OR-Tools  9.3
lu_factorization.h
Go to the documentation of this file.
1// Copyright 2010-2021 Google LLC
2// Licensed under the Apache License, Version 2.0 (the "License");
3// you may not use this file except in compliance with the License.
4// You may obtain a copy of the License at
5//
6// http://www.apache.org/licenses/LICENSE-2.0
7//
8// Unless required by applicable law or agreed to in writing, software
9// distributed under the License is distributed on an "AS IS" BASIS,
10// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
11// See the License for the specific language governing permissions and
12// limitations under the License.
13
14#ifndef OR_TOOLS_GLOP_LU_FACTORIZATION_H_
15#define OR_TOOLS_GLOP_LU_FACTORIZATION_H_
16
18#include "ortools/glop/parameters.pb.h"
19#include "ortools/glop/status.h"
25#include "ortools/util/stats.h"
26
27namespace operations_research {
28namespace glop {
29
30// An LU-Factorization class encapsulating the LU factorization data and
31// algorithms. The actual algorithm is in markowitz.h and .cc. This class holds
32// all the Solve() functions that deal with the permutations and the L and U
33// factors once they are computed.
35 public:
37
38 // Returns true if the LuFactorization is a factorization of the identity
39 // matrix. In this state, all the Solve() functions will work for any
40 // vector dimension.
41 bool IsIdentityFactorization() { return is_identity_factorization_; }
42
43 // Clears internal data structure and reset this class to the factorization
44 // of an identity matrix.
45 void Clear();
46
47 // Computes an LU-decomposition for a given matrix B. If for some reason,
48 // there was an error, then the factorization is reset to the one of the
49 // identity matrix, and an error is reported.
50 //
51 // Note(user): Since a client must use the result, there is little chance of
52 // it being confused by this revert to identity factorization behavior. The
53 // reason behind it is that this way, calling any public function of this
54 // class will never cause a crash of the program.
55 ABSL_MUST_USE_RESULT Status
56 ComputeFactorization(const CompactSparseMatrixView& compact_matrix);
57
58 // Given a set of columns, find a maximum linearly independent subset that can
59 // be factorized in a stable way, and complete it into a square matrix using
60 // slack columns. The initial set can have less, more or the same number of
61 // columns as the number of rows.
63 const std::vector<ColIndex>& candidates);
64
65 // Returns the column permutation used by the LU factorization.
66 const ColumnPermutation& GetColumnPermutation() const { return col_perm_; }
67
68 // Sets the column permutation to the identity permutation. The idea is that
69 // the column permutation can be incorporated in the basis RowToColMapping,
70 // and once this is done, then a client can call this and effectively remove
71 // the need for a column permutation on each solve.
73 col_perm_.clear();
74 inverse_col_perm_.clear();
75 }
76
77 // Solves 'B.x = b', x initially contains b, and is replaced by 'B^{-1}.b'.
78 // Since P.B.Q^{-1} = L.U, we have B = P^{-1}.L.U.Q.
79 // 1/ Solve P^{-1}.y = b for y by computing y = P.b,
80 // 2/ solve L.z = y for z,
81 // 3/ solve U.t = z for t,
82 // 4/ finally solve Q.x = t, by computing x = Q^{-1}.t.
83 void RightSolve(DenseColumn* x) const;
84
85 // Solves 'y.B = r', y initially contains r, and is replaced by r.B^{-1}.
86 // Internally, it takes x = y^T, b = r^T and solves B^T.x = b.
87 // We have P.B.Q^{-1} = P.B.Q^T = L.U, thus (L.U)^T = Q.B^T.P^T.
88 // Therefore B^T = Q^{-1}.U^T.L^T.P^T.P^{-1} = Q^{-1}.U^T.L^T.P
89 // The procedure is thus:
90 // 1/ Solve Q^{-1}.y = b for y, by computing y = Q.b,
91 // 2/ solve U^T.z = y for z,
92 // 3/ solve L^T.t = z for t,
93 // 4/ finally, solve P.x = t for x by computing x = P^{-1}.t.
94 void LeftSolve(DenseRow* y) const;
95
96 // More fine-grained right/left solve functions that may exploit the initial
97 // non-zeros of the input vector if non-empty. Note that a solve involving L
98 // actually solves P^{-1}.L and a solve involving U actually solves U.Q. To
99 // solve a system with the initial matrix B, one needs to call:
100 // - RightSolveL() and then RightSolveU() for a right solve (B.x = initial x).
101 // - LeftSolveU() and then LeftSolveL() for a left solve (y.B = initial y).
105
106 // Specialized version of LeftSolveL() that may exploit the initial non_zeros
107 // of y if it is non empty. Moreover, if result_before_permutation is not
108 // NULL, it might be filled with the result just before row_perm_ is applied
109 // to it and true is returned. If result_before_permutation is not filled,
110 // then false is returned.
112 ScatteredColumn* result_before_permutation) const;
114
115 // Specialized version of RightSolveLWithNonZeros() that takes a SparseColumn
116 // or a ScatteredColumn as input. non_zeros will either be cleared or set to
117 // the non zeros of the result. Important: the output x must be of the correct
118 // size and all zero.
