primal_dual_hybrid_gradient.cc

Go to the documentation of this file.

// Copyright 2010-2025 Google LLC
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
 
// We solve a QP, which we call the "original QP", by applying preprocessing
// including presolve and rescaling, which produces a new QP that we call the
// "working QP". We then solve the working QP using the Primal Dual Hybrid
// Gradient algorithm (PDHG). The optimality criteria are evaluated using the
// original QP. There are three main modules in this file:
// * The free function `PrimalDualHybridGradient()`, which is the user API, and
//   is responsible for input validation that doesn't use
//   ShardedQuadraticProgram, creating a `PreprocessSolver`, and calling
//   `PreprocessSolver::PreprocessAndSolve()`.
// * The class `PreprocessSolver`, which is responsible for everything that
//   touches the original QP, including input validation that uses
//   ShardedQuadraticProgram, the preprocessing, converting solutions to the
//   working QP back to solutions to the original QP, and termination checks. It
//   also creates a `Solver` object and calls `Solver::Solve()`.
// * The class `Solver`, which is responsible for everything that only touches
//   the working QP. It keeps a pointer to `PreprocessSolver` and calls methods
//   on it when it needs access to the original QP, e.g. termination checks.
//   When feasibility polishing is enabled the main solve's `Solver` object
//   creates additional `Solver` objects periodically to do the feasibility
//   polishing (in `Solver::TryPrimalPolishing()` and
//   `Solver::TryDualPolishing()`).
// The main reason for having two separate classes `PreprocessSolver` and
// `Solver` is the fact that feasibility polishing mode uses a single
// `PreprocessSolver` object with multiple `Solver` objects.
 
#include "ortools/pdlp/primal_dual_hybrid_gradient.h"
 
#include <algorithm>
#include <atomic>
#include <cmath>
#include <cstdint>
#include <functional>
#include <limits>
#include <optional>
#include <random>
#include <string>
#include <utility>
#include <vector>
 
#include "Eigen/Core"
#include "Eigen/SparseCore"
#include "absl/algorithm/container.h"
#include "absl/base/nullability.h"
#include "absl/log/check.h"
#include "absl/log/log.h"
#include "absl/status/status.h"
#include "absl/status/statusor.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_format.h"
#include "absl/strings/string_view.h"
#include "absl/time/clock.h"
#include "absl/time/time.h"
#include "google/protobuf/repeated_ptr_field.h"
#include "ortools/base/mathutil.h"
#include "ortools/base/timer.h"
#include "ortools/glop/parameters.pb.h"
#include "ortools/glop/preprocessor.h"
#include "ortools/linear_solver/linear_solver.pb.h"
#include "ortools/lp_data/lp_data.h"
#include "ortools/lp_data/lp_types.h"
#include "ortools/lp_data/proto_utils.h"
#include "ortools/pdlp/iteration_stats.h"
#include "ortools/pdlp/quadratic_program.h"
#include "ortools/pdlp/sharded_optimization_utils.h"
#include "ortools/pdlp/sharded_quadratic_program.h"
#include "ortools/pdlp/sharder.h"
#include "ortools/pdlp/solve_log.pb.h"
#include "ortools/pdlp/solvers.pb.h"
#include "ortools/pdlp/solvers_proto_validation.h"
#include "ortools/pdlp/termination.h"
#include "ortools/pdlp/trust_region.h"
#include "ortools/util/logging.h"
 
namespace operations_research::pdlp {
 
namespace {
 
using ::Eigen::VectorXd;
using ::operations_research::SolverLogger;
 
using IterationStatsCallback =
    std::function<void(const IterationCallbackInfo&)>;
 
// Computes a `num_threads` that is capped by the problem size and `num_shards`,
// if specified, to avoid creating unusable threads.
int NumThreads(const int num_threads, const int num_shards,
               const QuadraticProgram& qp, SolverLogger& logger) {
  int capped_num_threads = num_threads;
  if (num_shards > 0) {
    capped_num_threads = std::min(capped_num_threads, num_shards);
  }
  const int64_t problem_limit = std::max(qp.variable_lower_bounds.size(),
                                         qp.constraint_lower_bounds.size());
  capped_num_threads =
      static_cast<int>(std::min(int64_t{capped_num_threads}, problem_limit));
  capped_num_threads = std::max(capped_num_threads, 1);
  if (capped_num_threads != num_threads) {
    SOLVER_LOG(&logger, "WARNING: Reducing num_threads from ", num_threads,
               " to ", capped_num_threads,
               " because additional threads would be useless.");
  }
  return capped_num_threads;
}
 
// If `num_shards` is positive, returns it. Otherwise returns a reasonable
// number of shards to use with `ShardedQuadraticProgram` for the given
// `num_threads`.
int NumShards(const int num_threads, const int num_shards) {
  if (num_shards > 0) return num_shards;
  return num_threads == 1 ? 1 : 4 * num_threads;
}
 
std::string ConvergenceInformationString(
    const ConvergenceInformation& convergence_information,
    const RelativeConvergenceInformation& relative_information,
    const OptimalityNorm residual_norm) {
  constexpr absl::string_view kFormatStr =
      "%#12.6g %#12.6g %#12.6g | %#12.6g %#12.6g %#12.6g | %#12.6g %#12.6g | "
      "%#12.6g %#12.6g";
  switch (residual_norm) {
    case OPTIMALITY_NORM_L_INF:
      return absl::StrFormat(
          kFormatStr, relative_information.relative_l_inf_primal_residual,
          relative_information.relative_l_inf_dual_residual,
          relative_information.relative_optimality_gap,
          convergence_information.l_inf_primal_residual(),
          convergence_information.l_inf_dual_residual(),
          convergence_information.primal_objective() -
              convergence_information.dual_objective(),
          convergence_information.primal_objective(),
          convergence_information.dual_objective(),
          convergence_information.l2_primal_variable(),
          convergence_information.l2_dual_variable());
    case OPTIMALITY_NORM_L2:
      return absl::StrFormat(kFormatStr,
                             relative_information.relative_l2_primal_residual,
                             relative_information.relative_l2_dual_residual,
                             relative_information.relative_optimality_gap,
                             convergence_information.l2_primal_residual(),
                             convergence_information.l2_dual_residual(),
                             convergence_information.primal_objective() -
                                 convergence_information.dual_objective(),
                             convergence_information.primal_objective(),
                             convergence_information.dual_objective(),
                             convergence_information.l2_primal_variable(),
                             convergence_information.l2_dual_variable());
    case OPTIMALITY_NORM_L_INF_COMPONENTWISE:
      return absl::StrFormat(
          kFormatStr,
          convergence_information.l_inf_componentwise_primal_residual(),
          convergence_information.l_inf_componentwise_dual_residual(),
          relative_information.relative_optimality_gap,
          convergence_information.l_inf_primal_residual(),
          convergence_information.l_inf_dual_residual(),
          convergence_information.primal_objective() -
              convergence_information.dual_objective(),
          convergence_information.primal_objective(),
          convergence_information.dual_objective(),
          convergence_information.l2_primal_variable(),
          convergence_information.l2_dual_variable());
    case OPTIMALITY_NORM_UNSPECIFIED:
      LOG(FATAL) << "Residual norm not specified.";
  }
  LOG(FATAL) << "Invalid residual norm " << residual_norm << ".";
}
 
std::string ConvergenceInformationShortString(
    const ConvergenceInformation& convergence_information,
    const RelativeConvergenceInformation& relative_information,
    const OptimalityNorm residual_norm) {
  constexpr absl::string_view kFormatStr =
      "%#10.4g %#10.4g %#10.4g | %#10.4g %#10.4g";
  switch (residual_norm) {
    case OPTIMALITY_NORM_L_INF:
      return absl::StrFormat(
          kFormatStr, relative_information.relative_l_inf_primal_residual,
          relative_information.relative_l_inf_dual_residual,
          relative_information.relative_optimality_gap,
          convergence_information.primal_objective(),
          convergence_information.dual_objective());
    case OPTIMALITY_NORM_L2:
      return absl::StrFormat(kFormatStr,
                             relative_information.relative_l2_primal_residual,
                             relative_information.relative_l2_dual_residual,
                             relative_information.relative_optimality_gap,
                             convergence_information.primal_objective(),
                             convergence_information.dual_objective());
    case OPTIMALITY_NORM_L_INF_COMPONENTWISE:
      return absl::StrFormat(
          kFormatStr,
          convergence_information.l_inf_componentwise_primal_residual(),
          convergence_information.l_inf_componentwise_dual_residual(),
          relative_information.relative_optimality_gap,
          convergence_information.primal_objective(),
          convergence_information.dual_objective());
    case OPTIMALITY_NORM_UNSPECIFIED:
      LOG(FATAL) << "Residual norm not specified.";
  }
  LOG(FATAL) << "Invalid residual norm " << residual_norm << ".";
}
 
// Logs a description of `iter_stats`, based on the
// `iter_stats.convergence_information` entry with
// `.candidate_type()==preferred_candidate` if one exists, otherwise based on
// the first value, if any. `termination_criteria.optimality_norm` determines
// which residual norms from `iter_stats.convergence_information` are used.
void LogIterationStats(int verbosity_level, bool use_feasibility_polishing,
                       IterationType iteration_type,
                       const IterationStats& iter_stats,
                       const TerminationCriteria& termination_criteria,
                       const QuadraticProgramBoundNorms& bound_norms,
                       PointType preferred_candidate, SolverLogger& logger) {
  std::string iteration_string =
      verbosity_level >= 3
          ? absl::StrFormat("%6d %8.1f %6.1f", iter_stats.iteration_number(),
                            iter_stats.cumulative_kkt_matrix_passes(),
                            iter_stats.cumulative_time_sec())
          : absl::StrFormat("%6d %6.1f", iter_stats.iteration_number(),
                            iter_stats.cumulative_time_sec());
  auto convergence_information =
      GetConvergenceInformation(iter_stats, preferred_candidate);
  if (!convergence_information.has_value() &&
      iter_stats.convergence_information_size() > 0) {
    convergence_information = iter_stats.convergence_information(0);
  }
  const char* phase_string = [&]() {
    if (use_feasibility_polishing) {
      switch (iteration_type) {
        case IterationType::kNormal:
          return "n ";
        case IterationType::kPrimalFeasibility:
          return "p ";
        case IterationType::kDualFeasibility:
          return "d ";
        case IterationType::kNormalTermination:
        case IterationType::kFeasibilityPolishingTermination:
        case IterationType::kPresolveTermination:
          return "t ";
      }
    } else {
      return "";
    }
  }();
  if (convergence_information.has_value()) {
    const char* iterate_string = [&]() {
      if (verbosity_level >= 4) {
        switch (convergence_information->candidate_type()) {
          case POINT_TYPE_CURRENT_ITERATE:
            return "C ";
          case POINT_TYPE_AVERAGE_ITERATE:
            return "A ";
          case POINT_TYPE_ITERATE_DIFFERENCE:
            return "D ";
          case POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION:
            return "F ";
          default:
            return "? ";
        }
      } else {
        return "";
      }
    }();
    const RelativeConvergenceInformation relative_information =
        ComputeRelativeResiduals(
            EffectiveOptimalityCriteria(termination_criteria),
            *convergence_information, bound_norms);
    std::string convergence_string =
        verbosity_level >= 3
            ? ConvergenceInformationString(
                  *convergence_information, relative_information,
                  termination_criteria.optimality_norm())
            : ConvergenceInformationShortString(
                  *convergence_information, relative_information,
                  termination_criteria.optimality_norm());
    SOLVER_LOG(&logger, phase_string, iterate_string, iteration_string, " | ",
               convergence_string);
  } else {
    // No convergence information, just log the basic work stats.
    SOLVER_LOG(&logger, phase_string, verbosity_level >= 4 ? "? " : "",
               iteration_string);
  }
}
 
void LogIterationStatsHeader(int verbosity_level,
                             bool use_feasibility_polishing,
                             SolverLogger& logger) {
  std::string work_labels =
      verbosity_level >= 3
          ? absl::StrFormat("%6s %8s %6s", "iter#", "kkt_pass", "time")
          : absl::StrFormat("%6s %6s", "iter#", "time");
  std::string convergence_labels =
      verbosity_level >= 3
          ? absl::StrFormat(
                "%12s %12s %12s | %12s %12s %12s | %12s %12s | %12s %12s",
                "rel_prim_res", "rel_dual_res", "rel_gap", "prim_resid",
                "dual_resid", "obj_gap", "prim_obj", "dual_obj", "prim_var_l2",
                "dual_var_l2")
          : absl::StrFormat("%10s %10s %10s | %10s %10s", "rel_p_res",
                            "rel_d_res", "rel_gap", "prim_obj", "dual_obj");
  SOLVER_LOG(&logger, use_feasibility_polishing ? "f " : "",
             verbosity_level >= 4 ? "I " : "", work_labels, " | ",
             convergence_labels);
}
 
enum class InnerStepOutcome {
  kSuccessful,
  kForceNumericalTermination,
};
 
// Makes the closing changes to `solve_log` and builds a `SolverResult`.
// NOTE: `primal_solution`, `dual_solution`, and `solve_log` are passed by
// value. To avoid unnecessary copying, move these objects (i.e. use
// `std::move()`).
SolverResult ConstructSolverResult(VectorXd primal_solution,
                                   VectorXd dual_solution,
                                   const IterationStats& stats,
                                   TerminationReason termination_reason,
                                   PointType output_type, SolveLog solve_log) {
  solve_log.set_iteration_count(stats.iteration_number());
  solve_log.set_termination_reason(termination_reason);
  solve_log.set_solution_type(output_type);
  solve_log.set_solve_time_sec(stats.cumulative_time_sec());
  *solve_log.mutable_solution_stats() = stats;
  return SolverResult{.primal_solution = std::move(primal_solution),
                      .dual_solution = std::move(dual_solution),
                      .solve_log = std::move(solve_log)};
}
 
// See comment at top of file.
class PreprocessSolver {
 public:
  // Assumes that `qp` and `params` are valid.
  // Note that the `qp` is intentionally passed by value.
  // `logger` is captured, and must outlive the class.
  // NOTE: Many `PreprocessSolver` methods accept a `params` argument. This is
  // passed as an argument instead of stored as a member variable to support
  // using different `params` in different contexts with the same
  // `PreprocessSolver` object.
  explicit PreprocessSolver(QuadraticProgram qp,
                            const PrimalDualHybridGradientParams& params,
                            SolverLogger* logger);
 
  // Not copyable or movable (because `glop::MainLpPreprocessor` isn't).
  PreprocessSolver(const PreprocessSolver&) = delete;
  PreprocessSolver& operator=(const PreprocessSolver&) = delete;
  PreprocessSolver(PreprocessSolver&&) = delete;
  PreprocessSolver& operator=(PreprocessSolver&&) = delete;
 
  // Zero is used if `initial_solution` is nullopt. If `interrupt_solve` is not
  // nullptr, then the solver will periodically check if
  // `interrupt_solve->load()` is true, in which case the solve will terminate
  // with `TERMINATION_REASON_INTERRUPTED_BY_USER`. Ownership is not
  // transferred. If `iteration_stats_callback` is not nullptr, then at each
  // termination step (when iteration stats are logged),
  // `iteration_stats_callback` will also be called with those iteration stats.
  SolverResult PreprocessAndSolve(
      const PrimalDualHybridGradientParams& params,
      std::optional<PrimalAndDualSolution> initial_solution,
      const std::atomic<bool>* interrupt_solve,
      IterationStatsCallback iteration_stats_callback);
 
  // Returns a `TerminationReasonAndPointType` when the termination criteria are
  // satisfied, otherwise returns nothing. The pointers to working_* can be
  // nullptr if an iterate of that type is not available. For the iterate types
  // that are available, uses the primal and dual vectors to compute solution
  // statistics and adds them to the stats proto.
  // `full_stats` is used for checking work-based termination criteria,
  // but `stats` is updated with convergence information.
  // NOTE: The primal and dual input pairs should be scaled solutions.
  // TODO(user): Move this long argument list into a struct.
  std::optional<TerminationReasonAndPointType>
  UpdateIterationStatsAndCheckTermination(
      const PrimalDualHybridGradientParams& params,
      bool force_numerical_termination, const VectorXd& working_primal_current,
      const VectorXd& working_dual_current,
      const VectorXd* working_primal_average,
      const VectorXd* working_dual_average,
      const VectorXd* working_primal_delta, const VectorXd* working_dual_delta,
      const VectorXd& last_primal_start_point,
      const VectorXd& last_dual_start_point,
      const std::atomic<bool>* interrupt_solve, IterationType iteration_type,
      const IterationStats& full_stats, IterationStats& stats);
 
  // Computes solution statistics for the primal and dual input pair, which
  // should be a scaled solution. If `convergence_information != nullptr`,
  // populates it with convergence information, and similarly if
  // `infeasibility_information != nullptr`, populates it with infeasibility
  // information.
  void ComputeConvergenceAndInfeasibilityFromWorkingSolution(
      const PrimalDualHybridGradientParams& params,
      const VectorXd& working_primal, const VectorXd& working_dual,
      PointType candidate_type, ConvergenceInformation* convergence_information,
      InfeasibilityInformation* infeasibility_information) const;
 
  // Returns a `SolverResult` for the original problem, given a `SolverResult`
  // from the scaled or preprocessed problem. Also computes the reduced costs.
  // NOTE: `result` is passed by value. To avoid unnecessary copying, move this
  // object (i.e. use `std::move()`).
  SolverResult ConstructOriginalSolverResult(
      const PrimalDualHybridGradientParams& params, SolverResult result,
      SolverLogger& logger) const;
 
  const ShardedQuadraticProgram& ShardedWorkingQp() const {
    return sharded_qp_;
  }
 
  // Swaps `variable_lower_bounds` and `variable_upper_bounds` with the ones on
  // the sharded quadratic program.
  void SwapVariableBounds(VectorXd& variable_lower_bounds,
                          VectorXd& variable_upper_bounds) {
    sharded_qp_.SwapVariableBounds(variable_lower_bounds,
                                   variable_upper_bounds);
  }
 
