1*bf2c3715SXin Li // This file is part of Eigen, a lightweight C++ template library 2*bf2c3715SXin Li // for linear algebra. 3*bf2c3715SXin Li // 4*bf2c3715SXin Li // Copyright (C) 2008-2010 Gael Guennebaud <[email protected]> 5*bf2c3715SXin Li // Copyright (C) 2010 Jitse Niesen <[email protected]> 6*bf2c3715SXin Li // 7*bf2c3715SXin Li // This Source Code Form is subject to the terms of the Mozilla 8*bf2c3715SXin Li // Public License v. 2.0. If a copy of the MPL was not distributed 9*bf2c3715SXin Li // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 10*bf2c3715SXin Li 11*bf2c3715SXin Li #ifndef EIGEN_SELFADJOINTEIGENSOLVER_H 12*bf2c3715SXin Li #define EIGEN_SELFADJOINTEIGENSOLVER_H 13*bf2c3715SXin Li 14*bf2c3715SXin Li #include "./Tridiagonalization.h" 15*bf2c3715SXin Li 16*bf2c3715SXin Li namespace Eigen { 17*bf2c3715SXin Li 18*bf2c3715SXin Li template<typename _MatrixType> 19*bf2c3715SXin Li class GeneralizedSelfAdjointEigenSolver; 20*bf2c3715SXin Li 21*bf2c3715SXin Li namespace internal { 22*bf2c3715SXin Li template<typename SolverType,int Size,bool IsComplex> struct direct_selfadjoint_eigenvalues; 23*bf2c3715SXin Li 24*bf2c3715SXin Li template<typename MatrixType, typename DiagType, typename SubDiagType> 25*bf2c3715SXin Li EIGEN_DEVICE_FUNC 26*bf2c3715SXin Li ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec); 27*bf2c3715SXin Li } 28*bf2c3715SXin Li 29*bf2c3715SXin Li /** \eigenvalues_module \ingroup Eigenvalues_Module 30*bf2c3715SXin Li * 31*bf2c3715SXin Li * 32*bf2c3715SXin Li * \class SelfAdjointEigenSolver 33*bf2c3715SXin Li * 34*bf2c3715SXin Li * \brief Computes eigenvalues and eigenvectors of selfadjoint matrices 35*bf2c3715SXin Li * 36*bf2c3715SXin Li * \tparam _MatrixType the type of the matrix of which we are computing the 37*bf2c3715SXin Li * eigendecomposition; this is expected to be an instantiation of the Matrix 38*bf2c3715SXin Li * class template. 39*bf2c3715SXin Li * 40*bf2c3715SXin Li * A matrix \f$ A \f$ is selfadjoint if it equals its adjoint. For real 41*bf2c3715SXin Li * matrices, this means that the matrix is symmetric: it equals its 42*bf2c3715SXin Li * transpose. This class computes the eigenvalues and eigenvectors of a 43*bf2c3715SXin Li * selfadjoint matrix. These are the scalars \f$ \lambda \f$ and vectors 44*bf2c3715SXin Li * \f$ v \f$ such that \f$ Av = \lambda v \f$. The eigenvalues of a 45*bf2c3715SXin Li * selfadjoint matrix are always real. If \f$ D \f$ is a diagonal matrix with 46*bf2c3715SXin Li * the eigenvalues on the diagonal, and \f$ V \f$ is a matrix with the 47*bf2c3715SXin Li * eigenvectors as its columns, then \f$ A = V D V^{-1} \f$. This is called the 48*bf2c3715SXin Li * eigendecomposition. 49*bf2c3715SXin Li * 50*bf2c3715SXin Li * For a selfadjoint matrix, \f$ V \f$ is unitary, meaning its inverse is equal 51*bf2c3715SXin Li * to its adjoint, \f$ V^{-1} = V^{\dagger} \f$. If \f$ A \f$ is real, then 52*bf2c3715SXin Li * \f$ V \f$ is also real and therefore orthogonal, meaning its inverse is 53*bf2c3715SXin Li * equal to its transpose, \f$ V^{-1} = V^T \f$. 54*bf2c3715SXin Li * 55*bf2c3715SXin Li * The algorithm exploits the fact that the matrix is selfadjoint, making it 56*bf2c3715SXin Li * faster and more accurate than the general purpose eigenvalue algorithms 57*bf2c3715SXin Li * implemented in EigenSolver and ComplexEigenSolver. 58*bf2c3715SXin Li * 59*bf2c3715SXin Li * Only the \b lower \b triangular \b part of the input matrix is referenced. 60*bf2c3715SXin Li * 61*bf2c3715SXin Li * Call the function compute() to compute the eigenvalues and eigenvectors of 62*bf2c3715SXin Li * a given matrix. Alternatively, you can use the 63*bf2c3715SXin Li * SelfAdjointEigenSolver(const MatrixType&, int) constructor which computes 64*bf2c3715SXin Li * the eigenvalues and eigenvectors at construction time. Once the eigenvalue 65*bf2c3715SXin Li * and eigenvectors are computed, they can be retrieved with the eigenvalues() 66*bf2c3715SXin Li * and eigenvectors() functions. 67*bf2c3715SXin Li * 68*bf2c3715SXin Li * The documentation for SelfAdjointEigenSolver(const MatrixType&, int) 69*bf2c3715SXin Li * contains an example of the typical use of this class. 70*bf2c3715SXin Li * 71*bf2c3715SXin Li * To solve the \em generalized eigenvalue problem \f$ Av = \lambda Bv \f$ and 72*bf2c3715SXin Li * the likes, see the class GeneralizedSelfAdjointEigenSolver. 73*bf2c3715SXin Li * 74*bf2c3715SXin Li * \sa MatrixBase::eigenvalues(), class EigenSolver, class ComplexEigenSolver 75*bf2c3715SXin Li */ 76*bf2c3715SXin Li template<typename _MatrixType> class SelfAdjointEigenSolver 77*bf2c3715SXin Li { 78*bf2c3715SXin Li public: 79*bf2c3715SXin Li 80*bf2c3715SXin Li typedef _MatrixType MatrixType; 81*bf2c3715SXin Li enum { 82*bf2c3715SXin Li Size = MatrixType::RowsAtCompileTime, 83*bf2c3715SXin Li ColsAtCompileTime = MatrixType::ColsAtCompileTime, 84*bf2c3715SXin Li Options = MatrixType::Options, 85*bf2c3715SXin Li MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime 86*bf2c3715SXin Li }; 87*bf2c3715SXin Li 88*bf2c3715SXin Li /** \brief Scalar type for matrices of type \p _MatrixType. */ 89*bf2c3715SXin Li typedef typename MatrixType::Scalar Scalar; 90*bf2c3715SXin Li typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 91*bf2c3715SXin Li 92*bf2c3715SXin Li typedef Matrix<Scalar,Size,Size,ColMajor,MaxColsAtCompileTime,MaxColsAtCompileTime> EigenvectorsType; 93*bf2c3715SXin Li 94*bf2c3715SXin Li /** \brief Real scalar type for \p _MatrixType. 95*bf2c3715SXin Li * 96*bf2c3715SXin Li * This is just \c Scalar if #Scalar is real (e.g., \c float or 97*bf2c3715SXin Li * \c double), and the type of the real part of \c Scalar if #Scalar is 98*bf2c3715SXin Li * complex. 99*bf2c3715SXin Li */ 100*bf2c3715SXin Li typedef typename NumTraits<Scalar>::Real RealScalar; 101*bf2c3715SXin Li 102*bf2c3715SXin Li friend struct internal::direct_selfadjoint_eigenvalues<SelfAdjointEigenSolver,Size,NumTraits<Scalar>::IsComplex>; 103*bf2c3715SXin Li 104*bf2c3715SXin Li /** \brief Type for vector of eigenvalues as returned by eigenvalues(). 