1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
5// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
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_FULLPIVOTINGHOUSEHOLDERQR_H
12#define EIGEN_FULLPIVOTINGHOUSEHOLDERQR_H
13
14namespace Eigen {
15
16namespace internal {
17
18template<typename _MatrixType> struct traits<FullPivHouseholderQR<_MatrixType> >
19 : traits<_MatrixType>
20{
21  enum { Flags = 0 };
22};
23
24template<typename MatrixType> struct FullPivHouseholderQRMatrixQReturnType;
25
26template<typename MatrixType>
27struct traits<FullPivHouseholderQRMatrixQReturnType<MatrixType> >
28{
29  typedef typename MatrixType::PlainObject ReturnType;
30};
31
32} // end namespace internal
33
34/** \ingroup QR_Module
35  *
36  * \class FullPivHouseholderQR
37  *
38  * \brief Householder rank-revealing QR decomposition of a matrix with full pivoting
39  *
40  * \tparam _MatrixType the type of the matrix of which we are computing the QR decomposition
41  *
42  * This class performs a rank-revealing QR decomposition of a matrix \b A into matrices \b P, \b P', \b Q and \b R
43  * such that
44  * \f[
45  *  \mathbf{P} \, \mathbf{A} \, \mathbf{P}' = \mathbf{Q} \, \mathbf{R}
46  * \f]
47  * by using Householder transformations. Here, \b P and \b P' are permutation matrices, \b Q a unitary matrix
48  * and \b R an upper triangular matrix.
49  *
50  * This decomposition performs a very prudent full pivoting in order to be rank-revealing and achieve optimal
51  * numerical stability. The trade-off is that it is slower than HouseholderQR and ColPivHouseholderQR.
52  *
53  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
54  *
55  * \sa MatrixBase::fullPivHouseholderQr()
56  */
57template<typename _MatrixType> class FullPivHouseholderQR
58{
59  public:
60
61    typedef _MatrixType MatrixType;
62    enum {
63      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
64      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
65      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
66      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
67    };
68    typedef typename MatrixType::Scalar Scalar;
69    typedef typename MatrixType::RealScalar RealScalar;
70    // FIXME should be int
71    typedef typename MatrixType::StorageIndex StorageIndex;
72    typedef internal::FullPivHouseholderQRMatrixQReturnType<MatrixType> MatrixQReturnType;
73    typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
74    typedef Matrix<StorageIndex, 1,
75                   EIGEN_SIZE_MIN_PREFER_DYNAMIC(ColsAtCompileTime,RowsAtCompileTime), RowMajor, 1,
76                   EIGEN_SIZE_MIN_PREFER_FIXED(MaxColsAtCompileTime,MaxRowsAtCompileTime)> IntDiagSizeVectorType;
77    typedef PermutationMatrix<ColsAtCompileTime, MaxColsAtCompileTime> PermutationType;
78    typedef typename internal::plain_row_type<MatrixType>::type RowVectorType;
79    typedef typename internal::plain_col_type<MatrixType>::type ColVectorType;
80    typedef typename MatrixType::PlainObject PlainObject;
81
82    /** \brief Default Constructor.
83      *
84      * The default constructor is useful in cases in which the user intends to
85      * perform decompositions via FullPivHouseholderQR::compute(const MatrixType&).
86      */
87    FullPivHouseholderQR()
88      : m_qr(),
89        m_hCoeffs(),
90        m_rows_transpositions(),
91        m_cols_transpositions(),
92        m_cols_permutation(),
93        m_temp(),
94        m_isInitialized(false),
95        m_usePrescribedThreshold(false) {}
96
97    /** \brief Default Constructor with memory preallocation
98      *
99      * Like the default constructor but with preallocation of the internal data
100      * according to the specified problem \a size.
