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 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_LU_H
11 #define EIGEN_LU_H
12 
13 namespace Eigen {
14 
15 /** \ingroup LU_Module
16   *
17   * \class FullPivLU
18   *
19   * \brief LU decomposition of a matrix with complete pivoting, and related features
20   *
21   * \param MatrixType the type of the matrix of which we are computing the LU decomposition
22   *
23   * This class represents a LU decomposition of any matrix, with complete pivoting: the matrix A is
24   * decomposed as \f$ A = P^{-1} L U Q^{-1} \f$ where L is unit-lower-triangular, U is
25   * upper-triangular, and P and Q are permutation matrices. This is a rank-revealing LU
26   * decomposition. The eigenvalues (diagonal coefficients) of U are sorted in such a way that any
27   * zeros are at the end.
28   *
29   * This decomposition provides the generic approach to solving systems of linear equations, computing
30   * the rank, invertibility, inverse, kernel, and determinant.
31   *
32   * This LU decomposition is very stable and well tested with large matrices. However there are use cases where the SVD
33   * decomposition is inherently more stable and/or flexible. For example, when computing the kernel of a matrix,
34   * working with the SVD allows to select the smallest singular values of the matrix, something that
35   * the LU decomposition doesn't see.
36   *
37   * The data of the LU decomposition can be directly accessed through the methods matrixLU(),
38   * permutationP(), permutationQ().
39   *
40   * As an exemple, here is how the original matrix can be retrieved:
41   * \include class_FullPivLU.cpp
42   * Output: \verbinclude class_FullPivLU.out
43   *
44   * \sa MatrixBase::fullPivLu(), MatrixBase::determinant(), MatrixBase::inverse()
45   */
46 template<typename _MatrixType> class FullPivLU
47 {
48   public:
49     typedef _MatrixType MatrixType;
50     enum {
51       RowsAtCompileTime = MatrixType::RowsAtCompileTime,
52       ColsAtCompileTime = MatrixType::ColsAtCompileTime,
53       Options = MatrixType::Options,
54       MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
55       MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
56     };
57     typedef typename MatrixType::Scalar Scalar;
58     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
59     typedef typename internal::traits<MatrixType>::StorageKind StorageKind;
60     typedef typename MatrixType::Index Index;
61     typedef typename internal::plain_row_type<MatrixType, Index>::type IntRowVectorType;
62     typedef typename internal::plain_col_type<MatrixType, Index>::type IntColVectorType;
63     typedef PermutationMatrix<ColsAtCompileTime, MaxColsAtCompileTime> PermutationQType;
64     typedef PermutationMatrix<RowsAtCompileTime, MaxRowsAtCompileTime> PermutationPType;
65 
66     /**
67       * \brief Default Constructor.
68       *
69       * The default constructor is useful in cases in which the user intends to
70       * perform decompositions via LU::compute(const MatrixType&).
71       */
72     FullPivLU();
73 
74     /** \brief Default Constructor with memory preallocation
75       *
76       * Like the default constructor but with preallocation of the internal data
77       * according to the specified problem \a size.
78       * \sa FullPivLU()
79       */
80     FullPivLU(Index rows, Index cols);
81 
82     /** Constructor.
83       *
84       * \param matrix the matrix of which to compute the LU decomposition.
85       *               It is required to be nonzero.
86       */
87     FullPivLU(const MatrixType& matrix);
88 
89     /** Computes the LU decomposition of the given matrix.
90       *
91       * \param matrix the matrix of which to compute the LU decomposition.
92       *               It is required to be nonzero.
93       *
94       * \returns a reference to *this
95       */
96     FullPivLU& compute(const MatrixType& matrix);
97 
98     /** \returns the LU decomposition matrix: the upper-triangular part is U, the
99       * unit-lower-triangular part is L (at least for square matrices; in the non-square
100       * case, special care is needed, see the documentation of class FullPivLU).
