1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
5 // Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 #ifndef EIGEN_PARTIALLU_H
12 #define EIGEN_PARTIALLU_H
13 
14 namespace Eigen {
15 
16 /** \ingroup LU_Module
17   *
18   * \class PartialPivLU
19   *
20   * \brief LU decomposition of a matrix with partial pivoting, and related features
21   *
22   * \param MatrixType the type of the matrix of which we are computing the LU decomposition
23   *
24   * This class represents a LU decomposition of a \b square \b invertible matrix, with partial pivoting: the matrix A
25   * is decomposed as A = PLU where L is unit-lower-triangular, U is upper-triangular, and P
26   * is a permutation matrix.
27   *
28   * Typically, partial pivoting LU decomposition is only considered numerically stable for square invertible
29   * matrices. Thus LAPACK's dgesv and dgesvx require the matrix to be square and invertible. The present class
30   * does the same. It will assert that the matrix is square, but it won't (actually it can't) check that the
31   * matrix is invertible: it is your task to check that you only use this decomposition on invertible matrices.
32   *
33   * The guaranteed safe alternative, working for all matrices, is the full pivoting LU decomposition, provided
34   * by class FullPivLU.
35   *
36   * This is \b not a rank-revealing LU decomposition. Many features are intentionally absent from this class,
37   * such as rank computation. If you need these features, use class FullPivLU.
38   *
39   * This LU decomposition is suitable to invert invertible matrices. It is what MatrixBase::inverse() uses
40   * in the general case.
41   * On the other hand, it is \b not suitable to determine whether a given matrix is invertible.
42   *
43   * The data of the LU decomposition can be directly accessed through the methods matrixLU(), permutationP().
44   *
45   * \sa MatrixBase::partialPivLu(), MatrixBase::determinant(), MatrixBase::inverse(), MatrixBase::computeInverse(), class FullPivLU
46   */
47 template<typename _MatrixType> class PartialPivLU
48 {
49   public:
50 
51     typedef _MatrixType MatrixType;
52     enum {
53       RowsAtCompileTime = MatrixType::RowsAtCompileTime,
54       ColsAtCompileTime = MatrixType::ColsAtCompileTime,
55       Options = MatrixType::Options,
56       MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
57       MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
58     };
59     typedef typename MatrixType::Scalar Scalar;
60     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
61     typedef typename internal::traits<MatrixType>::StorageKind StorageKind;
62     typedef typename MatrixType::Index Index;
63     typedef PermutationMatrix<RowsAtCompileTime, MaxRowsAtCompileTime> PermutationType;
64     typedef Transpositions<RowsAtCompileTime, MaxRowsAtCompileTime> TranspositionType;
65 
66 
67     /**
68     * \brief Default Constructor.
69     *
70     * The default constructor is useful in cases in which the user intends to
71     * perform decompositions via PartialPivLU::compute(const MatrixType&).
72     */
73     PartialPivLU();
74 
75     /** \brief Default Constructor with memory preallocation
76       *
77       * Like the default constructor but with preallocation of the internal data
78       * according to the specified problem \a size.
79       * \sa PartialPivLU()
80       */
81     PartialPivLU(Index size);
82 
83     /** Constructor.
84       *
85       * \param matrix the matrix of which to compute the LU decomposition.
86       *
87       * \warning The matrix should have full rank (e.g. if it's square, it should be invertible).
88       * If you need to deal with non-full rank, use class FullPivLU instead.
89       */
90     PartialPivLU(const MatrixType& matrix);
91 
92     PartialPivLU& compute(const MatrixType& matrix);
93 
94     /** \returns the LU decomposition matrix: the upper-triangular part is U, the
95       * unit-lower-triangular part is L (at least for square matrices; in the non-square
96       * case, special care is needed, see the documentation of class FullPivLU).
97       *
98       * \sa matrixL(), matrixU()
99       */
matrixLU()100     inline const MatrixType& matrixLU() const
101     {
102       eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
103       return m_lu;
104     }
105 
106     /** \returns the permutation matrix P.
