Computes the analytic gradient and a central finite-difference approximation
at the same parameter vector, then reports the relative error for each
component.
Usage
verify_gradient(par, fn, gr, eps = 1e-05)
Arguments
- par
Parameter vector to check at.
- fn
Objective function (scalar-valued).
- gr
Gradient function (returns vector same length as par).
- eps
Finite difference step size (default 1e-5).
Value
List with components:
- analytic
Analytic gradient vector.
- numeric
Numerical gradient vector.
- rel_err
Relative error for each component.
- max_rel_err
Maximum relative error across all components.