101       *
102       * \sa matrixL(), matrixU()
103       */
matrixLU()104     inline const MatrixType& matrixLU() const
105     {
106       eigen_assert(m_isInitialized && "LU is not initialized.");
107       return m_lu;
108     }
109 
110     /** \returns the number of nonzero pivots in the LU decomposition.
111       * Here nonzero is meant in the exact sense, not in a fuzzy sense.
112       * So that notion isn't really intrinsically interesting, but it is
113       * still useful when implementing algorithms.
114       *
115       * \sa rank()
116       */
nonzeroPivots()117     inline Index nonzeroPivots() const
118     {
119       eigen_assert(m_isInitialized && "LU is not initialized.");
120       return m_nonzero_pivots;
121     }
122 
123     /** \returns the absolute value of the biggest pivot, i.e. the biggest
124       *          diagonal coefficient of U.
125       */
maxPivot()126     RealScalar maxPivot() const { return m_maxpivot; }
127 
128     /** \returns the permutation matrix P
129       *
130       * \sa permutationQ()
131       */
permutationP()132     inline const PermutationPType& permutationP() const
133     {
134       eigen_assert(m_isInitialized && "LU is not initialized.");
135       return m_p;
136     }
137 
138     /** \returns the permutation matrix Q
139       *
140       * \sa permutationP()
141       */
permutationQ()142     inline const PermutationQType& permutationQ() const
143     {
144       eigen_assert(m_isInitialized && "LU is not initialized.");
145       return m_q;
146     }
147 
148     /** \returns the kernel of the matrix, also called its null-space. The columns of the returned matrix
149       * will form a basis of the kernel.
150       *
151       * \note If the kernel has dimension zero, then the returned matrix is a column-vector filled with zeros.
152       *
153       * \note This method has to determine which pivots should be considered nonzero.
154       *       For that, it uses the threshold value that you can control by calling
155       *       setThreshold(const RealScalar&).
156       *
157       * Example: \include FullPivLU_kernel.cpp
158       * Output: \verbinclude FullPivLU_kernel.out
159       *
160       * \sa image()
161       */
kernel()162     inline const internal::kernel_retval<FullPivLU> kernel() const
163     {
164       eigen_assert(m_isInitialized && "LU is not initialized.");
165       return internal::kernel_retval<FullPivLU>(*this);
166     }
167 
168     /** \returns the image of the matrix, also called its column-space. The columns of the returned matrix
169       * will form a basis of the kernel.
170       *
171       * \param originalMatrix the original matrix, of which *this is the LU decomposition.
172       *                       The reason why it is needed to pass it here, is that this allows
173       *                       a large optimization, as otherwise this method would need to reconstruct it
174       *                       from the LU decomposition.
175       *
176       * \note If the image has dimension zero, then the returned matrix is a column-vector filled with zeros.
177       *
178       * \note This method has to determine which pivots should be considered nonzero.
179       *       For that, it uses the threshold value that you can control by calling
180       *       setThreshold(const RealScalar&).
181       *
182       * Example: \include FullPivLU_image.cpp
183       * Output: \verbinclude FullPivLU_image.out
184       *
185       * \sa kernel()
186       */
187     inline const internal::image_retval<FullPivLU>
image(const MatrixType & originalMatrix)188       image(const MatrixType& originalMatrix) const
189     {
190       eigen_assert(m_isInitialized && "LU is not initialized.");
191       return internal::image_retval<FullPivLU>(*this, originalMatrix);
192     }
193 
194     /** \return a solution x to the equation Ax=b, where A is the matrix of which
195       * *this is the LU decomposition.
196       *
197       * \param b the right-hand-side of the equation to solve. Can be a vector or a matrix,
198       *          the only requirement in order for the equation to make sense is that
199       *          b.rows()==A.rows(), where A is the matrix of which *this is the LU decomposition.
200       *
201       * \returns a solution.
