lpsolver 0.1.0 → 0.2.0

data/ext/lpsolver-highs/highs/presolve/HighsPostsolveStack.cpp

@@ -0,0 +1,1368 @@
+/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
+/*                                                                       */
+/*    This file is part of the HiGHS linear optimization suite           */
+/*                                                                       */
+/*    Available as open-source under the MIT License                     */
+/*                                                                       */
+/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
+#include "presolve/HighsPostsolveStack.h"
+
+#include <numeric>
+
+#include "lp_data/HConst.h"
+#include "lp_data/HighsModelUtils.h"  // For debugging #2001
+#include "lp_data/HighsOptions.h"
+#include "util/HighsCDouble.h"
+#include "util/HighsUtils.h"
+
+namespace presolve {
+
+void HighsPostsolveStack::initializeIndexMaps(HighsInt numRow,
+                                              HighsInt numCol) {
+  origNumRow = numRow;
+  origNumCol = numCol;
+
+  origRowIndex.resize(numRow);
+  std::iota(origRowIndex.begin(), origRowIndex.end(), 0);
+
+  origColIndex.resize(numCol);
+  std::iota(origColIndex.begin(), origColIndex.end(), 0);
+
+  linearlyTransformable.resize(numCol, true);
+}
+
+void HighsPostsolveStack::compressIndexMaps(
+    const std::vector<HighsInt>& newRowIndex,
+    const std::vector<HighsInt>& newColIndex) {
+  // loop over rows, decrease row counter for deleted rows (marked with -1),
+  // store original index at new index position otherwise
+  HighsInt numRow = origRowIndex.size();
+  for (size_t i = 0; i != newRowIndex.size(); ++i) {
+    if (newRowIndex[i] == -1)
+      --numRow;
+    else
+      origRowIndex[newRowIndex[i]] = origRowIndex[i];
+  }
+  // resize original index array to new size
+  origRowIndex.resize(numRow);
+
+  // now compress the column array
+  HighsInt numCol = origColIndex.size();
+  for (size_t i = 0; i != newColIndex.size(); ++i) {
+    if (newColIndex[i] == -1)
+      --numCol;
+    else
+      origColIndex[newColIndex[i]] = origColIndex[i];
+  }
+  origColIndex.resize(numCol);
+}
+
+void HighsPostsolveStack::LinearTransform::undo(const HighsOptions& options,
+                                                HighsSolution& solution) const {
+  solution.col_value[col] *= scale;
+  solution.col_value[col] += constant;
+
+  if (solution.dual_valid) solution.col_dual[col] /= scale;
+}
+
+void HighsPostsolveStack::LinearTransform::transformToPresolvedSpace(
+    std::vector<double>& primalSol) const {
+  primalSol[col] -= constant;
+  primalSol[col] /= scale;
+}
+
+static HighsBasisStatus computeRowStatus(double dual,
+                                         HighsPostsolveStack::RowType rowType) {
+  if (rowType == HighsPostsolveStack::RowType::kEq)
+    return dual < 0 ? HighsBasisStatus::kUpper : HighsBasisStatus::kLower;
+  else if (rowType == HighsPostsolveStack::RowType::kGeq)
+    return HighsBasisStatus::kLower;
+  else
+    return HighsBasisStatus::kUpper;
+}
+
+void HighsPostsolveStack::FreeColSubstitution::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& rowValues,
+    const std::vector<Nonzero>& colValues, HighsSolution& solution,
+    HighsBasis& basis) {
+  // a (removed) cut may have been used in this reduction.
+  bool isModelRow = static_cast<size_t>(row) < solution.row_value.size();
+
+  // compute primal values
+  double colCoef = 0;
+  HighsCDouble rowValue = 0;
+  for (const auto& rowVal : rowValues) {
+    if (rowVal.index == col)
+      colCoef = rowVal.value;
+    else
+      rowValue += rowVal.value * solution.col_value[rowVal.index];
+  }
+
+  assert(colCoef != 0);
+  // Row values aren't fully postsolved, so why do this?
+  if (isModelRow)
+    solution.row_value[row] =
+        double(rowValue + colCoef * solution.col_value[col]);
+  solution.col_value[col] = double((rhs - rowValue) / colCoef);
+
+  // if no dual values requested, return here
+  if (!solution.dual_valid) return;
+
+  // compute the row dual value such that reduced cost of basic column is 0
+  if (isModelRow) {
+    solution.row_dual[row] = 0;
+    HighsCDouble dualval = colCost;
+    for (const auto& colVal : colValues) {
+      if (static_cast<size_t>(colVal.index) < solution.row_dual.size())
+        dualval -= colVal.value * solution.row_dual[colVal.index];
+    }
+    solution.row_dual[row] = double(dualval / colCoef);
+  }
+
+  solution.col_dual[col] = 0;
+
+  // set basis status if necessary
+  if (!basis.valid) return;
+
+  basis.col_status[col] = HighsBasisStatus::kBasic;
+  if (isModelRow)
+    basis.row_status[row] = computeRowStatus(solution.row_dual[row], rowType);
+}
+
+static HighsBasisStatus computeStatus(double dual, HighsBasisStatus& status,
+                                      double dual_feasibility_tolerance) {
+  if (dual > dual_feasibility_tolerance)
+    status = HighsBasisStatus::kLower;
+  else if (dual < -dual_feasibility_tolerance)
+    status = HighsBasisStatus::kUpper;
+
+  return status;
+}
+
+static HighsBasisStatus computeStatus(double dual,
+                                      double dual_feasibility_tolerance) {
+  if (dual > dual_feasibility_tolerance)
+    return HighsBasisStatus::kLower;
+  else if (dual < -dual_feasibility_tolerance)
+    return HighsBasisStatus::kUpper;
+  else
+    return HighsBasisStatus::kBasic;
+}
+
+void HighsPostsolveStack::DoubletonEquation::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& colValues,
+    HighsSolution& solution, HighsBasis& basis) const {
+  // retrieve the row and column index, the row side and the two
+  // coefficients then compute the primal values
+  solution.col_value[colSubst] =
+      double((rhs - HighsCDouble(coef) * solution.col_value[col]) / coefSubst);
+
+  // can only do primal postsolve, stop here
+  if (row == -1 || !solution.dual_valid) return;
+
+  const HighsBasisStatus colStatus =
+      !basis.valid
+          ? computeStatus(solution.col_dual[col],
+                          options.dual_feasibility_tolerance)
+          : computeStatus(solution.col_dual[col], basis.col_status[col],
+                          options.dual_feasibility_tolerance);
+
+  // a (removed) cut may have been used in this reduction.
+  bool isModelRow = static_cast<size_t>(row) < solution.row_value.size();
+
+  // compute the current dual values of the row and the substituted column
+  // before deciding on which column becomes basic
+  // for each entry in a row i of the substituted column we added the
+  // doubleton equation row with scale -a_i/substCoef. Therefore the dual
+  // multiplier of this row i implicitly increases the dual multiplier of this
+  // doubleton equation row with that scale.
