Home | History | Annotate | Download | only in SVD
      1 // This file is part of Eigen, a lightweight C++ template library
      2 // for linear algebra.
      3 //
      4 // Copyright (C) 2009-2010 Benoit Jacob <jacob.benoit.1 (at) gmail.com>
      5 // Copyright (C) 2013-2014 Gael Guennebaud <gael.guennebaud (at) 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_JACOBISVD_H
     12 #define EIGEN_JACOBISVD_H
     13 
     14 namespace Eigen {
     15 
     16 namespace internal {
     17 // forward declaration (needed by ICC)
     18 // the empty body is required by MSVC
     19 template<typename MatrixType, int QRPreconditioner,
     20          bool IsComplex = NumTraits<typename MatrixType::Scalar>::IsComplex>
     21 struct svd_precondition_2x2_block_to_be_real {};
     22 
     23 /*** QR preconditioners (R-SVD)
     24  ***
     25  *** Their role is to reduce the problem of computing the SVD to the case of a square matrix.
     26  *** This approach, known as R-SVD, is an optimization for rectangular-enough matrices, and is a requirement for
     27  *** JacobiSVD which by itself is only able to work on square matrices.
     28  ***/
     29 
     30 enum { PreconditionIfMoreColsThanRows, PreconditionIfMoreRowsThanCols };
     31 
     32 template<typename MatrixType, int QRPreconditioner, int Case>
     33 struct qr_preconditioner_should_do_anything
     34 {
     35   enum { a = MatrixType::RowsAtCompileTime != Dynamic &&
     36              MatrixType::ColsAtCompileTime != Dynamic &&
     37              MatrixType::ColsAtCompileTime <= MatrixType::RowsAtCompileTime,
     38          b = MatrixType::RowsAtCompileTime != Dynamic &&
     39              MatrixType::ColsAtCompileTime != Dynamic &&
     40              MatrixType::RowsAtCompileTime <= MatrixType::ColsAtCompileTime,
     41          ret = !( (QRPreconditioner == NoQRPreconditioner) ||
     42                   (Case == PreconditionIfMoreColsThanRows && bool(a)) ||
     43                   (Case == PreconditionIfMoreRowsThanCols && bool(b)) )
     44   };
     45 };
     46 
     47 template<typename MatrixType, int QRPreconditioner, int Case,
     48          bool DoAnything = qr_preconditioner_should_do_anything<MatrixType, QRPreconditioner, Case>::ret
     49 > struct qr_preconditioner_impl {};
     50 
     51 template<typename MatrixType, int QRPreconditioner, int Case>
     52 class qr_preconditioner_impl<MatrixType, QRPreconditioner, Case, false>
     53 {
     54 public:
     55   void allocate(const JacobiSVD<MatrixType, QRPreconditioner>&) {}
     56   bool run(JacobiSVD<MatrixType, QRPreconditioner>&, const MatrixType&)
     57   {
     58     return false;
     59   }
     60 };
     61 
     62 /*** preconditioner using FullPivHouseholderQR ***/
     63 
     64 template<typename MatrixType>
     65 class qr_preconditioner_impl<MatrixType, FullPivHouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true>
     66 {
     67 public:
     68   typedef typename MatrixType::Scalar Scalar;
     69   enum
     70   {
     71     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
     72     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime
     73   };
     74   typedef Matrix<Scalar, 1, RowsAtCompileTime, RowMajor, 1, MaxRowsAtCompileTime> WorkspaceType;
     75 
     76   void allocate(const JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd)
     77   {
     78     if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols())
     79     {
     80       m_qr.~QRType();
     81       ::new (&m_qr) QRType(svd.rows(), svd.cols());
     82     }
     83     if (svd.m_computeFullU) m_workspace.resize(svd.rows());
     84   }
     85 
     86   bool run(JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
     87   {
     88     if(matrix.rows() > matrix.cols())
     89     {
     90       m_qr.compute(matrix);
     91       svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>();
     92       if(svd.m_computeFullU) m_qr.matrixQ().evalTo(svd.m_matrixU, m_workspace);
     93       if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation();
     94       return true;
     95     }
     96     return false;
     97   }
     98 private:
     99   typedef FullPivHouseholderQR<MatrixType> QRType;
    100   QRType m_qr;
    101   WorkspaceType m_workspace;
    102 };
    103 
    104 template<typename MatrixType>
    105 class qr_preconditioner_impl<MatrixType, FullPivHouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true>
    106 {
    107 public:
    108   typedef typename MatrixType::Scalar Scalar;
    109   enum
    110   {
    111     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
    112     ColsAtCompileTime = MatrixType::ColsAtCompileTime,
    113     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
    114     MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
    115     TrOptions = RowsAtCompileTime==1 ? (MatrixType::Options & ~(RowMajor))
