cuts.cc

// Copyright 2010-2021 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.

#include "ortools/sat/cuts.h"

#include <algorithm>
#include <cmath>
#include <cstdint>
#include <functional>
#include <limits>
#include <memory>
#include <string>
#include <utility>
#include <vector>

#include "ortools/algorithms/knapsack_solver_for_cuts.h"
#include "ortools/base/integral_types.h"
#include "ortools/base/stl_util.h"
#include "ortools/base/strong_vector.h"
#include "ortools/sat/integer.h"
#include "ortools/sat/linear_constraint.h"
#include "ortools/sat/linear_constraint_manager.h"
#include "ortools/sat/sat_base.h"
#include "ortools/sat/util.h"
#include "ortools/util/time_limit.h"

namespace operations_research {
namespace sat {

namespace {

// Minimum amount of violation of the cut constraint by the solution. This
// is needed to avoid numerical issues and adding cuts with minor effect.
const double kMinCutViolation = 1e-4;

// Returns the lp value of a Literal.
double GetLiteralLpValue(
    const Literal lit,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    const IntegerEncoder* encoder) {
  const IntegerVariable direct_view = encoder->GetLiteralView(lit);
  if (direct_view != kNoIntegerVariable) {
    return lp_values[direct_view];
  }
  const IntegerVariable opposite_view = encoder->GetLiteralView(lit.Negated());
  DCHECK_NE(opposite_view, kNoIntegerVariable);
  return 1.0 - lp_values[opposite_view];
}

// Returns a constraint that disallow all given variables to be at their current
// upper bound. The arguments must form a non-trival constraint of the form
// sum terms (coeff * var) <= upper_bound.
LinearConstraint GenerateKnapsackCutForCover(
    const std::vector<IntegerVariable>& vars,
    const std::vector<IntegerValue>& coeffs, const IntegerValue upper_bound,
    const IntegerTrail& integer_trail) {
  CHECK_EQ(vars.size(), coeffs.size());
  CHECK_GT(vars.size(), 0);
  LinearConstraint cut;
  IntegerValue cut_upper_bound = IntegerValue(0);
  IntegerValue max_coeff = coeffs[0];
  // slack = \sum_{i}(coeffs[i] * upper_bound[i]) - upper_bound.
  IntegerValue slack = -upper_bound;
  for (int i = 0; i < vars.size(); ++i) {
    const IntegerValue var_upper_bound =
        integer_trail.LevelZeroUpperBound(vars[i]);
    cut_upper_bound += var_upper_bound;
    cut.vars.push_back(vars[i]);
    cut.coeffs.push_back(IntegerValue(1));
    max_coeff = std::max(max_coeff, coeffs[i]);
    slack += coeffs[i] * var_upper_bound;
  }
  CHECK_GT(slack, 0.0) << "Invalid cover for knapsack cut.";
  cut_upper_bound -= CeilRatio(slack, max_coeff);
  cut.lb = kMinIntegerValue;
  cut.ub = cut_upper_bound;
  VLOG(2) << "Generated Knapsack Constraint:" << cut.DebugString();
  return cut;
}

bool SolutionSatisfiesConstraint(
    const LinearConstraint& constraint,
    const absl::StrongVector<IntegerVariable, double>& lp_values) {
  const double activity = ComputeActivity(constraint, lp_values);
  const double tolerance = 1e-6;
  return (activity <= ToDouble(constraint.ub) + tolerance &&
          activity >= ToDouble(constraint.lb) - tolerance)
             ? true
             : false;
}

bool SmallRangeAndAllCoefficientsMagnitudeAreTheSame(
    const LinearConstraint& constraint, IntegerTrail* integer_trail) {
  if (constraint.vars.empty()) return true;

  const int64_t magnitude = std::abs(constraint.coeffs[0].value());
  for (int i = 1; i < constraint.coeffs.size(); ++i) {
    const IntegerVariable var = constraint.vars[i];
    if (integer_trail->LevelZeroUpperBound(var) -
            integer_trail->LevelZeroLowerBound(var) >
        1) {
      return false;
    }
    if (std::abs(constraint.coeffs[i].value()) != magnitude) {
      return false;
    }
  }
  return true;
}

bool AllVarsTakeIntegerValue(
    const std::vector<IntegerVariable> vars,
    const absl::StrongVector<IntegerVariable, double>& lp_values) {
  for (IntegerVariable var : vars) {
    if (std::abs(lp_values[var] - std::round(lp_values[var])) > 1e-6) {
      return false;
    }
  }
  return true;
}

// Returns smallest cover size for the given constraint taking into account
// level zero bounds. Smallest Cover size is computed as follows.
// 1. Compute the upper bound if all variables are shifted to have zero lower
//    bound.
// 2. Sort all terms (coefficient * shifted upper bound) in non decreasing
//    order.
// 3. Add terms in cover until term sum is smaller or equal to upper bound.
// 4. Add the last item which violates the upper bound. This forms the smallest
//    cover. Return the size of this cover.
int GetSmallestCoverSize(const LinearConstraint& constraint,
                         const IntegerTrail& integer_trail) {
  IntegerValue ub = constraint.ub;
  std::vector<IntegerValue> sorted_terms;
  for (int i = 0; i < constraint.vars.size(); ++i) {
    const IntegerValue coeff = constraint.coeffs[i];
    const IntegerVariable var = constraint.vars[i];
    const IntegerValue var_ub = integer_trail.LevelZeroUpperBound(var);
    const IntegerValue var_lb = integer_trail.LevelZeroLowerBound(var);
    ub -= var_lb * coeff;
    sorted_terms.push_back(coeff * (var_ub - var_lb));
  }
  std::sort(sorted_terms.begin(), sorted_terms.end(),
            std::greater<IntegerValue>());
  int smallest_cover_size = 0;
  IntegerValue sorted_term_sum = IntegerValue(0);
  while (sorted_term_sum <= ub &&
         smallest_cover_size < constraint.vars.size()) {
    sorted_term_sum += sorted_terms[smallest_cover_size++];
  }
  return smallest_cover_size;
}

bool ConstraintIsEligibleForLifting(const LinearConstraint& constraint,
                                    const IntegerTrail& integer_trail) {
  for (const IntegerVariable var : constraint.vars) {
    if (integer_trail.LevelZeroLowerBound(var) != IntegerValue(0) ||
        integer_trail.LevelZeroUpperBound(var) != IntegerValue(1)) {
      return false;
    }
  }
  return true;
}
}  // namespace

bool LiftKnapsackCut(
    const LinearConstraint& constraint,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    const std::vector<IntegerValue>& cut_vars_original_coefficients,
    const IntegerTrail& integer_trail, TimeLimit* time_limit,
    LinearConstraint* cut) {
  std::set<IntegerVariable> vars_in_cut;
  for (IntegerVariable var : cut->vars) {
    vars_in_cut.insert(var);
  }

  std::vector<std::pair<IntegerValue, IntegerVariable>> non_zero_vars;
  std::vector<std::pair<IntegerValue, IntegerVariable>> zero_vars;
  for (int i = 0; i < constraint.vars.size(); ++i) {
    const IntegerVariable var = constraint.vars[i];
    if (integer_trail.LevelZeroLowerBound(var) != IntegerValue(0) ||
        integer_trail.LevelZeroUpperBound(var) != IntegerValue(1)) {
      continue;
    }
    if (vars_in_cut.find(var) != vars_in_cut.end()) continue;
    const IntegerValue coeff = constraint.coeffs[i];
    if (lp_values[var] <= 1e-6) {
      zero_vars.push_back({coeff, var});
    } else {
      non_zero_vars.push_back({coeff, var});
    }
  }

  // Decide lifting sequence (nonzeros, zeros in nonincreasing order
  // of coefficient ).
  std::sort(non_zero_vars.rbegin(), non_zero_vars.rend());
  std::sort(zero_vars.rbegin(), zero_vars.rend());

  std::vector<std::pair<IntegerValue, IntegerVariable>> lifting_sequence(
      std::move(non_zero_vars));

  lifting_sequence.insert(lifting_sequence.end(), zero_vars.begin(),
                          zero_vars.end());

  // Form Knapsack.
  std::vector<double> lifting_profits;
  std::vector<double> lifting_weights;
  for (int i = 0; i < cut->vars.size(); ++i) {
    lifting_profits.push_back(ToDouble(cut->coeffs[i]));
    lifting_weights.push_back(ToDouble(cut_vars_original_coefficients[i]));
  }

  // Lift the cut.
  bool is_lifted = false;
  bool is_solution_optimal = false;
  KnapsackSolverForCuts knapsack_solver("Knapsack cut lifter");
  for (auto entry : lifting_sequence) {
    is_solution_optimal = false;
    const IntegerValue var_original_coeff = entry.first;
    const IntegerVariable var = entry.second;
    const IntegerValue lifting_capacity = constraint.ub - entry.first;
    if (lifting_capacity <= IntegerValue(0)) continue;
    knapsack_solver.Init(lifting_profits, lifting_weights,
                         ToDouble(lifting_capacity));
    knapsack_solver.set_node_limit(100);
    // NOTE: Since all profits and weights are integer, solution of
    // knapsack is also integer.
    // TODO(user): Use an integer solver or heuristic.
    knapsack_solver.Solve(time_limit, &is_solution_optimal);
    const double knapsack_upper_bound =
        std::round(knapsack_solver.GetUpperBound());
    const IntegerValue cut_coeff =
        cut->ub - static_cast<int64_t>(knapsack_upper_bound);
    if (cut_coeff > IntegerValue(0)) {
      is_lifted = true;
      cut->vars.push_back(var);
      cut->coeffs.push_back(cut_coeff);
      lifting_profits.push_back(ToDouble(cut_coeff));
      lifting_weights.push_back(ToDouble(var_original_coeff));
    }
  }
  return is_lifted;
}