202       *
203       * \note_about_checking_solutions
204       *
205       * \note_about_arbitrary_choice_of_solution
206       * \note_about_using_kernel_to_study_multiple_solutions
207       *
208       * Example: \include FullPivLU_solve.cpp
209       * Output: \verbinclude FullPivLU_solve.out
210       *
211       * \sa TriangularView::solve(), kernel(), inverse()
212       */
213     template<typename Rhs>
214     inline const internal::solve_retval<FullPivLU, Rhs>
solve(const MatrixBase<Rhs> & b)215     solve(const MatrixBase<Rhs>& b) const
216     {
217       eigen_assert(m_isInitialized && "LU is not initialized.");
218       return internal::solve_retval<FullPivLU, Rhs>(*this, b.derived());
219     }
220 
221     /** \returns the determinant of the matrix of which
222       * *this is the LU decomposition. It has only linear complexity
223       * (that is, O(n) where n is the dimension of the square matrix)
224       * as the LU decomposition has already been computed.
225       *
226       * \note This is only for square matrices.
227       *
228       * \note For fixed-size matrices of size up to 4, MatrixBase::determinant() offers
229       *       optimized paths.
230       *
231       * \warning a determinant can be very big or small, so for matrices
232       * of large enough dimension, there is a risk of overflow/underflow.
233       *
234       * \sa MatrixBase::determinant()
235       */
236     typename internal::traits<MatrixType>::Scalar determinant() const;
237 
238     /** Allows to prescribe a threshold to be used by certain methods, such as rank(),
239       * who need to determine when pivots are to be considered nonzero. This is not used for the
240       * LU decomposition itself.
241       *
242       * When it needs to get the threshold value, Eigen calls threshold(). By default, this
243       * uses a formula to automatically determine a reasonable threshold.
244       * Once you have called the present method setThreshold(const RealScalar&),
245       * your value is used instead.
246       *
247       * \param threshold The new value to use as the threshold.
248       *
249       * A pivot will be considered nonzero if its absolute value is strictly greater than
250       *  \f$ \vert pivot \vert \leqslant threshold \times \vert maxpivot \vert \f$
251       * where maxpivot is the biggest pivot.
252       *
253       * If you want to come back to the default behavior, call setThreshold(Default_t)
254       */
setThreshold(const RealScalar & threshold)255     FullPivLU& setThreshold(const RealScalar& threshold)
256     {
257       m_usePrescribedThreshold = true;
258       m_prescribedThreshold = threshold;
259       return *this;
260     }
261 
262     /** Allows to come back to the default behavior, letting Eigen use its default formula for
263       * determining the threshold.
264       *
265       * You should pass the special object Eigen::Default as parameter here.
266       * \code lu.setThreshold(Eigen::Default); \endcode
267       *
268       * See the documentation of setThreshold(const RealScalar&).
269       */
setThreshold(Default_t)270     FullPivLU& setThreshold(Default_t)
271     {
272       m_usePrescribedThreshold = false;
273       return *this;
274     }
275 
276     /** Returns the threshold that will be used by certain methods such as rank().
277       *
278       * See the documentation of setThreshold(const RealScalar&).
279       */
threshold()280     RealScalar threshold() const
281     {
282       eigen_assert(m_isInitialized || m_usePrescribedThreshold);
283       return m_usePrescribedThreshold ? m_prescribedThreshold
284       // this formula comes from experimenting (see "LU precision tuning" thread on the list)
285       // and turns out to be identical to Higham's formula used already in LDLt.
286                                       : NumTraits<Scalar>::epsilon() * m_lu.diagonalSize();
287     }
288 
289     /** \returns the rank of the matrix of which *this is the LU decomposition.
290       *
291       * \note This method has to determine which pivots should be considered nonzero.
292       *       For that, it uses the threshold value that you can control by calling
293       *       setThreshold(const RealScalar&).