  // Swaps `constraint_lower_bounds` and `constraint_upper_bounds` with the ones
  // on the sharded quadratic program.
  void SwapConstraintBounds(VectorXd& constraint_lower_bounds,
                            VectorXd& constraint_upper_bounds) {
    sharded_qp_.SwapConstraintBounds(constraint_lower_bounds,
                                     constraint_upper_bounds);
  }
 
  // Swaps `objective` with the objective vector on the quadratic program.
  // Swapping `objective_matrix` is not yet supported because it hasn't been
  // needed.
  void SwapObjectiveVector(VectorXd& objective) {
    sharded_qp_.SwapObjectiveVector(objective);
  }
 
  const QuadraticProgramBoundNorms& OriginalBoundNorms() const {
    return original_bound_norms_;
  }
 
  SolverLogger& Logger() { return logger_; }
 
 private:
  struct PresolveInfo {
    explicit PresolveInfo(ShardedQuadraticProgram original_qp,
                          const PrimalDualHybridGradientParams& params)
        : preprocessor_parameters(PreprocessorParameters(params)),
          preprocessor(&preprocessor_parameters),
          sharded_original_qp(std::move(original_qp)),
          trivial_col_scaling_vec(
              OnesVector(sharded_original_qp.PrimalSharder())),
          trivial_row_scaling_vec(
              OnesVector(sharded_original_qp.DualSharder())) {}
 
    glop::GlopParameters preprocessor_parameters;
    glop::MainLpPreprocessor preprocessor;
    ShardedQuadraticProgram sharded_original_qp;
    bool presolved_problem_was_maximization = false;
    const VectorXd trivial_col_scaling_vec, trivial_row_scaling_vec;
  };
 
  // TODO(user): experiment with different preprocessor types.
  static glop::GlopParameters PreprocessorParameters(
      const PrimalDualHybridGradientParams& params);
 
  // If presolve is enabled, moves `sharded_qp_` to
  // `presolve_info_.sharded_original_qp` and computes the presolved linear
  // program and installs it in `sharded_qp_`. Clears `initial_solution` if
  // presolve is enabled. If presolve solves the problem completely returns the
  // appropriate `TerminationReason`. Otherwise returns nullopt. If presolve
  // is disabled or an error occurs modifies nothing and returns nullopt.
  std::optional<TerminationReason> ApplyPresolveIfEnabled(
      const PrimalDualHybridGradientParams& params,
      std::optional<PrimalAndDualSolution>* initial_solution);
 
  void ComputeAndApplyRescaling(const PrimalDualHybridGradientParams& params,
                                VectorXd& starting_primal_solution,
                                VectorXd& starting_dual_solution);
 
  void LogQuadraticProgramStats(const QuadraticProgramStats& stats) const;
 
  double InitialPrimalWeight(const PrimalDualHybridGradientParams& params,
                             double l2_norm_primal_linear_objective,
                             double l2_norm_constraint_bounds) const;
 
  PrimalAndDualSolution RecoverOriginalSolution(
      PrimalAndDualSolution working_solution) const;
 
  // Adds one entry of `PointMetadata` to `stats` using the input solutions.
  void AddPointMetadata(const PrimalDualHybridGradientParams& params,
                        const VectorXd& primal_solution,
                        const VectorXd& dual_solution, PointType point_type,
                        const VectorXd& last_primal_start_point,
                        const VectorXd& last_dual_start_point,
                        IterationStats& stats) const;
 
  const QuadraticProgram& Qp() const { return sharded_qp_.Qp(); }
 
  const int num_threads_;
  const int num_shards_;
 
  // The bound norms of the original problem.
  QuadraticProgramBoundNorms original_bound_norms_;
 
  // This is the QP that PDHG is run on. It is modified by presolve and
  // rescaling, if those are enabled, and then serves as the
  // `ShardedWorkingQp()` when calling `Solver::Solve()`. The original problem
  // is available in `presolve_info_->sharded_original_qp` if
  // `presolve_info_.has_value()`, and otherwise can be obtained by undoing the
  // scaling of `sharded_qp_` by `col_scaling_vec_` and `row_scaling_vec_`.
  ShardedQuadraticProgram sharded_qp_;
 
  // Set iff presolve is enabled.
  std::optional<PresolveInfo> presolve_info_;
 
  // The scaling vectors that map the original (or presolved) quadratic program
  // to the working version. See
  // `ShardedQuadraticProgram::RescaleQuadraticProgram()` for details.
  VectorXd col_scaling_vec_;
  VectorXd row_scaling_vec_;
 
  // A counter used to trigger writing iteration headers.
  int log_counter_ = 0;
  absl::Time time_of_last_log_ = absl::InfinitePast();
  SolverLogger& logger_;
  IterationStatsCallback iteration_stats_callback_;
};
 
// See comment at top of file.
class Solver {
 public:
  // `preprocess_solver` should not be nullptr, and the `PreprocessSolver`
  // object must outlive this `Solver` object. Ownership is not transferred.
  explicit Solver(const PrimalDualHybridGradientParams& params,
                  VectorXd starting_primal_solution,
                  VectorXd starting_dual_solution, double initial_step_size,
                  double initial_primal_weight,
                  PreprocessSolver* preprocess_solver);
 
  // Not copyable or movable (because there are const members).
  Solver(const Solver&) = delete;
  Solver& operator=(const Solver&) = delete;
  Solver(Solver&&) = delete;
  Solver& operator=(Solver&&) = delete;
 
  const QuadraticProgram& WorkingQp() const { return ShardedWorkingQp().Qp(); }
 
  const ShardedQuadraticProgram& ShardedWorkingQp() const {
    return preprocess_solver_->ShardedWorkingQp();
  }
 
  // Runs PDHG iterations on the instance that has been initialized in `Solver`.
  // If `interrupt_solve` is not nullptr, then the solver will periodically
  // check if `interrupt_solve->load()` is true, in which case the solve will
  // terminate with `TERMINATION_REASON_INTERRUPTED_BY_USER`. Ownership is not
  // transferred.
  // `solve_log` should contain initial problem statistics.
  // On return, `SolveResult.reduced_costs` will be empty, and the solution will
  // be to the preprocessed/scaled problem.
  SolverResult Solve(IterationType iteration_type,
                     const std::atomic<bool>* interrupt_solve,
                     SolveLog solve_log);
 
 private:
  struct NextSolutionAndDelta {
    VectorXd value;
    // `delta` is `value` - current_solution.
    VectorXd delta;
  };
 
  struct DistanceBasedRestartInfo {
    double distance_moved_last_restart_period;
    int length_of_last_restart_period;
  };
 
  // Movement terms (weighted squared norms of primal and dual deltas) larger
  // than this cause termination because iterates are diverging, and likely to
  // cause infinite and NaN values.
  constexpr static double kDivergentMovement = 1.0e100;
 
  // The total number of iterations in feasibility polishing is at most
  // `4 * iterations_completed_ / kFeasibilityIterationFraction`.
  // One factor of two is because there are both primal and dual feasibility
  // polishing phases, and the other factor of two is because
  // `next_feasibility_polishing_iteration` increases by a factor of 2 each
  // feasibility polishing phase, so the sum of iteration limits is at most
  // twice the last value.
  constexpr static int kFeasibilityIterationFraction = 8;
 
  // Attempts to solve primal and dual feasibility subproblems starting at the
  // average iterate, for at most `iteration_limit` iterations each. If
  // successful, returns a `SolverResult`, otherwise nullopt. Appends
  // information about the polishing phases to
  // `solve_log.feasibility_polishing_details`.
  std::optional<SolverResult> TryFeasibilityPolishing(
      int iteration_limit, const std::atomic<bool>* interrupt_solve,
      SolveLog& solve_log);
 
  // Tries to find primal feasibility, adds the solve log details to
  // `solve_log.feasibility_polishing_details`, and returns the result.
  SolverResult TryPrimalPolishing(VectorXd starting_primal_solution,
                                  int iteration_limit,
                                  const std::atomic<bool>* interrupt_solve,
                                  SolveLog& solve_log);
 
  // Tries to find dual feasibility, adds the solve log details to
  // `solve_log.feasibility_polishing_details`, and returns the result.
  SolverResult TryDualPolishing(VectorXd starting_dual_solution,
                                int iteration_limit,
                                const std::atomic<bool>* interrupt_solve,
                                SolveLog& solve_log);
 
  NextSolutionAndDelta ComputeNextPrimalSolution(double primal_step_size) const;
 
  NextSolutionAndDelta ComputeNextDualSolution(
      double dual_step_size, double extrapolation_factor,
      const NextSolutionAndDelta& next_primal_solution,
      const VectorXd* absl_nullable next_primal_product = nullptr) const;
 
  std::pair<double, double> ComputeMovementTerms(
      const VectorXd& delta_primal, const VectorXd& delta_dual) const;
 
  double ComputeMovement(const VectorXd& delta_primal,
                         const VectorXd& delta_dual) const;
 
  double ComputeNonlinearity(const VectorXd& delta_primal,
                             const VectorXd& next_dual_product) const;
 
  // Sets current_primal_product_ and current_dual_product_ based on
  // current_primal_solution_ and current_dual_solution_ respectively.
  void SetCurrentPrimalAndDualProducts();
 
  // Creates all the simple-to-compute statistics in stats.
  IterationStats CreateSimpleIterationStats(RestartChoice restart_used) const;
 
  // Returns the total work so far from the main iterations and feasibility
  // polishing phases.
  IterationStats TotalWorkSoFar(const SolveLog& solve_log) const;
 
  RestartChoice ChooseRestartToApply(bool is_major_iteration);
 
  VectorXd PrimalAverage() const;
 
  VectorXd DualAverage() const;
 
  double ComputeNewPrimalWeight() const;
 
  // Picks the primal and dual solutions according to `output_type`, and makes
  // the closing changes to `solve_log`. This function should only be called
  // once when the solver is finishing its execution.
  // NOTE: `primal_solution` and `dual_solution` are used as the output except
  // when `output_type` is `POINT_TYPE_CURRENT_ITERATE` or
  // `POINT_TYPE_ITERATE_DIFFERENCE`, in which case the values are computed from
  // `Solver` data.
  // NOTE: `primal_solution`, `dual_solution`, and `solve_log` are passed by
  // value. To avoid unnecessary copying, move these objects (i.e. use
  // `std::move()`).
  SolverResult PickSolutionAndConstructSolverResult(
      VectorXd primal_solution, VectorXd dual_solution,
      const IterationStats& stats, TerminationReason termination_reason,
      PointType output_type, SolveLog solve_log) const;
 
  double DistanceTraveledFromLastStart(const VectorXd& primal_solution,
                                       const VectorXd& dual_solution) const;
 
  LocalizedLagrangianBounds ComputeLocalizedBoundsAtCurrent() const;
 
  LocalizedLagrangianBounds ComputeLocalizedBoundsAtAverage() const;
 
  // Applies the given `RestartChoice`. If a restart is chosen, updates the
  // state of the algorithm accordingly and computes a new primal weight.
  void ApplyRestartChoice(RestartChoice restart_to_apply);
 
  std::optional<SolverResult> MajorIterationAndTerminationCheck(
      IterationType iteration_type, bool force_numerical_termination,
      const std::atomic<bool>* interrupt_solve,
      const IterationStats& work_from_feasibility_polishing,
      SolveLog& solve_log);
 
  bool ShouldDoAdaptiveRestartHeuristic(double candidate_normalized_gap) const;
 
  RestartChoice DetermineDistanceBasedRestartChoice() const;
 
  void ResetAverageToCurrent();
 
  void LogNumericalTermination(const Eigen::VectorXd& primal_delta,
                               const Eigen::VectorXd& dual_delta) const;
 
  void LogInnerIterationLimitHit() const;
 
  // Takes a step based on the Malitsky and Pock linesearch algorithm.
  // (https://arxiv.org/pdf/1608.08883.pdf)
  // The current implementation is provably convergent (at an optimal rate)
  // for LP programs (provided we do not change the primal weight at every major
  // iteration). Further, we have observed that this rule is very sensitive to
  // the parameter choice whenever we apply the primal weight recomputation
  // heuristic.
  InnerStepOutcome TakeMalitskyPockStep();
 
  // Takes a step based on the adaptive heuristic presented in Section 3.1 of
  // https://arxiv.org/pdf/2106.04756.pdf (further generalized to QP).
  InnerStepOutcome TakeAdaptiveStep();
 
  // Takes a constant-size step.
  InnerStepOutcome TakeConstantSizeStep();
 
  const PrimalDualHybridGradientParams params_;
 
  VectorXd current_primal_solution_;
  VectorXd current_dual_solution_;
  VectorXd current_primal_delta_;
  VectorXd current_dual_delta_;
 
  ShardedWeightedAverage primal_average_;
  ShardedWeightedAverage dual_average_;
 
  double step_size_;
  double primal_weight_;
 
  PreprocessSolver* preprocess_solver_;
 
  // For Malitsky-Pock linesearch only: `step_size_` / previous_step_size
  double ratio_last_two_step_sizes_;
  // For adaptive restarts only.
  double normalized_gap_at_last_trial_ =
      std::numeric_limits<double>::infinity();
  // For adaptive restarts only.
  double normalized_gap_at_last_restart_ =
      std::numeric_limits<double>::infinity();
 
  // `preprocessing_time_sec_` is added to the current value of `timer_` when
  // computing `cumulative_time_sec` in iteration stats.
  double preprocessing_time_sec_;
  WallTimer timer_;
  int iterations_completed_;
  int num_rejected_steps_;
  // A cache of `constraint_matrix * current_primal_solution_`.
  // Malitsky-Pock linesearch only.
  std::optional<VectorXd> current_primal_product_;
  // A cache of `constraint_matrix.transpose() * current_dual_solution_`.
  VectorXd current_dual_product_;
  // The primal point at which the algorithm was last restarted from, or
  // the initial primal starting point if no restart has occurred.
  VectorXd last_primal_start_point_;
  // The dual point at which the algorithm was last restarted from, or
  // the initial dual starting point if no restart has occurred.
  VectorXd last_dual_start_point_;
  // Information for deciding whether to trigger a distance-based restart.
  // The distances are initialized to +inf to force a restart during the first
  // major iteration check.
  DistanceBasedRestartInfo distance_based_restart_info_ = {
      .distance_moved_last_restart_period =
          std::numeric_limits<double>::infinity(),
      .length_of_last_restart_period = 1,
  };
};
 
PreprocessSolver::PreprocessSolver(QuadraticProgram qp,
                                   const PrimalDualHybridGradientParams& params,
                                   SolverLogger* logger)
    : num_threads_(
          NumThreads(params.num_threads(), params.num_shards(), qp, *logger)),
      num_shards_(NumShards(num_threads_, params.num_shards())),
      sharded_qp_(std::move(qp), num_threads_, num_shards_,
                  params.scheduler_type(), nullptr),
      logger_(*logger) {}
 
SolverResult ErrorSolverResult(const TerminationReason reason,
                               const std::string& message,
                               SolverLogger& logger) {
  SolveLog error_log;
  error_log.set_termination_reason(reason);
  error_log.set_termination_string(message);
  SOLVER_LOG(&logger,
             "The solver did not run because of invalid input: ", message);
  return SolverResult{.solve_log = error_log};
}
 
