if self.results.hypothesis_rejected[opstr]:
sstvds[opstr] = self.tvds[opstr]
self.sstvds = sstvds
if verbosity >= 3:
print("complete.")
if verbosity >= 3:
print("\n--- Results ---\n")
if verbosity >= 1:
if self.inconsistent_datasets_detected:
print("The datasets are INCONSISTENT at {0:.2f}% significance.".format(self.significance * 100))
print(" - Details:")
print(" - The aggregate log-likelihood ratio test is "
"significant at {0:.2f} standard deviations.".format(self.aggregate_nsigma))
print(" - The aggregate log-likelihood ratio test "
"standard deviations signficance threshold is {0:.2f}".format(self.aggregate_nsigma_threshold))
print(
" - The number of sequences with data that is "
"inconsistent is {0}".format(self.number_of_significant_sequences))
if len(self.dataset_list_or_multidataset) == 2 and self.number_of_significant_sequences > 0:
max_SSTVD_gs, max_SSTVD = self.get_maximum_SSTVD()
print(" - The maximum SSTVD over all sequences is {0:.2f}".format(max_SSTVD))
if verbosity >= 2:
print(" - The maximum SSTVD was observed for {}".format(max_SSTVD_gs))
else:
print("Statistical hypothesis tests did NOT find inconsistency "
"between the datasets at {0:.2f}% significance.".format(self.significance * 100))
return
def get_TVD(self, circuit):
"""
Returns the observed total variation distacnce (TVD) for the specified circuit.
This is only possible if the comparison is between two sets of data. See Eq. (19)
in "Probing context-dependent errors in quantum processors", by Rudinger et al. for the
definition of this observed TVD.
This is a quantification for the "amount" of context dependence for this circuit (see also,
get_JSD(), get_SSTVD() and get_SSJSD()).
Parameters
----------
circuit : Circuit
The circuit to return the TVD of.
Returns
-------
float
The TVD for the specified circuit.
"""
try: assert len(self.dataset_list_or_multidataset) == 2
except: raise ValueError("The TVD is only defined for comparisons between two datasets!")
return self.tvds[circuit]
def get_SSTVD(self, circuit):
"""
Returns the "statistically significant total variation distacnce" (SSTVD) for the specified
circuit. This is only possible if the comparison is between two sets of data. The SSTVD
is None if the circuit has not been found to have statistically significant variation.
Otherwise it is equal to the observed TVD. See Eq. (20) and surrounding discussion in
"Probing context-dependent errors in quantum processors", by Rudinger et al., for more information.
This is a quantification for the "amount" of context dependence for this circuit (see also,
get_JSD(), get_TVD() and get_SSJSD()).
Parameters
----------
circuit : Circuit
The circuit to return the SSTVD of.
Returns
-------
float
The SSTVD for the specified circuit.
"""
try: assert len(self.dataset_list_or_multidataset) == 2
except: raise ValueError("Can only compute TVD between two datasets.")
assert(self.sstvds is not None), "The SSTVDS have not been calculated! Run the .implement() method first!"
return self.sstvds.get(circuit, None)
def get_maximum_SSTVD(self):
"""
Returns the maximum, over circuits, of the "statistically significant total variation distance"
(SSTVD). This is only possible if the comparison is between two sets of data. See the .get_SSTVD()
method for information on SSTVD.
Returns
-------
float
The circuit associated with the maximum SSTVD, and the SSTVD of that circuit.
"""
try: assert len(self.dataset_list_or_multidataset) == 2
except: raise ValueError("Can only compute TVD between two datasets.")
assert(self.sstvds is not None), "The SSTVDS have not been calculated! Run the .implement() method first!"
if len(self.sstvds) == 0:
return None, None
else:
index = _np.argmax(list(self.sstvds.values()))
max_sstvd_gs = list(self.sstvds.keys())[index]
max_sstvd = self.sstvds[max_sstvd_gs]
return max_sstvd_gs, max_sstvd
def get_pvalue(self, circuit):
"""
Returns the pvalue for the log-likelihood ratio test for the specified circuit.
