Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 1 | // This file is part of Eigen, a lightweight C++ template library |
| 2 | // for linear algebra. |
| 3 | // |
| 4 | // Copyright (C) 2012, 2013 Chen-Pang He <jdh8@ms63.hinet.net> |
| 5 | // |
| 6 | // This Source Code Form is subject to the terms of the Mozilla |
| 7 | // Public License v. 2.0. If a copy of the MPL was not distributed |
| 8 | // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. |
| 9 | |
| 10 | #ifndef EIGEN_MATRIX_POWER |
| 11 | #define EIGEN_MATRIX_POWER |
| 12 | |
| 13 | namespace Eigen { |
| 14 | |
| 15 | template<typename MatrixType> class MatrixPower; |
| 16 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 17 | /** |
| 18 | * \ingroup MatrixFunctions_Module |
| 19 | * |
| 20 | * \brief Proxy for the matrix power of some matrix. |
| 21 | * |
| 22 | * \tparam MatrixType type of the base, a matrix. |
| 23 | * |
| 24 | * This class holds the arguments to the matrix power until it is |
| 25 | * assigned or evaluated for some other reason (so the argument |
| 26 | * should not be changed in the meantime). It is the return type of |
| 27 | * MatrixPower::operator() and related functions and most of the |
| 28 | * time this is the only way it is used. |
| 29 | */ |
| 30 | /* TODO This class is only used by MatrixPower, so it should be nested |
| 31 | * into MatrixPower, like MatrixPower::ReturnValue. However, my |
| 32 | * compiler complained about unused template parameter in the |
| 33 | * following declaration in namespace internal. |
| 34 | * |
| 35 | * template<typename MatrixType> |
| 36 | * struct traits<MatrixPower<MatrixType>::ReturnValue>; |
| 37 | */ |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 38 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 39 | class MatrixPowerParenthesesReturnValue : public ReturnByValue< MatrixPowerParenthesesReturnValue<MatrixType> > |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 40 | { |
| 41 | public: |
| 42 | typedef typename MatrixType::RealScalar RealScalar; |
| 43 | typedef typename MatrixType::Index Index; |
| 44 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 45 | /** |
| 46 | * \brief Constructor. |
| 47 | * |
| 48 | * \param[in] pow %MatrixPower storing the base. |
| 49 | * \param[in] p scalar, the exponent of the matrix power. |
| 50 | */ |
| 51 | MatrixPowerParenthesesReturnValue(MatrixPower<MatrixType>& pow, RealScalar p) : m_pow(pow), m_p(p) |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 52 | { } |
| 53 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 54 | /** |
| 55 | * \brief Compute the matrix power. |
| 56 | * |
| 57 | * \param[out] result |
| 58 | */ |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 59 | template<typename ResultType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 60 | inline void evalTo(ResultType& result) const |
| 61 | { m_pow.compute(result, m_p); } |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 62 | |
| 63 | Index rows() const { return m_pow.rows(); } |
| 64 | Index cols() const { return m_pow.cols(); } |
| 65 | |
| 66 | private: |
| 67 | MatrixPower<MatrixType>& m_pow; |
| 68 | const RealScalar m_p; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 69 | }; |
| 70 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 71 | /** |
| 72 | * \ingroup MatrixFunctions_Module |
| 73 | * |
| 74 | * \brief Class for computing matrix powers. |
| 75 | * |
| 76 | * \tparam MatrixType type of the base, expected to be an instantiation |
| 77 | * of the Matrix class template. |
| 78 | * |
| 79 | * This class is capable of computing triangular real/complex matrices |
| 80 | * raised to a power in the interval \f$ (-1, 1) \f$. |
| 81 | * |
| 82 | * \note Currently this class is only used by MatrixPower. One may |
| 83 | * insist that this be nested into MatrixPower. This class is here to |
| 84 | * faciliate future development of triangular matrix functions. |
| 85 | */ |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 86 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 87 | class MatrixPowerAtomic : internal::noncopyable |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 88 | { |
| 89 | private: |
| 90 | enum { |
| 91 | RowsAtCompileTime = MatrixType::RowsAtCompileTime, |