// If `check_excessively_small_values`, in addition to checking for extreme
// values that PDLP can't handle, also check for extremely small constraint
// bounds, variable bound gaps, and objective vector values that PDLP can handle
// but that cause problems for other code such as glop's presolve.
std::optional<SolverResult> CheckProblemStats(
    const QuadraticProgramStats& problem_stats, const double objective_offset,
    bool check_excessively_small_values, SolverLogger& logger) {
  const double kExcessiveInputValue = 1e50;
  const double kExcessivelySmallInputValue = 1e-50;
  const double kMaxDynamicRange = 1e20;
  if (std::isnan(problem_stats.constraint_matrix_l2_norm())) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "Constraint matrix has a NAN.", logger);
  }
  if (problem_stats.constraint_matrix_abs_max() > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Constraint matrix has a non-zero with absolute value ",
                     problem_stats.constraint_matrix_abs_max(),
                     " which exceeds limit of ", kExcessiveInputValue, "."),
        logger);
  }
  if (problem_stats.constraint_matrix_abs_max() >
      kMaxDynamicRange * problem_stats.constraint_matrix_abs_min()) {
    SOLVER_LOG(
        &logger, "WARNING: Constraint matrix has largest absolute value ",
        problem_stats.constraint_matrix_abs_max(),
        " and smallest non-zero absolute value ",
        problem_stats.constraint_matrix_abs_min(), " performance may suffer.");
  }
  if (problem_stats.constraint_matrix_col_min_l_inf_norm() > 0 &&
      problem_stats.constraint_matrix_col_min_l_inf_norm() <
          kExcessivelySmallInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Constraint matrix has a column with Linf norm ",
                     problem_stats.constraint_matrix_col_min_l_inf_norm(),
                     " which is less than limit of ",
                     kExcessivelySmallInputValue, "."),
        logger);
  }
  if (problem_stats.constraint_matrix_row_min_l_inf_norm() > 0 &&
      problem_stats.constraint_matrix_row_min_l_inf_norm() <
          kExcessivelySmallInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Constraint matrix has a row with Linf norm ",
                     problem_stats.constraint_matrix_row_min_l_inf_norm(),
                     " which is less than limit of ",
                     kExcessivelySmallInputValue, "."),
        logger);
  }
  if (std::isnan(problem_stats.combined_bounds_l2_norm())) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "Constraint bounds vector has a NAN.", logger);
  }
  if (problem_stats.combined_bounds_max() > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Combined constraint bounds vector has a non-zero with "
                     "absolute value ",
                     problem_stats.combined_bounds_max(),
                     " which exceeds limit of ", kExcessiveInputValue, "."),
        logger);
  }
  if (check_excessively_small_values &&
      problem_stats.combined_bounds_min() > 0 &&
      problem_stats.combined_bounds_min() < kExcessivelySmallInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Combined constraint bounds vector has a non-zero with "
                     "absolute value ",
                     problem_stats.combined_bounds_min(),
                     " which is less than the limit of ",
                     kExcessivelySmallInputValue, "."),
        logger);
  }
  if (problem_stats.combined_bounds_max() >
      kMaxDynamicRange * problem_stats.combined_bounds_min()) {
    SOLVER_LOG(&logger,
               "WARNING: Combined constraint bounds vector has largest "
               "absolute value ",
               problem_stats.combined_bounds_max(),
               " and smallest non-zero absolute value ",
               problem_stats.combined_bounds_min(),
               "; performance may suffer.");
  }
  if (std::isnan(problem_stats.combined_variable_bounds_l2_norm())) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "Variable bounds vector has a NAN.", logger);
  }
  if (problem_stats.combined_variable_bounds_max() > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Combined variable bounds vector has a non-zero with "
                     "absolute value ",
                     problem_stats.combined_variable_bounds_max(),
                     " which exceeds limit of ", kExcessiveInputValue, "."),
        logger);
  }
  if (check_excessively_small_values &&
      problem_stats.combined_variable_bounds_min() > 0 &&
      problem_stats.combined_variable_bounds_min() <
          kExcessivelySmallInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Combined variable bounds vector has a non-zero with "
                     "absolute value ",
                     problem_stats.combined_variable_bounds_min(),
                     " which is less than the limit of ",
                     kExcessivelySmallInputValue, "."),
        logger);
  }
  if (problem_stats.combined_variable_bounds_max() >
      kMaxDynamicRange * problem_stats.combined_variable_bounds_min()) {
    SOLVER_LOG(
        &logger,
        "WARNING: Combined variable bounds vector has largest absolute value ",
        problem_stats.combined_variable_bounds_max(),
        " and smallest non-zero absolute value ",
        problem_stats.combined_variable_bounds_min(),
        "; performance may suffer.");
  }
  if (problem_stats.variable_bound_gaps_max() >
      kMaxDynamicRange * problem_stats.variable_bound_gaps_min()) {
    SOLVER_LOG(&logger,
               "WARNING: Variable bound gap vector has largest absolute value ",
               problem_stats.variable_bound_gaps_max(),
               " and smallest non-zero absolute value ",
               problem_stats.variable_bound_gaps_min(),
               "; performance may suffer.");
  }
  if (std::isnan(objective_offset)) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "Objective offset is NAN.", logger);
  }
  if (std::abs(objective_offset) > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Objective offset ", objective_offset,
                     " has absolute value which exceeds limit of ",
                     kExcessiveInputValue, "."),
        logger);
  }
  if (std::isnan(problem_stats.objective_vector_l2_norm())) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "Objective vector has a NAN.", logger);
  }
  if (problem_stats.objective_vector_abs_max() > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Objective vector has a non-zero with absolute value ",
                     problem_stats.objective_vector_abs_max(),
                     " which exceeds limit of ", kExcessiveInputValue, "."),
        logger);
  }
  if (check_excessively_small_values &&
      problem_stats.objective_vector_abs_min() > 0 &&
      problem_stats.objective_vector_abs_min() < kExcessivelySmallInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Objective vector has a non-zero with absolute value ",
                     problem_stats.objective_vector_abs_min(),
                     " which is less than the limit of ",
                     kExcessivelySmallInputValue, "."),
        logger);
  }
  if (problem_stats.objective_vector_abs_max() >
      kMaxDynamicRange * problem_stats.objective_vector_abs_min()) {
    SOLVER_LOG(&logger, "WARNING: Objective vector has largest absolute value ",
               problem_stats.objective_vector_abs_max(),
               " and smallest non-zero absolute value ",
               problem_stats.objective_vector_abs_min(),
               "; performance may suffer.");
  }
  if (std::isnan(problem_stats.objective_matrix_l2_norm())) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "Objective matrix has a NAN.", logger);
  }
  if (problem_stats.objective_matrix_abs_max() > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        absl::StrCat("Objective matrix has a non-zero with absolute value ",
                     problem_stats.objective_matrix_abs_max(),
                     " which exceeds limit of ", kExcessiveInputValue, "."),
        logger);
  }
  if (problem_stats.objective_matrix_abs_max() >
      kMaxDynamicRange * problem_stats.objective_matrix_abs_min()) {
    SOLVER_LOG(&logger, "WARNING: Objective matrix has largest absolute value ",
               problem_stats.objective_matrix_abs_max(),
               " and smallest non-zero absolute value ",
               problem_stats.objective_matrix_abs_min(),
               "; performance may suffer.");
  }
  return std::nullopt;
}
 
std::optional<SolverResult> CheckInitialSolution(
    const ShardedQuadraticProgram& sharded_qp,
    const PrimalAndDualSolution& initial_solution, SolverLogger& logger) {
  const double kExcessiveInputValue = 1e50;
  if (initial_solution.primal_solution.size() != sharded_qp.PrimalSize()) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_INITIAL_SOLUTION,
        absl::StrCat("Initial primal solution has size ",
                     initial_solution.primal_solution.size(),
                     " which differs from problem primal size ",
                     sharded_qp.PrimalSize()),
        logger);
  }
  if (std::isnan(
          Norm(initial_solution.primal_solution, sharded_qp.PrimalSharder()))) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_INITIAL_SOLUTION,
                             "Initial primal solution has a NAN.", logger);
  }
  if (const double norm = LInfNorm(initial_solution.primal_solution,
                                   sharded_qp.PrimalSharder());
      norm > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_INITIAL_SOLUTION,
        absl::StrCat(
            "Initial primal solution has an entry with absolute value ", norm,
            " which exceeds limit of ", kExcessiveInputValue),
        logger);
  }
  if (initial_solution.dual_solution.size() != sharded_qp.DualSize()) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_INITIAL_SOLUTION,
        absl::StrCat("Initial dual solution has size ",
                     initial_solution.dual_solution.size(),
                     " which differs from problem dual size ",
                     sharded_qp.DualSize()),
        logger);
  }
  if (std::isnan(
          Norm(initial_solution.dual_solution, sharded_qp.DualSharder()))) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_INITIAL_SOLUTION,
                             "Initial dual solution has a NAN.", logger);
  }
  if (const double norm =
          LInfNorm(initial_solution.dual_solution, sharded_qp.DualSharder());
      norm > kExcessiveInputValue) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_INITIAL_SOLUTION,
        absl::StrCat("Initial dual solution has an entry with absolute value ",
                     norm, " which exceeds limit of ", kExcessiveInputValue),
        logger);
  }
  return std::nullopt;
}
 
SolverResult PreprocessSolver::PreprocessAndSolve(
    const PrimalDualHybridGradientParams& params,
    std::optional<PrimalAndDualSolution> initial_solution,
    const std::atomic<bool>* interrupt_solve,
    IterationStatsCallback iteration_stats_callback) {
  WallTimer timer;
  timer.Start();
  SolveLog solve_log;
  if (params.verbosity_level() >= 1) {
    SOLVER_LOG(&logger_, "Solving with PDLP parameters: ", params);
  }
  if (Qp().problem_name.has_value()) {
    solve_log.set_instance_name(*Qp().problem_name);
  }
  *solve_log.mutable_params() = params;
  sharded_qp_.ReplaceLargeConstraintBoundsWithInfinity(
      params.infinite_constraint_bound_threshold());
  if (!HasValidBounds(sharded_qp_)) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PROBLEM,
        "The input problem has invalid bounds (after replacing large "
        "constraint bounds with infinity): some variable or constraint has "
        "lower_bound > upper_bound, lower_bound == inf, or upper_bound == "
        "-inf.",
        logger_);
  }
  if (Qp().objective_matrix.has_value() &&
      !sharded_qp_.PrimalSharder().ParallelTrueForAllShards(
          [&](const Sharder::Shard& shard) -> bool {
            return (shard(Qp().objective_matrix->diagonal()).array() >= 0.0)
                .all();
          })) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "The objective is not convex (i.e., the objective "
                             "matrix contains negative or NAN entries).",
                             logger_);
  }
  *solve_log.mutable_original_problem_stats() = ComputeStats(sharded_qp_);
  const QuadraticProgramStats& original_problem_stats =
      solve_log.original_problem_stats();
  if (auto maybe_result =
          CheckProblemStats(original_problem_stats, Qp().objective_offset,
                            params.presolve_options().use_glop(), logger_);
      maybe_result.has_value()) {
    return *maybe_result;
  }
  if (initial_solution.has_value()) {
    if (auto maybe_result =
            CheckInitialSolution(sharded_qp_, *initial_solution, logger_);
        maybe_result.has_value()) {
      return *maybe_result;
    }
  }
  original_bound_norms_ = BoundNormsFromProblemStats(original_problem_stats);
  const std::string preprocessing_string = absl::StrCat(
      params.presolve_options().use_glop() ? "presolving and " : "",
      "rescaling:");
  if (params.verbosity_level() >= 1) {
    SOLVER_LOG(&logger_, "Problem stats before ", preprocessing_string);
    LogQuadraticProgramStats(solve_log.original_problem_stats());
  }
  iteration_stats_callback_ = std::move(iteration_stats_callback);
  std::optional<TerminationReason> maybe_terminate =
      ApplyPresolveIfEnabled(params, &initial_solution);
  if (maybe_terminate.has_value()) {
    // Glop also feeds zero primal and dual solutions when the preprocessor
    // has a non-INIT status. When the preprocessor status is optimal the
    // vectors have length 0. When the status is something else the lengths
    // may be non-zero, but that's OK since we don't promise to produce a
    // meaningful solution in that case.
    IterationStats iteration_stats;
    iteration_stats.set_cumulative_time_sec(timer.Get());
    solve_log.set_preprocessing_time_sec(iteration_stats.cumulative_time_sec());
    VectorXd working_primal = ZeroVector(sharded_qp_.PrimalSharder());
    VectorXd working_dual = ZeroVector(sharded_qp_.DualSharder());
    ComputeConvergenceAndInfeasibilityFromWorkingSolution(
        params, working_primal, working_dual, POINT_TYPE_PRESOLVER_SOLUTION,
        iteration_stats.add_convergence_information(),
        iteration_stats.add_infeasibility_information());
    std::optional<TerminationReasonAndPointType> earned_termination =
        CheckIterateTerminationCriteria(params.termination_criteria(),
                                        iteration_stats, original_bound_norms_,
                                        /*force_numerical_termination=*/false);
    if (!earned_termination.has_value()) {
      earned_termination = CheckSimpleTerminationCriteria(
          params.termination_criteria(), iteration_stats, interrupt_solve);
    }
    TerminationReason final_termination_reason;
    if (earned_termination.has_value() &&
        (earned_termination->reason == TERMINATION_REASON_OPTIMAL ||
         earned_termination->reason == TERMINATION_REASON_PRIMAL_INFEASIBLE ||
         earned_termination->reason == TERMINATION_REASON_DUAL_INFEASIBLE)) {
      final_termination_reason = earned_termination->reason;
    } else {
      if (*maybe_terminate == TERMINATION_REASON_OPTIMAL) {
        final_termination_reason = TERMINATION_REASON_NUMERICAL_ERROR;
        SOLVER_LOG(
            &logger_,
            "WARNING: Presolve claimed to solve the LP optimally but the "
            "solution doesn't satisfy the optimality criteria.");
      } else {
        final_termination_reason = *maybe_terminate;
      }
    }
    return ConstructOriginalSolverResult(
        params,
        ConstructSolverResult(
            std::move(working_primal), std::move(working_dual),
            std::move(iteration_stats), final_termination_reason,
            POINT_TYPE_PRESOLVER_SOLUTION, std::move(solve_log)),
        logger_);
  }
 
  VectorXd starting_primal_solution;
  VectorXd starting_dual_solution;
  // The current solution is updated by `ComputeAndApplyRescaling`.
  if (initial_solution.has_value()) {
    starting_primal_solution = std::move(initial_solution->primal_solution);
    starting_dual_solution = std::move(initial_solution->dual_solution);
  } else {
    SetZero(sharded_qp_.PrimalSharder(), starting_primal_solution);
    SetZero(sharded_qp_.DualSharder(), starting_dual_solution);
  }
  // The following projections are necessary since all our checks assume that
  // the primal and dual variable bounds are satisfied.
  ProjectToPrimalVariableBounds(sharded_qp_, starting_primal_solution);
  ProjectToDualVariableBounds(sharded_qp_, starting_dual_solution);
 
  ComputeAndApplyRescaling(params, starting_primal_solution,
                           starting_dual_solution);
  *solve_log.mutable_preprocessed_problem_stats() = ComputeStats(sharded_qp_);
  if (params.verbosity_level() >= 1) {
    SOLVER_LOG(&logger_, "Problem stats after ", preprocessing_string);
    LogQuadraticProgramStats(solve_log.preprocessed_problem_stats());
  }
 
  double step_size = 0.0;
  if (params.linesearch_rule() ==
      PrimalDualHybridGradientParams::CONSTANT_STEP_SIZE_RULE) {
    std::mt19937 random(1);
    double inverse_step_size;
    const auto lipschitz_result =
        EstimateMaximumSingularValueOfConstraintMatrix(
            sharded_qp_, std::nullopt, std::nullopt,
            /*desired_relative_error=*/0.2, /*failure_probability=*/0.0005,
            random);
    // With high probability, `lipschitz_result.singular_value` is within
    // +/- `lipschitz_result.estimated_relative_error
    //     * lipschitz_result.singular_value`
    const double lipschitz_term_upper_bound =
        lipschitz_result.singular_value /
        (1.0 - lipschitz_result.estimated_relative_error);
    inverse_step_size = lipschitz_term_upper_bound;
    step_size = inverse_step_size > 0.0 ? 1.0 / inverse_step_size : 1.0;
  } else {
    // This initial step size is designed to err on the side of being too big.
    // This is because
    //  (i) too-big steps are rejected and hence don't hurt us other than
    //  wasting
    //      an iteration and
    // (ii) the step size adjustment algorithm shrinks the step size as far as
    //      needed in a single iteration but raises it slowly.
    // The tiny constant is there to keep the step size finite in the case of a
    // trivial LP with no constraints.
    step_size =
        1.0 /
        std::max(
            1.0e-20,
            solve_log.preprocessed_problem_stats().constraint_matrix_abs_max());
  }
  step_size *= params.initial_step_size_scaling();
 
  const double primal_weight = InitialPrimalWeight(
      params, solve_log.preprocessed_problem_stats().objective_vector_l2_norm(),
      solve_log.preprocessed_problem_stats().combined_bounds_l2_norm());
 
  Solver solver(params, starting_primal_solution, starting_dual_solution,
                step_size, primal_weight, this);
  solve_log.set_preprocessing_time_sec(timer.Get());
  SolverResult result = solver.Solve(IterationType::kNormal, interrupt_solve,
                                     std::move(solve_log));
  return ConstructOriginalSolverResult(params, std::move(result), logger_);
}
 
glop::GlopParameters PreprocessSolver::PreprocessorParameters(
    const PrimalDualHybridGradientParams& params) {
  glop::GlopParameters glop_params;
  // TODO(user): Test if dualization helps or hurts performance.
  glop_params.set_solve_dual_problem(glop::GlopParameters::NEVER_DO);
  // Experiments show that this preprocessing step can hurt because it relaxes
  // variable bounds.
  glop_params.set_use_implied_free_preprocessor(false);
  // We do our own scaling.
  glop_params.set_use_scaling(false);
  if (params.presolve_options().has_glop_parameters()) {
    glop_params.MergeFrom(params.presolve_options().glop_parameters());
  }
  return glop_params;
}
 