LinearConstraint GetPreprocessedLinearConstraint(
    const LinearConstraint& constraint,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    const IntegerTrail& integer_trail) {
  IntegerValue ub = constraint.ub;
  LinearConstraint constraint_with_left_vars;
  for (int i = 0; i < constraint.vars.size(); ++i) {
    const IntegerVariable var = constraint.vars[i];
    const IntegerValue var_ub = integer_trail.LevelZeroUpperBound(var);
    const IntegerValue coeff = constraint.coeffs[i];
    if (ToDouble(var_ub) - lp_values[var] <= 1.0 - kMinCutViolation) {
      constraint_with_left_vars.vars.push_back(var);
      constraint_with_left_vars.coeffs.push_back(coeff);
    } else {
      // Variable not in cut
      const IntegerValue var_lb = integer_trail.LevelZeroLowerBound(var);
      ub -= coeff * var_lb;
    }
  }
  constraint_with_left_vars.ub = ub;
  constraint_with_left_vars.lb = constraint.lb;
  return constraint_with_left_vars;
}

bool ConstraintIsTriviallyTrue(const LinearConstraint& constraint,
                               const IntegerTrail& integer_trail) {
  IntegerValue term_sum = IntegerValue(0);
  for (int i = 0; i < constraint.vars.size(); ++i) {
    const IntegerVariable var = constraint.vars[i];
    const IntegerValue var_ub = integer_trail.LevelZeroUpperBound(var);
    const IntegerValue coeff = constraint.coeffs[i];
    term_sum += coeff * var_ub;
  }
  if (term_sum <= constraint.ub) {
    VLOG(2) << "Filtered by cover filter";
    return true;
  }
  return false;
}

bool CanBeFilteredUsingCutLowerBound(
    const LinearConstraint& preprocessed_constraint,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    const IntegerTrail& integer_trail) {
  std::vector<double> variable_upper_bound_distances;
  for (const IntegerVariable var : preprocessed_constraint.vars) {
    const IntegerValue var_ub = integer_trail.LevelZeroUpperBound(var);
    variable_upper_bound_distances.push_back(ToDouble(var_ub) - lp_values[var]);
  }
  // Compute the min cover size.
  const int smallest_cover_size =
      GetSmallestCoverSize(preprocessed_constraint, integer_trail);

  std::nth_element(
      variable_upper_bound_distances.begin(),
      variable_upper_bound_distances.begin() + smallest_cover_size - 1,
      variable_upper_bound_distances.end());
  double cut_lower_bound = 0.0;
  for (int i = 0; i < smallest_cover_size; ++i) {
    cut_lower_bound += variable_upper_bound_distances[i];
  }
  if (cut_lower_bound >= 1.0 - kMinCutViolation) {
    VLOG(2) << "Filtered by kappa heuristic";
    return true;
  }
  return false;
}

double GetKnapsackUpperBound(std::vector<KnapsackItem> items,
                             const double capacity) {
  // Sort items by value by weight ratio.
  std::sort(items.begin(), items.end(), std::greater<KnapsackItem>());
  double left_capacity = capacity;
  double profit = 0.0;
  for (const KnapsackItem item : items) {
    if (item.weight <= left_capacity) {
      profit += item.profit;
      left_capacity -= item.weight;
    } else {
      profit += (left_capacity / item.weight) * item.profit;
      break;
    }
  }
  return profit;
}

bool CanBeFilteredUsingKnapsackUpperBound(
    const LinearConstraint& constraint,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    const IntegerTrail& integer_trail) {
  std::vector<KnapsackItem> items;
  double capacity = -ToDouble(constraint.ub) - 1.0;
  double sum_variable_profit = 0;
  for (int i = 0; i < constraint.vars.size(); ++i) {
    const IntegerVariable var = constraint.vars[i];
    const IntegerValue var_ub = integer_trail.LevelZeroUpperBound(var);
    const IntegerValue var_lb = integer_trail.LevelZeroLowerBound(var);
    const IntegerValue coeff = constraint.coeffs[i];
    KnapsackItem item;
    item.profit = ToDouble(var_ub) - lp_values[var];
    item.weight = ToDouble(coeff * (var_ub - var_lb));
    items.push_back(item);
    capacity += ToDouble(coeff * var_ub);
    sum_variable_profit += item.profit;
  }

  // Return early if the required upper bound is negative since all the profits
  // are non negative.
  if (sum_variable_profit - 1.0 + kMinCutViolation < 0.0) return false;

  // Get the knapsack upper bound.
  const double knapsack_upper_bound =
      GetKnapsackUpperBound(std::move(items), capacity);
  if (knapsack_upper_bound < sum_variable_profit - 1.0 + kMinCutViolation) {
    VLOG(2) << "Filtered by knapsack upper bound";
    return true;
  }
  return false;
}

bool CanFormValidKnapsackCover(
    const LinearConstraint& preprocessed_constraint,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    const IntegerTrail& integer_trail) {
  if (ConstraintIsTriviallyTrue(preprocessed_constraint, integer_trail)) {
    return false;
  }
  if (CanBeFilteredUsingCutLowerBound(preprocessed_constraint, lp_values,
                                      integer_trail)) {
    return false;
  }
  if (CanBeFilteredUsingKnapsackUpperBound(preprocessed_constraint, lp_values,
                                           integer_trail)) {
    return false;
  }
  return true;
}

void ConvertToKnapsackForm(const LinearConstraint& constraint,
                           std::vector<LinearConstraint>* knapsack_constraints,
                           IntegerTrail* integer_trail) {
  // If all coefficient are the same, the generated knapsack cuts cannot be
  // stronger than the constraint itself. However, when we substitute variables
  // using the implication graph, this is not longer true. So we only skip
  // constraints with same coeff and no substitutions.
  if (SmallRangeAndAllCoefficientsMagnitudeAreTheSame(constraint,
                                                      integer_trail)) {
    return;
  }
  if (constraint.ub < kMaxIntegerValue) {
    LinearConstraint canonical_knapsack_form;

    // Negate the variables with negative coefficients.
    for (int i = 0; i < constraint.vars.size(); ++i) {
      const IntegerVariable var = constraint.vars[i];
      const IntegerValue coeff = constraint.coeffs[i];
      if (coeff > IntegerValue(0)) {
        canonical_knapsack_form.AddTerm(var, coeff);
      } else {
        canonical_knapsack_form.AddTerm(NegationOf(var), -coeff);
      }
    }
    canonical_knapsack_form.ub = constraint.ub;
    canonical_knapsack_form.lb = kMinIntegerValue;
    knapsack_constraints->push_back(canonical_knapsack_form);
  }

  if (constraint.lb > kMinIntegerValue) {
    LinearConstraint canonical_knapsack_form;

    // Negate the variables with positive coefficients.
    for (int i = 0; i < constraint.vars.size(); ++i) {
      const IntegerVariable var = constraint.vars[i];
      const IntegerValue coeff = constraint.coeffs[i];
      if (coeff > IntegerValue(0)) {
        canonical_knapsack_form.AddTerm(NegationOf(var), coeff);
      } else {
        canonical_knapsack_form.AddTerm(var, -coeff);
      }
    }
    canonical_knapsack_form.ub = -constraint.lb;
    canonical_knapsack_form.lb = kMinIntegerValue;
    knapsack_constraints->push_back(canonical_knapsack_form);
  }
}

// TODO(user): This is no longer used as we try to separate all cut with
// knapsack now, remove.
CutGenerator CreateKnapsackCoverCutGenerator(
    const std::vector<LinearConstraint>& base_constraints,
    const std::vector<IntegerVariable>& vars, Model* model) {
  CutGenerator result;
  result.vars = vars;

  IntegerTrail* integer_trail = model->GetOrCreate<IntegerTrail>();
  std::vector<LinearConstraint> knapsack_constraints;
  for (const LinearConstraint& constraint : base_constraints) {
    // There is often a lot of small linear base constraints and it doesn't seem
    // super useful to generate cuts for constraints of size 2. Any valid cut
    // of size 1 should be already infered by the propagation.
    //
    // TODO(user): The case of size 2 is a bit less clear. investigate more if
    // it is useful.
    if (constraint.vars.size() <= 2) continue;

    ConvertToKnapsackForm(constraint, &knapsack_constraints, integer_trail);
  }
  VLOG(1) << "#knapsack constraints: " << knapsack_constraints.size();

  // Note(user): for Knapsack cuts, it seems always advantageous to replace a
  // variable X by a TIGHT lower bound of the form "coeff * binary + lb". This
  // will not change "covers" but can only result in more violation by the
  // current LP solution.
  ImpliedBoundsProcessor implied_bounds_processor(
      vars, integer_trail, model->GetOrCreate<ImpliedBounds>());

  // TODO(user): do not add generator if there are no knapsack constraints.
  result.generate_cuts = [implied_bounds_processor, knapsack_constraints, vars,
                          model, integer_trail](
                             const absl::StrongVector<IntegerVariable, double>&
                                 lp_values,
                             LinearConstraintManager* manager) mutable {
    // TODO(user): When we use implied-bound substitution, we might still infer
    // an interesting cut even if all variables are integer. See if we still
    // want to skip all such constraints.
    if (AllVarsTakeIntegerValue(vars, lp_values)) return true;

    KnapsackSolverForCuts knapsack_solver(
        "Knapsack on demand cover cut generator");
    int64_t skipped_constraints = 0;
    LinearConstraint mutable_constraint;

    // Iterate through all knapsack constraints.
    implied_bounds_processor.RecomputeCacheAndSeparateSomeImpliedBoundCuts(
        lp_values);
    for (const LinearConstraint& constraint : knapsack_constraints) {
      if (model->GetOrCreate<TimeLimit>()->LimitReached()) break;
      VLOG(2) << "Processing constraint: " << constraint.DebugString();

      mutable_constraint = constraint;
      implied_bounds_processor.ProcessUpperBoundedConstraint(
          lp_values, &mutable_constraint);
      MakeAllCoefficientsPositive(&mutable_constraint);

      const LinearConstraint preprocessed_constraint =
          GetPreprocessedLinearConstraint(mutable_constraint, lp_values,
                                          *integer_trail);
      if (preprocessed_constraint.vars.empty()) continue;

      if (!CanFormValidKnapsackCover(preprocessed_constraint, lp_values,
                                     *integer_trail)) {
        skipped_constraints++;
        continue;
      }

      // Profits are (upper_bounds[i] - lp_values[i]) for knapsack variables.
      std::vector<double> profits;
      profits.reserve(preprocessed_constraint.vars.size());

      // Weights are (coeffs[i] * (upper_bound[i] - lower_bound[i])).
      std::vector<double> weights;
      weights.reserve(preprocessed_constraint.vars.size());

      double capacity = -ToDouble(preprocessed_constraint.ub) - 1.0;

      // Compute and store the sum of variable profits. This is the constant
      // part of the objective of the problem we are trying to solve. Hence
      // this part is not supplied to the knapsack_solver and is subtracted
      // when we receive the knapsack solution.
      double sum_variable_profit = 0;

      // Compute the profits, the weights and the capacity for the knapsack
      // instance.
      for (int i = 0; i < preprocessed_constraint.vars.size(); ++i) {
        const IntegerVariable var = preprocessed_constraint.vars[i];
        const double coefficient = ToDouble(preprocessed_constraint.coeffs[i]);
        const double var_ub = ToDouble(integer_trail->LevelZeroUpperBound(var));
        const double var_lb = ToDouble(integer_trail->LevelZeroLowerBound(var));
        const double variable_profit = var_ub - lp_values[var];
        profits.push_back(variable_profit);

        sum_variable_profit += variable_profit;

        const double weight = coefficient * (var_ub - var_lb);
        weights.push_back(weight);
        capacity += weight + coefficient * var_lb;
      }
      if (capacity < 0.0) continue;

      std::vector<IntegerVariable> cut_vars;
      std::vector<IntegerValue> cut_vars_original_coefficients;

      VLOG(2) << "Knapsack size: " << profits.size();
      knapsack_solver.Init(profits, weights, capacity);

      // Set the time limit for the knapsack solver.
      const double time_limit_for_knapsack_solver =
          model->GetOrCreate<TimeLimit>()->GetTimeLeft();

      // Solve the instance and subtract the constant part to compute the
      // sum_of_distance_to_ub_for_vars_in_cover.
      // TODO(user): Consider solving the instance approximately.
      bool is_solution_optimal = false;
      knapsack_solver.set_solution_upper_bound_threshold(
          sum_variable_profit - 1.0 + kMinCutViolation);
      // TODO(user): Consider providing lower bound threshold as
      // sum_variable_profit - 1.0 + kMinCutViolation.
      // TODO(user): Set node limit for knapsack solver.
      auto time_limit_for_solver =
          absl::make_unique<TimeLimit>(time_limit_for_knapsack_solver);
      const double sum_of_distance_to_ub_for_vars_in_cover =
          sum_variable_profit -
          knapsack_solver.Solve(time_limit_for_solver.get(),
                                &is_solution_optimal);
      if (is_solution_optimal) {
        VLOG(2) << "Knapsack Optimal solution found yay !";
      }
      if (time_limit_for_solver->LimitReached()) {
        VLOG(1) << "Knapsack Solver run out of time limit.";
      }
      if (sum_of_distance_to_ub_for_vars_in_cover < 1.0 - kMinCutViolation) {
        // Constraint is eligible for the cover.