| 92 | MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime |
| 93 | }; |
| 94 | typedef typename MatrixType::Scalar Scalar; |
| 95 | typedef typename MatrixType::RealScalar RealScalar; |
| 96 | typedef std::complex<RealScalar> ComplexScalar; |
| 97 | typedef typename MatrixType::Index Index; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 98 | typedef Block<MatrixType,Dynamic,Dynamic> ResultType; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 99 | |
| 100 | const MatrixType& m_A; |
| 101 | RealScalar m_p; |
| 102 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 103 | void computePade(int degree, const MatrixType& IminusT, ResultType& res) const; |
| 104 | void compute2x2(ResultType& res, RealScalar p) const; |
| 105 | void computeBig(ResultType& res) const; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 106 | static int getPadeDegree(float normIminusT); |
| 107 | static int getPadeDegree(double normIminusT); |
| 108 | static int getPadeDegree(long double normIminusT); |
| 109 | static ComplexScalar computeSuperDiag(const ComplexScalar&, const ComplexScalar&, RealScalar p); |
| 110 | static RealScalar computeSuperDiag(RealScalar, RealScalar, RealScalar p); |
| 111 | |
| 112 | public: |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 113 | /** |
| 114 | * \brief Constructor. |
| 115 | * |
| 116 | * \param[in] T the base of the matrix power. |
| 117 | * \param[in] p the exponent of the matrix power, should be in |
| 118 | * \f$ (-1, 1) \f$. |
| 119 | * |
| 120 | * The class stores a reference to T, so it should not be changed |
| 121 | * (or destroyed) before evaluation. Only the upper triangular |
| 122 | * part of T is read. |
| 123 | */ |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 124 | MatrixPowerAtomic(const MatrixType& T, RealScalar p); |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 125 | |
| 126 | /** |
| 127 | * \brief Compute the matrix power. |
| 128 | * |
| 129 | * \param[out] res \f$ A^p \f$ where A and p are specified in the |
| 130 | * constructor. |
| 131 | */ |
| 132 | void compute(ResultType& res) const; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 133 | }; |
| 134 | |
| 135 | template<typename MatrixType> |
| 136 | MatrixPowerAtomic<MatrixType>::MatrixPowerAtomic(const MatrixType& T, RealScalar p) : |
| 137 | m_A(T), m_p(p) |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 138 | { |
| 139 | eigen_assert(T.rows() == T.cols()); |
| 140 | eigen_assert(p > -1 && p < 1); |
| 141 | } |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 142 | |
| 143 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 144 | void MatrixPowerAtomic<MatrixType>::compute(ResultType& res) const |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 145 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 146 | using std::pow; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 147 | switch (m_A.rows()) { |
| 148 | case 0: |
| 149 | break; |
| 150 | case 1: |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 151 | res(0,0) = pow(m_A(0,0), m_p); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 152 | break; |
| 153 | case 2: |
| 154 | compute2x2(res, m_p); |
| 155 | break; |
| 156 | default: |
| 157 | computeBig(res); |
| 158 | } |
| 159 | } |
| 160 | |
| 161 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 162 | void MatrixPowerAtomic<MatrixType>::computePade(int degree, const MatrixType& IminusT, ResultType& res) const |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 163 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 164 | int i = 2*degree; |
| 165 | res = (m_p-degree) / (2*i-2) * IminusT; |
| 166 | |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 167 | for (--i; i; --i) { |
| 168 | res = (MatrixType::Identity(IminusT.rows(), IminusT.cols()) + res).template triangularView<Upper>() |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 169 | .solve((i==1 ? -m_p : i&1 ? (-m_p-i/2)/(2*i) : (m_p-i/2)/(2*i-2)) * IminusT).eval(); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 170 | } |
| 171 | res += MatrixType::Identity(IminusT.rows(), IminusT.cols()); |
| 172 | } |
| 173 | |
| 174 | // This function assumes that res has the correct size (see bug 614) |
