stimflow `time_reversed()` doesn't support noisy measurement

[cduck]

Chunk.time_reversed() is a very handy way to time-reverse any stim circuit, including the measurements and detectors. However, it does not handle noisy measurement gates M(0.01) 1 2 3.

Example:

import stim, stimflow as sf
c = stim.Circuit('''
R 0 1
X_ERROR(0.01) 0 1
H 0
CNOT 0 1
M(0.02) 0 1
DETECTOR rec[-1] rec[-2]
''')

c_rev = sf.Chunk(c, flows=[], q2i={}).time_reversed().circuit
print(c_rev)

Actual output:

M(0.02) 1 0    <- This line is wrong
CX 0 1
H 0
X_ERROR(0.01) 0 1
M 1 0
DETECTOR rec[-2]

Expected output:

R 1 0
X_ERROR(0.02) 1 0
CX 0 1
H 0
X_ERROR(0.01) 0 1
M 1 0
DETECTOR rec[-2]

Tested with

Python 3.10.18
stim==1.16.0
stimflow==1.17.dev1784161823

[Strilanc]

This is maybe besides the point, but why are you time reversing a circuit that already has noise in it? I would have assumed that a noise model would be adding noise based on the type of gate, but time reversing doesn't preserve gate type, so you'd always want to add noise after time reversing never before.

Strictly speaking the output here is correct, in that it preserves the fault tolerance structure, but I agree it would be nicer if it turned it into a reset. I'm a bit worried that it might cause some surprising stuff to happen w.r.t. tagging (e.g. the user may expect the tag to only flow to the error instruction or expect it to only flow to the reset and so is surprised when it follows both since stim has no way to know the intended semantics). I had a preference for, as much as possible, one time forward instruction to turn into one time reversed instruction. This case would violate that... though again I agree the reset does seem like the nicer result in this case. And the user can still opt out by turning on the no-measure-into-reset option.

[Strilanc]

I think the diff is very simple btw, basically just this:

        if (!dont_turn_measurements_into_resets && rev.xs[q].empty() && rev.zs[q].empty() &&
            rev.rec_bits.contains(rev.num_measurements_in_past - 1) && inst.args.empty()) {
            // Noiseless measurements with past-dependence and no future-dependence become resets.
            inverted_circuit.safe_append(CircuitInstruction(reset, inst.args, &t, inst.tag));
        } else {

into this:

        if (!dont_turn_measurements_into_resets && rev.xs[q].empty() && rev.zs[q].empty() &&
            rev.rec_bits.contains(rev.num_measurements_in_past - 1)) {
            // Measurements with past-dependence and no future-dependence become resets.
            inverted_circuit.safe_append(CircuitInstruction(reset, {}, &t, inst.tag));
            if (!inst.args.empty()) {
                // Noise on the measurement becomes noise after the reset.
                inverted_circuit.safe_append(CircuitInstruction(err, inst.args, &t, inst.tag));
            }
        } else {

[cduck]

I agree that inverting a circuit with noise is kind of silly but it's a quick shortcut if your noise is time-symmetric anyways. But my use case today was processing circuits containing M(0) instructions (don't ask...) which also fails, despite not technically containing noise.

It would be nice if the gate translation is one-to-one so how about a R(x_error) instruction 😁?

[Strilanc]

I could certainly special case the probability=0 case with no worries about it.

[cduck]

I'm fine with that solution.

Oh, Circuit.time_reversed_for_flows() was the method I should have used. Chunk.time_reversed() doesn't have dont_turn_measurements_into_resets anyways.