Add Sampler API for expectation values #3910

95-martin-orion · 2021-03-12T21:52:22Z

Stage 2 of the Expectation Value API (#3492). This is a rewrite of #3700.

This PR introduces the sample_expectation_values method and its default implementation, which injects gates to ensure measurement occurs in the correct basis. Hardware and software samplers can use this method as-is or implement their own version for custom behavior (e.g. server-side calculation of expectation values).

mpharrigan

(I would have marked this "comment" but then the "see changes since last reviewed" doesn't work, so you get the angry red X)

It surprised me to see this implemented this way. This is halfway towards the sophisticated cirq.work set of routines. I would have expected this to serve as a simple, reference implementation. Specifically, I was surprised to see grouping and parameterized circuits showing up here, although I'm delighted that some of the functionality in cirq.work was useful. In that vein, we should converge on how to turn the results from a sampler to expectation values. I'm not so sure about the current bit twiddling in this PR.

The reason why I would have expected a very simple "reference implementation" is that if you wanted any degree of sophistication, it would probably make sense to just wrap the high-level cirq.work entry functions. That is, this function body would be like two lines. Of course that high-level API isn't merged yet. You can preview it e.g. https://github.com/quantumlib/Cirq/pull/3647/files#diff-da3dfc139c4c6d4a898a2ad1b919f0e3c72880da2c2606288077bb15ae80b630R562 here. (Line 562). The upside is the signature (given the rest of the default arguments) matches the API you've proposed here.

Edit: none of this to say that the current PR is bad. It looks pretty straightforward.

cirq/work/sampler.py

mpharrigan · 2021-03-17T22:16:33Z

cirq/work/sampler.py

+        obs_pstrings: List[ops.PauliString] = []
+        for psum in psums:
+            pstring_aggregate = {}
+            for pstring in psum:
+                pstring_aggregate.update(pstring.items())
+            obs_pstrings.append(ops.PauliString(pstring_aggregate))


what is this section doing? Factor out into a private helper function?

The purpose of this is to get InitObsSettings for each PauliSum. Since observables_to_settings only accepts PauliStrings, I perform this conversion for each PauliSum provided.

If it makes sense to you, I could fold this into observables_to_settings and have it accept a mix of PauliString and PauliSum as observables.

cirq/work/sampler.py

mpharrigan · 2021-03-17T22:25:14Z

cirq/work/sampler.py

+            for sweep_idx, pr in enumerate(study.to_resolvers(input_params)):
+                subseries = series
+                for k in pr:
+                    subseries = subseries & (samples[k] == pr.value_of(k))
+                int_results = samples.loc[subseries, 'z']
+                for setting in grouped_settings:
+                    obs_idx = obs_settings.index(setting)
+                    psum = psums[obs_idx]
+                    ev = 0
+                    for pstr in psum:
+                        sum_val = 0
+                        for int_result in int_results:
+                            val = 1
+                            for q in pstr.qubits:
+                                bit_idx = num_qubits - qmap[q] - 1
+                                if (int_result >> bit_idx) & 1:
+                                    val *= -1
+                            sum_val += val
+                        ev += sum_val * pstr.coefficient
+                    results[sweep_idx][obs_idx] = ev / num_samples


this section is very difficult to read and I suspect much slower than a vectorized numpy solution would be. Maybe factoring parts into helpfully-named helper functions or comments could go a long way?

Since you're already using some of the cirq.work functionality, consider using observable_measurement_data._stats_from_measurements

Done. Also split off a helper function, but with _stats_from_measurements there's not much left to split off.

95-martin-orion · 2021-03-18T17:48:02Z

In addition to the changes noted above, I added a plug in the method description for the full observable measurements toolkit, including the selling points listed in #2781.

mpharrigan · 2021-03-19T00:14:11Z

For Simulators that implement the Sampler interface, have you considered just using the analytic expectation values and drawing samples from a binomial distribution?

95-martin-orion · 2021-03-19T15:15:58Z

For Simulators that implement the Sampler interface, have you considered just using the analytic expectation values and drawing samples from a binomial distribution?

I suppose we could do this? It seems orthogonal to this PR, though - those simulators have access to this implementation for now (since SimulatesSamples inherits from Sampler) and can override it if a faster method is required.

As an aside: what is the use case for this? IIUC, such a simulator could implement simulate_expectation_values and return the exact EVs instead of "sampling" inexact EVs this way.

dstrain115 · 2021-04-08T13:02:39Z

cirq/work/sampler.py

+from cirq.work import group_settings_greedy, observables_to_settings
+from cirq.work.observable_measurement import (
+    _get_params_for_setting,
+    _with_parameterized_layers,


This is bad practice to import private functions from other modules. It breaks naming assumptions. We should either change the name to indicate that it is no longer private or remove the import (or create a new method that is public in those modules).

I agree, although this behavior is consistent with the observable measurements imports (all files are under cirq/work):

Cirq/cirq/work/observable_measurement_data.py

Lines 23 to 29 in 845836a

from cirq.work.observable_settings import (

InitObsSetting,

_max_weight_observable,

_max_weight_state,

_MeasurementSpec,

zeros_state,

)

Should I send a "prequel" PR that fixes this before moving forward with this one?

dstrain115 · 2021-04-08T13:03:30Z

cirq/work/sampler.py

+    param_resolver: 'cirq.ParamResolverOrSimilarType',
+    mask: Optional[pd.Series] = None,
+) -> pd.Series:
+    """Generates a 'mask' for a given parameterization of a sample.