105*bf2c3715SXin Li * 106*bf2c3715SXin Li * This is a column vector with entries of type #RealScalar. 107*bf2c3715SXin Li * The length of the vector is the size of \p _MatrixType. 108*bf2c3715SXin Li */ 109*bf2c3715SXin Li typedef typename internal::plain_col_type<MatrixType, RealScalar>::type RealVectorType; 110*bf2c3715SXin Li typedef Tridiagonalization<MatrixType> TridiagonalizationType; 111*bf2c3715SXin Li typedef typename TridiagonalizationType::SubDiagonalType SubDiagonalType; 112*bf2c3715SXin Li 113*bf2c3715SXin Li /** \brief Default constructor for fixed-size matrices. 114*bf2c3715SXin Li * 115*bf2c3715SXin Li * The default constructor is useful in cases in which the user intends to 116*bf2c3715SXin Li * perform decompositions via compute(). This constructor 117*bf2c3715SXin Li * can only be used if \p _MatrixType is a fixed-size matrix; use 118*bf2c3715SXin Li * SelfAdjointEigenSolver(Index) for dynamic-size matrices. 119*bf2c3715SXin Li * 120*bf2c3715SXin Li * Example: \include SelfAdjointEigenSolver_SelfAdjointEigenSolver.cpp 121*bf2c3715SXin Li * Output: \verbinclude SelfAdjointEigenSolver_SelfAdjointEigenSolver.out 122*bf2c3715SXin Li */ 123*bf2c3715SXin Li EIGEN_DEVICE_FUNC 124*bf2c3715SXin Li SelfAdjointEigenSolver() 125*bf2c3715SXin Li : m_eivec(), 126*bf2c3715SXin Li m_eivalues(), 127*bf2c3715SXin Li m_subdiag(), 128*bf2c3715SXin Li m_hcoeffs(), 129*bf2c3715SXin Li m_info(InvalidInput), 130*bf2c3715SXin Li m_isInitialized(false), 131*bf2c3715SXin Li m_eigenvectorsOk(false) 132*bf2c3715SXin Li { } 133*bf2c3715SXin Li 134*bf2c3715SXin Li /** \brief Constructor, pre-allocates memory for dynamic-size matrices. 135*bf2c3715SXin Li * 136*bf2c3715SXin Li * \param [in] size Positive integer, size of the matrix whose 137*bf2c3715SXin Li * eigenvalues and eigenvectors will be computed. 138*bf2c3715SXin Li * 139*bf2c3715SXin Li * This constructor is useful for dynamic-size matrices, when the user 140*bf2c3715SXin Li * intends to perform decompositions via compute(). The \p size 141*bf2c3715SXin Li * parameter is only used as a hint. It is not an error to give a wrong 142*bf2c3715SXin Li * \p size, but it may impair performance. 143*bf2c3715SXin Li * 144*bf2c3715SXin Li * \sa compute() for an example 145*bf2c3715SXin Li */ 146*bf2c3715SXin Li EIGEN_DEVICE_FUNC 147*bf2c3715SXin Li explicit SelfAdjointEigenSolver(Index size) 148*bf2c3715SXin Li : m_eivec(size, size), 149*bf2c3715SXin Li m_eivalues(size), 150*bf2c3715SXin Li m_subdiag(size > 1 ? size - 1 : 1), 151*bf2c3715SXin Li m_hcoeffs(size > 1 ? size - 1 : 1), 152*bf2c3715SXin Li m_isInitialized(false), 153*bf2c3715SXin Li m_eigenvectorsOk(false) 154*bf2c3715SXin Li {} 155*bf2c3715SXin Li 156*bf2c3715SXin Li /** \brief Constructor; computes eigendecomposition of given matrix. 157*bf2c3715SXin Li * 158*bf2c3715SXin Li * \param[in] matrix Selfadjoint matrix whose eigendecomposition is to 159*bf2c3715SXin Li * be computed. Only the lower triangular part of the matrix is referenced. 160*bf2c3715SXin Li * \param[in] options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly. 161*bf2c3715SXin Li * 162*bf2c3715SXin Li * This constructor calls compute(const MatrixType&, int) to compute the 163*bf2c3715SXin Li * eigenvalues of the matrix \p matrix. The eigenvectors are computed if 164*bf2c3715SXin Li * \p options equals #ComputeEigenvectors. 165*bf2c3715SXin Li * 166*bf2c3715SXin Li * Example: \include SelfAdjointEigenSolver_SelfAdjointEigenSolver_MatrixType.cpp 167*bf2c3715SXin Li * Output: \verbinclude SelfAdjointEigenSolver_SelfAdjointEigenSolver_MatrixType.out 168*bf2c3715SXin Li * 169*bf2c3715SXin Li * \sa compute(const MatrixType&, int) 170*bf2c3715SXin Li */ 171*bf2c3715SXin Li template<typename InputType> 172*bf2c3715SXin Li EIGEN_DEVICE_FUNC 173*bf2c3715SXin Li explicit SelfAdjointEigenSolver(const EigenBase<InputType>& matrix, int options = ComputeEigenvectors) 174*bf2c3715SXin Li : m_eivec(matrix.rows(), matrix.cols()), 175*bf2c3715SXin Li m_eivalues(matrix.cols()), 176*bf2c3715SXin Li m_subdiag(matrix.rows() > 1 ? matrix.rows() - 1 : 1), 177*bf2c3715SXin Li m_hcoeffs(matrix.cols() > 1 ? matrix.cols() - 1 : 1), 178*bf2c3715SXin Li m_isInitialized(false), 179*bf2c3715SXin Li m_eigenvectorsOk(false) 180*bf2c3715SXin Li { 181*bf2c3715SXin Li compute(matrix.derived(), options); 182*bf2c3715SXin Li } 183*bf2c3715SXin Li 184*bf2c3715SXin Li /** \brief Computes eigendecomposition of given matrix. 185*bf2c3715SXin Li * 186*bf2c3715SXin Li * \param[in] matrix Selfadjoint matrix whose eigendecomposition is to 187*bf2c3715SXin Li * be computed. Only the lower triangular part of the matrix is referenced. 188*bf2c3715SXin Li * \param[in] options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly. 189*bf2c3715SXin Li * \returns Reference to \c *this 190*bf2c3715SXin Li * 191*bf2c3715SXin Li * This function computes the eigenvalues of \p matrix. The eigenvalues() 192*bf2c3715SXin Li * function can be used to retrieve them. If \p options equals #ComputeEigenvectors, 193*bf2c3715SXin Li * then the eigenvectors are also computed and can be retrieved by 194*bf2c3715SXin Li * calling eigenvectors(). 195*bf2c3715SXin Li * 196*bf2c3715SXin Li * This implementation uses a symmetric QR algorithm. The matrix is first 197*bf2c3715SXin Li * reduced to tridiagonal form using the Tridiagonalization class. The 198*bf2c3715SXin Li * tridiagonal matrix is then brought to diagonal form with implicit 199*bf2c3715SXin Li * symmetric QR steps with Wilkinson shift. Details can be found in 200*bf2c3715SXin Li * Section 8.3 of Golub \& Van Loan, <i>%Matrix Computations</i>. 201*bf2c3715SXin Li * 202*bf2c3715SXin Li * The cost of the computation is about \f$ 9n^3 \f$ if the eigenvectors 203*bf2c3715SXin Li * are required and \f$ 4n^3/3 \f$ if they are not required. 204*bf2c3715SXin Li * 205*bf2c3715SXin Li * This method reuses the memory in the SelfAdjointEigenSolver object that 206*bf2c3715SXin Li * was allocated when the object was constructed, if the size of the 207*bf2c3715SXin Li * matrix does not change. 