Free cookie consent management tool by TermsFeed Policy Generator

source: branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Eigen/Eigen/src/QR/ColPivHouseholderQR.h @ 9562

Last change on this file since 9562 was 9562, checked in by gkronber, 12 years ago

#1967 worked on Gaussian process evolution.

File size: 19.9 KB
Line 
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_COLPIVOTINGHOUSEHOLDERQR_H
12#define EIGEN_COLPIVOTINGHOUSEHOLDERQR_H
13
14namespace Eigen {
15
16/** \ingroup QR_Module
17  *
18  * \class ColPivHouseholderQR
19  *
20  * \brief Householder rank-revealing QR decomposition of a matrix with column-pivoting
21  *
22  * \param MatrixType the type of the matrix of which we are computing the QR decomposition
23  *
24  * This class performs a rank-revealing QR decomposition of a matrix \b A into matrices \b P, \b Q and \b R
25  * such that
26  * \f[
27  *  \mathbf{A} \, \mathbf{P} = \mathbf{Q} \, \mathbf{R}
28  * \f]
29  * by using Householder transformations. Here, \b P is a permutation matrix, \b Q a unitary matrix and \b R an
30  * upper triangular matrix.
31  *
32  * This decomposition performs column pivoting in order to be rank-revealing and improve
33  * numerical stability. It is slower than HouseholderQR, and faster than FullPivHouseholderQR.
34  *
35  * \sa MatrixBase::colPivHouseholderQr()
36  */
37template<typename _MatrixType> class ColPivHouseholderQR
38{
39  public:
40
41    typedef _MatrixType MatrixType;
42    enum {
43      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
44      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
45      Options = MatrixType::Options,
46      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
47      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
48    };
49    typedef typename MatrixType::Scalar Scalar;
50    typedef typename MatrixType::RealScalar RealScalar;
51    typedef typename MatrixType::Index Index;
52    typedef Matrix<Scalar, RowsAtCompileTime, RowsAtCompileTime, Options, MaxRowsAtCompileTime, MaxRowsAtCompileTime> MatrixQType;
53    typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
54    typedef PermutationMatrix<ColsAtCompileTime, MaxColsAtCompileTime> PermutationType;
55    typedef typename internal::plain_row_type<MatrixType, Index>::type IntRowVectorType;
56    typedef typename internal::plain_row_type<MatrixType>::type RowVectorType;
57    typedef typename internal::plain_row_type<MatrixType, RealScalar>::type RealRowVectorType;
58    typedef typename HouseholderSequence<MatrixType,HCoeffsType>::ConjugateReturnType HouseholderSequenceType;
59   
60  private:
61   
62    typedef typename PermutationType::Index PermIndexType;
63   
64  public:
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 ColPivHouseholderQR::compute(const MatrixType&).
71    */
72    ColPivHouseholderQR()
73      : m_qr(),
74        m_hCoeffs(),
75        m_colsPermutation(),
76        m_colsTranspositions(),
77        m_temp(),
78        m_colSqNorms(),
79        m_isInitialized(false) {}
80
81    /** \brief Default Constructor with memory preallocation
82      *
83      * Like the default constructor but with preallocation of the internal data
84      * according to the specified problem \a size.
85      * \sa ColPivHouseholderQR()
86      */
87    ColPivHouseholderQR(Index rows, Index cols)
88      : m_qr(rows, cols),
89        m_hCoeffs((std::min)(rows,cols)),
90        m_colsPermutation(PermIndexType(cols)),
91        m_colsTranspositions(cols),
92        m_temp(cols),
93        m_colSqNorms(cols),
94        m_isInitialized(false),
95        m_usePrescribedThreshold(false) {}
96
97    ColPivHouseholderQR(const MatrixType& matrix)
98      : m_qr(matrix.rows(), matrix.cols()),
99        m_hCoeffs((std::min)(matrix.rows(),matrix.cols())),
100        m_colsPermutation(PermIndexType(matrix.cols())),
101        m_colsTranspositions(matrix.cols()),
102        m_temp(matrix.cols()),
103        m_colSqNorms(matrix.cols()),
104        m_isInitialized(false),
105        m_usePrescribedThreshold(false)
106    {
107      compute(matrix);
108    }
109
110    /** This method finds a solution x to the equation Ax=b, where A is the matrix of which
111      * *this is the QR decomposition, if any exists.