TerminationReason GlopStatusToTerminationReason(
    const glop::ProblemStatus glop_status, SolverLogger& logger) {
  switch (glop_status) {
    case glop::ProblemStatus::OPTIMAL:
      return TERMINATION_REASON_OPTIMAL;
    case glop::ProblemStatus::INVALID_PROBLEM:
      return TERMINATION_REASON_INVALID_PROBLEM;
    case glop::ProblemStatus::ABNORMAL:
    case glop::ProblemStatus::IMPRECISE:
      return TERMINATION_REASON_NUMERICAL_ERROR;
    case glop::ProblemStatus::PRIMAL_INFEASIBLE:
    case glop::ProblemStatus::DUAL_INFEASIBLE:
    case glop::ProblemStatus::INFEASIBLE_OR_UNBOUNDED:
    case glop::ProblemStatus::DUAL_UNBOUNDED:
    case glop::ProblemStatus::PRIMAL_UNBOUNDED:
      return TERMINATION_REASON_PRIMAL_OR_DUAL_INFEASIBLE;
    default:
      SOLVER_LOG(&logger, "WARNING: Unexpected preprocessor status ",
                 glop_status);
      return TERMINATION_REASON_OTHER;
  }
}
 
std::optional<TerminationReason> PreprocessSolver::ApplyPresolveIfEnabled(
    const PrimalDualHybridGradientParams& params,
    std::optional<PrimalAndDualSolution>* const initial_solution) {
  const bool presolve_enabled = params.presolve_options().use_glop();
  if (!presolve_enabled) {
    return std::nullopt;
  }
  if (!IsLinearProgram(Qp())) {
    SOLVER_LOG(&logger_,
               "WARNING: Skipping presolve, which is only supported for linear "
               "programs");
    return std::nullopt;
  }
  absl::StatusOr<MPModelProto> model = QpToMpModelProto(Qp());
  if (!model.ok()) {
    SOLVER_LOG(&logger_,
               "WARNING: Skipping presolve because of error converting to "
               "MPModelProto: ",
               model.status());
    return std::nullopt;
  }
  if (initial_solution->has_value()) {
    SOLVER_LOG(&logger_,
               "WARNING: Ignoring initial solution. Initial solutions are "
               "ignored when presolve is on.");
    initial_solution->reset();
  }
  glop::LinearProgram glop_lp;
  glop::MPModelProtoToLinearProgram(*model, &glop_lp);
  // Save RAM.
  model->Clear();
  presolve_info_.emplace(std::move(sharded_qp_), params);
  // To simplify our code we ignore the return value indicating whether
  // postprocessing is required. We simply call `RecoverSolution()`
  // unconditionally, which may do nothing.
  presolve_info_->preprocessor.Run(&glop_lp);
  presolve_info_->presolved_problem_was_maximization =
      glop_lp.IsMaximizationProblem();
  MPModelProto output;
  glop::LinearProgramToMPModelProto(glop_lp, &output);
  // This will only fail if given an invalid LP, which shouldn't happen.
  absl::StatusOr<QuadraticProgram> presolved_qp =
      QpFromMpModelProto(output, /*relax_integer_variables=*/false);
  CHECK_OK(presolved_qp.status());
  // `MPModelProto` doesn't support scaling factors, so if `glop_lp` has an
  // `objective_scaling_factor` it won't be set in output and `presolved_qp`.
  // The scaling factor of `presolved_qp` isn't actually used anywhere, but we
  // set it for completeness.
  presolved_qp->objective_scaling_factor = glop_lp.objective_scaling_factor();
  sharded_qp_ = ShardedQuadraticProgram(std::move(*presolved_qp), num_threads_,
                                        num_shards_, params.scheduler_type());
  // A status of `INIT` means the preprocessor created a (usually) smaller
  // problem that needs solving. Other statuses mean the preprocessor solved
  // the problem completely.
  if (presolve_info_->preprocessor.status() != glop::ProblemStatus::INIT) {
    col_scaling_vec_ = OnesVector(sharded_qp_.PrimalSharder());
    row_scaling_vec_ = OnesVector(sharded_qp_.DualSharder());
    return GlopStatusToTerminationReason(presolve_info_->preprocessor.status(),
                                         logger_);
  }
  return std::nullopt;
}
 
void PreprocessSolver::ComputeAndApplyRescaling(
    const PrimalDualHybridGradientParams& params,
    VectorXd& starting_primal_solution, VectorXd& starting_dual_solution) {
  ScalingVectors scaling = ApplyRescaling(
      RescalingOptions{.l_inf_ruiz_iterations = params.l_inf_ruiz_iterations(),
                       .l2_norm_rescaling = params.l2_norm_rescaling()},
      sharded_qp_);
  row_scaling_vec_ = std::move(scaling.row_scaling_vec);
  col_scaling_vec_ = std::move(scaling.col_scaling_vec);
 
  CoefficientWiseQuotientInPlace(col_scaling_vec_, sharded_qp_.PrimalSharder(),
                                 starting_primal_solution);
  CoefficientWiseQuotientInPlace(row_scaling_vec_, sharded_qp_.DualSharder(),
                                 starting_dual_solution);
}
 
void PreprocessSolver::LogQuadraticProgramStats(
    const QuadraticProgramStats& stats) const {
  SOLVER_LOG(&logger_,
             absl::StrFormat("There are %i variables, %i constraints, and %i "
                             "constraint matrix nonzeros.",
                             stats.num_variables(), stats.num_constraints(),
                             stats.constraint_matrix_num_nonzeros()));
  if (Qp().constraint_matrix.nonZeros() > 0) {
    SOLVER_LOG(&logger_,
               absl::StrFormat("Absolute values of nonzero constraint matrix "
                               "elements: largest=%f, "
                               "smallest=%f, avg=%f",
                               stats.constraint_matrix_abs_max(),
                               stats.constraint_matrix_abs_min(),
                               stats.constraint_matrix_abs_avg()));
    SOLVER_LOG(
        &logger_,
        absl::StrFormat("Constraint matrix, infinity norm: max(row & col)=%f, "
                        "min_col=%f, min_row=%f",
                        stats.constraint_matrix_abs_max(),
                        stats.constraint_matrix_col_min_l_inf_norm(),
                        stats.constraint_matrix_row_min_l_inf_norm()));
    SOLVER_LOG(
        &logger_,
        absl::StrFormat(
            "Constraint bounds statistics (max absolute value per row): "
            "largest=%f, smallest=%f, avg=%f, l2_norm=%f",
            stats.combined_bounds_max(), stats.combined_bounds_min(),
            stats.combined_bounds_avg(), stats.combined_bounds_l2_norm()));
  }
  if (!IsLinearProgram(Qp())) {
    SOLVER_LOG(&logger_,
               absl::StrFormat("There are %i nonzero diagonal coefficients in "
                               "the objective matrix.",
                               stats.objective_matrix_num_nonzeros()));
    SOLVER_LOG(
        &logger_,
        absl::StrFormat(
            "Absolute values of nonzero objective matrix elements: largest=%f, "
            "smallest=%f, avg=%f",
            stats.objective_matrix_abs_max(), stats.objective_matrix_abs_min(),
            stats.objective_matrix_abs_avg()));
  }
  SOLVER_LOG(&logger_, absl::StrFormat("Absolute values of objective vector "
                                       "elements: largest=%f, smallest=%f, "
                                       "avg=%f, l2_norm=%f",
                                       stats.objective_vector_abs_max(),
                                       stats.objective_vector_abs_min(),
                                       stats.objective_vector_abs_avg(),
                                       stats.objective_vector_l2_norm()));
  SOLVER_LOG(
      &logger_,
      absl::StrFormat(
          "Gaps between variable upper and lower bounds: #finite=%i of %i, "
          "largest=%f, smallest=%f, avg=%f",
          stats.variable_bound_gaps_num_finite(), stats.num_variables(),
          stats.variable_bound_gaps_max(), stats.variable_bound_gaps_min(),
          stats.variable_bound_gaps_avg()));
}
 
double PreprocessSolver::InitialPrimalWeight(
    const PrimalDualHybridGradientParams& params,
    const double l2_norm_primal_linear_objective,
    const double l2_norm_constraint_bounds) const {
  if (params.has_initial_primal_weight()) {
    return params.initial_primal_weight();
  }
  if (l2_norm_primal_linear_objective > 0.0 &&
      l2_norm_constraint_bounds > 0.0) {
    // The hand-wavy motivation for this choice is that the objective vector
    // has units of (objective units)/(primal units) and the constraint
    // bounds vector has units of (objective units)/(dual units),
    // therefore this ratio has units (dual units)/(primal units). By
    // dimensional analysis, these are the same units as the primal weight.
    return l2_norm_primal_linear_objective / l2_norm_constraint_bounds;
  } else {
    return 1.0;
  }
}
 
PrimalAndDualSolution PreprocessSolver::RecoverOriginalSolution(
    PrimalAndDualSolution working_solution) const {
  glop::ProblemSolution glop_solution(glop::RowIndex{0}, glop::ColIndex{0});
  if (presolve_info_.has_value()) {
    // We compute statuses relative to the working problem so we can detect when
    // variables are at their bounds without floating-point roundoff induced by
    // scaling.
    glop_solution = internal::ComputeStatuses(Qp(), working_solution);
  }
  CoefficientWiseProductInPlace(col_scaling_vec_, sharded_qp_.PrimalSharder(),
                                working_solution.primal_solution);
  CoefficientWiseProductInPlace(row_scaling_vec_, sharded_qp_.DualSharder(),
                                working_solution.dual_solution);
  if (presolve_info_.has_value()) {
    glop_solution.primal_values =
        glop::DenseRow(working_solution.primal_solution.begin(),
                       working_solution.primal_solution.end());
    glop_solution.dual_values =
        glop::DenseColumn(working_solution.dual_solution.begin(),
                          working_solution.dual_solution.end());
    // We got the working QP by calling `LinearProgramToMPModelProto()` and
    // `QpFromMpModelProto()`. We need to negate the duals if the LP resulting
    // from presolve was a max problem.
    if (presolve_info_->presolved_problem_was_maximization) {
      for (glop::RowIndex i{0}; i < glop_solution.dual_values.size(); ++i) {
        glop_solution.dual_values[i] *= -1;
      }
    }
    presolve_info_->preprocessor.RecoverSolution(&glop_solution);
    PrimalAndDualSolution solution;
    solution.primal_solution =
        Eigen::Map<Eigen::VectorXd>(glop_solution.primal_values.data(),
                                    glop_solution.primal_values.size().value());
    solution.dual_solution =
        Eigen::Map<Eigen::VectorXd>(glop_solution.dual_values.data(),
                                    glop_solution.dual_values.size().value());
    // We called `QpToMpModelProto()` and `MPModelProtoToLinearProgram()` to
    // convert our original QP into input for glop's preprocessor. The former
    // multiplies the objective vector by `objective_scaling_factor`, which
    // multiplies the duals by that factor as well. To undo this we divide by
    // `objective_scaling_factor`.
    solution.dual_solution /=
        presolve_info_->sharded_original_qp.Qp().objective_scaling_factor;
    // Glop's preprocessor sometimes violates the primal bounds constraints. To
    // be safe we project both primal and dual.
    ProjectToPrimalVariableBounds(presolve_info_->sharded_original_qp,
                                  solution.primal_solution);
    ProjectToDualVariableBounds(presolve_info_->sharded_original_qp,
                                solution.dual_solution);
    return solution;
  } else {
    return working_solution;
  }
}
 
void SetActiveSetInformation(const ShardedQuadraticProgram& sharded_qp,
                             const VectorXd& primal_solution,
                             const VectorXd& dual_solution,
                             const VectorXd& primal_start_point,
                             const VectorXd& dual_start_point,
                             PointMetadata& metadata) {
  CHECK_EQ(primal_solution.size(), sharded_qp.PrimalSize());
  CHECK_EQ(dual_solution.size(), sharded_qp.DualSize());
  CHECK_EQ(primal_start_point.size(), sharded_qp.PrimalSize());
  CHECK_EQ(dual_start_point.size(), sharded_qp.DualSize());
 
  const QuadraticProgram& qp = sharded_qp.Qp();
  metadata.set_active_primal_variable_count(
      static_cast<int64_t>(sharded_qp.PrimalSharder().ParallelSumOverShards(
          [&](const Sharder::Shard& shard) {
            const auto primal_shard = shard(primal_solution);
            const auto lower_bound_shard = shard(qp.variable_lower_bounds);
            const auto upper_bound_shard = shard(qp.variable_upper_bounds);
            return (primal_shard.array() > lower_bound_shard.array() &&
                    primal_shard.array() < upper_bound_shard.array())
                .count();
          })));
 
  // Most of the computation from the previous `ParallelSumOverShards` is
  // duplicated here. However the overhead shouldn't be too large, and using
  // `ParallelSumOverShards` is simpler than just using `ParallelForEachShard`.
  metadata.set_active_primal_variable_change(
      static_cast<int64_t>(sharded_qp.PrimalSharder().ParallelSumOverShards(
          [&](const Sharder::Shard& shard) {
            const auto primal_shard = shard(primal_solution);
            const auto primal_start_shard = shard(primal_start_point);
            const auto lower_bound_shard = shard(qp.variable_lower_bounds);
            const auto upper_bound_shard = shard(qp.variable_upper_bounds);
            return ((primal_shard.array() > lower_bound_shard.array() &&
                     primal_shard.array() < upper_bound_shard.array()) !=
                    (primal_start_shard.array() > lower_bound_shard.array() &&
                     primal_start_shard.array() < upper_bound_shard.array()))
                .count();
          })));
 
  metadata.set_active_dual_variable_count(
      static_cast<int64_t>(sharded_qp.DualSharder().ParallelSumOverShards(
          [&](const Sharder::Shard& shard) {
            const auto dual_shard = shard(dual_solution);
            const auto lower_bound_shard = shard(qp.constraint_lower_bounds);
            const auto upper_bound_shard = shard(qp.constraint_upper_bounds);
            const double kInfinity = std::numeric_limits<double>::infinity();
            return (dual_shard.array() != 0.0 ||
                    (lower_bound_shard.array() == -kInfinity &&
                     upper_bound_shard.array() == kInfinity))
                .count();
          })));
 
  metadata.set_active_dual_variable_change(
      static_cast<int64_t>(sharded_qp.DualSharder().ParallelSumOverShards(
          [&](const Sharder::Shard& shard) {
            const auto dual_shard = shard(dual_solution);
            const auto dual_start_shard = shard(dual_start_point);
            const auto lower_bound_shard = shard(qp.constraint_lower_bounds);
            const auto upper_bound_shard = shard(qp.constraint_upper_bounds);
            const double kInfinity = std::numeric_limits<double>::infinity();
            return ((dual_shard.array() != 0.0 ||
                     (lower_bound_shard.array() == -kInfinity &&
                      upper_bound_shard.array() == kInfinity)) !=
                    (dual_start_shard.array() != 0.0 ||
                     (lower_bound_shard.array() == -kInfinity &&
                      upper_bound_shard.array() == kInfinity)))
                .count();
          })));
}
 
void PreprocessSolver::AddPointMetadata(
    const PrimalDualHybridGradientParams& params,
    const VectorXd& primal_solution, const VectorXd& dual_solution,
    PointType point_type, const VectorXd& last_primal_start_point,
    const VectorXd& last_dual_start_point, IterationStats& stats) const {
  PointMetadata metadata;
  metadata.set_point_type(point_type);
  std::vector<int> random_projection_seeds(
      params.random_projection_seeds().begin(),
      params.random_projection_seeds().end());
  SetRandomProjections(sharded_qp_, primal_solution, dual_solution,
                       random_projection_seeds, metadata);
  if (point_type != POINT_TYPE_ITERATE_DIFFERENCE) {
    SetActiveSetInformation(sharded_qp_, primal_solution, dual_solution,
                            last_primal_start_point, last_dual_start_point,
                            metadata);
  }
  *stats.add_point_metadata() = metadata;
}
 
std::optional<TerminationReasonAndPointType>
PreprocessSolver::UpdateIterationStatsAndCheckTermination(
    const PrimalDualHybridGradientParams& params,
    bool force_numerical_termination, const VectorXd& working_primal_current,
    const VectorXd& working_dual_current,
    const VectorXd* working_primal_average,
    const VectorXd* working_dual_average, const VectorXd* working_primal_delta,
    const VectorXd* working_dual_delta, const VectorXd& last_primal_start_point,
    const VectorXd& last_dual_start_point,
    const std::atomic<bool>* interrupt_solve,
    const IterationType iteration_type, const IterationStats& full_stats,
    IterationStats& stats) {
  ComputeConvergenceAndInfeasibilityFromWorkingSolution(
      params, working_primal_current, working_dual_current,
      POINT_TYPE_CURRENT_ITERATE, stats.add_convergence_information(),
      stats.add_infeasibility_information());
  AddPointMetadata(params, working_primal_current, working_dual_current,
                   POINT_TYPE_CURRENT_ITERATE, last_primal_start_point,
                   last_dual_start_point, stats);
  if (working_primal_average != nullptr && working_dual_average != nullptr) {
    ComputeConvergenceAndInfeasibilityFromWorkingSolution(
        params, *working_primal_average, *working_dual_average,
        POINT_TYPE_AVERAGE_ITERATE, stats.add_convergence_information(),
        stats.add_infeasibility_information());
    AddPointMetadata(params, *working_primal_average, *working_dual_average,
                     POINT_TYPE_AVERAGE_ITERATE, last_primal_start_point,
                     last_dual_start_point, stats);
  }
  // Undoing presolve doesn't work for iterate differences.
  if (!presolve_info_.has_value() && working_primal_delta != nullptr &&
      working_dual_delta != nullptr) {
    ComputeConvergenceAndInfeasibilityFromWorkingSolution(
        params, *working_primal_delta, *working_dual_delta,
        POINT_TYPE_ITERATE_DIFFERENCE, nullptr,
        stats.add_infeasibility_information());
    AddPointMetadata(params, *working_primal_delta, *working_dual_delta,
                     POINT_TYPE_ITERATE_DIFFERENCE, last_primal_start_point,
                     last_dual_start_point, stats);
  }
  constexpr int kLogEvery = 15;
  absl::Time logging_time = absl::Now();
  if (params.verbosity_level() >= 2 &&
      (params.log_interval_seconds() == 0.0 ||
       logging_time - time_of_last_log_ >=
           absl::Seconds(params.log_interval_seconds()))) {
    if (log_counter_ == 0) {
      LogIterationStatsHeader(params.verbosity_level(),
                              params.use_feasibility_polishing(), logger_);
    }
    LogIterationStats(params.verbosity_level(),
                      params.use_feasibility_polishing(), iteration_type, stats,
                      params.termination_criteria(), original_bound_norms_,
                      POINT_TYPE_AVERAGE_ITERATE, logger_);
    if (params.verbosity_level() >= 4) {
      // If the convergence information doesn't contain an average iterate, the
      // previous `LogIterationStats()` will report some other iterate (usually
      // the current one) so don't repeat logging it.
      if (GetConvergenceInformation(stats, POINT_TYPE_AVERAGE_ITERATE) !=
          std::nullopt) {
        LogIterationStats(
            params.verbosity_level(), params.use_feasibility_polishing(),
            iteration_type, stats, params.termination_criteria(),
            original_bound_norms_, POINT_TYPE_CURRENT_ITERATE, logger_);
      }
    }
    time_of_last_log_ = logging_time;
    if (++log_counter_ >= kLogEvery) {
      log_counter_ = 0;
    }
  }
  if (iteration_stats_callback_ != nullptr) {
    iteration_stats_callback_(
        {.iteration_type = iteration_type,
         .termination_criteria = params.termination_criteria(),
         .iteration_stats = stats,
         .bound_norms = original_bound_norms_});
  }
 
  if (const auto termination = CheckIterateTerminationCriteria(
          params.termination_criteria(), stats, original_bound_norms_,
          force_numerical_termination);
      termination.has_value()) {
    return termination;
  }
  return CheckSimpleTerminationCriteria(params.termination_criteria(),
                                        full_stats, interrupt_solve);
}
 
void PreprocessSolver::ComputeConvergenceAndInfeasibilityFromWorkingSolution(
    const PrimalDualHybridGradientParams& params,
    const VectorXd& working_primal, const VectorXd& working_dual,
    PointType candidate_type, ConvergenceInformation* convergence_information,
    InfeasibilityInformation* infeasibility_information) const {
  const TerminationCriteria::DetailedOptimalityCriteria criteria =
      EffectiveOptimalityCriteria(params.termination_criteria());
  const double primal_epsilon_ratio =
      EpsilonRatio(criteria.eps_optimal_primal_residual_absolute(),
                   criteria.eps_optimal_primal_residual_relative());
  const double dual_epsilon_ratio =
      EpsilonRatio(criteria.eps_optimal_dual_residual_absolute(),
                   criteria.eps_optimal_dual_residual_relative());
  if (presolve_info_.has_value()) {
    // Undoing presolve doesn't make sense for iterate differences.
    CHECK_NE(candidate_type, POINT_TYPE_ITERATE_DIFFERENCE);
 