208*bf2c3715SXin Li * 209*bf2c3715SXin Li * Example: \include SelfAdjointEigenSolver_compute_MatrixType.cpp 210*bf2c3715SXin Li * Output: \verbinclude SelfAdjointEigenSolver_compute_MatrixType.out 211*bf2c3715SXin Li * 212*bf2c3715SXin Li * \sa SelfAdjointEigenSolver(const MatrixType&, int) 213*bf2c3715SXin Li */ 214*bf2c3715SXin Li template<typename InputType> 215*bf2c3715SXin Li EIGEN_DEVICE_FUNC 216*bf2c3715SXin Li SelfAdjointEigenSolver& compute(const EigenBase<InputType>& matrix, int options = ComputeEigenvectors); 217*bf2c3715SXin Li 218*bf2c3715SXin Li /** \brief Computes eigendecomposition of given matrix using a closed-form algorithm 219*bf2c3715SXin Li * 220*bf2c3715SXin Li * This is a variant of compute(const MatrixType&, int options) which 221*bf2c3715SXin Li * directly solves the underlying polynomial equation. 222*bf2c3715SXin Li * 223*bf2c3715SXin Li * Currently only 2x2 and 3x3 matrices for which the sizes are known at compile time are supported (e.g., Matrix3d). 224*bf2c3715SXin Li * 225*bf2c3715SXin Li * This method is usually significantly faster than the QR iterative algorithm 226*bf2c3715SXin Li * but it might also be less accurate. It is also worth noting that 227*bf2c3715SXin Li * for 3x3 matrices it involves trigonometric operations which are 228*bf2c3715SXin Li * not necessarily available for all scalar types. 229*bf2c3715SXin Li * 230*bf2c3715SXin Li * For the 3x3 case, we observed the following worst case relative error regarding the eigenvalues: 231*bf2c3715SXin Li * - double: 1e-8 232*bf2c3715SXin Li * - float: 1e-3 233*bf2c3715SXin Li * 234*bf2c3715SXin Li * \sa compute(const MatrixType&, int options) 235*bf2c3715SXin Li */ 236*bf2c3715SXin Li EIGEN_DEVICE_FUNC 237*bf2c3715SXin Li SelfAdjointEigenSolver& computeDirect(const MatrixType& matrix, int options = ComputeEigenvectors); 238*bf2c3715SXin Li 239*bf2c3715SXin Li /** 240*bf2c3715SXin Li *\brief Computes the eigen decomposition from a tridiagonal symmetric matrix 241*bf2c3715SXin Li * 242*bf2c3715SXin Li * \param[in] diag The vector containing the diagonal of the matrix. 243*bf2c3715SXin Li * \param[in] subdiag The subdiagonal of the matrix. 244*bf2c3715SXin Li * \param[in] options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly. 245*bf2c3715SXin Li * \returns Reference to \c *this 246*bf2c3715SXin Li * 247*bf2c3715SXin Li * This function assumes that the matrix has been reduced to tridiagonal form. 248*bf2c3715SXin Li * 249*bf2c3715SXin Li * \sa compute(const MatrixType&, int) for more information 250*bf2c3715SXin Li */ 251*bf2c3715SXin Li SelfAdjointEigenSolver& computeFromTridiagonal(const RealVectorType& diag, const SubDiagonalType& subdiag , int options=ComputeEigenvectors); 252*bf2c3715SXin Li 253*bf2c3715SXin Li /** \brief Returns the eigenvectors of given matrix. 254*bf2c3715SXin Li * 255*bf2c3715SXin Li * \returns A const reference to the matrix whose columns are the eigenvectors. 256*bf2c3715SXin Li * 257*bf2c3715SXin Li * \pre The eigenvectors have been computed before. 258*bf2c3715SXin Li * 259*bf2c3715SXin Li * Column \f$ k \f$ of the returned matrix is an eigenvector corresponding 260*bf2c3715SXin Li * to eigenvalue number \f$ k \f$ as returned by eigenvalues(). The 261*bf2c3715SXin Li * eigenvectors are normalized to have (Euclidean) norm equal to one. If 262*bf2c3715SXin Li * this object was used to solve the eigenproblem for the selfadjoint 263*bf2c3715SXin Li * matrix \f$ A \f$, then the matrix returned by this function is the 264*bf2c3715SXin Li * matrix \f$ V \f$ in the eigendecomposition \f$ A = V D V^{-1} \f$. 265*bf2c3715SXin Li * 266*bf2c3715SXin Li * For a selfadjoint matrix, \f$ V \f$ is unitary, meaning its inverse is equal 267*bf2c3715SXin Li * to its adjoint, \f$ V^{-1} = V^{\dagger} \f$. If \f$ A \f$ is real, then 268*bf2c3715SXin Li * \f$ V \f$ is also real and therefore orthogonal, meaning its inverse is 269*bf2c3715SXin Li * equal to its transpose, \f$ V^{-1} = V^T \f$. 270*bf2c3715SXin Li * 271*bf2c3715SXin Li * Example: \include SelfAdjointEigenSolver_eigenvectors.cpp 272*bf2c3715SXin Li * Output: \verbinclude SelfAdjointEigenSolver_eigenvectors.out 273*bf2c3715SXin Li * 274*bf2c3715SXin Li * \sa eigenvalues() 275*bf2c3715SXin Li */ 276*bf2c3715SXin Li EIGEN_DEVICE_FUNC 277*bf2c3715SXin Li const EigenvectorsType& eigenvectors() const 278*bf2c3715SXin Li { 279*bf2c3715SXin Li eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); 280*bf2c3715SXin Li eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues."); 281*bf2c3715SXin Li return m_eivec; 282*bf2c3715SXin Li } 283*bf2c3715SXin Li 284*bf2c3715SXin Li /** \brief Returns the eigenvalues of given matrix. 285*bf2c3715SXin Li * 286*bf2c3715SXin Li * \returns A const reference to the column vector containing the eigenvalues. 287*bf2c3715SXin Li * 288*bf2c3715SXin Li * \pre The eigenvalues have been computed before. 289*bf2c3715SXin Li * 290*bf2c3715SXin Li * The eigenvalues are repeated according to their algebraic multiplicity, 291*bf2c3715SXin Li * so there are as many eigenvalues as rows in the matrix. The eigenvalues 292*bf2c3715SXin Li * are sorted in increasing order. 293*bf2c3715SXin Li * 294*bf2c3715SXin Li * Example: \include SelfAdjointEigenSolver_eigenvalues.cpp 295*bf2c3715SXin Li * Output: \verbinclude SelfAdjointEigenSolver_eigenvalues.out 296*bf2c3715SXin Li * 297*bf2c3715SXin Li * \sa eigenvectors(), MatrixBase::eigenvalues() 298*bf2c3715SXin Li */ 299*bf2c3715SXin Li EIGEN_DEVICE_FUNC 300*bf2c3715SXin Li const RealVectorType& eigenvalues() const 301*bf2c3715SXin Li { 302*bf2c3715SXin Li eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); 303*bf2c3715SXin Li return m_eivalues; 304*bf2c3715SXin Li } 305*bf2c3715SXin Li 306*bf2c3715SXin Li /** \brief Computes the positive-definite square root of the matrix. 307*bf2c3715SXin Li * 308*bf2c3715SXin Li * \returns the positive-definite square root of the matrix 309*bf2c3715SXin Li * 310*bf2c3715SXin Li * \pre The eigenvalues and eigenvectors of a positive-definite matrix 311*bf2c3715SXin Li * have been computed before. 