Gauge detectors stop backpropagating at resets, leading to DEM depending on gate order

[asierpinor]

When a circuit contains a gauge detector it seems like it stops backpropagating after it hits a reset. This doesn't happen when the detector is deterministic.

Plotting with circuit.diagram("detslice-with-ops") looks like this:

The blue area is the gauge detector, and the red area is a normal detector.

As a consequence of this, errors that happen before the reset don't affect the blue gauge detector, which means that the DEM might be different depending on the order of operations.

If this is intended behavior it would be good to document it, but I would expect the behavior to be the same for deterministic and non-deterministic detectors. If so it would be great to fix it.

I uploaded the full example here: https://github.com/asierpinor/issue-stim-gauge-detectors

I should say that circuit.detecting_regions() does give the right detecting regions for both.

[Strilanc]

Oof, another reason I wish I'd never added gauge detectors.

You're right this is a bug in the context of having gauge detectors. The propagation is truncated in diagrams because it made it easier to debug broken circuits (by preventing wrong detectors at the end of the circuit from spraying color everywhere at the start of the circuit). Now I have to decide between making the normal case harder vs the corner case correct.

The diagrams are kindof never going to be good for gauge detectors, because even if the propagation continued backwards it would be showing too much. Places where D1*D2 (the actual detector) are not present will have D1 present and D2 present, and so will show colors.

[asierpinor]

I have found gauge detectors useful in circuits that have genuine non-deterministic observables such as circuits with T gates through teleportation (e.g. magic state distillation), because Stim can at least be used to derive the DEM (after replacing T by S, for example), which is useful for decoding. Or at least I haven't found a way to avoid gauge detectors for such cases. I guess Stim wasn't intended for such cases, but I assume there must be Clifford circuits that would benefit from gauge detectors in a similar way too, I just don't have examples.

So I personally don't mind so much about backpropagation for the diagrams, more for the DEM, but for the diagrams maybe truncating gauge detectors could be an optional parameter.

(In my example, D1*D2 is not a detector -- it's a silly example to capture the idea)

[Strilanc]

@asierpinor You can use OBSERVABLE_INCLUDE(0) Z0 (with Pauli term targets) to force observables into or around specific places. That might solve your use case.