    116               : ColsAtCompileTime==1 ? (MatrixType::Options |   RowMajor)
    117               : MatrixType::Options
    118   };
    119   typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, TrOptions, MaxColsAtCompileTime, MaxRowsAtCompileTime>
    120           TransposeTypeWithSameStorageOrder;
    121 
    122   void allocate(const JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd)
    123   {
    124     if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols())
    125     {
    126       m_qr.~QRType();
    127       ::new (&m_qr) QRType(svd.cols(), svd.rows());
    128     }
    129     m_adjoint.resize(svd.cols(), svd.rows());
    130     if (svd.m_computeFullV) m_workspace.resize(svd.cols());
    131   }
    132 
    133   bool run(JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
    134   {
    135     if(matrix.cols() > matrix.rows())
    136     {
    137       m_adjoint = matrix.adjoint();
    138       m_qr.compute(m_adjoint);
    139       svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint();
    140       if(svd.m_computeFullV) m_qr.matrixQ().evalTo(svd.m_matrixV, m_workspace);
    141       if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation();
    142       return true;
    143     }
    144     else return false;
    145   }
    146 private:
    147   typedef FullPivHouseholderQR<TransposeTypeWithSameStorageOrder> QRType;
    148   QRType m_qr;
    149   TransposeTypeWithSameStorageOrder m_adjoint;
    150   typename internal::plain_row_type<MatrixType>::type m_workspace;
    151 };
    152 
    153 /*** preconditioner using ColPivHouseholderQR ***/
    154 
    155 template<typename MatrixType>
    156 class qr_preconditioner_impl<MatrixType, ColPivHouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true>
    157 {
    158 public:
    159   void allocate(const JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd)
    160   {
    161     if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols())
    162     {
    163       m_qr.~QRType();
    164       ::new (&m_qr) QRType(svd.rows(), svd.cols());
    165     }
    166     if (svd.m_computeFullU) m_workspace.resize(svd.rows());
    167     else if (svd.m_computeThinU) m_workspace.resize(svd.cols());
    168   }
    169 
    170   bool run(JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
    171   {
    172     if(matrix.rows() > matrix.cols())
    173     {
    174       m_qr.compute(matrix);
    175       svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>();
    176       if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace);
    177       else if(svd.m_computeThinU)
    178       {
    179         svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols());
    180         m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace);
    181       }
    182       if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation();
    183       return true;
    184     }
    185     return false;
    186   }
    187 
    188 private:
    189   typedef ColPivHouseholderQR<MatrixType> QRType;
    190   QRType m_qr;
    191   typename internal::plain_col_type<MatrixType>::type m_workspace;
    192 };
    193 
    194 template<typename MatrixType>
    195 class qr_preconditioner_impl<MatrixType, ColPivHouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true>
    196 {
    197 public:
    198   typedef typename MatrixType::Scalar Scalar;
    199   enum
    200   {
    201     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
    202     ColsAtCompileTime = MatrixType::ColsAtCompileTime,
    203     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
    204     MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
    205     TrOptions = RowsAtCompileTime==1 ? (MatrixType::Options & ~(RowMajor))
    206               : ColsAtCompileTime==1 ? (MatrixType::Options |   RowMajor)
    207               : MatrixType::Options
    208   };
    209 
    210   typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, TrOptions, MaxColsAtCompileTime, MaxRowsAtCompileTime>
    211           TransposeTypeWithSameStorageOrder;
    212 
    213   void allocate(const JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd)
    214   {
    215     if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols())
    216     {
    217       m_qr.~QRType();
    218       ::new (&m_qr) QRType(svd.cols(), svd.rows());
    219     }
    220     if (svd.m_computeFullV) m_workspace.resize(svd.cols());
    221     else if (svd.m_computeThinV) m_workspace.resize(svd.rows());
    222     m_adjoint.resize(svd.cols(), svd.rows());
    223   }
    224 
    225   bool run(JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
    226   {
    227     if(matrix.cols() > matrix.rows())
    228     {
    229       m_adjoint = matrix.adjoint();
    230       m_qr.compute(m_adjoint);
    231 
    232       svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint();
    233       if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace);