    PrimalAndDualSolution original = RecoverOriginalSolution(
        {.primal_solution = working_primal, .dual_solution = working_dual});
    if (convergence_information != nullptr) {
      *convergence_information = ComputeConvergenceInformation(
          params, presolve_info_->sharded_original_qp,
          presolve_info_->trivial_col_scaling_vec,
          presolve_info_->trivial_row_scaling_vec, original.primal_solution,
          original.dual_solution, primal_epsilon_ratio, dual_epsilon_ratio,
          candidate_type);
    }
    if (infeasibility_information != nullptr) {
      VectorXd primal_copy =
          CloneVector(original.primal_solution,
                      presolve_info_->sharded_original_qp.PrimalSharder());
      ProjectToPrimalVariableBounds(presolve_info_->sharded_original_qp,
                                    primal_copy,
                                    /*use_feasibility_bounds=*/true);
      // Only iterate differences should be able to violate the dual variable
      // bounds, and iterate differences aren't used when presolve is enabled.
      *infeasibility_information = ComputeInfeasibilityInformation(
          params, presolve_info_->sharded_original_qp,
          presolve_info_->trivial_col_scaling_vec,
          presolve_info_->trivial_row_scaling_vec, primal_copy,
          original.dual_solution, original.primal_solution, candidate_type);
    }
  } else {
    if (convergence_information != nullptr) {
      *convergence_information = ComputeConvergenceInformation(
          params, sharded_qp_, col_scaling_vec_, row_scaling_vec_,
          working_primal, working_dual, primal_epsilon_ratio,
          dual_epsilon_ratio, candidate_type);
    }
    if (infeasibility_information != nullptr) {
      VectorXd primal_copy =
          CloneVector(working_primal, sharded_qp_.PrimalSharder());
      ProjectToPrimalVariableBounds(sharded_qp_, primal_copy,
                                    /*use_feasibility_bounds=*/true);
      if (candidate_type == POINT_TYPE_ITERATE_DIFFERENCE) {
        // Iterate differences might not satisfy the dual variable bounds.
        VectorXd dual_copy =
            CloneVector(working_dual, sharded_qp_.DualSharder());
        ProjectToDualVariableBounds(sharded_qp_, dual_copy);
        *infeasibility_information = ComputeInfeasibilityInformation(
            params, sharded_qp_, col_scaling_vec_, row_scaling_vec_,
            primal_copy, dual_copy, working_primal, candidate_type);
      } else {
        *infeasibility_information = ComputeInfeasibilityInformation(
            params, sharded_qp_, col_scaling_vec_, row_scaling_vec_,
            primal_copy, working_dual, working_primal, candidate_type);
      }
    }
  }
}
 
// `result` is used both as the input and as the temporary that will be
// returned.
SolverResult PreprocessSolver::ConstructOriginalSolverResult(
    const PrimalDualHybridGradientParams& params, SolverResult result,
    SolverLogger& logger) const {
  const bool use_zero_primal_objective =
      result.solve_log.termination_reason() ==
      TERMINATION_REASON_PRIMAL_INFEASIBLE;
  if (presolve_info_.has_value()) {
    // Transform the solutions so they match the original unscaled problem.
    PrimalAndDualSolution original_solution = RecoverOriginalSolution(
        {.primal_solution = std::move(result.primal_solution),
         .dual_solution = std::move(result.dual_solution)});
    result.primal_solution = std::move(original_solution.primal_solution);
    if (result.solve_log.termination_reason() ==
        TERMINATION_REASON_DUAL_INFEASIBLE) {
      ProjectToPrimalVariableBounds(presolve_info_->sharded_original_qp,
                                    result.primal_solution,
                                    /*use_feasibility_bounds=*/true);
    }
    // If presolve is enabled, no projection of the dual is performed when
    // checking for infeasibility.
    result.dual_solution = std::move(original_solution.dual_solution);
    // `RecoverOriginalSolution` doesn't recover reduced costs so we need to
    // compute them with respect to the original problem.
    result.reduced_costs = ReducedCosts(
        params, presolve_info_->sharded_original_qp, result.primal_solution,
        result.dual_solution, use_zero_primal_objective);
  } else {
    if (result.solve_log.termination_reason() ==
        TERMINATION_REASON_DUAL_INFEASIBLE) {
      ProjectToPrimalVariableBounds(sharded_qp_, result.primal_solution,
                                    /*use_feasibility_bounds=*/true);
    }
    if (result.solve_log.termination_reason() ==
        TERMINATION_REASON_PRIMAL_INFEASIBLE) {
      ProjectToDualVariableBounds(sharded_qp_, result.dual_solution);
    }
    result.reduced_costs =
        ReducedCosts(params, sharded_qp_, result.primal_solution,
                     result.dual_solution, use_zero_primal_objective);
    // Transform the solutions so they match the original unscaled problem.
    CoefficientWiseProductInPlace(col_scaling_vec_, sharded_qp_.PrimalSharder(),
                                  result.primal_solution);
    CoefficientWiseProductInPlace(row_scaling_vec_, sharded_qp_.DualSharder(),
                                  result.dual_solution);
    CoefficientWiseQuotientInPlace(
        col_scaling_vec_, sharded_qp_.PrimalSharder(), result.reduced_costs);
  }
  IterationType iteration_type;
  switch (result.solve_log.solution_type()) {
    case POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION:
      iteration_type = IterationType::kFeasibilityPolishingTermination;
      break;
    case POINT_TYPE_PRESOLVER_SOLUTION:
      iteration_type = IterationType::kPresolveTermination;
      break;
    default:
      iteration_type = IterationType::kNormalTermination;
      break;
  }
  if (iteration_stats_callback_ != nullptr) {
    iteration_stats_callback_(
        {.iteration_type = iteration_type,
         .termination_criteria = params.termination_criteria(),
         .iteration_stats = result.solve_log.solution_stats(),
         .bound_norms = original_bound_norms_});
  }
 
  if (params.verbosity_level() >= 1) {
    SOLVER_LOG(&logger, "Termination reason: ",
               TerminationReason_Name(result.solve_log.termination_reason()));
    SOLVER_LOG(&logger, "Solution point type: ",
               PointType_Name(result.solve_log.solution_type()));
    SOLVER_LOG(&logger, "Final solution stats:");
    LogIterationStatsHeader(params.verbosity_level(),
                            params.use_feasibility_polishing(), logger);
    LogIterationStats(params.verbosity_level(),
                      params.use_feasibility_polishing(), iteration_type,
                      result.solve_log.solution_stats(),
                      params.termination_criteria(), original_bound_norms_,
                      result.solve_log.solution_type(), logger);
    const auto& convergence_info = GetConvergenceInformation(
        result.solve_log.solution_stats(), result.solve_log.solution_type());
    if (convergence_info.has_value()) {
      if (std::isfinite(convergence_info->corrected_dual_objective())) {
        SOLVER_LOG(&logger, "Dual objective after infeasibility correction: ",
                   convergence_info->corrected_dual_objective());
      }
    }
  }
  return result;
}
 
Solver::Solver(const PrimalDualHybridGradientParams& params,
               VectorXd starting_primal_solution,
               VectorXd starting_dual_solution, const double initial_step_size,
               const double initial_primal_weight,
               PreprocessSolver* preprocess_solver)
    : params_(params),
      current_primal_solution_(std::move(starting_primal_solution)),
      current_dual_solution_(std::move(starting_dual_solution)),
      primal_average_(&preprocess_solver->ShardedWorkingQp().PrimalSharder()),
      dual_average_(&preprocess_solver->ShardedWorkingQp().DualSharder()),
      step_size_(initial_step_size),
      primal_weight_(initial_primal_weight),
      preprocess_solver_(preprocess_solver) {}
 
Solver::NextSolutionAndDelta Solver::ComputeNextPrimalSolution(
    double primal_step_size) const {
  const int64_t primal_size = ShardedWorkingQp().PrimalSize();
  NextSolutionAndDelta result = {
      .value = VectorXd(primal_size),
      .delta = VectorXd(primal_size),
  };
  const QuadraticProgram& qp = WorkingQp();
  // This computes the primal portion of the PDHG algorithm:
  // argmin_x[gradient(f)(`current_primal_solution_`)^T x + g(x)
  //   + `current_dual_solution_`^T K x
  //   + (0.5 / `primal_step_size`) * norm(x - `current_primal_solution_`)^2]
  // See Sections 2 - 3 of Chambolle and Pock and the comment in the header.
  // We omitted the constant terms from Chambolle and Pock's (7).
  // This minimization is easy to do in closed form since it can be separated
  // into independent problems for each of the primal variables.
  ShardedWorkingQp().PrimalSharder().ParallelForEachShard(
      [&](const Sharder::Shard& shard) {
        if (!IsLinearProgram(qp)) {
          // TODO(user): Does changing this to auto (so it becomes an
          // Eigen deferred result), or inlining it below, change performance?
          const VectorXd diagonal_scaling =
              primal_step_size *
                  shard(qp.objective_matrix->diagonal()).array() +
              1.0;
          shard(result.value) =
              (shard(current_primal_solution_) -
               primal_step_size *
                   (shard(qp.objective_vector) - shard(current_dual_product_)))
                  // Scale i-th element by 1 / (1 + `primal_step_size` * Q_{ii})
                  .cwiseQuotient(diagonal_scaling)
                  .cwiseMin(shard(qp.variable_upper_bounds))
                  .cwiseMax(shard(qp.variable_lower_bounds));
        } else {
          // The formula in the LP case is simplified for better performance.
          shard(result.value) =
              (shard(current_primal_solution_) -
               primal_step_size *
                   (shard(qp.objective_vector) - shard(current_dual_product_)))
                  .cwiseMin(shard(qp.variable_upper_bounds))
                  .cwiseMax(shard(qp.variable_lower_bounds));
        }
        shard(result.delta) =
            shard(result.value) - shard(current_primal_solution_);
      });
  return result;
}
 
Solver::NextSolutionAndDelta Solver::ComputeNextDualSolution(
    double dual_step_size, double extrapolation_factor,
    const NextSolutionAndDelta& next_primal_solution,
    const VectorXd* absl_nullable next_primal_product) const {
  const int64_t dual_size = ShardedWorkingQp().DualSize();
  NextSolutionAndDelta result = {
      .value = VectorXd(dual_size),
      .delta = VectorXd(dual_size),
  };
  const QuadraticProgram& qp = WorkingQp();
  std::optional<VectorXd> extrapolated_primal;
  if (!next_primal_product) {
    extrapolated_primal.emplace(ShardedWorkingQp().PrimalSize());
    ShardedWorkingQp().PrimalSharder().ParallelForEachShard(
        [&](const Sharder::Shard& shard) {
          shard(*extrapolated_primal) =
              (shard(next_primal_solution.value) +
               extrapolation_factor * shard(next_primal_solution.delta));
        });
  }
  ShardedWorkingQp().TransposedConstraintMatrixSharder().ParallelForEachShard(
      [&](const Sharder::Shard& shard) {
        VectorXd temp;
        if (next_primal_product) {
          CHECK(current_primal_product_.has_value());
          temp = shard(current_dual_solution_) -
                 dual_step_size *
                     (-extrapolation_factor * shard(*current_primal_product_) +
                      (extrapolation_factor + 1) * shard(*next_primal_product));
        } else {
          temp = shard(current_dual_solution_) -
                 dual_step_size *
                     shard(ShardedWorkingQp().TransposedConstraintMatrix())
                         .transpose() *
                     extrapolated_primal.value();
        }
 
        // Each element of the argument of `.cwiseMin()` is the critical point
        // of the respective 1D minimization problem if it's negative.
        // Likewise the argument to the `.cwiseMax()` is the critical point if
        // positive.
        shard(result.value) =
            VectorXd::Zero(temp.size())
                .cwiseMin(temp +
                          dual_step_size * shard(qp.constraint_upper_bounds))
                .cwiseMax(temp +
                          dual_step_size * shard(qp.constraint_lower_bounds));
        shard(result.delta) =
            (shard(result.value) - shard(current_dual_solution_));
      });
  return result;
}
 
std::pair<double, double> Solver::ComputeMovementTerms(
    const VectorXd& delta_primal, const VectorXd& delta_dual) const {
  return {SquaredNorm(delta_primal, ShardedWorkingQp().PrimalSharder()),
          SquaredNorm(delta_dual, ShardedWorkingQp().DualSharder())};
}
 
double Solver::ComputeMovement(const VectorXd& delta_primal,
                               const VectorXd& delta_dual) const {
  const auto [primal_squared_norm, dual_squared_norm] =
      ComputeMovementTerms(delta_primal, delta_dual);
  return (0.5 * primal_weight_ * primal_squared_norm) +
         (0.5 / primal_weight_) * dual_squared_norm;
}
 
double Solver::ComputeNonlinearity(const VectorXd& delta_primal,
                                   const VectorXd& next_dual_product) const {
  // Lemma 1 in Chambolle and Pock includes a term with L_f, the Lipshitz
  // constant of f. This is zero in our formulation.
  return ShardedWorkingQp().PrimalSharder().ParallelSumOverShards(
      [&](const Sharder::Shard& shard) {
        return -shard(delta_primal)
                    .dot(shard(next_dual_product) -
                         shard(current_dual_product_));
      });
}
 
void Solver::SetCurrentPrimalAndDualProducts() {
  if (params_.linesearch_rule() ==
      PrimalDualHybridGradientParams::MALITSKY_POCK_LINESEARCH_RULE) {
    current_primal_product_ = TransposedMatrixVectorProduct(
        ShardedWorkingQp().TransposedConstraintMatrix(),
        current_primal_solution_,
        ShardedWorkingQp().TransposedConstraintMatrixSharder());
  } else {
    current_primal_product_.reset();
  }
  current_dual_product_ = TransposedMatrixVectorProduct(
      WorkingQp().constraint_matrix, current_dual_solution_,
      ShardedWorkingQp().ConstraintMatrixSharder());
}
 
IterationStats Solver::CreateSimpleIterationStats(
    RestartChoice restart_used) const {
  IterationStats stats;
  double num_kkt_passes_per_rejected_step = 1.0;
  if (params_.linesearch_rule() ==
      PrimalDualHybridGradientParams::MALITSKY_POCK_LINESEARCH_RULE) {
    num_kkt_passes_per_rejected_step = 0.5;
  }
  stats.set_iteration_number(iterations_completed_);
  stats.set_cumulative_rejected_steps(num_rejected_steps_);
  // TODO(user): This formula doesn't account for kkt passes in major
  // iterations.
  stats.set_cumulative_kkt_matrix_passes(iterations_completed_ +
                                         num_kkt_passes_per_rejected_step *
                                             num_rejected_steps_);
  stats.set_cumulative_time_sec(preprocessing_time_sec_ + timer_.Get());
  stats.set_restart_used(restart_used);
  stats.set_step_size(step_size_);
  stats.set_primal_weight(primal_weight_);
  return stats;
}
 
double Solver::DistanceTraveledFromLastStart(
    const VectorXd& primal_solution, const VectorXd& dual_solution) const {
  return std::sqrt((0.5 * primal_weight_) *
                       SquaredDistance(primal_solution,
                                       last_primal_start_point_,
                                       ShardedWorkingQp().PrimalSharder()) +
                   (0.5 / primal_weight_) *
                       SquaredDistance(dual_solution, last_dual_start_point_,
                                       ShardedWorkingQp().DualSharder()));
}
 
LocalizedLagrangianBounds Solver::ComputeLocalizedBoundsAtCurrent() const {
  const double distance_traveled_by_current = DistanceTraveledFromLastStart(
      current_primal_solution_, current_dual_solution_);
  return ComputeLocalizedLagrangianBounds(
      ShardedWorkingQp(), current_primal_solution_, current_dual_solution_,
      PrimalDualNorm::kEuclideanNorm, primal_weight_,
      distance_traveled_by_current,
      /*primal_product=*/current_primal_product_.has_value()
          ? &current_primal_product_.value()
          : nullptr,
      &current_dual_product_, params_.use_diagonal_qp_trust_region_solver(),
      params_.diagonal_qp_trust_region_solver_tolerance());
}
 
LocalizedLagrangianBounds Solver::ComputeLocalizedBoundsAtAverage() const {
  // TODO(user): These vectors are recomputed again for termination checks
  // and again if we eventually restart to the average.
  VectorXd average_primal = PrimalAverage();
  VectorXd average_dual = DualAverage();
 
  const double distance_traveled_by_average =
      DistanceTraveledFromLastStart(average_primal, average_dual);
 
  return ComputeLocalizedLagrangianBounds(
      ShardedWorkingQp(), average_primal, average_dual,
      PrimalDualNorm::kEuclideanNorm, primal_weight_,
      distance_traveled_by_average,
      /*primal_product=*/nullptr, /*dual_product=*/nullptr,
      params_.use_diagonal_qp_trust_region_solver(),
      params_.diagonal_qp_trust_region_solver_tolerance());
}
 
bool AverageHasBetterPotential(
    const LocalizedLagrangianBounds& local_bounds_at_average,
    const LocalizedLagrangianBounds& local_bounds_at_current) {
  return BoundGap(local_bounds_at_average) /
             MathUtil::Square(local_bounds_at_average.radius) <
         BoundGap(local_bounds_at_current) /
             MathUtil::Square(local_bounds_at_current.radius);
}
 
double NormalizedGap(
    const LocalizedLagrangianBounds& local_bounds_at_candidate) {
  const double distance_traveled_by_candidate =
      local_bounds_at_candidate.radius;
  return BoundGap(local_bounds_at_candidate) / distance_traveled_by_candidate;
}
 