        IntegerValue constraint_ub_for_cut = preprocessed_constraint.ub;
        std::set<IntegerVariable> vars_in_cut;
        for (int i = 0; i < preprocessed_constraint.vars.size(); ++i) {
          const IntegerVariable var = preprocessed_constraint.vars[i];
          const IntegerValue coefficient = preprocessed_constraint.coeffs[i];
          if (!knapsack_solver.best_solution(i)) {
            cut_vars.push_back(var);
            cut_vars_original_coefficients.push_back(coefficient);
            vars_in_cut.insert(var);
          } else {
            const IntegerValue var_lb = integer_trail->LevelZeroLowerBound(var);
            constraint_ub_for_cut -= coefficient * var_lb;
          }
        }
        LinearConstraint cut = GenerateKnapsackCutForCover(
            cut_vars, cut_vars_original_coefficients, constraint_ub_for_cut,
            *integer_trail);

        // Check if the constraint has only binary variables.
        bool is_lifted = false;
        if (ConstraintIsEligibleForLifting(cut, *integer_trail)) {
          if (LiftKnapsackCut(mutable_constraint, lp_values,
                              cut_vars_original_coefficients, *integer_trail,
                              model->GetOrCreate<TimeLimit>(), &cut)) {
            is_lifted = true;
          }
        }

        CHECK(!SolutionSatisfiesConstraint(cut, lp_values));
        manager->AddCut(cut, is_lifted ? "LiftedKnapsack" : "Knapsack",
                        lp_values);
      }
    }
    if (skipped_constraints > 0) {
      VLOG(2) << "Skipped constraints: " << skipped_constraints;
    }
    return true;
  };

  return result;
}

// Compute the larger t <= max_t such that t * rhs_remainder >= divisor / 2.
//
// This is just a separate function as it is slightly faster to compute the
// result only once.
IntegerValue GetFactorT(IntegerValue rhs_remainder, IntegerValue divisor,
                        IntegerValue max_t) {
  DCHECK_GE(max_t, 1);
  return rhs_remainder == 0
             ? max_t
             : std::min(max_t, CeilRatio(divisor / 2, rhs_remainder));
}

std::function<IntegerValue(IntegerValue)> GetSuperAdditiveRoundingFunction(
    IntegerValue rhs_remainder, IntegerValue divisor, IntegerValue t,
    IntegerValue max_scaling) {
  DCHECK_GE(max_scaling, 1);
  DCHECK_GE(t, 1);

  // Adjust after the multiplication by t.
  rhs_remainder *= t;
  DCHECK_LT(rhs_remainder, divisor);

  // Make sure we don't have an integer overflow below. Note that we assume that
  // divisor and the maximum coeff magnitude are not too different (maybe a
  // factor 1000 at most) so that the final result will never overflow.
  max_scaling =
      std::min(max_scaling, std::numeric_limits<int64_t>::max() / divisor);

  const IntegerValue size = divisor - rhs_remainder;
  if (max_scaling == 1 || size == 1) {
    // TODO(user): Use everywhere a two step computation to avoid overflow?
    // First divide by divisor, then multiply by t. For now, we limit t so that
    // we never have an overflow instead.
    return [t, divisor](IntegerValue coeff) {
      return FloorRatio(t * coeff, divisor);
    };
  } else if (size <= max_scaling) {
    return [size, rhs_remainder, t, divisor](IntegerValue coeff) {
      const IntegerValue t_coeff = t * coeff;
      const IntegerValue ratio = FloorRatio(t_coeff, divisor);
      const IntegerValue remainder = PositiveRemainder(t_coeff, divisor);
      const IntegerValue diff = remainder - rhs_remainder;
      return size * ratio + std::max(IntegerValue(0), diff);
    };
  } else if (max_scaling.value() * rhs_remainder.value() < divisor) {
    // Because of our max_t limitation, the rhs_remainder might stay small.
    //
    // If it is "too small" we cannot use the code below because it will not be
    // valid. So we just divide divisor into max_scaling bucket. The
    // rhs_remainder will be in the bucket 0.
    //
    // Note(user): This seems the same as just increasing t, modulo integer
    // overflows. Maybe we should just always do the computation like this so
    // that we can use larger t even if coeff is close to kint64max.
    return [t, divisor, max_scaling](IntegerValue coeff) {
      const IntegerValue t_coeff = t * coeff;
      const IntegerValue ratio = FloorRatio(t_coeff, divisor);
      const IntegerValue remainder = PositiveRemainder(t_coeff, divisor);
      const IntegerValue bucket = FloorRatio(remainder * max_scaling, divisor);
      return max_scaling * ratio + bucket;
    };
  } else {
    // We divide (size = divisor - rhs_remainder) into (max_scaling - 1) buckets
    // and increase the function by 1 / max_scaling for each of them.
    //
    // Note that for different values of max_scaling, we get a family of
    // functions that do not dominate each others. So potentially, a max scaling
    // as low as 2 could lead to the better cut (this is exactly the Letchford &
    // Lodi function).
    //
    // Another interesting fact, is that if we want to compute the maximum alpha
    // for a constraint with 2 terms like:
    //    divisor * Y + (ratio * divisor + remainder) * X
    //               <= rhs_ratio * divisor + rhs_remainder
    // so that we have the cut:
    //              Y + (ratio + alpha) * X  <= rhs_ratio
    // This is the same as computing the maximum alpha such that for all integer
    // X > 0 we have CeilRatio(alpha * divisor * X, divisor)
    //    <= CeilRatio(remainder * X - rhs_remainder, divisor).
    // We can prove that this alpha is of the form (n - 1) / n, and it will
    // be reached by such function for a max_scaling of n.
    //
    // TODO(user): This function is not always maximal when
    // size % (max_scaling - 1) == 0. Improve?
    return [size, rhs_remainder, t, divisor, max_scaling](IntegerValue coeff) {
      const IntegerValue t_coeff = t * coeff;
      const IntegerValue ratio = FloorRatio(t_coeff, divisor);
      const IntegerValue remainder = PositiveRemainder(t_coeff, divisor);
      const IntegerValue diff = remainder - rhs_remainder;
      const IntegerValue bucket =
          diff > 0 ? CeilRatio(diff * (max_scaling - 1), size)
                   : IntegerValue(0);
      return max_scaling * ratio + bucket;
    };
  }
}

// TODO(user): This has been optimized a bit, but we can probably do even better
// as it still takes around 25% percent of the run time when all the cuts are on
// for the opm*mps.gz problems and others.
void IntegerRoundingCutHelper::ComputeCut(
    RoundingOptions options, const std::vector<double>& lp_values,
    const std::vector<IntegerValue>& lower_bounds,
    const std::vector<IntegerValue>& upper_bounds,
    ImpliedBoundsProcessor* ib_processor, LinearConstraint* cut) {
  const int size = lp_values.size();
  if (size == 0) return;
  CHECK_EQ(lower_bounds.size(), size);
  CHECK_EQ(upper_bounds.size(), size);
  CHECK_EQ(cut->vars.size(), size);
  CHECK_EQ(cut->coeffs.size(), size);
  CHECK_EQ(cut->lb, kMinIntegerValue);

  // To optimize the computation of the best divisor below, we only need to
  // look at the indices with a shifted lp value that is not close to zero.
  //
  // TODO(user): sort by decreasing lp_values so that our early abort test in
  // the critical loop below has more chance of returning early? I tried but it
  // didn't seems to change much though.
  relevant_indices_.clear();
  relevant_lp_values_.clear();
  relevant_coeffs_.clear();
  relevant_bound_diffs_.clear();
  divisors_.clear();
  adjusted_coeffs_.clear();

  // Compute the maximum magnitude for non-fixed variables.
  IntegerValue max_magnitude(0);
  for (int i = 0; i < size; ++i) {
    if (lower_bounds[i] == upper_bounds[i]) continue;
    const IntegerValue magnitude = IntTypeAbs(cut->coeffs[i]);
    max_magnitude = std::max(max_magnitude, magnitude);
  }

  // Shift each variable using its lower/upper bound so that no variable can
  // change sign. We eventually do a change of variable to its negation so
  // that all variable are non-negative.
  bool overflow = false;
  change_sign_at_postprocessing_.assign(size, false);
  for (int i = 0; i < size; ++i) {
    if (cut->coeffs[i] == 0) continue;
    const IntegerValue magnitude = IntTypeAbs(cut->coeffs[i]);

    // We might change them below.
    IntegerValue lb = lower_bounds[i];
    double lp_value = lp_values[i];

    const IntegerValue ub = upper_bounds[i];
    const IntegerValue bound_diff =
        IntegerValue(CapSub(ub.value(), lb.value()));
    // Note that since we use ToDouble() this code works fine with lb/ub at
    // min/max integer value.
    //
    // TODO(user): Experiments with different heuristics. Other solver also
    // seems to try a bunch of possibilities in a "postprocess" phase once
    // the divisor is chosen. Try that.
    {
      // when the magnitude of the entry become smaller and smaller we bias
      // towards a positive coefficient. This is because after rounding this
      // will likely become zero instead of -divisor and we need the lp value
      // to be really close to its bound to compensate.
      const double lb_dist = std::abs(lp_value - ToDouble(lb));
      const double ub_dist = std::abs(lp_value - ToDouble(ub));
      const double bias =
          std::max(1.0, 0.1 * ToDouble(max_magnitude) / ToDouble(magnitude));
      if ((bias * lb_dist > ub_dist && cut->coeffs[i] < 0) ||
          (lb_dist > bias * ub_dist && cut->coeffs[i] > 0)) {
        change_sign_at_postprocessing_[i] = true;
        cut->coeffs[i] = -cut->coeffs[i];
        lp_value = -lp_value;
        lb = -ub;
      }
    }

    // Always shift to lb.
    // coeff * X = coeff * (X - shift) + coeff * shift.
    lp_value -= ToDouble(lb);
    if (!AddProductTo(-cut->coeffs[i], lb, &cut->ub)) {
      overflow = true;
      break;
    }