    234       else if(svd.m_computeThinV)
    235       {
    236         svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows());
    237         m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace);
    238       }
    239       if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation();
    240       return true;
    241     }
    242     else return false;
    243   }
    244 
    245 private:
    246   typedef ColPivHouseholderQR<TransposeTypeWithSameStorageOrder> QRType;
    247   QRType m_qr;
    248   TransposeTypeWithSameStorageOrder m_adjoint;
    249   typename internal::plain_row_type<MatrixType>::type m_workspace;
    250 };
    251 
    252 /*** preconditioner using HouseholderQR ***/
    253 
    254 template<typename MatrixType>
    255 class qr_preconditioner_impl<MatrixType, HouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true>
    256 {
    257 public:
    258   void allocate(const JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd)
    259   {
    260     if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols())
    261     {
    262       m_qr.~QRType();
    263       ::new (&m_qr) QRType(svd.rows(), svd.cols());
    264     }
    265     if (svd.m_computeFullU) m_workspace.resize(svd.rows());
    266     else if (svd.m_computeThinU) m_workspace.resize(svd.cols());
    267   }
    268 
    269   bool run(JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd, const MatrixType& matrix)
    270   {
    271     if(matrix.rows() > matrix.cols())
    272     {
    273       m_qr.compute(matrix);
    274       svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>();
    275       if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace);
    276       else if(svd.m_computeThinU)
    277       {
    278         svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols());
    279         m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace);
    280       }
    281       if(svd.computeV()) svd.m_matrixV.setIdentity(matrix.cols(), matrix.cols());
    282       return true;
    283     }
    284     return false;
    285   }
    286 private:
    287   typedef HouseholderQR<MatrixType> QRType;
    288   QRType m_qr;
    289   typename internal::plain_col_type<MatrixType>::type m_workspace;
    290 };
    291 
    292 template<typename MatrixType>
    293 class qr_preconditioner_impl<MatrixType, HouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true>
    294 {
    295 public:
    296   typedef typename MatrixType::Scalar Scalar;
    297   enum
    298   {
    299     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
    300     ColsAtCompileTime = MatrixType::ColsAtCompileTime,
    301     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
    302     MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
    303     Options = MatrixType::Options
    304   };
    305 
    306   typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, Options, MaxColsAtCompileTime, MaxRowsAtCompileTime>
    307           TransposeTypeWithSameStorageOrder;
    308 
    309   void allocate(const JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd)
    310   {
    311     if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols())
    312     {
    313       m_qr.~QRType();
    314       ::new (&m_qr) QRType(svd.cols(), svd.rows());
    315     }
    316     if (svd.m_computeFullV) m_workspace.resize(svd.cols());
    317     else if (svd.m_computeThinV) m_workspace.resize(svd.rows());
    318     m_adjoint.resize(svd.cols(), svd.rows());
    319   }
    320 
    321   bool run(JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd, const MatrixType& matrix)
    322   {
    323     if(matrix.cols() > matrix.rows())
    324     {
    325       m_adjoint = matrix.adjoint();
    326       m_qr.compute(m_adjoint);
    327 
    328       svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint();
    329       if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace);
    330       else if(svd.m_computeThinV)
    331       {
    332         svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows());
    333         m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace);
    334       }
    335       if(svd.computeU()) svd.m_matrixU.setIdentity(matrix.rows(), matrix.rows());
    336       return true;
    337     }
    338     else return false;
    339   }
    340 
    341 private:
    342   typedef HouseholderQR<TransposeTypeWithSameStorageOrder> QRType;
    343   QRType m_qr;
    344   TransposeTypeWithSameStorageOrder m_adjoint;
    345   typename internal::plain_row_type<MatrixType>::type m_workspace;
    346 };
    347 
    348 /*** 2x2 SVD implementation
    349  ***
    350  *** JacobiSVD consists in performing a series of 2x2 SVD subproblems
    351  ***/
    352 
    353 template<typename MatrixType, int QRPreconditioner>
    354 struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, false>
