Free cookie consent management tool by TermsFeed Policy Generator

source: branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Eigen/Eigen/src/Cholesky/LLT.h @ 10586

Last change on this file since 10586 was 9562, checked in by gkronber, 12 years ago

#1967 worked on Gaussian process evolution.

File size: 15.6 KB
Line 
1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
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_LLT_H
11#define EIGEN_LLT_H
12
13namespace Eigen {
14
15namespace internal{
16template<typename MatrixType, int UpLo> struct LLT_Traits;
17}
18
19/** \ingroup Cholesky_Module
20  *
21  * \class LLT
22  *
23  * \brief Standard Cholesky decomposition (LL^T) of a matrix and associated features
24  *
25  * \param MatrixType the type of the matrix of which we are computing the LL^T Cholesky decomposition
26  * \param UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper.
27  *             The other triangular part won't be read.
28  *
29  * This class performs a LL^T Cholesky decomposition of a symmetric, positive definite
30  * matrix A such that A = LL^* = U^*U, where L is lower triangular.
31  *
32  * While the Cholesky decomposition is particularly useful to solve selfadjoint problems like  D^*D x = b,
33  * for that purpose, we recommend the Cholesky decomposition without square root which is more stable
34  * and even faster. Nevertheless, this standard Cholesky decomposition remains useful in many other
35  * situations like generalised eigen problems with hermitian matrices.
36  *
37  * Remember that Cholesky decompositions are not rank-revealing. This LLT decomposition is only stable on positive definite matrices,
38  * use LDLT instead for the semidefinite case. Also, do not use a Cholesky decomposition to determine whether a system of equations
39  * has a solution.
40  *
41  * Example: \include LLT_example.cpp
42  * Output: \verbinclude LLT_example.out
43  *   
44  * \sa MatrixBase::llt(), class LDLT
45  */
46 /* HEY THIS DOX IS DISABLED BECAUSE THERE's A BUG EITHER HERE OR IN LDLT ABOUT THAT (OR BOTH)
47  * Note that during the decomposition, only the upper triangular part of A is considered. Therefore,
48  * the strict lower part does not have to store correct values.
49  */
50template<typename _MatrixType, int _UpLo> class LLT
51{
52  public:
53    typedef _MatrixType MatrixType;
54    enum {
55      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
56      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
57      Options = MatrixType::Options,
58      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
59    };
60    typedef typename MatrixType::Scalar Scalar;
61    typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
62    typedef typename MatrixType::Index Index;
63
64    enum {
65      PacketSize = internal::packet_traits<Scalar>::size,
66      AlignmentMask = int(PacketSize)-1,
67      UpLo = _UpLo
68    };
69
70    typedef internal::LLT_Traits<MatrixType,UpLo> Traits;
71
72    /**
73      * \brief Default Constructor.
74      *
75      * The default constructor is useful in cases in which the user intends to
76      * perform decompositions via LLT::compute(const MatrixType&).
77      */
78    LLT() : m_matrix(), m_isInitialized(false) {}
79
80    /** \brief Default Constructor with memory preallocation
81      *
82      * Like the default constructor but with preallocation of the internal data
83      * according to the specified problem \a size.
84      * \sa LLT()
85      */
86    LLT(Index size) : m_matrix(size, size),
87                    m_isInitialized(false) {}
88
89    LLT(const MatrixType& matrix)
90      : m_matrix(matrix.rows(), matrix.cols()),
91        m_isInitialized(false)
92    {
93      compute(matrix);
94    }
95
96    /** \returns a view of the upper triangular matrix U */
97    inline typename Traits::MatrixU matrixU() const
98    {
99      eigen_assert(m_isInitialized && "LLT is not initialized.");
100      return Traits::getU(m_matrix);
101    }
102
103    /** \returns a view of the lower triangular matrix L */
104    inline typename Traits::MatrixL matrixL() const
105    {
106      eigen_assert(m_isInitialized && "LLT is not initialized.");
107      return Traits::getL(m_matrix);
108    }
109
110    /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A.
111      *
112      * Since this LLT class assumes anyway that the matrix A is invertible, the solution
113      * theoretically exists and is unique regardless of b.
114      *
115      * Example: \include LLT_solve.cpp
116      * Output: \verbinclude LLT_solve.out
117      *
118      * \sa solveInPlace(), MatrixBase::llt()
119      */
120    template<typename Rhs>
121    inline const internal::solve_retval<LLT, Rhs>
122    solve(const MatrixBase<Rhs>& b) const
123    {
124      eigen_assert(m_isInitialized && "LLT is not initialized.");
125      eigen_assert(m_matrix.rows()==b.rows()
126                && "LLT::solve(): invalid number of rows of the right hand side matrix b");
127      return internal::solve_retval<LLT, Rhs>(*this, b.derived());
128    }
129
130    #ifdef EIGEN2_SUPPORT
131    template<typename OtherDerived, typename ResultType>
132    bool solve(const MatrixBase<OtherDerived>& b, ResultType *result) const
133    {
134      *result = this->solve(b);
135      return true;
136    }
137   
138    bool isPositiveDefinite() const { return true; }
139    #endif
140
141    template<typename Derived>
142    void solveInPlace(MatrixBase<Derived> &bAndX) const;
143
144    LLT& compute(const MatrixType& matrix);
145
146    /** \returns the LLT decomposition matrix
147      *
148      * TODO: document the storage layout
149      */
150    inline const MatrixType& matrixLLT() const
151    {
152      eigen_assert(m_isInitialized && "LLT is not initialized.");
153      return m_matrix;
154    }
155
156    MatrixType reconstructedMatrix() const;
157
158
159    /** \brief Reports whether previous computation was successful.
160      *
161      * \returns \c Success if computation was succesful,
162      *          \c NumericalIssue if the matrix.appears to be negative.
163      */
164    ComputationInfo info() const
165    {
166      eigen_assert(m_isInitialized && "LLT is not initialized.");
167      return m_info;
168    }
169
170    inline Index rows() const { return m_matrix.rows(); }
171    inline Index cols() const { return m_matrix.cols(); }
172
173    template<typename VectorType>
174    LLT rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
175
176  protected:
177    /** \internal
178      * Used to compute and store L
179      * The strict upper part is not used and even not initialized.
180      */
181    MatrixType m_matrix;
182    bool m_isInitialized;
183    ComputationInfo m_info;
184};
185
186namespace internal {
187
188template<typename Scalar, int UpLo> struct llt_inplace;
189
190template<typename MatrixType, typename VectorType>
191static typename MatrixType::Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma)
192{
193  typedef typename MatrixType::Scalar Scalar;
194  typedef typename MatrixType::RealScalar RealScalar;
195  typedef typename MatrixType::Index Index;
196  typedef typename MatrixType::ColXpr ColXpr;
197  typedef typename internal::remove_all<ColXpr>::type ColXprCleaned;
198  typedef typename ColXprCleaned::SegmentReturnType ColXprSegment;
199  typedef Matrix<Scalar,Dynamic,1> TempVectorType;
200  typedef typename TempVectorType::SegmentReturnType TempVecSegment;
201
202  int n = mat.cols();
203  eigen_assert(mat.rows()==n && vec.size()==n);
204
205  TempVectorType temp;
206
207  if(sigma>0)
208  {
209    // This version is based on Givens rotations.
210    // It is faster than the other one below, but only works for updates,
211    // i.e., for sigma > 0
212    temp = sqrt(sigma) * vec;
213
214    for(int i=0; i<n; ++i)
215    {
216      JacobiRotation<Scalar> g;
217      g.makeGivens(mat(i,i), -temp(i), &mat(i,i));
218
219      int rs = n-i-1;
220      if(rs>0)
221      {
222        ColXprSegment x(mat.col(i).tail(rs));
223        TempVecSegment y(temp.tail(rs));
224        apply_rotation_in_the_plane(x, y, g);
225      }
226    }
227  }
228  else
229  {
230    temp = vec;
231    RealScalar beta = 1;
232    for(int j=0; j<n; ++j)
233    {
234      RealScalar Ljj = real(mat.coeff(j,j));
235      RealScalar dj = abs2(Ljj);
236      Scalar wj = temp.coeff(j);
237      RealScalar swj2 = sigma*abs2(wj);
238      RealScalar gamma = dj*beta + swj2;
239
240      RealScalar x = dj + swj2/beta;
241      if (x<=RealScalar(0))
242        return j;
243      RealScalar nLjj = sqrt(x);
244      mat.coeffRef(j,j) = nLjj;
245      beta += swj2/dj;
246
247      // Update the terms of L
248      Index rs = n-j-1;
249      if(rs)
250      {
251        temp.tail(rs) -= (wj/Ljj) * mat.col(j).tail(rs);
252        if(gamma != 0)
253          mat.col(j).tail(rs) = (nLjj/Ljj) * mat.col(j).tail(rs) + (nLjj * sigma*conj(wj)/gamma)*temp.tail(rs);
254      }
255    }
256  }
257  return -1;
258}
259
260template<typename Scalar> struct llt_inplace<Scalar, Lower>
261{
262  typedef typename NumTraits<Scalar>::Real RealScalar;
263  template<typename MatrixType>
264  static typename MatrixType::Index unblocked(MatrixType& mat)
265  {
266    typedef typename MatrixType::Index Index;
267   
268    eigen_assert(mat.rows()==mat.cols());
269    const Index size = mat.rows();
270    for(Index k = 0; k < size; ++k)
271    {
272      Index rs = size-k-1; // remaining size
273
274      Block<MatrixType,Dynamic,1> A21(mat,k+1,k,rs,1);
275      Block<MatrixType,1,Dynamic> A10(mat,k,0,1,k);
276      Block<MatrixType,Dynamic,Dynamic> A20(mat,k+1,0,rs,k);
277
278      RealScalar x = real(mat.coeff(k,k));
279      if (k>0) x -= A10.squaredNorm();
280      if (x<=RealScalar(0))
281        return k;
282      mat.coeffRef(k,k) = x = sqrt(x);
283      if (k>0 && rs>0) A21.noalias() -= A20 * A10.adjoint();
284      if (rs>0) A21 *= RealScalar(1)/x;
285    }
286    return -1;
287  }
288
289  template<typename MatrixType>
290  static typename MatrixType::Index blocked(MatrixType& m)
291  {
292    typedef typename MatrixType::Index Index;
293    eigen_assert(m.rows()==m.cols());
294    Index size = m.rows();
295    if(size<32)
296      return unblocked(m);
297
298    Index blockSize = size/8;
299    blockSize = (blockSize/16)*16;
300    blockSize = (std::min)((std::max)(blockSize,Index(8)), Index(128));
301
302    for (Index k=0; k<size; k+=blockSize)
303    {
304      // partition the matrix:
305      //       A00 |  -  |  -
306      // lu  = A10 | A11 |  -
307      //       A20 | A21 | A22
308      Index bs = (std::min)(blockSize, size-k);
309      Index rs = size - k - bs;
310      Block<MatrixType,Dynamic,Dynamic> A11(m,k,   k,   bs,bs);
311      Block<MatrixType,Dynamic,Dynamic> A21(m,k+bs,k,   rs,bs);
312      Block<MatrixType,Dynamic,Dynamic> A22(m,k+bs,k+bs,rs,rs);
313
314      Index ret;
315      if((ret=unblocked(A11))>=0) return k+ret;
316      if(rs>0) A11.adjoint().template triangularView<Upper>().template solveInPlace<OnTheRight>(A21);
317      if(rs>0) A22.template selfadjointView<Lower>().rankUpdate(A21,-1); // bottleneck
318    }
319    return -1;
320  }
321
322  template<typename MatrixType, typename VectorType>
323  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
324  {
325    return Eigen::internal::llt_rank_update_lower(mat, vec, sigma);
326  }
327};
328 
329template<typename Scalar> struct llt_inplace<Scalar, Upper>
330{
331  typedef typename NumTraits<Scalar>::Real RealScalar;
332
333  template<typename MatrixType>
334  static EIGEN_STRONG_INLINE typename MatrixType::Index unblocked(MatrixType& mat)
335  {
336    Transpose<MatrixType> matt(mat);
337    return llt_inplace<Scalar, Lower>::unblocked(matt);
338  }
339  template<typename MatrixType>
340  static EIGEN_STRONG_INLINE typename MatrixType::Index blocked(MatrixType& mat)
341  {
342    Transpose<MatrixType> matt(mat);
343    return llt_inplace<Scalar, Lower>::blocked(matt);
344  }
345  template<typename MatrixType, typename VectorType>
346  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
347  {
348    Transpose<MatrixType> matt(mat);
349    return llt_inplace<Scalar, Lower>::rankUpdate(matt, vec.conjugate(), sigma);
350  }
351};
352
353template<typename MatrixType> struct LLT_Traits<MatrixType,Lower>
354{
355  typedef const TriangularView<const MatrixType, Lower> MatrixL;
356  typedef const TriangularView<const typename MatrixType::AdjointReturnType, Upper> MatrixU;
357  static inline MatrixL getL(const MatrixType& m) { return m; }
358  static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
359  static bool inplace_decomposition(MatrixType& m)
360  { return llt_inplace<typename MatrixType::Scalar, Lower>::blocked(m)==-1; }
361};
362
363template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
364{
365  typedef const TriangularView<const typename MatrixType::AdjointReturnType, Lower> MatrixL;
366  typedef const TriangularView<const MatrixType, Upper> MatrixU;
367  static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
368  static inline MatrixU getU(const MatrixType& m) { return m; }
369  static bool inplace_decomposition(MatrixType& m)
370  { return llt_inplace<typename MatrixType::Scalar, Upper>::blocked(m)==-1; }
371};
372
373} // end namespace internal
374
375/** Computes / recomputes the Cholesky decomposition A = LL^* = U^*U of \a matrix
376  *
377  * \returns a reference to *this
378  *
379  * Example: \include TutorialLinAlgComputeTwice.cpp
380  * Output: \verbinclude TutorialLinAlgComputeTwice.out
381  */
382template<typename MatrixType, int _UpLo>
383LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
384{
385  eigen_assert(a.rows()==a.cols());
386  const Index size = a.rows();
387  m_matrix.resize(size, size);
388  m_matrix = a;
389
390  m_isInitialized = true;
391  bool ok = Traits::inplace_decomposition(m_matrix);
392  m_info = ok ? Success : NumericalIssue;
393
394  return *this;
395}
396
397/** Performs a rank one update (or dowdate) of the current decomposition.
398  * If A = LL^* before the rank one update,
399  * then after it we have LL^* = A + sigma * v v^* where \a v must be a vector
400  * of same dimension.
401  */
402template<typename _MatrixType, int _UpLo>
403template<typename VectorType>
404LLT<_MatrixType,_UpLo> LLT<_MatrixType,_UpLo>::rankUpdate(const VectorType& v, const RealScalar& sigma)
405{
406  EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorType);
407  eigen_assert(v.size()==m_matrix.cols());
408  eigen_assert(m_isInitialized);
409  if(internal::llt_inplace<typename MatrixType::Scalar, UpLo>::rankUpdate(m_matrix,v,sigma)>=0)
410    m_info = NumericalIssue;
411  else
412    m_info = Success;
413
414  return *this;
415}
416   
417namespace internal {
418template<typename _MatrixType, int UpLo, typename Rhs>
419struct solve_retval<LLT<_MatrixType, UpLo>, Rhs>
420  : solve_retval_base<LLT<_MatrixType, UpLo>, Rhs>
421{
422  typedef LLT<_MatrixType,UpLo> LLTType;
423  EIGEN_MAKE_SOLVE_HELPERS(LLTType,Rhs)
424
425  template<typename Dest> void evalTo(Dest& dst) const
426  {
427    dst = rhs();
428    dec().solveInPlace(dst);
429  }
430};
431}
432
433/** \internal use x = llt_object.solve(x);
434  *
435  * This is the \em in-place version of solve().
436  *
437  * \param bAndX represents both the right-hand side matrix b and result x.
438  *
439  * \returns true always! If you need to check for existence of solutions, use another decomposition like LU, QR, or SVD.
440  *
441  * This version avoids a copy when the right hand side matrix b is not
442  * needed anymore.
443  *
444  * \sa LLT::solve(), MatrixBase::llt()
445  */
446template<typename MatrixType, int _UpLo>
447template<typename Derived>
448void LLT<MatrixType,_UpLo>::solveInPlace(MatrixBase<Derived> &bAndX) const
449{
450  eigen_assert(m_isInitialized && "LLT is not initialized.");
451  eigen_assert(m_matrix.rows()==bAndX.rows());
452  matrixL().solveInPlace(bAndX);
453  matrixU().solveInPlace(bAndX);
454}
455
456/** \returns the matrix represented by the decomposition,
457 * i.e., it returns the product: L L^*.
458 * This function is provided for debug purpose. */
459template<typename MatrixType, int _UpLo>
460MatrixType LLT<MatrixType,_UpLo>::reconstructedMatrix() const
461{
462  eigen_assert(m_isInitialized && "LLT is not initialized.");
463  return matrixL() * matrixL().adjoint().toDenseMatrix();
464}
465
466/** \cholesky_module
467  * \returns the LLT decomposition of \c *this
468  */
469template<typename Derived>
470inline const LLT<typename MatrixBase<Derived>::PlainObject>
471MatrixBase<Derived>::llt() const
472{
473  return LLT<PlainObject>(derived());
474}
475
476/** \cholesky_module
477  * \returns the LLT decomposition of \c *this
478  */
479template<typename MatrixType, unsigned int UpLo>
480inline const LLT<typename SelfAdjointView<MatrixType, UpLo>::PlainObject, UpLo>
481SelfAdjointView<MatrixType, UpLo>::llt() const
482{
483  return LLT<PlainObject,UpLo>(m_matrix);
484}
485
486} // end namespace Eigen
487
488#endif // EIGEN_LLT_H
Note: See TracBrowser for help on using the repository browser.