312*bf2c3715SXin Li * 313*bf2c3715SXin Li * The square root of a positive-definite matrix \f$ A \f$ is the 314*bf2c3715SXin Li * positive-definite matrix whose square equals \f$ A \f$. This function 315*bf2c3715SXin Li * uses the eigendecomposition \f$ A = V D V^{-1} \f$ to compute the 316*bf2c3715SXin Li * square root as \f$ A^{1/2} = V D^{1/2} V^{-1} \f$. 317*bf2c3715SXin Li * 318*bf2c3715SXin Li * Example: \include SelfAdjointEigenSolver_operatorSqrt.cpp 319*bf2c3715SXin Li * Output: \verbinclude SelfAdjointEigenSolver_operatorSqrt.out 320*bf2c3715SXin Li * 321*bf2c3715SXin Li * \sa operatorInverseSqrt(), <a href="unsupported/group__MatrixFunctions__Module.html">MatrixFunctions Module</a> 322*bf2c3715SXin Li */ 323*bf2c3715SXin Li EIGEN_DEVICE_FUNC 324*bf2c3715SXin Li MatrixType operatorSqrt() const 325*bf2c3715SXin Li { 326*bf2c3715SXin Li eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); 327*bf2c3715SXin Li eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues."); 328*bf2c3715SXin Li return m_eivec * m_eivalues.cwiseSqrt().asDiagonal() * m_eivec.adjoint(); 329*bf2c3715SXin Li } 330*bf2c3715SXin Li 331*bf2c3715SXin Li /** \brief Computes the inverse square root of the matrix. 332*bf2c3715SXin Li * 333*bf2c3715SXin Li * \returns the inverse positive-definite square root of the matrix 334*bf2c3715SXin Li * 335*bf2c3715SXin Li * \pre The eigenvalues and eigenvectors of a positive-definite matrix 336*bf2c3715SXin Li * have been computed before. 337*bf2c3715SXin Li * 338*bf2c3715SXin Li * This function uses the eigendecomposition \f$ A = V D V^{-1} \f$ to 339*bf2c3715SXin Li * compute the inverse square root as \f$ V D^{-1/2} V^{-1} \f$. This is 340*bf2c3715SXin Li * cheaper than first computing the square root with operatorSqrt() and 341*bf2c3715SXin Li * then its inverse with MatrixBase::inverse(). 342*bf2c3715SXin Li * 343*bf2c3715SXin Li * Example: \include SelfAdjointEigenSolver_operatorInverseSqrt.cpp 344*bf2c3715SXin Li * Output: \verbinclude SelfAdjointEigenSolver_operatorInverseSqrt.out 345*bf2c3715SXin Li * 346*bf2c3715SXin Li * \sa operatorSqrt(), MatrixBase::inverse(), <a href="unsupported/group__MatrixFunctions__Module.html">MatrixFunctions Module</a> 347*bf2c3715SXin Li */ 348*bf2c3715SXin Li EIGEN_DEVICE_FUNC 349*bf2c3715SXin Li MatrixType operatorInverseSqrt() const 350*bf2c3715SXin Li { 351*bf2c3715SXin Li eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); 352*bf2c3715SXin Li eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues."); 353*bf2c3715SXin Li return m_eivec * m_eivalues.cwiseInverse().cwiseSqrt().asDiagonal() * m_eivec.adjoint(); 354*bf2c3715SXin Li } 355*bf2c3715SXin Li 356*bf2c3715SXin Li /** \brief Reports whether previous computation was successful. 357*bf2c3715SXin Li * 358*bf2c3715SXin Li * \returns \c Success if computation was successful, \c NoConvergence otherwise. 359*bf2c3715SXin Li */ 360*bf2c3715SXin Li EIGEN_DEVICE_FUNC 361*bf2c3715SXin Li ComputationInfo info() const 362*bf2c3715SXin Li { 363*bf2c3715SXin Li eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); 364*bf2c3715SXin Li return m_info; 365*bf2c3715SXin Li } 366*bf2c3715SXin Li 367*bf2c3715SXin Li /** \brief Maximum number of iterations. 368*bf2c3715SXin Li * 369*bf2c3715SXin Li * The algorithm terminates if it does not converge within m_maxIterations * n iterations, where n 370*bf2c3715SXin Li * denotes the size of the matrix. This value is currently set to 30 (copied from LAPACK). 371*bf2c3715SXin Li */ 372*bf2c3715SXin Li static const int m_maxIterations = 30; 373*bf2c3715SXin Li 374*bf2c3715SXin Li protected: 375*bf2c3715SXin Li static EIGEN_DEVICE_FUNC 376*bf2c3715SXin Li void check_template_parameters() 377*bf2c3715SXin Li { 378*bf2c3715SXin Li EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar); 379*bf2c3715SXin Li } 380*bf2c3715SXin Li 381*bf2c3715SXin Li EigenvectorsType m_eivec; 382*bf2c3715SXin Li RealVectorType m_eivalues; 383*bf2c3715SXin Li typename TridiagonalizationType::SubDiagonalType m_subdiag; 384*bf2c3715SXin Li typename TridiagonalizationType::CoeffVectorType m_hcoeffs; 385*bf2c3715SXin Li ComputationInfo m_info; 386*bf2c3715SXin Li bool m_isInitialized; 387*bf2c3715SXin Li bool m_eigenvectorsOk; 388*bf2c3715SXin Li }; 389*bf2c3715SXin Li 390*bf2c3715SXin Li namespace internal { 391*bf2c3715SXin Li /** \internal 392*bf2c3715SXin Li * 393*bf2c3715SXin Li * \eigenvalues_module \ingroup Eigenvalues_Module 394*bf2c3715SXin Li * 395*bf2c3715SXin Li * Performs a QR step on a tridiagonal symmetric matrix represented as a 396*bf2c3715SXin Li * pair of two vectors \a diag and \a subdiag. 397*bf2c3715SXin Li * 398*bf2c3715SXin Li * \param diag the diagonal part of the input selfadjoint tridiagonal matrix 399*bf2c3715SXin Li * \param subdiag the sub-diagonal part of the input selfadjoint tridiagonal matrix 400*bf2c3715SXin Li * \param start starting index of the submatrix to work on 401*bf2c3715SXin Li * \param end last+1 index of the submatrix to work on 402*bf2c3715SXin Li * \param matrixQ pointer to the column-major matrix holding the eigenvectors, can be 0 403*bf2c3715SXin Li * \param n size of the input matrix 404*bf2c3715SXin Li * 405*bf2c3715SXin Li * For compilation efficiency reasons, this procedure does not use eigen expression 406*bf2c3715SXin Li * for its arguments. 407*bf2c3715SXin Li * 408*bf2c3715SXin Li * Implemented from Golub's "Matrix Computations", algorithm 8.3.2: 409*bf2c3715SXin Li * "implicit symmetric QR step with Wilkinson shift" 410*bf2c3715SXin Li */ 411*bf2c3715SXin Li template<int StorageOrder,typename RealScalar, typename Scalar, typename Index> 412*bf2c3715SXin Li EIGEN_DEVICE_FUNC 413*bf2c3715SXin Li static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n); 414*bf2c3715SXin Li } 415*bf2c3715SXin Li 416*bf2c3715SXin Li template<typename MatrixType> 417*bf2c3715SXin Li template<typename InputType> 418*bf2c3715SXin Li EIGEN_DEVICE_FUNC 419*bf2c3715SXin Li SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType> 420*bf2c3715SXin Li ::compute(const EigenBase<InputType>& a_matrix, int options) 421*bf2c3715SXin Li { 422*bf2c3715SXin Li check_template_parameters(); 423*bf2c3715SXin Li 424*bf2c3715SXin Li const InputType &matrix(a_matrix.derived()); 425*bf2c3715SXin Li 426*bf2c3715SXin Li EIGEN_USING_STD(abs); 427*bf2c3715SXin Li eigen_assert(matrix.cols() == matrix.rows()); 428*bf2c3715SXin Li eigen_assert((options&~(EigVecMask|GenEigMask))==0 429*bf2c3715SXin Li && (options&EigVecMask)!