// TODO(user): Review / cleanup adaptive heuristic.
bool Solver::ShouldDoAdaptiveRestartHeuristic(
    double candidate_normalized_gap) const {
  const double gap_reduction_ratio =
      candidate_normalized_gap / normalized_gap_at_last_restart_;
  if (gap_reduction_ratio < params_.sufficient_reduction_for_restart()) {
    return true;
  }
  if (gap_reduction_ratio < params_.necessary_reduction_for_restart() &&
      candidate_normalized_gap > normalized_gap_at_last_trial_) {
    // We've made the "necessary" amount of progress, and iterates appear to
    // be getting worse, so restart.
    return true;
  }
  return false;
}
 
RestartChoice Solver::DetermineDistanceBasedRestartChoice() const {
  // The following checks are safeguards that normally should not be triggered.
  if (primal_average_.NumTerms() == 0) {
    return RESTART_CHOICE_NO_RESTART;
  } else if (distance_based_restart_info_.length_of_last_restart_period == 0) {
    return RESTART_CHOICE_RESTART_TO_AVERAGE;
  }
  const int restart_period_length = primal_average_.NumTerms();
  const double distance_moved_this_restart_period_by_average =
      DistanceTraveledFromLastStart(primal_average_.ComputeAverage(),
                                    dual_average_.ComputeAverage());
  const double distance_moved_last_restart_period =
      distance_based_restart_info_.distance_moved_last_restart_period;
 
  // A restart should be triggered when the normalized distance traveled by
  // the average is at least a constant factor smaller than the last.
  // TODO(user): Experiment with using `.necessary_reduction_for_restart()`
  // as a heuristic when deciding if a restart should be triggered.
  if ((distance_moved_this_restart_period_by_average / restart_period_length) <
      params_.sufficient_reduction_for_restart() *
          (distance_moved_last_restart_period /
           distance_based_restart_info_.length_of_last_restart_period)) {
    // Restart at current solution when it yields a smaller normalized potential
    // function value than the average (heuristic suggested by ohinder@).
    if (AverageHasBetterPotential(ComputeLocalizedBoundsAtAverage(),
                                  ComputeLocalizedBoundsAtCurrent())) {
      return RESTART_CHOICE_RESTART_TO_AVERAGE;
    } else {
      return RESTART_CHOICE_WEIGHTED_AVERAGE_RESET;
    }
  } else {
    return RESTART_CHOICE_NO_RESTART;
  }
}
 
RestartChoice Solver::ChooseRestartToApply(const bool is_major_iteration) {
  if (!primal_average_.HasNonzeroWeight() &&
      !dual_average_.HasNonzeroWeight()) {
    return RESTART_CHOICE_NO_RESTART;
  }
  // TODO(user): This forced restart is very important for the performance of
  // `ADAPTIVE_HEURISTIC`. Test if the impact comes primarily from the first
  // forced restart (which would unseat a good initial starting point that could
  // prevent restarts early in the solve) or if it's really needed for the full
  // duration of the solve. If it is really needed, should we then trigger major
  // iterations on powers of two?
  const int restart_length = primal_average_.NumTerms();
  if (restart_length >= iterations_completed_ / 2 &&
      params_.restart_strategy() ==
          PrimalDualHybridGradientParams::ADAPTIVE_HEURISTIC) {
    if (AverageHasBetterPotential(ComputeLocalizedBoundsAtAverage(),
                                  ComputeLocalizedBoundsAtCurrent())) {
      return RESTART_CHOICE_RESTART_TO_AVERAGE;
    } else {
      return RESTART_CHOICE_WEIGHTED_AVERAGE_RESET;
    }
  }
  if (is_major_iteration) {
    switch (params_.restart_strategy()) {
      case PrimalDualHybridGradientParams::NO_RESTARTS:
        return RESTART_CHOICE_WEIGHTED_AVERAGE_RESET;
      case PrimalDualHybridGradientParams::EVERY_MAJOR_ITERATION:
        return RESTART_CHOICE_RESTART_TO_AVERAGE;
      case PrimalDualHybridGradientParams::ADAPTIVE_HEURISTIC: {
        const LocalizedLagrangianBounds local_bounds_at_average =
            ComputeLocalizedBoundsAtAverage();
        const LocalizedLagrangianBounds local_bounds_at_current =
            ComputeLocalizedBoundsAtCurrent();
        double normalized_gap;
        RestartChoice choice;
        if (AverageHasBetterPotential(local_bounds_at_average,
                                      local_bounds_at_current)) {
          normalized_gap = NormalizedGap(local_bounds_at_average);
          choice = RESTART_CHOICE_RESTART_TO_AVERAGE;
        } else {
          normalized_gap = NormalizedGap(local_bounds_at_current);
          choice = RESTART_CHOICE_WEIGHTED_AVERAGE_RESET;
        }
        if (ShouldDoAdaptiveRestartHeuristic(normalized_gap)) {
          return choice;
        } else {
          normalized_gap_at_last_trial_ = normalized_gap;
          return RESTART_CHOICE_NO_RESTART;
        }
      }
      case PrimalDualHybridGradientParams::ADAPTIVE_DISTANCE_BASED: {
        return DetermineDistanceBasedRestartChoice();
      }
      default:
        LOG(FATAL) << "Unrecognized restart_strategy "
                   << params_.restart_strategy();
        return RESTART_CHOICE_UNSPECIFIED;
    }
  } else {
    return RESTART_CHOICE_NO_RESTART;
  }
}
 
VectorXd Solver::PrimalAverage() const {
  if (primal_average_.HasNonzeroWeight()) {
    return primal_average_.ComputeAverage();
  } else {
    return current_primal_solution_;
  }
}
 
VectorXd Solver::DualAverage() const {
  if (dual_average_.HasNonzeroWeight()) {
    return dual_average_.ComputeAverage();
  } else {
    return current_dual_solution_;
  }
}
 
double Solver::ComputeNewPrimalWeight() const {
  const double primal_distance =
      Distance(current_primal_solution_, last_primal_start_point_,
               ShardedWorkingQp().PrimalSharder());
  const double dual_distance =
      Distance(current_dual_solution_, last_dual_start_point_,
               ShardedWorkingQp().DualSharder());
  // This choice of a nonzero tolerance balances performance and numerical
  // issues caused by very huge or very tiny weights. It was picked as the best
  // among {0.0, 1.0e-20, 2.0e-16, 1.0e-10, 1.0e-5} on the preprocessed MIPLIB
  // dataset. The effect of changing this value is relatively minor overall.
  constexpr double kNonzeroTol = 1.0e-10;
  if (primal_distance <= kNonzeroTol || primal_distance >= 1.0 / kNonzeroTol ||
      dual_distance <= kNonzeroTol || dual_distance >= 1.0 / kNonzeroTol) {
    return primal_weight_;
  }
  const double smoothing_param = params_.primal_weight_update_smoothing();
  const double unsmoothed_new_primal_weight = dual_distance / primal_distance;
  const double new_primal_weight =
      std::exp(smoothing_param * std::log(unsmoothed_new_primal_weight) +
               (1.0 - smoothing_param) * std::log(primal_weight_));
  if (params_.verbosity_level() >= 4) {
    SOLVER_LOG(&preprocess_solver_->Logger(), "New computed primal weight is ",
               new_primal_weight, " at iteration ", iterations_completed_);
  }
  return new_primal_weight;
}
 
SolverResult Solver::PickSolutionAndConstructSolverResult(
    VectorXd primal_solution, VectorXd dual_solution,
    const IterationStats& stats, TerminationReason termination_reason,
    PointType output_type, SolveLog solve_log) const {
  switch (output_type) {
    case POINT_TYPE_CURRENT_ITERATE:
      AssignVector(current_primal_solution_, ShardedWorkingQp().PrimalSharder(),
                   primal_solution);
      AssignVector(current_dual_solution_, ShardedWorkingQp().DualSharder(),
                   dual_solution);
      break;
    case POINT_TYPE_ITERATE_DIFFERENCE:
      AssignVector(current_primal_delta_, ShardedWorkingQp().PrimalSharder(),
                   primal_solution);
      AssignVector(current_dual_delta_, ShardedWorkingQp().DualSharder(),
                   dual_solution);
      break;
    case POINT_TYPE_AVERAGE_ITERATE:
    case POINT_TYPE_PRESOLVER_SOLUTION:
      break;
    default:
      // Default to average whenever `output_type` is `POINT_TYPE_NONE`.
      output_type = POINT_TYPE_AVERAGE_ITERATE;
      break;
  }
  return ConstructSolverResult(
      std::move(primal_solution), std::move(dual_solution), stats,
      termination_reason, output_type, std::move(solve_log));
}
 
void Solver::ApplyRestartChoice(const RestartChoice restart_to_apply) {
  switch (restart_to_apply) {
    case RESTART_CHOICE_UNSPECIFIED:
    case RESTART_CHOICE_NO_RESTART:
      return;
    case RESTART_CHOICE_WEIGHTED_AVERAGE_RESET:
      if (params_.verbosity_level() >= 4) {
        SOLVER_LOG(&preprocess_solver_->Logger(),
                   "Restarted to current on iteration ", iterations_completed_,
                   " after ", primal_average_.NumTerms(), " iterations");
      }
      break;
    case RESTART_CHOICE_RESTART_TO_AVERAGE:
      if (params_.verbosity_level() >= 4) {
        SOLVER_LOG(&preprocess_solver_->Logger(),
                   "Restarted to average on iteration ", iterations_completed_,
                   " after ", primal_average_.NumTerms(), " iterations");
      }
      current_primal_solution_ = primal_average_.ComputeAverage();
      current_dual_solution_ = dual_average_.ComputeAverage();
      SetCurrentPrimalAndDualProducts();
      break;
  }
  primal_weight_ = ComputeNewPrimalWeight();
  ratio_last_two_step_sizes_ = 1;
  if (params_.restart_strategy() ==
      PrimalDualHybridGradientParams::ADAPTIVE_HEURISTIC) {
    // It's important for the theory that the distances here are calculated
    // given the new primal weight.
    const LocalizedLagrangianBounds local_bounds_at_last_restart =
        ComputeLocalizedBoundsAtCurrent();
    const double distance_traveled_since_last_restart =
        local_bounds_at_last_restart.radius;
    normalized_gap_at_last_restart_ = BoundGap(local_bounds_at_last_restart) /
                                      distance_traveled_since_last_restart;
    normalized_gap_at_last_trial_ = std::numeric_limits<double>::infinity();
  } else if (params_.restart_strategy() ==
             PrimalDualHybridGradientParams::ADAPTIVE_DISTANCE_BASED) {
    // Update parameters for distance-based restarts.
    distance_based_restart_info_ = {
        .distance_moved_last_restart_period = DistanceTraveledFromLastStart(
            current_primal_solution_, current_dual_solution_),
        .length_of_last_restart_period = primal_average_.NumTerms()};
  }
  primal_average_.Clear();
  dual_average_.Clear();
  AssignVector(current_primal_solution_, ShardedWorkingQp().PrimalSharder(),
               /*dest=*/last_primal_start_point_);
  AssignVector(current_dual_solution_, ShardedWorkingQp().DualSharder(),
               /*dest=*/last_dual_start_point_);
}
 
// Adds the work (`iteration_number`, `cumulative_kkt_matrix_passes`,
// `cumulative_rejected_steps`, and `cumulative_time_sec`) from
// `additional_work_stats` to `stats` and returns the result.
// `stats` is intentionally passed by value.
IterationStats AddWorkStats(IterationStats stats,
                            const IterationStats& additional_work_stats) {
  stats.set_iteration_number(stats.iteration_number() +
                             additional_work_stats.iteration_number());
  stats.set_cumulative_kkt_matrix_passes(
      stats.cumulative_kkt_matrix_passes() +
      additional_work_stats.cumulative_kkt_matrix_passes());
  stats.set_cumulative_rejected_steps(
      stats.cumulative_rejected_steps() +
      additional_work_stats.cumulative_rejected_steps());
  stats.set_cumulative_time_sec(stats.cumulative_time_sec() +
                                additional_work_stats.cumulative_time_sec());
  return stats;
}
 
// Accumulates and returns the work (`iteration_number`,
// `cumulative_kkt_matrix_passes`, `cumulative_rejected_steps`, and
// `cumulative_time_sec`) from `solve_log.feasibility_polishing_details`.
IterationStats WorkFromFeasibilityPolishing(const SolveLog& solve_log) {
  IterationStats result;
  for (const FeasibilityPolishingDetails& feasibility_polishing_detail :
       solve_log.feasibility_polishing_details()) {
    result = AddWorkStats(std::move(result),
                          feasibility_polishing_detail.solution_stats());
  }
  return result;
}
 
bool TerminationReasonIsInterrupted(const TerminationReason reason) {
  return reason == TERMINATION_REASON_INTERRUPTED_BY_USER;
}
 
bool TerminationReasonIsWorkLimitNotInterrupted(
    const TerminationReason reason) {
  return reason == TERMINATION_REASON_ITERATION_LIMIT ||
         reason == TERMINATION_REASON_TIME_LIMIT ||
         reason == TERMINATION_REASON_KKT_MATRIX_PASS_LIMIT;
}
 
// Note: `TERMINATION_REASON_INTERRUPTED_BY_USER` is treated as a work limit
// (that was determined in real-time by the user).
bool TerminationReasonIsWorkLimit(const TerminationReason reason) {
  return TerminationReasonIsWorkLimitNotInterrupted(reason) ||
         TerminationReasonIsInterrupted(reason);
}
 
bool DoFeasibilityPolishingAfterLimitsReached(
    const PrimalDualHybridGradientParams& params,
    const TerminationReason reason) {
  if (TerminationReasonIsWorkLimitNotInterrupted(reason)) {
    return params.apply_feasibility_polishing_after_limits_reached();
  }
  if (TerminationReasonIsInterrupted(reason)) {
    return params.apply_feasibility_polishing_if_solver_is_interrupted();
  }
  return false;
}
 
std::optional<SolverResult> Solver::MajorIterationAndTerminationCheck(
    const IterationType iteration_type, const bool force_numerical_termination,
    const std::atomic<bool>* interrupt_solve,
    const IterationStats& work_from_feasibility_polishing,
    SolveLog& solve_log) {
  const int major_iteration_cycle =
      iterations_completed_ % params_.major_iteration_frequency();
  const bool is_major_iteration =
      major_iteration_cycle == 0 && iterations_completed_ > 0;
  // Just decide what to do for now. The actual restart, if any, is
  // performed after the termination check.
  const RestartChoice restart = force_numerical_termination
                                    ? RESTART_CHOICE_NO_RESTART
                                    : ChooseRestartToApply(is_major_iteration);
  IterationStats stats = CreateSimpleIterationStats(restart);
  IterationStats full_work_stats =
      AddWorkStats(stats, work_from_feasibility_polishing);
  std::optional<TerminationReasonAndPointType> simple_termination_reason =
      CheckSimpleTerminationCriteria(params_.termination_criteria(),
                                     full_work_stats, interrupt_solve);
  const bool check_termination =
      major_iteration_cycle % params_.termination_check_frequency() == 0 ||
      simple_termination_reason.has_value() || force_numerical_termination;
  // We check termination on every major iteration.
  DCHECK(!is_major_iteration || check_termination);
  if (check_termination) {
    // Check for termination and update iteration stats with both simple and
    // solution statistics. The later are computationally harder to compute and
    // hence only computed here.
    VectorXd primal_average = PrimalAverage();
    VectorXd dual_average = DualAverage();
 
    const std::optional<TerminationReasonAndPointType>
        maybe_termination_reason =
            preprocess_solver_->UpdateIterationStatsAndCheckTermination(
                params_, force_numerical_termination, current_primal_solution_,
                current_dual_solution_,
                primal_average_.HasNonzeroWeight() ? &primal_average : nullptr,
                dual_average_.HasNonzeroWeight() ? &dual_average : nullptr,
                current_primal_delta_.size() > 0 ? &current_primal_delta_
                                                 : nullptr,
                current_dual_delta_.size() > 0 ? &current_dual_delta_ : nullptr,
                last_primal_start_point_, last_dual_start_point_,
                interrupt_solve, iteration_type, full_work_stats, stats);
    if (params_.record_iteration_stats()) {
      *solve_log.add_iteration_stats() = stats;
    }
    // We've terminated.
    if (maybe_termination_reason.has_value()) {
      if (iteration_type == IterationType::kNormal &&
          DoFeasibilityPolishingAfterLimitsReached(
              params_, maybe_termination_reason->reason)) {
        const std::optional<SolverResult> feasibility_result =
            TryFeasibilityPolishing(
                iterations_completed_ / kFeasibilityIterationFraction,
                interrupt_solve, solve_log);
        if (feasibility_result.has_value()) {
          LOG(INFO) << "Returning result from feasibility polishing after "
                       "limits reached";
          return *feasibility_result;
        }
      }
      IterationStats terminating_full_stats =
          AddWorkStats(stats, work_from_feasibility_polishing);
      return PickSolutionAndConstructSolverResult(
          std::move(primal_average), std::move(dual_average),
          terminating_full_stats, maybe_termination_reason->reason,
          maybe_termination_reason->type, std::move(solve_log));
    }
  } else if (params_.record_iteration_stats()) {
    // Record simple iteration stats only.
    *solve_log.add_iteration_stats() = stats;
  }
  ApplyRestartChoice(restart);
  return std::nullopt;
}
 
void Solver::ResetAverageToCurrent() {
  primal_average_.Clear();
  dual_average_.Clear();
  primal_average_.Add(current_primal_solution_, /*weight=*/1.0);
  dual_average_.Add(current_dual_solution_, /*weight=*/1.0);
}
 
void Solver::LogNumericalTermination(const Eigen::VectorXd& primal_delta,
                                     const Eigen::VectorXd& dual_delta) const {
  if (params_.verbosity_level() >= 2) {
    auto [primal_squared_norm, dual_squared_norm] =
        ComputeMovementTerms(primal_delta, dual_delta);
    SOLVER_LOG(&preprocess_solver_->Logger(),
               "Forced numerical termination at iteration ",
               iterations_completed_, " with primal delta squared norm ",
               primal_squared_norm, " dual delta squared norm ",
               dual_squared_norm, " primal weight ", primal_weight_);
  }
}
 
void Solver::LogInnerIterationLimitHit() const {
  SOLVER_LOG(&preprocess_solver_->Logger(),
             "WARNING: Inner iteration limit reached at iteration ",
             iterations_completed_);
}
 
