blob: e363e779dea5da37ff6aab6b4b61a5ac6c5bd205 [file] [log] [blame]
Brian Silverman72890c22015-09-19 14:37:37 -04001// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
Austin Schuh189376f2018-12-20 22:11:15 +11004// Copyright (C) 2011, 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
Brian Silverman72890c22015-09-19 14:37:37 -04005//
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_SQUARE_ROOT
11#define EIGEN_MATRIX_SQUARE_ROOT
12
13namespace Eigen {
14
Austin Schuh189376f2018-12-20 22:11:15 +110015namespace internal {
Brian Silverman72890c22015-09-19 14:37:37 -040016
17// pre: T.block(i,i,2,2) has complex conjugate eigenvalues
18// post: sqrtT.block(i,i,2,2) is square root of T.block(i,i,2,2)
Austin Schuh189376f2018-12-20 22:11:15 +110019template <typename MatrixType, typename ResultType>
Austin Schuhc55b0172022-02-20 17:52:35 -080020void matrix_sqrt_quasi_triangular_2x2_diagonal_block(const MatrixType& T, Index i, ResultType& sqrtT)
Brian Silverman72890c22015-09-19 14:37:37 -040021{
22 // TODO: This case (2-by-2 blocks with complex conjugate eigenvalues) is probably hidden somewhere
23 // in EigenSolver. If we expose it, we could call it directly from here.
Austin Schuh189376f2018-12-20 22:11:15 +110024 typedef typename traits<MatrixType>::Scalar Scalar;
Brian Silverman72890c22015-09-19 14:37:37 -040025 Matrix<Scalar,2,2> block = T.template block<2,2>(i,i);
26 EigenSolver<Matrix<Scalar,2,2> > es(block);
27 sqrtT.template block<2,2>(i,i)
28 = (es.eigenvectors() * es.eigenvalues().cwiseSqrt().asDiagonal() * es.eigenvectors().inverse()).real();
29}
30
31// pre: block structure of T is such that (i,j) is a 1x1 block,
32// all blocks of sqrtT to left of and below (i,j) are correct
33// post: sqrtT(i,j) has the correct value
Austin Schuh189376f2018-12-20 22:11:15 +110034template <typename MatrixType, typename ResultType>
Austin Schuhc55b0172022-02-20 17:52:35 -080035void matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
Brian Silverman72890c22015-09-19 14:37:37 -040036{
Austin Schuh189376f2018-12-20 22:11:15 +110037 typedef typename traits<MatrixType>::Scalar Scalar;
Brian Silverman72890c22015-09-19 14:37:37 -040038 Scalar tmp = (sqrtT.row(i).segment(i+1,j-i-1) * sqrtT.col(j).segment(i+1,j-i-1)).value();
39 sqrtT.coeffRef(i,j) = (T.coeff(i,j) - tmp) / (sqrtT.coeff(i,i) + sqrtT.coeff(j,j));
40}
41
42// similar to compute1x1offDiagonalBlock()
Austin Schuh189376f2018-12-20 22:11:15 +110043template <typename MatrixType, typename ResultType>
Austin Schuhc55b0172022-02-20 17:52:35 -080044void matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
Brian Silverman72890c22015-09-19 14:37:37 -040045{
Austin Schuh189376f2018-12-20 22:11:15 +110046 typedef typename traits<MatrixType>::Scalar Scalar;
Brian Silverman72890c22015-09-19 14:37:37 -040047 Matrix<Scalar,1,2> rhs = T.template block<1,2>(i,j);
48 if (j-i > 1)
49 rhs -= sqrtT.block(i, i+1, 1, j-i-1) * sqrtT.block(i+1, j, j-i-1, 2);