Parameters
----------
circuit : Circuit
The circuit to return the p-value of.
Returns
-------
float
The p-value of the specified circuit.
"""
return self.pVals[circuit]
def get_pvalue_pseudothreshold(self):
"""
Returns the (multi-test-adjusted) statistical significance pseudo-threshold for the per-sequence
p-values (obtained from the log-likehood ratio test). This is a "pseudo-threshold", because it
is data-dependent in general, but all the per-sequence p-values below this value are statistically
significant. This quantity is given by Eq. (9) in "Probing context-dependent errors in quantum
processors", by Rudinger et al.
Returns
-------
float
The statistical significance pseudo-threshold for the per-sequence p-value.
"""
assert(self.pVal_pseudothreshold is not None), \
"This has not yet been calculated! Run the .implement() method first!"
return self.pVal_pseudothreshold
def get_LLR(self, circuit):
"""
Returns the log-likelihood ratio (LLR) for the input circuit.
This is the quantity defined in Eq (4) of "Probing context-dependent
errors in quantum processors", by Rudinger et al.
Parameters
----------
circuit : Circuit
The circuit to return the LLR of.
Returns
-------
float
The LLR of the specified circuit.
"""
return self.llrs[circuit]
def get_LLR_pseudothreshold(self):
"""
Returns the (multi-test-adjusted) statistical significance pseudo-threshold for the per-sequence
log-likelihood ratio (LLR). This is a "pseudo-threshold", because it is data-dependent in
general, but all LLRs above this value are statistically significant. This quantity is given
by Eq (10) in "Probing context-dependent errors in quantum processors", by Rudinger et al.
Returns
-------
float
The statistical significance pseudo-threshold for per-sequence LLR.
"""
assert(self.llr_pseudothreshold is not None), \
"This has not yet been calculated! Run the .implement() method first!"
return self.llr_pseudothreshold
def get_JSD(self, circuit):
"""
Returns the observed Jensen-Shannon divergence (JSD) between "contexts" for
the specified circuit. The JSD is a rescaling of the LLR, given by dividing
the LLR by 2*N where N is the total number of counts (summed over contexts) for
this circuit. This quantity is given by Eq (15) in "Probing context-dependent
errors in quantum processors", Rudinger et al.
This is a quantification for the "amount" of context dependence for this circuit (see also,
get_TVD(), get_SSTVD() and get_SSJSD()).
Parameters
----------
circuit : Circuit
The circuit to return the JSD of
Returns
-------
float
The JSD of the specified circuit.
"""
return self.jsds[circuit]
def get_JSD_pseudothreshold(self):
"""
Returns the statistical significance pseudo-threshold for the Jensen-Shannon divergence (JSD)
between "contexts". This is a rescaling of the pseudo-threshold for the LLR, returned by the
method .get_LLR_pseudothreshold(); see that method for more details. This threshold is also given by
Eq (17) in "Probing context-dependent errors in quantum processors", by Rudinger et al.
Note that this pseudo-threshold is not defined if the total number of counts (summed over
contexts) for a sequence varies between sequences.
Returns
-------
float
The pseudo-threshold for the JSD of a circuit, if well-defined.
"""
assert(self.fixed_totalcount_data), \
"The JSD only has a pseudo-threshold when there is the same number of total counts per sequence!"
assert(self.jsd_pseudothreshold is not None), \
"This has not yet been calculated! Run the .implement() method first!"
return self.jsd_pseudothreshold
def get_SSJSD(self, circuit):
"""
Returns the "statistically significanet Jensen-Shannon divergence" (SSJSD) between "contexts" for
the specified circuit. This is the JSD of the circuit (see .get_JSD()), if the circuit
has been found to be context dependent, and otherwise it is None. This quantity is the JSD version
of the SSTVD given in Eq. (20) of "Probing context-dependent errors in quantum processors", by Rudinger
et al.
This is a quantification for the "amount" of context dependence for this circuit (see also,
get_TVD(), get_SSTVD() and get_SSJSD()).