+  HighsCDouble rowDual = 0.0;
+  if (isModelRow) {
+    solution.row_dual[row] = 0;
+    for (const auto& colVal : colValues) {
+      if (static_cast<size_t>(colVal.index) < solution.row_dual.size())
+        rowDual -= colVal.value * solution.row_dual[colVal.index];
+    }
+    rowDual /= coefSubst;
+    solution.row_dual[row] = double(rowDual);
+  }
+  // the equation was also added to the objective, so the current values need
+  // to be changed
+  solution.col_dual[colSubst] = substCost;
+  solution.col_dual[col] += substCost * coef / coefSubst;
+
+  if ((upperTightened && colStatus == HighsBasisStatus::kUpper) ||
+      (lowerTightened && colStatus == HighsBasisStatus::kLower)) {
+    // column must get zero reduced cost as the current bound cannot be used
+    // so alter the dual multiplier of the row to make the dual multiplier of
+    // column zero
+    double rowDualDelta = solution.col_dual[col] / coef;
+    if (isModelRow) solution.row_dual[row] = double(rowDual + rowDualDelta);
+    solution.col_dual[col] = 0.0;
+    solution.col_dual[colSubst] = double(
+        HighsCDouble(solution.col_dual[colSubst]) - rowDualDelta * coefSubst);
+
+    if (basis.valid) {
+      if ((std::signbit(coef) == std::signbit(coefSubst) &&
+           basis.col_status[col] == HighsBasisStatus::kUpper) ||
+          (std::signbit(coef) != std::signbit(coefSubst) &&
+           basis.col_status[col] == HighsBasisStatus::kLower))
+        basis.col_status[colSubst] = HighsBasisStatus::kLower;
+      else
+        basis.col_status[colSubst] = HighsBasisStatus::kUpper;
+      basis.col_status[col] = HighsBasisStatus::kBasic;
+    }
+  } else {
+    // otherwise make the reduced cost of the substituted column zero and make
+    // that column basic
+    double rowDualDelta = solution.col_dual[colSubst] / coefSubst;
+    if (isModelRow) solution.row_dual[row] = double(rowDual + rowDualDelta);
+    solution.col_dual[colSubst] = 0.0;
+    solution.col_dual[col] =
+        double(HighsCDouble(solution.col_dual[col]) - rowDualDelta * coef);
+    if (basis.valid) basis.col_status[colSubst] = HighsBasisStatus::kBasic;
+  }
+
+  if (!basis.valid) return;
+
+  if (isModelRow)
+    basis.row_status[row] = computeRowStatus(solution.row_dual[row], rowType);
+}
+
+void HighsPostsolveStack::EqualityRowAddition::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& eqRowValues,
+    HighsSolution& solution, HighsBasis& basis) const {
+  // (removed) cuts may have been used in this reduction.
+  if (static_cast<size_t>(row) >= solution.row_value.size() ||
+      static_cast<size_t>(addedEqRow) >= solution.row_value.size())
+    return;
+
+  // nothing more to do if the row is zero in the dual solution or there is
+  // no dual solution
+  if (!solution.dual_valid || solution.row_dual[row] == 0.0) return;
+
+  // the dual multiplier of the row implicitly increases the dual multiplier
+  // of the equation with the scale the equation was added with
+  solution.row_dual[addedEqRow] =
+      double(HighsCDouble(eqRowScale) * solution.row_dual[row] +
+             solution.row_dual[addedEqRow]);
+
+  assert(!basis.valid);
+}
+
+void HighsPostsolveStack::EqualityRowAdditions::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& eqRowValues,
+    const std::vector<Nonzero>& targetRows, HighsSolution& solution,
+    HighsBasis& basis) const {
+  // a (removed) cut may have been used in this reduction.
+  if (static_cast<size_t>(addedEqRow) >= solution.row_value.size()) return;
+
+  // nothing more to do if the row is zero in the dual solution or there is
+  // no dual solution
+  if (!solution.dual_valid) return;
+
+  // the dual multiplier of the rows where the eq row was added implicitly
+  // increases the dual multiplier of the equation with the scale that was
+  // used for adding the equation
+  HighsCDouble eqRowDual = solution.row_dual[addedEqRow];
+  for (const auto& targetRow : targetRows) {
+    if (static_cast<size_t>(targetRow.index) < solution.row_dual.size())
+      eqRowDual +=
+          HighsCDouble(targetRow.value) * solution.row_dual[targetRow.index];
+  }
+  solution.row_dual[addedEqRow] = double(eqRowDual);
+
+  assert(!basis.valid);
+}
+
+void HighsPostsolveStack::ForcingColumn::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& colValues,
+    HighsSolution& solution, HighsBasis& basis) const {
+  HighsInt nonbasicRow = -1;
+  HighsBasisStatus nonbasicRowStatus = HighsBasisStatus::kNonbasic;
+  double colValFromNonbasicRow = colBound;
+
+  HighsInt debug_num_use_row_value = 0;
+  const bool debug_report = false;
+
+  auto computeColVal = [&](HighsInt direction) {
+    // choose solution value that makes all rows feasible
+    for (const auto& colVal : colValues) {
+      // Row values aren't fully postsolved, so how can this work?
+      debug_num_use_row_value++;
+      if (static_cast<size_t>(colVal.index) < solution.row_value.size()) {
+        double colValFromRow = solution.row_value[colVal.index] / colVal.value;
+        if (direction * colValFromRow > direction * colValFromNonbasicRow) {
+          nonbasicRow = colVal.index;
+          colValFromNonbasicRow = colValFromRow;
+          nonbasicRowStatus = direction * colVal.value > 0
+                                  ? HighsBasisStatus::kLower
+                                  : HighsBasisStatus::kUpper;
+        }
+      }
+    }
+    // round solution value if column is integer-constrained
+    if (nonbasicRow != -1 && colIntegral)
+      colValFromNonbasicRow =
+          direction * std::ceil(direction * colValFromNonbasicRow -
+                                options.mip_feasibility_tolerance);
+  };
+
+  if (atInfiniteUpper) {
+    // choose largest value as then all rows are feasible
+    computeColVal(HighsInt{1});
+  } else {
+    // choose smallest value, as then all rows are feasible
+    computeColVal(HighsInt{-1});
+  }
+  if (debug_num_use_row_value && debug_report) {
+    printf(
+        "HighsPostsolveStack::ForcingColumn::undo Using %d unknown row "
+        "activit%s\n",
+        int(debug_num_use_row_value),
+        debug_num_use_row_value > 1 ? "ies" : "y");
+  }
+
+  solution.col_value[col] = colValFromNonbasicRow;
+
+  if (!solution.dual_valid) return;
+
+  solution.col_dual[col] = 0.0;
+
+  if (!basis.valid) return;
+  if (nonbasicRow == -1) {
+    basis.col_status[col] =
+        atInfiniteUpper ? HighsBasisStatus::kLower : HighsBasisStatus::kUpper;
+  } else {
+    basis.col_status[col] = HighsBasisStatus::kBasic;
+    basis.row_status[nonbasicRow] = nonbasicRowStatus;
+  }
+}
+
+void HighsPostsolveStack::ForcingColumnRemovedRow::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& rowValues,
+    HighsSolution& solution, HighsBasis& basis) const {
+  // a (removed) cut may have been used in this reduction.