InnerStepOutcome Solver::TakeMalitskyPockStep() {
  InnerStepOutcome outcome = InnerStepOutcome::kSuccessful;
  const double primal_step_size = step_size_ / primal_weight_;
  NextSolutionAndDelta next_primal_solution =
      ComputeNextPrimalSolution(primal_step_size);
  // The theory by Malitsky and Pock holds for any new_step_size in the interval
  // [`step_size`, `step_size` * sqrt(1 + `ratio_last_two_step_sizes_`)].
  // `dilating_coeff` determines where in this interval the new step size lands.
  // NOTE: Malitsky and Pock use theta for `ratio_last_two_step_sizes`.
  double dilating_coeff =
      1 + (params_.malitsky_pock_parameters().step_size_interpolation() *
           (sqrt(1 + ratio_last_two_step_sizes_) - 1));
  double new_primal_step_size = primal_step_size * dilating_coeff;
  double step_size_downscaling =
      params_.malitsky_pock_parameters().step_size_downscaling_factor();
  double contraction_factor =
      params_.malitsky_pock_parameters().linesearch_contraction_factor();
  const double dual_weight = primal_weight_ * primal_weight_;
  int inner_iterations = 0;
  VectorXd next_primal_product(current_dual_solution_.size());
  ShardedWorkingQp().TransposedConstraintMatrixSharder().ParallelForEachShard(
      [&](const Sharder::Shard& shard) {
        shard(next_primal_product) =
            shard(ShardedWorkingQp().TransposedConstraintMatrix()).transpose() *
            next_primal_solution.value;
      });
 
  for (bool accepted_step = false; !accepted_step; ++inner_iterations) {
    if (inner_iterations >= 60) {
      LogInnerIterationLimitHit();
      ResetAverageToCurrent();
      outcome = InnerStepOutcome::kForceNumericalTermination;
      break;
    }
    const double new_last_two_step_sizes_ratio =
        new_primal_step_size / primal_step_size;
    NextSolutionAndDelta next_dual_solution = ComputeNextDualSolution(
        dual_weight * new_primal_step_size, new_last_two_step_sizes_ratio,
        next_primal_solution, &next_primal_product);
 
    VectorXd next_dual_product = TransposedMatrixVectorProduct(
        WorkingQp().constraint_matrix, next_dual_solution.value,
        ShardedWorkingQp().ConstraintMatrixSharder());
    double delta_dual_norm =
        Norm(next_dual_solution.delta, ShardedWorkingQp().DualSharder());
    double delta_dual_prod_norm =
        Distance(current_dual_product_, next_dual_product,
                 ShardedWorkingQp().PrimalSharder());
    if (primal_weight_ * new_primal_step_size * delta_dual_prod_norm <=
        contraction_factor * delta_dual_norm) {
      // Accept new_step_size as a good step.
      step_size_ = new_primal_step_size * primal_weight_;
      ratio_last_two_step_sizes_ = new_last_two_step_sizes_ratio;
      // Malitsky and Pock guarantee uses a nonsymmetric weighted average,
      // the primal variable average involves the initial point, while the dual
      // doesn't. See Theorem 2 in https://arxiv.org/pdf/1608.08883.pdf for
      // details.
      if (!primal_average_.HasNonzeroWeight()) {
        primal_average_.Add(
            current_primal_solution_,
            /*weight=*/new_primal_step_size * new_last_two_step_sizes_ratio);
      }
 
      current_primal_solution_ = std::move(next_primal_solution.value);
      current_dual_solution_ = std::move(next_dual_solution.value);
      current_dual_product_ = std::move(next_dual_product);
      current_primal_product_ = std::move(next_primal_product);
      primal_average_.Add(current_primal_solution_,
                          /*weight=*/new_primal_step_size);
      dual_average_.Add(current_dual_solution_,
                        /*weight=*/new_primal_step_size);
      const double movement =
          ComputeMovement(next_primal_solution.delta, next_dual_solution.delta);
      if (movement == 0.0) {
        LogNumericalTermination(next_primal_solution.delta,
                                next_dual_solution.delta);
        ResetAverageToCurrent();
        outcome = InnerStepOutcome::kForceNumericalTermination;
      } else if (movement > kDivergentMovement) {
        LogNumericalTermination(next_primal_solution.delta,
                                next_dual_solution.delta);
        outcome = InnerStepOutcome::kForceNumericalTermination;
      }
      current_primal_delta_ = std::move(next_primal_solution.delta);
      current_dual_delta_ = std::move(next_dual_solution.delta);
      break;
    } else {
      new_primal_step_size = step_size_downscaling * new_primal_step_size;
    }
  }
  // `inner_iterations` isn't incremented for the accepted step.
  num_rejected_steps_ += inner_iterations;
  return outcome;
}
 
InnerStepOutcome Solver::TakeAdaptiveStep() {
  InnerStepOutcome outcome = InnerStepOutcome::kSuccessful;
  int inner_iterations = 0;
  for (bool accepted_step = false; !accepted_step; ++inner_iterations) {
    if (inner_iterations >= 60) {
      LogInnerIterationLimitHit();
      ResetAverageToCurrent();
      outcome = InnerStepOutcome::kForceNumericalTermination;
      break;
    }
    const double primal_step_size = step_size_ / primal_weight_;
    const double dual_step_size = step_size_ * primal_weight_;
    NextSolutionAndDelta next_primal_solution =
        ComputeNextPrimalSolution(primal_step_size);
    NextSolutionAndDelta next_dual_solution = ComputeNextDualSolution(
        dual_step_size, /*extrapolation_factor=*/1.0, next_primal_solution);
    const double movement =
        ComputeMovement(next_primal_solution.delta, next_dual_solution.delta);
    if (movement == 0.0) {
      LogNumericalTermination(next_primal_solution.delta,
                              next_dual_solution.delta);
      ResetAverageToCurrent();
      outcome = InnerStepOutcome::kForceNumericalTermination;
      break;
    } else if (movement > kDivergentMovement) {
      LogNumericalTermination(next_primal_solution.delta,
                              next_dual_solution.delta);
      outcome = InnerStepOutcome::kForceNumericalTermination;
      break;
    }
    VectorXd next_dual_product = TransposedMatrixVectorProduct(
        WorkingQp().constraint_matrix, next_dual_solution.value,
        ShardedWorkingQp().ConstraintMatrixSharder());
    const double nonlinearity =
        ComputeNonlinearity(next_primal_solution.delta, next_dual_product);
 
    // See equation (5) in https://arxiv.org/pdf/2106.04756.pdf.
    const double step_size_limit =
        nonlinearity > 0 ? movement / nonlinearity
                         : std::numeric_limits<double>::infinity();
 
    if (step_size_ <= step_size_limit) {
      current_primal_solution_ = std::move(next_primal_solution.value);
      current_dual_solution_ = std::move(next_dual_solution.value);
      current_dual_product_ = std::move(next_dual_product);
      current_primal_product_.reset();
      current_primal_delta_ = std::move(next_primal_solution.delta);
      current_dual_delta_ = std::move(next_dual_solution.delta);
      primal_average_.Add(current_primal_solution_, /*weight=*/step_size_);
      dual_average_.Add(current_dual_solution_, /*weight=*/step_size_);
      accepted_step = true;
    }
    const double total_steps_attempted =
        num_rejected_steps_ + inner_iterations + iterations_completed_ + 1;
    // Our step sizes are a factor 1 - (`total_steps_attempted` + 1)^(-
    // `step_size_reduction_exponent`) smaller than they could be as a margin to
    // reduce rejected steps.
    // The std::isinf() test protects against NAN if std::pow() == 1.0.
    const double first_term =
        std::isinf(step_size_limit)
            ? step_size_limit
            : (1 - std::pow(total_steps_attempted + 1.0,
                            -params_.adaptive_linesearch_parameters()
                                 .step_size_reduction_exponent())) *
                  step_size_limit;
    const double second_term =
        (1 + std::pow(total_steps_attempted + 1.0,
                      -params_.adaptive_linesearch_parameters()
                           .step_size_growth_exponent())) *
        step_size_;
    // From the first term when we have to reject a step, `step_size_`
    // decreases by a factor of at least 1 - (`total_steps_attempted` + 1)^(-
    // `step_size_reduction_exponent`). From the second term we increase
    // `step_size_` by a factor of at most 1 + (`total_steps_attempted` +
    // 1)^(-`step_size_growth_exponent`) Therefore if more than order
    // (`total_steps_attempted` + 1)^(`step_size_reduction_exponent`
    // - `step_size_growth_exponent`) fraction of the time we have a rejected
    // step, we overall decrease `step_size_`. When `step_size_` is
    // sufficiently small we stop having rejected steps.
    step_size_ = std::min(first_term, second_term);
  }
  // `inner_iterations` is incremented for the accepted step.
  num_rejected_steps_ += inner_iterations - 1;
  return outcome;
}
 
InnerStepOutcome Solver::TakeConstantSizeStep() {
  const double primal_step_size = step_size_ / primal_weight_;
  const double dual_step_size = step_size_ * primal_weight_;
  NextSolutionAndDelta next_primal_solution =
      ComputeNextPrimalSolution(primal_step_size);
  NextSolutionAndDelta next_dual_solution = ComputeNextDualSolution(
      dual_step_size, /*extrapolation_factor=*/1.0, next_primal_solution);
  const double movement =
      ComputeMovement(next_primal_solution.delta, next_dual_solution.delta);
  if (movement == 0.0) {
    LogNumericalTermination(next_primal_solution.delta,
                            next_dual_solution.delta);
    ResetAverageToCurrent();
    return InnerStepOutcome::kForceNumericalTermination;
  } else if (movement > kDivergentMovement) {
    LogNumericalTermination(next_primal_solution.delta,
                            next_dual_solution.delta);
    return InnerStepOutcome::kForceNumericalTermination;
  }
  VectorXd next_dual_product = TransposedMatrixVectorProduct(
      WorkingQp().constraint_matrix, next_dual_solution.value,
      ShardedWorkingQp().ConstraintMatrixSharder());
  current_primal_solution_ = std::move(next_primal_solution.value);
  current_dual_solution_ = std::move(next_dual_solution.value);
  current_dual_product_ = std::move(next_dual_product);
  current_primal_product_.reset();
  current_primal_delta_ = std::move(next_primal_solution.delta);
  current_dual_delta_ = std::move(next_dual_solution.delta);
  primal_average_.Add(current_primal_solution_, /*weight=*/step_size_);
  dual_average_.Add(current_dual_solution_, /*weight=*/step_size_);
  return InnerStepOutcome::kSuccessful;
}
 
IterationStats Solver::TotalWorkSoFar(const SolveLog& solve_log) const {
  IterationStats stats = CreateSimpleIterationStats(RESTART_CHOICE_NO_RESTART);
  IterationStats full_stats =
      AddWorkStats(stats, WorkFromFeasibilityPolishing(solve_log));
  return full_stats;
}
 
FeasibilityPolishingDetails BuildFeasibilityPolishingDetails(
    PolishingPhaseType phase_type, int iteration_count,
    const PrimalDualHybridGradientParams& params, const SolveLog& solve_log) {
  FeasibilityPolishingDetails details;
  details.set_polishing_phase_type(phase_type);
  details.set_main_iteration_count(iteration_count);
  *details.mutable_params() = params;
  details.set_termination_reason(solve_log.termination_reason());
  details.set_iteration_count(solve_log.iteration_count());
  details.set_solve_time_sec(solve_log.solve_time_sec());
  *details.mutable_solution_stats() = solve_log.solution_stats();
  details.set_solution_type(solve_log.solution_type());
  absl::c_copy(solve_log.iteration_stats(),
               google::protobuf::RepeatedPtrFieldBackInserter(
                   details.mutable_iteration_stats()));
  return details;
}
 
std::optional<SolverResult> Solver::TryFeasibilityPolishing(
    const int iteration_limit, const std::atomic<bool>* interrupt_solve,
    SolveLog& solve_log) {
  TerminationCriteria::DetailedOptimalityCriteria optimality_criteria =
      EffectiveOptimalityCriteria(params_.termination_criteria());
 
  VectorXd average_primal = PrimalAverage();
  VectorXd average_dual = DualAverage();
 
  ConvergenceInformation first_convergence_info;
  preprocess_solver_->ComputeConvergenceAndInfeasibilityFromWorkingSolution(
      params_, average_primal, average_dual, POINT_TYPE_AVERAGE_ITERATE,
      &first_convergence_info, nullptr);
 
  // NOTE: Even though the objective gap sometimes is reduced by feasibility
  // polishing, it is usually increased, and an experiment (on MIPLIB2017)
  // shows that this test reduces the iteration count by 3-4% on average.
  if (!ObjectiveGapMet(optimality_criteria, first_convergence_info)) {
    std::optional<TerminationReasonAndPointType> simple_termination_reason =
        CheckSimpleTerminationCriteria(params_.termination_criteria(),
                                       TotalWorkSoFar(solve_log),
                                       interrupt_solve);
    if (!(simple_termination_reason.has_value() &&
          DoFeasibilityPolishingAfterLimitsReached(
              params_, simple_termination_reason->reason))) {
      if (params_.verbosity_level() >= 2) {
        SOLVER_LOG(&preprocess_solver_->Logger(),
                   "Skipping feasibility polishing because the objective gap "
                   "is too large.");
      }
      return std::nullopt;
    }
  }
 
  if (params_.verbosity_level() >= 2) {
    SOLVER_LOG(&preprocess_solver_->Logger(),
               "Starting primal feasibility polishing");
  }
  SolverResult primal_result = TryPrimalPolishing(
      std::move(average_primal), iteration_limit, interrupt_solve, solve_log);
 
  if (params_.verbosity_level() >= 2) {
    SOLVER_LOG(
        &preprocess_solver_->Logger(),
        "Primal feasibility polishing termination reason: ",
        TerminationReason_Name(primal_result.solve_log.termination_reason()));
  }
  if (TerminationReasonIsWorkLimit(
          primal_result.solve_log.termination_reason())) {
    // Have we also reached the overall work limit? If so, consider finishing
    // the final polishing phase.
    std::optional<TerminationReasonAndPointType> simple_termination_reason =
        CheckSimpleTerminationCriteria(params_.termination_criteria(),
                                       TotalWorkSoFar(solve_log),
                                       interrupt_solve);
    if (!(simple_termination_reason.has_value() &&
          DoFeasibilityPolishingAfterLimitsReached(
              params_, simple_termination_reason->reason))) {
      return std::nullopt;
    }
  } else if (primal_result.solve_log.termination_reason() !=
             TERMINATION_REASON_OPTIMAL) {
    // Note: `TERMINATION_REASON_PRIMAL_INFEASIBLE` could happen normally, but
    // we haven't ensured that the correct solution is returned in that case, so
    // we ignore the polishing result indicating infeasibility.
    // `TERMINATION_REASON_NUMERICAL_ERROR` can occur, but would be surprising
    // and interesting. Other termination reasons are probably bugs.
    SOLVER_LOG(&preprocess_solver_->Logger(),
               "WARNING: Primal feasibility polishing terminated with error ",
               primal_result.solve_log.termination_reason());
    return std::nullopt;
  }
 
  if (params_.verbosity_level() >= 2) {
    SOLVER_LOG(&preprocess_solver_->Logger(),
               "Starting dual feasibility polishing");
  }
  SolverResult dual_result = TryDualPolishing(
      std::move(average_dual), iteration_limit, interrupt_solve, solve_log);
 
  if (params_.verbosity_level() >= 2) {
    SOLVER_LOG(
        &preprocess_solver_->Logger(),
        "Dual feasibility polishing termination reason: ",
        TerminationReason_Name(dual_result.solve_log.termination_reason()));
  }
 