=EigVecMask 430*bf2c3715SXin Li && "invalid option parameter"); 431*bf2c3715SXin Li bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; 432*bf2c3715SXin Li Index n = matrix.cols(); 433*bf2c3715SXin Li m_eivalues.resize(n,1); 434*bf2c3715SXin Li 435*bf2c3715SXin Li if(n==1) 436*bf2c3715SXin Li { 437*bf2c3715SXin Li m_eivec = matrix; 438*bf2c3715SXin Li m_eivalues.coeffRef(0,0) = numext::real(m_eivec.coeff(0,0)); 439*bf2c3715SXin Li if(computeEigenvectors) 440*bf2c3715SXin Li m_eivec.setOnes(n,n); 441*bf2c3715SXin Li m_info = Success; 442*bf2c3715SXin Li m_isInitialized = true; 443*bf2c3715SXin Li m_eigenvectorsOk = computeEigenvectors; 444*bf2c3715SXin Li return *this; 445*bf2c3715SXin Li } 446*bf2c3715SXin Li 447*bf2c3715SXin Li // declare some aliases 448*bf2c3715SXin Li RealVectorType& diag = m_eivalues; 449*bf2c3715SXin Li EigenvectorsType& mat = m_eivec; 450*bf2c3715SXin Li 451*bf2c3715SXin Li // map the matrix coefficients to [-1:1] to avoid over- and underflow. 452*bf2c3715SXin Li mat = matrix.template triangularView<Lower>(); 453*bf2c3715SXin Li RealScalar scale = mat.cwiseAbs().maxCoeff(); 454*bf2c3715SXin Li if(scale==RealScalar(0)) scale = RealScalar(1); 455*bf2c3715SXin Li mat.template triangularView<Lower>() /= scale; 456*bf2c3715SXin Li m_subdiag.resize(n-1); 457*bf2c3715SXin Li m_hcoeffs.resize(n-1); 458*bf2c3715SXin Li internal::tridiagonalization_inplace(mat, diag, m_subdiag, m_hcoeffs, computeEigenvectors); 459*bf2c3715SXin Li 460*bf2c3715SXin Li m_info = internal::computeFromTridiagonal_impl(diag, m_subdiag, m_maxIterations, computeEigenvectors, m_eivec); 461*bf2c3715SXin Li 462*bf2c3715SXin Li // scale back the eigen values 463*bf2c3715SXin Li m_eivalues *= scale; 464*bf2c3715SXin Li 465*bf2c3715SXin Li m_isInitialized = true; 466*bf2c3715SXin Li m_eigenvectorsOk = computeEigenvectors; 467*bf2c3715SXin Li return *this; 468*bf2c3715SXin Li } 469*bf2c3715SXin Li 470*bf2c3715SXin Li template<typename MatrixType> 471*bf2c3715SXin Li SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType> 472*bf2c3715SXin Li ::computeFromTridiagonal(const RealVectorType& diag, const SubDiagonalType& subdiag , int options) 473*bf2c3715SXin Li { 474*bf2c3715SXin Li //TODO : Add an option to scale the values beforehand 475*bf2c3715SXin Li bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; 476*bf2c3715SXin Li 477*bf2c3715SXin Li m_eivalues = diag; 478*bf2c3715SXin Li m_subdiag = subdiag; 479*bf2c3715SXin Li if (computeEigenvectors) 480*bf2c3715SXin Li { 481*bf2c3715SXin Li m_eivec.setIdentity(diag.size(), diag.size()); 482*bf2c3715SXin Li } 483*bf2c3715SXin Li m_info = internal::computeFromTridiagonal_impl(m_eivalues, m_subdiag, m_maxIterations, computeEigenvectors, m_eivec); 484*bf2c3715SXin Li 485*bf2c3715SXin Li m_isInitialized = true; 486*bf2c3715SXin Li m_eigenvectorsOk = computeEigenvectors; 487*bf2c3715SXin Li return *this; 488*bf2c3715SXin Li } 489*bf2c3715SXin Li 490*bf2c3715SXin Li namespace internal { 491*bf2c3715SXin Li /** 492*bf2c3715SXin Li * \internal 493*bf2c3715SXin Li * \brief Compute the eigendecomposition from a tridiagonal matrix 494*bf2c3715SXin Li * 495*bf2c3715SXin Li * \param[in,out] diag : On input, the diagonal of the matrix, on output the eigenvalues 496*bf2c3715SXin Li * \param[in,out] subdiag : The subdiagonal part of the matrix (entries are modified during the decomposition) 497*bf2c3715SXin Li * \param[in] maxIterations : the maximum number of iterations 498*bf2c3715SXin Li * \param[in] computeEigenvectors : whether the eigenvectors have to be computed or not 499*bf2c3715SXin Li * \param[out] eivec : The matrix to store the eigenvectors if computeEigenvectors==true. Must be allocated on input. 500*bf2c3715SXin Li * \returns \c Success or \c NoConvergence 501*bf2c3715SXin Li */ 502*bf2c3715SXin Li template<typename MatrixType, typename DiagType, typename SubDiagType> 503*bf2c3715SXin Li EIGEN_DEVICE_FUNC 504*bf2c3715SXin Li ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec) 505*bf2c3715SXin Li { 506*bf2c3715SXin Li ComputationInfo info; 507*bf2c3715SXin Li typedef typename MatrixType::Scalar Scalar; 508*bf2c3715SXin Li 509*bf2c3715SXin Li Index n = diag.size(); 510*bf2c3715SXin Li Index end = n-1; 511*bf2c3715SXin Li Index start = 0; 512*bf2c3715SXin Li Index iter = 0; // total number of iterations 513*bf2c3715SXin Li 514*bf2c3715SXin Li typedef typename DiagType::RealScalar RealScalar; 515*bf2c3715SXin Li const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)(); 516*bf2c3715SXin Li const RealScalar precision_inv = RealScalar(1)/NumTraits<RealScalar>::epsilon(); 517*bf2c3715SXin Li while (end>0) 518*bf2c3715SXin Li { 519*bf2c3715SXin Li for (Index i = start; i<end; ++i) { 520*bf2c3715SXin Li if (numext::abs(subdiag[i]) < considerAsZero) { 521*bf2c3715SXin Li subdiag[i] = RealScalar(0); 522*bf2c3715SXin Li } else { 523*bf2c3715SXin Li // abs(subdiag[i]) <= epsilon * sqrt(abs(diag[i]) + abs(diag[i+1])) 524*bf2c3715SXin Li // Scaled to prevent underflows. 525*bf2c3715SXin Li const RealScalar scaled_subdiag = precision_inv * subdiag[i]; 526*bf2c3715SXin Li if (scaled_subdiag * scaled_subdiag <= (numext::abs(diag[i])+numext::abs(diag[i+1]))) { 527*bf2c3715SXin Li subdiag[i] = RealScalar(0); 528*bf2c3715SXin Li } 529*bf2c3715SXin Li } 530*bf2c3715SXin Li } 531*bf2c3715SXin Li 532*bf2c3715SXin Li // find the largest unreduced block at the end of the matrix. 533*bf2c3715SXin Li while (end>0 && subdiag[end-1]==RealScalar(0)) 534*bf2c3715SXin Li { 535*bf2c3715SXin Li end--; 536*bf2c3715SXin Li } 537*bf2c3715SXin Li if (end<=0) 538*bf2c3715SXin Li break; 539*bf2c3715SXin Li 540*bf2c3715SXin Li // if we spent too many iterations, we give up 541*bf2c3715SXin Li iter++; 542*bf2c3715SXin Li if(iter > maxIterations * n) break; 543*bf2c3715SXin Li 544*bf2c3715SXin Li start = end - 1; 545*bf2c3715SXin Li while (start>0 && subdiag[start-1]!