    // Deal with fixed variable, no need to shift back in this case, we can
    // just remove the term.
    if (bound_diff == 0) {
      cut->coeffs[i] = IntegerValue(0);
      lp_value = 0.0;
    }

    if (std::abs(lp_value) > 1e-2) {
      relevant_coeffs_.push_back(cut->coeffs[i]);
      relevant_indices_.push_back(i);
      relevant_lp_values_.push_back(lp_value);
      relevant_bound_diffs_.push_back(bound_diff);
      divisors_.push_back(magnitude);
    }
  }

  // TODO(user): Maybe this shouldn't be called on such constraint.
  if (relevant_coeffs_.empty()) {
    VLOG(2) << "Issue, nothing to cut.";
    *cut = LinearConstraint(IntegerValue(0), IntegerValue(0));
    return;
  }
  CHECK_NE(max_magnitude, 0);

  // Our heuristic will try to generate a few different cuts, and we will keep
  // the most violated one scaled by the l2 norm of the relevant position.
  //
  // TODO(user): Experiment for the best value of this initial violation
  // threshold. Note also that we use the l2 norm on the restricted position
  // here. Maybe we should change that? On that note, the L2 norm usage seems a
  // bit weird to me since it grows with the number of term in the cut. And
  // often, we already have a good cut, and we make it stronger by adding extra
  // terms that do not change its activity.
  //
  // The discussion above only concern the best_scaled_violation initial value.
  // The remainder_threshold allows to not consider cuts for which the final
  // efficacity is clearly lower than 1e-3 (it is a bound, so we could generate
  // cuts with a lower efficacity than this).
  double best_scaled_violation = 0.01;
  const IntegerValue remainder_threshold(max_magnitude / 1000);

  // The cut->ub might have grown quite a bit with the bound substitution, so
  // we need to include it too since we will apply the rounding function on it.
  max_magnitude = std::max(max_magnitude, IntTypeAbs(cut->ub));

  // Make sure that when we multiply the rhs or the coefficient by a factor t,
  // we do not have an integer overflow. Actually, we need a bit more room
  // because we might round down a value to the next multiple of
  // max_magnitude.
  const IntegerValue threshold = kMaxIntegerValue / 2;
  if (overflow || max_magnitude >= threshold) {
    VLOG(2) << "Issue, overflow.";
    *cut = LinearConstraint(IntegerValue(0), IntegerValue(0));
    return;
  }
  const IntegerValue max_t = threshold / max_magnitude;

  // There is no point trying twice the same divisor or a divisor that is too
  // small. Note that we use a higher threshold than the remainder_threshold
  // because we can boost the remainder thanks to our adjusting heuristic below
  // and also because this allows to have cuts with a small range of
  // coefficients.
  //
  // TODO(user): Note that the std::sort() is visible in some cpu profile.
  {
    int new_size = 0;
    const IntegerValue divisor_threshold = max_magnitude / 10;
    for (int i = 0; i < divisors_.size(); ++i) {
      if (divisors_[i] <= divisor_threshold) continue;
      divisors_[new_size++] = divisors_[i];
    }
    divisors_.resize(new_size);
  }
  gtl::STLSortAndRemoveDuplicates(&divisors_, std::greater<IntegerValue>());

  // TODO(user): Avoid quadratic algorithm? Note that we are quadratic in
  // relevant_indices not the full cut->coeffs.size(), but this is still too
  // much on some problems.
  IntegerValue best_divisor(0);
  for (const IntegerValue divisor : divisors_) {
    // Skip if we don't have the potential to generate a good enough cut.
    const IntegerValue initial_rhs_remainder =
        cut->ub - FloorRatio(cut->ub, divisor) * divisor;
    if (initial_rhs_remainder <= remainder_threshold) continue;

    IntegerValue temp_ub = cut->ub;
    adjusted_coeffs_.clear();

    // We will adjust coefficient that are just under an exact multiple of
    // divisor to an exact multiple. This is meant to get rid of small errors
    // that appears due to rounding error in our exact computation of the
    // initial constraint given to this class.
    //
    // Each adjustement will cause the initial_rhs_remainder to increase, and we
    // do not want to increase it above divisor. Our threshold below guarantees
    // this. Note that the higher the rhs_remainder becomes, the more the
    // function f() has a chance to reduce the violation, so it is not always a
    // good idea to use all the slack we have between initial_rhs_remainder and
    // divisor.
    //
    // TODO(user): If possible, it might be better to complement these
    // variables. Even if the adjusted lp_values end up larger, if we loose less
    // when taking f(), then we will have a better violation.
    const IntegerValue adjust_threshold =
        (divisor - initial_rhs_remainder - 1) / IntegerValue(size);
    if (adjust_threshold > 0) {
      // Even before we finish the adjust, we can have a lower bound on the
      // activily loss using this divisor, and so we can abort early. This is
      // similar to what is done below in the function.
      bool early_abort = false;
      double loss_lb = 0.0;
      const double threshold = ToDouble(initial_rhs_remainder);

      for (int i = 0; i < relevant_coeffs_.size(); ++i) {
        // Compute the difference of coeff with the next multiple of divisor.
        const IntegerValue coeff = relevant_coeffs_[i];
        const IntegerValue remainder =
            CeilRatio(coeff, divisor) * divisor - coeff;

        if (divisor - remainder <= initial_rhs_remainder) {
          // We do not know exactly f() yet, but it will always round to the
          // floor of the division by divisor in this case.
          loss_lb += ToDouble(divisor - remainder) * relevant_lp_values_[i];
          if (loss_lb >= threshold) {
            early_abort = true;
            break;
          }
        }

        // Adjust coeff of the form k * divisor - epsilon.
        const IntegerValue diff = relevant_bound_diffs_[i];
        if (remainder > 0 && remainder <= adjust_threshold &&
            CapProd(diff.value(), remainder.value()) <= adjust_threshold) {
          temp_ub += remainder * diff;
          adjusted_coeffs_.push_back({i, coeff + remainder});
        }
      }

      if (early_abort) continue;
    }

    // Create the super-additive function f().
    const IntegerValue rhs_remainder =
        temp_ub - FloorRatio(temp_ub, divisor) * divisor;
    if (rhs_remainder == 0) continue;

    const auto f = GetSuperAdditiveRoundingFunction(
        rhs_remainder, divisor, GetFactorT(rhs_remainder, divisor, max_t),
        options.max_scaling);

    // As we round coefficients, we will compute the loss compared to the
    // current scaled constraint activity. As soon as this loss crosses the
    // slack, then we known that there is no violation and we can abort early.
    //
    // TODO(user): modulo the scaling, we could compute the exact threshold
    // using our current best cut. Note that we also have to account the change
    // in slack due to the adjust code above.
    const double scaling = ToDouble(f(divisor)) / ToDouble(divisor);
    const double threshold = scaling * ToDouble(rhs_remainder);
    double loss = 0.0;

    // Apply f() to the cut and compute the cut violation. Note that it is
    // okay to just look at the relevant indices since the other have a lp
    // value which is almost zero. Doing it like this is faster, and even if
    // the max_magnitude might be off it should still be relevant enough.
    double violation = -ToDouble(f(temp_ub));
    double l2_norm = 0.0;
    bool early_abort = false;
    int adjusted_coeffs_index = 0;
    for (int i = 0; i < relevant_coeffs_.size(); ++i) {
      IntegerValue coeff = relevant_coeffs_[i];

      // Adjust coeff according to our previous computation if needed.
      if (adjusted_coeffs_index < adjusted_coeffs_.size() &&
          adjusted_coeffs_[adjusted_coeffs_index].first == i) {
        coeff = adjusted_coeffs_[adjusted_coeffs_index].second;
        adjusted_coeffs_index++;
      }

      if (coeff == 0) continue;
      const IntegerValue new_coeff = f(coeff);
      const double new_coeff_double = ToDouble(new_coeff);
      const double lp_value = relevant_lp_values_[i];

      l2_norm += new_coeff_double * new_coeff_double;
      violation += new_coeff_double * lp_value;
      loss += (scaling * ToDouble(coeff) - new_coeff_double) * lp_value;
      if (loss >= threshold) {
        early_abort = true;
        break;
      }
    }
    if (early_abort) continue;

    // Here we scale by the L2 norm over the "relevant" positions. This seems
    // to work slighly better in practice.
    violation /= sqrt(l2_norm);
    if (violation > best_scaled_violation) {
      best_scaled_violation = violation;
      best_divisor = divisor;
    }
  }

  if (best_divisor == 0) {
    *cut = LinearConstraint(IntegerValue(0), IntegerValue(0));
    return;
  }

  // Adjust coefficients.
  //
  // TODO(user): It might make sense to also adjust the one with a small LP
  // value, but then the cut will be slighlty different than the one we computed
  // above. Try with and without maybe?
  const IntegerValue initial_rhs_remainder =
      cut->ub - FloorRatio(cut->ub, best_divisor) * best_divisor;
  const IntegerValue adjust_threshold =
      (best_divisor - initial_rhs_remainder - 1) / IntegerValue(size);
  if (adjust_threshold > 0) {
    for (int i = 0; i < relevant_indices_.size(); ++i) {
      const int index = relevant_indices_[i];
      const IntegerValue diff = relevant_bound_diffs_[i];
      if (diff > adjust_threshold) continue;

      // Adjust coeff of the form k * best_divisor - epsilon.
      const IntegerValue coeff = cut->coeffs[index];
      const IntegerValue remainder =
          CeilRatio(coeff, best_divisor) * best_divisor - coeff;
      if (CapProd(diff.value(), remainder.value()) <= adjust_threshold) {
        cut->ub += remainder * diff;
        cut->coeffs[index] += remainder;
      }
    }
  }

  // Create the super-additive function f().
  //
  // TODO(user): Try out different rounding function and keep the best. We can
  // change max_t and max_scaling. It might not be easy to choose which cut is
  // the best, but we can at least know for sure if one dominate the other
  // completely. That is, if for all coeff f(coeff)/f(divisor) is greater than
  // or equal to the same value for another function f.
  const IntegerValue rhs_remainder =
      cut->ub - FloorRatio(cut->ub, best_divisor) * best_divisor;
  IntegerValue factor_t = GetFactorT(rhs_remainder, best_divisor, max_t);
  auto f = GetSuperAdditiveRoundingFunction(rhs_remainder, best_divisor,
                                            factor_t, options.max_scaling);

  // Look amongst all our possible function f() for one that dominate greedily
  // our current best one. Note that we prefer lower scaling factor since that
  // result in a cut with lower coefficients.
  remainders_.clear();
  for (int i = 0; i < size; ++i) {
    const IntegerValue coeff = cut->coeffs[i];
    const IntegerValue r =
        coeff - FloorRatio(coeff, best_divisor) * best_divisor;
    if (r > rhs_remainder) remainders_.push_back(r);
  }
  gtl::STLSortAndRemoveDuplicates(&remainders_);
  if (remainders_.size() <= 100) {
    best_rs_.clear();
    for (const IntegerValue r : remainders_) {
      best_rs_.push_back(f(r));
    }
    IntegerValue best_d = f(best_divisor);

