Typically gauge optimizations are performed with respect to the set of ideal target gates and spam operations. This is convenient, since you need to specify the ideal targets as points of comparison, but not always the best approach. Particularly when you expect all or some of the gate estimates to either substantially differ from the ideal operations or differ, even by small amounts, in particular ways from the ideal operations, it can be hugely aid later interpretation to specify a non-ideal Model
as the target for gauge-optimization. By separating the "ideal targets" from the "gauge optimization targets", you're able to tell the gauge optimizer what gates you think you have, including any known errors. This can result in a gauge-optimized estimate which is much more sensible and straightforward to interpet.
For example, gauge transformations can slosh error between the SPAM operations and the non-unital parts of gates. If you know your gates are slightly non-unital you can include this information in the gauge-optimization-target (by specifying a Model
which is slightly non-unital) and obtain a resulting estimate of low SPAM-error and slightly non-unital gates. If you just used the ideal (unital) target gates, the gauge-optimizer, which is often setup to care more about matching gate than SPAM ops, could have sloshed all the error into the SPAM ops, resulting in a confusing estimate that indicates perfectly unital gates and horrible SPAM operations.
This example demonstrates how to separately specify the gauge-optimization-target Model
. There are two places where you might want to do this: 1) when calling pygsti.run_long_sequence_gst
, to direct the gauge-optimization it performs, or 2) when calling estimate.add_gaugeoptimized
to add a gauge-optimized version of an estimate after the main GST algorithms have been run.
In both cases, a dictionary of gauge-optimization "parameters" (really just a dictionary of arguments for pygsti.gaugeopt_to_target
) is required, and one simply needs to set the targetModel
argument of pygsti.gaugeopt_to_target
by specifying targetModel
within the parameter dictionary. We demonstrate this below.
First, we'll setup a standard GST analysis as usual except we'll create a mdl_guess
model that is meant to be an educated guess at what we expect the estimate to be. We'll gauge optimize to mdl_guess
instead of the usual target_model
:
import pygsti
from pygsti.modelpacks import smq1Q_XYI
#Generate some fake data (all usual stuff here)
exp_design = smq1Q_XYI.create_gst_experiment_design(max_max_length=4)
mdl_datagen = smq1Q_XYI.target_model().depolarize(op_noise=0.1, spam_noise=0.001)
ds = pygsti.data.simulate_data(mdl_datagen, exp_design.all_circuits_needing_data, num_samples=1000, seed=1234)
data = pygsti.protocols.ProtocolData(exp_design, ds)
mdl_guess = smq1Q_XYI.target_model()
mdl_guess[('Gxpi2',0)].depolarize(0.1)
# GST with standard "ideal target" gauge optimization
results1 = pygsti.protocols.StandardGST("full TP").run(data)
# GST with our guess as the gauge optimization target
gaugeopt_suite = pygsti.protocols.GSTGaugeOptSuite(gaugeopt_suite_names=['stdgaugeopt'],
gaugeopt_target=mdl_guess)
results2 = pygsti.protocols.StandardGST("full TP", gaugeopt_suite).run(data)
After running both the "ideal-target" and "mdl_guess-target" gauge optimizations, we can compare them with the ideal targets and the data-generating gates themselves. We see that using mdl_guess
results in a similar frobenius distance to the ideal targets, a slightly closer estimate to the data-generating model, and reflects our expectation that the Gx
gate is slightly worse than the other gates.
target_model = smq1Q_XYI.target_model()
mdl_1 = results1.estimates['full TP'].models['stdgaugeopt']
mdl_2 = results2.estimates['full TP'].models['stdgaugeopt']
print("Diff between ideal and ideal-target-gauge-opt = ", mdl_1.frobeniusdist(target_model))
print("Diff between ideal and mdl_guess-gauge-opt = ", mdl_2.frobeniusdist(target_model))
print("Diff between data-gen and ideal-target-gauge-opt = ", mdl_1.frobeniusdist(mdl_datagen))
print("Diff between data-gen and mdl_guess-gauge-opt = ", mdl_2.frobeniusdist(mdl_datagen))
print("Diff between ideal-target-GO and mdl_guess-GO = ", mdl_1.frobeniusdist(mdl_2))
print("\nPer-op difference between ideal and ideal-target-GO")
print(mdl_1.strdiff(target_model))
print("\nPer-op difference between ideal and mdl_guess-GO")
print(mdl_2.strdiff(target_model))
Results
¶We can also include our mdl_guess
as the targetModel
when adding a new gauge-optimized result. See other examples for more info on using add_gaugeoptimized
.
results1.estimates['full TP'].add_gaugeoptimized(results2.estimates['full TP'].goparameters['stdgaugeopt'],
label="using mdl_guess")
mdl_1b = results1.estimates['full TP'].models['using mdl_guess']
print(mdl_1b.frobeniusdist(mdl_2)) # gs1b is the same as gs2