+  if (static_cast<size_t>(row) >= solution.row_value.size()) return;
+
+  // we use the row value as storage for the scaled value implied on the
+  // column dual
+  HighsCDouble val = rhs;
+  for (const auto& rowVal : rowValues)
+    val -= rowVal.value * solution.col_value[rowVal.index];
+
+  // Row values aren't fully postsolved, so why do this?
+  solution.row_value[row] = double(val);
+
+  if (solution.dual_valid) solution.row_dual[row] = 0.0;
+  if (basis.valid) basis.row_status[row] = HighsBasisStatus::kBasic;
+}
+
+void HighsPostsolveStack::SingletonRow::undo(const HighsOptions& options,
+                                             HighsSolution& solution,
+                                             HighsBasis& basis) const {
+  // a (removed) cut may have been used in this reduction.
+  bool isModelRow = static_cast<size_t>(row) < solution.row_value.size();
+
+  // nothing to do if the rows dual value is zero in the dual solution or
+  // there is no dual solution
+  if (!solution.dual_valid) return;
+
+  const HighsBasisStatus colStatus =
+      !basis.valid
+          ? computeStatus(solution.col_dual[col],
+                          options.dual_feasibility_tolerance)
+          : computeStatus(solution.col_dual[col], basis.col_status[col],
+                          options.dual_feasibility_tolerance);
+
+  if ((!colLowerTightened || colStatus != HighsBasisStatus::kLower) &&
+      (!colUpperTightened || colStatus != HighsBasisStatus::kUpper)) {
+    // the tightened bound is not used in the basic solution
+    // hence we simply make the row basic and give it a dual multiplier of 0
+    if (isModelRow) {
+      if (basis.valid) basis.row_status[row] = HighsBasisStatus::kBasic;
+      solution.row_dual[row] = 0;
+    }
+    return;
+  }
+
+  // choose the row dual value such that the columns reduced cost becomes
+  // zero
+  if (isModelRow) solution.row_dual[row] = solution.col_dual[col] / coef;
+  solution.col_dual[col] = 0;
+
+  if (!basis.valid) return;
+
+  if (isModelRow) {
+    switch (colStatus) {
+      case HighsBasisStatus::kLower:
+        assert(colLowerTightened);
+        if (coef > 0)
+          // tightened lower bound comes from row lower bound
+          basis.row_status[row] = HighsBasisStatus::kLower;
+        else
+          // tightened lower bound comes from row upper bound
+          basis.row_status[row] = HighsBasisStatus::kUpper;
+
+        break;
+      case HighsBasisStatus::kUpper:
+        if (coef > 0)
+          // tightened upper bound comes from row lower bound
+          basis.row_status[row] = HighsBasisStatus::kUpper;
+        else
+          // tightened lower bound comes from row upper bound
+          basis.row_status[row] = HighsBasisStatus::kLower;
+        break;
+      default:
+        assert(false);
+    }
+  }
+
+  // column becomes basic
+  basis.col_status[col] = HighsBasisStatus::kBasic;
+}
+
+// column fixed to lower or upper bound
+void HighsPostsolveStack::FixedCol::undo(const HighsOptions& options,
+                                         const std::vector<Nonzero>& colValues,
+                                         HighsSolution& solution,
+                                         HighsBasis& basis) const {
+  // set solution value
+  solution.col_value[col] = fixValue;
+
+  if (!solution.dual_valid) return;
+
+  // compute reduced cost
+
+  HighsCDouble reducedCost = colCost;
+  for (const auto& colVal : colValues) {
+    if (static_cast<size_t>(colVal.index) < solution.row_dual.size())
+      reducedCost -= colVal.value * solution.row_dual[colVal.index];
+  }
+
+  solution.col_dual[col] = double(reducedCost);
+
+  // set basis status
+  if (basis.valid) {
+    basis.col_status[col] = fixType;
+    if (basis.col_status[col] == HighsBasisStatus::kNonbasic)
+      basis.col_status[col] = solution.col_dual[col] >= 0
+                                  ? HighsBasisStatus::kLower
+                                  : HighsBasisStatus::kUpper;
+  }
+}
+
+void HighsPostsolveStack::RedundantRow::undo(const HighsOptions& options,
+                                             HighsSolution& solution,
+                                             HighsBasis& basis) const {
+  // a (removed) cut may have been used in this reduction.
+  if (static_cast<size_t>(row) >= solution.row_value.size()) return;
+
+  // set row dual to zero if dual solution requested
+  if (!solution.dual_valid) return;
+
+  solution.row_dual[row] = 0.0;
+
+  if (basis.valid) basis.row_status[row] = HighsBasisStatus::kBasic;
+}
+
+void HighsPostsolveStack::ForcingRow::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& rowValues,
+    HighsSolution& solution, HighsBasis& basis) const {
+  if (!solution.dual_valid) return;
+
+  // compute the row dual multiplier and determine the new basic column
+  HighsInt basicCol = -1;
+  double dualDelta = 0;
+  HighsInt direction = rowType == RowType::kLeq ? 1 : -1;
+  for (const auto& rowVal : rowValues) {
+    double colDual = solution.col_dual[rowVal.index] - rowVal.value * dualDelta;
+    if (direction * colDual * rowVal.value < 0) {
+      // column is dual infeasible, decrease the row dual such that its
+      // reduced cost become zero and remember this column as the new basic
+      // column for this row
+      dualDelta = solution.col_dual[rowVal.index] / rowVal.value;
+      basicCol = rowVal.index;
+    }
+  }
+
+  if (basicCol != -1) {
+    bool isModelRow = static_cast<size_t>(row) < solution.row_dual.size();
+    if (isModelRow) solution.row_dual[row] = solution.row_dual[row] + dualDelta;
+    for (const auto& rowVal : rowValues) {
+      solution.col_dual[rowVal.index] =
+          double(solution.col_dual[rowVal.index] -
+                 HighsCDouble(dualDelta) * rowVal.value);
+    }
+    solution.col_dual[basicCol] = 0;
+
+    if (basis.valid) {
+      if (isModelRow)
+        basis.row_status[row] =
+            (rowType == RowType::kGeq ? HighsBasisStatus::kLower
+                                      : HighsBasisStatus::kUpper);
+
+      basis.col_status[basicCol] = HighsBasisStatus::kBasic;
+    }
+  }
+}
+
+void HighsPostsolveStack::DuplicateRow::undo(const HighsOptions& options,
+                                             HighsSolution& solution,
+                                             HighsBasis& basis) const {
+  // (removed) cuts may have been used in this reduction.