50 Matrix<Scalar,2,2> A = sqrtT.coeff(i,i) * Matrix<Scalar,2,2>::Identity();
51 A += sqrtT.template block<2,2>(j,j).transpose();
52 sqrtT.template block<1,2>(i,j).transpose() = A.fullPivLu().solve(rhs.transpose());
53}
54
55// similar to compute1x1offDiagonalBlock()
Austin Schuh189376f2018-12-20 22:11:15 +110056template <typename MatrixType, typename ResultType>
Austin Schuhc55b0172022-02-20 17:52:35 -080057void matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
Brian Silverman72890c22015-09-19 14:37:37 -040058{
Austin Schuh189376f2018-12-20 22:11:15 +110059 typedef typename traits<MatrixType>::Scalar Scalar;
Brian Silverman72890c22015-09-19 14:37:37 -040060 Matrix<Scalar,2,1> rhs = T.template block<2,1>(i,j);
61 if (j-i > 2)
62 rhs -= sqrtT.block(i, i+2, 2, j-i-2) * sqrtT.block(i+2, j, j-i-2, 1);
63 Matrix<Scalar,2,2> A = sqrtT.coeff(j,j) * Matrix<Scalar,2,2>::Identity();
64 A += sqrtT.template block<2,2>(i,i);
65 sqrtT.template block<2,1>(i,j) = A.fullPivLu().solve(rhs);
66}
67
Brian Silverman72890c22015-09-19 14:37:37 -040068// solves the equation A X + X B = C where all matrices are 2-by-2
69template <typename MatrixType>
Austin Schuh189376f2018-12-20 22:11:15 +110070void matrix_sqrt_quasi_triangular_solve_auxiliary_equation(MatrixType& X, const MatrixType& A, const MatrixType& B, const MatrixType& C)
Brian Silverman72890c22015-09-19 14:37:37 -040071{
Austin Schuh189376f2018-12-20 22:11:15 +110072 typedef typename traits<MatrixType>::Scalar Scalar;
Brian Silverman72890c22015-09-19 14:37:37 -040073 Matrix<Scalar,4,4> coeffMatrix = Matrix<Scalar,4,4>::Zero();
74 coeffMatrix.coeffRef(0,0) = A.coeff(0,0) + B.coeff(0,0);
75 coeffMatrix.coeffRef(1,1) = A.coeff(0,0) + B.coeff(1,1);
76 coeffMatrix.coeffRef(2,2) = A.coeff(1,1) + B.coeff(0,0);
77 coeffMatrix.coeffRef(3,3) = A.coeff(1,1) + B.coeff(1,1);
78 coeffMatrix.coeffRef(0,1) = B.coeff(1,0);
79 coeffMatrix.coeffRef(0,2) = A.coeff(0,1);
80 coeffMatrix.coeffRef(1,0) = B.coeff(0,1);
81 coeffMatrix.coeffRef(1,3) = A.coeff(0,1);
82 coeffMatrix.coeffRef(2,0) = A.coeff(1,0);
83 coeffMatrix.coeffRef(2,3) = B.coeff(1,0);
84 coeffMatrix.coeffRef(3,1) = A.coeff(1,0);
85 coeffMatrix.coeffRef(3,2) = B.coeff(0,1);
Austin Schuh189376f2018-12-20 22:11:15 +110086
Brian Silverman72890c22015-09-19 14:37:37 -040087 Matrix<Scalar,4,1> rhs;
88 rhs.coeffRef(0) = C.coeff(0,0);
89 rhs.coeffRef(1) = C.coeff(0,1);
90 rhs.coeffRef(2) = C.coeff(1,0);
91 rhs.coeffRef(3) = C.coeff(1,1);
Austin Schuh189376f2018-12-20 22:11:15 +110092
Brian Silverman72890c22015-09-19 14:37:37 -040093 Matrix<Scalar,4,1> result;
94 result = coeffMatrix.fullPivLu().solve(rhs);
95
96 X.coeffRef(0,0) = result.coeff(0);
97 X.coeffRef(0,1) = result.coeff(1);
98 X.coeffRef(1,0) = result.coeff(2);
99 X.coeffRef(1,1) = result.coeff(3);
100}
101
Austin Schuh189376f2018-12-20 22:11:15 +1100102// similar to compute1x1offDiagonalBlock()
103template <typename MatrixType, typename ResultType>