| 175 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 176 | void MatrixPowerAtomic<MatrixType>::compute2x2(ResultType& res, RealScalar p) const |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 177 | { |
| 178 | using std::abs; |
| 179 | using std::pow; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 180 | res.coeffRef(0,0) = pow(m_A.coeff(0,0), p); |
| 181 | |
| 182 | for (Index i=1; i < m_A.cols(); ++i) { |
| 183 | res.coeffRef(i,i) = pow(m_A.coeff(i,i), p); |
| 184 | if (m_A.coeff(i-1,i-1) == m_A.coeff(i,i)) |
| 185 | res.coeffRef(i-1,i) = p * pow(m_A.coeff(i,i), p-1); |
| 186 | else if (2*abs(m_A.coeff(i-1,i-1)) < abs(m_A.coeff(i,i)) || 2*abs(m_A.coeff(i,i)) < abs(m_A.coeff(i-1,i-1))) |
| 187 | res.coeffRef(i-1,i) = (res.coeff(i,i)-res.coeff(i-1,i-1)) / (m_A.coeff(i,i)-m_A.coeff(i-1,i-1)); |
| 188 | else |
| 189 | res.coeffRef(i-1,i) = computeSuperDiag(m_A.coeff(i,i), m_A.coeff(i-1,i-1), p); |
| 190 | res.coeffRef(i-1,i) *= m_A.coeff(i-1,i); |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 195 | void MatrixPowerAtomic<MatrixType>::computeBig(ResultType& res) const |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 196 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 197 | using std::ldexp; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 198 | const int digits = std::numeric_limits<RealScalar>::digits; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 199 | const RealScalar maxNormForPade = digits <= 24? 4.3386528e-1L // single precision |
| 200 | : digits <= 53? 2.789358995219730e-1L // double precision |
| 201 | : digits <= 64? 2.4471944416607995472e-1L // extended precision |
| 202 | : digits <= 106? 1.1016843812851143391275867258512e-1L // double-double |
| 203 | : 9.134603732914548552537150753385375e-2L; // quadruple precision |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 204 | MatrixType IminusT, sqrtT, T = m_A.template triangularView<Upper>(); |
| 205 | RealScalar normIminusT; |
| 206 | int degree, degree2, numberOfSquareRoots = 0; |
| 207 | bool hasExtraSquareRoot = false; |
| 208 | |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 209 | for (Index i=0; i < m_A.cols(); ++i) |
| 210 | eigen_assert(m_A(i,i) != RealScalar(0)); |
| 211 | |
| 212 | while (true) { |
| 213 | IminusT = MatrixType::Identity(m_A.rows(), m_A.cols()) - T; |
| 214 | normIminusT = IminusT.cwiseAbs().colwise().sum().maxCoeff(); |
| 215 | if (normIminusT < maxNormForPade) { |
| 216 | degree = getPadeDegree(normIminusT); |
| 217 | degree2 = getPadeDegree(normIminusT/2); |
| 218 | if (degree - degree2 <= 1 || hasExtraSquareRoot) |
| 219 | break; |
| 220 | hasExtraSquareRoot = true; |
| 221 | } |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 222 | matrix_sqrt_triangular(T, sqrtT); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 223 | T = sqrtT.template triangularView<Upper>(); |
| 224 | ++numberOfSquareRoots; |
| 225 | } |
| 226 | computePade(degree, IminusT, res); |
| 227 | |
| 228 | for (; numberOfSquareRoots; --numberOfSquareRoots) { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 229 | compute2x2(res, ldexp(m_p, -numberOfSquareRoots)); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 230 | res = res.template triangularView<Upper>() * res; |
| 231 | } |
| 232 | compute2x2(res, m_p); |
| 233 | } |
| 234 | |
| 235 | template<typename MatrixType> |
| 236 | inline int MatrixPowerAtomic<MatrixType>::getPadeDegree(float normIminusT) |
| 237 | { |
| 238 | const float maxNormForPade[] = { 2.8064004e-1f /* degree = 3 */ , 4.3386528e-1f }; |
| 239 | int degree = 3; |
| 240 | for (; degree <= 4; ++degree) |
| 241 | if (normIminusT <= maxNormForPade[degree - 3]) |
| 242 | break; |
| 243 | return degree; |
| 244 | } |
| 245 | |
| 246 | template<typename MatrixType> |
| 247 | inline int MatrixPowerAtomic<MatrixType>::getPadeDegree(double normIminusT) |
| 248 | { |
| 249 | const double maxNormForPade[] = { 1.884160592658218e-2 /* degree = 3 */ , 6.038881904059573e-2, 1.239917516308172e-1, |
| 250 | 1.999045567181744e-1, 2.789358995219730e-1 }; |
| 251 | int degree = 3; |
| 252 | for (; degree <= 7; ++degree) |
| 253 | if (normIminusT <= maxNormForPade[degree - 3]) |
| 254 | break; |
| 255 | return degree; |
| 256 | } |
| 257 | |
| 258 | template<typename MatrixType> |
| 259 | inline int MatrixPowerAtomic<MatrixType>::getPadeDegree(long double normIminusT) |
