:py:mod:`pygsti.data.datacomparator` ==================================== .. py:module:: pygsti.data.datacomparator .. autoapi-nested-parse:: Defines the DataComparator class used to compare multiple DataSets. Module Contents --------------- Classes ~~~~~~~ .. autoapisummary:: pygsti.data.datacomparator.DataComparator .. py:class:: DataComparator(dataset_list_or_multidataset, circuits='all', op_exclusions=None, op_inclusions=None, ds_names=None, allow_bad_circuits=False) A comparison between multiple data, presumably taken in different contexts. This object can be used to run 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 data, 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. 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 run 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 run 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. 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. .. py:property:: maximum_sstvd 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 .sstvd() method for information on SSTVD. Returns ------- float The circuit associated with the maximum SSTVD, and the SSTVD of that circuit. .. py:property:: pvalue_pseudothreshold Returns the (multi-test-adjusted) statistical significance pseudo-threshold for the per-sequence p-values. The p-values under consideration are those 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. .. py:property:: llr_pseudothreshold Returns the statistical significance pseudo-threshold for the per-sequence log-_likelihood ratio (LLR). This results has been multi-test-adjusted. 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. .. py:property:: jsd_pseudothreshold 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 .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. .. py:property:: aggregate_llr Returns the "aggregate" log-_likelihood ratio (LLR). This values compares 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. .. py:property:: aggregate_llr_threshold The (multi-test-adjusted) statistical significance threshold for the "aggregate" log-_likelihood ratio (LLR). Above this value, the LLR is significant. See .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. .. py:property:: aggregate_pvalue Returns the p-value for the "aggregate" log-_likelihood ratio (LLR). This compares 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 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. .. py:property:: aggregate_pvalue_threshold The (multi-test-adjusted) statistical significance threshold for the p-value of the "aggregate" LLR. Here, LLR refers to the log-_likelihood ratio. Below this p-value the LLR would be deemed significant. See the .aggregate_pvalue method for more details. Returns ------- float The statistical significance threshold for the p-value of the "aggregate" LLR. .. py:property:: aggregate_nsigma The number of standard deviations the "aggregate" LLR is above the context-independent mean. More specifically, 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 . .. py:property:: aggregate_nsigma_threshold The significance threshold above which the signed standard deviations of the aggregate LLR is significant. The (multi-test-adjusted) statistical significance threshold for the signed standard deviations of the the "aggregate" log-_likelihood ratio (LLR). See the .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. .. py:attribute:: dataset_list_or_multidataset .. py:attribute:: pVals .. py:attribute:: pVals_pseudothreshold :value: 'None' .. py:attribute:: llrs .. py:attribute:: llrs_pseudothreshold :value: 'None' .. py:attribute:: jsds .. py:attribute:: op_exclusions :value: 'None' .. py:attribute:: op_inclusions :value: 'None' .. py:attribute:: pVals0 :value: "''" .. py:attribute:: dof .. py:attribute:: num_strs :value: '0' .. py:attribute:: DS_names :value: 'None' .. py:attribute:: aggregate_pVal .. py:attribute:: aggregate_pVal_threshold :value: 'None' .. py:attribute:: sstvds :value: 'None' .. py:attribute:: pVal_pseudothreshold :value: 'None' .. py:method:: run(significance=0.05, per_circuit_correction='Hochberg', aggregate_test_weighting=0.5, pass_alpha=True, verbosity=2) Runs statistical hypothesis testing. This detects 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_circuit_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_circuit_correction` is set to the default value, or another option that controls the FWER of the per-sequence comparion (see below). per_circuit_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 .. py:method:: tvd(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, jsd(), sstvd() and ssjsd()). Parameters ---------- circuit : Circuit The circuit to return the TVD of. Returns ------- float The TVD for the specified circuit. .. py:method:: sstvd(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, jsd(), _tvd() and ssjsd()). Parameters ---------- circuit : Circuit The circuit to return the SSTVD of. Returns ------- float The SSTVD for the specified circuit. .. py:method:: pvalue(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. .. py:method:: llr(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. .. py:method:: jsd(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, _tvd(), sstvd() and ssjsd()). Parameters ---------- circuit : Circuit The circuit to return the JSD of Returns ------- float The JSD of the specified circuit. .. py:method:: ssjsd(circuit) Returns the statistically significant Jensen-Shannon divergence" (SSJSD) between "contexts" for `circuit`. This is the JSD of the circuit (see .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, _tvd(), sstvd() and ssjsd()). Parameters ---------- circuit : Circuit The circuit to return the JSD of Returns ------- float The JSD of the specified circuit. .. py:method:: worst_circuits(number) Returns the "worst" circuits that have the smallest p-values. Parameters ---------- 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.