// This file is part of Eigen, a lightweight C++ template library // for linear algebra. // // Copyright (C) 2008 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. #ifndef EIGEN2_SVD_H #define EIGEN2_SVD_H namespace Eigen { /** \ingroup SVD_Module * \nonstableyet * * \class SVD * * \brief Standard SVD decomposition of a matrix and associated features * * \param MatrixType the type of the matrix of which we are computing the SVD decomposition * * This class performs a standard SVD decomposition of a real matrix A of size \c M x \c N * with \c M \>= \c N. * * * \sa MatrixBase::SVD() */ template class SVD { private: typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; enum { PacketSize = internal::packet_traits::size, AlignmentMask = int(PacketSize)-1, MinSize = EIGEN_SIZE_MIN_PREFER_DYNAMIC(MatrixType::RowsAtCompileTime, MatrixType::ColsAtCompileTime) }; typedef Matrix ColVector; typedef Matrix RowVector; typedef Matrix MatrixUType; typedef Matrix MatrixVType; typedef Matrix SingularValuesType; public: SVD() {} // a user who relied on compiler-generated default compiler reported problems with MSVC in 2.0.7 SVD(const MatrixType& matrix) : m_matU(matrix.rows(), (std::min)(matrix.rows(), matrix.cols())), m_matV(matrix.cols(),matrix.cols()), m_sigma((std::min)(matrix.rows(),matrix.cols())) { compute(matrix); } template bool solve(const MatrixBase &b, ResultType* result) const; const MatrixUType& matrixU() const { return m_matU; } const SingularValuesType& singularValues() const { return m_sigma; } const MatrixVType& matrixV() const { return m_matV; } void compute(const MatrixType& matrix); SVD& sort(); template void computeUnitaryPositive(UnitaryType *unitary, PositiveType *positive) const; template void computePositiveUnitary(PositiveType *positive, UnitaryType *unitary) const; template void computeRotationScaling(RotationType *unitary, ScalingType *positive) const; template void computeScalingRotation(ScalingType *positive, RotationType *unitary) const; protected: /** \internal */ MatrixUType m_matU; /** \internal */ MatrixVType m_matV; /** \internal */ SingularValuesType m_sigma; }; /** Computes / recomputes the SVD decomposition A = U S V^* of \a matrix * * \note this code has been adapted from JAMA (public domain) */ template void SVD::compute(const MatrixType& matrix) { const int m = matrix.rows(); const int n = matrix.cols(); const int nu = (std::min)(m,n); ei_assert(m>=n && "In Eigen 2.0, SVD only works for MxN matrices with M>=N. Sorry!"); ei_assert(m>1 && "In Eigen 2.0, SVD doesn't work on 1x1 matrices"); m_matU.resize(m, nu); m_matU.setZero(); m_sigma.resize((std::min)(m,n)); m_matV.resize(n,n); RowVector e(n); ColVector work(m); MatrixType matA(matrix); const bool wantu = true; const bool wantv = true; int i=0, j=0, k=0; // Reduce A to bidiagonal form, storing the diagonal elements // in s and the super-diagonal elements in e. int nct = (std::min)(m-1,n); int nrt = (std::max)(0,(std::min)(n-2,m)); for (k = 0; k < (std::max)(nct,nrt); ++k) { if (k < nct) { // Compute the transformation for the k-th column and // place the k-th diagonal in m_sigma[k]. m_sigma[k] = matA.col(k).end(m-k).norm(); if (m_sigma[k] != 0.0) // FIXME { if (matA(k,k) < 0.0) m_sigma[k] = -m_sigma[k]; matA.col(k).end(m-k) /= m_sigma[k]; matA(k,k) += 1.0; } m_sigma[k] = -m_sigma[k]; } for (j = k+1; j < n; ++j) { if ((k < nct) && (m_sigma[k] != 0.0)) { // Apply the transformation. Scalar t = matA.col(k).end(m-k).eigen2_dot(matA.col(j).end(m-k)); // FIXME dot product or cwise prod + .sum() ?? t = -t/matA(k,k); matA.col(j).end(m-k) += t * matA.col(k).end(m-k); } // Place the k-th row of A into e for the // subsequent calculation of the row transformation. e[j] = matA(k,j); } // Place the transformation in U for subsequent back multiplication. if (wantu & (k < nct)) m_matU.col(k).end(m-k) = matA.col(k).end(m-k); if (k < nrt) { // Compute the k-th row transformation and place the // k-th super-diagonal in e[k]. e[k] = e.end(n-k-1).norm(); if (e[k] != 0.0) { if (e[k+1] < 0.0) e[k] = -e[k]; e.end(n-k-1) /= e[k]; e[k+1] += 1.0; } e[k] = -e[k]; if ((k+1 < m) & (e[k] != 0.0)) { // Apply the transformation. work.end(m-k-1) = matA.corner(BottomRight,m-k-1,n-k-1) * e.end(n-k-1); for (j = k+1; j < n; ++j) matA.col(j).end(m-k-1) += (-e[j]/e[k+1]) * work.end(m-k-1); } // Place the transformation in V for subsequent back multiplication. if (wantv) m_matV.col(k).end(n-k-1) = e.end(n-k-1); } } // Set up the final bidiagonal matrix or order p. int p = (std::min)(n,m+1); if (nct < n) m_sigma[nct] = matA(nct,nct); if (m < p) m_sigma[p-1] = 0.0; if (nrt+1 < p) e[nrt] = matA(nrt,p-1); e[p-1] = 0.0; // If required, generate U. if (wantu) { for (j = nct; j < nu; ++j) { m_matU.col(j).setZero(); m_matU(j,j) = 1.0; } for (k = nct-1; k >= 0; k--) { if (m_sigma[k] != 0.0) { for (j = k+1; j < nu; ++j) { Scalar t = m_matU.col(k).end(m-k).eigen2_dot(m_matU.col(j).end(m-k)); // FIXME is it really a dot product we want ? t = -t/m_matU(k,k); m_matU.col(j).end(m-k) += t * m_matU.col(k).end(m-k); } m_matU.col(k).end(m-k) = - m_matU.col(k).end(m-k); m_matU(k,k) = Scalar(1) + m_matU(k,k); if (k-1>0) m_matU.col(k).start(k-1).setZero(); } else { m_matU.col(k).setZero(); m_matU(k,k) = 1.0; } } } // If required, generate V. if (wantv) { for (k = n-1; k >= 0; k--) { if ((k < nrt) & (e[k] != 0.0)) { for (j = k+1; j < nu; ++j) { Scalar t = m_matV.col(k).end(n-k-1).eigen2_dot(m_matV.col(j).end(n-k-1)); // FIXME is it really a dot product we want ? t = -t/m_matV(k+1,k); m_matV.col(j).end(n-k-1) += t * m_matV.col(k).end(n-k-1); } } m_matV.col(k).setZero(); m_matV(k,k) = 1.0; } } // Main iteration loop for the singular values. int pp = p-1; int iter = 0; Scalar eps = ei_pow(Scalar(2),ei_is_same_type::ret ? Scalar(-23) : Scalar(-52)); while (p > 0) { int k=0; int kase=0; // Here is where a test for too many iterations would go. // This section of the program inspects for // negligible elements in the s and e arrays. On // completion the variables kase and k are set as follows. // kase = 1 if s(p) and e[k-1] are negligible and k

= -1; --k) { if (k == -1) break; if (ei_abs(e[k]) <= eps*(ei_abs(m_sigma[k]) + ei_abs(m_sigma[k+1]))) { e[k] = 0.0; break; } } if (k == p-2) { kase = 4; } else { int ks; for (ks = p-1; ks >= k; --ks) { if (ks == k) break; Scalar t = (ks != p ? ei_abs(e[ks]) : Scalar(0)) + (ks != k+1 ? ei_abs(e[ks-1]) : Scalar(0)); if (ei_abs(m_sigma[ks]) <= eps*t) { m_sigma[ks] = 0.0; break; } } if (ks == k) { kase = 3; } else if (ks == p-1) { kase = 1; } else { kase = 2; k = ks; } } ++k; // Perform the task indicated by kase. switch (kase) { // Deflate negligible s(p). case 1: { Scalar