| 260 | { |
| 261 | #if LDBL_MANT_DIG == 53 |
| 262 | const int maxPadeDegree = 7; |
| 263 | const double maxNormForPade[] = { 1.884160592658218e-2L /* degree = 3 */ , 6.038881904059573e-2L, 1.239917516308172e-1L, |
| 264 | 1.999045567181744e-1L, 2.789358995219730e-1L }; |
| 265 | #elif LDBL_MANT_DIG <= 64 |
| 266 | const int maxPadeDegree = 8; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 267 | const long double maxNormForPade[] = { 6.3854693117491799460e-3L /* degree = 3 */ , 2.6394893435456973676e-2L, |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 268 | 6.4216043030404063729e-2L, 1.1701165502926694307e-1L, 1.7904284231268670284e-1L, 2.4471944416607995472e-1L }; |
| 269 | #elif LDBL_MANT_DIG <= 106 |
| 270 | const int maxPadeDegree = 10; |
| 271 | const double maxNormForPade[] = { 1.0007161601787493236741409687186e-4L /* degree = 3 */ , |
| 272 | 1.0007161601787493236741409687186e-3L, 4.7069769360887572939882574746264e-3L, 1.3220386624169159689406653101695e-2L, |
| 273 | 2.8063482381631737920612944054906e-2L, 4.9625993951953473052385361085058e-2L, 7.7367040706027886224557538328171e-2L, |
| 274 | 1.1016843812851143391275867258512e-1L }; |
| 275 | #else |
| 276 | const int maxPadeDegree = 10; |
| 277 | const double maxNormForPade[] = { 5.524506147036624377378713555116378e-5L /* degree = 3 */ , |
| 278 | 6.640600568157479679823602193345995e-4L, 3.227716520106894279249709728084626e-3L, |
| 279 | 9.619593944683432960546978734646284e-3L, 2.134595382433742403911124458161147e-2L, |
| 280 | 3.908166513900489428442993794761185e-2L, 6.266780814639442865832535460550138e-2L, |
| 281 | 9.134603732914548552537150753385375e-2L }; |
| 282 | #endif |
| 283 | int degree = 3; |
| 284 | for (; degree <= maxPadeDegree; ++degree) |
| 285 | if (normIminusT <= maxNormForPade[degree - 3]) |
| 286 | break; |
| 287 | return degree; |
| 288 | } |
| 289 | |
| 290 | template<typename MatrixType> |
| 291 | inline typename MatrixPowerAtomic<MatrixType>::ComplexScalar |
| 292 | MatrixPowerAtomic<MatrixType>::computeSuperDiag(const ComplexScalar& curr, const ComplexScalar& prev, RealScalar p) |
| 293 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 294 | using std::ceil; |
| 295 | using std::exp; |
| 296 | using std::log; |
| 297 | using std::sinh; |
| 298 | |
| 299 | ComplexScalar logCurr = log(curr); |
| 300 | ComplexScalar logPrev = log(prev); |
| 301 | int unwindingNumber = ceil((numext::imag(logCurr - logPrev) - RealScalar(EIGEN_PI)) / RealScalar(2*EIGEN_PI)); |
| 302 | ComplexScalar w = numext::log1p((curr-prev)/prev)/RealScalar(2) + ComplexScalar(0, EIGEN_PI*unwindingNumber); |
| 303 | return RealScalar(2) * exp(RealScalar(0.5) * p * (logCurr + logPrev)) * sinh(p * w) / (curr - prev); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 304 | } |
| 305 | |
| 306 | template<typename MatrixType> |
| 307 | inline typename MatrixPowerAtomic<MatrixType>::RealScalar |
| 308 | MatrixPowerAtomic<MatrixType>::computeSuperDiag(RealScalar curr, RealScalar prev, RealScalar p) |
| 309 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 310 | using std::exp; |
| 311 | using std::log; |
| 312 | using std::sinh; |
| 313 | |
| 314 | RealScalar w = numext::log1p((curr-prev)/prev)/RealScalar(2); |
| 315 | return 2 * exp(p * (log(curr) + log(prev)) / 2) * sinh(p * w) / (curr - prev); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 316 | } |
| 317 | |
| 318 | /** |
| 319 | * \ingroup MatrixFunctions_Module |
| 320 | * |
| 321 | * \brief Class for computing matrix powers. |
| 322 | * |
| 323 | * \tparam MatrixType type of the base, expected to be an instantiation |
| 324 | * of the Matrix class template. |
| 325 | * |
| 326 | * This class is capable of computing real/complex matrices raised to |
| 327 | * an arbitrary real power. Meanwhile, it saves the result of Schur |
| 328 | * decomposition if an non-integral power has even been calculated. |
| 329 | * Therefore, if you want to compute multiple (>= 2) matrix powers |
| 330 | * for the same matrix, using the class directly is more efficient than |
| 331 | * calling MatrixBase::pow(). |
| 332 | * |
| 333 | * Example: |
| 334 | * \include MatrixPower_optimal.cpp |
| 335 | * Output: \verbinclude MatrixPower_optimal.out |
| 336 | */ |