294       */
rank()295     inline Index rank() const
296     {
297       using std::abs;
298       eigen_assert(m_isInitialized && "LU is not initialized.");
299       RealScalar premultiplied_threshold = abs(m_maxpivot) * threshold();
300       Index result = 0;
301       for(Index i = 0; i < m_nonzero_pivots; ++i)
302         result += (abs(m_lu.coeff(i,i)) > premultiplied_threshold);
303       return result;
304     }
305 
306     /** \returns the dimension of the kernel of the matrix of which *this is the LU decomposition.
307       *
308       * \note This method has to determine which pivots should be considered nonzero.
309       *       For that, it uses the threshold value that you can control by calling
310       *       setThreshold(const RealScalar&).
311       */
dimensionOfKernel()312     inline Index dimensionOfKernel() const
313     {
314       eigen_assert(m_isInitialized && "LU is not initialized.");
315       return cols() - rank();
316     }
317 
318     /** \returns true if the matrix of which *this is the LU decomposition represents an injective
319       *          linear map, i.e. has trivial kernel; false otherwise.
320       *
321       * \note This method has to determine which pivots should be considered nonzero.
322       *       For that, it uses the threshold value that you can control by calling
323       *       setThreshold(const RealScalar&).
324       */
isInjective()325     inline bool isInjective() const
326     {
327       eigen_assert(m_isInitialized && "LU is not initialized.");
328       return rank() == cols();
329     }
330 
331     /** \returns true if the matrix of which *this is the LU decomposition represents a surjective
332       *          linear map; false otherwise.
333       *
334       * \note This method has to determine which pivots should be considered nonzero.
335       *       For that, it uses the threshold value that you can control by calling
336       *       setThreshold(const RealScalar&).
337       */
isSurjective()338     inline bool isSurjective() const
339     {
340       eigen_assert(m_isInitialized && "LU is not initialized.");
341       return rank() == rows();
342     }
343 
344     /** \returns true if the matrix of which *this is the LU decomposition is invertible.
345       *
346       * \note This method has to determine which pivots should be considered nonzero.
347       *       For that, it uses the threshold value that you can control by calling
348       *       setThreshold(const RealScalar&).
349       */
isInvertible()350     inline bool isInvertible() const
351     {
352       eigen_assert(m_isInitialized && "LU is not initialized.");
353       return isInjective() && (m_lu.rows() == m_lu.cols());
354     }
355 
356     /** \returns the inverse of the matrix of which *this is the LU decomposition.
357       *
358       * \note If this matrix is not invertible, the returned matrix has undefined coefficients.
359       *       Use isInvertible() to first determine whether this matrix is invertible.
360       *
361       * \sa MatrixBase::inverse()
362       */
inverse()363     inline const internal::solve_retval<FullPivLU,typename MatrixType::IdentityReturnType> inverse() const
364     {
365       eigen_assert(m_isInitialized && "LU is not initialized.");
366       eigen_assert(m_lu.rows() == m_lu.cols() && "You can't take the inverse of a non-square matrix!");
367       return internal::solve_retval<FullPivLU,typename MatrixType::IdentityReturnType>
368                (*this, MatrixType::Identity(m_lu.rows(), m_lu.cols()));
369     }
370 
371     MatrixType reconstructedMatrix() const;
372 
rows()373     inline Index rows() const { return m_lu.rows(); }
cols()374     inline Index cols() const { return m_lu.cols(); }
375 
376   protected:
377     MatrixType m_lu;
378     PermutationPType m_p;
379     PermutationQType m_q;
380     IntColVectorType m_rowsTranspositions;
381     IntRowVectorType m_colsTranspositions;
382     Index m_det_pq, m_nonzero_pivots;