    // Note that the complexity seems high 100 * 2 * options.max_scaling, but
    // this only run on cuts that are already efficient and the inner loop tend
    // to abort quickly. I didn't see this code in the cpu profile so far.
    for (const IntegerValue t :
         {IntegerValue(1), GetFactorT(rhs_remainder, best_divisor, max_t)}) {
      for (IntegerValue s(2); s <= options.max_scaling; ++s) {
        const auto g =
            GetSuperAdditiveRoundingFunction(rhs_remainder, best_divisor, t, s);
        int num_strictly_better = 0;
        rs_.clear();
        const IntegerValue d = g(best_divisor);
        for (int i = 0; i < best_rs_.size(); ++i) {
          const IntegerValue temp = g(remainders_[i]);
          if (temp * best_d < best_rs_[i] * d) break;
          if (temp * best_d > best_rs_[i] * d) num_strictly_better++;
          rs_.push_back(temp);
        }
        if (rs_.size() == best_rs_.size() && num_strictly_better > 0) {
          f = g;
          factor_t = t;
          best_rs_ = rs_;
          best_d = d;
        }
      }
    }
  }

  // Starts to apply f() to the cut. We only apply it to the rhs here, the
  // coefficient will be done after the potential lifting of some Booleans.
  cut->ub = f(cut->ub);
  tmp_terms_.clear();

  // Lift some implied bounds Booleans. Note that we will add them after
  // "size" so they will be ignored in the second loop below.
  num_lifted_booleans_ = 0;
  if (ib_processor != nullptr) {
    for (int i = 0; i < size; ++i) {
      const IntegerValue coeff = cut->coeffs[i];
      if (coeff == 0) continue;

      IntegerVariable var = cut->vars[i];
      if (change_sign_at_postprocessing_[i]) {
        var = NegationOf(var);
      }

      const ImpliedBoundsProcessor::BestImpliedBoundInfo info =
          ib_processor->GetCachedImpliedBoundInfo(var);

      // Avoid overflow.
      if (CapProd(CapProd(std::abs(coeff.value()), factor_t.value()),
                  info.bound_diff.value()) ==
          std::numeric_limits<int64_t>::max()) {
        continue;
      }

      // Because X = bound_diff * B + S
      // We can replace coeff * X by the expression before applying f:
      //   = f(coeff * bound_diff) * B + f(coeff) * [X - bound_diff * B]
      //   = f(coeff) * X + (f(coeff * bound_diff) - f(coeff) * bound_diff] B
      // So we can "lift" B into the cut.
      const IntegerValue coeff_b =
          f(coeff * info.bound_diff) - f(coeff) * info.bound_diff;
      CHECK_GE(coeff_b, 0);
      if (coeff_b == 0) continue;

      ++num_lifted_booleans_;
      if (info.is_positive) {
        tmp_terms_.push_back({info.bool_var, coeff_b});
      } else {
        tmp_terms_.push_back({info.bool_var, -coeff_b});
        cut->ub = CapAdd(-coeff_b.value(), cut->ub.value());
      }
    }
  }

  // Apply f() to the cut.
  //
  // Remove the bound shifts so the constraint is expressed in the original
  // variables.
  for (int i = 0; i < size; ++i) {
    IntegerValue coeff = cut->coeffs[i];
    if (coeff == 0) continue;
    coeff = f(coeff);
    if (coeff == 0) continue;
    if (change_sign_at_postprocessing_[i]) {
      cut->ub = IntegerValue(
          CapAdd((coeff * -upper_bounds[i]).value(), cut->ub.value()));
      tmp_terms_.push_back({cut->vars[i], -coeff});
    } else {
      cut->ub = IntegerValue(
          CapAdd((coeff * lower_bounds[i]).value(), cut->ub.value()));
      tmp_terms_.push_back({cut->vars[i], coeff});
    }
  }

  // Basic post-processing.
  CleanTermsAndFillConstraint(&tmp_terms_, cut);
  RemoveZeroTerms(cut);
  DivideByGCD(cut);
}

bool CoverCutHelper::TrySimpleKnapsack(
    const LinearConstraint base_ct, const std::vector<double>& lp_values,
    const std::vector<IntegerValue>& lower_bounds,
    const std::vector<IntegerValue>& upper_bounds) {
  const int base_size = lp_values.size();

  // Fill terms with a rewrite of the base constraint where all coeffs &
  // variables are positive by using either (X - LB) or (UB - X) as new
  // variables.
  terms_.clear();
  IntegerValue rhs = base_ct.ub;
  IntegerValue sum_of_diff(0);
  IntegerValue max_base_magnitude(0);
  for (int i = 0; i < base_size; ++i) {
    const IntegerValue coeff = base_ct.coeffs[i];
    const IntegerValue positive_coeff = IntTypeAbs(coeff);
    max_base_magnitude = std::max(max_base_magnitude, positive_coeff);
    const IntegerValue bound_diff = upper_bounds[i] - lower_bounds[i];
    if (!AddProductTo(positive_coeff, bound_diff, &sum_of_diff)) {
      return false;
    }
    const IntegerValue diff = positive_coeff * bound_diff;
    if (coeff > 0) {
      if (!AddProductTo(-coeff, lower_bounds[i], &rhs)) return false;
      terms_.push_back(
          {i, ToDouble(upper_bounds[i]) - lp_values[i], positive_coeff, diff});
    } else {
      if (!AddProductTo(-coeff, upper_bounds[i], &rhs)) return false;
      terms_.push_back(
          {i, lp_values[i] - ToDouble(lower_bounds[i]), positive_coeff, diff});
    }
  }

  // Try a simple cover heuristic.
  // Look for violated CUT of the form: sum (UB - X) or (X - LB) >= 1.
  double activity = 0.0;
  int new_size = 0;
  std::sort(terms_.begin(), terms_.end(), [](const Term& a, const Term& b) {
    if (a.dist_to_max_value == b.dist_to_max_value) {
      // Prefer low coefficients if the distance is the same.
      return a.positive_coeff < b.positive_coeff;
    }
    return a.dist_to_max_value < b.dist_to_max_value;
  });
  for (int i = 0; i < terms_.size(); ++i) {
    const Term& term = terms_[i];
    activity += term.dist_to_max_value;

    // As an heuristic we select all the term so that the sum of distance
    // to the upper bound is <= 1.0. If the corresponding rhs is negative, then
    // we will have a cut of violation at least 0.0. Note that this violation
    // can be improved by the lifting.
    //
    // TODO(user): experiment with different threshold (even greater than one).
    // Or come up with an algo that incorporate the lifting into the heuristic.
    if (activity > 1.0) {
      new_size = i;  // before this entry.
      break;
    }

    rhs -= term.diff;
  }

  // If the rhs is now negative, we have a cut.
  //
  // Note(user): past this point, now that a given "base" cover has been chosen,
  // we basically compute the cut (of the form sum X <= bound) with the maximum
  // possible violation. Note also that we lift as much as possible, so we don't
  // necessarily optimize for the cut efficacity though. But we do get a
  // stronger cut.
  if (rhs >= 0) return false;
  if (new_size == 0) return false;

  // Transform to a minimal cover. We want to greedily remove the largest coeff
  // first, so we have more chance for the "lifting" below which can increase
  // the cut violation. If the coeff are the same, we prefer to remove high
  // distance from upper bound first.
  //
  // We compute the cut at the same time.
  terms_.resize(new_size);
  std::sort(terms_.begin(), terms_.end(), [](const Term& a, const Term& b) {
    if (a.positive_coeff == b.positive_coeff) {
      return a.dist_to_max_value > b.dist_to_max_value;
    }
    return a.positive_coeff > b.positive_coeff;
  });
  in_cut_.assign(base_ct.vars.size(), false);
  cut_.ClearTerms();
  cut_.lb = kMinIntegerValue;
  cut_.ub = IntegerValue(-1);
  IntegerValue max_coeff(0);
  for (const Term term : terms_) {
    if (term.diff + rhs < 0) {
      rhs += term.diff;
      continue;
    }
    in_cut_[term.index] = true;
    max_coeff = std::max(max_coeff, term.positive_coeff);
    cut_.vars.push_back(base_ct.vars[term.index]);
    if (base_ct.coeffs[term.index] > 0) {
      cut_.coeffs.push_back(IntegerValue(1));
      cut_.ub += upper_bounds[term.index];
    } else {
      cut_.coeffs.push_back(IntegerValue(-1));
      cut_.ub -= lower_bounds[term.index];
    }
  }

  // In case the max_coeff variable is not binary, it might be possible to
  // tighten the cut a bit more.
  //
  // Note(user): I never observed this on the miplib so far.
  if (max_coeff == 0) return true;
  if (max_coeff < -rhs) {
    const IntegerValue m = FloorRatio(-rhs - 1, max_coeff);
    rhs += max_coeff * m;
    cut_.ub -= m;
  }
  CHECK_LT(rhs, 0);

  // Lift all at once the variables not used in the cover.
  //
  // We have a cut of the form sum_i X_i <= b that we will lift into
  // sum_i scaling X_i + sum f(base_coeff_j) X_j <= b * scaling.
  //
  // Using the super additivity of f() and how we construct it,
  // we know that: sum_j base_coeff_j X_j <= N * max_coeff + (max_coeff - slack)
  // implies that: sum_j f(base_coeff_j) X_j <= N * scaling.
  //
  // 1/ cut > b -(N+1)  => original sum + (N+1) * max_coeff >= rhs + slack
  // 2/ rewrite 1/ as : scaling * cut >= scaling * b - scaling * N => ...
  // 3/ lift > N * scaling => lift_sum > N * max_coeff + (max_coeff - slack)
  // And adding 2/ + 3/ we prove what we want:
  // cut * scaling + lift > b * scaling => original_sum + lift_sum > rhs.
  const IntegerValue slack = -rhs;
  const IntegerValue remainder = max_coeff - slack;
  max_base_magnitude = std::max(max_base_magnitude, IntTypeAbs(cut_.ub));
  const IntegerValue max_scaling(std::min(
      IntegerValue(60), FloorRatio(kMaxIntegerValue, max_base_magnitude)));
  const auto f = GetSuperAdditiveRoundingFunction(remainder, max_coeff,
                                                  IntegerValue(1), max_scaling);

  const IntegerValue scaling = f(max_coeff);
  if (scaling > 1) {
    for (int i = 0; i < cut_.coeffs.size(); ++i) cut_.coeffs[i] *= scaling;
    cut_.ub *= scaling;
  }

  num_lifting_ = 0;
  for (int i = 0; i < base_size; ++i) {
    if (in_cut_[i]) continue;
    const IntegerValue positive_coeff = IntTypeAbs(base_ct.coeffs[i]);
    const IntegerValue new_coeff = f(positive_coeff);
    if (new_coeff == 0) continue;

    ++num_lifting_;
    if (base_ct.coeffs[i] > 0) {
      // Add new_coeff * (X - LB)
      cut_.coeffs.push_back(new_coeff);
      cut_.vars.push_back(base_ct.vars[i]);
      cut_.ub += lower_bounds[i] * new_coeff;
    } else {
      // Add new_coeff * (UB - X)
      cut_.coeffs.push_back(-new_coeff);
      cut_.vars.push_back(base_ct.vars[i]);
      cut_.ub -= upper_bounds[i] * new_coeff;
    }
  }

  if (scaling > 1) DivideByGCD(&cut_);
  return true;
}

CutGenerator CreatePositiveMultiplicationCutGenerator(IntegerVariable z,
                                                      IntegerVariable x,
                                                      IntegerVariable y,
                                                      int linearization_level,
                                                      Model* model) {
  CutGenerator result;
  result.vars = {z, x, y};