| 337 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 338 | class MatrixPower : internal::noncopyable |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 339 | { |
| 340 | private: |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 341 | typedef typename MatrixType::Scalar Scalar; |
| 342 | typedef typename MatrixType::RealScalar RealScalar; |
| 343 | typedef typename MatrixType::Index Index; |
| 344 | |
| 345 | public: |
| 346 | /** |
| 347 | * \brief Constructor. |
| 348 | * |
| 349 | * \param[in] A the base of the matrix power. |
| 350 | * |
| 351 | * The class stores a reference to A, so it should not be changed |
| 352 | * (or destroyed) before evaluation. |
| 353 | */ |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 354 | explicit MatrixPower(const MatrixType& A) : |
| 355 | m_A(A), |
| 356 | m_conditionNumber(0), |
| 357 | m_rank(A.cols()), |
| 358 | m_nulls(0) |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 359 | { eigen_assert(A.rows() == A.cols()); } |
| 360 | |
| 361 | /** |
| 362 | * \brief Returns the matrix power. |
| 363 | * |
| 364 | * \param[in] p exponent, a real scalar. |
| 365 | * \return The expression \f$ A^p \f$, where A is specified in the |
| 366 | * constructor. |
| 367 | */ |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 368 | const MatrixPowerParenthesesReturnValue<MatrixType> operator()(RealScalar p) |
| 369 | { return MatrixPowerParenthesesReturnValue<MatrixType>(*this, p); } |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 370 | |
| 371 | /** |
| 372 | * \brief Compute the matrix power. |
| 373 | * |
| 374 | * \param[in] p exponent, a real scalar. |
| 375 | * \param[out] res \f$ A^p \f$ where A is specified in the |
| 376 | * constructor. |
| 377 | */ |
| 378 | template<typename ResultType> |
| 379 | void compute(ResultType& res, RealScalar p); |
| 380 | |
| 381 | Index rows() const { return m_A.rows(); } |
| 382 | Index cols() const { return m_A.cols(); } |
| 383 | |
| 384 | private: |
| 385 | typedef std::complex<RealScalar> ComplexScalar; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 386 | typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0, |
| 387 | MatrixType::RowsAtCompileTime, MatrixType::ColsAtCompileTime> ComplexMatrix; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 388 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 389 | /** \brief Reference to the base of matrix power. */ |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 390 | typename MatrixType::Nested m_A; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 391 | |
| 392 | /** \brief Temporary storage. */ |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 393 | MatrixType m_tmp; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 394 | |
| 395 | /** \brief Store the result of Schur decomposition. */ |
| 396 | ComplexMatrix m_T, m_U; |
| 397 | |
| 398 | /** \brief Store fractional power of m_T. */ |
| 399 | ComplexMatrix m_fT; |
| 400 | |
| 401 | /** |
| 402 | * \brief Condition number of m_A. |
| 403 | * |
| 404 | * It is initialized as 0 to avoid performing unnecessary Schur |
| 405 | * decomposition, which is the bottleneck. |
| 406 | */ |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 407 | RealScalar m_conditionNumber; |
| 408 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 409 | /** \brief Rank of m_A. */ |
| 410 | Index m_rank; |
| 411 | |
| 412 | /** \brief Rank deficiency of m_A. */ |
| 413 | Index m_nulls; |
| 414 | |
| 415 | /** |
| 416 | * \brief Split p into integral part and fractional part. |
| 417 | * |
| 418 | * \param[in] p The exponent. |
| 419 | * \param[out] p The fractional part ranging in \f$ (-1, 1) \f$. |
| 420 | * \param[out] intpart The integral part. |
| 421 | * |
| 422 | * Only if the fractional part is nonzero, it calls initialize(). |
| 423 | */ |
| 424 | void split(RealScalar& p, RealScalar& intpart); |
| 425 | |
| 426 | /** \brief Perform Schur decomposition for fractional power. */ |
| 427 | void initialize(); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 428 | |
| 429 | template<typename ResultType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 430 | void computeIntPower(ResultType& res, RealScalar p); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 431 | |