121 ScatteredColumn* x) const;
122
123 // Specialized version of RightSolveLWithNonZeros() where x is originally
124 // equal to 'a' permuted by row_perm_. Note that 'a' is only used for DCHECK.
126 ScatteredColumn* x) const;
127
128 // Specialized version of LeftSolveU() for an unit right-hand side.
129 // non_zeros will either be cleared or set to the non zeros of the results.
130 // It also returns the value of col permuted by Q (which is the position
131 // of the unit-vector rhs in the solve system: y.U = rhs).
132 // Important: the output y must be of the correct size and all zero.
133 ColIndex LeftSolveUForUnitRow(ColIndex col, ScatteredRow* y) const;
134
135 // Returns the given column of U.
136 // It will only be valid until the next call to GetColumnOfU().
137 const SparseColumn& GetColumnOfU(ColIndex col) const;
138
139 // Returns the norm of B^{-1}.a
141
142 // Returns the norm of (B^T)^{-1}.e_row where e is an unit vector.
143 Fractional DualEdgeSquaredNorm(RowIndex row) const;
144
145 // The fill-in of the LU-factorization is defined as the sum of the number
146 // of entries of both the lower- and upper-triangular matrices L and U minus
147 // the number of entries in the initial matrix B.
148 //
149 // This returns the number of entries in lower + upper as the percentage of
150 // the number of entries in B.
151 double GetFillInPercentage(const CompactSparseMatrixView& matrix) const;
152
153 // Returns the number of entries in L + U.
154 // If the factorization is the identity, this returns 0.
155 EntryIndex NumberOfEntries() const;
156
157 // Computes the determinant of the input matrix B.
158 // Since P.B.Q^{-1} = L.U, det(P) * det(B) * det(Q^{-1}) = det(L) * det(U).
159 // det(L) = 1 since L is a lower-triangular matrix with 1 on the diagonal.
160 // det(P) = +1 or -1 (by definition it is the sign of the permutation P)
161 // det(Q^{-1}) = +1 or -1 (the sign of the permutation Q^{-1})
162 // Finally det(U) = product of the diagonal elements of U, since U is an
163 // upper-triangular matrix.
164 // Taking all this into account:
165 // det(B) = sign(P) * sign(Q^{-1}) * prod_i u_ii .
167
168 // Computes the 1-norm of the inverse of the input matrix B.
169 // For this we iteratively solve B.x = e_j, where e_j is the jth unit vector.
170 // The result of this computation is the jth column of B^-1.
171 // The 1-norm |B| is defined as max_j sum_i |a_ij|
172 // http://en.wikipedia.org/wiki/Matrix_norm
174
175 // Computes the infinity-norm of the inverse of the input matrix B.
176 // The infinity-norm |B| is defined as max_i sum_j |a_ij|
177 // http://en.wikipedia.org/wiki/Matrix_norm
179
180 // Computes the condition number of the input matrix B.
181 // For a given norm, this is the matrix norm times the norm of its inverse.
182 //
183 // Note that because the LuFactorization class does not keep the
184 // non-factorized matrix in memory, it needs to be passed to these functions.
185 // It is up to the client to pass exactly the same matrix as the one used
186 // for ComputeFactorization().
187 //
188 // TODO(user): separate this from LuFactorization.
190 const CompactSparseMatrixView& matrix) const;
192 const CompactSparseMatrixView& matrix) const;
194
195 // Sets the current parameters.
196 void SetParameters(const GlopParameters& parameters) {
197 parameters_ = parameters;
198 markowitz_.SetParameters(parameters);
199 }
200
201 // Returns a string containing the statistics for this class.
202 std::string StatString() const {
203 return stats_.StatString() + markowitz_.StatString();
204 }
205
206 // This is only used for testing and in debug mode.
207 // TODO(user): avoid the matrix conversion by multiplying TriangularMatrix
208 // directly.
210 SparseMatrix temp_lower, temp_upper;
211 lower_.CopyToSparseMatrix(&temp_lower);
212 upper_.CopyToSparseMatrix(&temp_upper);
213 product->PopulateFromProduct(temp_lower, temp_upper);
214 }
215
216 // Returns the deterministic time of the last factorization.
218
219 // Visible for testing.
220 const RowPermutation& row_perm() const { return row_perm_; }
222 return inverse_col_perm_;
223 }
224
225 private:
226 // Statistics about this class.
227 struct Stats : public StatsGroup {
228 Stats()
229 : StatsGroup("LuFactorization"),
230 basis_num_entries("basis_num_entries", this),
231 lu_fill_in("lu_fill_in", this) {}
232 IntegerDistribution basis_num_entries;
233 RatioDistribution lu_fill_in;
234 };
235
236 // Internal function used in the left solve functions.
237 void LeftSolveScratchpad() const;
238
239 // Internal function used in the right solve functions
240 template <typename Column>
241 void RightSolveLInternal(const Column& b, ScatteredColumn* x) const;
242
243 // Fills transpose_upper_ from upper_.
