Handle qudits in drop_terminal_measurements

[daxfohl]

Explicitly set the dimension of the X's and I's corresponding to the invert_mask in drop_terminal_measurements to avoid the ValueError that was raised when the invert_mask was applied to qudits.

Fixes #5994

Note the weird MatrixGate swaps the 0 and 1 dimensions on qudits, leaving other dimensions unchanged. This behavior is consistent with the invert_mask implementation at https://github.com/quantumlib/cirq/blob/3fefe2984a1203c0bf647c1ea84f4882b05f8477/cirq-core/cirq/sim/simulator.py#L823, which has been in place since 2020. If there's a DimensionSwapGate or something, we could use that here instead, but I didn't see one.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.86%. Comparing base (0361a81) to head (91bb645).
Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6879      +/-   ##
==========================================
- Coverage   97.87%   97.86%   -0.01%     
==========================================
  Files        1084     1084              
  Lines       94420    94447      +27     
==========================================
+ Hits        92409    92435      +26     
- Misses       2011     2012       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[daxfohl]

glitchy test again? It failed on cirq-core/cirq/experiments/z_phase_calibration_test.py::test_calibrate_z_phases_no_options on Ubuntu, but works fine on my Ubuntu instance. @mhucka

[mhucka]

Yeah, I get that one too. I opened issue #6906 and will try to get to the bottom of it.

[dstrain115]

It's a little ridiculous we don't have a proper qudit X gate, but I suppose that's not the fault of this PR. LGTM.

[daxfohl]

@dstrain115 X and Z already generalize to qudits: #4919. X on qudits is equivalent to q = (q + 1) % dim. The existing (since 2019) implementation of invert_mask on a measurement doesn't behave like that on qudits though; instead it swaps the 0 and 1 states, leaving other states untouched. So that's what the behavior of the matrix gate here replicates.

I just remembered I had a PR that adapts gates to a subspace of a qudit: #4783. That would be a generalization of what we'd need here. To swap 0,1, cirq.DimensionAdapterGate(cirq.X, [(q.dim, (0, 1))]) would do the trick. I don't think this one use case calls for reviving that PR though.

[dstrain115]

@daxfohl Ah, ok. The definition I have seen of a qudit X gate is like X_0,1 which would swap the |0> and |1> subspaces. I am not sure there's a great standard for this since qudit operations are a less trod ground. Thanks for the flurry of changes lately!

[daxfohl]

@dstrain115 Yeah, there was some discussion on the original issue #3190, that both approaches were reasonable, but the group seemed to converge on the +1 mod dim approach. In particular this comment from @cduck: #3190 (comment). The way qudit X is implemented is equivalent to the old PlusGate that we had in tests everywhere, except it's an EigenGate and thus defined on non-integer exponents (and has interesting properties there, as shown in some of the tests).

The DimensionAdapterGate from the comment above could be used for the on(qudit.subspace[...]) approach (see the test_diagram test in that PR). @cduck also has a simpler but more targeted adapter SingleQuditSubspaceGate that's specific to the [0,1] subspace, from which a FlipGate that swaps the [0,1] subspace is defined in just a few lines: https://github.com/cduck/cirqtools/blob/master/cirqtools/qudit/common_gates.py#L120.

(I'm not advocating for pulling in any of those gates; just leaving this as an FYI in case it comes up in the future).