+  if (static_cast<size_t>(row) >= solution.row_value.size()) return;
+  bool duplicateIsModelRow =
+      static_cast<size_t>(duplicateRow) < solution.row_value.size();
+
+  if (!solution.dual_valid) return;
+  if (!rowUpperTightened && !rowLowerTightened) {
+    // simple case of row2 being redundant, in which case it just gets a
+    // dual multiplier of 0 and is made basic
+    if (duplicateIsModelRow) {
+      solution.row_dual[duplicateRow] = 0.0;
+      if (basis.valid)
+        basis.row_status[duplicateRow] = HighsBasisStatus::kBasic;
+    }
+    return;
+  }
+
+  const HighsBasisStatus rowStatus =
+      !basis.valid
+          ? computeStatus(solution.row_dual[row],
+                          options.dual_feasibility_tolerance)
+          : computeStatus(solution.row_dual[row], basis.row_status[row],
+                          options.dual_feasibility_tolerance);
+
+  auto computeRowDualAndStatus = [&](bool tighened) {
+    if (tighened) {
+      if (duplicateIsModelRow) {
+        solution.row_dual[duplicateRow] =
+            solution.row_dual[row] / duplicateRowScale;
+        if (basis.valid) {
+          if (duplicateRowScale > 0)
+            basis.row_status[duplicateRow] = HighsBasisStatus::kUpper;
+          else
+            basis.row_status[duplicateRow] = HighsBasisStatus::kLower;
+        }
+      }
+      solution.row_dual[row] = 0.0;
+      if (basis.valid) basis.row_status[row] = HighsBasisStatus::kBasic;
+    } else if (duplicateIsModelRow) {
+      solution.row_dual[duplicateRow] = 0.0;
+      if (basis.valid)
+        basis.row_status[duplicateRow] = HighsBasisStatus::kBasic;
+    }
+  };
+
+  // at least one bound of the row was tightened by using the bound of the
+  // scaled parallel row, hence we might need to make the parallel row
+  // nonbasic and the row basic
+
+  switch (rowStatus) {
+    case HighsBasisStatus::kBasic:
+      // if row is basic the parallel row is also basic
+      if (duplicateIsModelRow) {
+        solution.row_dual[duplicateRow] = 0.0;
+        if (basis.valid)
+          basis.row_status[duplicateRow] = HighsBasisStatus::kBasic;
+      }
+      break;
+    case HighsBasisStatus::kUpper:
+      // if row sits on its upper bound, and the row upper bound was
+      // tightened using the parallel row we make the row basic and
+      // transfer its dual value to the parallel row with the proper scale
+      computeRowDualAndStatus(rowUpperTightened);
+      break;
+    case HighsBasisStatus::kLower:
+      computeRowDualAndStatus(rowLowerTightened);
+      break;
+    default:
+      assert(false);
+  }
+}
+
+void HighsPostsolveStack::DuplicateColumn::undo(const HighsOptions& options,
+                                                HighsSolution& solution,
+                                                HighsBasis& basis) const {
+  const bool debug_report = false;
+  const double mergeVal = solution.col_value[col];
+
+  auto okResidual = [&](const double x, const double y) {
+    const double check_mergeVal = x + colScale * y;
+    const double residual = std::fabs(check_mergeVal - mergeVal);
+    const bool ok_residual = residual <= options.primal_feasibility_tolerance;
+    if (!ok_residual && debug_report) {
+      printf(
+          "HighsPostsolveStack::DuplicateColumn::undo %g + %g.%g = %g != %g: "
+          "residual = %g\n",
+          x, colScale, y, check_mergeVal, mergeVal, residual);
+    }
+    return ok_residual;
+  };
+
+  auto isAtBound = [&](const double value, const double bound) {
+    if (value < bound - options.primal_feasibility_tolerance) return false;
+    if (value <= bound + options.primal_feasibility_tolerance) return true;
+    return false;
+  };
+
+  //  const bool ok_merge = okMerge(options.mip_feasibility_tolerance);
+  //  assert(ok_merge);
+  //
+  // the column dual of the duplicate column is easily computed by scaling
+  // since col * colScale yields the coefficient values and cost of the
+  // duplicate column.
+  if (solution.dual_valid)
+    solution.col_dual[duplicateCol] = solution.col_dual[col] * colScale;
+
+  if (basis.valid) {
+    // do postsolve using basis status if a basis is available: if the
+    // merged column is nonbasic, we can just set both columns to
+    // appropriate nonbasic status and value
+    //
+    // Undoing z = x + a.y
+    //
+    // Since x became z, its basis status is unchanged
+    //
+    // For a > 0, z\in [x_l + a.y_l, x_u + a.y_u]
+    //
+    // If z is nonbasic at its lower (upper) bound, set y to be
+    // nonbasic at its lower (upper) bound
+    //
+    // For a < 0, z\in [x_l + a.y_u, x_u + a.y_l]
+    //
+    // If z is nonbasic at lower (upper) bound, set y to be nonbasic
+    // at its upper (lower) bounds
+    //
+    // Check for perturbations
+    switch (basis.col_status[col]) {
+      case HighsBasisStatus::kLower: {
+        solution.col_value[col] = colLower;
+        if (colScale > 0) {
+          basis.col_status[duplicateCol] = HighsBasisStatus::kLower;
+          solution.col_value[duplicateCol] = duplicateColLower;
+        } else {
+          basis.col_status[duplicateCol] = HighsBasisStatus::kUpper;
+          solution.col_value[duplicateCol] = duplicateColUpper;
+        }
+        // nothing else to do
+        assert(okResidual(solution.col_value[col],
+                          solution.col_value[duplicateCol]));
+        return;
+      }
+      case HighsBasisStatus::kUpper: {
+        solution.col_value[col] = colUpper;
+        if (colScale > 0) {
+          basis.col_status[duplicateCol] = HighsBasisStatus::kUpper;
+          solution.col_value[duplicateCol] = duplicateColUpper;
+        } else {
+          basis.col_status[duplicateCol] = HighsBasisStatus::kLower;
+          solution.col_value[duplicateCol] = duplicateColLower;
+        }
+        // nothing else to do
+        assert(okResidual(solution.col_value[col],
+                          solution.col_value[duplicateCol]));
+        return;
+      }
+      case HighsBasisStatus::kZero: {
+        solution.col_value[col] = 0.0;
+        basis.col_status[duplicateCol] = HighsBasisStatus::kZero;
+        solution.col_value[duplicateCol] = 0.0;
+        // nothing else to do
+        assert(okResidual(solution.col_value[col],
+                          solution.col_value[duplicateCol]));
+        return;
+      }
+      case HighsBasisStatus::kBasic:
+      case HighsBasisStatus::kNonbasic:;
+    }
+    // Nonbasic cases should have been considered; basic case
+    // considered later
+    assert(basis.col_status[col] == HighsBasisStatus::kBasic);
+  }
+
+  // either no basis for postsolve, or column status is basic. One of
+  // the two columns must become nonbasic. In case of integrality it is
+  // simpler to choose col, since it has a coefficient of +1 in the equation z
+  // = col + colScale * duplicateCol where the merged column is z and is
+  // currently using the index of col. The duplicateCol can have a positive or
+  // negative coefficient. So for postsolve, we first start out with col
+  // sitting at the lower bound and compute the corresponding value for the
+  // duplicate column as (z - colLower)/colScale. Then the following things
+  // might happen:
+  // - case 1: the value computed for duplicateCol is within the bounds
+  // - case 1.1: duplicateCol is continuous -> accept value, make col nonbasic
+  // at lower and duplicateCol basic
+  // - case 1.2: duplicateCol is integer -> accept value if integer feasible,
+  // otherwise round down and compute value of col as
+  // col = z - colScale * duplicateCol
+  // - case 2: the value for duplicateCol violates the column bounds: make it
+  // sit at the bound that is violated
+  //           and compute the value of col as col = z - colScale *
+  //           duplicateCol for basis postsolve col is basic and duplicateCol
+  //           nonbasic at lower/upper depending on which bound is violated.
+
+  if (colLower != -kHighsInf)
+    solution.col_value[col] = colLower;
+  else
+    solution.col_value[col] = std::min(0.0, colUpper);
+  solution.col_value[duplicateCol] = static_cast<double>(
+      (static_cast<HighsCDouble>(mergeVal) - solution.col_value[col]) /
+      colScale);
+
+  bool recomputeCol = false;
+
+  // Set any basis status for duplicateCol to kNonbasic to check that
+  // it is set
+  if (basis.valid) basis.col_status[duplicateCol] = HighsBasisStatus::kNonbasic;
+
+  if (solution.col_value[duplicateCol] > duplicateColUpper) {
+    solution.col_value[duplicateCol] = duplicateColUpper;
+    recomputeCol = true;
+    if (basis.valid) basis.col_status[duplicateCol] = HighsBasisStatus::kUpper;
+  } else if (solution.col_value[duplicateCol] < duplicateColLower) {
+    solution.col_value[duplicateCol] = duplicateColLower;
+    recomputeCol = true;
+    if (basis.valid) basis.col_status[duplicateCol] = HighsBasisStatus::kLower;
+  } else if (duplicateColIntegral) {
+    // Doesn't set basis.col_status[duplicateCol], so assume no basis
+    assert(!basis.valid);
+    if (fractionality(solution.col_value[duplicateCol]) >
+        options.mip_feasibility_tolerance) {
+      solution.col_value[duplicateCol] =
+          std::floor(solution.col_value[duplicateCol]);
+      recomputeCol = true;
+    }
+  }
+
+  if (recomputeCol) {
+    solution.col_value[col] =
+        mergeVal - colScale * solution.col_value[duplicateCol];
+    if (!duplicateColIntegral && colIntegral) {
+      // if column is integral and duplicateCol is not we need to make sure
+      // we split the values into an integral one for col
+      //
+      // Doesn't set basis.col_status[duplicateCol], so assume no basis
+      assert(!basis.valid);
+      solution.col_value[col] = std::ceil(solution.col_value[col] -
+                                          options.mip_feasibility_tolerance);
+      solution.col_value[duplicateCol] = static_cast<double>(
+          (static_cast<HighsCDouble>(mergeVal) - solution.col_value[col]) /
+          colScale);
+    }
+  } else {
+    // setting col to its lower bound yielded a feasible value for
+    // duplicateCol - not necessarily!
+    if (basis.valid) {
+      // This makes duplicateCol basic
+      basis.col_status[duplicateCol] = basis.col_status[col];
+      basis.col_status[col] = HighsBasisStatus::kLower;
+      assert(basis.col_status[duplicateCol] == HighsBasisStatus::kBasic);
+    }
+  }
+  // Check that any basis status for duplicateCol has been set
+  if (basis.valid)
+    assert(basis.col_status[duplicateCol] != HighsBasisStatus::kNonbasic);
+
+  auto hasError = [&]() {
+    bool illegal_duplicateCol_lower =
+        solution.col_value[duplicateCol] <
+        duplicateColLower - options.mip_feasibility_tolerance;
+    bool illegal_duplicateCol_upper =
+        solution.col_value[duplicateCol] >
+        duplicateColUpper + options.mip_feasibility_tolerance;
+    bool illegal_col_lower =
+        solution.col_value[col] < colLower - options.mip_feasibility_tolerance;
+    bool illegal_col_upper =
+        solution.col_value[col] > colUpper + options.mip_feasibility_tolerance;
+    bool illegal_residual =
+        !okResidual(solution.col_value[col], solution.col_value[duplicateCol]);
+    return (illegal_duplicateCol_lower || illegal_duplicateCol_upper ||
+            illegal_col_lower || illegal_col_upper || illegal_residual);
+  };
+
+  if (hasError()) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: col = %d(%g), duplicateCol = %d(%g)\n"
+          "%g\n%g\n%g %g %d\n%g %g %d\n",
+          int(col), solution.col_value[col], int(duplicateCol),
+          solution.col_value[duplicateCol], mergeVal, colScale, colLower,
+          colUpper, colIntegral, duplicateColLower, duplicateColUpper,
+          duplicateColIntegral);
+    // Fix error due to undo
+    undoFix(options, solution, mergeVal);
+  } else {
+    return;
+  }
+  const bool allow_assert = false;
+  if (allow_assert) assert(!hasError());
+  // Following undoFix, set any basis status, ideally keeping col basic
+  if (basis.valid) {
+    bool duplicateCol_basic = false;
+    if (duplicateColLower <= -kHighsInf && duplicateColUpper >= kHighsInf) {
+      // duplicateCol is free, so may be zero
+      if (solution.col_value[duplicateCol] == 0) {
+        basis.col_status[col] = HighsBasisStatus::kBasic;
+        basis.col_status[duplicateCol] = HighsBasisStatus::kZero;
+      } else {
+        duplicateCol_basic = true;
+      }
+    } else if (isAtBound(solution.col_value[duplicateCol], duplicateColLower)) {
+      basis.col_status[col] = HighsBasisStatus::kBasic;
+      basis.col_status[duplicateCol] = HighsBasisStatus::kLower;
+    } else if (isAtBound(solution.col_value[duplicateCol], duplicateColUpper)) {
+      basis.col_status[col] = HighsBasisStatus::kBasic;
+      basis.col_status[duplicateCol] = HighsBasisStatus::kUpper;
+    } else {
+      // duplicateCol is not free or at a bound, so must be basic
+      duplicateCol_basic = true;
+    }
+    if (duplicateCol_basic) {
+      // duplicateCol must be basic
+      basis.col_status[duplicateCol] = HighsBasisStatus::kBasic;
+      // Hopefully col can be nonbasic
+      if (isAtBound(solution.col_value[col], colLower)) {
+        basis.col_status[col] = HighsBasisStatus::kLower;
+      } else if (isAtBound(solution.col_value[col], colUpper)) {
+        basis.col_status[col] = HighsBasisStatus::kUpper;
+      } else {
+        basis.col_status[col] = HighsBasisStatus::kNonbasic;
+        if (debug_report)
+          printf(
+              "When demerging, neither col nor duplicateCol can be "
+              "nonbasic\n");
+        if (kAllowDeveloperAssert) assert(666 == 999);
+      }
+    }
+  }
+}
+
+bool HighsPostsolveStack::DuplicateColumn::okMerge(
+    const double tolerance) const {
+  // When merging x and y to x+a.y, not all values of a are permitted,
+  // since it must be possible to map back onto feasible values of x
+  // and y.
