dense_regularized_ldlt.hpp

// Copyright (c) Sleipnir contributors
 
#pragma once
 
#include <Eigen/Cholesky>
#include <Eigen/Core>
 
#include "sleipnir/optimization/solver/util/inertia.hpp"
 
// See docs/algorithms.md#Works_cited for citation definitions
 
namespace slp {
 
template <typename Scalar>
class DenseRegularizedLDLT {
 public:
  using DenseMatrix = Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic>;
  using DenseVector = Eigen::Vector<Scalar, Eigen::Dynamic>;
 
  DenseRegularizedLDLT(int num_decision_variables, int num_equality_constraints)
      : m_num_decision_variables{num_decision_variables},
        m_num_equality_constraints{num_equality_constraints} {}
 
  DenseRegularizedLDLT(int num_decision_variables, int num_equality_constraints,
                       Scalar γ_min)
      : m_num_decision_variables{num_decision_variables},
        m_num_equality_constraints{num_equality_constraints},
        m_γ_min{γ_min} {}
 
  Eigen::ComputationInfo info() const { return m_info; }
 
  DenseRegularizedLDLT& compute(const DenseMatrix& lhs) {
    // The regularization procedure is based on algorithm B.1 of [1]
 
    m_info = m_solver.compute(lhs).info();
 
    if (m_info == Eigen::Success) {
      auto D = m_solver.vectorD();
 
      // If the inertia is ideal and D from LDLT is sufficiently far from zero,
      // don't regularize the system
      if (Inertia{D} == ideal_inertia &&
          (D.cwiseAbs().array() >= Scalar(1e-4)).all()) {
        m_prev_δ = Scalar(0);
        m_prev_γ = Scalar(0);
        return *this;
      }
    }
 
    // Also regularize the Hessian. If the Hessian wasn't regularized in a
    // previous run of compute(), start at small values of δ and γ. Otherwise,
    // attempt a δ and γ half as big as the previous run so δ and γ can trend
    // downwards over time.
    Scalar δ = m_prev_δ == Scalar(0) ? Scalar(1e-4) : m_prev_δ / Scalar(2);
    Scalar γ = m_γ_min;
 
    while (true) {
      m_info = m_solver.compute(lhs + regularization(δ, γ)).info();
 
      if (m_info == Eigen::Success) {
        Inertia inertia{m_solver.vectorD()};
 
        if (inertia == ideal_inertia) {
          // If the inertia is ideal, report success
          m_prev_δ = δ;
          m_prev_γ = γ;
          return *this;
        } else if (inertia.zero > 0) {
          if (γ == Scalar(0)) {
            // If there's zero eigenvalues and γ = 0, increase γ
            γ = Scalar(1e-10);
          } else {
            // If there's zero eigenvalues and γ > 0, increase δ and γ
            δ *= Scalar(10);
            γ *= Scalar(10);
          }
        } else if (inertia.negative > ideal_inertia.negative) {
          // If there's too many negative eigenvalues, increase δ
          δ *= Scalar(10);
        } else if (inertia.positive > ideal_inertia.positive) {
          // If there's too many positive eigenvalues, increase γ
          γ = γ == Scalar(0) ? Scalar(1e-10) : γ * Scalar(10);
        }
      } else {
        // If the decomposition failed, increase δ and γ
        δ *= Scalar(10);
        γ = γ == Scalar(0) ? Scalar(1e-10) : γ * Scalar(10);
      }
 
      // If the Hessian perturbation is too high, report failure. This can be
      // caused by ill-conditioning.
      if (δ > Scalar(1e20) || γ > Scalar(1e20)) {
        m_info = Eigen::NumericalIssue;
        m_prev_δ = δ;
        m_prev_γ = γ;
        return *this;
      }
    }
  }
 
  template <typename Rhs>
  DenseVector solve(const Eigen::MatrixBase<Rhs>& rhs) const {
    return m_solver.solve(rhs);
  }
 
  template <typename Rhs>
  DenseVector solve(const Eigen::SparseMatrixBase<Rhs>& rhs) const {
    return m_solver.solve(rhs.toDense());
  }
 
  Scalar hessian_regularization() const { return m_prev_δ; }
 
  Scalar constraint_jacobian_regularization() const { return m_prev_γ; }
 
 private:
  using Solver = Eigen::LDLT<DenseMatrix>;
 
  Solver m_solver;
 
  Eigen::ComputationInfo m_info = Eigen::Success;
 
  int m_num_decision_variables = 0;
 
  int m_num_equality_constraints = 0;
 
  Scalar m_γ_min{1e-10};
 
  Inertia ideal_inertia{m_num_decision_variables, m_num_equality_constraints,
                        0};
 
  Scalar m_prev_δ{0};
 
  Scalar m_prev_γ{0};
 
  DenseMatrix regularization(Scalar δ, Scalar γ) const {
    DenseVector vec{m_num_decision_variables + m_num_equality_constraints};
    vec.segment(0, m_num_decision_variables).setConstant(δ);
    vec.segment(m_num_decision_variables, m_num_equality_constraints)
        .setConstant(-γ);
 
    return vec.asDiagonal().toDenseMatrix();
  }
};
 
}  // namespace slp