| 432 | template<typename ResultType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 433 | void computeFracPower(ResultType& res, RealScalar p); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 434 | |
| 435 | template<int Rows, int Cols, int Options, int MaxRows, int MaxCols> |
| 436 | static void revertSchur( |
| 437 | Matrix<ComplexScalar, Rows, Cols, Options, MaxRows, MaxCols>& res, |
| 438 | const ComplexMatrix& T, |
| 439 | const ComplexMatrix& U); |
| 440 | |
| 441 | template<int Rows, int Cols, int Options, int MaxRows, int MaxCols> |
| 442 | static void revertSchur( |
| 443 | Matrix<RealScalar, Rows, Cols, Options, MaxRows, MaxCols>& res, |
| 444 | const ComplexMatrix& T, |
| 445 | const ComplexMatrix& U); |
| 446 | }; |
| 447 | |
| 448 | template<typename MatrixType> |
| 449 | template<typename ResultType> |
| 450 | void MatrixPower<MatrixType>::compute(ResultType& res, RealScalar p) |
| 451 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 452 | using std::pow; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 453 | switch (cols()) { |
| 454 | case 0: |
| 455 | break; |
| 456 | case 1: |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 457 | res(0,0) = pow(m_A.coeff(0,0), p); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 458 | break; |
| 459 | default: |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 460 | RealScalar intpart; |
| 461 | split(p, intpart); |
| 462 | |
| 463 | res = MatrixType::Identity(rows(), cols()); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 464 | computeIntPower(res, intpart); |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 465 | if (p) computeFracPower(res, p); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 466 | } |
| 467 | } |
| 468 | |
| 469 | template<typename MatrixType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 470 | void MatrixPower<MatrixType>::split(RealScalar& p, RealScalar& intpart) |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 471 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 472 | using std::floor; |
| 473 | using std::pow; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 474 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 475 | intpart = floor(p); |
| 476 | p -= intpart; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 477 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 478 | // Perform Schur decomposition if it is not yet performed and the power is |
| 479 | // not an integer. |
| 480 | if (!m_conditionNumber && p) |
| 481 | initialize(); |
| 482 | |
| 483 | // Choose the more stable of intpart = floor(p) and intpart = ceil(p). |
| 484 | if (p > RealScalar(0.5) && p > (1-p) * pow(m_conditionNumber, p)) { |
| 485 | --p; |
| 486 | ++intpart; |
| 487 | } |
| 488 | } |
| 489 | |
| 490 | template<typename MatrixType> |
| 491 | void MatrixPower<MatrixType>::initialize() |
| 492 | { |
| 493 | const ComplexSchur<MatrixType> schurOfA(m_A); |
| 494 | JacobiRotation<ComplexScalar> rot; |
| 495 | ComplexScalar eigenvalue; |
| 496 | |
| 497 | m_fT.resizeLike(m_A); |
| 498 | m_T = schurOfA.matrixT(); |
| 499 | m_U = schurOfA.matrixU(); |
| 500 | m_conditionNumber = m_T.diagonal().array().abs().maxCoeff() / m_T.diagonal().array().abs().minCoeff(); |
| 501 | |
| 502 | // Move zero eigenvalues to the bottom right corner. |
| 503 | for (Index i = cols()-1; i>=0; --i) { |
| 504 | if (m_rank <= 2) |
| 505 | return; |
| 506 | if (m_T.coeff(i,i) == RealScalar(0)) { |
| 507 | for (Index j=i+1; j < m_rank; ++j) { |
| 508 | eigenvalue = m_T.coeff(j,j); |
| 509 | rot.makeGivens(m_T.coeff(j-1,j), eigenvalue); |
| 510 | m_T.applyOnTheRight(j-1, j, rot); |
| 511 | m_T.applyOnTheLeft(j-1, j, rot.adjoint()); |
| 512 | m_T.coeffRef(j-1,j-1) = eigenvalue; |
| 513 | m_T.coeffRef(j,j) = RealScalar(0); |
| 514 | m_U.applyOnTheRight(j-1, j, rot); |
| 515 | } |
| 516 | --m_rank; |
| 517 | } |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 518 | } |
| 519 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 520 | m_nulls = rows() - m_rank; |
| 521 | if (m_nulls) { |
| 522 | eigen_assert(m_T.bottomRightCorner(m_nulls, m_nulls).isZero() |
| 523 | && "Base of matrix power should be invertible or with a semisimple zero eigenvalue."); |
