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 (at) inria.fr> 5 // Copyright (C) 2010,2012 Jitse Niesen <jitse (at) maths.leeds.ac.uk> 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_REAL_SCHUR_H 12 #define EIGEN_REAL_SCHUR_H 13 14 #include "./HessenbergDecomposition.h" 15 16 namespace Eigen { 17 18 /** \eigenvalues_module \ingroup Eigenvalues_Module 19 * 20 * 21 * \class RealSchur 22 * 23 * \brief Performs a real Schur decomposition of a square matrix 24 * 25 * \tparam _MatrixType the type of the matrix of which we are computing the 26 * real Schur decomposition; this is expected to be an instantiation of the 27 * Matrix class template. 28 * 29 * Given a real square matrix A, this class computes the real Schur 30 * decomposition: \f$ A = U T U^T \f$ where U is a real orthogonal matrix and 31 * T is a real quasi-triangular matrix. An orthogonal matrix is a matrix whose 32 * inverse is equal to its transpose, \f$ U^{-1} = U^T \f$. A quasi-triangular 33 * matrix is a block-triangular matrix whose diagonal consists of 1-by-1 34 * blocks and 2-by-2 blocks with complex eigenvalues. The eigenvalues of the 35 * blocks on the diagonal of T are the same as the eigenvalues of the matrix 36 * A, and thus the real Schur decomposition is used in EigenSolver to compute 37 * the eigendecomposition of a matrix. 38 * 39 * Call the function compute() to compute the real Schur decomposition of a 40 * given matrix. Alternatively, you can use the RealSchur(const MatrixType&, bool) 41 * constructor which computes the real Schur decomposition at construction 42 * time. Once the decomposition is computed, you can use the matrixU() and 43 * matrixT() functions to retrieve the matrices U and T in the decomposition. 44 * 45 * The documentation of RealSchur(const MatrixType&, bool) contains an example 46 * of the typical use of this class. 47 * 48 * \note The implementation is adapted from 49 * <a href="http://math.nist.gov/javanumerics/jama/">JAMA</a> (public domain). 50 * Their code is based on EISPACK. 51 * 52 * \sa class ComplexSchur, class EigenSolver, class ComplexEigenSolver 53 */ 54 template<typename _MatrixType> class RealSchur 55 { 56 public: 57 typedef _MatrixType MatrixType; 58 enum { 59 RowsAtCompileTime = MatrixType::RowsAtCompileTime, 60 ColsAtCompileTime = MatrixType::ColsAtCompileTime, 61 Options = MatrixType::Options, 62 MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, 63 MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime 64 }; 65 typedef typename MatrixType::Scalar Scalar; 66 typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar; 67 typedef typename MatrixType::Index Index; 68 69 typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> EigenvalueType; 70 typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ColumnVectorType; 71 72 /** \brief Default constructor. 73 * 74 * \param [in] size Positive integer, size of the matrix whose Schur decomposition will be computed. 75 * 76 * The default constructor is useful in cases in which the user intends to 77 * perform decompositions via compute(). The \p size parameter is only 78 * used as a hint. It is not an error to give a wrong \p size, but it may 79 * impair performance. 80 * 81 * \sa compute() for an example. 82 */ 83 RealSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) 84 : m_matT(size, size), 85 m_matU(size, size), 86 m_workspaceVector(size), 87 m_hess(size), 88 m_isInitialized(false), 89 m_matUisUptodate(false), 90 m_maxIters(-1) 91 { } 92 93 /** \brief Constructor; computes real Schur decomposition of given matrix. 94 * 95 * \param[in] matrix Square matrix whose Schur decomposition is to be computed. 96 * \param[in] computeU If true, both T and U are computed; if false, only T is computed. 97 * 98 * This constructor calls compute() to compute the Schur decomposition. 