107       */
permutationP()108     inline const PermutationType& permutationP() const
109     {
110       eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
111       return m_p;
112     }
113 
114     /** This method returns the solution x to the equation Ax=b, where A is the matrix of which
115       * *this is the LU decomposition.
116       *
117       * \param b the right-hand-side of the equation to solve. Can be a vector or a matrix,
118       *          the only requirement in order for the equation to make sense is that
119       *          b.rows()==A.rows(), where A is the matrix of which *this is the LU decomposition.
120       *
121       * \returns the solution.
122       *
123       * Example: \include PartialPivLU_solve.cpp
124       * Output: \verbinclude PartialPivLU_solve.out
125       *
126       * Since this PartialPivLU class assumes anyway that the matrix A is invertible, the solution
127       * theoretically exists and is unique regardless of b.
128       *
129       * \sa TriangularView::solve(), inverse(), computeInverse()
130       */
131     template<typename Rhs>
132     inline const internal::solve_retval<PartialPivLU, Rhs>
solve(const MatrixBase<Rhs> & b)133     solve(const MatrixBase<Rhs>& b) const
134     {
135       eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
136       return internal::solve_retval<PartialPivLU, Rhs>(*this, b.derived());
137     }
138 
139     /** \returns the inverse of the matrix of which *this is the LU decomposition.
140       *
141       * \warning The matrix being decomposed here is assumed to be invertible. If you need to check for
142       *          invertibility, use class FullPivLU instead.
143       *
144       * \sa MatrixBase::inverse(), LU::inverse()
145       */
inverse()146     inline const internal::solve_retval<PartialPivLU,typename MatrixType::IdentityReturnType> inverse() const
147     {
148       eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
149       return internal::solve_retval<PartialPivLU,typename MatrixType::IdentityReturnType>
150                (*this, MatrixType::Identity(m_lu.rows(), m_lu.cols()));
151     }
152 
153     /** \returns the determinant of the matrix of which
154       * *this is the LU decomposition. It has only linear complexity
155       * (that is, O(n) where n is the dimension of the square matrix)
156       * as the LU decomposition has already been computed.
157       *
158       * \note For fixed-size matrices of size up to 4, MatrixBase::determinant() offers
159       *       optimized paths.
160       *
161       * \warning a determinant can be very big or small, so for matrices
162       * of large enough dimension, there is a risk of overflow/underflow.
163       *
164       * \sa MatrixBase::determinant()
165       */
166     typename internal::traits<MatrixType>::Scalar determinant() const;
167 
168     MatrixType reconstructedMatrix() const;
169 
rows()170     inline Index rows() const { return m_lu.rows(); }
cols()171     inline Index cols() const { return m_lu.cols(); }
172 
173   protected:
174     MatrixType m_lu;
175     PermutationType m_p;
176     TranspositionType m_rowsTranspositions;
177     Index m_det_p;
178     bool m_isInitialized;
179 };
180 
181 template<typename MatrixType>
PartialPivLU()182 PartialPivLU<MatrixType>::PartialPivLU()
183   : m_lu(),
184     m_p(),
185     m_rowsTranspositions(),
186     m_det_p(0),
187     m_isInitialized(false)
188 {
189 }
190 
191 template<typename MatrixType>
PartialPivLU(Index size)192 PartialPivLU<MatrixType>::PartialPivLU(Index size)
193   : m_lu(size, size),
194     m_p(size),
195     m_rowsTranspositions(size),
196     m_det_p(0),
197     m_isInitialized(false)
198 {
199 }
200 
201 template<typename MatrixType>
PartialPivLU(const MatrixType & matrix)202 PartialPivLU<MatrixType>::PartialPivLU(const MatrixType& matrix)
203   : m_lu(matrix.rows(), matrix.rows()),
204     m_p(matrix.rows()),
205     m_rowsTranspositions(matrix.rows()),
206     m_det_p(0),
207     m_isInitialized(false)
208 {
209   compute(matrix);
210 }
211 
212 namespace internal {
213 
214 /** \internal This is the blocked version of fullpivlu_unblocked() */
215 template<typename Scalar, int StorageOrder, typename PivIndex>
216 struct partial_lu_impl
217 {
218   // FIXME add a stride to Map, so that the following mapping becomes easier,
219   // another option would be to create an expression being able to automatically
220   // warp any Map, Matrix, and Block expressions as a unique type, but since that's exactly
221   // a Map + stride, why not adding a stride to Map, and convenient ctors from a Matrix,
222   // and Block.
