`psd_solve` leads to numerical errors
calebweinreb opened this issue · 0 comments
calebweinreb commented
The current implementation of psd_solve
leads to numerical instability in some cases and could be improved:
- Currently 1e-6 is added uniformly across the whole matrix. It would be more stable to just boost the diagonal.
- Currently
jnp.linalg.solve
is used which doesn't take advantage of the PSD structure.