f(e[p-2]); e[p-2] = 0.0; for (j = p-2; j >= k; --j) { Scalar t(numext::hypot(m_sigma[j],f)); Scalar cs(m_sigma[j]/t); Scalar sn(f/t); m_sigma[j] = t; if (j != k) { f = -sn*e[j-1]; e[j-1] = cs*e[j-1]; } if (wantv) { for (i = 0; i < n; ++i) { t = cs*m_matV(i,j) + sn*m_matV(i,p-1); m_matV(i,p-1) = -sn*m_matV(i,j) + cs*m_matV(i,p-1); m_matV(i,j) = t; } } } } break; // Split at negligible s(k). case 2: { Scalar f(e[k-1]); e[k-1] = 0.0; for (j = k; j < p; ++j) { Scalar t(numext::hypot(m_sigma[j],f)); Scalar cs( m_sigma[j]/t); Scalar sn(f/t); m_sigma[j] = t; f = -sn*e[j]; e[j] = cs*e[j]; if (wantu) { for (i = 0; i < m; ++i) { t = cs*m_matU(i,j) + sn*m_matU(i,k-1); m_matU(i,k-1) = -sn*m_matU(i,j) + cs*m_matU(i,k-1); m_matU(i,j) = t; } } } } break; // Perform one qr step. case 3: { // Calculate the shift. Scalar scale = (std::max)((std::max)((std::max)((std::max)( ei_abs(m_sigma[p-1]),ei_abs(m_sigma[p-2])),ei_abs(e[p-2])), ei_abs(m_sigma[k])),ei_abs(e[k])); Scalar sp = m_sigma[p-1]/scale; Scalar spm1 = m_sigma[p-2]/scale; Scalar epm1 = e[p-2]/scale; Scalar sk = m_sigma[k]/scale; Scalar ek = e[k]/scale; Scalar b = ((spm1 + sp)*(spm1 - sp) + epm1*epm1)/Scalar(2); Scalar c = (sp*epm1)*(sp*epm1); Scalar shift(0); if ((b != 0.0) || (c != 0.0)) { shift = ei_sqrt(b*b + c); if (b < 0.0) shift = -shift; shift = c/(b + shift); } Scalar f = (sk + sp)*(sk - sp) + shift; Scalar g = sk*ek; // Chase zeros. for (j = k; j < p-1; ++j) { Scalar t = numext::hypot(f,g); Scalar cs = f/t; Scalar sn = g/t; if (j != k) e[j-1] = t; f = cs*m_sigma[j] + sn*e[j]; e[j] = cs*e[j] - sn*m_sigma[j]; g = sn*m_sigma[j+1]; m_sigma[j+1] = cs*m_sigma[j+1]; if (wantv) { for (i = 0; i < n; ++i) { t = cs*m_matV(i,j) + sn*m_matV(i,j+1); m_matV(i,j+1) = -sn*m_matV(i,j) + cs*m_matV(i,j+1); m_matV(i,j) = t; } } t = numext::hypot(f,g); cs = f/t; sn = g/t; m_sigma[j] = t; f = cs*e[j] + sn*m_sigma[j+1]; m_sigma[j+1] = -sn*e[j] + cs*m_sigma[j+1]; g = sn*e[j+1]; e[j+1] = cs*e[j+1]; if (wantu && (j < m-1)) { for (i = 0; i < m; ++i) { t = cs*m_matU(i,j) + sn*m_matU(i,j+1); m_matU(i,j+1) = -sn*m_matU(i,j) + cs*m_matU(i,j+1); m_matU(i,j) = t; } } } e[p-2] = f; iter = iter + 1; } break; // Convergence. case 4: { // Make the singular values positive. if (m_sigma[k] <= 0.0) { m_sigma[k] = m_sigma[k] < Scalar(0) ? -m_sigma[k] : Scalar(0); if (wantv) m_matV.col(k).start(pp+1) = -m_matV.col(k).start(pp+1); } // Order the singular values. while (k < pp) { if (m_sigma[k] >= m_sigma[k+1]) break; Scalar t = m_sigma[k]; m_sigma[k] = m_sigma[k+1]; m_sigma[k+1] = t; if (wantv && (k < n-1)) m_matV.col(k).swap(m_matV.col(k+1)); if (wantu && (k < m-1)) m_matU.col(k).swap(m_matU.col(k+1)); ++k; } iter = 0; p--; } break; } // end big switch } // end iterations } template SVD& SVD::sort() { int mu = m_matU.rows(); int mv = m_matV.rows(); int n = m_matU.cols(); for (int i=0; i p) { k = j; p = m_sigma.coeff(j); } } if (k != i) { m_sigma.coeffRef(k) = m_sigma.coeff(i); // i.e. m_sigma.coeffRef(i) = p; // swaps the i-th and the k-th elements int j = mu; for(int s=0; j!=0; ++s, --j) std::swap(m_matU.coeffRef(s,i), m_matU.coeffRef(s,k)); j = mv; for (int s=0; j!