223   typedef Map<Matrix<Scalar, Dynamic, Dynamic, StorageOrder> > MapLU;
224   typedef Block<MapLU, Dynamic, Dynamic> MatrixType;
225   typedef Block<MatrixType,Dynamic,Dynamic> BlockType;
226   typedef typename MatrixType::RealScalar RealScalar;
227   typedef typename MatrixType::Index Index;
228 
229   /** \internal performs the LU decomposition in-place of the matrix \a lu
230     * using an unblocked algorithm.
231     *
232     * In addition, this function returns the row transpositions in the
233     * vector \a row_transpositions which must have a size equal to the number
234     * of columns of the matrix \a lu, and an integer \a nb_transpositions
235     * which returns the actual number of transpositions.
236     *
237     * \returns The index of the first pivot which is exactly zero if any, or a negative number otherwise.
238     */
unblocked_lupartial_lu_impl239   static Index unblocked_lu(MatrixType& lu, PivIndex* row_transpositions, PivIndex& nb_transpositions)
240   {
241     const Index rows = lu.rows();
242     const Index cols = lu.cols();
243     const Index size = (std::min)(rows,cols);
244     nb_transpositions = 0;
245     Index first_zero_pivot = -1;
246     for(Index k = 0; k < size; ++k)
247     {
248       Index rrows = rows-k-1;
249       Index rcols = cols-k-1;
250 
251       Index row_of_biggest_in_col;
252       RealScalar biggest_in_corner
253         = lu.col(k).tail(rows-k).cwiseAbs().maxCoeff(&row_of_biggest_in_col);
254       row_of_biggest_in_col += k;
255 
256       row_transpositions[k] = PivIndex(row_of_biggest_in_col);
257 
258       if(biggest_in_corner != RealScalar(0))
259       {
260         if(k != row_of_biggest_in_col)
261         {
262           lu.row(k).swap(lu.row(row_of_biggest_in_col));
263           ++nb_transpositions;
264         }
265 
266         // FIXME shall we introduce a safe quotient expression in cas 1/lu.coeff(k,k)
267         // overflow but not the actual quotient?
268         lu.col(k).tail(rrows) /= lu.coeff(k,k);
269       }
270       else if(first_zero_pivot==-1)
271       {
272         // the pivot is exactly zero, we record the index of the first pivot which is exactly 0,
273         // and continue the factorization such we still have A = PLU
274         first_zero_pivot = k;
275       }
276 
277       if(k<rows-1)
278         lu.bottomRightCorner(rrows,rcols).noalias() -= lu.col(k).tail(rrows) * lu.row(k).tail(rcols);
279     }
280     return first_zero_pivot;
281   }
282 
283   /** \internal performs the LU decomposition in-place of the matrix represented
284     * by the variables \a rows, \a cols, \a lu_data, and \a lu_stride using a
285     * recursive, blocked algorithm.
286     *
287     * In addition, this function returns the row transpositions in the
288     * vector \a row_transpositions which must have a size equal to the number
289     * of columns of the matrix \a lu, and an integer \a nb_transpositions
290     * which returns the actual number of transpositions.
291     *
292     * \returns The index of the first pivot which is exactly zero if any, or a negative number otherwise.