    355 {
    356   typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
    357   typedef typename MatrixType::RealScalar RealScalar;
    358   static bool run(typename SVD::WorkMatrixType&, SVD&, Index, Index, RealScalar&) { return true; }
    359 };
    360 
    361 template<typename MatrixType, int QRPreconditioner>
    362 struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, true>
    363 {
    364   typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
    365   typedef typename MatrixType::Scalar Scalar;
    366   typedef typename MatrixType::RealScalar RealScalar;
    367   static bool run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q, RealScalar& maxDiagEntry)
    368   {
    369     using std::sqrt;
    370     using std::abs;
    371     Scalar z;
    372     JacobiRotation<Scalar> rot;
    373     RealScalar n = sqrt(numext::abs2(work_matrix.coeff(p,p)) + numext::abs2(work_matrix.coeff(q,p)));
    374 
    375     const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
    376     const RealScalar precision = NumTraits<Scalar>::epsilon();
    377 
    378     if(n==0)
    379     {
    380       // make sure first column is zero
    381       work_matrix.coeffRef(p,p) = work_matrix.coeffRef(q,p) = Scalar(0);
    382 
    383       if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
    384       {
    385         // work_matrix.coeff(p,q) can be zero if work_matrix.coeff(q,p) is not zero but small enough to underflow when computing n
    386         z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
    387         work_matrix.row(p) *= z;
    388         if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z);
    389       }
    390       if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
    391       {
    392         z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
    393         work_matrix.row(q) *= z;
    394         if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z);
    395       }
    396       // otherwise the second row is already zero, so we have nothing to do.
    397     }
    398     else
    399     {
    400       rot.c() = conj(work_matrix.coeff(p,p)) / n;
    401       rot.s() = work_matrix.coeff(q,p) / n;
    402       work_matrix.applyOnTheLeft(p,q,rot);
    403       if(svd.computeU()) svd.m_matrixU.applyOnTheRight(p,q,rot.adjoint());
    404       if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
    405       {
    406         z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
    407         work_matrix.col(q) *= z;
    408         if(svd.computeV()) svd.m_matrixV.col(q) *= z;
    409       }
    410       if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
    411       {
    412         z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
    413         work_matrix.row(q) *= z;
    414         if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z);
    415       }
    416     }
    417 
    418     // update largest diagonal entry
    419     maxDiagEntry = numext::maxi<RealScalar>(maxDiagEntry,numext::maxi<RealScalar>(abs(work_matrix.coeff(p,p)), abs(work_matrix.coeff(q,q))));
    420     // and check whether the 2x2 block is already diagonal
    421     RealScalar threshold = numext::maxi<RealScalar>(considerAsZero, precision * maxDiagEntry);
    422     return abs(work_matrix.coeff(p,q))>threshold || abs(work_matrix.coeff(q,p)) > threshold;
    423   }
    424 };
    425 
    426 template<typename _MatrixType, int QRPreconditioner>
    427 struct traits<JacobiSVD<_MatrixType,QRPreconditioner> >
    428 {
    429   typedef _MatrixType MatrixType;
    430 };
    431 
    432 } // end namespace internal
    433 
    434 /** \ingroup SVD_Module
    435   *
    436   *
    437   * \class JacobiSVD
    438   *
    439   * \brief Two-sided Jacobi SVD decomposition of a rectangular matrix
    440   *
    441   * \tparam _MatrixType the type of the matrix of which we are computing the SVD decomposition
    442   * \tparam QRPreconditioner this optional parameter allows to specify the type of QR decomposition that will be used internally
    443   *                        for the R-SVD step for non-square matrices. See discussion of possible values below.
    444   *
    445   * SVD decomposition consists in decomposing any n-by-p matrix \a A as a product
    446   *   \f[ A = U S V^* \f]
    447   * where \a U is a n-by-n unitary, \a V is a p-by-p unitary, and \a S is a n-by-p real positive matrix which is zero outside of its main diagonal;
    448   * the diagonal entries of S are known as the \em singular \em values of \a A and the columns of \a U and \a V are known as the left
    449   * and right \em singular \em vectors of \a A respectively.
    450   *
    451   * Singular values are always sorted in decreasing order.
    452   *
    453   * This JacobiSVD decomposition computes only the singular values by default. If you want \a U or \a V, you need to ask for them explicitly.
    454   *
    455   * You can ask for only \em thin \a U or \a V to be computed, meaning the following. In case of a rectangular n-by-p matrix, letting \a m be the
    456   * smaller value among \a n and \a p, there are only \a m singular vectors; the remaining columns of \a U and \a V do not correspond to actual
    457   * singular vectors. Asking for \em thin \a U or \a V means asking for only their \a m first columns to be formed. So \a U is then a n-by-m matrix,
    458   * and \a V is then a p-by-m matrix. Notice that thin \a U and \a V are all you need for (least squares) solving.
    459   *
    460   * Here's an example demonstrating basic usage:
    461   * \include JacobiSVD_basic.cpp
    462   * Output: \verbinclude JacobiSVD_basic.out
    463   *
    464   * This JacobiSVD class is a two-sided Jacobi R-SVD decomposition, ensuring optimal reliability and accuracy. The downside is that it's slower than
    465   * bidiagonalizing SVD algorithms for large square matrices; however its complexity is still \f$ O(n^2p) \f$ where \a n is the smaller dimension and
    466   * \a p is the greater dimension, meaning that it is still of the same order of complexity as the faster bidiagonalizing R-SVD algorithms.
    467   * In particular, like any R-SVD, it takes advantage of non-squareness in that its complexity is only linear in the greater dimension.
    468   *
    469   * If the input matrix has inf or nan coefficients, the result of the computation is undefined, but the computation is guaranteed to
    470   * terminate in finite (and reasonable) time.
    471   *
    472   * The possible values for QRPreconditioner are:
    473   * \li ColPivHouseholderQRPreconditioner is the default. In practice it's very safe. It uses column-pivoting QR.
    474   * \li FullPivHouseholderQRPreconditioner, is the safest and slowest. It uses full-pivoting QR.
    475   *     Contrary to other QRs, it doesn't allow computing thin unitaries.
    476   * \li HouseholderQRPreconditioner is the fastest, and less safe and accurate than the pivoting variants. It uses non-pivoting QR.
    477   *     This is very similar in safety and accuracy to the bidiagonalization process used by bidiagonalizing SVD algorithms (since bidiagonalization
    478   *     is inherently non-pivoting). However the resulting SVD is still more reliable than bidiagonalizing SVDs because the Jacobi-based iterarive
    479   *     process is more reliable than the optimized bidiagonal SVD iterations.
    480   * \li NoQRPreconditioner allows not to use a QR preconditioner at all. This is useful if you know that you will only be computing
    481   *     JacobiSVD decompositions of square matrices. Non-square matrices require a QR preconditioner. Using this option will result in
    482   *     faster compilation and smaller executable code. It won't significantly speed up computation, since JacobiSVD is always checking
    483   *     if QR preconditioning is needed before applying it anyway.
    484   *
    485   * \sa MatrixBase::jacobiSvd()
    486   */
    487 template<typename _MatrixType, int QRPreconditioner> class JacobiSVD
    488  : public SVDBase<JacobiSVD<_MatrixType,QRPreconditioner> >
    489 {
    490     typedef SVDBase<JacobiSVD> Base;
    491   public:
    492 
    493     typedef _MatrixType MatrixType;
    494     typedef typename MatrixType::Scalar Scalar;
    495     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
    496     enum {
    497       RowsAtCompileTime = MatrixType::RowsAtCompileTime,
    498       ColsAtCompileTime = MatrixType::ColsAtCompileTime,
    499       DiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime,ColsAtCompileTime),
    500       MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
    501       MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
    502       MaxDiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(MaxRowsAtCompileTime,MaxColsAtCompileTime),
    503       MatrixOptions = MatrixType::Options
    504     };
    505 
    506     typedef typename Base::MatrixUType MatrixUType;
    507     typedef typename Base::MatrixVType MatrixVType;
    508     typedef typename Base::SingularValuesType SingularValuesType;
    509 
    510     typedef typename internal::plain_row_type<MatrixType>::type RowType;
    511     typedef typename internal::plain_col_type<MatrixType>::type ColType;
    512     typedef Matrix<Scalar, DiagSizeAtCompileTime, DiagSizeAtCompileTime,
    513                    MatrixOptions, MaxDiagSizeAtCompileTime, MaxDiagSizeAtCompileTime>
    514             WorkMatrixType;
    515 
    516     /** \brief Default Constructor.
    517       *
    518       * The default constructor is useful in cases in which the user intends to
    519       * perform decompositions via JacobiSVD::compute(const MatrixType&).
    520       */
    521     JacobiSVD()
    522     {}
    523 
    524 
    525     /** \brief Default Constructor with memory preallocation
    526       *
    527       * Like the default constructor but with preallocation of the internal data
    528       * according to the specified problem size.
    529       * \sa JacobiSVD()
    530       */
    531     JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0)
    532     {
    533       allocate(rows, cols, computationOptions);
    534     }
    535 
    536     /** \brief Constructor performing the decomposition of given matrix.
    537      *
    538      * \param matrix the matrix to decompose
    539      * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
    540      *                           By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU,
    541      *                           #ComputeFullV, #ComputeThinV.
    542      *
    543      * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
    544      * available with the (non-default) FullPivHouseholderQR preconditioner.
    545      */
    546     explicit JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0)
    547     {
    548       compute(matrix, computationOptions);
    549     }
    550 
    551     /** \brief Method performing the decomposition of given matrix using custom options.
    552      *
    553      * \param matrix the matrix to decompose
    554      * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
    555      *                           By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU,
    556      *                           #ComputeFullV, #ComputeThinV.
    557      *
    558      * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
    559      * available with the (non-default) FullPivHouseholderQR preconditioner.
    560      */
    561     JacobiSVD& compute(const MatrixType& matrix, unsigned int computationOptions);
    562 
    563     /** \brief Method performing the decomposition of given matrix using current options.
    564      *
    565      * \param matrix the matrix to decompose
    566      *
    567      * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int).
    568      */
    569     JacobiSVD& compute(const MatrixType& matrix)
    570     {
    571       return compute(matrix, m_computationOptions);
    572     }
    573 
    574     using Base::computeU;
    575     using Base::computeV;
    576     using Base::rows;
    577     using Base::cols;
    578     using Base::rank;
    579 
    580   private:
    581     void allocate(Index rows, Index cols, unsigned int computationOptions);
    582 
    583   protected:
    584     using Base::m_matrixU;
    585     using Base::m_matrixV;
    586     using Base::m_singularValues;
    587     using Base::m_isInitialized;
    588     using Base::m_isAllocated;
    589     using Base::m_usePrescribedThreshold;
    590     using Base::m_computeFullU;
    591     using Base::m_computeThinU;
    592     using Base::m_computeFullV;
    593     using Base::m_computeThinV;
    594     using Base::m_computationOptions;
    595     using Base::m_nonzeroSingularValues;
    596     using Base::m_rows;
    597     using Base::m_cols;
    598     using Base::m_diagSize;
    599     using Base::m_prescribedThreshold;
    600     WorkMatrixType m_workMatrix;
    601 
    602     template<typename __MatrixType, int _QRPreconditioner, bool _IsComplex>
    603     friend struct internal::svd_precondition_2x2_block_to_be_real;
    604     template<typename __MatrixType, int _QRPreconditioner, int _Case, bool _DoAnything>
    605     friend struct internal::qr_preconditioner_impl;
    606 
    607     internal::qr_preconditioner_impl<MatrixType, QRPreconditioner, internal::PreconditionIfMoreColsThanRows> m_qr_precond_morecols;
    608     internal::qr_preconditioner_impl<MatrixType, QRPreconditioner, internal::PreconditionIfMoreRowsThanCols> m_qr_precond_morerows;
    609     MatrixType m_scaledMatrix;
    610 };
    611 
    612 template<typename MatrixType, int QRPreconditioner>
    613 void JacobiSVD<MatrixType, QRPreconditioner>::allocate(Index rows, Index cols, unsigned int computationOptions)
    614 {
    615   eigen_assert(rows >= 0 && cols >= 0);
    616 
    617   if (m_isAllocated &&
    618       rows == m_rows &&
    619       cols == m_cols &&
    620       computationOptions == m_computationOptions)
    621   {
    622     return;
    623   }
    624 
    625   m_rows = rows;
    626   m_cols = cols;