244 void ComputeTransposeUpper();
245
246 // transpose_lower_ is only needed when we compute dual norms.
247 void ComputeTransposeLower() const;
248
249 // Computes R = P.B.Q^{-1} - L.U and returns false if the largest magnitude of
250 // the coefficients of P.B.Q^{-1} - L.U is greater than tolerance.
251 bool CheckFactorization(const CompactSparseMatrixView& matrix,
252 Fractional tolerance) const;
253
254 // Special case where we have nothing to do. This happens at the beginning
255 // when we start the problem with an all-slack basis and gives a good speedup
256 // on really easy problems. It is initially true and set to true each time we
257 // call Clear(). We set it to false if a call to ComputeFactorization()
258 // succeeds.
259 bool is_identity_factorization_;
260
261 // The triangular factor L and U (and its transpose).
262 TriangularMatrix lower_;
263 TriangularMatrix upper_;
264 TriangularMatrix transpose_upper_;
265
266 // The transpose of lower_. It is just used by DualEdgeSquaredNorm()
267 // and mutable so it can be lazily initialized.
268 mutable TriangularMatrix transpose_lower_;
269
270 // The column permutation Q and its inverse Q^{-1} in P.B.Q^{-1} = L.U.
271 ColumnPermutation col_perm_;
272 ColumnPermutation inverse_col_perm_;
273
274 // The row permutation P and its inverse P^{-1} in P.B.Q^{-1} = L.U.
275 RowPermutation row_perm_;
276 RowPermutation inverse_row_perm_;
277
278 // Temporary storage used by LeftSolve()/RightSolve().
279 mutable DenseColumn dense_column_scratchpad_;
280
281 // Temporary storage used by GetColumnOfU().
282 mutable SparseColumn column_of_upper_;
283
284 // Same as dense_column_scratchpad_ but this vector is always reset to zero by
285 // the functions that use it. non_zero_rows_ is used to track the
286 // non_zero_rows_ position of dense_column_scratchpad_.
287 mutable DenseColumn dense_zero_scratchpad_;
288 mutable std::vector<RowIndex> non_zero_rows_;
289
290 // Statistics, mutable so const functions can still update it.
291 mutable Stats stats_;
292
293 // Proto holding all the parameters of this algorithm.
294 GlopParameters parameters_;
295
296 // The class doing the Markowitz LU factorization.
297 Markowitz markowitz_;
298
299 DISALLOW_COPY_AND_ASSIGN(LuFactorization);
300};
301
302} // namespace glop
303} // namespace operations_research
304#endif // OR_TOOLS_GLOP_LU_FACTORIZATION_H_
void LeftSolveUWithNonZeros(ScatteredRow *y) const
const SparseColumn & GetColumnOfU(ColIndex col) const
RowToColMapping ComputeInitialBasis(const CompactSparseMatrix &matrix, const std::vector< ColIndex > &candidates)
void RightSolveLForColumnView(const ColumnView &b, ScatteredColumn *x) const
const ColumnPermutation & inverse_col_perm() const
void RightSolveLWithPermutedInput(const DenseColumn &a, ScatteredColumn *x) const
double GetFillInPercentage(const CompactSparseMatrixView &matrix) const
Fractional RightSolveSquaredNorm(const ColumnView &a) const
void RightSolveUWithNonZeros(ScatteredColumn *x) const
const ColumnPermutation & GetColumnPermutation() const
bool LeftSolveLWithNonZeros(ScatteredRow *y, ScatteredColumn *result_before_permutation) const
ColIndex LeftSolveUForUnitRow(ColIndex col, ScatteredRow *y) const
Fractional DualEdgeSquaredNorm(RowIndex row) const
void RightSolveLForScatteredColumn(const ScatteredColumn &b, ScatteredColumn *x) const
void RightSolveLWithNonZeros(ScatteredColumn *x) const
Fractional ComputeInfinityNormConditionNumber(const CompactSparseMatrixView &matrix) const
void ComputeLowerTimesUpper(SparseMatrix *product) const
ABSL_MUST_USE_RESULT Status ComputeFactorization(const CompactSparseMatrixView &compact_matrix)
void SetParameters(const GlopParameters &parameters)
Fractional ComputeOneNormConditionNumber(const CompactSparseMatrixView &matrix) const
const RowPermutation & row_perm() const
void SetParameters(const GlopParameters &parameters)
Definition: markowitz.h:313
std::string StatString() const
Definition: markowitz.h:310
void PopulateFromProduct(const SparseMatrix &a, const SparseMatrix &b)
Definition: sparse.cc:250
void CopyToSparseMatrix(SparseMatrix *output) const
Definition: sparse.cc:757
int64_t b
int64_t a
SatParameters parameters
ColIndex col
Definition: markowitz.cc:183
RowIndex row
Definition: markowitz.cc:182
Permutation< ColIndex > ColumnPermutation
StrictITIVector< RowIndex, Fractional > DenseColumn
Definition: lp_types.h:332
Permutation< RowIndex > RowPermutation
Collection of objects used to extend the Constraint Solver library.