293     *
294     * \note This very low level interface using pointers, etc. is to:
295     *   1 - reduce the number of instanciations to the strict minimum
296     *   2 - avoid infinite recursion of the instanciations with Block<Block<Block<...> > >
297     */
298   static Index blocked_lu(Index rows, Index cols, Scalar* lu_data, Index luStride, PivIndex* row_transpositions, PivIndex& nb_transpositions, Index maxBlockSize=256)
299   {
300     MapLU lu1(lu_data,StorageOrder==RowMajor?rows:luStride,StorageOrder==RowMajor?luStride:cols);
301     MatrixType lu(lu1,0,0,rows,cols);
302 
303     const Index size = (std::min)(rows,cols);
304 
305     // if the matrix is too small, no blocking:
306     if(size<=16)
307     {
308       return unblocked_lu(lu, row_transpositions, nb_transpositions);
309     }
310 
311     // automatically adjust the number of subdivisions to the size
312     // of the matrix so that there is enough sub blocks:
313     Index blockSize;
314     {
315       blockSize = size/8;
316       blockSize = (blockSize/16)*16;
317       blockSize = (std::min)((std::max)(blockSize,Index(8)), maxBlockSize);
318     }
319 
320     nb_transpositions = 0;
321     Index first_zero_pivot = -1;
322     for(Index k = 0; k < size; k+=blockSize)
323     {
324       Index bs = (std::min)(size-k,blockSize); // actual size of the block
325       Index trows = rows - k - bs; // trailing rows
326       Index tsize = size - k - bs; // trailing size
327 
328       // partition the matrix:
329       //                          A00 | A01 | A02
330       // lu  = A_0 | A_1 | A_2 =  A10 | A11 | A12
331       //                          A20 | A21 | A22
332       BlockType A_0(lu,0,0,rows,k);
333       BlockType A_2(lu,0,k+bs,rows,tsize);
334       BlockType A11(lu,k,k,bs,bs);
335       BlockType A12(lu,k,k+bs,bs,tsize);
336       BlockType A21(lu,k+bs,k,trows,bs);
337       BlockType A22(lu,k+bs,k+bs,trows,tsize);
338 
339       PivIndex nb_transpositions_in_panel;
340       // recursively call the blocked LU algorithm on [A11^T A21^T]^T
341       // with a very small blocking size:
342       Index ret = blocked_lu(trows+bs, bs, &lu.coeffRef(k,k), luStride,
343                    row_transpositions+k, nb_transpositions_in_panel, 16);
344       if(ret>=0 && first_zero_pivot==-1)
345         first_zero_pivot = k+ret;
346 
347       nb_transpositions += nb_transpositions_in_panel;
348       // update permutations and apply them to A_0
349       for(Index i=k; i<k+bs; ++i)
350       {
351         Index piv = (row_transpositions[i] += k);
352         A_0.row(i).swap(A_0.row(piv));
353       }
354 
355       if(trows)
356       {
357         // apply permutations to A_2
358         for(Index i=k;i<k+bs; ++i)
359           A_2.row(i).swap(A_2.row(row_transpositions[i]));
360 
361         // A12 = A11^-1 A12
362         A11.template triangularView<UnitLower>().solveInPlace(A12);
363 
364         A22.noalias() -= A21 * A12;
365       }
366     }
367     return first_zero_pivot;
368   }
369 };
370 
371 /** \internal performs the LU decomposition with partial pivoting in-place.