=0) 546*bf2c3715SXin Li start--; 547*bf2c3715SXin Li 548*bf2c3715SXin Li internal::tridiagonal_qr_step<MatrixType::Flags&RowMajorBit ? RowMajor : ColMajor>(diag.data(), subdiag.data(), start, end, computeEigenvectors ? eivec.data() : (Scalar*)0, n); 549*bf2c3715SXin Li } 550*bf2c3715SXin Li if (iter <= maxIterations * n) 551*bf2c3715SXin Li info = Success; 552*bf2c3715SXin Li else 553*bf2c3715SXin Li info = NoConvergence; 554*bf2c3715SXin Li 555*bf2c3715SXin Li // Sort eigenvalues and corresponding vectors. 556*bf2c3715SXin Li // TODO make the sort optional ? 557*bf2c3715SXin Li // TODO use a better sort algorithm !! 558*bf2c3715SXin Li if (info == Success) 559*bf2c3715SXin Li { 560*bf2c3715SXin Li for (Index i = 0; i < n-1; ++i) 561*bf2c3715SXin Li { 562*bf2c3715SXin Li Index k; 563*bf2c3715SXin Li diag.segment(i,n-i).minCoeff(&k); 564*bf2c3715SXin Li if (k > 0) 565*bf2c3715SXin Li { 566*bf2c3715SXin Li numext::swap(diag[i], diag[k+i]); 567*bf2c3715SXin Li if(computeEigenvectors) 568*bf2c3715SXin Li eivec.col(i).swap(eivec.col(k+i)); 569*bf2c3715SXin Li } 570*bf2c3715SXin Li } 571*bf2c3715SXin Li } 572*bf2c3715SXin Li return info; 573*bf2c3715SXin Li } 574*bf2c3715SXin Li 575*bf2c3715SXin Li template<typename SolverType,int Size,bool IsComplex> struct direct_selfadjoint_eigenvalues 576*bf2c3715SXin Li { 577*bf2c3715SXin Li EIGEN_DEVICE_FUNC 578*bf2c3715SXin Li static inline void run(SolverType& eig, const typename SolverType::MatrixType& A, int options) 579*bf2c3715SXin Li { eig.compute(A,options); } 580*bf2c3715SXin Li }; 581*bf2c3715SXin Li 582*bf2c3715SXin Li template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3,false> 583*bf2c3715SXin Li { 584*bf2c3715SXin Li typedef typename SolverType::MatrixType MatrixType; 585*bf2c3715SXin Li typedef typename SolverType::RealVectorType VectorType; 586*bf2c3715SXin Li typedef typename SolverType::Scalar Scalar; 587*bf2c3715SXin Li typedef typename SolverType::EigenvectorsType EigenvectorsType; 588*bf2c3715SXin Li 589*bf2c3715SXin Li 590*bf2c3715SXin Li /** \internal 591*bf2c3715SXin Li * Computes the roots of the characteristic polynomial of \a m. 592*bf2c3715SXin Li * For numerical stability m.trace() should be near zero and to avoid over- or underflow m should be normalized. 593*bf2c3715SXin Li */ 594*bf2c3715SXin Li EIGEN_DEVICE_FUNC 595*bf2c3715SXin Li static inline void computeRoots(const MatrixType& m, VectorType& roots) 596*bf2c3715SXin Li { 597*bf2c3715SXin Li EIGEN_USING_STD(sqrt) 598*bf2c3715SXin Li EIGEN_USING_STD(atan2) 599*bf2c3715SXin Li EIGEN_USING_STD(cos) 600*bf2c3715SXin Li EIGEN_USING_STD(sin) 601*bf2c3715SXin Li const Scalar s_inv3 = Scalar(1)/Scalar(3); 602*bf2c3715SXin Li const Scalar s_sqrt3 = sqrt(Scalar(3)); 603*bf2c3715SXin Li 604*bf2c3715SXin Li // The characteristic equation is x^3 - c2*x^2 + c1*x - c0 = 0. The 605*bf2c3715SXin Li // eigenvalues are the roots to this equation, all guaranteed to be 606*bf2c3715SXin Li // real-valued, because the matrix is symmetric. 607*bf2c3715SXin Li Scalar c0 = m(0,0)*m(1,1)*m(2,2) + Scalar(2)*m(1,0)*m(2,0)*m(2,1) - m(0,0)*m(2,1)*m(2,1) - m(1,1)*m(2,0)*m(2,0) - m(2,2)*m(1,0)*m(1,0); 608*bf2c3715SXin Li Scalar c1 = m(0,0)*m(1,1) - m(1,0)*m(1,0) + m(0,0)*m(2,2) - m(2,0)*m(2,0) + m(1,1)*m(2,2) - m(2,1)*m(2,1); 609*bf2c3715SXin Li Scalar c2 = m(0,0) + m(1,1) + m(2,2); 610*bf2c3715SXin Li 611*bf2c3715SXin Li // Construct the parameters used in classifying the roots of the equation 612*bf2c3715SXin Li // and in solving the equation for the roots in closed form. 613*bf2c3715SXin Li Scalar c2_over_3 = c2*s_inv3; 614*bf2c3715SXin Li Scalar a_over_3 = (c2*c2_over_3 - c1)*s_inv3; 615*bf2c3715SXin Li a_over_3 = numext::maxi(a_over_3, Scalar(0)); 616*bf2c3715SXin Li 617*bf2c3715SXin Li Scalar half_b = Scalar(0.5)*(c0 + c2_over_3*(Scalar(2)*c2_over_3*c2_over_3 - c1)); 618*bf2c3715SXin Li 619*bf2c3715SXin Li Scalar q = a_over_3*a_over_3*a_over_3 - half_b*half_b; 620*bf2c3715SXin Li q = numext::maxi(q, Scalar(0)); 621*bf2c3715SXin Li 622*bf2c3715SXin Li // Compute the eigenvalues by solving for the roots of the polynomial. 623*bf2c3715SXin Li Scalar rho = sqrt(a_over_3); 624*bf2c3715SXin Li Scalar theta = atan2(sqrt(q),half_b)*s_inv3; // since sqrt(q) > 0, atan2 is in [0, pi] and theta is in [0, pi/3] 625*bf2c3715SXin Li Scalar cos_theta = cos(theta); 626*bf2c3715SXin Li Scalar sin_theta = sin(theta); 627*bf2c3715SXin Li // roots are already sorted, since cos is monotonically decreasing on [0, pi] 628*bf2c3715SXin Li roots(0) = c2_over_3 - rho*(cos_theta + s_sqrt3*sin_theta); // == 2*rho*cos(theta+2pi/3) 629*bf2c3715SXin Li roots(1) = c2_over_3 - rho*(cos_theta - s_sqrt3*sin_theta); // == 2*rho*cos(theta+ pi/3) 630*bf2c3715SXin Li roots(2) = c2_over_3 + Scalar(2)*rho*cos_theta; 631*bf2c3715SXin Li } 632*bf2c3715SXin Li 633*bf2c3715SXin Li EIGEN_DEVICE_FUNC 634*bf2c3715SXin Li static inline bool extract_kernel(MatrixType& mat, Ref<VectorType> res, Ref<VectorType> representative) 635*bf2c3715SXin Li { 636*bf2c3715SXin Li EIGEN_USING_STD(abs); 637*bf2c3715SXin Li EIGEN_USING_STD(sqrt); 638*bf2c3715SXin Li Index i0; 639*bf2c3715SXin Li // Find non-zero column i0 (by construction, there must exist a non zero coefficient on the diagonal): 640*bf2c3715SXin Li mat.diagonal().cwiseAbs().maxCoeff(&i0); 641*bf2c3715SXin Li // mat.col(i0) is a good candidate for an orthogonal vector to the current eigenvector, 642*bf2c3715SXin Li // so let's save it: 643*bf2c3715SXin Li representative = mat.col(i0); 644*bf2c3715SXin Li Scalar n0, n1; 645*bf2c3715SXin Li VectorType c0, c1; 646*bf2c3715SXin Li n0 = (c0 = representative.cross(mat.col((i0+1)%3))).squaredNorm(); 647*bf2c3715SXin Li n1 = (c1 = representative.cross(mat.col((i0+2)%3))).squaredNorm(); 648*bf2c3715SXin Li if(n0>n1) res = c0/sqrt(n0); 649*bf2c3715SXin Li else res = c1/sqrt(n1); 650*bf2c3715SXin Li 651*bf2c3715SXin Li return true; 652*bf2c3715SXin Li } 653*bf2c3715SXin Li 654*bf2c3715SXin Li EIGEN_DEVICE_FUNC 655*bf2c3715SXin Li static inline void run(SolverType& solver, const MatrixType& mat, int options) 656*bf2c3715SXin Li { 657*bf2c3715SXin Li eigen_assert(mat.cols() == 3 && mat.cols() == mat.rows()); 658*bf2c3715SXin Li eigen_assert((options&~(EigVecMask|GenEigMask))==0 659*bf2c3715SXin Li && (options&EigVecMask)!=EigVecMask 660*bf2c3715SXin Li && "invalid option parameter"); 661*bf2c3715SXin Li bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; 662*bf2c3715SXin Li 663*bf2c3715SXin Li EigenvectorsType& eivecs = solver.m_eivec; 664*bf2c3715SXin Li VectorType& eivals = solver.m_eivalues; 665*bf2c3715SXin Li 666*bf2c3715SXin Li // Shift the matrix to the mean eigenvalue and map the matrix coefficients to [-1:1] to avoid over- and underflow. 667*bf2c3715SXin Li Scalar shift = mat.trace() / Scalar(3); 668*bf2c3715SXin Li // TODO Avoid this copy. Currently it is necessary to suppress bogus values when determining maxCoeff and for computing the eigenvectors later 669*bf2c3715SXin Li MatrixType scaledMat = mat.template selfadjointView<Lower>(); 670*bf2c3715SXin Li scaledMat.diagonal().array() -= shift; 671*bf2c3715SXin Li Scalar scale = scaledMat.cwiseAbs().maxCoeff(); 672*bf2c3715SXin Li if(scale > 0) scaledMat /= scale; // TODO for scale==0 we could save the remaining operations 673*bf2c3715SXin Li 674*bf2c3715SXin Li // compute the eigenvalues 675*bf2c3715SXin Li computeRoots(scaledMat,eivals); 676*bf2c3715SXin Li 677*bf2c3715SXin Li // compute the eigenvectors 678*bf2c3715SXin Li if(computeEigenvectors) 679*bf2c3715SXin Li { 680*bf2c3715SXin Li if((eivals(2)-eivals(0))<=Eigen::NumTraits<Scalar>::epsilon()) 681*bf2c3715SXin Li { 682*bf2c3715SXin Li // All three eigenvalues are numerically the same 683*bf2c3715SXin Li eivecs.setIdentity(); 684*bf2c3715SXin Li } 685*bf2c3715SXin Li else 686*bf2c3715SXin Li { 687*bf2c3715SXin Li MatrixType tmp; 688*bf2c3715SXin Li tmp = scaledMat; 689*bf2c3715SXin Li 690*bf2c3715SXin Li // Compute the eigenvector of the most distinct eigenvalue 691*bf2c3715SXin Li Scalar d0 = eivals(2) - eivals(1); 692*bf2c3715SXin Li Scalar d1 = eivals(1) - eivals(0); 693*bf2c3715SXin Li Index k(0), l(2); 694*bf2c3715SXin Li if(d0 > d1) 695*bf2c3715SXin Li { 696*bf2c3715SXin Li numext::swap(k,l); 697*bf2c3715SXin Li d0 = d1; 698*bf2c3715SXin Li } 699*bf2c3715SXin Li 700*bf2c3715SXin Li // Compute the eigenvector of index k 701*bf2c3715SXin Li { 702*bf2c3715SXin Li tmp.diagonal().array () -= eivals(k); 703*bf2c3715SXin Li // By construction, 'tmp' is of rank 2, and its kernel corresponds to the respective eigenvector. 704*bf2c3715SXin Li extract_kernel(tmp, eivecs.col(k), eivecs.col(l)); 705*bf2c3715SXin Li } 706*bf2c3715SXin Li 707*bf2c3715SXin Li // Compute eigenvector of index l 708*bf2c3715SXin Li if(d0<=2*Eigen::NumTraits<Scalar>::epsilon()*d1) 709*bf2c3715SXin Li { 710*bf2c3715SXin Li // If d0 is too small, then the two other eigenvalues are numerically the same, 711*bf2c3715SXin Li // and thus we only have to ortho-normalize the near orthogonal vector we saved above. 712*bf2c3715SXin Li eivecs.col(l) -= eivecs.col(k).dot(eivecs.col(l))*eivecs.col(l); 713*bf2c3715SXin Li eivecs.col(l).normalize(); 714*bf2c3715SXin Li } 715*bf2c3715SXin Li else 716*bf2c3715SXin Li { 717*bf2c3715SXin Li tmp = scaledMat; 718*bf2c3715SXin Li tmp.diagonal().array () -= eivals(l); 719*bf2c3715SXin Li 720*bf2c3715SXin Li VectorType dummy; 721*bf2c3715SXin Li extract_kernel(tmp, eivecs.col(l), dummy); 722*bf2c3715SXin Li } 723*bf2c3715SXin Li 724*bf2c3715SXin Li // Compute last eigenvector from the other two 725*bf2c3715SXin Li eivecs.col(1) = eivecs.col(2).cross(eivecs.col(0)).normalized(); 726*bf2c3715SXin Li } 727*bf2c3715SXin Li } 728*bf2c3715SXin Li 729*bf2c3715SXin Li // Rescale back to the original size. 730*bf2c3715SXin Li eivals *= scale; 731*bf2c3715SXin Li eivals.array() += shift; 732*bf2c3715SXin Li 733*bf2c3715SXin Li solver.m_info = Success; 734*bf2c3715SXin Li solver.m_isInitialized = true; 735*bf2c3715SXin Li solver.m_eigenvectorsOk = computeEigenvectors; 736*bf2c3715SXin Li } 737*bf2c3715SXin Li }; 738*bf2c3715SXin Li 739*bf2c3715SXin Li // 2x2 direct eigenvalues decomposition, code from Hauke Heibel 740*bf2c3715SXin Li template<typename SolverType> 741*bf2c3715SXin Li struct direct_selfadjoint_eigenvalues<SolverType,2,false> 742*bf2c3715SXin Li { 743*bf2c3715SXin Li typedef typename SolverType::MatrixType MatrixType; 744*bf2c3715SXin Li typedef typename SolverType::RealVectorType VectorType; 745*bf2c3715SXin Li typedef typename SolverType::Scalar Scalar; 746*bf2c3715SXin Li typedef typename SolverType::EigenvectorsType EigenvectorsType; 747*bf2c3715SXin Li 748*bf2c3715SXin Li EIGEN_DEVICE_FUNC 749*bf2c3715SXin Li static inline void computeRoots(const MatrixType& m, VectorType& roots) 750*bf2c3715SXin Li { 751*bf2c3715SXin Li EIGEN_USING_STD(sqrt); 752*bf2c3715SXin Li const Scalar t0 = Scalar(0.5) * sqrt( numext::abs2(m(0,0)-m(1,1)) + Scalar(4)*numext::abs2(m(1,0))); 753*bf2c3715SXin Li const Scalar t1 = Scalar(0.5) * (m(0,0) + m(1,1)); 754*bf2c3715SXin Li roots(0) = t1 - t0; 755*bf2c3715SXin Li roots(1) = t1 + t0; 756*bf2c3715SXin Li } 757*bf2c3715SXin Li 758*bf2c3715SXin Li EIGEN_DEVICE_FUNC 759*bf2c3715SXin Li static inline void run(SolverType& solver, const MatrixType& mat, int options) 760*bf2c3715SXin Li { 761*bf2c3715SXin Li EIGEN_USING_STD(sqrt); 762*bf2c3715SXin Li EIGEN_USING_STD(abs); 763*bf2c3715SXin Li 764*bf2c3715SXin Li eigen_assert(mat.cols() == 2 && mat.cols() == mat.rows()); 765*bf2c3715SXin Li eigen_assert((options&~(EigVecMask|GenEigMask))==0 766*bf2c3715SXin Li && (options&EigVecMask)!=EigVecMask 767*bf2c3715SXin Li && "invalid option parameter"); 768*bf2c3715SXin Li bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; 769*bf2c3715SXin Li 770*bf2c3715SXin Li EigenvectorsType& eivecs = solver.m_eivec; 771*bf2c3715SXin Li VectorType& eivals = solver.m_eivalues; 772*bf2c3715SXin Li 773*bf2c3715SXin Li // Shift the matrix to the mean eigenvalue and map the matrix coefficients to [-1:1] to avoid over- and underflow. 774*bf2c3715SXin Li Scalar shift = mat.trace() / Scalar(2); 775*bf2c3715SXin Li MatrixType scaledMat = mat; 776*bf2c3715SXin Li scaledMat.coeffRef(0,1) = mat.coeff(1,0); 777*bf2c3715SXin Li scaledMat.diagonal().array() -= shift; 778*bf2c3715SXin Li Scalar scale = scaledMat.cwiseAbs().maxCoeff(); 779*bf2c3715SXin Li if(scale > Scalar(0)) 780*bf2c3715SXin Li scaledMat /= scale; 781*bf2c3715SXin Li 782*bf2c3715SXin Li // Compute the eigenvalues 783*bf2c3715SXin Li computeRoots(scaledMat,eivals); 784*bf2c3715SXin Li 785*bf2c3715SXin Li // compute the eigen vectors 786*bf2c3715SXin Li if(computeEigenvectors) 787*bf2c3715SXin Li { 788*bf2c3715SXin Li if((eivals(1)-eivals(0))<=abs(eivals(1))*Eigen::NumTraits<Scalar>::epsilon()) 789*bf2c3715SXin Li { 790*bf2c3715SXin Li eivecs.setIdentity(); 791*bf2c3715SXin Li } 792*bf2c3715SXin Li else 793*bf2c3715SXin Li { 794*bf2c3715SXin Li scaledMat.diagonal().array () -= eivals(1); 795*bf2c3715SXin Li Scalar a2 = numext::abs2(scaledMat(0,0)); 796*bf2c3715SXin Li Scalar c2 = numext::abs2(scaledMat(1,1)); 797*bf2c3715SXin Li Scalar b2 = numext::abs2(scaledMat(1,0)); 798*bf2c3715SXin Li if(a2>c2) 799*bf2c3715SXin Li { 800*bf2c3715SXin Li eivecs.col(1) << -scaledMat(1,0), scaledMat(0,0); 801*bf2c3715SXin Li eivecs.col(1) /= sqrt(a2+b2); 802*bf2c3715SXin Li } 803*bf2c3715SXin Li else 804*bf2c3715SXin Li { 805*bf2c3715SXin Li eivecs.col(1) << -scaledMat(1,1), scaledMat(1,0); 806*bf2c3715SXin Li eivecs.col(1) /= sqrt(c2+b2); 807*bf2c3715SXin Li } 808*bf2c3715SXin Li 809*bf2c3715SXin Li eivecs.col(0) << eivecs.col(1).unitOrthogonal(); 810*bf2c3715SXin Li } 811*bf2c3715SXin Li } 812*bf2c3715SXin Li 813*bf2c3715SXin Li // Rescale back to the original size. 814*bf2c3715SXin Li eivals *= scale; 815*bf2c3715SXin Li eivals.array() += shift; 816*bf2c3715SXin Li 817*bf2c3715SXin Li solver.m_info = Success; 818*bf2c3715SXin Li solver.m_isInitialized = true; 819*bf2c3715SXin Li solver.m_eigenvectorsOk = computeEigenvectors; 820*bf2c3715SXin Li } 821*bf2c3715SXin Li }; 822*bf2c3715SXin Li 823*bf2c3715SXin Li } 824*bf2c3715SXin Li 825*bf2c3715SXin Li template<typename MatrixType> 826*bf2c3715SXin Li EIGEN_DEVICE_FUNC 827*bf2c3715SXin Li SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType> 828*bf2c3715SXin Li ::computeDirect(const MatrixType& matrix, int options) 829*bf2c3715SXin Li { 830*bf2c3715SXin Li internal::direct_selfadjoint_eigenvalues<SelfAdjointEigenSolver,Size,NumTraits<Scalar>::IsComplex>::run(*this,matrix,options); 831*bf2c3715SXin Li return *this; 832*bf2c3715SXin Li } 833*bf2c3715SXin Li 834*bf2c3715SXin Li namespace internal { 835*bf2c3715SXin Li 836*bf2c3715SXin Li // Francis implicit QR step. 837*bf2c3715SXin Li template<int StorageOrder,typename RealScalar, typename Scalar, typename Index> 838*bf2c3715SXin Li EIGEN_DEVICE_FUNC 839*bf2c3715SXin Li static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n) 840*bf2c3715SXin Li { 841*bf2c3715SXin Li // Wilkinson Shift. 842*bf2c3715SXin Li RealScalar td = (diag[end-1] - diag[end])*RealScalar(0.5); 843*bf2c3715SXin Li RealScalar e = subdiag[end-1]; 844*bf2c3715SXin Li // Note that thanks to scaling, e^2 or td^2 cannot overflow, however they can still 845*bf2c3715SXin Li // underflow thus leading to inf/NaN values when using the following commented code: 846*bf2c3715SXin Li // RealScalar e2 = numext::abs2(subdiag[end-1]); 847*bf2c3715SXin Li // RealScalar mu = diag[end] - e2 / (td + (td>0 ? 1 : -1) * sqrt(td*td + e2)); 848*bf2c3715SXin Li // This explain the following, somewhat more complicated, version: 849*bf2c3715SXin Li RealScalar mu = diag[end]; 850*bf2c3715SXin Li if(td==RealScalar(0)) { 851*bf2c3715SXin Li mu -= numext::abs(e); 852*bf2c3715SXin Li } else if (e != RealScalar(0)) { 853*bf2c3715SXin Li const RealScalar e2 = numext::abs2(e); 854*bf2c3715SXin Li const RealScalar h = numext::hypot(td,e); 855*bf2c3715SXin Li if(e2 == RealScalar(0)) { 856*bf2c3715SXin Li mu -= e / ((td + (td>RealScalar(0) ? h : -h)) / e); 857*bf2c3715SXin Li } else { 858*bf2c3715SXin Li mu -= e2 / (td + (td>RealScalar(0) ? h : -h)); 859*bf2c3715SXin Li } 860*bf2c3715SXin Li } 861*bf2c3715SXin Li 862*bf2c3715SXin Li RealScalar x = diag[start] - mu; 863*bf2c3715SXin Li RealScalar z = subdiag[start]; 864*bf2c3715SXin Li // If z ever becomes zero, the Givens rotation will be the identity and 865*bf2c3715SXin Li // z will stay zero for all future iterations. 866*bf2c3715SXin Li for (Index k = start; k < end && z != RealScalar(0); ++k) 867*bf2c3715SXin Li { 868*bf2c3715SXin Li JacobiRotation<RealScalar> rot; 869*bf2c3715SXin Li rot.makeGivens(x, z); 870*bf2c3715SXin Li 871*bf2c3715SXin Li // do T = G' T G 872*bf2c3715SXin Li RealScalar sdk = rot.s() * diag[k] + rot.c() * subdiag[k]; 873*bf2c3715SXin Li RealScalar dkp1 = rot.s() * subdiag[k] + rot.c() * diag[k+1]; 874*bf2c3715SXin Li 875*bf2c3715SXin Li diag[k] = rot.c() * (rot.c() * diag[k] - rot.s() * subdiag[k]) - rot.s() * (rot.c() * subdiag[k] - rot.s() * diag[k+1]); 876*bf2c3715SXin Li diag[k+1] = rot.s() * sdk + rot.c() * dkp1; 877*bf2c3715SXin Li subdiag[k] = rot.c() * sdk - rot.s() * dkp1; 878*bf2c3715SXin Li 879*bf2c3715SXin Li if (k > start) 880*bf2c3715SXin Li subdiag[k - 1] = rot.c() * subdiag[k-1] - rot.s() * z; 881*bf2c3715SXin Li 882*bf2c3715SXin Li // "Chasing the bulge" to return to triangular form. 883*bf2c3715SXin Li x = subdiag[k]; 884*bf2c3715SXin Li if (k < end - 1) 885*bf2c3715SXin Li { 886*bf2c3715SXin Li z = -rot.s() * subdiag[k+1]; 887*bf2c3715SXin Li subdiag[k + 1] = rot.c() * subdiag[k+1]; 888*bf2c3715SXin Li } 889*bf2c3715SXin Li 890*bf2c3715SXin Li // apply the givens rotation to the unit matrix Q = Q * G 891*bf2c3715SXin Li if (matrixQ) 892*bf2c3715SXin Li { 893*bf2c3715SXin Li // FIXME if StorageOrder == RowMajor this operation is not very efficient 894*bf2c3715SXin Li Map<Matrix<Scalar,Dynamic,Dynamic,StorageOrder> > q(matrixQ,n,n); 895*bf2c3715SXin Li q.applyOnTheRight(k,k+1,rot); 896*bf2c3715SXin Li } 897*bf2c3715SXin Li } 898*bf2c3715SXin Li } 899*bf2c3715SXin Li 900*bf2c3715SXin Li } // end namespace internal 901*bf2c3715SXin Li 902*bf2c3715SXin Li } // end namespace Eigen 903*bf2c3715SXin Li 904*bf2c3715SXin Li #endif // EIGEN_SELFADJOINTEIGENSOLVER_H 905