  IntegerTrail* const integer_trail = model->GetOrCreate<IntegerTrail>();
  Trail* trail = model->GetOrCreate<Trail>();

  result.generate_cuts =
      [z, x, y, linearization_level, model, trail, integer_trail](
          const absl::StrongVector<IntegerVariable, double>& lp_values,
          LinearConstraintManager* manager) {
        if (trail->CurrentDecisionLevel() > 0 && linearization_level == 1) {
          return true;
        }
        const int64_t x_lb = integer_trail->LevelZeroLowerBound(x).value();
        const int64_t x_ub = integer_trail->LevelZeroUpperBound(x).value();
        const int64_t y_lb = integer_trail->LevelZeroLowerBound(y).value();
        const int64_t y_ub = integer_trail->LevelZeroUpperBound(y).value();

        // TODO(user): Compute a better bound (int_max / 4 ?).
        const int64_t kMaxSafeInteger = (int64_t{1} << 53) - 1;

        if (CapProd(x_ub, y_ub) >= kMaxSafeInteger) {
          VLOG(3) << "Potential overflow in PositiveMultiplicationCutGenerator";
          return true;
        }

        const double x_lp_value = lp_values[x];
        const double y_lp_value = lp_values[y];
        const double z_lp_value = lp_values[z];

        // TODO(user): As the bounds change monotonically, these cuts
        // dominate any previous one.  try to keep a reference to the cut and
        // replace it. Alternatively, add an API for a level-zero bound change
        // callback.

        // Cut -z + x_coeff * x + y_coeff* y <= rhs
        auto try_add_above_cut =
            [manager, z_lp_value, x_lp_value, y_lp_value, x, y, z, model,
             &lp_values](int64_t x_coeff, int64_t y_coeff, int64_t rhs) {
              if (-z_lp_value + x_lp_value * x_coeff + y_lp_value * y_coeff >=
                  rhs + kMinCutViolation) {
                LinearConstraintBuilder cut(model, /*lb=*/kMinIntegerValue,
                                            /*ub=*/IntegerValue(rhs));
                cut.AddTerm(z, IntegerValue(-1));
                if (x_coeff != 0) cut.AddTerm(x, IntegerValue(x_coeff));
                if (y_coeff != 0) cut.AddTerm(y, IntegerValue(y_coeff));
                manager->AddCut(cut.Build(), "PositiveProduct", lp_values);
              }
            };

        // Cut -z + x_coeff * x + y_coeff* y >= rhs
        auto try_add_below_cut =
            [manager, z_lp_value, x_lp_value, y_lp_value, x, y, z, model,
             &lp_values](int64_t x_coeff, int64_t y_coeff, int64_t rhs) {
              if (-z_lp_value + x_lp_value * x_coeff + y_lp_value * y_coeff <=
                  rhs - kMinCutViolation) {
                LinearConstraintBuilder cut(model, /*lb=*/IntegerValue(rhs),
                                            /*ub=*/kMaxIntegerValue);
                cut.AddTerm(z, IntegerValue(-1));
                if (x_coeff != 0) cut.AddTerm(x, IntegerValue(x_coeff));
                if (y_coeff != 0) cut.AddTerm(y, IntegerValue(y_coeff));
                manager->AddCut(cut.Build(), "PositiveProduct", lp_values);
              }
            };

        // McCormick relaxation of bilinear constraints. These 4 cuts are the
        // exact facets of the x * y polyhedron for a bounded x and y.
        //
        // Each cut correspond to plane that contains two of the line
        // (x=x_lb), (x=x_ub), (y=y_lb), (y=y_ub). The easiest to
        // understand them is to draw the x*y curves and see the 4
        // planes that correspond to the convex hull of the graph.
        try_add_above_cut(y_lb, x_lb, x_lb * y_lb);
        try_add_above_cut(y_ub, x_ub, x_ub * y_ub);
        try_add_below_cut(y_ub, x_lb, x_lb * y_ub);
        try_add_below_cut(y_lb, x_ub, x_ub * y_lb);
        return true;
      };

  return result;
}

CutGenerator CreateSquareCutGenerator(IntegerVariable y, IntegerVariable x,
                                      int linearization_level, Model* model) {
  CutGenerator result;
  result.vars = {y, x};

  Trail* trail = model->GetOrCreate<Trail>();
  IntegerTrail* integer_trail = model->GetOrCreate<IntegerTrail>();
  result.generate_cuts =
      [y, x, linearization_level, trail, integer_trail](
          const absl::StrongVector<IntegerVariable, double>& lp_values,
          LinearConstraintManager* manager) {
        if (trail->CurrentDecisionLevel() > 0 && linearization_level == 1) {
          return true;
        }
        const int64_t x_ub = integer_trail->LevelZeroUpperBound(x).value();
        const int64_t x_lb = integer_trail->LevelZeroLowerBound(x).value();

        if (x_lb == x_ub) return true;

        // Check for potential overflows.
        if (x_ub > (int64_t{1} << 31)) return true;
        DCHECK_GE(x_lb, 0);

        const double y_lp_value = lp_values[y];
        const double x_lp_value = lp_values[x];

        // First cut: target should be below the line:
        //     (x_lb, x_lb ^ 2) to (x_ub, x_ub ^ 2).
        // The slope of that line is (ub^2 - lb^2) / (ub - lb) = ub + lb.
        const int64_t y_lb = x_lb * x_lb;
        const int64_t above_slope = x_ub + x_lb;
        const double max_lp_y = y_lb + above_slope * (x_lp_value - x_lb);
        if (y_lp_value >= max_lp_y + kMinCutViolation) {
          // cut: y <= (x_lb + x_ub) * x - x_lb * x_ub
          LinearConstraint above_cut;
          above_cut.vars.push_back(y);
          above_cut.coeffs.push_back(IntegerValue(1));
          above_cut.vars.push_back(x);
          above_cut.coeffs.push_back(IntegerValue(-above_slope));
          above_cut.lb = kMinIntegerValue;
          above_cut.ub = IntegerValue(-x_lb * x_ub);
          manager->AddCut(above_cut, "SquareUpper", lp_values);
        }

        // Second cut: target should be above all the lines
        //     (value, value ^ 2) to (value + 1, (value + 1) ^ 2)
        // The slope of that line is 2 * value + 1
        //
        // Note that we only add one of these cuts. The one for x_lp_value in
        // [value, value + 1].
        const int64_t x_floor = static_cast<int64_t>(std::floor(x_lp_value));
        const int64_t below_slope = 2 * x_floor + 1;
        const double min_lp_y =
            below_slope * x_lp_value - x_floor - x_floor * x_floor;
        if (min_lp_y >= y_lp_value + kMinCutViolation) {
          // cut: y >= below_slope * (x - x_floor) + x_floor ^ 2
          //    : y >= below_slope * x - x_floor ^ 2 - x_floor
          LinearConstraint below_cut;
          below_cut.vars.push_back(y);
          below_cut.coeffs.push_back(IntegerValue(1));
          below_cut.vars.push_back(x);
          below_cut.coeffs.push_back(-IntegerValue(below_slope));
          below_cut.lb = IntegerValue(-x_floor - x_floor * x_floor);
          below_cut.ub = kMaxIntegerValue;
          manager->AddCut(below_cut, "SquareLower", lp_values);
        }
        return true;
      };

  return result;
}

void ImpliedBoundsProcessor::ProcessUpperBoundedConstraint(
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    LinearConstraint* cut) {
  ProcessUpperBoundedConstraintWithSlackCreation(
      /*substitute_only_inner_variables=*/false, IntegerVariable(0), lp_values,
      cut, nullptr);
}

ImpliedBoundsProcessor::BestImpliedBoundInfo
ImpliedBoundsProcessor::GetCachedImpliedBoundInfo(IntegerVariable var) {
  auto it = cache_.find(var);
  if (it != cache_.end()) return it->second;
  return BestImpliedBoundInfo();
}

ImpliedBoundsProcessor::BestImpliedBoundInfo
ImpliedBoundsProcessor::ComputeBestImpliedBound(
    IntegerVariable var,
    const absl::StrongVector<IntegerVariable, double>& lp_values) {
  auto it = cache_.find(var);
  if (it != cache_.end()) return it->second;
  BestImpliedBoundInfo result;
  const IntegerValue lb = integer_trail_->LevelZeroLowerBound(var);
  for (const ImpliedBoundEntry& entry :
       implied_bounds_->GetImpliedBounds(var)) {
    // Only process entries with a Boolean variable currently part of the LP
    // we are considering for this cut.
    //
    // TODO(user): the more we use cuts, the less it make sense to have a
    // lot of small independent LPs.
    if (!lp_vars_.contains(PositiveVariable(entry.literal_view))) {
      continue;
    }

    // The equation is X = lb + diff * Bool + Slack where Bool is in [0, 1]
    // and slack in [0, ub - lb].
    const IntegerValue diff = entry.lower_bound - lb;
    CHECK_GE(diff, 0);
    const double bool_lp_value = entry.is_positive
                                     ? lp_values[entry.literal_view]
                                     : 1.0 - lp_values[entry.literal_view];
    const double slack_lp_value =
        lp_values[var] - ToDouble(lb) - bool_lp_value * ToDouble(diff);

    // If the implied bound equation is not respected, we just add it
    // to implied_bound_cuts, and skip the entry for now.
    if (slack_lp_value < -1e-4) {
      LinearConstraint ib_cut;
      ib_cut.lb = kMinIntegerValue;
      std::vector<std::pair<IntegerVariable, IntegerValue>> terms;
      if (entry.is_positive) {
        // X >= Indicator * (bound - lb) + lb
        terms.push_back({entry.literal_view, diff});
        terms.push_back({var, IntegerValue(-1)});
        ib_cut.ub = -lb;
      } else {
        // X >= -Indicator * (bound - lb) + bound
        terms.push_back({entry.literal_view, -diff});
        terms.push_back({var, IntegerValue(-1)});
        ib_cut.ub = -entry.lower_bound;
      }
      CleanTermsAndFillConstraint(&terms, &ib_cut);
      ib_cut_pool_.AddCut(std::move(ib_cut), "IB", lp_values);
      continue;
    }

    // We look for tight implied bounds, and amongst the tightest one, we
    // prefer larger coefficient in front of the Boolean.
    if (slack_lp_value + 1e-4 < result.slack_lp_value ||
        (slack_lp_value < result.slack_lp_value + 1e-4 &&
         diff > result.bound_diff)) {
      result.bool_lp_value = bool_lp_value;
      result.slack_lp_value = slack_lp_value;

      result.bound_diff = diff;
      result.is_positive = entry.is_positive;
      result.bool_var = entry.literal_view;
    }
  }
  cache_[var] = result;
  return result;
}

void ImpliedBoundsProcessor::RecomputeCacheAndSeparateSomeImpliedBoundCuts(
    const absl::StrongVector<IntegerVariable, double>& lp_values) {
  cache_.clear();
  for (const IntegerVariable var :
       implied_bounds_->VariablesWithImpliedBounds()) {
    if (!lp_vars_.contains(PositiveVariable(var))) continue;
    ComputeBestImpliedBound(var, lp_values);
  }
}

void ImpliedBoundsProcessor::ProcessUpperBoundedConstraintWithSlackCreation(
    bool substitute_only_inner_variables, IntegerVariable first_slack,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    LinearConstraint* cut, std::vector<SlackInfo>* slack_infos) {
  if (cache_.empty()) return;  // Nothing to do.
  tmp_terms_.clear();
  IntegerValue new_ub = cut->ub;
  bool changed = false;