372   */
373 template<typename MatrixType, typename TranspositionType>
partial_lu_inplace(MatrixType & lu,TranspositionType & row_transpositions,typename TranspositionType::Index & nb_transpositions)374 void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, typename TranspositionType::Index& nb_transpositions)
375 {
376   eigen_assert(lu.cols() == row_transpositions.size());
377   eigen_assert((&row_transpositions.coeffRef(1)-&row_transpositions.coeffRef(0)) == 1);
378 
379   partial_lu_impl
380     <typename MatrixType::Scalar, MatrixType::Flags&RowMajorBit?RowMajor:ColMajor, typename TranspositionType::Index>
381     ::blocked_lu(lu.rows(), lu.cols(), &lu.coeffRef(0,0), lu.outerStride(), &row_transpositions.coeffRef(0), nb_transpositions);
382 }
383 
384 } // end namespace internal
385 
386 template<typename MatrixType>
compute(const MatrixType & matrix)387 PartialPivLU<MatrixType>& PartialPivLU<MatrixType>::compute(const MatrixType& matrix)
388 {
389   // the row permutation is stored as int indices, so just to be sure:
390   eigen_assert(matrix.rows()<NumTraits<int>::highest());
391 
392   m_lu = matrix;
393 
394   eigen_assert(matrix.rows() == matrix.cols() && "PartialPivLU is only for square (and moreover invertible) matrices");
395   const Index size = matrix.rows();
396 
397   m_rowsTranspositions.resize(size);
398 
399   typename TranspositionType::Index nb_transpositions;
400   internal::partial_lu_inplace(m_lu, m_rowsTranspositions, nb_transpositions);
401   m_det_p = (nb_transpositions%2) ? -1 : 1;
402 
403   m_p = m_rowsTranspositions;
404 
405   m_isInitialized = true;
406   return *this;
407 }
408 
409 template<typename MatrixType>
determinant()410 typename internal::traits<MatrixType>::Scalar PartialPivLU<MatrixType>::determinant() const
411 {
412   eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
413   return Scalar(m_det_p) * m_lu.diagonal().prod();
414 }
415 
416 /** \returns the matrix represented by the decomposition,
417  * i.e., it returns the product: P^{-1} L U.
418  * This function is provided for debug purpose. */
419 template<typename MatrixType>
reconstructedMatrix()420 MatrixType PartialPivLU<MatrixType>::reconstructedMatrix() const
421 {
422   eigen_assert(m_isInitialized && "LU is not initialized.");
423   // LU
424   MatrixType res = m_lu.template triangularView<UnitLower>().toDenseMatrix()
425                  * m_lu.template triangularView<Upper>();
426 
427   // P^{-1}(LU)
428   res = m_p.inverse() * res;
429 
430   return res;
431 }
432 
433 /***** Implementation of solve() *****************************************************/
434 
435 namespace internal {
436 
437 template<typename _MatrixType, typename Rhs>
438 struct solve_retval<PartialPivLU<_MatrixType>, Rhs>
439   : solve_retval_base<PartialPivLU<_MatrixType>, Rhs>
440 {
441   EIGEN_MAKE_SOLVE_HELPERS(PartialPivLU<_MatrixType>,Rhs)
442 
443   template<typename Dest> void evalTo(Dest& dst) const
444   {
445     /* The decomposition PA = LU can be rewritten as A = P^{-1} L U.
446     * So we proceed as follows:
447     * Step 1: compute c = Pb.
448     * Step 2: replace c by the solution x to Lx = c.
449     * Step 3: replace c by the solution x to Ux = c.
450     */
451 
452     eigen_assert(rhs().rows() == dec().matrixLU().rows());
453 
454     // Step 1
455     dst = dec().permutationP() * rhs();
456 
457     // Step 2
458     dec().matrixLU().template triangularView<UnitLower>().solveInPlace(dst);
459 
460     // Step 3
461     dec().matrixLU().template triangularView<Upper>().solveInPlace(dst);
462   }
463 };
464 
465 } // end namespace internal
466 
467 /******** MatrixBase methods *******/
468 
469 /** \lu_module
470   *
471   * \return the partial-pivoting LU decomposition of \c *this.
472   *
473   * \sa class PartialPivLU
474   */
475 template<typename Derived>
476 inline const PartialPivLU<typename MatrixBase<Derived>::PlainObject>
477 MatrixBase<Derived>::partialPivLu() const
478 {
479   return PartialPivLU<PlainObject>(eval());
480 }
481 
482 #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
483 /** \lu_module
484   *
485   * Synonym of partialPivLu().
486   *
487   * \return the partial-pivoting LU decomposition of \c *this.
488   *
489   * \sa class PartialPivLU
490   */
491 template<typename Derived>
492 inline const PartialPivLU<typename MatrixBase<Derived>::PlainObject>
493 MatrixBase<Derived>::lu() const
494 {
495   return PartialPivLU<PlainObject>(eval());
496 }
497 #endif
498 
499 } // end namespace Eigen
500 
501 #endif // EIGEN_PARTIALLU_H
502