Why I am continuously getting "failed to decompose hyperedges into edges" in Stim?

Question

I have written a Hypergraph product quantum error correction circuit on Stim. The code is $[\![65,9,4]\!]$. I have put circuit level noise and I am using the sinter BPOSD decoder package to decode.

I have set decompose_errors = True but even after that I continue to get the error failed to decompose hyperedges into edges. What could be the reason, and how can I resolve it?

Answer 1

It sounds like you don't want decomposition and so should set decompose_errors=False (the default in stim).

What version of sinter are you using? When you don't give it a dem it will try to use decomposition when generating one automatically, but it should be falling back to not decomposing when that fails. But this logic has changed over time.

Answer 2

It seems that your circuit experiment includes some error mechanisms that flip more than two detectors.

This means your syndrome decoding graph is actually an hypergraph, with some hyperedges connecting more than 2 vertices. These hyperedges can be a property of hypergraph product codes or a consequence of the circuit you consider (e.g. hyperedges appear in some experiments with the surface code).

Stim's default decoders are Minimal Weight Perfect Matching decoders, which only works with circuit experiments involving no hyperedges. The decompose_errors=True option uses a heuristic method that can in some cases replace the hyperedges with regular edges, but won't work in general. It seems that stim cannot decompose the hyperedges in your case.

The BpOsdDecoder (and probably the BeliefFindDecoder) from the ldpc library you linked should be able to handle syndrome decoding hypergraphs. The main issue appears to be that the Sinter version of this decoder does not let you choose to allow stim.DetectorErrorModel with hyperedges.

(To be precise, the function called at line 98 of this file would have to pass a second True-valued argument, see its definition, line 61)

I can see no direct solution that would go around it.

You can probably create an issue on GitHub to ask about this choice to not expose this option.

As a possible workaround, you can try to modify the SinterBpOsdDecoder class to pass a second True argument, by copy/pasting/editing its definition directly in your code.