383     RealScalar m_maxpivot, m_prescribedThreshold;
384     bool m_isInitialized, m_usePrescribedThreshold;
385 };
386 
387 template<typename MatrixType>
FullPivLU()388 FullPivLU<MatrixType>::FullPivLU()
389   : m_isInitialized(false), m_usePrescribedThreshold(false)
390 {
391 }
392 
393 template<typename MatrixType>
FullPivLU(Index rows,Index cols)394 FullPivLU<MatrixType>::FullPivLU(Index rows, Index cols)
395   : m_lu(rows, cols),
396     m_p(rows),
397     m_q(cols),
398     m_rowsTranspositions(rows),
399     m_colsTranspositions(cols),
400     m_isInitialized(false),
401     m_usePrescribedThreshold(false)
402 {
403 }
404 
405 template<typename MatrixType>
FullPivLU(const MatrixType & matrix)406 FullPivLU<MatrixType>::FullPivLU(const MatrixType& matrix)
407   : m_lu(matrix.rows(), matrix.cols()),
408     m_p(matrix.rows()),
409     m_q(matrix.cols()),
410     m_rowsTranspositions(matrix.rows()),
411     m_colsTranspositions(matrix.cols()),
412     m_isInitialized(false),
413     m_usePrescribedThreshold(false)
414 {
415   compute(matrix);
416 }
417 
418 template<typename MatrixType>
compute(const MatrixType & matrix)419 FullPivLU<MatrixType>& FullPivLU<MatrixType>::compute(const MatrixType& matrix)
420 {
421   // the permutations are stored as int indices, so just to be sure:
422   eigen_assert(matrix.rows()<=NumTraits<int>::highest() && matrix.cols()<=NumTraits<int>::highest());
423 
424   m_isInitialized = true;
425   m_lu = matrix;
426 
427   const Index size = matrix.diagonalSize();
428   const Index rows = matrix.rows();
429   const Index cols = matrix.cols();
430 
431   // will store the transpositions, before we accumulate them at the end.
432   // can't accumulate on-the-fly because that will be done in reverse order for the rows.
433   m_rowsTranspositions.resize(matrix.rows());
434   m_colsTranspositions.resize(matrix.cols());
435   Index number_of_transpositions = 0; // number of NONTRIVIAL transpositions, i.e. m_rowsTranspositions[i]!=i
436 
437   m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case)
438   m_maxpivot = RealScalar(0);
439 
440   for(Index k = 0; k < size; ++k)
441   {
442     // First, we need to find the pivot.
443 
444     // biggest coefficient in the remaining bottom-right corner (starting at row k, col k)
445     Index row_of_biggest_in_corner, col_of_biggest_in_corner;
446     RealScalar biggest_in_corner;
447     biggest_in_corner = m_lu.bottomRightCorner(rows-k, cols-k)
448                         .cwiseAbs()
449                         .maxCoeff(&row_of_biggest_in_corner, &col_of_biggest_in_corner);
450     row_of_biggest_in_corner += k; // correct the values! since they were computed in the corner,
451     col_of_biggest_in_corner += k; // need to add k to them.
452 
453     if(biggest_in_corner==RealScalar(0))
454     {
455       // before exiting, make sure to initialize the still uninitialized transpositions
456       // in a sane state without destroying what we already have.
457       m_nonzero_pivots = k;
458       for(Index i = k; i < size; ++i)
459       {
460         m_rowsTranspositions.coeffRef(i) = i;
461         m_colsTranspositions.coeffRef(i) = i;
462       }
463       break;
464     }
465 
466     if(biggest_in_corner > m_maxpivot) m_maxpivot = biggest_in_corner;
467 
468     // Now that we've found the pivot, we need to apply the row/col swaps to
469     // bring it to the location (k,k).
470 
471     m_rowsTranspositions.coeffRef(k) = row_of_biggest_in_corner;
472     m_colsTranspositions.coeffRef(k) = col_of_biggest_in_corner;
473     if(k != row_of_biggest_in_corner) {
474       m_lu.row(k).swap(m_lu.row(row_of_biggest_in_corner));
475       ++number_of_transpositions;
476     }
477     if(k != col_of_biggest_in_corner) {
478       m_lu.col(k).swap(m_lu.col(col_of_biggest_in_corner));
479       ++number_of_transpositions;
480     }
481 
482     // Now that the pivot is at the right location, we update the remaining
483     // bottom-right corner by Gaussian elimination.
