finite-difference.hpp

Go to the documentation of this file.

 
#pragma once
 
#include "proxsuite-nlp/function-base.hpp"
 
namespace proxsuite {
namespace nlp {
namespace autodiff {
 
enum FDLevel {
  TOC1 = 0, 
  TOC2 = 1  
};
 
enum FDType {
  BACKWARD, 
  CENTRAL,  
  FORWARD   
};
 
namespace internal {
 
template <typename Scalar> struct finite_difference_impl {
  PROXSUITE_NLP_DYNAMIC_TYPEDEFS(Scalar);
 
  static void computeJacobian(const ManifoldAbstractTpl<Scalar> &space,
                              const BaseFunctionTpl<Scalar> &func,
                              const Scalar fd_eps, const ConstVectorRef &x,
                              MatrixRef Jout) {
    VectorXs ei(func.ndx());
    VectorXs xplus = space.neutral();
    VectorXs xminus = space.neutral();
    ei.setZero();
    for (int i = 0; i < func.ndx(); i++) {
      ei(i) = fd_eps;
      space.integrate(x, ei, xplus);
      space.integrate(x, -ei, xminus);
      Jout.col(i) = (func(xplus) - func(xminus)) / (2 * fd_eps);
      ei(i) = 0.;
    }
  }
 
  static void vectorHessianProduct(const ManifoldAbstractTpl<Scalar> &space,
                                   const C1FunctionTpl<Scalar> &func,
                                   const Scalar fd_eps, const ConstVectorRef &x,
                                   const ConstVectorRef &v, MatrixRef Hout) {
    VectorXs ei(func.ndx());
    VectorXs xplus = space.neutral();
    VectorXs xminus = space.neutral();
    MatrixXs Jplus(func.nr(), func.ndx());
    MatrixXs Jminus(func.nr(), func.ndx());
    Jplus.setZero();
    Jminus.setZero();
    ei.setZero();
 
    for (int i = 0; i < func.ndx(); i++) {
      ei(i) = fd_eps;
      space.integrate(x, ei, xplus);
      space.integrate(x, -ei, xminus);
      func.computeJacobian(xplus, Jplus);
      func.computeJacobian(xminus, Jminus);
      Hout.col(i) = ((Jplus - Jminus) / (2 * fd_eps)).transpose() * v;
      ei(i) = 0.;
    }
  }
};
 
} // namespace internal
 
template <typename Scalar, FDLevel n = TOC1> struct finite_difference_wrapper;
 
template <typename _Scalar>
struct finite_difference_wrapper<_Scalar, TOC1> : C1FunctionTpl<_Scalar> {
  using Scalar = _Scalar;
  PROXSUITE_NLP_DYNAMIC_TYPEDEFS(Scalar);
 
  using Base = C1FunctionTpl<_Scalar>;
  using FuncType = BaseFunctionTpl<Scalar>;
 
  const ManifoldAbstractTpl<Scalar> &space;
  const FuncType &func;
  Scalar fd_eps;
 
  using Base::computeJacobian;
 
  finite_difference_wrapper(const ManifoldAbstractTpl<Scalar> &space,
                            const FuncType &func, const Scalar fd_eps)
      : Base(space, func.nr()), space(space), func(func), fd_eps(fd_eps) {}
 
  VectorXs operator()(const ConstVectorRef &x) const override {
    return func(x);
  }
 
  void computeJacobian(const ConstVectorRef &x, MatrixRef Jout) const override {
    return internal::finite_difference_impl<Scalar>::computeJacobian(
        space, func, fd_eps, x, Jout);
  }
};
 
template <typename _Scalar>
struct finite_difference_wrapper<_Scalar, TOC2> : C2FunctionTpl<_Scalar> {
  using Scalar = _Scalar;
  PROXSUITE_NLP_DYNAMIC_TYPEDEFS(Scalar);
 
  using Base = C2FunctionTpl<_Scalar>;
  using FuncType = C1FunctionTpl<Scalar>;
 
  const ManifoldAbstractTpl<Scalar> &space;
  const FuncType &func;
  Scalar fd_eps;
 
  using Base::computeJacobian;
 
  finite_difference_wrapper(const ManifoldAbstractTpl<Scalar> &space,
                            const FuncType &func, const Scalar fd_eps)
      : Base(space, func.nr()), space(space), func(func), fd_eps(fd_eps) {}
 
  VectorXs operator()(const ConstVectorRef &x) const override {
    return func(x);
  }
 
  void computeJacobian(const ConstVectorRef &x, MatrixRef Jout) const override {
    func.computeJacobian(x, Jout);
  }
 
  void vectorHessianProduct(const ConstVectorRef &x, const ConstVectorRef &v,
                            MatrixRef Hout) const override {
    internal::finite_difference_impl<Scalar>::vectorHessianProduct(
        space, func, fd_eps, x, v, Hout);
  }
};
 
} // namespace autodiff
} // namespace nlp
} // namespace proxsuite