Package de.labathome.trustconstr

Class HessianLinearOperator

java.lang.Object
de.labathome.trustconstr.HessianLinearOperator
All Implemented Interfaces:
BiFunction<Matrix,Object,Matrix>
public class HessianLinearOperator extends Object implements BiFunction<Matrix,Object,Matrix>
Adapter that exposes a matrix-free HessianProduct as a BiFunction returning an explicit Hessian Matrix.

Mirrors scipy's LinearOperator-wrapping of hessp in _minimize_trustregion_constr (minimize_trustregion_constr.py:36): given a callable hessp(x, p) returning H(x)*p for each vector p, the wrapper materialises H(x) as a dense n x n matrix by probing hessp along each Cartesian basis vector. This is O(n) hessp evaluations -- sufficient for the modest problem sizes trust-constr is typically used on, and a future hook for a true matrix-free path when downstream consumers can accept one.

  • Constructor Details

    • HessianLinearOperator

      public HessianLinearOperator(HessianProduct hessp, long nVars)
      Parameters:
      hessp - matrix-free Hessian-vector product evaluator (x, p, args) -> H(x)*p
      nVars - number of decision variables (= dimension of the Hessian)

  • Method Details