Add new method for single bitstring target when measuring stabilizerstate probabilities_dict #12147

DerekMiner · 2024-04-05T20:50:45Z

Allows a user to pass a single bitstring target value to get the probability value for. In cases where you have a large amount of qubits, this would take a very long time to retrieve the probability value for the single bitstring via the probabilities_dict method as it calculates all probabilities. A new method was added called probabilities_dict_from_bitstring, that allows the user to pass a single bitstring target value to retrieve a value for.
This allows using a large amount of qubits with a very quick retrieval time for the required value as it no longer has to calculate all the possibly probabilities.

According to @ShellyGarion

The main use-case that I can think of is of an n-qubit StabilizerState when n is large (say, n=100). In this case, the StabilizerState is quadratic in the number of qubits, while obtaining all the outcomes can potentially grow in some cases as 2^n (e.g. where there is an H gate on all qubits). Hence, it makes sense to take only one outcome, or several outcomes whose number grows at most polynomially in n (it doesn't make sense to take half the 2^n outcomes for example).

and @ginsparg

For a variety of purposes (including classical shadows), it is necessary to calculate only a single probability for some target bitstring via .get_probability(target)

If a user wants to retrieve multiple bitstring targets they can call the method multiple times for now, there doesn't appear to be a current use case for passing in multiple bitstrings at a time. This could easily be added without much effort, without use of caching support. Caching could be used if there was ever a use case for a large amount of bitstrings to be calculated to improve performance, but less then all possible bitstrings.

As suggested by @ShellyGarion I removed the performance measuring code that checks that it is faster to perform only a single measurement. This should be self apparent as performing tests with 200 qubits would possible take up to 2^200 outcomes (1.606938e+60) and likely wouldn't finish calculating these within our lifetime.

When I did have this performance measuring code in from the previous PR #11947, there was a huge performance increase measured with the checks I performed, roughly close to 1/amount_of_possible_outcomes of time, with a tiny bit of overhead from caching (but improved from not caching).

added new method probabilities_dict_from_bitstrings, moved logic from probabilities_dict to probabilities_dict_from_bitstrings with ability to pass target, probabilities_dict will pass a target of None.

…bability for qubit Added some timing of the tests to determine if the version with a specfic target runs faster then the non-targetted version (which must calculate all 2^n possibilities for stabilizer state test_probablities_dict_single_qubit. When given a target it can skip branches that were not wanting to be calculated

simplified the method probabilities_dict_from_bitstrings, less checking was needed, simpler handling of no target value passed in

…dict_qubits

In stabilizerstate, added the ability to cache previously calculated branches so they do not get recalculated when attempting to get target values, gives a better starting point to calculate from instead of starting from an unknown probability value of 1.0

when creating an array for the outcome in probability, it is only built if the cache key is not found

…disabling cache added more optimization for performance when performing calculations for branches using caching and targets, it will now also store the deterministic node values which gives a very slight performance increase, migrated to most efficient ways to store and retrieve cache values. The performance increase when using a large amount of qubits is huge, when using a very small amount of qubits the very slight over head of caching makes it about the same. Increasing test coverage for using targets with probabilities

implemented more accurate tests that verify probability caching

for some calculations the cached results were in the wrong place which could cause false results for certain test cases with 3+ qubits

overhead was too high compared to calculating

Fixed failing test that would randomly occur, the stabilizer state was not being cached as well, which would sometimes produce a failure. After caching the object to restore when calculating starting at a node down the branch, this failure no longer occurs. Need to optimize and clean up the code a bit, but functioning

…to test

…gle_bitstring

ShellyGarion

Thank you very much for this high quality PR @DerekMiner !
I only have a few minor comments.

releasenotes/notes/outcome_bitstring_target_for_probabilities_dict-e53f524d115bbcfc.yaml

test/python/quantum_info/states/test_stabilizerstate.py

…gle_bitstring

qiskit/quantum_info/states/stabilizerstate.py

Qiskit#12147 (comment) Qiskit#12147 (comment)

…ng' of https://github.com/DerekMiner/QisKit into add_target_stabilizerstate_probabilities_single_bitstring

Qiskit#12147 (comment) Qiskit#12147 (comment)

moved the helper function _get_probabilities_dict to the helper function area next to _get_probabilities Qiskit#12147 (comment)

Qiskit#12147 (comment)

added a test that does not have all equal probabilities and is not a GHZ state Qiskit#12147 (comment)

combine the large and medium tests into one due to taking too long to run the tests and lowered the amount of qubits

removed redundant test in test_probabilities_dict_medium_num_qubits, as this was combined with the large test

renamed test_probabilities_dict_medium_num_qubits in test_stabilizerstate to test_probabilities_dict_from_bitstring

…gle_bitstring

ShellyGarion

Looks great now, thanks for the efforts!
I only found last minor misprint.

qiskit/quantum_info/states/stabilizerstate.py

typo for description for word targetting, changed to targeting

ShellyGarion

Nice work! Thanks for your contribution to Qiskit!

…tate probabilities_dict (Qiskit#12147) * added new method, deterministic helper unction added new method probabilities_dict_from_bitstrings, moved logic from probabilities_dict to probabilities_dict_from_bitstrings with ability to pass target, probabilities_dict will pass a target of None. * Adding test timing, fixed algorithm for determining if should get probability for qubit Added some timing of the tests to determine if the version with a specfic target runs faster then the non-targetted version (which must calculate all 2^n possibilities for stabilizer state test_probablities_dict_single_qubit. When given a target it can skip branches that were not wanting to be calculated * Adding more tests, fixing deterministic issues with bad target * Prob of targets with 0.0 probability * target calc changes * Simpler way to get all targets * Need to improve performance with multiple targets * Simplified the probabilities_dict_from_bitstrings method simplified the method probabilities_dict_from_bitstrings, less checking was needed, simpler handling of no target value passed in * Adding tests for targets * Added tests to test_probablities_dict_qubits * Add performance boost check when using targets for test_probablities_dict_qubits * Added Caching to prevent duplicate branch calculations In stabilizerstate, added the ability to cache previously calculated branches so they do not get recalculated when attempting to get target values, gives a better starting point to calculate from instead of starting from an unknown probability value of 1.0 * improve performance by performing expensive operations only when needed when creating an array for the outcome in probability, it is only built if the cache key is not found * Optimization for caching, deterministic calcs for branches, enabling/disabling cache added more optimization for performance when performing calculations for branches using caching and targets, it will now also store the deterministic node values which gives a very slight performance increase, migrated to most efficient ways to store and retrieve cache values. The performance increase when using a large amount of qubits is huge, when using a very small amount of qubits the very slight over head of caching makes it about the same. Increasing test coverage for using targets with probabilities * improved tests, checking of caching vs not caching implemented more accurate tests that verify probability caching * added more tests and variations for stabilizerstate * Corrected deterministic probability calculations * Fixed bug with cache in wrong place for some calculations for some calculations the cached results were in the wrong place which could cause false results for certain test cases with 3+ qubits * added more test cases * Probabilities update for moved method call * Added test for random test_probs_random_subsystem * fixed single probabilities tests lost in previous merge * Removed caching of deterministic values overhead was too high compared to calculating * Fixed failing test with test_probs_random_subsystem Fixed failing test that would randomly occur, the stabilizer state was not being cached as well, which would sometimes produce a failure. After caching the object to restore when calculating starting at a node down the branch, this failure no longer occurs. Need to optimize and clean up the code a bit, but functioning * Adding tests back that were not test_probs_random_subsystem, removed to test * Created ProbabilityCache object for dealing with caching * Finished adding ProbabilityCache object and commenting * Added more documentation for stabilizer state * Fixing style with tox -eblack * commenting fixes in stabilizerstate * Added release note * Updated release notes * added __future__ to probabilitiescache, removed unused import stabilizerstate * run tox -eblack to correct probabilitiescache * Corrected lint issues in multiple classes fixed lint issues in test_probabilitiycache, test_stabilizerstate, stabilizerstate, and probabilitycache * fixed more pylint lint issues pylint -rn qiskit test tools Corrected in probabilitycache, stabilizerstate, test_probabilitycach, test_stabilizerstate * fixed more lint issues fixed with: tox -eblack verified with: black --check qiskit test tools examples setup.py pylint -rn qiskit test tools tox -elint * added a bit more variance allowance for performance tests renamed test_probabilitycache.py file to correct spelling added a bit more performance_varability_percent percent time allowed from 0.001 to 0.005 * Corrected Lint issues, added test for probabilitycache added test_probabilitycache test, fixed lint issues and type issues in stabilizerstate, test_stabilizerstate * Fixed remaining import ordering issue * readded ddt decorator for test_probabilitycache * Changed to using process_time_ns from monotonic for performance measuring * Added output when performance times fail to aid in debugging in test_stabilizerstate * moved performance end time out of same method to calculate times * changed method name in test_stabilizerstate changed probability_percent_of_calculated_branches to _probability_percent_of_calculated_branches * Added GC disabling between timed tests * Attempting using perf_counter_ns instead of process_time_ns * Updated commenting, moved 'X' outcome check moved 'X' in outcome out of the for loop for StabilizerState._get_probabilities to help performance * more comment updating * Changed performance timer based on OS type * Changing to time.thread_time_ns(), and raising performance varabilitiy by 0.005 using time.thread_time_ns() for benchmarking instead of time.perf_counter_ns() and time.process_time_ns() based on the OS in test_stabilizerstate. It seems that time is being counted that shouldn't be with the other methods that causes false reports of the work taking longer then it does. Raising the performance_varability_percent in test_stablilizerstate by 0.005 to 0.01 for more headroom for performance checking * changing time.thread_time() vs time.thread_time_ns() to see if compatible with windows also added one more target for test_4 to differentiate no caching against caching in test_stabilizerstate * Changing performance timer based on OS for test_stablizerstate perf_counter_ns for win32 and thread_time for all other OS * fixed small issue in probabilitycache, improved probabilitycache test fixed issue when checking the cache key that was not operating correctly, internally the key is always a str, but when a list[str] was passed in the check was not correct, so changed to check if an instance of a list, and will join the outcome val to a str for the internal cache key. Improved the probabilitycache to test having alternating keys of str and list[str] to test this internal key checking * Added correct type hinting to probabilitycache * removed typing hint module removed the use of typing module for type hinting, looks like python wants to get away from using this module for type hinting, also fixed some commenting issues in stabilizerstate, test_probabilitycache, test_stabilizerstate, and probabilitycache files * Fix test pylint failure of Value 'list' is unsubscriptable (unsubscriptable-object) fixing tests failing with pylint when lint checking is performed. This worked locally and wasn't failing, but when it runs in the CI/CD process it fails, so adding from __future__ import annotations to avoid this failure for test_probabilitycache and test_stabilizerstate as new tests include type hinting * Fix small comment error in test, Added use_caching to tests, simplified cache calls Fixed a small error in my commenting for the test_stabilizerstate.py and forced use_caching parm in the call of probabilities_dict_from_bitstrings in the tests, even tho it is used as default, to make sure it is clear in the tests. Simplfied the adding values to Probabilitycache, and simplified the code in StabilizerState._get_probabilities() setting values in cache and retrieving the state when using cache, made cache_key in probabilitycache "private" and changed method to _cache_key * Significant code refactoring in StabiizerState Significant code refactoring in stabilizerstate. I was able to condense the use the cache to a smaller area in the _get_probabilities helper method so it is easier to work with and more clear. I removed methods I created that were single use helper methods that were able to be condensed down into simpler code such as _branches_to_measure, _retrieve_deterministic_probability, _branches_to_measure, _is_qubit_deterministic. Since these are 1 time use methods it didn't make sense to keep them around. The _get_probabilities helper method is much simpler and clearer, closer to the original with less tweaks and functioning with the same performance, and more clarity. Most of the code changed at this point is adding tests, most of this commit was removing unnecessary code and refactoring * Fixed stabilizerstate cache key outcome not found there was a check to make sure a key could be found before that was removed in the refactoring, once and awhile you get a target that can't be found in the cache because of the algorithm of looking for a key by increasing the number of X from the left to the right. There are situations where the X would be in the middle such as 0X1 that won't be found. The algorithm in the probability cache could be changed to use BST which would very rarely help performance, the fix is to make sure a key can be found and then only use the cache if that is the case * Moved all caching retrieval outside of _get_probabilities for performance there is a slight overhead when recursively iterating through the _get_probabilities helper when having to check if caching should be retrieved. Realistically when you use caching it is just to get you a better starting point, so it only needs to be retrieved once. The cache inserting remains in the _get_probabilities helper to build up the cache, but the logic was significantly simplified in the _get_probabilities helper, and now only a 1 time cache retrieval is performed for each target via the probabilities_dict_from_bitstrings method * Simplified Bitstring method for probability cache, removed unnecessary classes Removed Probabilitycache class, test_probabilitiycache.py. Simplified test_stabilizerstate.py for new bitstring probability testing methods. changed method probabilities_dict_from_bitstrings to probabilities_dict_from_bitstring, and simplified by removing cache, and only allowing a single string value to be passed in * Commenting update for _probabilities_bitstring_verify updating the commenting for _probabilities_bitstring_verify, move to top of file * black lint formatter ran * Uncomment out line causing circular import when running locally I get a circular import for two_qubit_decompose.py on line 154 "_specializations = two_qubit_decompose.Specialization", I need to comment out to test locally, but did not mean to commit this. Simplified the stablizierstate self._get_probabilities method but removing an unnecessary check * Remove type hints from test_stabilizerstate * Updated test_stabilizerstate, removed helper test method, new release note update the tests in test_stabilizerstate to no longer use a helper method for checking all the bitstring measurements in. Created a _get_probabilities_dict helper method in stabilizer state to be used for the probabilities_dict and probabilities_dict_with_bitstring methods. This was needed so that they could share the logic of the measurements, but also so that you could enforce in the probabilities_dict_with_bitstring method that outcome_bitstring must be provided. Previouslly this was optional and the user could avoid passing in a value, but this would essentially make it function the same as probabilities_dict. Now with the helper it enforces they must provide a bitstring * Added large qubit test with h gates, reorder imports for test_stabilizerstate added new test method test_probabilities_dict_large_num_qubits with large amount of qubit tests using an outcome bitstring target, num_qubits=[25, 50, 100, 200, 300, 400, 500, 600, 750], normally get any results for 750 qubits would require ~5.922387e+225 full bitstring calculations, but if you want one target bitstring you can perform ~750 measurements for a single probability for one result and get the value quickly * moved methods out of loop that are unnecessary in test_stabilizerstate moded target dict calculation out of loop that is not necessary to recalculate through each sample loop calculation * broke out medium hgate test to own test method, issue with assertTrue type in test testing a variety of hgates that also do the full probabilities_dict calculations up to 9 qubits called test_probabilities_dict_hgate_medium_num_qubits, also changed a assertTrue typo to assertDictEqual in the test_probabilities_dict_large_num_qubits method * fixed commenting in stabilizerstate, removed unnecessary outcome_bitstring array access syntax fixed some commenting for probabilities_dict_from_bitstring, and removed unnecessary syntax for accessing the array, was outcome_bitstring[i : i + 1] now outcome_bitstring[i] in _get_probabilities * Condensed tests in test_stabilizerstate, added helper test class, more test coverage added class StabilizerStateTestingTools in file test_stabilizerstate, these methods needed to be moved outside of the TestStabilizerState class because they are also needed in the TestStabilizerStateExpectationValue class within the file. It did not make sense to put these helper methods in the QiskitTestCase class, or the BaseQiskitTestCase class as these helpers are specific to helping test a new method in the stabilizerstate class. I condensed redudnant code for checking the individual bitstrings into helper method _verify_individual_bitstrings which takes a target dict, uses each entry in the dict to attempt to get the single probability, and verify each, which drastically raises the test coverage of this new method probabilities_dict_from_bitstring. added helper method StabilizerStateTestingTools._bitstring_product which builds a dict of the product of 0, 1 for a lenth passed in, used in several test cases for testing the zero probabilities. In the test cases you will also notice cases where I update the target dict target.update(StabilizerStateTestingTools._bitstring_product( num_qubits, target)), which builds the rest of the bitstring values that will have a 0 probability, which is then passed into the _verify_individual_bitstrings method to verify they get the correct 0 value I reduced the number of qubits for the test_probabilities_dict_hgate_large_num_qubits from tests of num_qubits=[25, 50, 100, 200, 300, 400, 500, 600, 750] to num_qubits=[25, 50, 100, 200] due to memory issues on windows, and the largest use case I have received so far is 100 qubits, so going above that by 2x Overall test coverage should be dramatically increased now testing for 0 probabilities, checking every bitstring individually has a probability, and simplifying the code, removing a lot of repetative logic, moved to the helper methods * Fixed isues in release notes Qiskit#12147 (comment) Qiskit#12147 (comment) * Corrected Targetted to targeted in stablizerstate * Updated commenting about target Qiskit#12147 (comment) Qiskit#12147 (comment) * Moved helper function _get_probabilities_dict moved the helper function _get_probabilities_dict to the helper function area next to _get_probabilities Qiskit#12147 (comment) * differentiated description for probabilities Qiskit#12147 (comment) * Added test to test_stabilizerstate added a test that does not have all equal probabilities and is not a GHZ state Qiskit#12147 (comment) * combined 2 test methods combine the large and medium tests into one due to taking too long to run the tests and lowered the amount of qubits * removed unnecessary test removed redundant test in test_probabilities_dict_medium_num_qubits, as this was combined with the large test * fixed lint issues, renamed test method renamed test_probabilities_dict_medium_num_qubits in test_stabilizerstate to test_probabilities_dict_from_bitstring * fixed type for targetting in stabilizerstate typo for description for word targetting, changed to targeting --------- Co-authored-by: Shelly Garion <46566946+ShellyGarion@users.noreply.github.com>