| 524 | m_fT.bottomRows(m_nulls).fill(RealScalar(0)); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 525 | } |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 526 | } |
| 527 | |
| 528 | template<typename MatrixType> |
| 529 | template<typename ResultType> |
| 530 | void MatrixPower<MatrixType>::computeIntPower(ResultType& res, RealScalar p) |
| 531 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 532 | using std::abs; |
| 533 | using std::fmod; |
| 534 | RealScalar pp = abs(p); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 535 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 536 | if (p<0) |
| 537 | m_tmp = m_A.inverse(); |
| 538 | else |
| 539 | m_tmp = m_A; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 540 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 541 | while (true) { |
| 542 | if (fmod(pp, 2) >= 1) |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 543 | res = m_tmp * res; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 544 | pp /= 2; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 545 | if (pp < 1) |
| 546 | break; |
| 547 | m_tmp *= m_tmp; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 548 | } |
| 549 | } |
| 550 | |
| 551 | template<typename MatrixType> |
| 552 | template<typename ResultType> |
| 553 | void MatrixPower<MatrixType>::computeFracPower(ResultType& res, RealScalar p) |
| 554 | { |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 555 | Block<ComplexMatrix,Dynamic,Dynamic> blockTp(m_fT, 0, 0, m_rank, m_rank); |
| 556 | eigen_assert(m_conditionNumber); |
| 557 | eigen_assert(m_rank + m_nulls == rows()); |
| 558 | |
| 559 | MatrixPowerAtomic<ComplexMatrix>(m_T.topLeftCorner(m_rank, m_rank), p).compute(blockTp); |
| 560 | if (m_nulls) { |
| 561 | m_fT.topRightCorner(m_rank, m_nulls) = m_T.topLeftCorner(m_rank, m_rank).template triangularView<Upper>() |
| 562 | .solve(blockTp * m_T.topRightCorner(m_rank, m_nulls)); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 563 | } |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 564 | revertSchur(m_tmp, m_fT, m_U); |
| 565 | res = m_tmp * res; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 566 | } |
| 567 | |
| 568 | template<typename MatrixType> |
| 569 | template<int Rows, int Cols, int Options, int MaxRows, int MaxCols> |
| 570 | inline void MatrixPower<MatrixType>::revertSchur( |
| 571 | Matrix<ComplexScalar, Rows, Cols, Options, MaxRows, MaxCols>& res, |
| 572 | const ComplexMatrix& T, |
| 573 | const ComplexMatrix& U) |
| 574 | { res.noalias() = U * (T.template triangularView<Upper>() * U.adjoint()); } |
| 575 | |
| 576 | template<typename MatrixType> |
| 577 | template<int Rows, int Cols, int Options, int MaxRows, int MaxCols> |
| 578 | inline void MatrixPower<MatrixType>::revertSchur( |
| 579 | Matrix<RealScalar, Rows, Cols, Options, MaxRows, MaxCols>& res, |
| 580 | const ComplexMatrix& T, |
| 581 | const ComplexMatrix& U) |
| 582 | { res.noalias() = (U * (T.template triangularView<Upper>() * U.adjoint())).real(); } |
| 583 | |
| 584 | /** |
| 585 | * \ingroup MatrixFunctions_Module |
| 586 | * |
| 587 | * \brief Proxy for the matrix power of some matrix (expression). |
| 588 | * |
| 589 | * \tparam Derived type of the base, a matrix (expression). |
| 590 | * |
| 591 | * This class holds the arguments to the matrix power until it is |
| 592 | * assigned or evaluated for some other reason (so the argument |
| 593 | * should not be changed in the meantime). It is the return type of |
| 594 | * MatrixBase::pow() and related functions and most of the |
| 595 | * time this is the only way it is used. |
| 596 | */ |
| 597 | template<typename Derived> |
| 598 | class MatrixPowerReturnValue : public ReturnByValue< MatrixPowerReturnValue<Derived> > |
| 599 | { |
| 600 | public: |
| 601 | typedef typename Derived::PlainObject PlainObject; |
| 602 | typedef typename Derived::RealScalar RealScalar; |
| 603 | typedef typename Derived::Index Index; |
| 604 | |
| 605 | /** |
| 606 | * \brief Constructor. |
| 607 | * |
| 608 | * \param[in] A %Matrix (expression), the base of the matrix power. |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 609 | * \param[in] p real scalar, the exponent of the matrix power. |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 610 | */ |
| 611 | MatrixPowerReturnValue(const Derived& A, RealScalar p) : m_A(A), m_p(p) |