112      *
113      * \param b the right-hand-side of the equation to solve.
114      *
115      * \returns a solution.
116      *
117      * \note The case where b is a matrix is not yet implemented. Also, this
118      *       code is space inefficient.
119      *
120      * \note_about_checking_solutions
121      *
122      * \note_about_arbitrary_choice_of_solution
123      *
124      * Example: \include ColPivHouseholderQR_solve.cpp
125      * Output: \verbinclude ColPivHouseholderQR_solve.out
126      */
127    template<typename Rhs>
128    inline const internal::solve_retval<ColPivHouseholderQR, Rhs>
129    solve(const MatrixBase<Rhs>& b) const
130    {
131      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
132      return internal::solve_retval<ColPivHouseholderQR, Rhs>(*this, b.derived());
133    }
134
135    HouseholderSequenceType householderQ(void) const;
136
137    /** \returns a reference to the matrix where the Householder QR decomposition is stored
138      */
139    const MatrixType& matrixQR() const
140    {
141      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
142      return m_qr;
143    }
144
145    ColPivHouseholderQR& compute(const MatrixType& matrix);
146
147    const PermutationType& colsPermutation() const
148    {
149      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
150      return m_colsPermutation;
151    }
152
153    /** \returns the absolute value of the determinant of the matrix of which
154      * *this is the QR decomposition. It has only linear complexity
155      * (that is, O(n) where n is the dimension of the square matrix)
156      * as the QR decomposition has already been computed.
157      *
158      * \note This is only for square matrices.
159      *
160      * \warning a determinant can be very big or small, so for matrices
161      * of large enough dimension, there is a risk of overflow/underflow.
162      * One way to work around that is to use logAbsDeterminant() instead.
163      *
164      * \sa logAbsDeterminant(), MatrixBase::determinant()
165      */
166    typename MatrixType::RealScalar absDeterminant() const;
167
168    /** \returns the natural log of the absolute value of the determinant of the matrix of which
169      * *this is the QR decomposition. It has only linear complexity
170      * (that is, O(n) where n is the dimension of the square matrix)
171      * as the QR decomposition has already been computed.
172      *
173      * \note This is only for square matrices.
174      *
175      * \note This method is useful to work around the risk of overflow/underflow that's inherent
176      * to determinant computation.
177      *
178      * \sa absDeterminant(), MatrixBase::determinant()
179      */
180    typename MatrixType::RealScalar logAbsDeterminant() const;
181
182    /** \returns the rank of the matrix of which *this is the QR decomposition.
183      *
184      * \note This method has to determine which pivots should be considered nonzero.
185      *       For that, it uses the threshold value that you can control by calling
186      *       setThreshold(const RealScalar&).
187      */
188    inline Index rank() const
189    {
190      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
191      RealScalar premultiplied_threshold = internal::abs(m_maxpivot) * threshold();
192      Index result = 0;
193      for(Index i = 0; i < m_nonzero_pivots; ++i)
194        result += (internal::abs(m_qr.coeff(i,i)) > premultiplied_threshold);
195      return result;
196    }
197
198    /** \returns the dimension of the kernel of the matrix of which *this is the QR decomposition.
199      *
200      * \note This method has to determine which pivots should be considered nonzero.
201      *       For that, it uses the threshold value that you can control by calling
202      *       setThreshold(const RealScalar&).
203      */
204    inline Index dimensionOfKernel() const
205    {
206      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
207      return cols() - rank();
208    }
209
210    /** \returns true if the matrix of which *this is the QR decomposition represents an injective
211      *          linear map, i.e. has trivial kernel; false otherwise.
212      *
213      * \note This method has to determine which pivots should be considered nonzero.
214      *       For that, it uses the threshold value that you can control by calling
215      *       setThreshold(const RealScalar&).
216      */
217    inline bool isInjective() const
218    {
219      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
220      return rank() == cols();
221    }
222
223    /** \returns true if the matrix of which *this is the QR decomposition represents a surjective
224      *          linear map; false otherwise.
225      *
226      * \note This method has to determine which pivots should be considered nonzero.
227      *       For that, it uses the threshold value that you can control by calling
228      *       setThreshold(const RealScalar&).
229      */
230    inline bool isSurjective() const
231    {
232      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
233      return rank() == rows();
234    }
235
236    /** \returns true if the matrix of which *this is the QR decomposition is invertible.
237      *
238      * \note This method has to determine which pivots should be considered nonzero.
239      *       For that, it uses the threshold value that you can control by calling
240      *       setThreshold(const RealScalar&).
241      */
242    inline bool isInvertible() const
243    {
244      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
245      return isInjective() && isSurjective();
246    }
247
248    /** \returns the inverse of the matrix of which *this is the QR decomposition.
249      *
250      * \note If this matrix is not invertible, the returned matrix has undefined coefficients.
251      *       Use isInvertible() to first determine whether this matrix is invertible.
252      */
253    inline const
254    internal::solve_retval<ColPivHouseholderQR, typename MatrixType::IdentityReturnType>
255    inverse() const
256    {
257      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
258      return internal::solve_retval<ColPivHouseholderQR,typename MatrixType::IdentityReturnType>
259               (*this, MatrixType::Identity(m_qr.rows(), m_qr.cols()));
260    }
261
262    inline Index rows() const { return m_qr.rows(); }
263    inline Index cols() const { return m_qr.cols(); }
264    const HCoeffsType& hCoeffs() const { return m_hCoeffs; }
265
266    /** Allows to prescribe a threshold to be used by certain methods, such as rank(),
267      * who need to determine when pivots are to be considered nonzero. This is not used for the
268      * QR decomposition itself.
269      *
270      * When it needs to get the threshold value, Eigen calls threshold(). By default, this
271      * uses a formula to automatically determine a reasonable threshold.
272      * Once you have called the present method setThreshold(const RealScalar&),
273      * your value is used instead.
274      *
275      * \param threshold The new value to use as the threshold.
276      *
277      * A pivot will be considered nonzero if its absolute value is strictly greater than
278      *  \f$ \vert pivot \vert \leqslant threshold \times \vert maxpivot \vert \f$
279      * where maxpivot is the biggest pivot.
280      *
281      * If you want to come back to the default behavior, call setThreshold(Default_t)
282      */
283    ColPivHouseholderQR& setThreshold(const RealScalar& threshold)
284    {
285      m_usePrescribedThreshold = true;
286      m_prescribedThreshold = threshold;
287      return *this;
288    }
289
290    /** Allows to come back to the default behavior, letting Eigen use its default formula for
291      * determining the threshold.
292      *
293      * You should pass the special object Eigen::Default as parameter here.
294      * \code qr.setThreshold(Eigen::Default); \endcode
295      *
296      * See the documentation of setThreshold(const RealScalar&).
297      */
298    ColPivHouseholderQR& setThreshold(Default_t)
299    {
300      m_usePrescribedThreshold = false;
301      return *this;
302    }
303
304    /** Returns the threshold that will be used by certain methods such as rank().
305      *
306      * See the documentation of setThreshold(const RealScalar&).
307      */
308    RealScalar threshold() const
309    {
310      eigen_assert(m_isInitialized || m_usePrescribedThreshold);
311      return m_usePrescribedThreshold ? m_prescribedThreshold
312      // this formula comes from experimenting (see "LU precision tuning" thread on the list)
313      // and turns out to be identical to Higham's formula used already in LDLt.
314                                      : NumTraits<Scalar>::epsilon() * m_qr.diagonalSize();
315    }
316
317    /** \returns the number of nonzero pivots in the QR decomposition.
318      * Here nonzero is meant in the exact sense, not in a fuzzy sense.
319      * So that notion isn't really intrinsically interesting, but it is
320      * still useful when implementing algorithms.
321      *
322      * \sa rank()
323      */
324    inline Index nonzeroPivots() const
325    {
326      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
327      return m_nonzero_pivots;
328    }
329
330    /** \returns the absolute value of the biggest pivot, i.e. the biggest
331      *          diagonal coefficient of R.
332      */
333    RealScalar maxPivot() const { return m_maxpivot; }
334
335  protected:
336    MatrixType m_qr;
337    HCoeffsType m_hCoeffs;
338    PermutationType m_colsPermutation;
339    IntRowVectorType m_colsTranspositions;
340    RowVectorType m_temp;
341    RealRowVectorType m_colSqNorms;
342    bool m_isInitialized, m_usePrescribedThreshold;
343    RealScalar m_prescribedThreshold, m_maxpivot;
344    Index m_nonzero_pivots;
345    Index m_det_pq;
346};
347
348template<typename MatrixType>
349typename MatrixType::RealScalar ColPivHouseholderQR<MatrixType>::absDeterminant() const
350{
351  eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
352  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
353  return internal::abs(m_qr.diagonal().prod());
354}
355
356template<typename MatrixType>
357typename MatrixType::RealScalar ColPivHouseholderQR<MatrixType>::logAbsDeterminant() const
358{
359  eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
360  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
361  return m_qr.diagonal().cwiseAbs().array().log().sum();
362}
363
364template<typename MatrixType>
365ColPivHouseholderQR<MatrixType>& ColPivHouseholderQR<MatrixType>::compute(const MatrixType& matrix)
366{
367  Index rows = matrix.rows();
368  Index cols = matrix.cols();
369  Index size = matrix.diagonalSize();
370
371  m_qr = matrix;
372  m_hCoeffs.resize(size);
373
374  m_temp.resize(cols);
375
376  m_colsTranspositions.resize(matrix.cols());
377  Index number_of_transpositions = 0;
378
379  m_colSqNorms.resize(cols);
380  for(Index k = 0; k < cols; ++k)
381    m_colSqNorms.coeffRef(k) = m_qr.col(k).squaredNorm();
382
383  RealScalar threshold_helper = m_colSqNorms.maxCoeff() * internal::abs2(NumTraits<Scalar>::epsilon()) / RealScalar(rows);
384
385  m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case)
386  m_maxpivot = RealScalar(0);
387
388  for(Index k = 0; k < size; ++k)
389  {
390    // first, we look up in our table m_colSqNorms which column has the biggest squared norm
391    Index biggest_col_index;
392    RealScalar biggest_col_sq_norm = m_colSqNorms.tail(cols-k).maxCoeff(&biggest_col_index);
393    biggest_col_index += k;
394
395    // since our table m_colSqNorms accumulates imprecision at every step, we must now recompute
396    // the actual squared norm of the selected column.
397    // Note that not doing so does result in solve() sometimes returning inf/nan values
398    // when running the unit test with 1000 repetitions.
399    biggest_col_sq_norm = m_qr.col(biggest_col_index).tail(rows-k).squaredNorm();
400
401    // we store that back into our table: it can't hurt to correct our table.
402    m_colSqNorms.coeffRef(biggest_col_index) = biggest_col_sq_norm;
403
404    // if the current biggest column is smaller than epsilon times the initial biggest column,
405    // terminate to avoid generating nan/inf values.
406    // Note that here, if we test instead for "biggest == 0", we get a failure every 1000 (or so)
407    // repetitions of the unit test, with the result of solve() filled with large values of the order
408    // of 1/(size*epsilon).
409    if(biggest_col_sq_norm < threshold_helper * RealScalar(rows-k))
410    {
411      m_nonzero_pivots = k;
412      m_hCoeffs.tail(size-k).setZero();
413      m_qr.bottomRightCorner(rows-k,cols-k)
414          .template triangularView<StrictlyLower>()
415          .setZero();
416      break;
417    }
418
419    // apply the transposition to the columns
420    m_colsTranspositions.coeffRef(k) = biggest_col_index;
421    if(k != biggest_col_index) {
422      m_qr.col(k).swap(m_qr.col(biggest_col_index));
423      std::swap(m_colSqNorms.coeffRef(k), m_colSqNorms.coeffRef(biggest_col_index));
424      ++number_of_transpositions;
425    }
426
427    // generate the householder vector, store it below the diagonal
428    RealScalar beta;
429    m_qr.col(k).tail(rows-k).makeHouseholderInPlace(m_hCoeffs.coeffRef(k), beta);
430
431    // apply the householder transformation to the diagonal coefficient
432    m_qr.coeffRef(k,k) = beta;
433
434    // remember the maximum absolute value of diagonal coefficients
435    if(internal::abs(beta) > m_maxpivot) m_maxpivot = internal::abs(beta);
436
437    // apply the householder transformation
438    m_qr.bottomRightCorner(rows-k, cols-k-1)
439        .applyHouseholderOnTheLeft(m_qr.col(k).tail(rows-k-1), m_hCoeffs.coeffRef(k), &m_temp.coeffRef(k+1));
440
441    // update our table of squared norms of the columns
442    m_colSqNorms.tail(cols-k-1) -= m_qr.row(k).tail(cols-k-1).cwiseAbs2();
443  }
444
445  m_colsPermutation.setIdentity(PermIndexType(cols));
446  for(PermIndexType k = 0; k < m_nonzero_pivots; ++k)
447    m_colsPermutation.applyTranspositionOnTheRight(PermIndexType(k), PermIndexType(m_colsTranspositions.coeff(k)));
448
449  m_det_pq = (number_of_transpositions%2) ? -1 : 1;
450  m_isInitialized = true;
451
452  return *this;
453}
454
455namespace internal {
456
457template<typename _MatrixType, typename Rhs>
458struct solve_retval<ColPivHouseholderQR<_MatrixType>, Rhs>
459  : solve_retval_base<ColPivHouseholderQR<_MatrixType>, Rhs>
460{
461  EIGEN_MAKE_SOLVE_HELPERS(ColPivHouseholderQR<_MatrixType>,Rhs)
462
463  template<typename Dest> void evalTo(Dest& dst) const
464  {
465    eigen_assert(rhs().rows() == dec().rows());
466
467    const int cols = dec().cols(),
468    nonzero_pivots = dec().nonzeroPivots();
469
470    if(nonzero_pivots == 0)
471    {
472      dst.setZero();
473      return;
474    }
475
476    typename Rhs::PlainObject c(rhs());
477
478    // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T
479    c.applyOnTheLeft(householderSequence(dec().matrixQR(), dec().hCoeffs())
480                     .setLength(dec().nonzeroPivots())
481         .transpose()
482      );
483
484    dec().matrixQR()
485       .topLeftCorner(nonzero_pivots, nonzero_pivots)
486       .template triangularView<Upper>()
487       .solveInPlace(c.topRows(nonzero_pivots));
488
489
490    typename Rhs::PlainObject d(c);
491    d.topRows(nonzero_pivots)
492      = dec().matrixQR()
493       .topLeftCorner(nonzero_pivots, nonzero_pivots)
494       .template triangularView<Upper>()
495       * c.topRows(nonzero_pivots);
496
497    for(Index i = 0; i < nonzero_pivots; ++i) dst.row(dec().colsPermutation().indices().coeff(i)) = c.row(i);
498    for(Index i = nonzero_pivots; i < cols; ++i) dst.row(dec().colsPermutation().indices().coeff(i)).setZero();
499  }
500};
501
502} // end namespace internal
503
504/** \returns the matrix Q as a sequence of householder transformations */
505template<typename MatrixType>
506typename ColPivHouseholderQR<MatrixType>::HouseholderSequenceType ColPivHouseholderQR<MatrixType>
507  ::householderQ() const
508{
509  eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
510  return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate()).setLength(m_nonzero_pivots);
511}
512
513/** \return the column-pivoting Householder QR decomposition of \c *this.
514  *
515  * \sa class ColPivHouseholderQR
516  */
517template<typename Derived>
518const ColPivHouseholderQR<typename MatrixBase<Derived>::PlainObject>
519MatrixBase<Derived>::colPivHouseholderQr() const
520{
521  return ColPivHouseholderQR<PlainObject>(eval());
522}
523
524} // end namespace Eigen
525
526#endif // EIGEN_COLPIVOTINGHOUSEHOLDERQR_H
Note: See TracBrowser for help on using the repository browser.