Austin Schuhc55b0172022-02-20 17:52:35 -0800104void matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
Austin Schuh189376f2018-12-20 22:11:15 +1100105{
106 typedef typename traits<MatrixType>::Scalar Scalar;
107 Matrix<Scalar,2,2> A = sqrtT.template block<2,2>(i,i);
108 Matrix<Scalar,2,2> B = sqrtT.template block<2,2>(j,j);
109 Matrix<Scalar,2,2> C = T.template block<2,2>(i,j);
110 if (j-i > 2)
111 C -= sqrtT.block(i, i+2, 2, j-i-2) * sqrtT.block(i+2, j, j-i-2, 2);
112 Matrix<Scalar,2,2> X;
113 matrix_sqrt_quasi_triangular_solve_auxiliary_equation(X, A, B, C);
114 sqrtT.template block<2,2>(i,j) = X;
115}
116
117// pre: T is quasi-upper-triangular and sqrtT is a zero matrix of the same size
118// post: the diagonal blocks of sqrtT are the square roots of the diagonal blocks of T
119template <typename MatrixType, typename ResultType>
120void matrix_sqrt_quasi_triangular_diagonal(const MatrixType& T, ResultType& sqrtT)
121{
122 using std::sqrt;
123 const Index size = T.rows();
124 for (Index i = 0; i < size; i++) {
125 if (i == size - 1 || T.coeff(i+1, i) == 0) {
126 eigen_assert(T(i,i) >= 0);
127 sqrtT.coeffRef(i,i) = sqrt(T.coeff(i,i));
128 }
129 else {
130 matrix_sqrt_quasi_triangular_2x2_diagonal_block(T, i, sqrtT);
131 ++i;
132 }
133 }
134}
135
136// pre: T is quasi-upper-triangular and diagonal blocks of sqrtT are square root of diagonal blocks of T.
137// post: sqrtT is the square root of T.
138template <typename MatrixType, typename ResultType>
139void matrix_sqrt_quasi_triangular_off_diagonal(const MatrixType& T, ResultType& sqrtT)
140{
141 const Index size = T.rows();
142 for (Index j = 1; j < size; j++) {
143 if (T.coeff(j, j-1) != 0) // if T(j-1:j, j-1:j) is a 2-by-2 block
144 continue;
145 for (Index i = j-1; i >= 0; i--) {
146 if (i > 0 && T.coeff(i, i-1) != 0) // if T(i-1:i, i-1:i) is a 2-by-2 block
147 continue;
148 bool iBlockIs2x2 = (i < size - 1) && (T.coeff(i+1, i) != 0);
149 bool jBlockIs2x2 = (j < size - 1) && (T.coeff(j+1, j) != 0);
150 if (iBlockIs2x2 && jBlockIs2x2)
151 matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(T, i, j, sqrtT);
152 else if (iBlockIs2x2 && !jBlockIs2x2)
153 matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(T, i, j, sqrtT);
154 else if (!iBlockIs2x2 && jBlockIs2x2)
155 matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(T, i, j, sqrtT);
156 else if (!iBlockIs2x2 && !jBlockIs2x2)
157 matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(T, i, j, sqrtT);
158 }
159 }
160}
161
162} // end of namespace internal
Brian Silverman72890c22015-09-19 14:37:37 -0400163
164/** \ingroup MatrixFunctions_Module
Austin Schuh189376f2018-12-20 22:11:15 +1100165 * \brief Compute matrix square root of quasi-triangular matrix.
Brian Silverman72890c22015-09-19 14:37:37 -0400166 *
Austin Schuh189376f2018-12-20 22:11:15 +1100167 * \tparam MatrixType type of \p arg, the argument of matrix square root,
168 * expected to be an instantiation of the Matrix class template.
169 * \tparam ResultType type of \p result, where result is to be stored.
170 * \param[in] arg argument of matrix square root.
171 * \param[out] result matrix square root of upper Hessenberg part of \p arg.
172 *
173 * This function computes the square root of the upper quasi-triangular matrix stored in the upper
174 * Hessenberg part of \p arg. Only the upper Hessenberg part of \p result is updated, the rest is
175 * not touched. See MatrixBase::sqrt() for details on how this computation is implemented.
Brian Silverman72890c22015-09-19 14:37:37 -0400176 *
177 * \sa MatrixSquareRoot, MatrixSquareRootQuasiTriangular
178 */
Austin Schuh189376f2018-12-20 22:11:15 +1100179template <typename MatrixType, typename ResultType>
180void matrix_sqrt_quasi_triangular(const MatrixType &arg, ResultType &result)
Brian Silverman72890c22015-09-19 14:37:37 -0400181{
Austin Schuh189376f2018-12-20 22:11:15 +1100182 eigen_assert(arg.rows() == arg.cols());
183 result.resize(arg.rows(), arg.cols());
184 internal::matrix_sqrt_quasi_triangular_diagonal(arg, result);
185 internal::matrix_sqrt_quasi_triangular_off_diagonal(arg, result);
186}
Brian Silverman72890c22015-09-19 14:37:37 -0400187
Brian Silverman72890c22015-09-19 14:37:37 -0400188
Austin Schuh189376f2018-12-20 22:11:15 +1100189/** \ingroup MatrixFunctions_Module
190 * \brief Compute matrix square root of triangular matrix.
191 *
192 * \tparam MatrixType type of \p arg, the argument of matrix square root,
193 * expected to be an instantiation of the Matrix class template.
194 * \tparam ResultType type of \p result, where result is to be stored.
195 * \param[in] arg argument of matrix square root.
196 * \param[out] result matrix square root of upper triangular part of \p arg.
197 *
198 * Only the upper triangular part (including the diagonal) of \p result is updated, the rest is not
199 * touched. See MatrixBase::sqrt() for details on how this computation is implemented.
200 *
201 * \sa MatrixSquareRoot, MatrixSquareRootQuasiTriangular
202 */
203template <typename MatrixType, typename ResultType>
204void matrix_sqrt_triangular(const MatrixType &arg, ResultType &result)
Brian Silverman72890c22015-09-19 14:37:37 -0400205{
206 using std::sqrt;
Austin Schuhc55b0172022-02-20 17:52:35 -0800207 typedef typename MatrixType::Scalar Scalar;
Austin Schuh189376f2018-12-20 22:11:15 +1100208
209 eigen_assert(arg.rows() == arg.cols());
210
211 // Compute square root of arg and store it in upper triangular part of result
212 // This uses that the square root of triangular matrices can be computed directly.
213 result.resize(arg.rows(), arg.cols());
214 for (Index i = 0; i < arg.rows(); i++) {
215 result.coeffRef(i,i) = sqrt(arg.coeff(i,i));
216 }
217 for (Index j = 1; j < arg.cols(); j++) {
218 for (Index i = j-1; i >= 0; i--) {
Brian Silverman72890c22015-09-19 14:37:37 -0400219 // if i = j-1, then segment has length 0 so tmp = 0
220 Scalar tmp = (result.row(i).segment(i+1,j-i-1) * result.col(j).segment(i+1,j-i-1)).value();
221 // denominator may be zero if original matrix is singular
Austin Schuh189376f2018-12-20 22:11:15 +1100222 result.coeffRef(i,j) = (arg.coeff(i,j) - tmp) / (result.coeff(i,i) + result.coeff(j,j));
Brian Silverman72890c22015-09-19 14:37:37 -0400223 }
224 }
225}
226
227
Austin Schuh189376f2018-12-20 22:11:15 +1100228namespace internal {
229
Brian Silverman72890c22015-09-19 14:37:37 -0400230/** \ingroup MatrixFunctions_Module
Austin Schuh189376f2018-12-20 22:11:15 +1100231 * \brief Helper struct for computing matrix square roots of general matrices.
Brian Silverman72890c22015-09-19 14:37:37 -0400232 * \tparam MatrixType type of the argument of the matrix square root,
233 * expected to be an instantiation of the Matrix class template.
234 *
235 * \sa MatrixSquareRootTriangular, MatrixSquareRootQuasiTriangular, MatrixBase::sqrt()
236 */
237template <typename MatrixType, int IsComplex = NumTraits<typename internal::traits<MatrixType>::Scalar>::IsComplex>
Austin Schuh189376f2018-12-20 22:11:15 +1100238struct matrix_sqrt_compute
Brian Silverman72890c22015-09-19 14:37:37 -0400239{
Austin Schuh189376f2018-12-20 22:11:15 +1100240 /** \brief Compute the matrix square root
241 *
242 * \param[in] arg matrix whose square root is to be computed.
243 * \param[out] result square root of \p arg.
244 *
245 * See MatrixBase::sqrt() for details on how this computation is implemented.
246 */
247 template <typename ResultType> static void run(const MatrixType &arg, ResultType &result);
Brian Silverman72890c22015-09-19 14:37:37 -0400248};
249
250
251// ********** Partial specialization for real matrices **********
252
253template <typename MatrixType>
Austin Schuh189376f2018-12-20 22:11:15 +1100254struct matrix_sqrt_compute<MatrixType, 0>
Brian Silverman72890c22015-09-19 14:37:37 -0400255{
Austin Schuhc55b0172022-02-20 17:52:35 -0800256 typedef typename MatrixType::PlainObject PlainType;
Austin Schuh189376f2018-12-20 22:11:15 +1100257 template <typename ResultType>
258 static void run(const MatrixType &arg, ResultType &result)
259 {
260 eigen_assert(arg.rows() == arg.cols());
Brian Silverman72890c22015-09-19 14:37:37 -0400261
Austin Schuh189376f2018-12-20 22:11:15 +1100262 // Compute Schur decomposition of arg
Austin Schuhc55b0172022-02-20 17:52:35 -0800263 const RealSchur<PlainType> schurOfA(arg);
264 const PlainType& T = schurOfA.matrixT();
265 const PlainType& U = schurOfA.matrixU();
Brian Silverman72890c22015-09-19 14:37:37 -0400266
Austin Schuh189376f2018-12-20 22:11:15 +1100267 // Compute square root of T
Austin Schuhc55b0172022-02-20 17:52:35 -0800268 PlainType sqrtT = PlainType::Zero(arg.rows(), arg.cols());
Austin Schuh189376f2018-12-20 22:11:15 +1100269 matrix_sqrt_quasi_triangular(T, sqrtT);
Brian Silverman72890c22015-09-19 14:37:37 -0400270
Austin Schuh189376f2018-12-20 22:11:15 +1100271 // Compute square root of arg
272 result = U * sqrtT * U.adjoint();
273 }
Brian Silverman72890c22015-09-19 14:37:37 -0400274};
275
276
277// ********** Partial specialization for complex matrices **********
278
279template <typename MatrixType>
Austin Schuh189376f2018-12-20 22:11:15 +1100280struct matrix_sqrt_compute<MatrixType, 1>
Brian Silverman72890c22015-09-19 14:37:37 -0400281{
Austin Schuhc55b0172022-02-20 17:52:35 -0800282 typedef typename MatrixType::PlainObject PlainType;
Austin Schuh189376f2018-12-20 22:11:15 +1100283 template <typename ResultType>
284 static void run(const MatrixType &arg, ResultType &result)
285 {
286 eigen_assert(arg.rows() == arg.cols());
Brian Silverman72890c22015-09-19 14:37:37 -0400287
Austin Schuh189376f2018-12-20 22:11:15 +1100288 // Compute Schur decomposition of arg
Austin Schuhc55b0172022-02-20 17:52:35 -0800289 const ComplexSchur<PlainType> schurOfA(arg);
290 const PlainType& T = schurOfA.matrixT();
