You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, all Lanczos-based functions of matrices use an eigenvalue decomposition to evaluate f(H)v.
For some functions, e.g. log and exp, alternatives may be more reliable.
I propose either providing a matfun_dense: Callable option to matrix-function-vector products, or hard-code log, exp, sqrt, and perhaps others in funm.
To decide which ones to hard-code, we could refer to what jax.scipy.linalg offers in terms of dense matrix-functions.
The text was updated successfully, but these errors were encountered:
#201 takes care of most of this, but leaves some open todos for applications in trace estimation (which is why this issue remains open for a bit longer).
Edit: the open todo is extending e.g. integrand_funm_sym to expect different dense matrix functions. Currently, this is not the case.
Currently, all Lanczos-based functions of matrices use an eigenvalue decomposition to evaluate f(H)v.
For some functions, e.g. log and exp, alternatives may be more reliable.
I propose either providing a
matfun_dense: Callable
option to matrix-function-vector products, or hard-code log, exp, sqrt, and perhaps others infunm
.To decide which ones to hard-code, we could refer to what jax.scipy.linalg offers in terms of dense matrix-functions.
The text was updated successfully, but these errors were encountered: