lpsolver 0.1.0 → 0.2.1

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

@@ -0,0 +1,1921 @@
+/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
+/*                                                                       */
+/*    This file is part of the HiGHS linear optimization suite           */
+/*                                                                       */
+/*    Available as open-source under the MIT License                     */
+/*                                                                       */
+/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
+/**@file HighsSymmetry.cpp
+ * @brief Facilities for symmetry detection
+ * @author Leona Gottwald
+ */
+
+#include "presolve/HighsSymmetry.h"
+
+#include <algorithm>
+#include <numeric>
+
+#include "../extern/pdqsort/pdqsort.h"
+#include "mip/HighsCliqueTable.h"
+#include "mip/HighsDomain.h"
+#include "parallel/HighsParallel.h"
+#include "util/HighsDisjointSets.h"
+
+void HighsSymmetryDetection::removeFixPoints() {
+  Gend.resize(numVertices);
+  for (HighsInt i = 0; i < numVertices; ++i) {
+    Gend[i] =
+        std::partition(Gedge.begin() + Gstart[i], Gedge.begin() + Gstart[i + 1],
+                       [&](const std::pair<HighsInt, HighsUInt>& edge) {
+                         return cellSize(vertexToCell[edge.first]) > 1;
+                       }) -
+        Gedge.begin();
+    assert(Gend[i] >= Gstart[i] && Gend[i] <= Gstart[i + 1]);
+  }
+
+  HighsInt unitCellIndex = numVertices;
+  currentPartition.erase(
+      std::remove_if(currentPartition.begin(), currentPartition.end(),
+                     [&](HighsInt vertex) {
+                       if (cellSize(vertexToCell[vertex]) == 1) {
+                         --unitCellIndex;
+                         vertexToCell[vertex] = unitCellIndex;
+                         return true;
+                       }
+                       return false;
+                     }),
+      currentPartition.end());
+
+  for (HighsInt i = 0; i < numVertices; ++i) {
+    if (Gend[i] == Gstart[i + 1]) continue;
+
+    for (HighsInt j = Gend[i]; j < Gstart[i + 1]; ++j)
+      Gedge[j].first = vertexToCell[Gedge[j].first];
+  }
+
+  if ((HighsInt)currentPartition.size() < numVertices) {
+    numVertices = currentPartition.size();
+    if (numVertices == 0) {
+      numActiveCols = 0;
+      return;
+    }
+    currentPartitionLinks.resize(numVertices);
+    cellInRefinementQueue.assign(numVertices, false);
+    assert(refinementQueue.empty());
+    refinementQueue.clear();
+    HighsInt cellStart = 0;
+    HighsInt cellNumber = 0;
+    for (HighsInt i = 0; i < numVertices; ++i) {
+      HighsInt vertex = currentPartition[i];
+      // if the cell number is different to the current cell number this is the
+      // start of a new cell
+      if (cellNumber != vertexToCell[vertex]) {
+        // remember the number of this cell to identify its end
+        cellNumber = vertexToCell[vertex];
+        // set the link of the cell start to point to its end
+        currentPartitionLinks[cellStart] = i;
+        // remember start of this cell
+        cellStart = i;
+      }
+
+      // correct the vertexToCell array to store the start index of the
+      // cell, not its number
+      updateCellMembership(i, cellStart, false);
+    }
+
+    // set the column partition link of the last started cell to point past the
+    // end
+    assert((int)currentPartitionLinks.size() > 0);
+    currentPartitionLinks[cellStart] = numVertices;
+
+    numActiveCols =
+        std::partition_point(currentPartition.begin(), currentPartition.end(),
+                             [&](HighsInt v) { return v < numCol; }) -
+        currentPartition.begin();
+  } else
+    numActiveCols = numCol;
+}
+
+void HighsSymmetries::clear() {
+  permutationColumns.clear();
+  permutations.clear();
+  orbitPartition.clear();
+  orbitSize.clear();
+  columnPosition.clear();
+  linkCompressionStack.clear();
+  columnToOrbitope.clear();
+  orbitopes.clear();
+  numPerms = 0;
+  numGenerators = 0;
+}
+
+void HighsSymmetries::mergeOrbits(HighsInt v1, HighsInt v2) {
+  if (v1 == v2) return;
+
+  HighsInt orbit1 = getOrbit(v1);
+  HighsInt orbit2 = getOrbit(v2);
+
+  if (orbit1 == orbit2) return;
+
+  if (orbitSize[orbit2] < orbitSize[orbit1]) {
+    orbitPartition[orbit2] = orbit1;
+    orbitSize[orbit1] += orbitSize[orbit2];
+  } else {
+    orbitPartition[orbit1] = orbit2;
+    orbitSize[orbit2] += orbitSize[orbit1];
+  }
+
+  return;
+}
+
+HighsInt HighsSymmetries::getOrbit(HighsInt col) {
+  HighsInt i = columnPosition[col];
+  if (i == -1) return -1;
+  HighsInt orbit = orbitPartition[i];
+  if (orbit != orbitPartition[orbit]) {
+    do {
+      linkCompressionStack.push_back(i);
+      i = orbit;
+      orbit = orbitPartition[orbit];
+    } while (orbit != orbitPartition[orbit]);
+
+    do {
+      i = linkCompressionStack.back();
+      linkCompressionStack.pop_back();
+      orbitPartition[i] = orbit;
+    } while (!linkCompressionStack.empty());
+  }
+
+  return orbit;
+}
+
+HighsInt HighsSymmetries::propagateOrbitopes(HighsDomain& domain) const {
+  if (columnToOrbitope.size() == 0 || domain.getBranchDepth() == 0) return 0;
+
+  std::set<HighsInt> propagationOrbitopes;
+  const auto& domchgstack = domain.getDomainChangeStack();
+  for (HighsInt pos : domain.getBranchingPositions()) {
+    const HighsInt* orbitope = columnToOrbitope.find(domchgstack[pos].column);
+    if (orbitope) propagationOrbitopes.insert(*orbitope);
+  }
+
+  HighsInt numFixed = 0;
+  for (HighsInt i : propagationOrbitopes) {
+    numFixed += orbitopes[i].orbitalFixing(domain);
+    if (domain.infeasible()) break;
+  }
+
+  // if (numFixed)
+  //   printf("orbital fixing for full orbitope found %d fixings\n", numFixed);
+
+  return numFixed;
+}
+
+std::shared_ptr<const StabilizerOrbits>
+HighsSymmetries::computeStabilizerOrbits(const HighsDomain& localdom) {
+  const auto& domchgStack = localdom.getDomainChangeStack();
+  const auto& branchingPos = localdom.getBranchingPositions();
+
+  StabilizerOrbits stabilizerOrbits;
+  stabilizerOrbits.stabilizedCols.reserve(permutationColumns.size());
+  for (HighsInt i : branchingPos) {
+    HighsInt col = domchgStack[i].column;
+    if (columnPosition[col] == -1) continue;
+
+    assert(localdom.variableType(col) != HighsVarType::kContinuous);
+
+    // if we branch a variable upwards it is either binary and branched to one
+    // and needs to be stabilized or it is a general integer and needs to be
+    // stabilized regardless of branching direction.
+    // if we branch downwards we only need to stabilize on this
+    // branching column if it is a general integer
+    if (domchgStack[i].boundtype == HighsBoundType::kLower ||
+        !localdom.isGlobalBinary(col))
+      stabilizerOrbits.stabilizedCols.push_back(columnPosition[col]);
+  }
+
+  HighsInt permLength = permutationColumns.size();
+  orbitPartition.resize(permLength);
+  std::iota(orbitPartition.begin(), orbitPartition.end(), 0);
+  orbitSize.assign(permLength, 1);
+
+  for (HighsInt i = 0; i < numPerms; ++i) {
+    const HighsInt* perm = permutations.data() + i * permutationColumns.size();
+
+    bool permRespectsBranchings = true;
+    for (HighsInt i : stabilizerOrbits.stabilizedCols) {
+      if (permutationColumns[i] != perm[i]) {
+        permRespectsBranchings = false;
+        break;
+      }
+    }
+
+    if (!permRespectsBranchings) continue;
+
+    for (HighsInt j = 0; j < permLength; ++j) {
+      mergeOrbits(permutationColumns[j], perm[j]);
+    }
+  }
+
+  stabilizerOrbits.stabilizedCols.clear();
+
+  stabilizerOrbits.orbitCols.reserve(permLength);
+  for (HighsInt i = 0; i < permLength; ++i) {
+    if (localdom.variableType(permutationColumns[i]) ==
+        HighsVarType::kContinuous)
+      continue;
+    HighsInt orbit = getOrbit(permutationColumns[i]);
+    if (orbitSize[orbit] == 1)
+      stabilizerOrbits.stabilizedCols.push_back(permutationColumns[i]);
+    else if (localdom.isGlobalBinary(permutationColumns[i]))
+      stabilizerOrbits.orbitCols.push_back(permutationColumns[i]);
+  }
+  stabilizerOrbits.symmetries = this;
+  pdqsort(stabilizerOrbits.stabilizedCols.begin(),
+          stabilizerOrbits.stabilizedCols.end());
+  if (!stabilizerOrbits.orbitCols.empty()) {
+    pdqsort(stabilizerOrbits.orbitCols.begin(),
+            stabilizerOrbits.orbitCols.end(),
+            [&](HighsInt col1, HighsInt col2) {
+              return getOrbit(col1) < getOrbit(col2);
+            });
+    HighsInt numOrbitCols = stabilizerOrbits.orbitCols.size();
+    stabilizerOrbits.orbitStarts.reserve(numOrbitCols + 1);
+    stabilizerOrbits.orbitStarts.push_back(0);
+
+    for (HighsInt i = 1; i < numOrbitCols; ++i) {
+      if (getOrbit(stabilizerOrbits.orbitCols[i]) !=
+          getOrbit(stabilizerOrbits.orbitCols[i - 1]))
+        stabilizerOrbits.orbitStarts.push_back(i);
+    }
+    stabilizerOrbits.orbitStarts.push_back(numOrbitCols);
+  }
+
+  return std::make_shared<const StabilizerOrbits>(std::move(stabilizerOrbits));
+}
+
+HighsInt StabilizerOrbits::orbitalFixing(HighsDomain& domain) const {
+  HighsInt numFixed = symmetries->propagateOrbitopes(domain);
+  if (domain.infeasible() || orbitCols.empty()) return numFixed;
+
+  HighsInt numOrbits = orbitStarts.size() - 1;
+  for (HighsInt i = 0; i < numOrbits; ++i) {
+    HighsInt fixcol = -1;
+    for (HighsInt j = orbitStarts[i]; j < orbitStarts[i + 1]; ++j) {
+      if (domain.isFixed(orbitCols[j])) {
+        fixcol = orbitCols[j];
+        break;
+      }
+    }
+
+    if (fixcol != -1) {
+      HighsInt oldNumFixed = numFixed;
+      auto oldSize = domain.getDomainChangeStack().size();
+      if (domain.col_lower_[fixcol] == 1.0) {
+        for (HighsInt j = orbitStarts[i]; j < orbitStarts[i + 1]; ++j) {
+          if (domain.col_lower_[orbitCols[j]] == 1.0) continue;
+          ++numFixed;
+          domain.changeBound(HighsBoundType::kLower, orbitCols[j], 1.0,
+                             HighsDomain::Reason::unspecified());
+          if (domain.infeasible()) return numFixed;
+        }
+      } else {
+        for (HighsInt j = orbitStarts[i]; j < orbitStarts[i + 1]; ++j) {
+          if (domain.col_upper_[orbitCols[j]] == 0.0) continue;
+          ++numFixed;
+          domain.changeBound(HighsBoundType::kUpper, orbitCols[j], 0.0,
+                             HighsDomain::Reason::unspecified());
+          if (domain.infeasible()) return numFixed;
+        }
+      }
+
+      HighsInt newFixed = numFixed - oldNumFixed;
+
+      if (newFixed != 0) {
+        domain.propagate();
+        if (domain.infeasible()) return numFixed;
+        if ((HighsInt)(domain.getDomainChangeStack().size() - oldSize) >
+            newFixed)
+          i = -1;
+      }
+    }
+  }
+
+  return numFixed;
+}
+
+bool StabilizerOrbits::isStabilized(HighsInt col) const {
+  return symmetries->columnPosition[col] == -1 ||
+         std::binary_search(stabilizedCols.begin(), stabilizedCols.end(), col);
+}
+
+void HighsOrbitopeMatrix::determineOrbitopeType(HighsCliqueTable& cliquetable) {
+  for (HighsInt j = 0; j < rowLength; ++j) {
+    for (HighsInt i = 0; i < numRows; ++i) {
+      columnToRow.insert(entry(i, j), i);
+    }
+  }
+
+  rowIsSetPacking.assign(numRows, -1);
+  numSetPackingRows = 0;
+
+  for (HighsInt j = 1; j < rowLength; ++j) {
+    HighsInt* colj1 = &entry(0, j);
+
+    for (HighsInt j0 = 0; j0 < j; ++j0) {
+      HighsInt* colj0 = &entry(0, j0);
+
+      for (HighsInt i = 0; i < numRows; ++i) {
+        if (rowIsSetPacking[i] != -1) continue;
+
+        HighsInt xj0 = colj0[i];
+        HighsInt xj1 = colj1[i];
+
+        auto commonClique = cliquetable.findCommonClique({xj0, 1}, {xj1, 1});
+
+        if (commonClique.first == nullptr) {
+          rowIsSetPacking[i] = false;
+          continue;
+        }
+
+        HighsInt overlap = 0;
+
+        for (HighsInt k = 0; k < commonClique.second; ++k) {
+          if (commonClique.first[k].val == 0) continue;
+
+          HighsInt* cliqueColRow = columnToRow.find(commonClique.first[k].col);
+          if (cliqueColRow && *cliqueColRow == i) ++overlap;
+        }
+
+        if (overlap == rowLength) {
+          rowIsSetPacking[i] = true;
+          ++numSetPackingRows;
+          if (numSetPackingRows == numRows) break;
+        }
+      }
+
+      if (numSetPackingRows == numRows) break;
+    }
+
+    if (numSetPackingRows == numRows) break;
+  }
+
+  // now for the rows that do not have a set packing structure check
+  // if we have such structure when all columns in the row are negated
+
+  for (HighsInt i = 0; i < numRows; ++i) {
+    if (!rowIsSetPacking[i]) rowIsSetPacking[i] = -1;
+  }
+
+  for (HighsInt j = 1; j < rowLength; ++j) {
+    HighsInt* colj1 = &entry(0, j);
+
+    for (HighsInt j0 = 0; j0 < j; ++j0) {
+      HighsInt* colj0 = &entry(0, j0);
+
+      for (HighsInt i = 0; i < numRows; ++i) {
+        if (rowIsSetPacking[i] != -1) continue;
+
+        HighsInt xj0 = colj0[i];
+        HighsInt xj1 = colj1[i];
+
+        // now look for cliques with value 0
+        auto commonClique = cliquetable.findCommonClique({xj0, 0}, {xj1, 0});
+
+        if (commonClique.first == nullptr) {
+          rowIsSetPacking[i] = false;
+          continue;
+        }
+
+        HighsInt overlap = 0;
+
+        for (HighsInt k = 0; k < commonClique.second; ++k) {
+          // skip clique variables with values of 1
+          if (commonClique.first[k].val == 1) continue;
+
+          HighsInt* cliqueColRow = columnToRow.find(commonClique.first[k].col);
+          if (cliqueColRow && *cliqueColRow == i) ++overlap;
+        }
+
+        if (overlap == rowLength) {
+          // mark with value 2, for negated set packing row with at most one
+          // zero
+          rowIsSetPacking[i] = 2;
+          ++numSetPackingRows;
+          if (numSetPackingRows == numRows) break;
+        }
+      }
+
+      if (numSetPackingRows == numRows) break;
+    }
+
+    if (numSetPackingRows == numRows) break;
+  }
+}
+
+HighsInt HighsOrbitopeMatrix::getBranchingColumn(
+    const std::vector<double>& colLower, const std::vector<double>& colUpper,
+    HighsInt col) const {
+  const HighsInt* i = columnToRow.find(col);
+  if (i && rowIsSetPacking[*i]) {
+    for (HighsInt j = 0; j < rowLength; ++j) {
+      HighsInt branchCol = entry(*i, j);
+      if (branchCol == col) break;
+      if (colLower[branchCol] != colUpper[branchCol]) return branchCol;
+    }
+  }
+
+  return col;
+}
+
+HighsInt HighsOrbitopeMatrix::orbitalFixingForPackingOrbitope(
+    const std::vector<HighsInt>& rows, HighsDomain& domain) const {
+  HighsInt numDynamicRows = rows.size();
+
+  // printf("propagate packing orbitope\n");
+
+  std::vector<HighsInt> firstOneInRow(numDynamicRows, -1);
+
+  for (HighsInt j = 0; j < rowLength; ++j) {
+    for (HighsInt i = 0; i < numDynamicRows; ++i) {
+      if (firstOneInRow[i] != -1) continue;
+      HighsInt r = rows[i];
+      HighsInt colrj = entry(r, j);
+      if (rowIsSetPacking[r] == 1) {
+        if (domain.col_lower_[colrj] > 0.5) firstOneInRow[i] = j;
+      } else {
+        assert(rowIsSetPacking[r] == 2);
+        if (domain.col_upper_[colrj] < 0.5) firstOneInRow[i] = j;
+      }
+    }
+  }
+
+  // we start looping over the rows and keep the index j at the column that
+  // is the right most possible location for a 1 entry.
+  // For the first row this position is 0.
+
+  HighsInt j = 0;
+  HighsInt numFixed = 0;
+  for (HighsInt i = 0; i < numDynamicRows; ++i) {
+    // at this position we know that the last possible position
+    // of a 1-entry in this row is j. If we have a 1 behind j
+    // the face of the orbitope intersecting the set of initial fixings is empty
+    if (firstOneInRow[i] > j) {
+      domain.markInfeasible();
+      // printf("packing orbitope propagation found infeasibility\n");
+      return numFixed;
+    }
+    HighsInt col_ij = entry(rows[i], j);
+
+    bool negate_i = rowIsSetPacking[rows[i]] == 2;
+
+    bool notZeroFixed = negate_i ? domain.col_lower_[col_ij] < 0.5
+                                 : domain.col_upper_[col_ij] > 0.5;
+
+    // as long as the entry is fixed to zero
+    // the frontier stays at the same column
+    // if we encounter an entry that is not fixed to zero
+    // we need to proceed with the next column and found a frontier step
+    if (notZeroFixed) {
+      // found a frontier step. Now we first check for the current column
+      // if it can be fixed to 1.
+      // For this we check if we find an infeasibility in the case where this
+      // column was fixed to zero in which case the frontier would have stayed
+      // at j for the following rows.
+      HighsInt j0 = j;
+      for (HighsInt k = i + 1; k < numDynamicRows; ++k) {
+        if (firstOneInRow[k] > j0) {
+          if (negate_i)
+            domain.changeBound(HighsBoundType::kUpper, col_ij, 0.0,
+                               HighsDomain::Reason::unspecified());
+          else
+            domain.changeBound(HighsBoundType::kLower, col_ij, 1.0,
+                               HighsDomain::Reason::unspecified());
+
+          ++numFixed;
+          if (domain.infeasible()) {
+            // printf("packing orbitope propagation found infeasibility\n");
+            return numFixed;
+          }
+          break;
+        }
+
+        bool negate_k = rowIsSetPacking[rows[k]] == 2;
+        bool notZeroFixed = negate_k
+                                ? domain.col_lower_[entry(rows[k], j0)] < 0.5
+                                : domain.col_upper_[entry(rows[k], j0)] > 0.5;
+
+        if (notZeroFixed) {
+          ++j0;
+          if (j0 == rowLength) break;
+        }
+      }
+
+      ++j;
+      if (j == rowLength) break;
+
+      for (HighsInt k = 0; k <= i; ++k) {
+        // we should have checked this row at the frontier position
+        assert(firstOneInRow[k] < j);
+
+        HighsInt col_kj = entry(rows[k], j);
+        bool negate_k = rowIsSetPacking[rows[k]] == 2;
+
+        if (negate_k) {
+          if (domain.col_lower_[col_kj] > 0.5) continue;
+
+          domain.changeBound(HighsBoundType::kLower, col_kj, 1.0,
+                             HighsDomain::Reason::unspecified());
+        } else {
+          if (domain.col_upper_[col_kj] < 0.5) continue;
+
+          domain.changeBound(HighsBoundType::kUpper, col_kj, 0.0,
+                             HighsDomain::Reason::unspecified());
+        }
+
+        ++numFixed;
+        if (domain.infeasible()) {
+          // this can happen due to deductions from earlier fixings
+          // otherwise it would have been caught by the infeasibility
+          // check within the next loop that goes over i
+          // printf("packing orbitope propagation found infeasibility\n");
+          return numFixed;
+        }
+      }
+    }
+  }
+
+  // check if there are more columns that can be completely fixed to zero
+  while (++j < rowLength) {
+    for (HighsInt k = 0; k < numDynamicRows; ++k) {
+      // we should have checked this row at the frontier position
+      assert(firstOneInRow[k] < j);
+
+      HighsInt col_kj = entry(rows[k], j);
+      bool negate_k = rowIsSetPacking[rows[k]] == 2;
+
+      if (negate_k) {
+        if (domain.col_lower_[col_kj] > 0.5) continue;
+
+        domain.changeBound(HighsBoundType::kLower, col_kj, 1.0,
+                           HighsDomain::Reason::unspecified());
+      } else {
+        if (domain.col_upper_[col_kj] < 0.5) continue;
+
+        domain.changeBound(HighsBoundType::kUpper, col_kj, 0.0,
+                           HighsDomain::Reason::unspecified());
+      }
+      // printf("new fixed\n");
+      ++numFixed;
+      if (domain.infeasible()) {
+        // this can happen due to deductions from earlier fixings
+        // otherwise it would have been caught by the infeasibility
+        // check within the next loop that goes over i
+        // printf("packing orbitope propagation found infeasibility\n");
+        return numFixed;
+      }
+    }
+  }
+
+  if (!domain.infeasible() && numFixed) domain.propagate();
+
+  // if (numFixed)
+  //  printf("orbital fixing for packing case fixed %d columns\n", numFixed);
+
+  return numFixed;
+}
+
+HighsInt HighsOrbitopeMatrix::orbitalFixingForFullOrbitope(
+    const std::vector<HighsInt>& rows, HighsDomain& domain) const {
+  HighsInt numDynamicRows = rows.size();
+  std::vector<int8_t> Mminimal(numDynamicRows * rowLength, -1);
+
+  for (HighsInt j = 0; j < rowLength; ++j) {
+    for (HighsInt i = 0; i < numDynamicRows; ++i) {
+      HighsInt r = rows[i];
+      HighsInt colij = matrix[r + j * numRows];
+      if (domain.col_lower_[colij] == 1.0)
+        Mminimal[i + j * numDynamicRows] = 1;
+      else if (domain.col_upper_[colij] == 0.0)
+        Mminimal[i + j * numDynamicRows] = 0;
+    }
+  }
+
+  std::vector<int8_t> Mmaximal = Mminimal;
+
+  int8_t* MminimaljLast = Mminimal.data() + numDynamicRows * (rowLength - 1);
+  int8_t* MmaximaljFirst = Mmaximal.data();
+  for (HighsInt k = 0; k < numDynamicRows; ++k) {
+    if (MminimaljLast[k] == -1) MminimaljLast[k] = 0;
+    if (MmaximaljFirst[k] == -1) MmaximaljFirst[k] = 1;
+  }
+
+  auto i_fixed = [&](const int8_t* colj0, const int8_t* colj1) {
+    for (HighsInt i = 0; i < numDynamicRows; ++i) {
+      if (colj0[i] != -1 && colj1[i] != -1 && colj0[i] != colj1[i]) return i;
+    }
+
+    return numDynamicRows;
+  };
+
+  auto i_discr = [&](const int8_t* colj0, const int8_t* colj1, HighsInt i_f) {
+    for (HighsInt i = i_f; i >= 0; --i) {
+      if (colj0[i] != 0 && colj1[i] != 1) return i;
+    }
+
+    return HighsInt{-1};
+  };
+
+  for (HighsInt j = rowLength - 2; j >= 0; --j) {
+    int8_t* colj0 = Mminimal.data() + j * numDynamicRows;
+    int8_t* colj1 = colj0 + numDynamicRows;
+    HighsInt i_f = i_fixed(colj0, colj1);
+
+    if (i_f == numDynamicRows) {
+      for (HighsInt k = 0; k < numDynamicRows; ++k) {
+        int8_t isFree = (colj0[k] == -1);
+        colj0[k] += (isFree & colj1[k]) + isFree;
+      }
+    } else {
+      HighsInt i_d = i_discr(colj0, colj1, i_f);
+      if (i_d == -1) {
+        domain.markInfeasible();
+        return 0;
+      }
+
+      for (HighsInt k = 0; k < i_d; ++k) {
+        int8_t isFree = (colj0[k] == -1);
+        colj0[k] += (isFree & colj1[k]) + isFree;
+      }
+      colj0[i_d] = 1;
+      for (HighsInt k = i_d + 1; k < numDynamicRows; ++k)
+        colj0[k] += (colj0[k] == -1);
+    }
+  }
+
+  for (HighsInt j = 1; j < rowLength; ++j) {
+    int8_t* colj0 = Mmaximal.data() + j * numDynamicRows;
+    int8_t* colj1 = colj0 - numDynamicRows;
+    HighsInt i_f = i_fixed(colj1, colj0);
+
+    if (i_f == numDynamicRows) {
+      for (HighsInt k = 0; k < numDynamicRows; ++k) {
+        int8_t isFree = (colj0[k] == -1);
+        colj0[k] += (isFree & colj1[k]) + isFree;
+      }
+    } else {
+      HighsInt i_d = i_discr(colj1, colj0, i_f);
+      if (i_d == -1) {
+        domain.markInfeasible();
+        return 0;
+      }
+      for (HighsInt k = 0; k < i_d; ++k) {
+        int8_t isFree = (colj0[k] == -1);
+        colj0[k] += (isFree & colj1[k]) + isFree;
+      }
+      colj0[i_d] = 0;
+      for (HighsInt k = i_d + 1; k < numDynamicRows; ++k)
+        colj0[k] += 2 * (colj0[k] == -1);
+    }
+  }
+
+  HighsInt numFixed = 0;
+
+  for (HighsInt j = 0; j < rowLength; ++j) {
+    const int8_t* colMaximal = Mmaximal.data() + j * numDynamicRows;
+    const int8_t* colMinimal = Mminimal.data() + j * numDynamicRows;
+
+    for (HighsInt i = 0; i < numDynamicRows; ++i) {
+      if (colMinimal[i] != colMaximal[i]) {
+        assert(colMinimal[i] < colMaximal[i]);
+        break;
+      }
+
+      HighsInt r = rows[i];
+      HighsInt colrj = matrix[r + j * numRows];
+      if (domain.isFixed(colrj)) continue;
+
+      ++numFixed;
+      if (colMinimal[i] == 1)
+        domain.changeBound(HighsBoundType::kLower, colrj, 1.0,
+                           HighsDomain::Reason::unspecified());
+      else
+        domain.changeBound(HighsBoundType::kUpper, colrj, 0.0,
+                           HighsDomain::Reason::unspecified());
+      if (domain.infeasible()) break;
+    }
+    if (domain.infeasible()) break;
+  }
+
+  if (!domain.infeasible()) domain.propagate();
+
+  return numFixed;
+}
+
+HighsInt HighsOrbitopeMatrix::orbitalFixing(HighsDomain& domain) const {
+  std::vector<HighsInt> rows;
+  std::vector<uint8_t> rowUsed(numRows);
+
+  rows.reserve(numRows);
+
+  const auto& branchpos = domain.getBranchingPositions();
+  const auto& domchgstack = domain.getDomainChangeStack();
+
+  bool isPacking = true;
+  for (HighsInt pos : branchpos) {
+    const HighsInt* i = columnToRow.find(domchgstack[pos].column);
+    if (i && !rowUsed[*i]) {
+      rowUsed[*i] = true;
+      isPacking = isPacking && rowIsSetPacking[*i] != 0;
+      rows.push_back(*i);
+    }
+  }
+
+  if (rows.empty()) return 0;
+
+  if (isPacking) return orbitalFixingForPackingOrbitope(rows, domain);
+
+  return orbitalFixingForFullOrbitope(rows, domain);
+}
+
+void HighsSymmetryDetection::initializeGroundSet() {
+  vertexGroundSet = currentPartition;
+  pdqsort(vertexGroundSet.begin(), vertexGroundSet.end());
+  vertexPosition.resize(vertexToCell.size(), -1);
+  for (HighsInt i = 0; i < numVertices; ++i)
+    vertexPosition[vertexGroundSet[i]] = i;
+
+  orbitPartition.resize(numVertices);
+  std::iota(orbitPartition.begin(), orbitPartition.end(), 0);
+  orbitSize.assign(numVertices, 1);
+
+  automorphisms.resize(numVertices * 64);
+  numAutomorphisms = 0;
+  currNodeCertificate.reserve(numVertices);
+}
+
+bool HighsSymmetryDetection::mergeOrbits(HighsInt v1, HighsInt v2) {
+  if (v1 == v2) return false;
+
+  HighsInt orbit1 = getOrbit(v1);
+  HighsInt orbit2 = getOrbit(v2);
+
+  if (orbit1 == orbit2) return false;
+
+  if (orbit1 < orbit2) {
+    orbitPartition[orbit2] = orbit1;
+    orbitSize[orbit1] += orbitSize[orbit2];
+  } else {
+    orbitPartition[orbit1] = orbit2;
+    orbitSize[orbit2] += orbitSize[orbit1];
+  }
+
+  return true;
+}
+
+HighsInt HighsSymmetryDetection::getOrbit(HighsInt vertex) {
+  HighsInt i = vertexPosition[vertex];
+  HighsInt orbit = orbitPartition[i];
+  if (orbit != orbitPartition[orbit]) {
+    do {
+      linkCompressionStack.push_back(i);
+      i = orbit;
+      orbit = orbitPartition[orbit];
+    } while (orbit != orbitPartition[orbit]);
+
+    do {
+      i = linkCompressionStack.back();
+      linkCompressionStack.pop_back();
+      orbitPartition[i] = orbit;
+    } while (!linkCompressionStack.empty());
+  }
+
+  return orbit;
+}
+
+void HighsSymmetryDetection::initializeHashValues() {
+  for (HighsInt i = 0; i != numVertices; ++i) {
+    HighsInt cell = vertexToCell[i];
+    for (HighsInt j = Gstart[i]; j != Gend[i]; ++j) {
+      HighsHashHelpers::sparse_combine32(vertexHash[Gedge[j].first], cell,
+                                         Gedge[j].second);
+    }
+    markCellForRefinement(cell);
+  }
+}
+
+bool HighsSymmetryDetection::updateCellMembership(HighsInt i, HighsInt cell,
+                                                  bool markForRefinement) {
+  HighsInt vertex = currentPartition[i];
+  if (vertexToCell[vertex] != cell) {
+    // set new cell id
+    vertexToCell[vertex] = cell;
+    if (i != cell) currentPartitionLinks[i] = cell;
+
+    // update hashes of affected rows
+    if (markForRefinement) {
+      for (HighsInt j = Gstart[vertex]; j != Gend[vertex]; ++j) {
+        HighsInt edgeDestinationCell = vertexToCell[Gedge[j].first];
+        if (cellSize(edgeDestinationCell) == 1) continue;
+        HighsHashHelpers::sparse_combine32(vertexHash[Gedge[j].first], cell,
+                                           Gedge[j].second);
+        markCellForRefinement(edgeDestinationCell);
+      }
+    }
+
+    return true;
+  }
+
+  return false;
+}
+
+bool HighsSymmetryDetection::splitCell(HighsInt cell, HighsInt splitPoint) {
+  u32 hSplit = getVertexHash(currentPartition[splitPoint]);
+  u32 hCell = getVertexHash(currentPartition[cell]);
+
+  u32 certificateVal =
+      (HighsHashHelpers::pair_hash<0>(hSplit, hCell) +
+       HighsHashHelpers::pair_hash<1>(
+           cell, currentPartitionLinks[cell] - splitPoint) +
+       HighsHashHelpers::pair_hash<2>(splitPoint, splitPoint - cell)) >>
+      32;
+
+  // employ prefix pruning scheme as in bliss
+  if (!firstLeaveCertificate.empty()) {
+    firstLeavePrefixLen +=
+        (firstLeavePrefixLen == (HighsInt)currNodeCertificate.size()) *
+        (certificateVal == firstLeaveCertificate[currNodeCertificate.size()]);
+    bestLeavePrefixLen +=
+        (bestLeavePrefixLen == (HighsInt)currNodeCertificate.size()) *
+        (certificateVal == bestLeaveCertificate[currNodeCertificate.size()]);
+
+    // if the node certificate is not a prefix of the first leave's certificate
+    // and it comes lexicographically after the certificate value of the
+    // lexicographically smallest leave certificate we prune the node
+    if (firstLeavePrefixLen <= (HighsInt)currNodeCertificate.size() &&
+        bestLeavePrefixLen <= (HighsInt)currNodeCertificate.size()) {
+      u32 diffVal = bestLeavePrefixLen == (HighsInt)currNodeCertificate.size()
+                        ? certificateVal
+                        : currNodeCertificate[bestLeavePrefixLen];
+      if (diffVal > bestLeaveCertificate[bestLeavePrefixLen]) return false;
+    }
+  }
+
+  currentPartitionLinks[splitPoint] = currentPartitionLinks[cell];
+  currentPartitionLinks[cell] = splitPoint;
+  cellCreationStack.push_back(splitPoint);
+  currNodeCertificate.push_back(certificateVal);
+
+  return true;
+}
+
+void HighsSymmetryDetection::markCellForRefinement(HighsInt cell) {
+  if (cellSize(cell) == 1 || cellInRefinementQueue[cell]) return;
+
+  cellInRefinementQueue[cell] = true;
+  refinementQueue.push_back(cell);
+  std::push_heap(refinementQueue.begin(), refinementQueue.end(),
+                 std::greater<HighsInt>());
+}
+
+HighsSymmetryDetection::u32 HighsSymmetryDetection::getVertexHash(HighsInt v) {
+  const u32* h = vertexHash.find(v);
+  if (h) return *h;
+  return 0;
+}
+
+bool HighsSymmetryDetection::partitionRefinement() {
+  while (!refinementQueue.empty()) {
+    std::pop_heap(refinementQueue.begin(), refinementQueue.end(),
+                  std::greater<HighsInt>());
+
+    HighsInt cellStart = refinementQueue.back();
+    HighsInt firstCellStart = cellStart;
+    refinementQueue.pop_back();
+    cellInRefinementQueue[cellStart] = false;
+
+    if (cellSize(cellStart) == 1) continue;
+    HighsInt cellEnd = currentPartitionLinks[cellStart];
+    assert(cellEnd >= cellStart);
+
+    // first check which vertices do have updated hash values and put them to
+    // the end of the partition
+    HighsInt refineStart =
+        std::partition(
+            currentPartition.begin() + cellStart,
+            currentPartition.begin() + cellEnd,
+            [&](HighsInt v) { return vertexHash.find(v) == nullptr; }) -
+        currentPartition.begin();
+
+    // if there are none there is nothing to refine
+    if (refineStart == cellEnd) continue;
+
+    // sort the vertices that have updated hash values by their hash values
+    pdqsort(currentPartition.begin() + refineStart,
+            currentPartition.begin() + cellEnd, [&](HighsInt v1, HighsInt v2) {
+              return vertexHash[v1] < vertexHash[v2];
+            });
+
+    // if not all vertices have updated hash values directly create the first
+    // new cell at the start of the range that we want to refine
+    if (refineStart != cellStart) {
+      assert(refineStart != cellStart);
+      assert(refineStart != cellEnd);
+      if (!splitCell(cellStart, refineStart)) {
+        // node can be pruned, make sure hash values are cleared and queue is
+        // empty
+        for (HighsInt c : refinementQueue) cellInRefinementQueue[c] = false;
+        refinementQueue.clear();
+        vertexHash.clear();
+        return false;
+      }
+      cellStart = refineStart;
+      updateCellMembership(cellStart, cellStart);
+    }
+
+    // now update the remaining vertices
+    bool prune = false;
+    HighsInt i;
+    assert(vertexHash.find(currentPartition[cellStart]) != nullptr);
+    // store value of first hash
+    u64 lastHash = vertexHash[currentPartition[cellStart]];
+    for (i = cellStart + 1; i < cellEnd; ++i) {
+      HighsInt vertex = currentPartition[i];
+      // get this vertex hash value
+      u64 hash = vertexHash[vertex];
+
+      if (hash != lastHash) {
+        // hash values do not match -> start of new cell
+        if (!splitCell(cellStart, i)) {
+          // refinement process yielded bad prefix of certificate
+          // -> node can be pruned
+          prune = true;
+          break;
+        }
+        cellStart = i;
+        // remember hash value of this new cell under lastHash
+        lastHash = hash;
+      }
+
+      // update membership of vertex to new cell
+      updateCellMembership(i, cellStart);
+    }
+
+    if (prune) {
+      // node can be pruned, make sure hash values are cleared and queue is
+      // empty
+      for (HighsInt c : refinementQueue) cellInRefinementQueue[c] = false;
+      refinementQueue.clear();
+      vertexHash.clear();
+      currentPartitionLinks[firstCellStart] = cellEnd;
+
+      // undo possibly incomplete changes done to the cells
+      for (--i; i >= refineStart; --i)
+        updateCellMembership(i, firstCellStart, false);
+
+      return false;
+    }
+
+    assert(currentPartitionLinks[cellStart] == cellEnd);
+  }
+
+  vertexHash.clear();
+
+  return true;
+}
+
+HighsInt HighsSymmetryDetection::selectTargetCell() const {
+  HighsInt i = 0;
+  if (nodeStack.size() > 1) i = nodeStack[nodeStack.size() - 2].targetCell;
+
+  while (i < numVertices) {
+    if (cellSize(i) > 1) return i;
+
+    ++i;
+  }
+
+  return -1;
+}
+
+bool HighsSymmetryDetection::checkStoredAutomorphism(HighsInt vertex) const {
+  HighsInt numCheck = std::min(numAutomorphisms, (HighsInt)64);
+
+  for (HighsInt i = 0; i < numCheck; ++i) {
+    const HighsInt* automorphism = automorphisms.data() + i * numVertices;
+    bool automorphismUseful = true;
+    for (HighsInt j = nodeStack.size() - 2; j >= firstPathDepth; --j) {
+      HighsInt fixPos = vertexPosition[nodeStack[j].lastDistiguished];
+
+      if (automorphism[fixPos] != vertexGroundSet[fixPos]) {
+        automorphismUseful = false;
+        break;
+      }
+    }
+
+    if (!automorphismUseful) continue;
+
+    if (automorphism[vertexPosition[vertex]] < vertex) return false;
+  }
+
+  return true;
+}
+
+bool HighsSymmetryDetection::determineNextToDistinguish() {
+  Node& currNode = nodeStack.back();
+  distinguishCands.clear();
+  std::vector<HighsInt>::iterator cellStart;
+  std::vector<HighsInt>::iterator cellEnd;
+  cellStart = currentPartition.begin() + currNode.targetCell;
+  cellEnd =
+      currentPartition.begin() + currentPartitionLinks[currNode.targetCell];
+
+  if (currNode.lastDistiguished == -1) {
+    auto nextDistinguishPos = std::min_element(cellStart, cellEnd);
+    distinguishCands.push_back(&*nextDistinguishPos);
+  } else if ((HighsInt)nodeStack.size() > firstPathDepth) {
+    for (auto i = cellStart; i != cellEnd; ++i) {
+      if (*i > currNode.lastDistiguished && checkStoredAutomorphism(*i))
+        distinguishCands.push_back(&*i);
+    }
+    if (distinguishCands.empty()) return false;
+    auto nextDistinguishPos =
+        std::min_element(distinguishCands.begin(), distinguishCands.end(),
+                         [](HighsInt* a, HighsInt* b) { return *a < *b; });
+    std::swap(*distinguishCands.begin(), *nextDistinguishPos);
+    distinguishCands.resize(1);
+  } else {
+    for (auto i = cellStart; i != cellEnd; ++i) {
+      if (*i > currNode.lastDistiguished && vertexGroundSet[getOrbit(*i)] == *i)
+        distinguishCands.push_back(&*i);
+    }
+    if (distinguishCands.empty()) return false;
+    auto nextDistinguishPos =
+        std::min_element(distinguishCands.begin(), distinguishCands.end(),
+                         [](HighsInt* a, HighsInt* b) { return *a < *b; });
+    std::swap(*distinguishCands.begin(), *nextDistinguishPos);
+    distinguishCands.resize(1);
+  }
+
+  return true;
+}
+
+bool HighsSymmetryDetection::distinguishVertex(HighsInt targetCell) {
+  assert(distinguishCands.size() == 1u);
+  HighsInt targetCellEnd = currentPartitionLinks[targetCell];
+  HighsInt newCell = targetCellEnd - 1;
+  std::swap(*distinguishCands[0], currentPartition[newCell]);
+  nodeStack.back().lastDistiguished = currentPartition[newCell];
+
+  if (!splitCell(targetCell, newCell)) return false;
+
+  updateCellMembership(newCell, newCell);
+
+  return true;
+}
+
+void HighsSymmetryDetection::backtrack(HighsInt backtrackStackNewEnd,
+                                       HighsInt backtrackStackEnd) {
+  // we assume that we always backtrack from a leaf node, i.e. a discrete
+  // partition therefore we do not need to remember the values of the hash
+  // contributions as it is the indentity for each position and all new cells
+  // are on the cell creation stack.
+  for (HighsInt stackPos = backtrackStackEnd - 1;
+       stackPos >= backtrackStackNewEnd; --stackPos) {
+    HighsInt cell = cellCreationStack[stackPos];
+    // look up the cell start of the preceding cell with link compression
+    HighsInt newStart = getCellStart(cell - 1);
+    // remember the current end
+    HighsInt currEnd = currentPartitionLinks[cell];
+    // change the link to point to the start of the preceding cell
+    currentPartitionLinks[cell] = newStart;
+    // change the link of the start pointer of the preceding cell to point to
+    // the end of this cell
+    currentPartitionLinks[newStart] = currEnd;
+  }
+}
+
+void HighsSymmetryDetection::cleanupBacktrack(HighsInt cellCreationStackPos) {
+  // the links have been updated. Even though they might still not be fully
+  // compressed the cell starts will all point to the correct cell end and the
+  // lookup with path compression will give the correct start
+  for (HighsInt stackPos = cellCreationStack.size() - 1;
+       stackPos >= cellCreationStackPos; --stackPos) {
+    HighsInt cell = cellCreationStack[stackPos];
+
+    HighsInt cellStart = getCellStart(cell);
+    HighsInt cellEnd = currentPartitionLinks[cellStart];
+
+    for (HighsInt v = cell;
+         v < cellEnd && vertexToCell[currentPartition[v]] == cell; ++v)
+      updateCellMembership(v, cellStart, false);
+  }
+
+  cellCreationStack.resize(cellCreationStackPos);
+}
+
+HighsInt HighsSymmetryDetection::getCellStart(HighsInt pos) {
+  HighsInt startPos = currentPartitionLinks[pos];
+  if (startPos > pos) return pos;
+  if (currentPartitionLinks[startPos] < startPos) {
+    do {
+      linkCompressionStack.push_back(pos);
+      pos = startPos;
+      startPos = currentPartitionLinks[startPos];
+    } while (currentPartitionLinks[startPos] < startPos);
+
+    do {
+      currentPartitionLinks[linkCompressionStack.back()] = startPos;
+      linkCompressionStack.pop_back();
+    } while (!linkCompressionStack.empty());
+  }
+
+  return startPos;
+}
+
+void HighsSymmetryDetection::createNode() {
+  nodeStack.emplace_back();
+  nodeStack.back().stackStart = cellCreationStack.size();
+  nodeStack.back().certificateEnd = currNodeCertificate.size();
+  nodeStack.back().targetCell = -1;
+  nodeStack.back().lastDistiguished = -1;
+}
+
+struct MatrixColumn {
+  uint32_t cost;
+  uint32_t lb;
+  uint32_t ub;
+  uint32_t integral;
+  uint32_t len;
+
+  bool operator==(const MatrixColumn& other) const {
+    return std::memcmp(this, &other, sizeof(MatrixColumn)) == 0;
+  }
+};
+
+struct MatrixRow {
+  uint32_t lb;
+  uint32_t ub;
+  uint32_t len;
+
+  bool operator==(const MatrixRow& other) const {
+    return std::memcmp(this, &other, sizeof(MatrixRow)) == 0;
+  }
+};
+
+void HighsSymmetryDetection::loadModelAsGraph(const HighsLp& model,
+                                              double epsilon) {
+  this->model = &model;
+  numCol = model.num_col_;
+  numRow = model.num_row_;
+  numVertices = numRow + numCol;
+
+  cellInRefinementQueue.resize(numVertices);
+  vertexToCell.resize(numVertices);
+  refinementQueue.reserve(numVertices);
+  currNodeCertificate.reserve(numVertices);
+
+  HighsHashTable<MatrixColumn, HighsInt> columnSet;
+  HighsHashTable<MatrixRow, HighsInt> rowSet;
+  HighsMatrixColoring coloring(epsilon);
+  edgeBuffer.resize(numVertices);
+  // set up row and column based incidence matrix
+  HighsInt numNz = model.a_matrix_.index_.size();
+  Gedge.resize(2 * numNz);
+  std::transform(model.a_matrix_.index_.begin(), model.a_matrix_.index_.end(),
+                 Gedge.begin(), [&](HighsInt rowIndex) {
+                   return std::make_pair(rowIndex + numCol, HighsUInt{0});
+                 });
+
+  Gstart.resize(numVertices + 1);
+  std::copy(model.a_matrix_.start_.begin(), model.a_matrix_.start_.end(),
+            Gstart.begin());
+
+  // set up the column colors and count row sizes
+  std::vector<HighsInt> rowSizes(numRow);
+  for (HighsInt i = 0; i < numCol; ++i) {
+    for (HighsInt j = Gstart[i]; j < Gstart[i + 1]; ++j) {
+      Gedge[j].second = coloring.color(model.a_matrix_.value_[j]);
+      rowSizes[model.a_matrix_.index_[j]] += 1;
+    }
+  }
+
+  // next set up the row starts using the computed row sizes
+  HighsInt offset = numNz;
+  for (HighsInt i = 0; i < numRow; ++i) {
+    Gstart[numCol + i] = offset;
+    offset += rowSizes[i];
+  }
+  Gstart[numCol + numRow] = offset;
+
+  Gend.assign(Gstart.begin() + 1, Gstart.end());
+
+  // finally add the nonzeros to the row major matrix
+  for (HighsInt i = 0; i < numCol; ++i) {
+    for (HighsInt j = Gstart[i]; j < Gstart[i + 1]; ++j) {
+      HighsInt row = model.a_matrix_.index_[j];
+      HighsInt ARpos = Gstart[numCol + row + 1] - rowSizes[row];
+      rowSizes[row] -= 1;
+      Gedge[ARpos].first = i;
+      Gedge[ARpos].second = Gedge[j].second;
+    }
+  }
+
+  // loop over the columns and assign them a number that is distinct based on
+  // their upper/lower bounds, cost, and integrality status. Use the columnSet
+  // hash table to look up whether a column with similar properties exists and
+  // use the previous number in that case. The number is stored in the
+  // colToCell array which is subsequently used to sort an initial column
+  // permutation.
+  HighsInt indexOffset = numCol + 1;
+  for (HighsInt i = 0; i < numCol; ++i) {
+    MatrixColumn matrixCol;
+
+    matrixCol.cost = coloring.color(model.col_cost_[i]);
+    matrixCol.lb = coloring.color(model.col_lower_[i]);
+    matrixCol.ub = coloring.color(model.col_upper_[i]);
+    matrixCol.integral = (u32)model.integrality_[i];
+    matrixCol.len = Gstart[i + 1] - Gstart[i];
+
+    HighsInt* columnCell = &columnSet[matrixCol];
+
+    if (*columnCell == 0) {
+      *columnCell = columnSet.size();
+      if (model.col_lower_[i] != 0.0 || model.col_upper_[i] != 1.0 ||
+          model.integrality_[i] == HighsVarType::kContinuous)
+        *columnCell += indexOffset;
+    }
+
+    vertexToCell[i] = *columnCell;
+  }
+
+  indexOffset = 2 * numCol + 1;
+  for (HighsInt i = 0; i < numRow; ++i) {
+    MatrixRow matrixRow;
+
+    matrixRow.lb = coloring.color(model.row_lower_[i]);
+    matrixRow.ub = coloring.color(model.row_upper_[i]);
+    matrixRow.len = Gstart[numCol + i + 1] - Gstart[numCol + i];
+
+    HighsInt* rowCell = &rowSet[matrixRow];
+
+    if (*rowCell == 0) *rowCell = rowSet.size();
+
+    vertexToCell[numCol + i] = indexOffset + *rowCell;
+  }
+
+  // set up the initial partition array, sort by the colToCell value
+  // assigned above
+  currentPartition.resize(numVertices);
+  std::iota(currentPartition.begin(), currentPartition.end(), 0);
+  pdqsort(currentPartition.begin(), currentPartition.end(),
+          [&](HighsInt v1, HighsInt v2) {
+            return vertexToCell[v1] < vertexToCell[v2];
+          });
+
+  // now set up partition links and correct the colToCell array to the
+  // correct cell index
+  currentPartitionLinks.resize(numVertices);
+  HighsInt cellStart = 0;
+  HighsInt cellNumber = 0;
+  for (HighsInt i = 0; i < numVertices; ++i) {
+    HighsInt vertex = currentPartition[i];
+    // if the cell number is different to the current cell number this is the
+    // start of a new cell
+    if (cellNumber != vertexToCell[vertex]) {
+      // remember the number of this cell to identify its end
+      cellNumber = vertexToCell[vertex];
+      // set the link of the cell start to point to its end
+      currentPartitionLinks[cellStart] = i;
+      // remember start of this cell
+      cellStart = i;
+    }
+
+    // correct the colToCell array to not store the start index of the
+    // cell, not its number
+    vertexToCell[vertex] = cellStart;
+    // set the link of the column to the cellStart
+    currentPartitionLinks[i] = cellStart;
+  }
+
+  // set the column partition link of the last started cell to point past the
+  // end
+  currentPartitionLinks[cellStart] = numVertices;
+}
+
+HighsHashTable<std::tuple<HighsInt, HighsInt, HighsUInt>>
+HighsSymmetryDetection::dumpCurrentGraph() {
+  HighsHashTable<std::tuple<HighsInt, HighsInt, HighsUInt>> graphTriplets;
+
+  for (HighsInt i = 0; i < numCol; ++i) {
+    HighsInt colCell = vertexToCell[i];
+    for (HighsInt j = Gstart[i]; j != Gend[i]; ++j)
+      graphTriplets.insert(vertexToCell[Gedge[j].first], colCell,
+                           Gedge[j].second);
+    for (HighsInt j = Gend[i]; j != Gstart[i + 1]; ++j)
+      graphTriplets.insert(Gedge[j].first, colCell, Gedge[j].second);
+  }
+
+  return graphTriplets;
+}
+
+void HighsSymmetryDetection::switchToNextNode(HighsInt backtrackDepth) {
+  HighsInt stackEnd = cellCreationStack.size();
+  // we need to backtrack the datastructures
+  nodeStack.resize(backtrackDepth);
+  if (backtrackDepth == 0) return;
+  do {
+    Node& currNode = nodeStack.back();
+    backtrack(currNode.stackStart, stackEnd);
+    stackEnd = currNode.stackStart;
+    firstPathDepth = std::min((HighsInt)nodeStack.size(), firstPathDepth);
+    bestPathDepth = std::min((HighsInt)nodeStack.size(), bestPathDepth);
+    firstLeavePrefixLen =
+        std::min(currNode.certificateEnd, firstLeavePrefixLen);
+    bestLeavePrefixLen = std::min(currNode.certificateEnd, bestLeavePrefixLen);
+    currNodeCertificate.resize(currNode.certificateEnd);
+    if (!determineNextToDistinguish()) {
+      nodeStack.pop_back();
+      continue;
+    }
+
+    // call cleanup backtrack with the final stackEnd
+    // so that all hashes are up to date and the link arrays do not contain
+    // chains anymore
+    cleanupBacktrack(stackEnd);
+    HighsInt targetCell = currNode.targetCell;
+
+    if (!distinguishVertex(targetCell)) {
+      // if distinguishing the next vertex fails, it means that its certificate
+      // value is lexicographically larger than that of the best leave
+      nodeStack.pop_back();
+      continue;
+    }
+
+    if (!partitionRefinement()) {
+      stackEnd = cellCreationStack.size();
+      continue;
+    }
+
+    createNode();
+    break;
+  } while (!nodeStack.empty());
+}
+
+bool HighsSymmetryDetection::compareCurrentGraph(
+    const HighsHashTable<std::tuple<HighsInt, HighsInt, HighsUInt>>& otherGraph,
+    HighsInt& wrongCell) const {
+  for (HighsInt i = 0; i < numCol; ++i) {
+    HighsInt colCell = vertexToCell[i];
+
+    for (HighsInt j = Gstart[i]; j != Gend[i]; ++j)
+      if (!otherGraph.find(std::make_tuple(vertexToCell[Gedge[j].first],
+                                           colCell, Gedge[j].second))) {
+        // return which cell does not match in its neighbourhood as this should
+        // have been detected with the hashing it can very rarely happen due to
+        // a hash collision. In such a case we want to backtrack to the last
+        // time where we targeted this particular cell. Otherwise we could spent
+        // a long time searching for a matching leave value until every
+        // combination is exhausted and for each leave in this subtree the graph
+        // comparison will fail on this edge.
+        wrongCell = colCell;
+        return false;
+      }
+    for (HighsInt j = Gend[i]; j != Gstart[i + 1]; ++j)
+      if (!otherGraph.find(
+              std::make_tuple(Gedge[j].first, colCell, Gedge[j].second))) {
+        wrongCell = colCell;
+        return false;
+      }
+  }
+
+  return true;
+}
+
+bool HighsSymmetryDetection::isFromBinaryColumn(HighsInt pos) const {
+  if (pos >= numActiveCols) return false;
+
+  HighsInt col = currentPartition[pos];
+
+  if (model->col_lower_[col] != 0.0 || model->col_upper_[col] != 1.0 ||
+      model->integrality_[col] == HighsVarType::kContinuous)
+    return false;
+
+  return true;
+}
+
+HighsSymmetryDetection::ComponentData
+HighsSymmetryDetection::computeComponentData(
+    const HighsSymmetries& symmetries) {
+  ComponentData componentData;
+
+  componentData.components.reset(numActiveCols);
+  componentData.firstUnfixed.assign(symmetries.numPerms, -1);
+  componentData.numUnfixed.assign(symmetries.numPerms, 0);
+  for (HighsInt i = 0; i < symmetries.numPerms; ++i) {
+    const HighsInt* perm = symmetries.permutations.data() + i * numActiveCols;
+
+    for (HighsInt j = 0; j < numActiveCols; ++j) {
+      if (perm[j] != vertexGroundSet[j]) {
+        HighsInt pos = vertexPosition[perm[j]];
+        componentData.numUnfixed[i] += 1;
+        if (componentData.firstUnfixed[i] != -1)
+          componentData.components.merge(componentData.firstUnfixed[i], pos);
+        else
+          componentData.firstUnfixed[i] = pos;
+      }
+    }
+  }
+
+  componentData.componentSets.assign(vertexGroundSet.begin(),
+                                     vertexGroundSet.begin() + numActiveCols);
+  pdqsort(componentData.componentSets.begin(),
+          componentData.componentSets.end(), [&](HighsInt u, HighsInt v) {
+            HighsInt uComp = componentData.components.getSet(vertexPosition[u]);
+            HighsInt vComp = componentData.components.getSet(vertexPosition[v]);
+            return std::make_pair(
+                       componentData.components.getSetSize(uComp) == 1, uComp) <
+                   std::make_pair(
+                       componentData.components.getSetSize(vComp) == 1, vComp);
+          });
+
+  HighsInt currentComponentStart = -1;
+  HighsInt currentComponent = -1;
+  HighsHashTable<HighsInt> currComponentOrbits;
+  for (HighsInt i = 0; i < numActiveCols; ++i) {
+    HighsInt comp = componentData.components.getSet(
+        vertexPosition[componentData.componentSets[i]]);
+    if (componentData.components.getSetSize(comp) == 1) break;
+    if (comp != currentComponent) {
+      currentComponent = comp;
+      currentComponentStart = i;
+      componentData.componentStarts.push_back(currentComponentStart);
+      componentData.componentNumber.push_back(currentComponent);
+      componentData.componentNumOrbits.emplace_back();
+      currComponentOrbits.clear();
+    }
+
+    if (currComponentOrbits.insert(getOrbit(componentData.componentSets[i]))) {
+      ++componentData.componentNumOrbits.back();
+    }
+  }
+
+  componentData.permComponents.reserve(symmetries.numPerms);
+
+  for (HighsInt i = 0; i < symmetries.numPerms; ++i) {
+    if (componentData.firstUnfixed[i] == -1) continue;
+    componentData.permComponents.push_back(i);
+  }
+
+  pdqsort(componentData.permComponents.begin(),
+          componentData.permComponents.end(), [&](HighsInt i, HighsInt j) {
+            HighsInt seti =
+                componentData.components.getSet(componentData.firstUnfixed[i]);
+            HighsInt setj =
+                componentData.components.getSet(componentData.firstUnfixed[j]);
+            return std::make_pair(seti, componentData.numUnfixed[i]) <
+                   std::make_pair(setj, componentData.numUnfixed[j]);
+          });
+
+  currentComponentStart = -1;
+  currentComponent = -1;
+
+  HighsInt numUsedPerms = componentData.permComponents.size();
+
+  for (HighsInt i = 0; i < numUsedPerms; ++i) {
+    HighsInt p = componentData.permComponents[i];
+    HighsInt comp =
+        componentData.components.getSet(componentData.firstUnfixed[p]);
+    if (comp != currentComponent) {
+      currentComponent = comp;
+      currentComponentStart = i;
+      componentData.permComponentStarts.push_back(currentComponentStart);
+    }
+  }
+
+  assert(componentData.permComponentStarts.size() ==
+         componentData.componentStarts.size());
+  componentData.permComponentStarts.push_back(numUsedPerms);
+
+  // HighsInt numComponents = componentData.componentStarts.size();
+  // printf("found %d components\n", numComponents);
+  componentData.componentStarts.push_back(numActiveCols);
+
+  return componentData;
+}
+
+bool HighsSymmetryDetection::isFullOrbitope(const ComponentData& componentData,
+                                            HighsInt component,
+                                            HighsSymmetries& symmetries) {
+  HighsInt componentSize = componentData.componentStarts[component + 1] -
+                           componentData.componentStarts[component];
+  if (componentSize == 1) return false;
+
+  // check that component acts only on binary variables
+  for (HighsInt i = componentData.componentStarts[component];
+       i < componentData.componentStarts[component + 1]; ++i) {
+    HighsInt col = componentData.componentSets[i];
+    if (model->integrality_[col] == HighsVarType::kContinuous ||
+        model->col_lower_[col] != 0.0 || model->col_upper_[col] != 1.0)
+      return false;
+  }
+
+  // check that the number of unfixed variables in the first permutation is even
+  HighsInt p0 =
+      componentData
+          .permComponents[componentData.permComponentStarts[component]];
+  if (componentData.numUnfixed[p0] & 1) return false;
+
+  // check that the other permutations in the component have the same number of
+  // unfixed variables as the first one
+  for (HighsInt k = componentData.permComponentStarts[component] + 1;
+       k < componentData.permComponentStarts[component + 1]; ++k) {
+    HighsInt p = componentData.permComponents[k];
+    if (componentData.numUnfixed[p] != componentData.numUnfixed[p0]) {
+      // printf("number of unfixed cols in perm %d: %d\n", p,
+      //       componentData.numUnfixed[p]);
+      // printf("wrong number of unfixed columns in permutation\n");
+      return false;
+    }
+  }
+
+  // all unfixed variables in the permutation must be part of a two cycle
+  // and each two cycle defines one row of the orbitope. Hence the number
+  // of unfixed variables in each permutation must be even and the number of
+  // rows is the number of unfixed variables divided by two
+  HighsInt orbitopeNumRows = componentData.numUnfixed[p0] >> 1;
+
+  // if this component is a full orbitope the component size must be
+  // divisible by the number of orbits and the result is the size of each
+  // orbit
+  HighsInt orbitopeOrbitSize = componentSize / orbitopeNumRows;
+  if (orbitopeOrbitSize * orbitopeNumRows != componentSize) {
+    // printf("wrong number of orbits (%d orbits for component of size %d)\n",
+    //        orbitopeNumRows, componentSize);
+    return false;
+  }
+
+  // the number of permutations must be n-1 where n is the size of the
+  // orbits in the orbitope
+  HighsInt componentNumPerms =
+      componentData.permComponentStarts[component + 1] -
+      componentData.permComponentStarts[component];
+  if (componentNumPerms != (orbitopeOrbitSize - 1)) {
+    // printf("wrong number of perms\n");
+    return false;
+  }
+
+  // set up the first two columns of the orbitope matrix based on the first
+  // permutation.
+  HighsOrbitopeMatrix orbitopeMatrix;
+  orbitopeMatrix.matrix.resize(componentSize, -1);
+  orbitopeMatrix.numRows = orbitopeNumRows;
+  orbitopeMatrix.rowLength = orbitopeOrbitSize;
+  assert(componentSize == orbitopeMatrix.numRows * orbitopeMatrix.rowLength);
+
+  const HighsInt* perm = symmetries.permutations.data() + p0 * numActiveCols;
+  HighsHashTable<HighsInt> colSet;
+  HighsInt m = 0;
+  for (HighsInt j = 0; j < numActiveCols; ++j) {
+    HighsInt jImagePos = vertexPosition[perm[j]];
+    if (jImagePos <= j) continue;
+    if (m == orbitopeNumRows) return false;
+
+    // permutation should consist of two cycles
+    if (perm[jImagePos] != vertexGroundSet[j]) return false;
+
+    orbitopeMatrix.matrix[m] = vertexGroundSet[j];
+    orbitopeMatrix.matrix[orbitopeMatrix.numRows + m] = perm[j];
+
+    // Remember set of variables of the orbtiope matrix. Each variable should
+    // occur only once, otherwise the permutation do not work out. Since we
+
+    if (!colSet.insert(vertexGroundSet[j])) return false;
+    if (!colSet.insert(perm[j])) return false;
+    ++m;
+  }
+
+  // printf("set up first two columns of possible orbitope with permutation
+  // %d\n",
+  //        p0);
+
+  HighsInt numColsAdded = 2;
+  bool triedLeftExtension = false;
+
+  while (numColsAdded < orbitopeMatrix.rowLength) {
+    if (colSet.size() != numColsAdded * orbitopeMatrix.numRows) return false;
+
+    HighsInt* thisCol = &orbitopeMatrix(0, numColsAdded);
+    HighsInt* prevCol = &orbitopeMatrix(0, numColsAdded - 1);
+
+    bool foundCand = false;
+
+    while (true) {
+      HighsInt movePos = vertexPosition[prevCol[0]];
+      perm = nullptr;
+
+      for (HighsInt k = componentData.permComponentStarts[component] + 1;
+           k < componentData.permComponentStarts[component + 1]; ++k) {
+        HighsInt p = componentData.permComponents[k];
+        perm = symmetries.permutations.data() + p * numActiveCols;
+
+        if (perm[movePos] != vertexGroundSet[movePos] &&
+            !colSet.find(perm[movePos])) {
+          foundCand = true;
+          break;
+        }
+      }
+
+      if (!foundCand && !triedLeftExtension) {
+        // if we fail to find a permutation extending the last column directly
+        // after we set up the first two columns we might try to extend column
+        // zero instead of the previous one
+        prevCol = &orbitopeMatrix(0, 0);
+        triedLeftExtension = true;
+        continue;
+      }
+
+      break;
+    }
+
+    if (!foundCand) {
+      // printf("did not find next permutation moving col %d\n", prevCol[0]);
+      return false;
+    }
+
+    for (HighsInt j = 0; j < orbitopeMatrix.numRows; ++j) {
+      HighsInt nextVertex = perm[vertexPosition[prevCol[j]]];
+
+      thisCol[j] = nextVertex;
+
+      // check if this is a two cycle
+      if (perm[vertexPosition[nextVertex]] != prevCol[j]) return false;
+
+      if (!colSet.insert(thisCol[j])) {
+        // printf("col already exists\n");
+        return false;
+      }
+    }
+
+    ++numColsAdded;
+  }
+
+  if (colSet.size() != componentSize) {
+    // printf("not all columns of component are mapped\n");
+    return false;
+  }
+
+  for (HighsInt col : orbitopeMatrix.matrix)
+    symmetries.columnToOrbitope.insert(col, symmetries.orbitopes.size());
+
+  symmetries.orbitopes.emplace_back(std::move(orbitopeMatrix));
+
+  // printf("component %d is full orbitope: size %d and %d orbits\n", component,
+  //        componentSize, componentData.componentNumOrbits[component]);
+  return true;
+}
+
+bool HighsSymmetryDetection::initializeDetection() {
+  initializeHashValues();
+  partitionRefinement();
+  removeFixPoints();
+  if (numActiveCols == 0) return false;
+  return true;
+}
+
+void HighsSymmetryDetection::run(HighsSymmetries& symmetries) {
+  assert(numActiveCols != 0);
+  initializeGroundSet();
+  currNodeCertificate.clear();
+  cellCreationStack.clear();
+  createNode();
+  HighsInt maxPerms = 64000000 / numActiveCols;
+  HighsSplitDeque* workerDeque = HighsTaskExecutor::getThisWorkerDeque();
+  while (!nodeStack.empty()) {
+    HighsInt targetCell = selectTargetCell();
+    if (targetCell == -1) {
+      if (firstLeavePartition.empty()) {
+        firstLeavePartition = currentPartition;
+        firstLeaveCertificate = currNodeCertificate;
+        bestLeaveCertificate = currNodeCertificate;
+        firstLeaveGraph = dumpCurrentGraph();
+        firstPathDepth = nodeStack.size();
+        bestPathDepth = nodeStack.size();
+        firstLeavePrefixLen = currNodeCertificate.size();
+        bestLeavePrefixLen = currNodeCertificate.size();
+
+        HighsInt backtrackDepth = firstPathDepth - 1;
+        while (backtrackDepth > 0 &&
+               !isFromBinaryColumn(nodeStack[backtrackDepth - 1].targetCell))
+          --backtrackDepth;
+        switchToNextNode(backtrackDepth);
+      } else {
+        HighsInt wrongCell = -1;
+        HighsInt backtrackDepth = nodeStack.size() - 1;
+        assert(currNodeCertificate.size() == firstLeaveCertificate.size());
+        if (firstLeavePrefixLen == (HighsInt)currNodeCertificate.size() ||
+            bestLeavePrefixLen == (HighsInt)currNodeCertificate.size()) {
+          if (firstLeavePrefixLen == (HighsInt)currNodeCertificate.size() &&
+              compareCurrentGraph(firstLeaveGraph, wrongCell)) {
+            HighsInt k = (numAutomorphisms++) & 63;
+            HighsInt* permutation = automorphisms.data() + k * numVertices;
+            for (HighsInt i = 0; i < numVertices; ++i) {
+              HighsInt firstLeaveCol = firstLeavePartition[i];
+              permutation[vertexPosition[currentPartition[i]]] = firstLeaveCol;
+            }
+
+            bool report = false;
+            for (HighsInt i = 0; i < numVertices; ++i) {
+              if (mergeOrbits(permutation[i], vertexGroundSet[i]) &&
+                  i < numActiveCols) {
+                assert(permutation[i] < numCol);
+                report = true;
+              }
+            }
+
+            if (report) {
+              symmetries.permutations.insert(symmetries.permutations.end(),
+                                             permutation,
+                                             permutation + numActiveCols);
+              ++symmetries.numPerms;
+              if (symmetries.numPerms == maxPerms) break;
+            }
+            backtrackDepth = std::min(backtrackDepth, firstPathDepth);
+          } else if (!bestLeavePartition.empty() &&
+                     bestLeavePrefixLen ==
+                         (HighsInt)currNodeCertificate.size() &&
+                     compareCurrentGraph(bestLeaveGraph, wrongCell)) {
+            HighsInt k = (numAutomorphisms++) & 63;
+            HighsInt* permutation = automorphisms.data() + k * numVertices;
+            for (HighsInt i = 0; i < numVertices; ++i) {
+              HighsInt bestLeaveCol = bestLeavePartition[i];
+              permutation[vertexPosition[currentPartition[i]]] = bestLeaveCol;
+            }
+
+            bool report = false;
+            for (HighsInt i = 0; i < numVertices; ++i) {
+              if (mergeOrbits(permutation[i], vertexGroundSet[i]) &&
+                  i < numActiveCols) {
+                assert(permutation[i] < numCol);
+                report = true;
+              }
+            }
+
+            if (report) {
+              symmetries.permutations.insert(symmetries.permutations.end(),
+                                             permutation,
+                                             permutation + numActiveCols);
+              ++symmetries.numPerms;
+              if (symmetries.numPerms == maxPerms) break;
+            }
+
+            backtrackDepth = std::min(backtrackDepth, bestPathDepth);
+          } else if (bestLeavePrefixLen <
+                         (HighsInt)currNodeCertificate.size() &&
+                     currNodeCertificate[bestLeavePrefixLen] >
+                         bestLeaveCertificate[bestLeavePrefixLen]) {
+            // certificate value is lexicographically above the smallest one
+            // seen so far, so we might be able to backtrack to a higher level
+            HighsInt possibleBacktrackDepth = firstPathDepth - 1;
+            while (nodeStack[possibleBacktrackDepth].certificateEnd <=
+                   bestLeavePrefixLen)
+              ++possibleBacktrackDepth;
+
+            backtrackDepth = std::min(possibleBacktrackDepth, backtrackDepth);
+          } else {
+            // This case can be caused by a hash collision which was now
+            // detected in the graph comparison call. The graph comparison call
+            // will return the cell where the vertex neighbourhood caused a
+            // mismatch on the edges. This would have been detected by
+            // an exact partition refinement when we targeted that cell the last
+            // time, so that is where we can backtrack to.
+            HighsInt possibleBacktrackDepth;
+            for (possibleBacktrackDepth = backtrackDepth;
+                 possibleBacktrackDepth >= 0; --possibleBacktrackDepth) {
+              if (nodeStack[possibleBacktrackDepth].targetCell == wrongCell) {
+                backtrackDepth = possibleBacktrackDepth;
+                break;
+              }
+            }
+          }
+        } else {
+          // leave must have a lexicographically smaller certificate value
+          // than the current best leave, because its prefix length is smaller
+          // than the best leaves and it would have been already pruned if
+          // it's certificate value was larger unless it is equal to the first
+          // leaf nodes certificate value which is caught by the first case
+          // of the if condition. Hence, having a lexicographically smaller
+          // certificate value than the best leave is the only way to get
+          // here.
+          assert(bestLeaveCertificate[bestLeavePrefixLen] >
+                     currNodeCertificate[bestLeavePrefixLen] &&
+                 std::memcmp(bestLeaveCertificate.data(),
+                             currNodeCertificate.data(),
+                             bestLeavePrefixLen * sizeof(u32)) == 0);
+          bestLeaveCertificate = currNodeCertificate;
+          bestLeaveGraph = dumpCurrentGraph();
+          bestLeavePartition = currentPartition;
+          bestPathDepth = nodeStack.size();
+          bestLeavePrefixLen = currNodeCertificate.size();
+        }
+
+        switchToNextNode(backtrackDepth);
+      }
+
+      workerDeque->checkInterrupt();
+    } else {
+      Node& currNode = nodeStack.back();
+      currNode.targetCell = targetCell;
+      bool success = determineNextToDistinguish();
+      assert(success);
+      if (!distinguishVertex(targetCell)) {
+        switchToNextNode(nodeStack.size() - 1);
+        continue;
+      }
+      if (!partitionRefinement()) {
+        switchToNextNode(nodeStack.size());
+        continue;
+      }
+
+      createNode();
+    }
+  }
+
+  symmetries.numGenerators = symmetries.numPerms;
+  if (symmetries.numPerms > 0) {
+    vertexPosition.resize(numCol);
+
+    ComponentData componentData = computeComponentData(symmetries);
+    HighsInt numComponents = componentData.numComponents();
+
+    for (HighsInt i = 0; i < numComponents; ++i) {
+      if (componentData.componentSize(i) == 1) continue;
+
+      isFullOrbitope(componentData, i, symmetries);
+    }
+
+    HighsHashTable<HighsInt> deletedPerms;
+    for (HighsInt p = 0; p < symmetries.numPerms; ++p) {
+      HighsInt* perm = symmetries.permutations.data() + p * numActiveCols;
+      for (HighsInt i = 0; i < numActiveCols; ++i) {
+        if (perm[i] != vertexGroundSet[i] &&
+            symmetries.columnToOrbitope.find(perm[i])) {
+          deletedPerms.insert(p);
+          break;
+        }
+      }
+    }
+
+    // check which columns have non-trivial orbits
+    HighsInt numFixed = 0;
+    for (HighsInt i = 0; i < numActiveCols; ++i) {
+      if (orbitSize[getOrbit(vertexGroundSet[i])] == 1 ||
+          symmetries.columnToOrbitope.find(vertexGroundSet[i])) {
+        vertexPosition[vertexGroundSet[i]] = -1;
+        vertexGroundSet[i] = -1;
+        numFixed += 1;
+      }
+    }
+
+    if (numFixed != 0) {
+      // now compress symmetries and the groundset to only contain the unfixed
+      // columns and columns not handled by full orbitopes
+      HighsInt p = 0;
+      HighsInt* perms = symmetries.permutations.data();
+      HighsInt* permEnd =
+          symmetries.permutations.data() + symmetries.numPerms * numActiveCols;
+      HighsInt* permOutput = symmetries.permutations.data();
+      while (perms != permEnd) {
+        if (!deletedPerms.find(p)) {
+          for (HighsInt i = 0; i < numActiveCols; ++i) {
+            if (vertexGroundSet[i] == -1) continue;
+
+            *permOutput = perms[i];
+            ++permOutput;
+          }
+        } else {
+          --symmetries.numPerms;
+        }
+        perms += numActiveCols;
+        ++p;
+      }
+
+      HighsInt outPos = 0;
+      for (HighsInt i = 0; i < numActiveCols; ++i) {
+        if (vertexGroundSet[i] == -1) continue;
+
+        vertexGroundSet[outPos] = vertexGroundSet[i];
+        vertexPosition[vertexGroundSet[outPos]] = outPos;
+        outPos += 1;
+      }
+
+      numActiveCols -= numFixed;
+      assert(permOutput == symmetries.permutations.data() +
+                               symmetries.numPerms * numActiveCols);
+    }
+
+    vertexGroundSet.resize(numActiveCols);
+    symmetries.permutationColumns = std::move(vertexGroundSet);
+    symmetries.columnPosition = std::move(vertexPosition);
+    symmetries.permutations.resize(symmetries.numPerms * numActiveCols);
+  }
+}