+  //
+  // Assume WLOG that a > 0, x\in[x_l, x_u], y\in[y_l, y_u]
+  //
+  // Let z = x + a.y
+  //
+  // Range for z is [x_l+a.y_l, x_u+a.y_u]
+  //
+  // * If x and y are both integer:
+  //
+  // z will be integer and x+a.y must generate all integer values in
+  // [x_l+a.y_l, x_u+a.y_u]. Hence a must be an integer. If a >=
+  // (x_u-x_l)+2 then, since [a.y_l, a.y_u] contains integer multiples
+  // of a, some of the intervening integers don't correspond to a
+  // value of x. Hence a must be an integer and a <= (x_u-x_l)+1
+  //
+  // For example, if x and y are binary, then x+a.y is [0, 1, a,
+  // 1+a]. For this to be a continuous sequence of integers, we must
+  // have a <= 2.
+  //
+  // * If x is integer and y is continuous:
+  //
+  // z will be continuous and x+a.y must generate all values in
+  // [x_l+a.y_l, x_u+a.y_u]. Since [x_l, x_u] are integers, [a.y_l,
+  // a.y_u] = a[y_l, y_u] must be of length at least 1. Hence a must
+  // be at least 1/(y_u-y_l) in magnitude.
+  //
+  // * If x is continuous and y is integer:
+  //
+  // z will be continuous and x+a.y must generate all values in
+  // [x_l+a.y_l, x_u+a.y_u]. Since [a.y_l, a.y_u] contains integer
+  // multiples of a, the gaps between them must not exceed the length
+  // of [x_l, x_u]. Hence a must be at most x_u-x_l in
+  // magnitude.
+  //
+  // Observe that this is equivalent to requiring 1/a to be at least
+  // 1/(x_u-x_l) in magnitude, the symmetric result corresponding to
+  // the merge (1/a)x+y.
+  //
+  //  * If x and y are both continuous
+  //
+  // z will be continuous and x+a.y naturally generates all values in
+  // [x_l+a.y_l, x_u+a.y_u].
+
+  const double scale = colScale;
+  const bool x_int = colIntegral;
+  const bool y_int = duplicateColIntegral;
+  const double x_lo = x_int ? std::ceil(colLower - tolerance) : colLower;
+  const double x_up = x_int ? std::floor(colUpper + tolerance) : colUpper;
+  const double y_lo =
+      y_int ? std::ceil(duplicateColLower - tolerance) : duplicateColLower;
+  const double y_up =
+      y_int ? std::floor(duplicateColUpper + tolerance) : duplicateColUpper;
+  const double x_len = x_up - x_lo;
+  const double y_len = y_up - y_lo;
+  std::string newline = "\n";
+  bool ok_merge = true;
+  const bool debug_report = false;
+  if (scale == 0) {
+    if (debug_report)
+      printf("%sDuplicateColumn::checkMerge: Scale cannot be zero\n",
+             newline.c_str());
+    newline = "";
+    ok_merge = false;
+  }
+  const double abs_scale = std::fabs(scale);
+  if (x_int) {
+    if (y_int) {
+      // Scale must be integer and not exceed (x_u-x_l)+1 in magnitude
+      if (fractionality(scale) > tolerance) {
+        if (debug_report)
+          printf(
+              "%sDuplicateColumn::checkMerge: scale must be integer, but is "
+              "%g\n",
+              newline.c_str(), scale);
+        newline = "";
+        ok_merge = false;
+      }
+      double scale_limit = x_len + 1 + tolerance;
+      if (abs_scale > scale_limit) {
+        if (debug_report)
+          printf(
+              "%sDuplicateColumn::checkMerge: scale = %g, but |scale| cannot "
+              "exceed %g since x is [%g, %g]\n",
+              newline.c_str(), scale, scale_limit, x_lo, x_up);
+        newline = "";
+        ok_merge = false;
+      }
+    } else {  // y is continuous
+      if (debug_report)
+        printf("DuplicateColumn::checkMerge: x-integer; y-continuous\n");
+      // Scale must be at least 1/(y_u-y_l) in magnitude
+      if (y_len == 0) {
+        if (debug_report)
+          printf(
+              "%sDuplicateColumn::checkMerge: scale = %g is too small in "
+              "magnitude, as y is [%g, %g]\n",
+              newline.c_str(), scale, y_lo, y_up);
+        newline = "";
+        ok_merge = false;
+      } else {
+        double scale_limit = 1 / y_len;
+        if (abs_scale < scale_limit) {
+          if (debug_report)
+            printf(
+                "%sDuplicateColumn::checkMerge: scale = %g, but |scale| must "
+                "be "
+                "at least %g since y is [%g, %g]\n",
+                newline.c_str(), scale, scale_limit, y_lo, y_up);
+          newline = "";
+          ok_merge = false;
+        }
+      }
+    }
+  } else {
+    if (y_int) {
+      if (debug_report)
+        printf("DuplicateColumn::checkMerge: x-continuous; y-integer\n");
+      // Scale must be at most (x_u-x_l) in magnitude
+      double scale_limit = x_len;
+      if (abs_scale > scale_limit) {
+        if (debug_report)
+          printf(
+              "%sDuplicateColumn::checkMerge: scale = %g, but |scale| must "
+              "be "
+              "at "
+              "most %g since x is [%g, %g]\n",
+              newline.c_str(), scale, scale_limit, x_lo, x_up);
+        newline = "";
+        ok_merge = false;
+      }
+    } else {
+      // x and y are continuous
+      //	if (debug_report) printf("DuplicateColumn::checkMerge:
+      // x-continuous ;
+      // y-continuous\n");
+    }
+  }
+  return ok_merge;
+}
+
+void HighsPostsolveStack::DuplicateColumn::undoFix(
+    const HighsOptions& options, HighsSolution& solution,
+    const double mergeValue) const {
+  const double mip_feasibility_tolerance = options.mip_feasibility_tolerance;
+  const double primal_feasibility_tolerance =
+      options.primal_feasibility_tolerance;
+  std::vector<double>& col_value = solution.col_value;
+  const bool allow_assert = false;
+  const bool debug_report = false;
+  //=============================================================================================
+
+  auto isInteger = [&](const double v) {
+    return (fractionality(v) <= mip_feasibility_tolerance);
+  };
+
+  auto isFeasible = [&](const double l, const double v, const double u) {
+    return v >= l - primal_feasibility_tolerance &&
+           v <= u + primal_feasibility_tolerance;
+  };
+  const double scale = colScale;
+  const bool x_int = colIntegral;
+  const bool y_int = duplicateColIntegral;
+  const int x_ix = col;
+  const int y_ix = duplicateCol;
+  const double x_lo =
+      x_int ? std::ceil(colLower - mip_feasibility_tolerance) : colLower;
+  const double x_up =
+      x_int ? std::floor(colUpper + mip_feasibility_tolerance) : colUpper;
+  const double y_lo =
+      y_int ? std::ceil(duplicateColLower - mip_feasibility_tolerance)
+            : duplicateColLower;
+  const double y_up =