  // TODO(user): we could relax a bit this test.
  int64_t overflow_detection = 0;

  const int size = cut->vars.size();
  for (int i = 0; i < size; ++i) {
    IntegerVariable var = cut->vars[i];
    IntegerValue coeff = cut->coeffs[i];

    // Starts by positive coefficient.
    // TODO(user): Not clear this is best.
    if (coeff < 0) {
      coeff = -coeff;
      var = NegationOf(var);
    }

    // Find the best implied bound to use.
    // TODO(user): We could also use implied upper bound, that is try with
    // NegationOf(var).
    const BestImpliedBoundInfo info = GetCachedImpliedBoundInfo(var);
    const int old_size = tmp_terms_.size();

    // Shall we keep the original term ?
    bool keep_term = false;
    if (info.bool_var == kNoIntegerVariable) keep_term = true;
    if (CapProd(std::abs(coeff.value()), info.bound_diff.value()) ==
        std::numeric_limits<int64_t>::max()) {
      keep_term = true;
    }

    // TODO(user): On some problem, not replacing the variable at their bound
    // by an implied bounds seems beneficial. This is especially the case on
    // g200x740.mps.gz
    //
    // Note that in ComputeCut() the variable with an LP value at the bound do
    // not contribute to the cut efficacity (no loss) but do contribute to the
    // various heuristic based on the coefficient magnitude.
    if (substitute_only_inner_variables) {
      const IntegerValue lb = integer_trail_->LevelZeroLowerBound(var);
      const IntegerValue ub = integer_trail_->LevelZeroUpperBound(var);
      if (lp_values[var] - ToDouble(lb) < 1e-2) keep_term = true;
      if (ToDouble(ub) - lp_values[var] < 1e-2) keep_term = true;
    }

    // This is when we do not add slack.
    if (slack_infos == nullptr) {
      // We do not want to loose anything, so we only replace if the slack lp is
      // zero.
      if (info.slack_lp_value > 1e-6) keep_term = true;
    }

    if (keep_term) {
      tmp_terms_.push_back({var, coeff});
    } else {
      // Substitute.
      const IntegerValue lb = integer_trail_->LevelZeroLowerBound(var);
      const IntegerValue ub = integer_trail_->LevelZeroUpperBound(var);

      SlackInfo slack_info;
      slack_info.lp_value = info.slack_lp_value;
      slack_info.lb = 0;
      slack_info.ub = ub - lb;

      if (info.is_positive) {
        // X = Indicator * diff + lb + Slack
        tmp_terms_.push_back({info.bool_var, coeff * info.bound_diff});
        if (!AddProductTo(-coeff, lb, &new_ub)) {
          VLOG(2) << "Overflow";
          return;
        }
        if (slack_infos != nullptr) {
          tmp_terms_.push_back({first_slack, coeff});
          first_slack += 2;

          // slack = X - Indicator * info.bound_diff - lb;
          slack_info.terms.push_back({var, IntegerValue(1)});
          slack_info.terms.push_back({info.bool_var, -info.bound_diff});
          slack_info.offset = -lb;
          slack_infos->push_back(slack_info);
        }
      } else {
        // X = (1 - Indicator) * (diff) + lb + Slack
        // X = -Indicator * (diff) + lb + diff + Slack
        tmp_terms_.push_back({info.bool_var, -coeff * info.bound_diff});
        if (!AddProductTo(-coeff, lb + info.bound_diff, &new_ub)) {
          VLOG(2) << "Overflow";
          return;
        }
        if (slack_infos != nullptr) {
          tmp_terms_.push_back({first_slack, coeff});
          first_slack += 2;

          // slack = X + Indicator * info.bound_diff - lb - diff;
          slack_info.terms.push_back({var, IntegerValue(1)});
          slack_info.terms.push_back({info.bool_var, +info.bound_diff});
          slack_info.offset = -lb - info.bound_diff;
          slack_infos->push_back(slack_info);
        }
      }
      changed = true;
    }

    // Add all the new terms coefficient to the overflow detection to avoid
    // issue when merging terms referring to the same variable.
    for (int i = old_size; i < tmp_terms_.size(); ++i) {
      overflow_detection =
          CapAdd(overflow_detection, std::abs(tmp_terms_[i].second.value()));
    }
  }

  if (overflow_detection >= kMaxIntegerValue) {
    VLOG(2) << "Overflow";
    return;
  }
  if (!changed) return;

  // Update the cut.
  //
  // Note that because of our overflow_detection variable, there should be
  // no integer overflow when we merge identical terms.
  cut->lb = kMinIntegerValue;  // Not relevant.
  cut->ub = new_ub;
  CleanTermsAndFillConstraint(&tmp_terms_, cut);
}

bool ImpliedBoundsProcessor::DebugSlack(IntegerVariable first_slack,
                                        const LinearConstraint& initial_cut,
                                        const LinearConstraint& cut,
                                        const std::vector<SlackInfo>& info) {
  tmp_terms_.clear();
  IntegerValue new_ub = cut.ub;
  for (int i = 0; i < cut.vars.size(); ++i) {
    // Simple copy for non-slack variables.
    if (cut.vars[i] < first_slack) {
      tmp_terms_.push_back({cut.vars[i], cut.coeffs[i]});
      continue;
    }

    // Replace slack by its definition.
    const IntegerValue multiplier = cut.coeffs[i];
    const int index = (cut.vars[i].value() - first_slack.value()) / 2;
    for (const std::pair<IntegerVariable, IntegerValue>& term :
         info[index].terms) {
      tmp_terms_.push_back({term.first, term.second * multiplier});
    }
    new_ub -= multiplier * info[index].offset;
  }

  LinearConstraint tmp_cut;
  tmp_cut.lb = kMinIntegerValue;  // Not relevant.
  tmp_cut.ub = new_ub;
  CleanTermsAndFillConstraint(&tmp_terms_, &tmp_cut);
  MakeAllVariablesPositive(&tmp_cut);

  // We need to canonicalize the initial_cut too for comparison. Note that we
  // only use this for debug, so we don't care too much about the memory and
  // extra time.
  // TODO(user): Expose CanonicalizeConstraint() from the manager.
  LinearConstraint tmp_copy;
  tmp_terms_.clear();
  for (int i = 0; i < initial_cut.vars.size(); ++i) {
    tmp_terms_.push_back({initial_cut.vars[i], initial_cut.coeffs[i]});
  }
  tmp_copy.lb = kMinIntegerValue;  // Not relevant.
  tmp_copy.ub = new_ub;
  CleanTermsAndFillConstraint(&tmp_terms_, &tmp_copy);
  MakeAllVariablesPositive(&tmp_copy);

  if (tmp_cut == tmp_copy) return true;

  LOG(INFO) << first_slack;
  LOG(INFO) << tmp_copy.DebugString();
  LOG(INFO) << cut.DebugString();
  LOG(INFO) << tmp_cut.DebugString();
  return false;
}

namespace {

void TryToGenerateAllDiffCut(
    const std::vector<std::pair<double, IntegerVariable>>& sorted_vars_lp,
    const IntegerTrail& integer_trail,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    LinearConstraintManager* manager) {
  Domain current_union;
  std::vector<IntegerVariable> current_set_vars;
  double sum = 0.0;
  for (auto value_var : sorted_vars_lp) {
    sum += value_var.first;
    const IntegerVariable var = value_var.second;
    // TODO(user): The union of the domain of the variable being considered
    // does not give the tightest bounds, try to get better bounds.
    current_union =
        current_union.UnionWith(integer_trail.InitialVariableDomain(var));
    current_set_vars.push_back(var);
    const int64_t required_min_sum =
        SumOfKMinValueInDomain(current_union, current_set_vars.size());
    const int64_t required_max_sum =
        SumOfKMaxValueInDomain(current_union, current_set_vars.size());
    if (sum < required_min_sum || sum > required_max_sum) {
      LinearConstraint cut;
      for (IntegerVariable var : current_set_vars) {
        cut.AddTerm(var, IntegerValue(1));
      }
      cut.lb = IntegerValue(required_min_sum);
      cut.ub = IntegerValue(required_max_sum);
      manager->AddCut(cut, "all_diff", lp_values);
      // NOTE: We can extend the current set but it is more helpful to generate
      // the cut on a different set of variables so we reset the counters.
      sum = 0.0;
      current_set_vars.clear();
      current_union = Domain();
    }
  }
}

}  // namespace

CutGenerator CreateAllDifferentCutGenerator(
    const std::vector<IntegerVariable>& vars, Model* model) {
  CutGenerator result;
  result.vars = vars;
  IntegerTrail* integer_trail = model->GetOrCreate<IntegerTrail>();
  Trail* trail = model->GetOrCreate<Trail>();
  result.generate_cuts =
      [vars, integer_trail, trail](
          const absl::StrongVector<IntegerVariable, double>& lp_values,
          LinearConstraintManager* manager) {
        // These cuts work at all levels but the generator adds too many cuts on
        // some instances and degrade the performance so we only use it at level
        // 0.
        if (trail->CurrentDecisionLevel() > 0) return true;
        std::vector<std::pair<double, IntegerVariable>> sorted_vars;
        for (const IntegerVariable var : vars) {
          if (integer_trail->LevelZeroLowerBound(var) ==
              integer_trail->LevelZeroUpperBound(var)) {
            continue;
          }
          sorted_vars.push_back(std::make_pair(lp_values[var], var));
        }
        std::sort(sorted_vars.begin(), sorted_vars.end());
        TryToGenerateAllDiffCut(sorted_vars, *integer_trail, lp_values,
                                manager);
        // Other direction.
        std::reverse(sorted_vars.begin(), sorted_vars.end());
        TryToGenerateAllDiffCut(sorted_vars, *integer_trail, lp_values,
                                manager);
        return true;
      };
  VLOG(1) << "Created all_diff cut generator of size: " << vars.size();
  return result;
}

namespace {
// Returns max((w2i - w1i)*Li, (w2i - w1i)*Ui).
IntegerValue MaxCornerDifference(const IntegerVariable var,
                                 const IntegerValue w1_i,
                                 const IntegerValue w2_i,
                                 const IntegerTrail& integer_trail) {
  const IntegerValue lb = integer_trail.LevelZeroLowerBound(var);
  const IntegerValue ub = integer_trail.LevelZeroUpperBound(var);
  return std::max((w2_i - w1_i) * lb, (w2_i - w1_i) * ub);
}