What is a mask for a parameterization? I think this should be explained.

dstrain115 · 2021-04-08T13:06:07Z

cirq/work/sampler.py

+
+        This is a minimal implementation for measuring observables, and is best
+        reserved for simple use cases. For more complex use cases, consider
+        upgrading to `cirq/work/observable_measurement.py`. Additional features


Change to the user visible name instead of the source code name, such as cirq.work.measure_grouped_settings

dstrain115 · 2021-04-08T13:08:30Z

cirq/work/sampler.py

+              number of repetitions
+            - Readout error symmetrization and mitigation
+
+        This method can be run on any device or simulator that supports circuit


Consider reordering or clarifying this. It's unclear which "this method" you are referring to (observable measurement or sample_expectation_values). Maybe consider putting the pointing to observable measurement at the end. I think that makes more sense to explain what this function does before pointing out alternatives.

Clarified which method this refers to. I prefer to keep this paragraph at the top of the docstring, since all but the simplest use cases can benefit from using the Observable Measurement toolkit instead.

To that end, there's an argument to be made for dropping this API altogether: it doesn't provide functionality outside what will be possible with the Observable Measurement toolkit, and potentially distracts users from the "correct" tools for getting expectation values. As it stands, it primarily serves as a simple reference implementation, as pointed out above.

dstrain115 · 2021-04-08T13:11:12Z

cirq/work/sampler.py

+        qubits = ops.QubitOrder.DEFAULT.order_for(program.all_qubits())
+        qmap = {q: i for i, q in enumerate(qubits)}
+        num_qubits = len(qubits)
+        psums = (


Optional: consider avoiding abbreviations and use pauli_sums to be more clear. Optional since this does make the pauli_string_aggregate variable name kind of long.

95-martin-orion · 2021-04-13T23:18:05Z

Discussed offline: putting this on hold in lieu of full Observable Measurement toolkit. To avoid distracting users from that toolkit, we may end up dropping this PR.

balopat · 2021-07-01T16:16:42Z

ping @95-martin-orion - if the decision is to go with the OM toolkit, maybe you could team up with Matt to clean up those PRs? :)

95-martin-orion · 2021-07-01T16:46:14Z

if the decision is to go with the OM toolkit, maybe you could team up with Matt to clean up those PRs? :)

Can confirm that this was the decision we reached. Reading #3647 it looks like the next PR is still in development - @mpharrigan feel free to assign me to review on that when it comes out.

95-martin-orion · 2021-07-01T16:46:41Z

Closing this PR as discontinued.

Add Sampler API for expectation values

f1b3727

95-martin-orion requested a review from mpharrigan March 12, 2021 21:52

95-martin-orion requested review from cduck, vtomole and a team as code owners March 12, 2021 21:52

google-cla bot added the cla: yes Makes googlebot stop complaining. label Mar 12, 2021

This was referenced Mar 17, 2021

Expectation layer won't work with remote simulators tensorflow/quantum#490

Closed

SampledExpectation layer with Cirq backend tensorflow/quantum#506

Closed

mpharrigan requested changes Mar 17, 2021

View reviewed changes

Use more functions.

e4de6a7

95-martin-orion requested a review from mpharrigan March 25, 2021 01:01

95-martin-orion added the Ready for Re-Review For when reviewers take their time. label Apr 5, 2021

dstrain115 reviewed Apr 8, 2021

View reviewed changes

balopat removed the Ready for Re-Review For when reviewers take their time. label Jun 28, 2021

balopat assigned mpharrigan Jul 1, 2021

95-martin-orion closed this Jul 1, 2021

mpharrigan added the area/expectation-value label Aug 17, 2021

mpharrigan mentioned this pull request Aug 17, 2021

[Obs] 4.5 - High-level API #4392

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add Sampler API for expectation values #3910

Add Sampler API for expectation values #3910

95-martin-orion commented Mar 12, 2021

mpharrigan left a comment •

edited

Loading

mpharrigan Mar 17, 2021

95-martin-orion Mar 18, 2021

mpharrigan Mar 17, 2021

95-martin-orion Mar 18, 2021

95-martin-orion commented Mar 18, 2021

mpharrigan commented Mar 19, 2021

95-martin-orion commented Mar 19, 2021

dstrain115 Apr 8, 2021

95-martin-orion Apr 12, 2021

dstrain115 Apr 8, 2021

dstrain115 Apr 8, 2021

dstrain115 Apr 8, 2021

95-martin-orion Apr 12, 2021

dstrain115 Apr 8, 2021

95-martin-orion commented Apr 13, 2021

balopat commented Jul 1, 2021

95-martin-orion commented Jul 1, 2021

95-martin-orion commented Jul 1, 2021

	from cirq.work.observable_settings import (
	InitObsSetting,
	_max_weight_observable,
	_max_weight_state,
	_MeasurementSpec,
	zeros_state,
	)

Add Sampler API for expectation values #3910

Add Sampler API for expectation values #3910

Conversation

95-martin-orion commented Mar 12, 2021

mpharrigan left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

95-martin-orion commented Mar 18, 2021

mpharrigan commented Mar 19, 2021

95-martin-orion commented Mar 19, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

95-martin-orion commented Apr 13, 2021

balopat commented Jul 1, 2021

95-martin-orion commented Jul 1, 2021

95-martin-orion commented Jul 1, 2021

mpharrigan left a comment •

edited

Loading