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