484 
485     if(k<rows-1)
486       m_lu.col(k).tail(rows-k-1) /= m_lu.coeff(k,k);
487     if(k<size-1)
488       m_lu.block(k+1,k+1,rows-k-1,cols-k-1).noalias() -= m_lu.col(k).tail(rows-k-1) * m_lu.row(k).tail(cols-k-1);
489   }
490 
491   // the main loop is over, we still have to accumulate the transpositions to find the
492   // permutations P and Q
493 
494   m_p.setIdentity(rows);
495   for(Index k = size-1; k >= 0; --k)
496     m_p.applyTranspositionOnTheRight(k, m_rowsTranspositions.coeff(k));
497 
498   m_q.setIdentity(cols);
499   for(Index k = 0; k < size; ++k)
500     m_q.applyTranspositionOnTheRight(k, m_colsTranspositions.coeff(k));
501 
502   m_det_pq = (number_of_transpositions%2) ? -1 : 1;
503   return *this;
504 }
505 
506 template<typename MatrixType>
determinant()507 typename internal::traits<MatrixType>::Scalar FullPivLU<MatrixType>::determinant() const
508 {
509   eigen_assert(m_isInitialized && "LU is not initialized.");
510   eigen_assert(m_lu.rows() == m_lu.cols() && "You can't take the determinant of a non-square matrix!");
511   return Scalar(m_det_pq) * Scalar(m_lu.diagonal().prod());
512 }
513 
514 /** \returns the matrix represented by the decomposition,
515  * i.e., it returns the product: \f$ P^{-1} L U Q^{-1} \f$.
516  * This function is provided for debug purposes. */
517 template<typename MatrixType>
reconstructedMatrix()518 MatrixType FullPivLU<MatrixType>::reconstructedMatrix() const
519 {
520   eigen_assert(m_isInitialized && "LU is not initialized.");
521   const Index smalldim = (std::min)(m_lu.rows(), m_lu.cols());
522   // LU
523   MatrixType res(m_lu.rows(),m_lu.cols());
524   // FIXME the .toDenseMatrix() should not be needed...
525   res = m_lu.leftCols(smalldim)
526             .template triangularView<UnitLower>().toDenseMatrix()
527       * m_lu.topRows(smalldim)
528             .template triangularView<Upper>().toDenseMatrix();
529 
530   // P^{-1}(LU)
531   res = m_p.inverse() * res;
532 
533   // (P^{-1}LU)Q^{-1}
534   res = res * m_q.inverse();
535 
536   return res;
537 }
538 
539 /********* Implementation of kernel() **************************************************/
540 
541 namespace internal {
542 template<typename _MatrixType>
543 struct kernel_retval<FullPivLU<_MatrixType> >
544   : kernel_retval_base<FullPivLU<_MatrixType> >
545 {
546   EIGEN_MAKE_KERNEL_HELPERS(FullPivLU<_MatrixType>)
547 
548   enum { MaxSmallDimAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(
549             MatrixType::MaxColsAtCompileTime,
550             MatrixType::MaxRowsAtCompileTime)
551   };
552 
553   template<typename Dest> void evalTo(Dest& dst) const
554   {
555     using std::abs;
556     const Index cols = dec().matrixLU().cols(), dimker = cols - rank();
557     if(dimker == 0)
558     {
559       // The Kernel is just {0}, so it doesn't have a basis properly speaking, but let's
560       // avoid crashing/asserting as that depends on floating point calculations. Let's
561       // just return a single column vector filled with zeros.
562       dst.setZero();
563       return;
564     }
565 
566     /* Let us use the following lemma:
567       *
568       * Lemma: If the matrix A has the LU decomposition PAQ = LU,
569       * then Ker A = Q(Ker U).
