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
We should use the first version in the file (reverse nested in forward), because it's linear. Doing it with finite diffs and multiply will be quadratic (as will be using the second version in the source file).
The motivation is the optimizer (conjugate gradient trust region method) used by the following paper, which uses trust-ncg in scipi.optimize.minimize:
Expose the version of the Hessian-vector product functional from
stan-dev/math
.Here's the functional we need to expose. We want the first one that uses reverse-mode nested in forward mode):
https://github.com/stan-dev/math/blob/develop/stan/math/mix/functor/hessian_times_vector.hpp
We should use the first version in the file (reverse nested in forward), because it's linear. Doing it with finite diffs and multiply will be quadratic (as will be using the second version in the source file).
The motivation is the optimizer (conjugate gradient trust region method) used by the following paper, which uses
trust-ncg
inscipi.optimize.minimize
:https://arxiv.org/pdf/2304.05527.pdf
@rgiordan said that they used PyMC because it provided Hessian-vector products.
The text was updated successfully, but these errors were encountered: