• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
5 // Copyright (C) 2009 Keir Mierle <mierle@gmail.com>
6 // Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
7 // Copyright (C) 2011 Timothy E. Holy <tim.holy@gmail.com >
8 //
9 // This Source Code Form is subject to the terms of the Mozilla
10 // Public License v. 2.0. If a copy of the MPL was not distributed
11 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
12 
13 #ifndef EIGEN_LDLT_H
14 #define EIGEN_LDLT_H
15 
16 namespace Eigen {
17 
18 namespace internal {
19 template<typename MatrixType, int UpLo> struct LDLT_Traits;
20 }
21 
22 /** \ingroup Cholesky_Module
23   *
24   * \class LDLT
25   *
26   * \brief Robust Cholesky decomposition of a matrix with pivoting
27   *
28   * \param MatrixType the type of the matrix of which to compute the LDL^T Cholesky decomposition
29   * \param UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper.
30   *             The other triangular part won't be read.
31   *
32   * Perform a robust Cholesky decomposition of a positive semidefinite or negative semidefinite
33   * matrix \f$ A \f$ such that \f$ A =  P^TLDL^*P \f$, where P is a permutation matrix, L
34   * is lower triangular with a unit diagonal and D is a diagonal matrix.
35   *
36   * The decomposition uses pivoting to ensure stability, so that L will have
37   * zeros in the bottom right rank(A) - n submatrix. Avoiding the square root
38   * on D also stabilizes the computation.
39   *
40   * Remember that Cholesky decompositions are not rank-revealing. Also, do not use a Cholesky
41   * decomposition to determine whether a system of equations has a solution.
42   *
43   * \sa MatrixBase::ldlt(), class LLT
44   */
45 template<typename _MatrixType, int _UpLo> class LDLT
46 {
47   public:
48     typedef _MatrixType MatrixType;
49     enum {
50       RowsAtCompileTime = MatrixType::RowsAtCompileTime,
51       ColsAtCompileTime = MatrixType::ColsAtCompileTime,
52       Options = MatrixType::Options & ~RowMajorBit, // these are the options for the TmpMatrixType, we need a ColMajor matrix here!
53       MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
54       MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
55       UpLo = _UpLo
56     };
57     typedef typename MatrixType::Scalar Scalar;
58     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
59     typedef typename MatrixType::Index Index;
60     typedef Matrix<Scalar, RowsAtCompileTime, 1, Options, MaxRowsAtCompileTime, 1> TmpMatrixType;
61 
62     typedef Transpositions<RowsAtCompileTime, MaxRowsAtCompileTime> TranspositionType;
63     typedef PermutationMatrix<RowsAtCompileTime, MaxRowsAtCompileTime> PermutationType;
64 
65     typedef internal::LDLT_Traits<MatrixType,UpLo> Traits;
66 
67     /** \brief Default Constructor.
68       *
69       * The default constructor is useful in cases in which the user intends to
70       * perform decompositions via LDLT::compute(const MatrixType&).
71       */
LDLT()72     LDLT() : m_matrix(), m_transpositions(), m_isInitialized(false) {}
73 
74     /** \brief Default Constructor with memory preallocation
75       *
76       * Like the default constructor but with preallocation of the internal data
77       * according to the specified problem \a size.
78       * \sa LDLT()
79       */
LDLT(Index size)80     LDLT(Index size)
81       : m_matrix(size, size),
82         m_transpositions(size),
83         m_temporary(size),
84         m_isInitialized(false)
85     {}
86 
87     /** \brief Constructor with decomposition
88       *
89       * This calculates the decomposition for the input \a matrix.
90       * \sa LDLT(Index size)
91       */
LDLT(const MatrixType & matrix)92     LDLT(const MatrixType& matrix)
93       : m_matrix(matrix.rows(), matrix.cols()),
94         m_transpositions(matrix.rows()),
95         m_temporary(matrix.rows()),
96         m_isInitialized(false)
97     {
98       compute(matrix);
99     }
100 
101     /** Clear any existing decomposition
102      * \sa rankUpdate(w,sigma)
103      */
setZero()104     void setZero()
105     {
106       m_isInitialized = false;
107     }
108 
109     /** \returns a view of the upper triangular matrix U */
matrixU()110     inline typename Traits::MatrixU matrixU() const
111     {
112       eigen_assert(m_isInitialized && "LDLT is not initialized.");
113       return Traits::getU(m_matrix);
114     }
115 
116     /** \returns a view of the lower triangular matrix L */
matrixL()117     inline typename Traits::MatrixL matrixL() const
118     {
119       eigen_assert(m_isInitialized && "LDLT is not initialized.");
120       return Traits::getL(m_matrix);
121     }
122 
123     /** \returns the permutation matrix P as a transposition sequence.
124       */
transpositionsP()125     inline const TranspositionType& transpositionsP() const
126     {
127       eigen_assert(m_isInitialized && "LDLT is not initialized.");
128       return m_transpositions;
129     }
130 
131     /** \returns the coefficients of the diagonal matrix D */
vectorD()132     inline Diagonal<const MatrixType> vectorD() const
133     {
134       eigen_assert(m_isInitialized && "LDLT is not initialized.");
135       return m_matrix.diagonal();
136     }
137 
138     /** \returns true if the matrix is positive (semidefinite) */
isPositive()139     inline bool isPositive() const
140     {
141       eigen_assert(m_isInitialized && "LDLT is not initialized.");
142       return m_sign == 1;
143     }
144 
145     #ifdef EIGEN2_SUPPORT
isPositiveDefinite()146     inline bool isPositiveDefinite() const
147     {
148       return isPositive();
149     }
150     #endif
151 
152     /** \returns true if the matrix is negative (semidefinite) */
isNegative(void)153     inline bool isNegative(void) const
154     {
155       eigen_assert(m_isInitialized && "LDLT is not initialized.");
156       return m_sign == -1;
157     }
158 
159     /** \returns a solution x of \f$ A x = b \f$ using the current decomposition of A.
160       *
161       * This function also supports in-place solves using the syntax <tt>x = decompositionObject.solve(x)</tt> .
162       *
163       * \note_about_checking_solutions
164       *
165       * More precisely, this method solves \f$ A x = b \f$ using the decomposition \f$ A = P^T L D L^* P \f$
166       * by solving the systems \f$ P^T y_1 = b \f$, \f$ L y_2 = y_1 \f$, \f$ D y_3 = y_2 \f$,
167       * \f$ L^* y_4 = y_3 \f$ and \f$ P x = y_4 \f$ in succession. If the matrix \f$ A \f$ is singular, then
168       * \f$ D \f$ will also be singular (all the other matrices are invertible). In that case, the
169       * least-square solution of \f$ D y_3 = y_2 \f$ is computed. This does not mean that this function
170       * computes the least-square solution of \f$ A x = b \f$ is \f$ A \f$ is singular.
171       *
172       * \sa MatrixBase::ldlt()
173       */
174     template<typename Rhs>
175     inline const internal::solve_retval<LDLT, Rhs>
solve(const MatrixBase<Rhs> & b)176     solve(const MatrixBase<Rhs>& b) const
177     {
178       eigen_assert(m_isInitialized && "LDLT is not initialized.");
179       eigen_assert(m_matrix.rows()==b.rows()
180                 && "LDLT::solve(): invalid number of rows of the right hand side matrix b");
181       return internal::solve_retval<LDLT, Rhs>(*this, b.derived());
182     }
183 
184     #ifdef EIGEN2_SUPPORT
185     template<typename OtherDerived, typename ResultType>
solve(const MatrixBase<OtherDerived> & b,ResultType * result)186     bool solve(const MatrixBase<OtherDerived>& b, ResultType *result) const
187     {
188       *result = this->solve(b);
189       return true;
190     }
191     #endif
192 
193     template<typename Derived>
194     bool solveInPlace(MatrixBase<Derived> &bAndX) const;
195 
196     LDLT& compute(const MatrixType& matrix);
197 
198     template <typename Derived>
199     LDLT& rankUpdate(const MatrixBase<Derived>& w,RealScalar alpha=1);
200 
201     /** \returns the internal LDLT decomposition matrix
202       *
203       * TODO: document the storage layout
204       */
matrixLDLT()205     inline const MatrixType& matrixLDLT() const
206     {
207       eigen_assert(m_isInitialized && "LDLT is not initialized.");
208       return m_matrix;
209     }
210 
211     MatrixType reconstructedMatrix() const;
212 
rows()213     inline Index rows() const { return m_matrix.rows(); }
cols()214     inline Index cols() const { return m_matrix.cols(); }
215 
216     /** \brief Reports whether previous computation was successful.
217       *
218       * \returns \c Success if computation was succesful,
219       *          \c NumericalIssue if the matrix.appears to be negative.
220       */
info()221     ComputationInfo info() const
222     {
223       eigen_assert(m_isInitialized && "LDLT is not initialized.");
224       return Success;
225     }
226 
227   protected:
228 
229     /** \internal
230       * Used to compute and store the Cholesky decomposition A = L D L^* = U^* D U.
231       * The strict upper part is used during the decomposition, the strict lower
232       * part correspond to the coefficients of L (its diagonal is equal to 1 and
233       * is not stored), and the diagonal entries correspond to D.
234       */
235     MatrixType m_matrix;
236     TranspositionType m_transpositions;
237     TmpMatrixType m_temporary;
238     int m_sign;
239     bool m_isInitialized;
240 };
241 
242 namespace internal {
243 
244 template<int UpLo> struct ldlt_inplace;
245 
246 template<> struct ldlt_inplace<Lower>
247 {
248   template<typename MatrixType, typename TranspositionType, typename Workspace>
249   static bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, int* sign=0)
250   {
251     typedef typename MatrixType::Scalar Scalar;
252     typedef typename MatrixType::RealScalar RealScalar;
253     typedef typename MatrixType::Index Index;
254     eigen_assert(mat.rows()==mat.cols());
255     const Index size = mat.rows();
256 
257     if (size <= 1)
258     {
259       transpositions.setIdentity();
260       if(sign)
261         *sign = real(mat.coeff(0,0))>0 ? 1:-1;
262       return true;
263     }
264 
265     RealScalar cutoff(0), biggest_in_corner;
266 
267     for (Index k = 0; k < size; ++k)
268     {
269       // Find largest diagonal element
270       Index index_of_biggest_in_corner;
271       biggest_in_corner = mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner);
272       index_of_biggest_in_corner += k;
273 
274       if(k == 0)
275       {
276         // The biggest overall is the point of reference to which further diagonals
277         // are compared; if any diagonal is negligible compared
278         // to the largest overall, the algorithm bails.
279         cutoff = abs(NumTraits<Scalar>::epsilon() * biggest_in_corner);
280 
281         if(sign)
282           *sign = real(mat.diagonal().coeff(index_of_biggest_in_corner)) > 0 ? 1 : -1;
283       }
284 
285       // Finish early if the matrix is not full rank.
286       if(biggest_in_corner < cutoff)
287       {
288         for(Index i = k; i < size; i++) transpositions.coeffRef(i) = i;
289         break;
290       }
291 
292       transpositions.coeffRef(k) = index_of_biggest_in_corner;
293       if(k != index_of_biggest_in_corner)
294       {
295         // apply the transposition while taking care to consider only
296         // the lower triangular part
297         Index s = size-index_of_biggest_in_corner-1; // trailing size after the biggest element
298         mat.row(k).head(k).swap(mat.row(index_of_biggest_in_corner).head(k));
299         mat.col(k).tail(s).swap(mat.col(index_of_biggest_in_corner).tail(s));
300         std::swap(mat.coeffRef(k,k),mat.coeffRef(index_of_biggest_in_corner,index_of_biggest_in_corner));
301         for(int i=k+1;i<index_of_biggest_in_corner;++i)
302         {
303           Scalar tmp = mat.coeffRef(i,k);
304           mat.coeffRef(i,k) = conj(mat.coeffRef(index_of_biggest_in_corner,i));
305           mat.coeffRef(index_of_biggest_in_corner,i) = conj(tmp);
306         }
307         if(NumTraits<Scalar>::IsComplex)
308           mat.coeffRef(index_of_biggest_in_corner,k) = conj(mat.coeff(index_of_biggest_in_corner,k));
309       }
310 
311       // partition the matrix:
312       //       A00 |  -  |  -
313       // lu  = A10 | A11 |  -
314       //       A20 | A21 | A22
315       Index rs = size - k - 1;
316       Block<MatrixType,Dynamic,1> A21(mat,k+1,k,rs,1);
317       Block<MatrixType,1,Dynamic> A10(mat,k,0,1,k);
318       Block<MatrixType,Dynamic,Dynamic> A20(mat,k+1,0,rs,k);
319 
320       if(k>0)
321       {
322         temp.head(k) = mat.diagonal().head(k).asDiagonal() * A10.adjoint();
323         mat.coeffRef(k,k) -= (A10 * temp.head(k)).value();
324         if(rs>0)
325           A21.noalias() -= A20 * temp.head(k);
326       }
327       if((rs>0) && (abs(mat.coeffRef(k,k)) > cutoff))
328         A21 /= mat.coeffRef(k,k);
329     }
330 
331     return true;
332   }
333 
334   // Reference for the algorithm: Davis and Hager, "Multiple Rank
335   // Modifications of a Sparse Cholesky Factorization" (Algorithm 1)
336   // Trivial rearrangements of their computations (Timothy E. Holy)
337   // allow their algorithm to work for rank-1 updates even if the
338   // original matrix is not of full rank.
339   // Here only rank-1 updates are implemented, to reduce the
340   // requirement for intermediate storage and improve accuracy
341   template<typename MatrixType, typename WDerived>
342   static bool updateInPlace(MatrixType& mat, MatrixBase<WDerived>& w, typename MatrixType::RealScalar sigma=1)
343   {
344     using internal::isfinite;
345     typedef typename MatrixType::Scalar Scalar;
346     typedef typename MatrixType::RealScalar RealScalar;
347     typedef typename MatrixType::Index Index;
348 
349     const Index size = mat.rows();
350     eigen_assert(mat.cols() == size && w.size()==size);
351 
352     RealScalar alpha = 1;
353 
354     // Apply the update
355     for (Index j = 0; j < size; j++)
356     {
357       // Check for termination due to an original decomposition of low-rank
358       if (!(isfinite)(alpha))
359         break;
360 
361       // Update the diagonal terms
362       RealScalar dj = real(mat.coeff(j,j));
363       Scalar wj = w.coeff(j);
364       RealScalar swj2 = sigma*abs2(wj);
365       RealScalar gamma = dj*alpha + swj2;
366 
367       mat.coeffRef(j,j) += swj2/alpha;
368       alpha += swj2/dj;
369 
370 
371       // Update the terms of L
372       Index rs = size-j-1;
373       w.tail(rs) -= wj * mat.col(j).tail(rs);
374       if(gamma != 0)
375         mat.col(j).tail(rs) += (sigma*conj(wj)/gamma)*w.tail(rs);
376     }
377     return true;
378   }
379 
380   template<typename MatrixType, typename TranspositionType, typename Workspace, typename WType>
381   static bool update(MatrixType& mat, const TranspositionType& transpositions, Workspace& tmp, const WType& w, typename MatrixType::RealScalar sigma=1)
382   {
383     // Apply the permutation to the input w
384     tmp = transpositions * w;
385 
386     return ldlt_inplace<Lower>::updateInPlace(mat,tmp,sigma);
387   }
388 };
389 
390 template<> struct ldlt_inplace<Upper>
391 {
392   template<typename MatrixType, typename TranspositionType, typename Workspace>
393   static EIGEN_STRONG_INLINE bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, int* sign=0)
394   {
395     Transpose<MatrixType> matt(mat);
396     return ldlt_inplace<Lower>::unblocked(matt, transpositions, temp, sign);
397   }
398 
399   template<typename MatrixType, typename TranspositionType, typename Workspace, typename WType>
400   static EIGEN_STRONG_INLINE bool update(MatrixType& mat, TranspositionType& transpositions, Workspace& tmp, WType& w, typename MatrixType::RealScalar sigma=1)
401   {
402     Transpose<MatrixType> matt(mat);
403     return ldlt_inplace<Lower>::update(matt, transpositions, tmp, w.conjugate(), sigma);
404   }
405 };
406 
407 template<typename MatrixType> struct LDLT_Traits<MatrixType,Lower>
408 {
409   typedef const TriangularView<const MatrixType, UnitLower> MatrixL;
410   typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitUpper> MatrixU;
411   static inline MatrixL getL(const MatrixType& m) { return m; }
412   static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
413 };
414 
415 template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
416 {
417   typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitLower> MatrixL;
418   typedef const TriangularView<const MatrixType, UnitUpper> MatrixU;
419   static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
420   static inline MatrixU getU(const MatrixType& m) { return m; }
421 };
422 
423 } // end namespace internal
424 
425 /** Compute / recompute the LDLT decomposition A = L D L^* = U^* D U of \a matrix
426   */
427 template<typename MatrixType, int _UpLo>
428 LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
429 {
430   eigen_assert(a.rows()==a.cols());
431   const Index size = a.rows();
432 
433   m_matrix = a;
434 
435   m_transpositions.resize(size);
436   m_isInitialized = false;
437   m_temporary.resize(size);
438 
439   internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, &m_sign);
440 
441   m_isInitialized = true;
442   return *this;
443 }
444 
445 /** Update the LDLT decomposition:  given A = L D L^T, efficiently compute the decomposition of A + sigma w w^T.
446  * \param w a vector to be incorporated into the decomposition.
447  * \param sigma a scalar, +1 for updates and -1 for "downdates," which correspond to removing previously-added column vectors. Optional; default value is +1.
448  * \sa setZero()
449   */
450 template<typename MatrixType, int _UpLo>
451 template<typename Derived>
452 LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w,typename NumTraits<typename MatrixType::Scalar>::Real sigma)
453 {
454   const Index size = w.rows();
455   if (m_isInitialized)
456   {
457     eigen_assert(m_matrix.rows()==size);
458   }
459   else
460   {
461     m_matrix.resize(size,size);
462     m_matrix.setZero();
463     m_transpositions.resize(size);
464     for (Index i = 0; i < size; i++)
465       m_transpositions.coeffRef(i) = i;
466     m_temporary.resize(size);
467     m_sign = sigma>=0 ? 1 : -1;
468     m_isInitialized = true;
469   }
470 
471   internal::ldlt_inplace<UpLo>::update(m_matrix, m_transpositions, m_temporary, w, sigma);
472 
473   return *this;
474 }
475 
476 namespace internal {
477 template<typename _MatrixType, int _UpLo, typename Rhs>
478 struct solve_retval<LDLT<_MatrixType,_UpLo>, Rhs>
479   : solve_retval_base<LDLT<_MatrixType,_UpLo>, Rhs>
480 {
481   typedef LDLT<_MatrixType,_UpLo> LDLTType;
482   EIGEN_MAKE_SOLVE_HELPERS(LDLTType,Rhs)
483 
484   template<typename Dest> void evalTo(Dest& dst) const
485   {
486     eigen_assert(rhs().rows() == dec().matrixLDLT().rows());
487     // dst = P b
488     dst = dec().transpositionsP() * rhs();
489 
490     // dst = L^-1 (P b)
491     dec().matrixL().solveInPlace(dst);
492 
493     // dst = D^-1 (L^-1 P b)
494     // more precisely, use pseudo-inverse of D (see bug 241)
495     using std::abs;
496     using std::max;
497     typedef typename LDLTType::MatrixType MatrixType;
498     typedef typename LDLTType::Scalar Scalar;
499     typedef typename LDLTType::RealScalar RealScalar;
500     const Diagonal<const MatrixType> vectorD = dec().vectorD();
501     RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() * NumTraits<Scalar>::epsilon(),
502 				 RealScalar(1) / NumTraits<RealScalar>::highest()); // motivated by LAPACK's xGELSS
503     for (Index i = 0; i < vectorD.size(); ++i) {
504       if(abs(vectorD(i)) > tolerance)
505 	dst.row(i) /= vectorD(i);
506       else
507 	dst.row(i).setZero();
508     }
509 
510     // dst = L^-T (D^-1 L^-1 P b)
511     dec().matrixU().solveInPlace(dst);
512 
513     // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b
514     dst = dec().transpositionsP().transpose() * dst;
515   }
516 };
517 }
518 
519 /** \internal use x = ldlt_object.solve(x);
520   *
521   * This is the \em in-place version of solve().
522   *
523   * \param bAndX represents both the right-hand side matrix b and result x.
524   *
525   * \returns true always! If you need to check for existence of solutions, use another decomposition like LU, QR, or SVD.
526   *
527   * This version avoids a copy when the right hand side matrix b is not
528   * needed anymore.
529   *
530   * \sa LDLT::solve(), MatrixBase::ldlt()
531   */
532 template<typename MatrixType,int _UpLo>
533 template<typename Derived>
534 bool LDLT<MatrixType,_UpLo>::solveInPlace(MatrixBase<Derived> &bAndX) const
535 {
536   eigen_assert(m_isInitialized && "LDLT is not initialized.");
537   const Index size = m_matrix.rows();
538   eigen_assert(size == bAndX.rows());
539 
540   bAndX = this->solve(bAndX);
541 
542   return true;
543 }
544 
545 /** \returns the matrix represented by the decomposition,
546  * i.e., it returns the product: P^T L D L^* P.
547  * This function is provided for debug purpose. */
548 template<typename MatrixType, int _UpLo>
549 MatrixType LDLT<MatrixType,_UpLo>::reconstructedMatrix() const
550 {
551   eigen_assert(m_isInitialized && "LDLT is not initialized.");
552   const Index size = m_matrix.rows();
553   MatrixType res(size,size);
554 
555   // P
556   res.setIdentity();
557   res = transpositionsP() * res;
558   // L^* P
559   res = matrixU() * res;
560   // D(L^*P)
561   res = vectorD().asDiagonal() * res;
562   // L(DL^*P)
563   res = matrixL() * res;
564   // P^T (LDL^*P)
565   res = transpositionsP().transpose() * res;
566 
567   return res;
568 }
569 
570 /** \cholesky_module
571   * \returns the Cholesky decomposition with full pivoting without square root of \c *this
572   */
573 template<typename MatrixType, unsigned int UpLo>
574 inline const LDLT<typename SelfAdjointView<MatrixType, UpLo>::PlainObject, UpLo>
575 SelfAdjointView<MatrixType, UpLo>::ldlt() const
576 {
577   return LDLT<PlainObject,UpLo>(m_matrix);
578 }
579 
580 /** \cholesky_module
581   * \returns the Cholesky decomposition with full pivoting without square root of \c *this
582   */
583 template<typename Derived>
584 inline const LDLT<typename MatrixBase<Derived>::PlainObject>
585 MatrixBase<Derived>::ldlt() const
586 {
587   return LDLT<PlainObject>(derived());
588 }
589 
590 } // end namespace Eigen
591 
592 #endif // EIGEN_LDLT_H
593