    627   m_isInitialized = false;
    628   m_isAllocated = true;
    629   m_computationOptions = computationOptions;
    630   m_computeFullU = (computationOptions & ComputeFullU) != 0;
    631   m_computeThinU = (computationOptions & ComputeThinU) != 0;
    632   m_computeFullV = (computationOptions & ComputeFullV) != 0;
    633   m_computeThinV = (computationOptions & ComputeThinV) != 0;
    634   eigen_assert(!(m_computeFullU && m_computeThinU) && "JacobiSVD: you can't ask for both full and thin U");
    635   eigen_assert(!(m_computeFullV && m_computeThinV) && "JacobiSVD: you can't ask for both full and thin V");
    636   eigen_assert(EIGEN_IMPLIES(m_computeThinU || m_computeThinV, MatrixType::ColsAtCompileTime==Dynamic) &&
    637               "JacobiSVD: thin U and V are only available when your matrix has a dynamic number of columns.");
    638   if (QRPreconditioner == FullPivHouseholderQRPreconditioner)
    639   {
    640       eigen_assert(!(m_computeThinU || m_computeThinV) &&
    641               "JacobiSVD: can't compute thin U or thin V with the FullPivHouseholderQR preconditioner. "
    642               "Use the ColPivHouseholderQR preconditioner instead.");
    643   }
    644   m_diagSize = (std::min)(m_rows, m_cols);
    645   m_singularValues.resize(m_diagSize);
    646   if(RowsAtCompileTime==Dynamic)
    647     m_matrixU.resize(m_rows, m_computeFullU ? m_rows
    648                             : m_computeThinU ? m_diagSize
    649                             : 0);
    650   if(ColsAtCompileTime==Dynamic)
    651     m_matrixV.resize(m_cols, m_computeFullV ? m_cols
    652                             : m_computeThinV ? m_diagSize
    653                             : 0);
    654   m_workMatrix.resize(m_diagSize, m_diagSize);
    655 
    656   if(m_cols>m_rows)   m_qr_precond_morecols.allocate(*this);
    657   if(m_rows>m_cols)   m_qr_precond_morerows.allocate(*this);
    658   if(m_rows!=m_cols)  m_scaledMatrix.resize(rows,cols);
    659 }
    660 
    661 template<typename MatrixType, int QRPreconditioner>
    662 JacobiSVD<MatrixType, QRPreconditioner>&
    663 JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsigned int computationOptions)
    664 {
    665   using std::abs;
    666   allocate(matrix.rows(), matrix.cols(), computationOptions);
    667 
    668   // currently we stop when we reach precision 2*epsilon as the last bit of precision can require an unreasonable number of iterations,
    669   // only worsening the precision of U and V as we accumulate more rotations
    670   const RealScalar precision = RealScalar(2) * NumTraits<Scalar>::epsilon();
    671 
    672   // limit for denormal numbers to be considered zero in order to avoid infinite loops (see bug 286)
    673   const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
    674 
    675   // Scaling factor to reduce over/under-flows
    676   RealScalar scale = matrix.cwiseAbs().maxCoeff();
    677   if(scale==RealScalar(0)) scale = RealScalar(1);
    678 
    679   /*** step 1. The R-SVD step: we use a QR decomposition to reduce to the case of a square matrix */
    680 
    681   if(m_rows!=m_cols)
    682   {
    683     m_scaledMatrix = matrix / scale;
    684     m_qr_precond_morecols.run(*this, m_scaledMatrix);
    685     m_qr_precond_morerows.run(*this, m_scaledMatrix);
    686   }
    687   else
    688   {
    689     m_workMatrix = matrix.block(0,0,m_diagSize,m_diagSize) / scale;
    690     if(m_computeFullU) m_matrixU.setIdentity(m_rows,m_rows);
    691     if(m_computeThinU) m_matrixU.setIdentity(m_rows,m_diagSize);
    692     if(m_computeFullV) m_matrixV.setIdentity(m_cols,m_cols);
    693     if(m_computeThinV) m_matrixV.setIdentity(m_cols, m_diagSize);
    694   }
    695 
    696   /*** step 2. The main Jacobi SVD iteration. ***/
    697   RealScalar maxDiagEntry = m_workMatrix.cwiseAbs().diagonal().maxCoeff();
    698 
    699   bool finished = false;
    700   while(!finished)
    701   {
    702     finished = true;
    703 
    704     // do a sweep: for all index pairs (p,q), perform SVD of the corresponding 2x2 sub-matrix
    705 
    706     for(Index p = 1; p < m_diagSize; ++p)
    707     {
    708       for(Index q = 0; q < p; ++q)
    709       {
    710         // if this 2x2 sub-matrix is not diagonal already...
    711         // notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't
    712         // keep us iterating forever. Similarly, small denormal numbers are considered zero.