| 612 | { } |
| 613 | |
| 614 | /** |
| 615 | * \brief Compute the matrix power. |
| 616 | * |
| 617 | * \param[out] result \f$ A^p \f$ where \p A and \p p are as in the |
| 618 | * constructor. |
| 619 | */ |
| 620 | template<typename ResultType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 621 | inline void evalTo(ResultType& result) const |
| 622 | { MatrixPower<PlainObject>(m_A.eval()).compute(result, m_p); } |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 623 | |
| 624 | Index rows() const { return m_A.rows(); } |
| 625 | Index cols() const { return m_A.cols(); } |
| 626 | |
| 627 | private: |
| 628 | const Derived& m_A; |
| 629 | const RealScalar m_p; |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 630 | }; |
| 631 | |
| 632 | /** |
| 633 | * \ingroup MatrixFunctions_Module |
| 634 | * |
| 635 | * \brief Proxy for the matrix power of some matrix (expression). |
| 636 | * |
| 637 | * \tparam Derived type of the base, a matrix (expression). |
| 638 | * |
| 639 | * This class holds the arguments to the matrix power until it is |
| 640 | * assigned or evaluated for some other reason (so the argument |
| 641 | * should not be changed in the meantime). It is the return type of |
| 642 | * MatrixBase::pow() and related functions and most of the |
| 643 | * time this is the only way it is used. |
| 644 | */ |
| 645 | template<typename Derived> |
| 646 | class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerReturnValue<Derived> > |
| 647 | { |
| 648 | public: |
| 649 | typedef typename Derived::PlainObject PlainObject; |
| 650 | typedef typename std::complex<typename Derived::RealScalar> ComplexScalar; |
| 651 | typedef typename Derived::Index Index; |
| 652 | |
| 653 | /** |
| 654 | * \brief Constructor. |
| 655 | * |
| 656 | * \param[in] A %Matrix (expression), the base of the matrix power. |
| 657 | * \param[in] p complex scalar, the exponent of the matrix power. |
| 658 | */ |
| 659 | MatrixComplexPowerReturnValue(const Derived& A, const ComplexScalar& p) : m_A(A), m_p(p) |
| 660 | { } |
| 661 | |
| 662 | /** |
| 663 | * \brief Compute the matrix power. |
| 664 | * |
| 665 | * Because \p p is complex, \f$ A^p \f$ is simply evaluated as \f$ |
| 666 | * \exp(p \log(A)) \f$. |
| 667 | * |
| 668 | * \param[out] result \f$ A^p \f$ where \p A and \p p are as in the |
| 669 | * constructor. |
| 670 | */ |
| 671 | template<typename ResultType> |
| 672 | inline void evalTo(ResultType& result) const |
| 673 | { result = (m_p * m_A.log()).exp(); } |
| 674 | |
| 675 | Index rows() const { return m_A.rows(); } |
| 676 | Index cols() const { return m_A.cols(); } |
| 677 | |
| 678 | private: |
| 679 | const Derived& m_A; |
| 680 | const ComplexScalar m_p; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 681 | }; |
| 682 | |
| 683 | namespace internal { |
| 684 | |
| 685 | template<typename MatrixPowerType> |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 686 | struct traits< MatrixPowerParenthesesReturnValue<MatrixPowerType> > |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 687 | { typedef typename MatrixPowerType::PlainObject ReturnType; }; |
| 688 | |
| 689 | template<typename Derived> |
| 690 | struct traits< MatrixPowerReturnValue<Derived> > |
| 691 | { typedef typename Derived::PlainObject ReturnType; }; |
| 692 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 693 | template<typename Derived> |
| 694 | struct traits< MatrixComplexPowerReturnValue<Derived> > |
| 695 | { typedef typename Derived::PlainObject ReturnType; }; |
| 696 | |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 697 | } |
| 698 | |
| 699 | template<typename Derived> |
| 700 | const MatrixPowerReturnValue<Derived> MatrixBase<Derived>::pow(const RealScalar& p) const |
| 701 | { return MatrixPowerReturnValue<Derived>(derived(), p); } |
| 702 | |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame^] | 703 | template<typename Derived> |
| 704 | const MatrixComplexPowerReturnValue<Derived> MatrixBase<Derived>::pow(const std::complex<RealScalar>& p) const |
| 705 | { return MatrixComplexPowerReturnValue<Derived>(derived(), p); } |
| 706 | |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 707 | } // namespace Eigen |
| 708 | |
| 709 | #endif // EIGEN_MATRIX_POWER |