Implement an SVD-truncated or regularized version of O2GF action solver

[adrn]

Describe your changes

This has been a curiosity of mine for too long, but finally decided to try implementing a test of the idea.

The main idea here starts with the O2GF action estimation method by @jlsanders. This method solves for the generating function and its derivatives that transforms from "toy" actions and angles to the "correct" action/angle values for an integrated orbit. This works by approximating the generating function as a Fourier series in the toy angles, and solves for the coefficients using a matrix solve. This is unstable when the integrated orbit is short or sparsely sampled because the design matrix can be ill-conditioned or the system underdetermined.

The demo implementation here replaces the standard linalg solve with a version that uses a low-rank approximation of A (via SVD), where the rank is set by the magnitude of the singular values. This effectively applies a low-pass filter to the Fourier modes of the generating function, which may reduce over-fitting and make the solve better conditioned.

I tried it on one orbit and it seems to work, but the performance gain over integrating an orbit for longer and using the standard method isn't huge. Specifically: orbit 1 integrated for 10 radial orbital periods with 128 steps per period, orbit 2 integrated for 4 radial orbital periods with 32 steps per period. The low-rank solve on orbit 2 was ~30% faster than the standard O2GF on orbit 1.

cc @nstarman and @jnibauer because I was mainly scoping this out as a potential addition to Galax, but now I think that standard O2GF is a first priority.

🤷

Checklist

• Did you add tests?
• Did you add documentation for your changes?
• Did you reference any relevant issues?
• Did you add a changelog entry? (see CHANGES.rst)
• Are the CI tests passing?
• Is the milestone set?