570       *
571       * Proof: trivial: just keep in mind that P, Q, L are invertible.
572       */
573 
574     /* Thus, all we need to do is to compute Ker U, and then apply Q.
575       *
576       * U is upper triangular, with eigenvalues sorted so that any zeros appear at the end.
577       * Thus, the diagonal of U ends with exactly
578       * dimKer zero's. Let us use that to construct dimKer linearly
579       * independent vectors in Ker U.
580       */
581 
582     Matrix<Index, Dynamic, 1, 0, MaxSmallDimAtCompileTime, 1> pivots(rank());
583     RealScalar premultiplied_threshold = dec().maxPivot() * dec().threshold();
584     Index p = 0;
585     for(Index i = 0; i < dec().nonzeroPivots(); ++i)
586       if(abs(dec().matrixLU().coeff(i,i)) > premultiplied_threshold)
587         pivots.coeffRef(p++) = i;
588     eigen_internal_assert(p == rank());
589 
590     // we construct a temporaty trapezoid matrix m, by taking the U matrix and
591     // permuting the rows and cols to bring the nonnegligible pivots to the top of
592     // the main diagonal. We need that to be able to apply our triangular solvers.
593     // FIXME when we get triangularView-for-rectangular-matrices, this can be simplified
594     Matrix<typename MatrixType::Scalar, Dynamic, Dynamic, MatrixType::Options,
595            MaxSmallDimAtCompileTime, MatrixType::MaxColsAtCompileTime>
596       m(dec().matrixLU().block(0, 0, rank(), cols));
597     for(Index i = 0; i < rank(); ++i)
598     {
599       if(i) m.row(i).head(i).setZero();
600       m.row(i).tail(cols-i) = dec().matrixLU().row(pivots.coeff(i)).tail(cols-i);
601     }
602     m.block(0, 0, rank(), rank());
603     m.block(0, 0, rank(), rank()).template triangularView<StrictlyLower>().setZero();
604     for(Index i = 0; i < rank(); ++i)
605       m.col(i).swap(m.col(pivots.coeff(i)));
606 
607     // ok, we have our trapezoid matrix, we can apply the triangular solver.
608     // notice that the math behind this suggests that we should apply this to the
609     // negative of the RHS, but for performance we just put the negative sign elsewhere, see below.
610     m.topLeftCorner(rank(), rank())
611      .template triangularView<Upper>().solveInPlace(
612         m.topRightCorner(rank(), dimker)
613       );
614 
615     // now we must undo the column permutation that we had applied!
616     for(Index i = rank()-1; i >= 0; --i)
617       m.col(i).swap(m.col(pivots.coeff(i)));
618 
619     // see the negative sign in the next line, that's what we were talking about above.
620     for(Index i = 0; i < rank(); ++i) dst.row(dec().permutationQ().indices().coeff(i)) = -m.row(i).tail(dimker);
621     for(Index i = rank(); i < cols; ++i) dst.row(dec().permutationQ().indices().coeff(i)).setZero();
622     for(Index k = 0; k < dimker; ++k) dst.coeffRef(dec().permutationQ().indices().coeff(rank()+k), k) = Scalar(1);
623   }
624 };
625 
626 /***** Implementation of image() *****************************************************/
627 
628 template<typename _MatrixType>
629 struct image_retval<FullPivLU<_MatrixType> >
630   : image_retval_base<FullPivLU<_MatrixType> >
631 {
632   EIGEN_MAKE_IMAGE_HELPERS(FullPivLU<_MatrixType>)
633 
634   enum { MaxSmallDimAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(
635             MatrixType::MaxColsAtCompileTime,
636             MatrixType::MaxRowsAtCompileTime)
637   };
638 
639   template<typename Dest> void evalTo(Dest& dst) const
640   {
641     using std::abs;
642     if(rank() == 0)
643     {
644       // The Image is just {0}, so it doesn't have a basis properly speaking, but let's
645       // avoid crashing/asserting as that depends on floating point calculations. Let's
646       // just return a single column vector filled with zeros.