DerekMiner added 30 commits February 23, 2024 09:41

added new method, deterministic helper unction

82716cd

added new method probabilities_dict_from_bitstrings, moved logic from probabilities_dict to probabilities_dict_from_bitstrings with ability to pass target, probabilities_dict will pass a target of None.

Adding more tests, fixing deterministic issues with bad target

0407cf9

Prob of targets with 0.0 probability

d68f819

target calc changes

4f2e508

Simpler way to get all targets

81a8b88

Need to improve performance with multiple targets

2bcdff0

Simplified the probabilities_dict_from_bitstrings method

24ae34b

simplified the method probabilities_dict_from_bitstrings, less checking was needed, simpler handling of no target value passed in

Adding tests for targets

57cd066

Added tests to test_probablities_dict_qubits

6811775

Add performance boost check when using targets for test_probablities_…

09835fb

…dict_qubits

improve performance by performing expensive operations only when needed

6ba22ef

when creating an array for the outcome in probability, it is only built if the cache key is not found

improved tests, checking of caching vs not caching

4d0a2b5

implemented more accurate tests that verify probability caching

added more tests and variations for stabilizerstate

6cee48e

Corrected deterministic probability calculations

37316cf

Fixed bug with cache in wrong place for some calculations

8734b60

for some calculations the cached results were in the wrong place which could cause false results for certain test cases with 3+ qubits

added more test cases

baf39ff

Merge branch 'main' into add_target_stablizerstate_probabilities_simple

3b5c068

Probabilities update for moved method call

5589458

Added test for random test_probs_random_subsystem

213391a

fixed single probabilities tests lost in previous merge

c429bcc

Removed caching of deterministic values

7970128

overhead was too high compared to calculating

Adding tests back that were not test_probs_random_subsystem, removed …

79a4731

…to test

Created ProbabilityCache object for dealing with caching

acc9569

Finished adding ProbabilityCache object and commenting

e738982

Added more documentation for stabilizer state

b40142b

Merge branch 'main' into add_target_stablizerstate_probabilities_simple

a28f332

Merge branch 'main' into add_target_stabilizerstate_probabilities_sin…

47e0a05

…gle_bitstring

DerekMiner changed the title ~~[WIP] Add new method for single bitstring target when measuring stabilizerstate probabilities_dict~~ Add new method for single bitstring target when measuring stabilizerstate probabilities_dict Apr 18, 2024