=0; ++s, --j) std::swap(m_matV.coeffRef(s,i), m_matV.coeffRef(s,k)); } } return *this; } /** \returns the solution of \f$ A x = b \f$ using the current SVD decomposition of A. * The parts of the solution corresponding to zero singular values are ignored. * * \sa MatrixBase::svd(), LU::solve(), LLT::solve() */ template template bool SVD::solve(const MatrixBase &b, ResultType* result) const { ei_assert(b.rows() == m_matU.rows()); Scalar maxVal = m_sigma.cwise().abs().maxCoeff(); for (int j=0; j aux = m_matU.transpose() * b.col(j); for (int i = 0; i col(j) = m_matV * aux; } return true; } /** Computes the polar decomposition of the matrix, as a product unitary x positive. * * If either pointer is zero, the corresponding computation is skipped. * * Only for square matrices. * * \sa computePositiveUnitary(), computeRotationScaling() */ template template void SVD::computeUnitaryPositive(UnitaryType *unitary, PositiveType *positive) const { ei_assert(m_matU.cols() == m_matV.cols() && "Polar decomposition is only for square matrices"); if(unitary) *unitary = m_matU * m_matV.adjoint(); if(positive) *positive = m_matV * m_sigma.asDiagonal() * m_matV.adjoint(); } /** Computes the polar decomposition of the matrix, as a product positive x unitary. * * If either pointer is zero, the corresponding computation is skipped. * * Only for square matrices. * * \sa computeUnitaryPositive(), computeRotationScaling() */ template template void SVD::computePositiveUnitary(UnitaryType *positive, PositiveType *unitary) const { ei_assert(m_matU.rows() == m_matV.rows() && "Polar decomposition is only for square matrices"); if(unitary) *unitary = m_matU * m_matV.adjoint(); if(positive) *positive = m_matU * m_sigma.asDiagonal() * m_matU.adjoint(); } /** decomposes the matrix as a product rotation x scaling, the scaling being * not necessarily positive. * * If either pointer is zero, the corresponding computation is skipped. * * This method requires the Geometry module. * * \sa computeScalingRotation(), computeUnitaryPositive() */ template template void SVD::computeRotationScaling(RotationType *rotation, ScalingType *scaling) const { ei_assert(m_matU.rows() == m_matV.rows() && "Polar decomposition is only for square matrices"); Scalar x = (m_matU * m_matV.adjoint()).determinant(); // so x has absolute value 1 Matrix sv(m_sigma); sv.coeffRef(0) *= x; if(scaling) scaling->lazyAssign(m_matV * sv.asDiagonal() * m_matV.adjoint()); if(rotation) { MatrixType m(m_matU); m.col(0) /= x; rotation->lazyAssign(m * m_matV.adjoint()); } } /** decomposes the matrix as a product scaling x rotation, the scaling being * not necessarily positive. * * If either pointer is zero, the corresponding computation is skipped. * * This method requires the Geometry module. * * \sa computeRotationScaling(), computeUnitaryPositive() */ template template void SVD::computeScalingRotation(ScalingType *scaling, RotationType *rotation) const { ei_assert(m_matU.rows() == m_matV.rows() && "Polar decomposition is only for square matrices"); Scalar x = (m_matU * m_matV.adjoint()).determinant(); // so x has absolute value 1 Matrix sv(m_sigma); sv.coeffRef(0) *= x; if(scaling) scaling->lazyAssign(m_matU * sv.asDiagonal() * m_matU.adjoint()); if(rotation) { MatrixType m(m_matU); m.col(0) /= x; rotation->lazyAssign(m * m_matV.adjoint()); } } /** \svd_module * \returns the SVD decomposition of \c *this */ template inline SVD::PlainObject> MatrixBase::svd() const { return SVD(derived()); } } // end namespace Eigen #endif // EIGEN2_SVD_H