Using Deterministic Errors in Stim Detector Circuit and decoding with Pymatching: X_ERROR(1) failing in PyMatching

Question

the Question

Is there a standard way to use deterministic errors (so, for instance, X_ERROR(1)) in Stim with PyMatching? Currently, sometimes if I use X_ERROR(1) in my circuit, my pymatching decoder fails with ValueError: maximum absolute edge weight of 16777215 exceeded.

I apologize if there is an obvious solution in the docs. I wasn't able to find one.

I think this failure makes sense because...

So using X_ERROR(1) sets 𝑝 = 1.

Reading what Craig explains here, "Ultimately the ... error probability 𝑝 becomes a weight 𝑤=lg(𝑝/(1−𝑝))", it makes sense X_ERROR(1) would cause problems given that lg(1/(1-1)) is a weird divide-by-zero edge case*.

And it looks like this is what is happening because, indeed, my matching graphs end up with an edge of weight: -inf. Example: [(0, None, {'fault_ids': set(), 'weight': -inf, 'error_probability': 1.0})] error.

And I think this is why I get the ValueError: maximum absolute edge weight of 16777215 exceeded

(In fact, I'm now confused why there are some circuits I can write with X_ERROR(1) that don't fail in pymatching.)

*Technically, in that post, he's talking about how pymatching combines edges. I'm making the assumption that even a single edge probability will be converted into weight: 𝑤=lg(𝑝/(1−𝑝))

Hack-arounds ...

• Running X_ERROR(0.99) does stop the error message but it makes doing fault enumeration a little weirder so I was hoping there would be another solution I just wasn't aware of.

• I can't use a normal X gate because I'm using DETECTORs (per documentation DETECTORs will treat X differently from X_ERROR(1))

A minimum failing example:

###### imports 
import stim 
import pymatching 
import numpy as np

###### Boilerplate decompose, run, and decode circuit function 
def run_decode(circuit, num_shots):
    model = circuit.detector_error_model(decompose_errors=True)
    matching = pymatching.Matching.from_detector_error_model(model)
    sampler = circuit.compile_detector_sampler()
    syndrome, actual_observables = sampler.sample(
        shots=num_shots, separate_observables=True)
    predicted_observables = matching.decode_batch(syndrome)
    error_rate = np.sum(np.any(predicted_observables !=
                        actual_observables, axis=1))/num_shots

    return error_rate


###### FAILING FUNCTION 
circ = stim.Circuit("""R 0
X 0
X_ERROR(1) 0
M 0 
DETECTOR rec[-1]""")

run_decode(circ, 10_000)

The edges of the matching graph here look like: [(0, None, {'fault_ids': set(), 'weight': -inf, 'error_probability': 1.0})] and the code fails with: ValueError: maximum absolute edge weight of 16777215 exceeded.

Please let me know if you need anymore information. I'll be happy to respond.

Answer 1

If you're trying to enumerate how errors are corrected, you shouldn't be telling pymatching which errors you are inserting. It will just end up correcting them perfectly (once this probability 1 crashing bug is fixed), because you've told it exactly which errors have happened by giving it a model that contains exactly those errors with probability 1.

What you should do is make one version of the circuit that has a normal noise model. Derive the detector error model given to pymatching from that variant of the circuit. Then have a second variant of the circuit with only your forced errors. Use that to generate the detection event samples given to pymatching.