fraction_to_the_boundary_rule.hpp

// Copyright (c) Sleipnir contributors
 
#pragma once
 
#include <Eigen/Core>
 
// See docs/algorithms.md#Works_cited for citation definitions
 
namespace slp {
 
template <typename Scalar>
Scalar fraction_to_the_boundary_rule(
    const Eigen::Vector<Scalar, Eigen::Dynamic>& x,
    const Eigen::Vector<Scalar, Eigen::Dynamic>& p, Scalar τ) {
  // The fraction-to-the-boundary rule is defined as:
  //
  //   α = max(α ∈ (0, 1] : x + αp ≥ (1 − τ)x)   (1)
  //
  // where x and τ are positive. Rearranging the inequality in (1) gives
  //
  //   x + αp ≥ (1 − τ)x
  //   x + αp ≥ x − τx
  //   αp ≥ −τx                                  (2)
  //
  // (2) is false if p < 0 and α is sufficiently large. Let p < 0.
  //
  //   αp ≥ −τx
  //   α ≤ −τxᵢ/pᵢ for i in range(x.rows())      (3)
  //
  // When (2) is false, find the largest α for which (3) is true.
  Scalar α(1);
  for (int i = 0; i < x.rows(); ++i) {
    if (α * p[i] < -τ * x[i]) {
      // α = −τx/p is (3)'s upper bound
      α = -τ * x[i] / p[i];
    }
  }
 
  return α;
}
 
}  // namespace slp