    713         RealScalar threshold = numext::maxi<RealScalar>(considerAsZero, precision * maxDiagEntry);
    714         if(abs(m_workMatrix.coeff(p,q))>threshold || abs(m_workMatrix.coeff(q,p)) > threshold)
    715         {
    716           finished = false;
    717           // perform SVD decomposition of 2x2 sub-matrix corresponding to indices p,q to make it diagonal
    718           // the complex to real operation returns true if the updated 2x2 block is not already diagonal
    719           if(internal::svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner>::run(m_workMatrix, *this, p, q, maxDiagEntry))
    720           {
    721             JacobiRotation<RealScalar> j_left, j_right;
    722             internal::real_2x2_jacobi_svd(m_workMatrix, p, q, &j_left, &j_right);
    723 
    724             // accumulate resulting Jacobi rotations
    725             m_workMatrix.applyOnTheLeft(p,q,j_left);
    726             if(computeU()) m_matrixU.applyOnTheRight(p,q,j_left.transpose());
    727 
    728             m_workMatrix.applyOnTheRight(p,q,j_right);
    729             if(computeV()) m_matrixV.applyOnTheRight(p,q,j_right);
    730 
    731             // keep track of the largest diagonal coefficient
    732             maxDiagEntry = numext::maxi<RealScalar>(maxDiagEntry,numext::maxi<RealScalar>(abs(m_workMatrix.coeff(p,p)), abs(m_workMatrix.coeff(q,q))));
    733           }
    734         }
    735       }
    736     }
    737   }
    738 
    739   /*** step 3. The work matrix is now diagonal, so ensure it's positive so its diagonal entries are the singular values ***/
    740 
    741   for(Index i = 0; i < m_diagSize; ++i)
    742   {
    743     // For a complex matrix, some diagonal coefficients might note have been
    744     // treated by svd_precondition_2x2_block_to_be_real, and the imaginary part
    745     // of some diagonal entry might not be null.
    746     if(NumTraits<Scalar>::IsComplex && abs(numext::imag(m_workMatrix.coeff(i,i)))>considerAsZero)
    747     {
    748       RealScalar a = abs(m_workMatrix.coeff(i,i));
    749       m_singularValues.coeffRef(i) = abs(a);
    750       if(computeU()) m_matrixU.col(i) *= m_workMatrix.coeff(i,i)/a;
    751     }
    752     else
    753     {
    754       // m_workMatrix.coeff(i,i) is already real, no difficulty:
    755       RealScalar a = numext::real(m_workMatrix.coeff(i,i));
    756       m_singularValues.coeffRef(i) = abs(a);
    757       if(computeU() && (a<RealScalar(0))) m_matrixU.col(i) = -m_matrixU.col(i);
    758     }
    759   }
    760 
    761   m_singularValues *= scale;
    762 
    763   /*** step 4. Sort singular values in descending order and compute the number of nonzero singular values ***/
    764 
    765   m_nonzeroSingularValues = m_diagSize;
    766   for(Index i = 0; i < m_diagSize; i++)
    767   {
    768     Index pos;
    769     RealScalar maxRemainingSingularValue = m_singularValues.tail(m_diagSize-i).maxCoeff(&pos);
    770     if(maxRemainingSingularValue == RealScalar(0))
    771     {
    772       m_nonzeroSingularValues = i;
    773       break;
    774     }
    775     if(pos)
    776     {
    777       pos += i;
    778       std::swap(m_singularValues.coeffRef(i), m_singularValues.coeffRef(pos));
    779       if(computeU()) m_matrixU.col(pos).swap(m_matrixU.col(i));
    780       if(computeV()) m_matrixV.col(pos).swap(m_matrixV.col(i));
    781     }
    782   }
    783 
    784   m_isInitialized = true;
    785   return *this;
    786 }
    787 
    788 /** \svd_module
    789   *
    790   * \return the singular value decomposition of \c *this computed by two-sided
    791   * Jacobi transformations.
    792   *
    793   * \sa class JacobiSVD
    794   */
    795 template<typename Derived>
    796 JacobiSVD<typename MatrixBase<Derived>::PlainObject>
    797 MatrixBase<Derived>::jacobiSvd(unsigned int computationOptions) const
    798 {
    799   return JacobiSVD<PlainObject>(*this, computationOptions);
    800 }
    801 
    802 } // end namespace Eigen
    803 
    804 #endif // EIGEN_JACOBISVD_H
    805