Parameters
----------
circuit : Circuit
The circuit to return the JSD of
Returns
-------
float
The JSD of the specified circuit.
"""
assert(self.llr_pseudothreshold is not None), \
"The hypothsis testing has not been implemented yet! Run the .implement() method first!"
if self.results.hypothesis_rejected[circuit]:
return self.jsds[circuit]
else:
return None
def get_aggregate_LLR(self):
"""
Returns the "aggregate" log-likelihood ratio (LLR), comparing the null
hypothesis of no context dependence in *any* sequence with the full model
of arbitrary context dependence. This is the sum of the per-sequence LLRs, and
it is defined in Eq (11) of "Probing context-dependent errors in
quantum processors", by Rudinger et al.
Returns
-------
float
The aggregate LLR.
"""
return self.aggregate_llr
def get_aggregate_LLR_threshold(self):
"""
Returns the (multi-test-adjusted) statistical significance threshold for the
"aggregate" log-likelihood ratio (LLR), above which this LLR is significant.
See .get_aggregate_LLR() for more details. This quantity is the LLR version
of the quantity defined in Eq (14) of "Probing context-dependent errors in
quantum processors", by Rudinger et al.
Returns
-------
float
The threshold above which the aggregate LLR is statistically significant.
"""
assert(self.aggregate_llr_threshold is not None), \
"This has not yet been calculated! Run the .implement() method first!"
return self.aggregate_llr_threshold
def get_aggregate_pvalue(self):
"""
Returns the p-value for the "aggregate" log-likelihood ratio (LLR), comparing the null
hypothesis of no context dependence in any sequence with the full model of arbitrary
dependence. This LLR is defined in Eq (11) in "Probing context-dependent errors in
quantum processors", by Rudinger et al., and it is converted to a p-value via Wilks'
theorem (see discussion therein).
Note that this p-value is often zero to machine precision, when there is context dependence,
so a more useful number is often returned by get_aggregate_nsigma() (that quantity is equivalent to
this p-value but expressed on a different scale).
Returns
-------
float
The p-value of the aggregate LLR.
"""
return self.aggregate_pVal
def get_aggregate_pvalue_threshold(self):
"""
Returns the (multi-test-adjusted) statistical significance threshold for the p-value of
the "aggregate" log-likelihood ratio (LLR), below which this p-value is significant.
See the .get_aggregate_pvalue() method for more details.
Returns
-------
float
The statistical significance threshold for the p-value of the "aggregate" LLR.
"""
assert(self.aggregate_pVal_threshold is not None), \
"This has not yet been calculated! Run the .implement() method first!"
return self.aggregate_pVal_threshold
def get_aggregate_nsigma(self):
"""
Returns the number of standard deviations above the context-independent mean that the "aggregate"
log-likelihood ratio (LLR) is. This quantity is defined in Eq (13) of "Probing context-dependent
errors in quantum processors", by Rudinger et al.
Returns
-------
float
The number of signed standard deviations of the aggregate LLR .
"""
return self.aggregate_nsigma
def get_aggregate_nsigma_threshold(self):
"""
Returns the (multi-test-adjusted) statistical significance threshold for the signed standard
deviations of the the "aggregate" log-likelihood ratio (LLR). See the .get_aggregate_nsigma()
method for more details. This quantity is defined in Eq (14) of "Probing context-dependent errors
in quantum processors", by Rudinger et al.
Returns
-------
float
The statistical significance threshold above which the signed standard deviations
of the aggregate LLR is significant.
"""
assert(self.aggregate_nsigma_threshold is not None), \
"This has not yet been calculated! Run the .implement() method first!"
return self.aggregate_nsigma_threshold
def get_worst_circuits(self, number):
"""
Returns the "worst" circuits that have the smallest p-values.
Parmeters
---------
number : int
The number of circuits to return.
Returns
-------
List
A list of tuples containing the worst `number` circuits along
with the correpsonding p-values.
"""
worst_strings = sorted(self.pVals.items(), key=lambda kv: kv[1])[:number]
return worst_strings