647       dst.setZero();
648       return;
649     }
650 
651     Matrix<Index, Dynamic, 1, 0, MaxSmallDimAtCompileTime, 1> pivots(rank());
652     RealScalar premultiplied_threshold = dec().maxPivot() * dec().threshold();
653     Index p = 0;
654     for(Index i = 0; i < dec().nonzeroPivots(); ++i)
655       if(abs(dec().matrixLU().coeff(i,i)) > premultiplied_threshold)
656         pivots.coeffRef(p++) = i;
657     eigen_internal_assert(p == rank());
658 
659     for(Index i = 0; i < rank(); ++i)
660       dst.col(i) = originalMatrix().col(dec().permutationQ().indices().coeff(pivots.coeff(i)));
661   }
662 };
663 
664 /***** Implementation of solve() *****************************************************/
665 
666 template<typename _MatrixType, typename Rhs>
667 struct solve_retval<FullPivLU<_MatrixType>, Rhs>
668   : solve_retval_base<FullPivLU<_MatrixType>, Rhs>
669 {
670   EIGEN_MAKE_SOLVE_HELPERS(FullPivLU<_MatrixType>,Rhs)
671 
672   template<typename Dest> void evalTo(Dest& dst) const
673   {
674     /* The decomposition PAQ = LU can be rewritten as A = P^{-1} L U Q^{-1}.
675      * So we proceed as follows:
676      * Step 1: compute c = P * rhs.
677      * Step 2: replace c by the solution x to Lx = c. Exists because L is invertible.
678      * Step 3: replace c by the solution x to Ux = c. May or may not exist.
679      * Step 4: result = Q * c;
680      */
681 
682     const Index rows = dec().rows(), cols = dec().cols(),
683               nonzero_pivots = dec().nonzeroPivots();
684     eigen_assert(rhs().rows() == rows);
685     const Index smalldim = (std::min)(rows, cols);
686 
687     if(nonzero_pivots == 0)
688     {
689       dst.setZero();
690       return;
691     }
692 
693     typename Rhs::PlainObject c(rhs().rows(), rhs().cols());
694 
695     // Step 1
696     c = dec().permutationP() * rhs();
697 
698     // Step 2
699     dec().matrixLU()
700         .topLeftCorner(smalldim,smalldim)
701         .template triangularView<UnitLower>()
702         .solveInPlace(c.topRows(smalldim));
703     if(rows>cols)
704     {
705       c.bottomRows(rows-cols)
706         -= dec().matrixLU().bottomRows(rows-cols)
707          * c.topRows(cols);
708     }
709 
710     // Step 3
711     dec().matrixLU()
712         .topLeftCorner(nonzero_pivots, nonzero_pivots)
713         .template triangularView<Upper>()
714         .solveInPlace(c.topRows(nonzero_pivots));
715 
716     // Step 4
717     for(Index i = 0; i < nonzero_pivots; ++i)
718       dst.row(dec().permutationQ().indices().coeff(i)) = c.row(i);
719     for(Index i = nonzero_pivots; i < dec().matrixLU().cols(); ++i)
720       dst.row(dec().permutationQ().indices().coeff(i)).setZero();
721   }
722 };
723 
724 } // end namespace internal
725 
726 /******* MatrixBase methods *****************************************************************/
727 
728 /** \lu_module
729   *
730   * \return the full-pivoting LU decomposition of \c *this.
731   *
732   * \sa class FullPivLU
733   */
734 template<typename Derived>
735 inline const FullPivLU<typename MatrixBase<Derived>::PlainObject>
736 MatrixBase<Derived>::fullPivLu() const
737 {
738   return FullPivLU<PlainObject>(eval());
739 }
740 
741 } // end namespace Eigen
742 
743 #endif // EIGEN_LU_H
744