Go to the documentation of this file.00001
00002
00003
00004
00005
00006
00007
00008
00009
00010 #ifndef EIGEN_MATRIX_FUNCTION_ATOMIC
00011 #define EIGEN_MATRIX_FUNCTION_ATOMIC
00012
00013 namespace Eigen {
00014
00023 template <typename MatrixType>
00024 class MatrixFunctionAtomic
00025 {
00026 public:
00027
00028 typedef typename MatrixType::Scalar Scalar;
00029 typedef typename MatrixType::Index Index;
00030 typedef typename NumTraits<Scalar>::Real RealScalar;
00031 typedef typename internal::stem_function<Scalar>::type StemFunction;
00032 typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
00033
00037 MatrixFunctionAtomic(StemFunction f) : m_f(f) { }
00038
00043 MatrixType compute(const MatrixType& A);
00044
00045 private:
00046
00047
00048 MatrixFunctionAtomic(const MatrixFunctionAtomic&);
00049 MatrixFunctionAtomic& operator=(const MatrixFunctionAtomic&);
00050
00051 void computeMu();
00052 bool taylorConverged(Index s, const MatrixType& F, const MatrixType& Fincr, const MatrixType& P);
00053
00055 StemFunction* m_f;
00056
00058 Index m_Arows;
00059
00061 Scalar m_avgEival;
00062
00064 MatrixType m_Ashifted;
00065
00067 RealScalar m_mu;
00068 };
00069
00070 template <typename MatrixType>
00071 MatrixType MatrixFunctionAtomic<MatrixType>::compute(const MatrixType& A)
00072 {
00073
00074 m_Arows = A.rows();
00075 m_avgEival = A.trace() / Scalar(RealScalar(m_Arows));
00076 m_Ashifted = A - m_avgEival * MatrixType::Identity(m_Arows, m_Arows);
00077 computeMu();
00078 MatrixType F = m_f(m_avgEival, 0) * MatrixType::Identity(m_Arows, m_Arows);
00079 MatrixType P = m_Ashifted;
00080 MatrixType Fincr;
00081 for (Index s = 1; s < 1.1 * m_Arows + 10; s++) {
00082 Fincr = m_f(m_avgEival, static_cast<int>(s)) * P;
00083 F += Fincr;
00084 P = Scalar(RealScalar(1.0/(s + 1))) * P * m_Ashifted;
00085 if (taylorConverged(s, F, Fincr, P)) {
00086 return F;
00087 }
00088 }
00089 eigen_assert("Taylor series does not converge" && 0);
00090 return F;
00091 }
00092
00094 template <typename MatrixType>
00095 void MatrixFunctionAtomic<MatrixType>::computeMu()
00096 {
00097 const MatrixType N = MatrixType::Identity(m_Arows, m_Arows) - m_Ashifted;
00098 VectorType e = VectorType::Ones(m_Arows);
00099 N.template triangularView<Upper>().solveInPlace(e);
00100 m_mu = e.cwiseAbs().maxCoeff();
00101 }
00102
00104 template <typename MatrixType>
00105 bool MatrixFunctionAtomic<MatrixType>::taylorConverged(Index s, const MatrixType& F,
00106 const MatrixType& Fincr, const MatrixType& P)
00107 {
00108 const Index n = F.rows();
00109 const RealScalar F_norm = F.cwiseAbs().rowwise().sum().maxCoeff();
00110 const RealScalar Fincr_norm = Fincr.cwiseAbs().rowwise().sum().maxCoeff();
00111 if (Fincr_norm < NumTraits<Scalar>::epsilon() * F_norm) {
00112 RealScalar delta = 0;
00113 RealScalar rfactorial = 1;
00114 for (Index r = 0; r < n; r++) {
00115 RealScalar mx = 0;
00116 for (Index i = 0; i < n; i++)
00117 mx = (std::max)(mx, std::abs(m_f(m_Ashifted(i, i) + m_avgEival, static_cast<int>(s+r))));
00118 if (r != 0)
00119 rfactorial *= RealScalar(r);
00120 delta = (std::max)(delta, mx / rfactorial);
00121 }
00122 const RealScalar P_norm = P.cwiseAbs().rowwise().sum().maxCoeff();
00123 if (m_mu * delta * P_norm < NumTraits<Scalar>::epsilon() * F_norm)
00124 return true;
00125 }
00126 return false;
00127 }
00128
00129 }
00130
00131 #endif // EIGEN_MATRIX_FUNCTION_ATOMIC