Stray edges in squared edge attribution / numerical issues with large Lambda[0] values

ApolloResearch / rib

Library for methods related to the Local Interaction Basis (LIB)

MIT License

2 stars 0 forks source link

Stray edges in squared edge attribution / numerical issues with large Lambda[0] values #276

Open stefan-apollo opened 6 months ago

stefan-apollo commented 6 months ago

As seen in OP report, the "squared" bases has some stray edges that the "functional" basis does not have.

Presumably due to wrong Lambdas.

stefan-apollo commented 6 months ago

We noticed that in the Jacobian code with new Lambdas we get a similar phenomenon of stray edges if Lambda[0] is large such as 1e9. See this comment. Either

We fixed the above bug and introduced a new one with a similar effect
Or it was the same bug all along

nix-apollo commented 5 months ago

This sometimes has to do with badly calculated integrals (because of low n_intervals). For instance, the integrals across ln_out are scaled by a constant factor equal to integral 1/sqrt(alpha) from 0 to 1 (=2). But with low number of n_intervals this integral can be evaluated as ~30 or so. This constant factor is squared when computing lambdas and edges, and leads to whacky values.

Some more info here.