+      y_int ? std::floor(duplicateColUpper + mip_feasibility_tolerance)
+            : duplicateColUpper;
+  if (kAllowDeveloperAssert) assert(scale);
+  double x_v = mergeValue;
+  double y_v = -kHighsInf;
+
+  auto checkIntVar = [](double z_lo, double z_up, double& z_0, double& z_d,
+                        double& z_1, bool& z_free) {
+    const double value_max = 1000;
+    z_0 = 0;
+    z_d = 0;
+    z_1 = 0;
+    z_free = false;
+    if (z_lo <= -kHighsInf) {
+      if (z_up >= kHighsInf) {
+        // z is free
+        z_free = true;
+        z_0 = 0;
+        z_d = 1.0;
+        z_1 = value_max;
+      } else {
+        // z is (-int, u]
+        z_0 = z_up;
+        z_d = -1.0;
+        z_1 = -value_max;
+      }
+    } else {
+      // z is [l, u] or [l, inf) respectively
+      z_0 = z_lo;
+      z_d = 1.0;
+      z_1 = z_up >= kHighsInf ? value_max : z_up;
+    }
+  };
+
+  auto computeValue = [mergeValue, scale](double value) {
+    return static_cast<double>((static_cast<HighsCDouble>(mergeValue) - value) /
+                               scale);
+  };
+
+  auto computeInvValue = [mergeValue, scale](double value) {
+    return static_cast<double>(static_cast<HighsCDouble>(mergeValue) -
+                               static_cast<HighsCDouble>(value) * scale);
+  };
+
+  auto findValue = [&](double& z_value, double z_0, double z_1, double z_delta,
+                       double& other_value, double other_lower,
+                       double other_upper, bool other_int) {
+    const double eps = 1e-8;
+    for (z_value = z_0;; z_value += z_delta) {
+      other_value = computeValue(z_value);
+      if (isFeasible(other_lower, other_value, other_upper) &&
+          (!other_int || isInteger(other_value)))
+        return true;
+      if (z_delta > 0 && z_value + z_delta >= z_1 + eps) return false;
+      if (z_delta < 0 && z_value + z_delta <= z_1 - eps) return false;
+    }
+    return false;
+  };
+
+  auto setXValue = [&](double value) {
+    x_v = value;
+    y_v = computeValue(value);
+  };
+
+  auto setYValue = [&](double value) {
+    y_v = value;
+    x_v = computeInvValue(value);
+  };
+
+  if (x_int) {
+    double x_0;
+    double x_d;
+    double x_1;
+    bool x_free;
+    checkIntVar(x_lo, x_up, x_0, x_d, x_1, x_free);
+    // x is integer, so look through its possible values to find a
+    // suitable y
+    if (x_free && debug_report) printf("DuplicateColumn::undo x is free\n");
+    if (debug_report)
+      printf("DuplicateColumn::undo Using x (%g; %g; %g)\n", x_0, x_d, x_1);
+    bool found_y = findValue(x_v, x_0, x_1, x_d, y_v, y_lo, y_up, y_int);
+    if (allow_assert) assert(found_y);
+  } else if (y_int) {
+    double y_0;
+    double y_d;
+    double y_1;
+    bool y_free;
+    checkIntVar(y_lo, y_up, y_0, y_d, y_1, y_free);
+    // y is integer, so look through its possible values to find a
+    // suitable x
+    if (y_free && debug_report) printf("DuplicateColumn::undo y is free\n");
+    if (debug_report)
+      printf("DuplicateColumn::undo Using y (%g; %g; %g)\n", y_0, y_d, y_1);
+    bool found_x = findValue(y_v, y_0, y_1, y_d, x_v, x_lo, x_up, x_int);
+    if (allow_assert) assert(found_x);
+  } else {
+    // x and y are both continuous
+    double v_m_a_ylo = 0;
+    double v_m_a_yup = 0;
+    if (y_lo <= -kHighsInf) {
+      v_m_a_ylo = scale > 0 ? kHighsInf : -kHighsInf;
+    } else {
+      v_m_a_ylo = computeInvValue(y_lo);
+    }
+    if (y_up >= kHighsInf) {
+      v_m_a_yup = scale > 0 ? -kHighsInf : kHighsInf;
+    } else {
+      v_m_a_yup = computeInvValue(y_up);
+    }
+    // Need to ensure that y puts x in [x_l, x_u]
+    if (scale > 0) {
+      if (debug_report)
+        printf("DuplicateColumn::undo [V-a(y_u), V-a(y_l)] == [%g, %g]\n",
+               v_m_a_yup, v_m_a_ylo);
+      // V-ay is in [V-a(y_u), V-a(y_l)] == [v_m_a_yup, v_m_a_ylo]
+      if (y_up < kHighsInf) {
+        // If v_m_a_yup is right of x_up+eps then [v_m_a_yup, v_m_a_ylo] is
+        // right of [x_lo-eps, x_up+eps] so there's no solution. [Could
+        // try v_m_a_ylo computed from y_lo-eps.]
+        if (kAllowDeveloperAssert)
+          assert(x_up + primal_feasibility_tolerance >= v_m_a_yup);
+        // This assignment is OK unless x_v < x_lo-eps
+        setYValue(y_up);
+        if (x_v < x_lo - primal_feasibility_tolerance) {
+          // Try y_v corresponding to x_lo
+          setXValue(x_lo);
+          if (y_v < y_lo - primal_feasibility_tolerance) {
+            // Very tight: use x_v on its margin and hope!
+            setXValue(x_lo - primal_feasibility_tolerance);
+          }
+        }
+      } else if (y_lo > -kHighsInf) {
+        // If v_m_a_ylo is left of x_lo-eps then [v_m_a_yup, v_m_a_ylo] is
+        // left of [x_lo-eps, x_up+eps] so there's no solution. [Could
+        // try v_m_a_yup computed from y_up+eps.]
+        if (kAllowDeveloperAssert)
+          assert(x_lo - primal_feasibility_tolerance <= v_m_a_ylo);
+        // This assignment is OK unless x_v > x_up-eps
+        setYValue(y_lo);
+        if (x_v > x_up + primal_feasibility_tolerance) {
+          // Try y_v corresponding to x_up
+          setXValue(x_up);
+          if (y_v > y_up + primal_feasibility_tolerance) {
+            // Very tight: use x_v on its margin and hope!
+            setXValue(x_up + primal_feasibility_tolerance);
+          }
+        }
+      } else {
+        // y is free, so use x_v = max(0, x_lo)
+        setXValue(std::max(0.0, x_lo));
+      }
+    } else {  // scale < 0
+      if (debug_report)
+        printf("DuplicateColumn::undo [V-a(y_l), V-a(y_u)] == [%g, %g]\n",
+               v_m_a_ylo, v_m_a_yup);
+      // V-ay is in [V-a(y_l), V-a(y_u)] == [v_m_a_ylo, v_m_a_yup]
+      if (y_lo > -kHighsInf) {
+        // If v_m_a_ylo is right of x_up+eps then [v_m_a_ylo, v_m_a_yup] is
+        // right of [x_lo-eps, x_up+eps] so there's no solution. [Could
+        // try v_m_a_ylo computed from y_up+eps.]
+        if (kAllowDeveloperAssert)
+          assert(x_up + primal_feasibility_tolerance >= v_m_a_ylo);
+        // This assignment is OK unless x_v < x_lo-eps
+        setYValue(y_lo);
+        if (x_v < x_lo - primal_feasibility_tolerance) {
+          // Try y_v corresponding to x_lo
+          setXValue(x_lo);
+          if (y_v > y_up + primal_feasibility_tolerance) {
+            // Very tight: use x_v on its margin and hope!
+            setXValue(x_lo - primal_feasibility_tolerance);
+          }
+        }
+      } else if (y_up < kHighsInf) {
+        // If v_m_a_yup is left of x_lo-eps then [v_m_a_ylo, v_m_a_yup] is
+        // left of [x_lo-eps, x_up+eps] so there's no solution. [Could
+        // try v_m_a_yup computed from y_lo-eps.]
+        if (kAllowDeveloperAssert)
+          assert(x_lo - primal_feasibility_tolerance <= v_m_a_yup);
+        // This assignment is OK unless x_v < x_lo-eps
+        setYValue(y_up);
+        if (x_v > x_up + primal_feasibility_tolerance) {
+          // Try y_v corresponding to x_up
+          setXValue(x_up);
+          if (y_v < y_lo - primal_feasibility_tolerance) {
+            // Very tight: use x_v on its margin and hope!
+            setXValue(x_up + primal_feasibility_tolerance);
+          }
+        }
+      } else {
+        // y is free, so use x_v = max(0, x_lo)
+        setXValue(std::max(0.0, x_lo));
+      }
+    }
+  }
+  const double residual_tolerance = 1e-12;
+  double residual = std::fabs(static_cast<double>(
+      static_cast<HighsCDouble>(x_v) - computeInvValue(y_v)));
+  const bool x_y_ok =
+      isFeasible(x_lo, x_v, x_up) && isFeasible(y_lo, y_v, y_up) &&
+      (!x_int || isInteger(x_v)) && (!y_int || isInteger(y_v)) &&
+      (std::fabs(x_v) < kHighsInf) && (std::fabs(y_v) < kHighsInf) &&
+      (residual <= residual_tolerance);
+
+  bool check;
+  check = isFeasible(x_lo, x_v, x_up);
+  if (!check) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: isFeasible(x_lo, x_v, x_up) is "
+          "false\n");
+    if (allow_assert) assert(check);
+  }
+  check = isFeasible(y_lo, y_v, y_up);
+  if (!check) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: isFeasible(y_lo, y_v, y_up) is "
+          "false\n");
+    if (allow_assert) assert(check);
+  }
+  check = !x_int || isInteger(x_v);
+  if (!check) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: !x_int || isInteger(x_v) is false\n");
+    if (allow_assert) assert(check);
+  }
+  check = !y_int || isInteger(y_v);
+  if (!check) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: !y_int || isInteger(y_v) is false\n");
+    if (allow_assert) assert(check);
+  }
+  check = std::fabs(x_v) < kHighsInf;
+  if (!check) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: std::fabs(x_v) < kHighsInf is "
+          "false\n");
+    if (allow_assert) assert(check);
+  }
+  check = std::fabs(y_v) < kHighsInf;
+  if (!check) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: std::fabs(y_v) < kHighsInf is "
+          "false\n");
+    if (allow_assert) assert(check);
+  }
+  check = residual <= residual_tolerance;
+  if (!check) {
+    if (debug_report)
+      printf(
+          "DuplicateColumn::undo error: residual <= residual_tolerance is "
+          "false\n");
+    if (allow_assert) assert(check);
+  }
+  check = residual <= residual_tolerance;
+  if (debug_report)
+    printf("DuplicateColumn::undo%s x = %g; y = %g to give x + (%g)y = %g",
+           x_y_ok ? "" : " ERROR", x_v, y_v, scale, mergeValue);
+  if (x_y_ok) {
+    if (debug_report) printf(": FIXED\n");
+  } else if (check) {
+    if (debug_report) printf("\n");
+  } else {
+    if (debug_report) printf(": residual = %g\n", residual);
+  }
+  //=============================================================================================
+  if (x_y_ok) {
+    col_value[x_ix] = x_v;
+    col_value[y_ix] = y_v;
+  }
+}
+
+void HighsPostsolveStack::DuplicateColumn::transformToPresolvedSpace(
+    std::vector<double>& primalSol) const {
+  primalSol[col] = primalSol[col] + colScale * primalSol[duplicateCol];
+}
+
+void HighsPostsolveStack::SlackColSubstitution::undo(
+    const HighsOptions& options, const std::vector<Nonzero>& rowValues,
+    HighsSolution& solution, HighsBasis& basis) {
+  bool debug_print = false;
+  // May have to determine row dual and basis status unless doing
+  // primal-only transformation in MIP solver, in which case row may
+  // no longer exist if it corresponds to a removed cut, so have to
+  // avoid exceeding array bounds of solution.row_value
+  bool isModelRow = static_cast<size_t>(row) < solution.row_value.size();
+
+  // compute primal values
+  double colCoef = 0;
+  HighsCDouble rowValue = 0;
+  for (const auto& rowVal : rowValues) {
+    if (rowVal.index == col)
+      colCoef = rowVal.value;
+    else
+      rowValue += rowVal.value * solution.col_value[rowVal.index];
+  }
+
+  assert(colCoef != 0);
+  // Row values aren't fully postsolved, so why do this?
+  if (isModelRow)
+    solution.row_value[row] =
+        double(rowValue + colCoef * solution.col_value[col]);
+
+  solution.col_value[col] = double((rhs - rowValue) / colCoef);
+
+  // If no dual values requested, return here
+  if (!solution.dual_valid) return;
+
+  // Row retains its dual value, and column has this dual value scaled by coeff
+  if (isModelRow) solution.col_dual[col] = -solution.row_dual[row] / colCoef;
+
+  // Set basis status if necessary
+  if (!basis.valid) return;
+
+  // If row is basic, then slack is basic, otherwise row retains its status
+  if (isModelRow) {
+    HighsBasisStatus save_row_basis_status = basis.row_status[row];
+    if (basis.row_status[row] == HighsBasisStatus::kBasic) {
+      basis.col_status[col] = HighsBasisStatus::kBasic;
+      basis.row_status[row] =
+          computeRowStatus(solution.row_dual[row], RowType::kEq);
+    } else if (basis.row_status[row] == HighsBasisStatus::kLower) {
+      basis.col_status[col] =
+          colCoef > 0 ? HighsBasisStatus::kUpper : HighsBasisStatus::kLower;
+    } else {
+      basis.col_status[col] =
+          colCoef > 0 ? HighsBasisStatus::kLower : HighsBasisStatus::kUpper;
+    }
+    if (debug_print)
+      printf(
+          "HighsPostsolveStack::SlackColSubstitution::undo OgRowStatus = %s; "
+          "RowStatus = %s; ColStatus = %s\n",
+          utilBasisStatusToString(save_row_basis_status).c_str(),
+          utilBasisStatusToString(basis.row_status[row]).c_str(),
+          utilBasisStatusToString(basis.col_status[col]).c_str());
+    if (basis.col_status[col] == HighsBasisStatus::kLower) {
+      assert(solution.col_dual[col] > -options.dual_feasibility_tolerance);
+    } else if (basis.col_status[col] == HighsBasisStatus::kUpper) {
+      assert(solution.col_dual[col] < options.dual_feasibility_tolerance);
+    }
+  } else {
+    basis.col_status[col] = HighsBasisStatus::kNonbasic;
+  }
+}
+
+}  // namespace presolve