// This is the coefficient of zk in the cut, where k = max_index.
// MPlusCoefficient_ki = max((wki - wI(i)i) * Li,
//                           (wki - wI(i)i) * Ui)
//                     = max corner difference for variable i,
//                       target expr I(i), max expr k.
// The coefficient of zk is Sum(i=1..n)(MPlusCoefficient_ki) + bk
IntegerValue MPlusCoefficient(
    const std::vector<IntegerVariable>& x_vars,
    const std::vector<LinearExpression>& exprs,
    const absl::StrongVector<IntegerVariable, int>& variable_partition,
    const int max_index, const IntegerTrail& integer_trail) {
  IntegerValue coeff = exprs[max_index].offset;
  // TODO(user): This algo is quadratic since GetCoefficientOfPositiveVar()
  // is linear. This can be optimized (better complexity) if needed.
  for (const IntegerVariable var : x_vars) {
    const int target_index = variable_partition[var];
    if (max_index != target_index) {
      coeff += MaxCornerDifference(
          var, GetCoefficientOfPositiveVar(var, exprs[target_index]),
          GetCoefficientOfPositiveVar(var, exprs[max_index]), integer_trail);
    }
  }
  return coeff;
}

// Compute the value of
// rhs = wI(i)i * xi + Sum(k=1..d)(MPlusCoefficient_ki * zk)
// for variable xi for given target index I(i).
double ComputeContribution(
    const IntegerVariable xi_var, const std::vector<IntegerVariable>& z_vars,
    const std::vector<LinearExpression>& exprs,
    const absl::StrongVector<IntegerVariable, double>& lp_values,
    const IntegerTrail& integer_trail, const int target_index) {
  CHECK_GE(target_index, 0);
  CHECK_LT(target_index, exprs.size());
  const LinearExpression& target_expr = exprs[target_index];
  const double xi_value = lp_values[xi_var];
  const IntegerValue wt_i = GetCoefficientOfPositiveVar(xi_var, target_expr);
  double contrib = ToDouble(wt_i) * xi_value;
  for (int expr_index = 0; expr_index < exprs.size(); ++expr_index) {
    if (expr_index == target_index) continue;
    const LinearExpression& max_expr = exprs[expr_index];
    const double z_max_value = lp_values[z_vars[expr_index]];
    const IntegerValue corner_value = MaxCornerDifference(
        xi_var, wt_i, GetCoefficientOfPositiveVar(xi_var, max_expr),
        integer_trail);
    contrib += ToDouble(corner_value) * z_max_value;
  }
  return contrib;
}
}  // namespace

CutGenerator CreateLinMaxCutGenerator(
    const IntegerVariable target, const std::vector<LinearExpression>& exprs,
    const std::vector<IntegerVariable>& z_vars, Model* model) {
  CutGenerator result;
  std::vector<IntegerVariable> x_vars;
  result.vars = {target};
  const int num_exprs = exprs.size();
  for (int i = 0; i < num_exprs; ++i) {
    result.vars.push_back(z_vars[i]);
    x_vars.insert(x_vars.end(), exprs[i].vars.begin(), exprs[i].vars.end());
  }
  gtl::STLSortAndRemoveDuplicates(&x_vars);
  // All expressions should only contain positive variables.
  DCHECK(std::all_of(x_vars.begin(), x_vars.end(), [](IntegerVariable var) {
    return VariableIsPositive(var);
  }));
  result.vars.insert(result.vars.end(), x_vars.begin(), x_vars.end());

  IntegerTrail* integer_trail = model->GetOrCreate<IntegerTrail>();
  result.generate_cuts =
      [x_vars, z_vars, target, num_exprs, exprs, integer_trail, model](
          const absl::StrongVector<IntegerVariable, double>& lp_values,
          LinearConstraintManager* manager) {
        absl::StrongVector<IntegerVariable, int> variable_partition(
            lp_values.size(), -1);
        absl::StrongVector<IntegerVariable, double> variable_partition_contrib(
            lp_values.size(), std::numeric_limits<double>::infinity());
        for (int expr_index = 0; expr_index < num_exprs; ++expr_index) {
          for (const IntegerVariable var : x_vars) {
            const double contribution = ComputeContribution(
                var, z_vars, exprs, lp_values, *integer_trail, expr_index);
            const double prev_contribution = variable_partition_contrib[var];
            if (contribution < prev_contribution) {
              variable_partition[var] = expr_index;
              variable_partition_contrib[var] = contribution;
            }
          }
        }

        LinearConstraintBuilder cut(model, /*lb=*/IntegerValue(0),
                                    /*ub=*/kMaxIntegerValue);
        double violation = lp_values[target];
        cut.AddTerm(target, IntegerValue(-1));

        for (const IntegerVariable xi_var : x_vars) {
          const int input_index = variable_partition[xi_var];
          const LinearExpression& expr = exprs[input_index];
          const IntegerValue coeff = GetCoefficientOfPositiveVar(xi_var, expr);
          if (coeff != IntegerValue(0)) {
            cut.AddTerm(xi_var, coeff);
          }
          violation -= ToDouble(coeff) * lp_values[xi_var];
        }
        for (int expr_index = 0; expr_index < num_exprs; ++expr_index) {
          const IntegerVariable z_var = z_vars[expr_index];
          const IntegerValue z_coeff = MPlusCoefficient(
              x_vars, exprs, variable_partition, expr_index, *integer_trail);
          if (z_coeff != IntegerValue(0)) {
            cut.AddTerm(z_var, z_coeff);
          }
          violation -= ToDouble(z_coeff) * lp_values[z_var];
        }
        if (violation > 1e-2) {
          manager->AddCut(cut.Build(), "LinMax", lp_values);
        }
        return true;
      };
  return result;
}

namespace {

IntegerValue EvaluateMaxAffine(
    const std::vector<std::pair<IntegerValue, IntegerValue>>& affines,
    IntegerValue x) {
  IntegerValue y = kMinIntegerValue;
  for (const auto& p : affines) {
    y = std::max(y, x * p.first + p.second);
  }
  return y;
}

}  // namespace

LinearConstraint BuildMaxAffineUpConstraint(
    const LinearExpression& target, IntegerVariable var,
    const std::vector<std::pair<IntegerValue, IntegerValue>>& affines,
    Model* model) {
  auto* integer_trail = model->GetOrCreate<IntegerTrail>();
  const IntegerValue x_min = integer_trail->LevelZeroLowerBound(var);
  const IntegerValue x_max = integer_trail->LevelZeroUpperBound(var);

  const IntegerValue y_at_min = EvaluateMaxAffine(affines, x_min);
  const IntegerValue y_at_max = EvaluateMaxAffine(affines, x_max);

  // TODO(user): Be careful to not have any integer overflow in any of
  // the formula used here.
  const IntegerValue delta_x = x_max - x_min;
  const IntegerValue delta_y = y_at_max - y_at_min;

  // target <= y_at_min + (delta_y / delta_x) * (var - x_min)
  // delta_x * target <= delta_x * y_at_min + delta_y * (var - x_min)
  // -delta_y * var + delta_x * target <= delta_x * y_at_min - delta_y * x_min
  const IntegerValue rhs = delta_x * y_at_min - delta_y * x_min;
  LinearConstraintBuilder lc(model, kMinIntegerValue, rhs);
  lc.AddLinearExpression(target, delta_x);
  lc.AddTerm(var, -delta_y);
  LinearConstraint ct = lc.Build();

  // Prevent to create constraints that can overflow.
  if (!ValidateLinearConstraintForOverflow(ct, *integer_trail)) {
    VLOG(2) << "Linear constraint can cause overflow: " << ct;

    // TODO(user): Change API instead of returning trivial constraint?
    ct.Clear();
  }

  return ct;
}

CutGenerator CreateMaxAffineCutGenerator(
    LinearExpression target, IntegerVariable var,
    std::vector<std::pair<IntegerValue, IntegerValue>> affines,
    const std::string cut_name, Model* model) {
  CutGenerator result;
  result.vars = target.vars;
  result.vars.push_back(var);
  gtl::STLSortAndRemoveDuplicates(&result.vars);

  IntegerTrail* integer_trail = model->GetOrCreate<IntegerTrail>();
  result.generate_cuts =
      [target, var, affines, cut_name, integer_trail, model](
          const absl::StrongVector<IntegerVariable, double>& lp_values,
          LinearConstraintManager* manager) {
        if (integer_trail->IsFixed(var)) return true;
        manager->AddCut(BuildMaxAffineUpConstraint(target, var, affines, model),
                        cut_name, lp_values);
        return true;
      };
  return result;
}

CutGenerator CreateCliqueCutGenerator(
    const std::vector<IntegerVariable>& base_variables, Model* model) {
  // Filter base_variables to only keep the one with a literal view, and
  // do the conversion.
  std::vector<IntegerVariable> variables;
  std::vector<Literal> literals;
  absl::flat_hash_map<LiteralIndex, IntegerVariable> positive_map;
  absl::flat_hash_map<LiteralIndex, IntegerVariable> negative_map;
  auto* integer_trail = model->GetOrCreate<IntegerTrail>();
  auto* encoder = model->GetOrCreate<IntegerEncoder>();
  for (const IntegerVariable var : base_variables) {
    if (integer_trail->LowerBound(var) != IntegerValue(0)) continue;
    if (integer_trail->UpperBound(var) != IntegerValue(1)) continue;
    const LiteralIndex literal_index = encoder->GetAssociatedLiteral(
        IntegerLiteral::GreaterOrEqual(var, IntegerValue(1)));
    if (literal_index != kNoLiteralIndex) {
      variables.push_back(var);
      literals.push_back(Literal(literal_index));
      positive_map[literal_index] = var;
      negative_map[Literal(literal_index).NegatedIndex()] = var;
    }
  }
  CutGenerator result;
  result.vars = variables;
  auto* implication_graph = model->GetOrCreate<BinaryImplicationGraph>();
  result.generate_cuts =
      [variables, literals, implication_graph, positive_map, negative_map,
       model](const absl::StrongVector<IntegerVariable, double>& lp_values,
              LinearConstraintManager* manager) {
        std::vector<double> packed_values;
        for (int i = 0; i < literals.size(); ++i) {
          packed_values.push_back(lp_values[variables[i]]);
        }
        const std::vector<std::vector<Literal>> at_most_ones =
            implication_graph->GenerateAtMostOnesWithLargeWeight(literals,
                                                                 packed_values);

        for (const std::vector<Literal>& at_most_one : at_most_ones) {
          // We need to express such "at most one" in term of the initial
          // variables, so we do not use the
          // LinearConstraintBuilder::AddLiteralTerm() here.
          LinearConstraintBuilder builder(
              model, IntegerValue(std::numeric_limits<int64_t>::min()),
              IntegerValue(1));
          for (const Literal l : at_most_one) {
            if (ContainsKey(positive_map, l.Index())) {
              builder.AddTerm(positive_map.at(l.Index()), IntegerValue(1));
            } else {
              // Add 1 - X to the linear constraint.
              builder.AddTerm(negative_map.at(l.Index()), IntegerValue(-1));
              builder.AddConstant(IntegerValue(1));
            }
          }

          manager->AddCut(builder.Build(), "clique", lp_values);
        }
        return true;
      };
  return result;
}

}  // namespace sat
}  // namespace operations_research