101      * \sa FullPivHouseholderQR()
102      */
103    FullPivHouseholderQR(Index rows, Index cols)
104      : m_qr(rows, cols),
105        m_hCoeffs((std::min)(rows,cols)),
106        m_rows_transpositions((std::min)(rows,cols)),
107        m_cols_transpositions((std::min)(rows,cols)),
108        m_cols_permutation(cols),
109        m_temp(cols),
110        m_isInitialized(false),
111        m_usePrescribedThreshold(false) {}
112
113    /** \brief Constructs a QR factorization from a given matrix
114      *
115      * This constructor computes the QR factorization of the matrix \a matrix by calling
116      * the method compute(). It is a short cut for:
117      *
118      * \code
119      * FullPivHouseholderQR<MatrixType> qr(matrix.rows(), matrix.cols());
120      * qr.compute(matrix);
121      * \endcode
122      *
123      * \sa compute()
124      */
125    template<typename InputType>
126    explicit FullPivHouseholderQR(const EigenBase<InputType>& matrix)
127      : m_qr(matrix.rows(), matrix.cols()),
128        m_hCoeffs((std::min)(matrix.rows(), matrix.cols())),
129        m_rows_transpositions((std::min)(matrix.rows(), matrix.cols())),
130        m_cols_transpositions((std::min)(matrix.rows(), matrix.cols())),
131        m_cols_permutation(matrix.cols()),
132        m_temp(matrix.cols()),
133        m_isInitialized(false),
134        m_usePrescribedThreshold(false)
135    {
136      compute(matrix.derived());
137    }
138
139    /** \brief Constructs a QR factorization from a given matrix
140      *
141      * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when \c MatrixType is a Eigen::Ref.
142      *
143      * \sa FullPivHouseholderQR(const EigenBase&)
144      */
145    template<typename InputType>
146    explicit FullPivHouseholderQR(EigenBase<InputType>& matrix)
147      : m_qr(matrix.derived()),
148        m_hCoeffs((std::min)(matrix.rows(), matrix.cols())),
149        m_rows_transpositions((std::min)(matrix.rows(), matrix.cols())),
150        m_cols_transpositions((std::min)(matrix.rows(), matrix.cols())),
151        m_cols_permutation(matrix.cols()),
152        m_temp(matrix.cols()),
153        m_isInitialized(false),
154        m_usePrescribedThreshold(false)
155    {
156      computeInPlace();
157    }
158
159    /** This method finds a solution x to the equation Ax=b, where A is the matrix of which
160      * \c *this is the QR decomposition.
161      *
162      * \param b the right-hand-side of the equation to solve.
163      *
164      * \returns the exact or least-square solution if the rank is greater or equal to the number of columns of A,
165      * and an arbitrary solution otherwise.
166      *
167      * \note_about_checking_solutions
168      *
169      * \note_about_arbitrary_choice_of_solution
170      *
171      * Example: \include FullPivHouseholderQR_solve.cpp
172      * Output: \verbinclude FullPivHouseholderQR_solve.out
173      */
174    template<typename Rhs>
175    inline const Solve<FullPivHouseholderQR, Rhs>
176    solve(const MatrixBase<Rhs>& b) const
177    {
178      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
179      return Solve<FullPivHouseholderQR, Rhs>(*this, b.derived());
180    }
181
182    /** \returns Expression object representing the matrix Q
183      */
184    MatrixQReturnType matrixQ(void) const;
185
186    /** \returns a reference to the matrix where the Householder QR decomposition is stored
187      */
188    const MatrixType& matrixQR() const
189    {
190      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
191      return m_qr;
192    }
193
194    template<typename InputType>
195    FullPivHouseholderQR& compute(const EigenBase<InputType>& matrix);
196
197    /** \returns a const reference to the column permutation matrix */
198    const PermutationType& colsPermutation() const
199    {
200      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
201      return m_cols_permutation;
202    }
203
204    /** \returns a const reference to the vector of indices representing the rows transpositions */
205    const IntDiagSizeVectorType& rowsTranspositions() const
206    {
207      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
208      return m_rows_transpositions;
209    }
210
211    /** \returns the absolute value of the determinant of the matrix of which
212      * *this is the QR decomposition. It has only linear complexity
213      * (that is, O(n) where n is the dimension of the square matrix)
214      * as the QR decomposition has already been computed.
215      *
216      * \note This is only for square matrices.
217      *
218      * \warning a determinant can be very big or small, so for matrices
219      * of large enough dimension, there is a risk of overflow/underflow.
220      * One way to work around that is to use logAbsDeterminant() instead.
221      *
222      * \sa logAbsDeterminant(), MatrixBase::determinant()
223      */
224    typename MatrixType::RealScalar absDeterminant() const;
225
226    /** \returns the natural log of the absolute value of the determinant of the matrix of which
227      * *this is the QR decomposition. It has only linear complexity
228      * (that is, O(n) where n is the dimension of the square matrix)
229      * as the QR decomposition has already been computed.
230      *
231      * \note This is only for square matrices.
232      *
233      * \note This method is useful to work around the risk of overflow/underflow that's inherent
234      * to determinant computation.
235      *
236      * \sa absDeterminant(), MatrixBase::determinant()
237      */
238    typename MatrixType::RealScalar logAbsDeterminant() const;
239
240    /** \returns the rank of the matrix of which *this is the QR decomposition.
241      *
242      * \note This method has to determine which pivots should be considered nonzero.
243      *       For that, it uses the threshold value that you can control by calling
244      *       setThreshold(const RealScalar&).
245      */
246    inline Index rank() const
247    {
248      using std::abs;
249      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
250      RealScalar premultiplied_threshold = abs(m_maxpivot) * threshold();
251      Index result = 0;
252      for(Index i = 0; i < m_nonzero_pivots; ++i)
253        result += (abs(m_qr.coeff(i,i)) > premultiplied_threshold);
254      return result;
255    }
256
257    /** \returns the dimension of the kernel of the matrix of which *this is the QR decomposition.
258      *
259      * \note This method has to determine which pivots should be considered nonzero.
260      *       For that, it uses the threshold value that you can control by calling
261      *       setThreshold(const RealScalar&).
262      */
263    inline Index dimensionOfKernel() const
264    {
265      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
266      return cols() - rank();
267    }
268
269    /** \returns true if the matrix of which *this is the QR decomposition represents an injective
270      *          linear map, i.e. has trivial kernel; false otherwise.
271      *
272      * \note This method has to determine which pivots should be considered nonzero.
273      *       For that, it uses the threshold value that you can control by calling
274      *       setThreshold(const RealScalar&).
275      */
276    inline bool isInjective() const
277    {
278      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
279      return rank() == cols();
280    }
281
282    /** \returns true if the matrix of which *this is the QR decomposition represents a surjective
283      *          linear map; false otherwise.
284      *
285      * \note This method has to determine which pivots should be considered nonzero.
286      *       For that, it uses the threshold value that you can control by calling
287      *       setThreshold(const RealScalar&).
288      */
289    inline bool isSurjective() const
290    {
291      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
292      return rank() == rows();
293    }
294
295    /** \returns true if the matrix of which *this is the QR decomposition is invertible.
296      *
297      * \note This method has to determine which pivots should be considered nonzero.
298      *       For that, it uses the threshold value that you can control by calling
299      *       setThreshold(const RealScalar&).
300      */
301    inline bool isInvertible() const
302    {
303      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
304      return isInjective() && isSurjective();
305    }
306
307    /** \returns the inverse of the matrix of which *this is the QR decomposition.
308      *
309      * \note If this matrix is not invertible, the returned matrix has undefined coefficients.
310      *       Use isInvertible() to first determine whether this matrix is invertible.
311      */
312    inline const Inverse<FullPivHouseholderQR> inverse() const
313    {
314      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
315      return Inverse<FullPivHouseholderQR>(*this);
316    }
317
318    inline Index rows() const { return m_qr.rows(); }
319    inline Index cols() const { return m_qr.cols(); }
320
321    /** \returns a const reference to the vector of Householder coefficients used to represent the factor \c Q.
322      *
323      * For advanced uses only.
324      */
325    const HCoeffsType& hCoeffs() const { return m_hCoeffs; }
326
327    /** Allows to prescribe a threshold to be used by certain methods, such as rank(),
328      * who need to determine when pivots are to be considered nonzero. This is not used for the
329      * QR decomposition itself.
330      *
331      * When it needs to get the threshold value, Eigen calls threshold(). By default, this
332      * uses a formula to automatically determine a reasonable threshold.
333      * Once you have called the present method setThreshold(const RealScalar&),
334      * your value is used instead.
335      *
336      * \param threshold The new value to use as the threshold.
337      *
338      * A pivot will be considered nonzero if its absolute value is strictly greater than
339      *  \f$ \vert pivot \vert \leqslant threshold \times \vert maxpivot \vert \f$
340      * where maxpivot is the biggest pivot.
341      *
342      * If you want to come back to the default behavior, call setThreshold(Default_t)
343      */
344    FullPivHouseholderQR& setThreshold(const RealScalar& threshold)
345    {
346      m_usePrescribedThreshold = true;
347      m_prescribedThreshold = threshold;
348      return *this;
349    }
350
351    /** Allows to come back to the default behavior, letting Eigen use its default formula for
352      * determining the threshold.
353      *
354      * You should pass the special object Eigen::Default as parameter here.
355      * \code qr.setThreshold(Eigen::Default); \endcode
356      *
357      * See the documentation of setThreshold(const RealScalar&).
358      */
359    FullPivHouseholderQR& setThreshold(Default_t)
360    {
361      m_usePrescribedThreshold = false;
362      return *this;
363    }
364
365    /** Returns the threshold that will be used by certain methods such as rank().
366      *
367      * See the documentation of setThreshold(const RealScalar&).
368      */
369    RealScalar threshold() const
370    {
371      eigen_assert(m_isInitialized || m_usePrescribedThreshold);
372      return m_usePrescribedThreshold ? m_prescribedThreshold
373      // this formula comes from experimenting (see "LU precision tuning" thread on the list)
374      // and turns out to be identical to Higham's formula used already in LDLt.
375                                      : NumTraits<Scalar>::epsilon() * RealScalar(m_qr.diagonalSize());
376    }
377
378    /** \returns the number of nonzero pivots in the QR decomposition.
379      * Here nonzero is meant in the exact sense, not in a fuzzy sense.
380      * So that notion isn't really intrinsically interesting, but it is
381      * still useful when implementing algorithms.
382      *
383      * \sa rank()
384      */
385    inline Index nonzeroPivots() const
386    {
387      eigen_assert(m_isInitialized && "LU is not initialized.");
388      return m_nonzero_pivots;
389    }
390
391    /** \returns the absolute value of the biggest pivot, i.e. the biggest
392      *          diagonal coefficient of U.
393      */
394    RealScalar maxPivot() const { return m_maxpivot; }
395
396    #ifndef EIGEN_PARSED_BY_DOXYGEN
397    template<typename RhsType, typename DstType>
398    EIGEN_DEVICE_FUNC
399    void _solve_impl(const RhsType &rhs, DstType &dst) const;
400    #endif
401
402  protected:
403
404    static void check_template_parameters()
405    {
406      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
407    }
408
409    void computeInPlace();
410
411    MatrixType m_qr;
412    HCoeffsType m_hCoeffs;
413    IntDiagSizeVectorType m_rows_transpositions;
414    IntDiagSizeVectorType m_cols_transpositions;
415    PermutationType m_cols_permutation;
416    RowVectorType m_temp;
417    bool m_isInitialized, m_usePrescribedThreshold;
418    RealScalar m_prescribedThreshold, m_maxpivot;
419    Index m_nonzero_pivots;
420    RealScalar m_precision;
421    Index m_det_pq;
422};
423
424template<typename MatrixType>
425typename MatrixType::RealScalar FullPivHouseholderQR<MatrixType>::absDeterminant() const
426{
427  using std::abs;
428  eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
429  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
430  return abs(m_qr.diagonal().prod());
431}
432
433template<typename MatrixType>
434typename MatrixType::RealScalar FullPivHouseholderQR<MatrixType>::logAbsDeterminant() const
435{
436  eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
437  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
438  return m_qr.diagonal().cwiseAbs().array().log().sum();
439}
440
441/** Performs the QR factorization of the given matrix \a matrix. The result of
442  * the factorization is stored into \c *this, and a reference to \c *this
443  * is returned.