291 const PlainType& U = schurOfA.matrixU();
Brian Silverman72890c22015-09-19 14:37:37 -0400292
Austin Schuh189376f2018-12-20 22:11:15 +1100293 // Compute square root of T
Austin Schuhc55b0172022-02-20 17:52:35 -0800294 PlainType sqrtT;
Austin Schuh189376f2018-12-20 22:11:15 +1100295 matrix_sqrt_triangular(T, sqrtT);
Brian Silverman72890c22015-09-19 14:37:37 -0400296
Austin Schuh189376f2018-12-20 22:11:15 +1100297 // Compute square root of arg
298 result = U * (sqrtT.template triangularView<Upper>() * U.adjoint());
299 }
Brian Silverman72890c22015-09-19 14:37:37 -0400300};
301
Austin Schuh189376f2018-12-20 22:11:15 +1100302} // end namespace internal
Brian Silverman72890c22015-09-19 14:37:37 -0400303
304/** \ingroup MatrixFunctions_Module
305 *
306 * \brief Proxy for the matrix square root of some matrix (expression).
307 *
308 * \tparam Derived Type of the argument to the matrix square root.
309 *
310 * This class holds the argument to the matrix square root until it
311 * is assigned or evaluated for some other reason (so the argument
312 * should not be changed in the meantime). It is the return type of
313 * MatrixBase::sqrt() and most of the time this is the only way it is
314 * used.
315 */
316template<typename Derived> class MatrixSquareRootReturnValue
317: public ReturnByValue<MatrixSquareRootReturnValue<Derived> >
318{
Austin Schuh189376f2018-12-20 22:11:15 +1100319 protected:
320 typedef typename internal::ref_selector<Derived>::type DerivedNested;
321
Brian Silverman72890c22015-09-19 14:37:37 -0400322 public:
323 /** \brief Constructor.
324 *
325 * \param[in] src %Matrix (expression) forming the argument of the
326 * matrix square root.
327 */
Austin Schuh189376f2018-12-20 22:11:15 +1100328 explicit MatrixSquareRootReturnValue(const Derived& src) : m_src(src) { }
Brian Silverman72890c22015-09-19 14:37:37 -0400329
330 /** \brief Compute the matrix square root.
331 *
332 * \param[out] result the matrix square root of \p src in the
333 * constructor.
334 */
335 template <typename ResultType>
336 inline void evalTo(ResultType& result) const
337 {
Austin Schuh189376f2018-12-20 22:11:15 +1100338 typedef typename internal::nested_eval<Derived, 10>::type DerivedEvalType;
339 typedef typename internal::remove_all<DerivedEvalType>::type DerivedEvalTypeClean;
340 DerivedEvalType tmp(m_src);
341 internal::matrix_sqrt_compute<DerivedEvalTypeClean>::run(tmp, result);
Brian Silverman72890c22015-09-19 14:37:37 -0400342 }
343
344 Index rows() const { return m_src.rows(); }
345 Index cols() const { return m_src.cols(); }
346
347 protected:
Austin Schuh189376f2018-12-20 22:11:15 +1100348 const DerivedNested m_src;
Brian Silverman72890c22015-09-19 14:37:37 -0400349};
350
351namespace internal {
352template<typename Derived>
353struct traits<MatrixSquareRootReturnValue<Derived> >
354{
355 typedef typename Derived::PlainObject ReturnType;
356};
357}
358
359template <typename Derived>
360const MatrixSquareRootReturnValue<Derived> MatrixBase<Derived>::sqrt() const
361{
362 eigen_assert(rows() == cols());
363 return MatrixSquareRootReturnValue<Derived>(derived());
364}
365
366} // end namespace Eigen
367
368#endif // EIGEN_MATRIX_FUNCTION