suggested extra input validation for OWD

[Krastanov]

Hi, folks,

We have been using your OWD implementation to test out our Multivariate Multicycle codes -- your tooling has been very valuable, thank you for making it.

In setting up some experiments for OWD, after running the experiments we noticed that some of the parameterization we gave was nonsensical, but the decoder silently reported "plausible" results. Here is what happened and a suggestion of ours for adding extra validation that makes this issue not "silent".

Let's say our code has c checks (i.e. a single-shot syndrome is a vector of length c). We ran a memory experiment with a single round, i.e. we had only c detectors. We incorrectly set the OWD decoder to have a window size of 2, i.e. bigger than the number of rounds, which does not make sense. Your already existing validation passed (the one linked below):

ldpc/src_python/ldpc/ckt_noise/base_overlapping_window_decoder.py

Line 41 in d342996

rounds = (self.window - self.commit) + self.decodings * self.commit

I think something that can detect the nonsensical result would be to check that the number of detectors divided by the number of checks is also integer (and equal to the number of rounds maybe).

[timohillmann]

Hi Stefan, thanks for raising this issue. Since I wrote the code for this, I've assigned myself and will try to check back on this regularly.
Can you elaborate on what plausible results mean? Also, I am trying to understand how to replicate your issue.
What are the values you chose for decodings and commit?

[Krastanov]

Hi, Timo! Regrettably we discarded that piece of code, but if I remember correctly it can be something like this:

parity_check_matrix = ... # something with 200 rows
checks = 200
detectors = 200 # corresponding to measuring the complete syndrome only once, i.e. a situation in which the number of rounds has to be 1, i.e. OWD should report an error because it can not run for a single round
commit = 1
window = 2 # impossible to achieve because we have extracted the syndrome only once
decodings = 1
rounds = 2-1+1 = 2

detectors % rounds passes, but I think we should also check detectors % checks directly, maybe even detector ÷ checks == rounds (especially given that the error message seems to imply it here

ldpc/src_python/ldpc/ckt_noise/base_overlapping_window_decoder.py

Line 47 in d342996

f"You expected {self.num_checks * rounds}"

)

[Krastanov]

cc @Fe-r-oz , just to keep you in the loop

[timohillmann]

Hi @Krastanov, I see what the issue is. I've created a PR with a safer version, see #102