Home | History | Annotate | Download | only in LU
      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 (at) 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       */
    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       */
    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       */
    126     RealScalar maxPivot() const { return m_maxpivot; }
    127 
    128     /** \returns the permutation matrix P
    129       *
    130       * \sa permutationQ()
    131       */
    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       */
    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       */
    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>
    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>
    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       */
    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       */
    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       */
    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       */
    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       */
    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       */
    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       */
    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       */
    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       */
    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 
    373     inline Index rows() const { return m_lu.rows(); }
    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>
    388 FullPivLU<MatrixType>::FullPivLU()
    389   : m_isInitialized(false), m_usePrescribedThreshold(false)
    390 {
    391 }
    392 
    393 template<typename MatrixType>
    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>
    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>
    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>
    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>
    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