99 * 100 * Example: \include RealSchur_RealSchur_MatrixType.cpp 101 * Output: \verbinclude RealSchur_RealSchur_MatrixType.out 102 */ 103 RealSchur(const MatrixType& matrix, bool computeU = true) 104 : m_matT(matrix.rows(),matrix.cols()), 105 m_matU(matrix.rows(),matrix.cols()), 106 m_workspaceVector(matrix.rows()), 107 m_hess(matrix.rows()), 108 m_isInitialized(false), 109 m_matUisUptodate(false), 110 m_maxIters(-1) 111 { 112 compute(matrix, computeU); 113 } 114 115 /** \brief Returns the orthogonal matrix in the Schur decomposition. 116 * 117 * \returns A const reference to the matrix U. 118 * 119 * \pre Either the constructor RealSchur(const MatrixType&, bool) or the 120 * member function compute(const MatrixType&, bool) has been called before 121 * to compute the Schur decomposition of a matrix, and \p computeU was set 122 * to true (the default value). 123 * 124 * \sa RealSchur(const MatrixType&, bool) for an example 125 */ 126 const MatrixType& matrixU() const 127 { 128 eigen_assert(m_isInitialized && "RealSchur is not initialized."); 129 eigen_assert(m_matUisUptodate && "The matrix U has not been computed during the RealSchur decomposition."); 130 return m_matU; 131 } 132 133 /** \brief Returns the quasi-triangular matrix in the Schur decomposition. 134 * 135 * \returns A const reference to the matrix T. 136 * 137 * \pre Either the constructor RealSchur(const MatrixType&, bool) or the 138 * member function compute(const MatrixType&, bool) has been called before 139 * to compute the Schur decomposition of a matrix. 140 * 141 * \sa RealSchur(const MatrixType&, bool) for an example 142 */ 143 const MatrixType& matrixT() const 144 { 145 eigen_assert(m_isInitialized && "RealSchur is not initialized."); 146 return m_matT; 147 } 148 149 /** \brief Computes Schur decomposition of given matrix. 150 * 151 * \param[in] matrix Square matrix whose Schur decomposition is to be computed. 152 * \param[in] computeU If true, both T and U are computed; if false, only T is computed. 153 * \returns Reference to \c *this 154 * 155 * The Schur decomposition is computed by first reducing the matrix to 156 * Hessenberg form using the class HessenbergDecomposition. The Hessenberg 157 * matrix is then reduced to triangular form by performing Francis QR 158 * iterations with implicit double shift. The cost of computing the Schur 159 * decomposition depends on the number of iterations; as a rough guide, it 160 * may be taken to be \f$25n^3\f$ flops if \a computeU is true and 161 * \f$10n^3\f$ flops if \a computeU is false. 162 * 163 * Example: \include RealSchur_compute.cpp 164 * Output: \verbinclude RealSchur_compute.out 165 * 166 * \sa compute(const MatrixType&, bool, Index) 167 */ 168 RealSchur& compute(const MatrixType& matrix, bool computeU = true); 169 170 /** \brief Computes Schur decomposition of a Hessenberg matrix H = Z T Z^T 171 * \param[in] matrixH Matrix in Hessenberg form H 172 * \param[in] matrixQ orthogonal matrix Q that transform a matrix A to H : A = Q H Q^T 173 * \param computeU Computes the matriX U of the Schur vectors 174 * \return Reference to \c *this 175 * 176 * This routine assumes that the matrix is already reduced in Hessenberg form matrixH 177 * using either the class HessenbergDecomposition or another mean. 178 * It computes the upper quasi-triangular matrix T of the Schur decomposition of H 179 * When computeU is true, this routine computes the matrix U such that 180 * A = U T U^T = (QZ) T (QZ)^T = Q H Q^T where A is the initial matrix 181 * 182 * NOTE Q is referenced if computeU is true; so, if the initial orthogonal matrix 183 * is not available, the user should give an identity matrix (Q.setIdentity()) 184 * 185 * \sa compute(const MatrixType&, bool) 186 */ 187 template<typename HessMatrixType, typename OrthMatrixType> 188 RealSchur& computeFromHessenberg(const HessMatrixType& matrixH, const OrthMatrixType& matrixQ, bool computeU); 189 /** \brief Reports whether previous computation was successful. 