  IterationStats full_stats = TotalWorkSoFar(solve_log);
  std::optional<TerminationReasonAndPointType> simple_termination_reason =
      CheckSimpleTerminationCriteria(params_.termination_criteria(), full_stats,
                                     interrupt_solve);
  if (TerminationReasonIsWorkLimit(
          dual_result.solve_log.termination_reason())) {
    // Have we also reached the overall work limit? If so, consider falling out
    // of the "if" test and returning the polishing solution anyway.
    if (simple_termination_reason.has_value() &&
        DoFeasibilityPolishingAfterLimitsReached(
            params_, simple_termination_reason->reason)) {
      preprocess_solver_->ComputeConvergenceAndInfeasibilityFromWorkingSolution(
          params_, primal_result.primal_solution, dual_result.dual_solution,
          POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION,
          full_stats.add_convergence_information(), nullptr);
      return ConstructSolverResult(
          std::move(primal_result.primal_solution),
          std::move(dual_result.dual_solution), full_stats,
          simple_termination_reason->reason,
          POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION, solve_log);
    } else {
      return std::nullopt;
    }
  } else if (dual_result.solve_log.termination_reason() !=
             TERMINATION_REASON_OPTIMAL) {
    // Note: The comment in the corresponding location when checking the
    // termination reason for primal feasibility polishing applies here
    // too.
    SOLVER_LOG(&preprocess_solver_->Logger(),
               "WARNING: Dual feasibility polishing terminated with error ",
               dual_result.solve_log.termination_reason());
    return std::nullopt;
  }
 
  preprocess_solver_->ComputeConvergenceAndInfeasibilityFromWorkingSolution(
      params_, primal_result.primal_solution, dual_result.dual_solution,
      POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION,
      full_stats.add_convergence_information(), nullptr);
  if (params_.verbosity_level() >= 2) {
    SOLVER_LOG(&preprocess_solver_->Logger(),
               "solution stats for polished solution:");
    LogIterationStatsHeader(params_.verbosity_level(),
                            /*use_feasibility_polishing=*/true,
                            preprocess_solver_->Logger());
    LogIterationStats(params_.verbosity_level(),
                      /*use_feasibility_polishing=*/true,
                      IterationType::kFeasibilityPolishingTermination,
                      full_stats, params_.termination_criteria(),
                      preprocess_solver_->OriginalBoundNorms(),
                      POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION,
                      preprocess_solver_->Logger());
  }
  std::optional<TerminationReasonAndPointType> earned_termination =
      CheckIterateTerminationCriteria(params_.termination_criteria(),
                                      full_stats,
                                      preprocess_solver_->OriginalBoundNorms(),
                                      /*force_numerical_termination=*/false);
  if (earned_termination.has_value() ||
      (simple_termination_reason.has_value() &&
       DoFeasibilityPolishingAfterLimitsReached(
           params_, simple_termination_reason->reason))) {
    return ConstructSolverResult(
        std::move(primal_result.primal_solution),
        std::move(dual_result.dual_solution), full_stats,
        earned_termination.has_value() ? earned_termination->reason
                                       : simple_termination_reason->reason,
        POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION, solve_log);
  }
  // Note: A typical termination check would now call
  // `CheckSimpleTerminationCriteria`. However, there is no obvious iterate to
  // use for a `SolverResult`. Instead, allow the main iteration loop to notice
  // if the termination criteria checking has caused us to meet the simple
  // termination criteria. This would be a rare case anyway, since dual
  // feasibility polishing didn't terminate because of a work limit, and
  // `full_stats` is created shortly after `TryDualPolishing` returned.
  return std::nullopt;
}
 
TerminationCriteria ReduceWorkLimitsByPreviousWork(
    TerminationCriteria criteria, const int iteration_limit,
    const IterationStats& previous_work,
    bool apply_feasibility_polishing_after_limits_reached) {
  if (apply_feasibility_polishing_after_limits_reached) {
    criteria.set_iteration_limit(iteration_limit);
    criteria.set_kkt_matrix_pass_limit(std::numeric_limits<double>::infinity());
    criteria.set_time_sec_limit(std::numeric_limits<double>::infinity());
  } else {
    criteria.set_iteration_limit(std::max(
        0, std::min(iteration_limit, criteria.iteration_limit() -
                                         previous_work.iteration_number())));
    criteria.set_kkt_matrix_pass_limit(
        std::max(0.0, criteria.kkt_matrix_pass_limit() -
                          previous_work.cumulative_kkt_matrix_passes()));
    criteria.set_time_sec_limit(std::max(
        0.0, criteria.time_sec_limit() - previous_work.cumulative_time_sec()));
  }
  return criteria;
}
 
SolverResult Solver::TryPrimalPolishing(
    VectorXd starting_primal_solution, const int iteration_limit,
    const std::atomic<bool>* interrupt_solve, SolveLog& solve_log) {
  PrimalDualHybridGradientParams primal_feasibility_params = params_;
  *primal_feasibility_params.mutable_termination_criteria() =
      ReduceWorkLimitsByPreviousWork(
          params_.termination_criteria(), iteration_limit,
          TotalWorkSoFar(solve_log),
          params_.apply_feasibility_polishing_after_limits_reached());
  if (params_.apply_feasibility_polishing_if_solver_is_interrupted()) {
    interrupt_solve = nullptr;
  }
 
  // This will save the original objective after the swap.
  VectorXd objective;
  SetZero(ShardedWorkingQp().PrimalSharder(), objective);
  preprocess_solver_->SwapObjectiveVector(objective);
 
  TerminationCriteria::DetailedOptimalityCriteria criteria =
      EffectiveOptimalityCriteria(params_.termination_criteria());
  const double kInfinity = std::numeric_limits<double>::infinity();
  criteria.set_eps_optimal_dual_residual_absolute(kInfinity);
  criteria.set_eps_optimal_dual_residual_relative(kInfinity);
  criteria.set_eps_optimal_objective_gap_absolute(kInfinity);
  criteria.set_eps_optimal_objective_gap_relative(kInfinity);
  *primal_feasibility_params.mutable_termination_criteria()
       ->mutable_detailed_optimality_criteria() = criteria;
  // TODO(user): Evaluate disabling primal weight updates for this primal
  // feasibility problem.
  VectorXd primal_feasibility_starting_dual;
  SetZero(ShardedWorkingQp().DualSharder(), primal_feasibility_starting_dual);
  Solver primal_solver(primal_feasibility_params,
                       std::move(starting_primal_solution),
                       std::move(primal_feasibility_starting_dual), step_size_,
                       primal_weight_, preprocess_solver_);
  SolveLog primal_solve_log;
  // Stop `timer_`, since the time in `primal_solver.Solve()` will be recorded
  // by its timer.
  timer_.Stop();
  SolverResult primal_result = primal_solver.Solve(
      IterationType::kPrimalFeasibility, interrupt_solve, primal_solve_log);
  timer_.Start();
  // Restore the original objective.
  preprocess_solver_->SwapObjectiveVector(objective);
 
  *solve_log.add_feasibility_polishing_details() =
      BuildFeasibilityPolishingDetails(
          POLISHING_PHASE_TYPE_PRIMAL_FEASIBILITY, iterations_completed_,
          primal_feasibility_params, primal_result.solve_log);
  return primal_result;
}
 
VectorXd MapFiniteValuesToZero(const Sharder& sharder, const VectorXd& input) {
  VectorXd output(input.size());
  const auto make_finite_values_zero = [](const double x) {
    return std::isfinite(x) ? 0.0 : x;
  };
  sharder.ParallelForEachShard([&](const Sharder::Shard& shard) {
    shard(output) = shard(input).unaryExpr(make_finite_values_zero);
  });
  return output;
}
 
SolverResult Solver::TryDualPolishing(VectorXd starting_dual_solution,
                                      const int iteration_limit,
                                      const std::atomic<bool>* interrupt_solve,
                                      SolveLog& solve_log) {
  PrimalDualHybridGradientParams dual_feasibility_params = params_;
  *dual_feasibility_params.mutable_termination_criteria() =
      ReduceWorkLimitsByPreviousWork(
          params_.termination_criteria(), iteration_limit,
          TotalWorkSoFar(solve_log),
          params_.apply_feasibility_polishing_after_limits_reached());
  if (params_.apply_feasibility_polishing_if_solver_is_interrupted()) {
    interrupt_solve = nullptr;
  }
 
  // These will initially contain the homogenous variable and constraint
  // bounds, but will contain the original variable and constraint bounds
  // after the swap.
  VectorXd constraint_lower_bounds = MapFiniteValuesToZero(
      ShardedWorkingQp().DualSharder(), WorkingQp().constraint_lower_bounds);
  VectorXd constraint_upper_bounds = MapFiniteValuesToZero(
      ShardedWorkingQp().DualSharder(), WorkingQp().constraint_upper_bounds);
  VectorXd variable_lower_bounds = MapFiniteValuesToZero(
      ShardedWorkingQp().PrimalSharder(), WorkingQp().variable_lower_bounds);
  VectorXd variable_upper_bounds = MapFiniteValuesToZero(
      ShardedWorkingQp().PrimalSharder(), WorkingQp().variable_upper_bounds);
  preprocess_solver_->SwapConstraintBounds(constraint_lower_bounds,
                                           constraint_upper_bounds);
  preprocess_solver_->SwapVariableBounds(variable_lower_bounds,
                                         variable_upper_bounds);
 
  TerminationCriteria::DetailedOptimalityCriteria criteria =
      EffectiveOptimalityCriteria(params_.termination_criteria());
  const double kInfinity = std::numeric_limits<double>::infinity();
  criteria.set_eps_optimal_primal_residual_absolute(kInfinity);
  criteria.set_eps_optimal_primal_residual_relative(kInfinity);
  criteria.set_eps_optimal_objective_gap_absolute(kInfinity);
  criteria.set_eps_optimal_objective_gap_relative(kInfinity);
  *dual_feasibility_params.mutable_termination_criteria()
       ->mutable_detailed_optimality_criteria() = criteria;
  // TODO(user): Evaluate disabling primal weight updates for this dual
  // feasibility problem.
  VectorXd dual_feasibility_starting_primal;
  SetZero(ShardedWorkingQp().PrimalSharder(), dual_feasibility_starting_primal);
  Solver dual_solver(dual_feasibility_params,
                     std::move(dual_feasibility_starting_primal),
                     std::move(starting_dual_solution), step_size_,
                     primal_weight_, preprocess_solver_);
  SolveLog dual_solve_log;
  // Stop `timer_`, since the time in `dual_solver.Solve()` will be recorded
  // by its timer.
  timer_.Stop();
  SolverResult dual_result = dual_solver.Solve(IterationType::kDualFeasibility,
                                               interrupt_solve, dual_solve_log);
  timer_.Start();
  // Restore the original bounds.
  preprocess_solver_->SwapConstraintBounds(constraint_lower_bounds,
                                           constraint_upper_bounds);
  preprocess_solver_->SwapVariableBounds(variable_lower_bounds,
                                         variable_upper_bounds);
  *solve_log.add_feasibility_polishing_details() =
      BuildFeasibilityPolishingDetails(
          POLISHING_PHASE_TYPE_DUAL_FEASIBILITY, iterations_completed_,
          dual_feasibility_params, dual_result.solve_log);
  return dual_result;
}
 
SolverResult Solver::Solve(const IterationType iteration_type,
                           const std::atomic<bool>* interrupt_solve,
                           SolveLog solve_log) {
  preprocessing_time_sec_ = solve_log.preprocessing_time_sec();
  timer_.Start();
  last_primal_start_point_ =
      CloneVector(current_primal_solution_, ShardedWorkingQp().PrimalSharder());
  last_dual_start_point_ =
      CloneVector(current_dual_solution_, ShardedWorkingQp().DualSharder());
  // Note: Any cached values computed here also need to be recomputed after a
  // restart.
 
  ratio_last_two_step_sizes_ = 1;
  SetCurrentPrimalAndDualProducts();
 
  // This is set to true if we can't proceed any more because of numerical
  // issues. We may or may not have found the optimal solution.
  bool force_numerical_termination = false;
 
  int next_feasibility_polishing_iteration = 100;
 
  num_rejected_steps_ = 0;
 
  IterationStats work_from_feasibility_polishing =
      WorkFromFeasibilityPolishing(solve_log);
  for (iterations_completed_ = 0;; ++iterations_completed_) {
    // This code performs the logic of the major iterations and termination
    // checks. It may modify the current solution and primal weight (e.g., when
    // performing a restart).
    const std::optional<SolverResult> maybe_result =
        MajorIterationAndTerminationCheck(
            iteration_type, force_numerical_termination, interrupt_solve,
            work_from_feasibility_polishing, solve_log);
    if (maybe_result.has_value()) {
      return maybe_result.value();
    }
 
    if (params_.use_feasibility_polishing() &&
        iteration_type == IterationType::kNormal &&
        iterations_completed_ >= next_feasibility_polishing_iteration) {
      const std::optional<SolverResult> feasibility_result =
          TryFeasibilityPolishing(
              iterations_completed_ / kFeasibilityIterationFraction,
              interrupt_solve, solve_log);
      if (feasibility_result.has_value()) {
        return *feasibility_result;
      }
      next_feasibility_polishing_iteration *= 2;
      // Update work to include new feasibility phases.
      work_from_feasibility_polishing = WorkFromFeasibilityPolishing(solve_log);
    }
 
    // TODO(user): If we use a step rule that could reject many steps in a
    // row, we should add a termination check within this loop also. For the
    // Malitsky and Pock rule, we perform a termination check and declare
    // NUMERICAL_ERROR whenever we hit 60 inner iterations.
    InnerStepOutcome outcome;
    switch (params_.linesearch_rule()) {
      case PrimalDualHybridGradientParams::MALITSKY_POCK_LINESEARCH_RULE:
        outcome = TakeMalitskyPockStep();
        break;
      case PrimalDualHybridGradientParams::ADAPTIVE_LINESEARCH_RULE:
        outcome = TakeAdaptiveStep();
        break;
      case PrimalDualHybridGradientParams::CONSTANT_STEP_SIZE_RULE:
        outcome = TakeConstantSizeStep();
        break;
      default:
        LOG(FATAL) << "Unrecognized linesearch rule "
                   << params_.linesearch_rule();
    }
    if (outcome == InnerStepOutcome::kForceNumericalTermination) {
      force_numerical_termination = true;
    }
  }  // loop over iterations
}
 
}  // namespace
 
SolverResult PrimalDualHybridGradient(
    QuadraticProgram qp, const PrimalDualHybridGradientParams& params,
    const std::atomic<bool>* interrupt_solve,
    std::function<void(const std::string&)> message_callback,
    IterationStatsCallback iteration_stats_callback) {
  return PrimalDualHybridGradient(std::move(qp), params, std::nullopt,
                                  interrupt_solve, std::move(message_callback),
                                  std::move(iteration_stats_callback));
}
 
// See comment at top of file.
SolverResult PrimalDualHybridGradient(
    QuadraticProgram qp, const PrimalDualHybridGradientParams& params,
    std::optional<PrimalAndDualSolution> initial_solution,
    const std::atomic<bool>* interrupt_solve,
    std::function<void(const std::string&)> message_callback,
    IterationStatsCallback iteration_stats_callback) {
  SolverLogger logger;
  logger.EnableLogging(true);
  if (message_callback) {
    logger.AddInfoLoggingCallback(std::move(message_callback));
  } else {
    logger.SetLogToStdOut(true);
  }
  const absl::Status params_status =
      ValidatePrimalDualHybridGradientParams(params);
  if (!params_status.ok()) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PARAMETER,
                             params_status.ToString(), logger);
  }
  if (!qp.constraint_matrix.isCompressed()) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "constraint_matrix must be in compressed format. "
                             "Call constraint_matrix.makeCompressed()",
                             logger);
  }
  const absl::Status dimensions_status = ValidateQuadraticProgramDimensions(qp);
  if (!dimensions_status.ok()) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             dimensions_status.ToString(), logger);
  }
  if (qp.objective_scaling_factor == 0) {
    return ErrorSolverResult(TERMINATION_REASON_INVALID_PROBLEM,
                             "The objective scaling factor cannot be zero.",
                             logger);
  }
  if (params.use_feasibility_polishing() && !IsLinearProgram(qp)) {
    return ErrorSolverResult(
        TERMINATION_REASON_INVALID_PARAMETER,
        "use_feasibility_polishing is only implemented for linear programs.",
        logger);
  }
  PreprocessSolver solver(std::move(qp), params, &logger);
  return solver.PreprocessAndSolve(params, std::move(initial_solution),
                                   interrupt_solve,
                                   std::move(iteration_stats_callback));
}
 
namespace internal {
 
glop::ProblemSolution ComputeStatuses(const QuadraticProgram& qp,
                                      const PrimalAndDualSolution& solution) {
  glop::ProblemSolution glop_solution(
      glop::RowIndex(solution.dual_solution.size()),
      glop::ColIndex(solution.primal_solution.size()));
  // This doesn't matter much as glop's preprocessor doesn't use this much.
  // We pick IMPRECISE since we are often calling this code early in the solve.
  glop_solution.status = glop::ProblemStatus::IMPRECISE;
  for (glop::RowIndex i{0}; i.value() < solution.dual_solution.size(); ++i) {
    if (qp.constraint_lower_bounds[i.value()] ==
        qp.constraint_upper_bounds[i.value()]) {
      glop_solution.constraint_statuses[i] =
          glop::ConstraintStatus::FIXED_VALUE;
    } else if (solution.dual_solution[i.value()] > 0) {
      glop_solution.constraint_statuses[i] =
          glop::ConstraintStatus::AT_LOWER_BOUND;
    } else if (solution.dual_solution[i.value()] < 0) {
      glop_solution.constraint_statuses[i] =
          glop::ConstraintStatus::AT_UPPER_BOUND;
    } else {
      glop_solution.constraint_statuses[i] = glop::ConstraintStatus::BASIC;
    }
  }
 
  for (glop::ColIndex i{0}; i.value() < solution.primal_solution.size(); ++i) {
    const bool at_lb = solution.primal_solution[i.value()] <=
                       qp.variable_lower_bounds[i.value()];
    const bool at_ub = solution.primal_solution[i.value()] >=
                       qp.variable_upper_bounds[i.value()];
    // Note that `ShardedWeightedAverage` is designed so that variables at their
    // bounds will be exactly at their bounds even with floating-point roundoff.
    if (at_lb) {
      if (at_ub) {
        glop_solution.variable_statuses[i] = glop::VariableStatus::FIXED_VALUE;
      } else {
        glop_solution.variable_statuses[i] =
            glop::VariableStatus::AT_LOWER_BOUND;
      }
    } else {
      if (at_ub) {
        glop_solution.variable_statuses[i] =
            glop::VariableStatus::AT_UPPER_BOUND;
      } else {
        glop_solution.variable_statuses[i] = glop::VariableStatus::BASIC;
      }
    }
  }
  return glop_solution;
}
 
}  // namespace internal
 
}  // namespace operations_research::pdlp