444  *
445  * \sa class FullPivHouseholderQR, FullPivHouseholderQR(const MatrixType&)
446  */
447template<typename MatrixType>
448template<typename InputType>
449FullPivHouseholderQR<MatrixType>& FullPivHouseholderQR<MatrixType>::compute(const EigenBase<InputType>& matrix)
450{
451  m_qr = matrix.derived();
452  computeInPlace();
453  return *this;
454}
455
456template<typename MatrixType>
457void FullPivHouseholderQR<MatrixType>::computeInPlace()
458{
459  check_template_parameters();
460
461  using std::abs;
462  Index rows = m_qr.rows();
463  Index cols = m_qr.cols();
464  Index size = (std::min)(rows,cols);
465
466
467  m_hCoeffs.resize(size);
468
469  m_temp.resize(cols);
470
471  m_precision = NumTraits<Scalar>::epsilon() * RealScalar(size);
472
473  m_rows_transpositions.resize(size);
474  m_cols_transpositions.resize(size);
475  Index number_of_transpositions = 0;
476
477  RealScalar biggest(0);
478
479  m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case)
480  m_maxpivot = RealScalar(0);
481
482  for (Index k = 0; k < size; ++k)
483  {
484    Index row_of_biggest_in_corner, col_of_biggest_in_corner;
485    typedef internal::scalar_score_coeff_op<Scalar> Scoring;
486    typedef typename Scoring::result_type Score;
487
488    Score score = m_qr.bottomRightCorner(rows-k, cols-k)
489                      .unaryExpr(Scoring())
490                      .maxCoeff(&row_of_biggest_in_corner, &col_of_biggest_in_corner);
491    row_of_biggest_in_corner += k;
492    col_of_biggest_in_corner += k;
493    RealScalar biggest_in_corner = internal::abs_knowing_score<Scalar>()(m_qr(row_of_biggest_in_corner, col_of_biggest_in_corner), score);
494    if(k==0) biggest = biggest_in_corner;
495
496    // if the corner is negligible, then we have less than full rank, and we can finish early
497    if(internal::isMuchSmallerThan(biggest_in_corner, biggest, m_precision))
498    {
499      m_nonzero_pivots = k;
500      for(Index i = k; i < size; i++)
501      {
502        m_rows_transpositions.coeffRef(i) = i;
503        m_cols_transpositions.coeffRef(i) = i;
504        m_hCoeffs.coeffRef(i) = Scalar(0);
505      }
506      break;
507    }
508
509    m_rows_transpositions.coeffRef(k) = row_of_biggest_in_corner;
510    m_cols_transpositions.coeffRef(k) = col_of_biggest_in_corner;
511    if(k != row_of_biggest_in_corner) {
512      m_qr.row(k).tail(cols-k).swap(m_qr.row(row_of_biggest_in_corner).tail(cols-k));
513      ++number_of_transpositions;
514    }
515    if(k != col_of_biggest_in_corner) {
516      m_qr.col(k).swap(m_qr.col(col_of_biggest_in_corner));
517      ++number_of_transpositions;
518    }
519
520    RealScalar beta;
521    m_qr.col(k).tail(rows-k).makeHouseholderInPlace(m_hCoeffs.coeffRef(k), beta);
522    m_qr.coeffRef(k,k) = beta;
523
524    // remember the maximum absolute value of diagonal coefficients
525    if(abs(beta) > m_maxpivot) m_maxpivot = abs(beta);
526
527    m_qr.bottomRightCorner(rows-k, cols-k-1)
528        .applyHouseholderOnTheLeft(m_qr.col(k).tail(rows-k-1), m_hCoeffs.coeffRef(k), &m_temp.coeffRef(k+1));
529  }
530
531  m_cols_permutation.setIdentity(cols);
532  for(Index k = 0; k < size; ++k)
533    m_cols_permutation.applyTranspositionOnTheRight(k, m_cols_transpositions.coeff(k));
534
535  m_det_pq = (number_of_transpositions%2) ? -1 : 1;
536  m_isInitialized = true;
537}
538
539#ifndef EIGEN_PARSED_BY_DOXYGEN
540template<typename _MatrixType>
541template<typename RhsType, typename DstType>
542void FullPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const
543{
544  eigen_assert(rhs.rows() == rows());
545  const Index l_rank = rank();
546
547  // FIXME introduce nonzeroPivots() and use it here. and more generally,
548  // make the same improvements in this dec as in FullPivLU.
549  if(l_rank==0)
550  {
551    dst.setZero();
552    return;
553  }
554
555  typename RhsType::PlainObject c(rhs);
556
557  Matrix<Scalar,1,RhsType::ColsAtCompileTime> temp(rhs.cols());
558  for (Index k = 0; k < l_rank; ++k)
559  {
560    Index remainingSize = rows()-k;
561    c.row(k).swap(c.row(m_rows_transpositions.coeff(k)));
562    c.bottomRightCorner(remainingSize, rhs.cols())
563      .applyHouseholderOnTheLeft(m_qr.col(k).tail(remainingSize-1),
564                               m_hCoeffs.coeff(k), &temp.coeffRef(0));
565  }
566
567  m_qr.topLeftCorner(l_rank, l_rank)
568      .template triangularView<Upper>()
569      .solveInPlace(c.topRows(l_rank));
570
571  for(Index i = 0; i < l_rank; ++i) dst.row(m_cols_permutation.indices().coeff(i)) = c.row(i);
572  for(Index i = l_rank; i < cols(); ++i) dst.row(m_cols_permutation.indices().coeff(i)).setZero();
573}
574#endif
575
576namespace internal {
577
578template<typename DstXprType, typename MatrixType>
579struct Assignment<DstXprType, Inverse<FullPivHouseholderQR<MatrixType> >, internal::assign_op<typename DstXprType::Scalar,typename FullPivHouseholderQR<MatrixType>::Scalar>, Dense2Dense>
580{
581  typedef FullPivHouseholderQR<MatrixType> QrType;
582  typedef Inverse<QrType> SrcXprType;
583  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename QrType::Scalar> &)
584  {
585    dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols()));
586  }
587};
588
589/** \ingroup QR_Module
590  *
591  * \brief Expression type for return value of FullPivHouseholderQR::matrixQ()
592  *
593  * \tparam MatrixType type of underlying dense matrix
594  */
595template<typename MatrixType> struct FullPivHouseholderQRMatrixQReturnType
596  : public ReturnByValue<FullPivHouseholderQRMatrixQReturnType<MatrixType> >
597{
598public:
599  typedef typename FullPivHouseholderQR<MatrixType>::IntDiagSizeVectorType IntDiagSizeVectorType;
600  typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
601  typedef Matrix<typename MatrixType::Scalar, 1, MatrixType::RowsAtCompileTime, RowMajor, 1,
602                 MatrixType::MaxRowsAtCompileTime> WorkVectorType;
603
604  FullPivHouseholderQRMatrixQReturnType(const MatrixType&       qr,
605                                        const HCoeffsType&      hCoeffs,
606                                        const IntDiagSizeVectorType& rowsTranspositions)
607    : m_qr(qr),
608      m_hCoeffs(hCoeffs),
609      m_rowsTranspositions(rowsTranspositions)
610  {}
611
612  template <typename ResultType>
613  void evalTo(ResultType& result) const
614  {
615    const Index rows = m_qr.rows();
616    WorkVectorType workspace(rows);
617    evalTo(result, workspace);
618  }
619
620  template <typename ResultType>
621  void evalTo(ResultType& result, WorkVectorType& workspace) const
622  {
623    using numext::conj;
624    // compute the product H'_0 H'_1 ... H'_n-1,
625    // where H_k is the k-th Householder transformation I - h_k v_k v_k'
626    // and v_k is the k-th Householder vector [1,m_qr(k+1,k), m_qr(k+2,k), ...]
627    const Index rows = m_qr.rows();
628    const Index cols = m_qr.cols();
629    const Index size = (std::min)(rows, cols);
630    workspace.resize(rows);
631    result.setIdentity(rows, rows);
632    for (Index k = size-1; k >= 0; k--)
633    {
634      result.block(k, k, rows-k, rows-k)
635            .applyHouseholderOnTheLeft(m_qr.col(k).tail(rows-k-1), conj(m_hCoeffs.coeff(k)), &workspace.coeffRef(k));
636      result.row(k).swap(result.row(m_rowsTranspositions.coeff(k)));
637    }
638  }
639
640  Index rows() const { return m_qr.rows(); }
641  Index cols() const { return m_qr.rows(); }
642
643protected:
644  typename MatrixType::Nested m_qr;
645  typename HCoeffsType::Nested m_hCoeffs;
646  typename IntDiagSizeVectorType::Nested m_rowsTranspositions;
647};
648
649// template<typename MatrixType>
650// struct evaluator<FullPivHouseholderQRMatrixQReturnType<MatrixType> >
651//  : public evaluator<ReturnByValue<FullPivHouseholderQRMatrixQReturnType<MatrixType> > >
652// {};
653
654} // end namespace internal
655
656template<typename MatrixType>
657inline typename FullPivHouseholderQR<MatrixType>::MatrixQReturnType FullPivHouseholderQR<MatrixType>::matrixQ() const
658{
659  eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
660  return MatrixQReturnType(m_qr, m_hCoeffs, m_rows_transpositions);
661}
662
663/** \return the full-pivoting Householder QR decomposition of \c *this.
664  *
665  * \sa class FullPivHouseholderQR
666  */
667template<typename Derived>
668const FullPivHouseholderQR<typename MatrixBase<Derived>::PlainObject>
669MatrixBase<Derived>::fullPivHouseholderQr() const
670{
671  return FullPivHouseholderQR<PlainObject>(eval());
672}
673
674} // end namespace Eigen
675
676#endif // EIGEN_FULLPIVOTINGHOUSEHOLDERQR_H
677