190 * 191 * \returns \c Success if computation was succesful, \c NoConvergence otherwise. 192 */ 193 ComputationInfo info() const 194 { 195 eigen_assert(m_isInitialized && "RealSchur is not initialized."); 196 return m_info; 197 } 198 199 /** \brief Sets the maximum number of iterations allowed. 200 * 201 * If not specified by the user, the maximum number of iterations is m_maxIterationsPerRow times the size 202 * of the matrix. 203 */ 204 RealSchur& setMaxIterations(Index maxIters) 205 { 206 m_maxIters = maxIters; 207 return *this; 208 } 209 210 /** \brief Returns the maximum number of iterations. */ 211 Index getMaxIterations() 212 { 213 return m_maxIters; 214 } 215 216 /** \brief Maximum number of iterations per row. 217 * 218 * If not otherwise specified, the maximum number of iterations is this number times the size of the 219 * matrix. It is currently set to 40. 220 */ 221 static const int m_maxIterationsPerRow = 40; 222 223 private: 224 225 MatrixType m_matT; 226 MatrixType m_matU; 227 ColumnVectorType m_workspaceVector; 228 HessenbergDecomposition<MatrixType> m_hess; 229 ComputationInfo m_info; 230 bool m_isInitialized; 231 bool m_matUisUptodate; 232 Index m_maxIters; 233 234 typedef Matrix<Scalar,3,1> Vector3s; 235 236 Scalar computeNormOfT(); 237 Index findSmallSubdiagEntry(Index iu, const Scalar& norm); 238 void splitOffTwoRows(Index iu, bool computeU, const Scalar& exshift); 239 void computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo); 240 void initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector); 241 void performFrancisQRStep(Index il, Index im, Index iu, bool computeU, const Vector3s& firstHouseholderVector, Scalar* workspace); 242 }; 243 244 245 template<typename MatrixType> 246 RealSchur<MatrixType>& RealSchur<MatrixType>::compute(const MatrixType& matrix, bool computeU) 247 { 248 eigen_assert(matrix.cols() == matrix.rows()); 249 Index maxIters = m_maxIters; 250 if (maxIters == -1) 251 maxIters = m_maxIterationsPerRow * matrix.rows(); 252 253 // Step 1. Reduce to Hessenberg form 254 m_hess.compute(matrix); 255 256 // Step 2. Reduce to real Schur form 257 computeFromHessenberg(m_hess.matrixH(), m_hess.matrixQ(), computeU); 258 259 return *this; 260 } 261 template<typename MatrixType> 262 template<typename HessMatrixType, typename OrthMatrixType> 263 RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMatrixType& matrixH, const OrthMatrixType& matrixQ, bool computeU) 264 { 265 m_matT = matrixH; 266 if(computeU) 267 m_matU = matrixQ; 268 269 Index maxIters = m_maxIters; 270 if (maxIters == -1) 271 maxIters = m_maxIterationsPerRow * matrixH.rows(); 272 m_workspaceVector.resize(m_matT.cols()); 273 Scalar* workspace = &m_workspaceVector.coeffRef(0); 274 275 // The matrix m_matT is divided in three parts. 276 // Rows 0,...,il-1 are decoupled from the rest because m_matT(il,il-1) is zero. 277 // Rows il,...,iu is the part we are working on (the active window). 278 // Rows iu+1,...,end are already brought in triangular form. 279 Index iu = m_matT.cols() - 1; 280 Index iter = 0; // iteration count for current eigenvalue 281 Index totalIter = 0; // iteration count for whole matrix 282 Scalar exshift(0); // sum of exceptional shifts 283 Scalar norm = computeNormOfT(); 284 285 if(norm!=0) 286 { 287 while (iu >= 0) 288 { 289 Index il = findSmallSubdiagEntry(iu, norm); 290 291 // Check for convergence 292 if (il == iu) // One root found 293 { 294 m_matT.coeffRef(iu,iu) = m_matT.coeff(iu,iu) + exshift; 295 if (iu > 0) 296 m_matT.coeffRef(iu, iu-1) = Scalar(0); 297 iu--; 298 iter = 0; 299 } 300 else if (il == iu-1) // Two roots found 301 { 302 splitOffTwoRows(iu, computeU, exshift); 303 iu -= 2; 304 iter = 0; 305 } 306 else // No convergence yet 307 { 308 // The firstHouseholderVector vector has to be initialized to something to get rid of a silly GCC warning (-O1 -Wall -DNDEBUG ) 309 Vector3s firstHouseholderVector(0,0,0), shiftInfo; 310 computeShift(iu, iter, exshift, shiftInfo); 311 iter = iter + 1; 312 totalIter = totalIter + 1; 313 if (totalIter > maxIters) break; 314 Index im; 315 initFrancisQRStep(il, iu, shiftInfo, im, firstHouseholderVector); 316 performFrancisQRStep(il, im, iu, computeU, firstHouseholderVector, workspace); 317 } 318 } 319 } 320 if(totalIter <= maxIters) 321 m_info = Success; 322 else 323 m_info = NoConvergence; 324 325 m_isInitialized = true; 326 m_matUisUptodate = computeU; 327 return *this; 328 } 329 330 /** \internal Computes and returns vector L1 norm of T */ 331 template<typename MatrixType> 332 inline typename MatrixType::Scalar RealSchur<MatrixType>::computeNormOfT() 333 { 334 const Index size = m_matT.cols(); 335 // FIXME to be efficient the following would requires a triangular reduxion code 336 // Scalar norm = m_matT.upper().cwiseAbs().sum() 337 // + m_matT.bottomLeftCorner(size-1,size-1).diagonal().cwiseAbs().sum(); 338 Scalar norm(0); 339 for (Index j = 0; j < size; ++j) 340 norm += m_matT.col(j).segment(0, (std::min)(size,j+2)).cwiseAbs().sum(); 341 return norm; 342 } 343 344 /** \internal Look for single small sub-diagonal element and returns its index */ 345 template<typename MatrixType> 346 inline typename MatrixType::Index RealSchur<MatrixType>::findSmallSubdiagEntry(Index iu, const Scalar& norm) 347 { 348 using std::abs; 349 Index res = iu; 350 while (res > 0) 351 { 352 Scalar s = abs(m_matT.coeff(res-1,res-1)) + abs(m_matT.coeff(res,res)); 353 if (s == 0.0) 354 s = norm; 355 if (abs(m_matT.coeff(res,res-1)) < NumTraits<Scalar>::epsilon() * s) 356 break; 357 res--; 358 } 359 return res; 360 } 361 362 /** \internal Update T given that rows iu-1 and iu decouple from the rest. */ 363 template<typename MatrixType> 364 inline void RealSchur<MatrixType>::splitOffTwoRows(Index iu, bool computeU, const Scalar& exshift) 365 { 366 using std::sqrt; 367 using std::abs; 368 const Index size = m_matT.cols(); 369 370 // The eigenvalues of the 2x2 matrix [a b; c d] are 371 // trace +/- sqrt(discr/4) where discr = tr^2 - 4*det, tr = a + d, det = ad - bc 372 Scalar p = Scalar(0.5) * (m_matT.coeff(iu-1,iu-1) - m_matT.coeff(iu,iu)); 373 Scalar q = p * p + m_matT.coeff(iu,iu-1) * m_matT.coeff(iu-1,iu); // q = tr^2 / 4 - det = discr/4 374 m_matT.coeffRef(iu,iu) += exshift; 375 m_matT.coeffRef(iu-1,iu-1) += exshift; 376 377 if (q >= Scalar(0)) // Two real eigenvalues 378 { 379 Scalar z = sqrt(abs(q)); 380 JacobiRotation<Scalar> rot; 381 if (p >= Scalar(0)) 382 rot.makeGivens(p + z, m_matT.coeff(iu, iu-1)); 383 else 384 rot.makeGivens(p - z, m_matT.coeff(iu, iu-1)); 385 386 m_matT.rightCols(size-iu+1).applyOnTheLeft(iu-1, iu, rot.adjoint()); 387 m_matT.topRows(iu+1).applyOnTheRight(iu-1, iu, rot); 388 m_matT.coeffRef(iu, iu-1) = Scalar(0); 389 if (computeU) 390 m_matU.applyOnTheRight(iu-1, iu, rot); 391 } 392 393 if (iu > 1) 394 m_matT.coeffRef(iu-1, iu-2) = Scalar(0); 395 } 396 397 /** \internal Form shift in shiftInfo, and update exshift if an exceptional shift is performed. */ 398 template<typename MatrixType> 399 inline void RealSchur<MatrixType>::computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo) 400 { 401 using std::sqrt; 402 using std::abs; 403 shiftInfo.coeffRef(0) = m_matT.coeff(iu,iu); 404 shiftInfo.coeffRef(1) = m_matT.coeff(iu-1,iu-1); 405 shiftInfo.coeffRef(2) = m_matT.coeff(iu,iu-1) * m_matT.coeff(iu-1,iu); 406 407 // Wilkinson's original ad hoc shift 408 if (iter == 10) 409 { 410 exshift += shiftInfo.coeff(0); 411 for (Index i = 0; i <= iu; ++i) 412 m_matT.coeffRef(i,i) -= shiftInfo.coeff(0); 413 Scalar s = abs(m_matT.coeff(iu,iu-1)) + abs(m_matT.coeff(iu-1,iu-2)); 414 shiftInfo.coeffRef(0) = Scalar(0.75) * s; 415 shiftInfo.coeffRef(1) = Scalar(0.75) * s; 416 shiftInfo.coeffRef(2) = Scalar(-0.4375) * s * s; 417 } 418 419 // MATLAB's new ad hoc shift 420 if (iter == 30) 421 { 422 Scalar s = (shiftInfo.coeff(1) - shiftInfo.coeff(0)) / Scalar(2.0); 423 s = s * s + shiftInfo.coeff(2); 424 if (s > Scalar(0)) 425 { 426 s = sqrt(s); 427 if (shiftInfo.coeff(1) < shiftInfo.coeff(0)) 428 s = -s; 429 s = s + (shiftInfo.coeff(1) - shiftInfo.coeff(0)) / Scalar(2.0); 430 s = shiftInfo.coeff(0) - shiftInfo.coeff(2) / s; 431 exshift += s; 432 for (Index i = 0; i <= iu; ++i) 433 m_matT.coeffRef(i,i) -= s; 434 shiftInfo.setConstant(Scalar(0.964)); 435 } 436 } 437 } 438 439 /** \internal Compute index im at which Francis QR step starts and the first Householder vector. */ 440 template<typename MatrixType> 441 inline void RealSchur<MatrixType>::initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector) 442 { 443 using std::abs; 444 Vector3s& v = firstHouseholderVector; // alias to save typing 445 446 for (im = iu-2; im >= il; --im) 447 { 448 const Scalar Tmm = m_matT.coeff(im,im); 449 const Scalar r = shiftInfo.coeff(0) - Tmm; 450 const Scalar s = shiftInfo.coeff(1) - Tmm; 451 v.coeffRef(0) = (r * s - shiftInfo.coeff(2)) / m_matT.coeff(im+1,im) + m_matT.coeff(im,im+1); 452 v.coeffRef(1) = m_matT.coeff(im+1,im+1) - Tmm - r - s; 453 v.coeffRef(2) = m_matT.coeff(im+2,im+1); 454 if (im == il) { 455 break; 456 } 457 const Scalar lhs = m_matT.coeff(im,im-1) * (abs(v.coeff(1)) + abs(v.coeff(2))); 458 const Scalar rhs = v.coeff(0) * (abs(m_matT.coeff(im-1,im-1)) + abs(Tmm) + abs(m_matT.coeff(im+1,im+1))); 459 if (abs(lhs) < NumTraits<Scalar>::epsilon() * rhs) 460 { 461 break; 462 } 463 } 464 } 465 466 /** \internal Perform a Francis QR step involving rows il:iu and columns im:iu. */ 467 template<typename MatrixType> 468 inline void RealSchur<MatrixType>::performFrancisQRStep(Index il, Index im, Index iu, bool computeU, const Vector3s& firstHouseholderVector, Scalar* workspace) 469 { 470 eigen_assert(im >= il); 471 eigen_assert(im <= iu-2); 472 473 const Index size = m_matT.cols(); 474 475 for (Index k = im; k <= iu-2; ++k) 476 { 477 bool firstIteration = (k == im); 478 479 Vector3s v; 480 if (firstIteration) 481 v = firstHouseholderVector; 482 else 483 v = m_matT.template block<3,1>(k,k-1); 484 485 Scalar tau, beta; 486 Matrix<Scalar, 2, 1> ess; 487 v.makeHouseholder(ess, tau, beta); 488 489 if (beta != Scalar(0)) // if v is not zero 490 { 491 if (firstIteration && k > il) 492 m_matT.coeffRef(k,k-1) = -m_matT.coeff(k,k-1); 493 else if (!firstIteration) 494 m_matT.coeffRef(k,k-1) = beta; 495 496 // These Householder transformations form the O(n^3) part of the algorithm 497 m_matT.block(k, k, 3, size-k).applyHouseholderOnTheLeft(ess, tau, workspace); 498 m_matT.block(0, k, (std::min)(iu,k+3) + 1, 3).applyHouseholderOnTheRight(ess, tau, workspace); 499 if (computeU) 500 m_matU.block(0, k, size, 3).applyHouseholderOnTheRight(ess, tau, workspace); 501 } 502 } 503 504 Matrix<Scalar, 2, 1> v = m_matT.template block<2,1>(iu-1, iu-2); 505 Scalar tau, beta; 506 Matrix<Scalar, 1, 1> ess; 507 v.makeHouseholder(ess, tau, beta); 508 509 if (beta != Scalar(0)) // if v is not zero 510 { 511 m_matT.coeffRef(iu-1, iu-2) = beta; 512 m_matT.block(iu-1, iu-1, 2, size-iu+1).applyHouseholderOnTheLeft(ess, tau, workspace); 513 m_matT.block(0, iu-1, iu+1, 2).applyHouseholderOnTheRight(ess, tau, workspace); 514 if (computeU) 515 m_matU.block(0, iu-1, size, 2).applyHouseholderOnTheRight(ess, tau, workspace); 516 } 517 518 // clean up pollution due to round-off errors 519 for (Index i = im+2; i <= iu; ++i) 520 { 521 m_matT.coeffRef(i,i-2) = Scalar(0); 522 if (i > im+2) 523 m_matT.coeffRef(i,i-3) = Scalar(0); 524 } 525 } 526 527 } // end namespace Eigen 528 529 #endif // EIGEN_REAL_SCHUR_H 530