• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
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_BASIC_PRECONDITIONERS_H
11 #define EIGEN_BASIC_PRECONDITIONERS_H
12 
13 namespace Eigen {
14 
15 /** \ingroup IterativeLinearSolvers_Module
16   * \brief A preconditioner based on the digonal entries
17   *
18   * This class allows to approximately solve for A.x = b problems assuming A is a diagonal matrix.
19   * In other words, this preconditioner neglects all off diagonal entries and, in Eigen's language, solves for:
20     \code
21     A.diagonal().asDiagonal() . x = b
22     \endcode
23   *
24   * \tparam _Scalar the type of the scalar.
25   *
26   * \implsparsesolverconcept
27   *
28   * This preconditioner is suitable for both selfadjoint and general problems.
29   * The diagonal entries are pre-inverted and stored into a dense vector.
30   *
31   * \note A variant that has yet to be implemented would attempt to preserve the norm of each column.
32   *
33   * \sa class LeastSquareDiagonalPreconditioner, class ConjugateGradient
34   */
35 template <typename _Scalar>
36 class DiagonalPreconditioner
37 {
38     typedef _Scalar Scalar;
39     typedef Matrix<Scalar,Dynamic,1> Vector;
40   public:
41     typedef typename Vector::StorageIndex StorageIndex;
42     enum {
43       ColsAtCompileTime = Dynamic,
44       MaxColsAtCompileTime = Dynamic
45     };
46 
DiagonalPreconditioner()47     DiagonalPreconditioner() : m_isInitialized(false) {}
48 
49     template<typename MatType>
DiagonalPreconditioner(const MatType & mat)50     explicit DiagonalPreconditioner(const MatType& mat) : m_invdiag(mat.cols())
51     {
52       compute(mat);
53     }
54 
rows()55     Index rows() const { return m_invdiag.size(); }
cols()56     Index cols() const { return m_invdiag.size(); }
57 
58     template<typename MatType>
analyzePattern(const MatType &)59     DiagonalPreconditioner& analyzePattern(const MatType& )
60     {
61       return *this;
62     }
63 
64     template<typename MatType>
factorize(const MatType & mat)65     DiagonalPreconditioner& factorize(const MatType& mat)
66     {
67       m_invdiag.resize(mat.cols());
68       for(int j=0; j<mat.outerSize(); ++j)
69       {
70         typename MatType::InnerIterator it(mat,j);
71         while(it && it.index()!=j) ++it;
72         if(it && it.index()==j && it.value()!=Scalar(0))
73           m_invdiag(j) = Scalar(1)/it.value();
74         else
75           m_invdiag(j) = Scalar(1);
76       }
77       m_isInitialized = true;
78       return *this;
79     }
80 
81     template<typename MatType>
compute(const MatType & mat)82     DiagonalPreconditioner& compute(const MatType& mat)
83     {
84       return factorize(mat);
85     }
86 
87     /** \internal */
88     template<typename Rhs, typename Dest>
_solve_impl(const Rhs & b,Dest & x)89     void _solve_impl(const Rhs& b, Dest& x) const
90     {
91       x = m_invdiag.array() * b.array() ;
92     }
93 
94     template<typename Rhs> inline const Solve<DiagonalPreconditioner, Rhs>
solve(const MatrixBase<Rhs> & b)95     solve(const MatrixBase<Rhs>& b) const
96     {
97       eigen_assert(m_isInitialized && "DiagonalPreconditioner is not initialized.");
98       eigen_assert(m_invdiag.size()==b.rows()
99                 && "DiagonalPreconditioner::solve(): invalid number of rows of the right hand side matrix b");
100       return Solve<DiagonalPreconditioner, Rhs>(*this, b.derived());
101     }
102 
info()103     ComputationInfo info() { return Success; }
104 
105   protected:
106     Vector m_invdiag;
107     bool m_isInitialized;
108 };
109 
110 /** \ingroup IterativeLinearSolvers_Module
111   * \brief Jacobi preconditioner for LeastSquaresConjugateGradient
112   *
113   * This class allows to approximately solve for A' A x  = A' b problems assuming A' A is a diagonal matrix.
114   * In other words, this preconditioner neglects all off diagonal entries and, in Eigen's language, solves for:
115     \code
116     (A.adjoint() * A).diagonal().asDiagonal() * x = b
117     \endcode
118   *
119   * \tparam _Scalar the type of the scalar.
120   *
121   * \implsparsesolverconcept
122   *
123   * The diagonal entries are pre-inverted and stored into a dense vector.
124   *
125   * \sa class LeastSquaresConjugateGradient, class DiagonalPreconditioner
126   */
127 template <typename _Scalar>
128 class LeastSquareDiagonalPreconditioner : public DiagonalPreconditioner<_Scalar>
129 {
130     typedef _Scalar Scalar;
131     typedef typename NumTraits<Scalar>::Real RealScalar;
132     typedef DiagonalPreconditioner<_Scalar> Base;
133     using Base::m_invdiag;
134   public:
135 
LeastSquareDiagonalPreconditioner()136     LeastSquareDiagonalPreconditioner() : Base() {}
137 
138     template<typename MatType>
LeastSquareDiagonalPreconditioner(const MatType & mat)139     explicit LeastSquareDiagonalPreconditioner(const MatType& mat) : Base()
140     {
141       compute(mat);
142     }
143 
144     template<typename MatType>
analyzePattern(const MatType &)145     LeastSquareDiagonalPreconditioner& analyzePattern(const MatType& )
146     {
147       return *this;
148     }
149 
150     template<typename MatType>
factorize(const MatType & mat)151     LeastSquareDiagonalPreconditioner& factorize(const MatType& mat)
152     {
153       // Compute the inverse squared-norm of each column of mat
154       m_invdiag.resize(mat.cols());
155       if(MatType::IsRowMajor)
156       {
157         m_invdiag.setZero();
158         for(Index j=0; j<mat.outerSize(); ++j)
159         {
160           for(typename MatType::InnerIterator it(mat,j); it; ++it)
161             m_invdiag(it.index()) += numext::abs2(it.value());
162         }
163         for(Index j=0; j<mat.cols(); ++j)
164           if(numext::real(m_invdiag(j))>RealScalar(0))
165             m_invdiag(j) = RealScalar(1)/numext::real(m_invdiag(j));
166       }
167       else
168       {
169         for(Index j=0; j<mat.outerSize(); ++j)
170         {
171           RealScalar sum = mat.innerVector(j).squaredNorm();
172           if(sum>RealScalar(0))
173             m_invdiag(j) = RealScalar(1)/sum;
174           else
175             m_invdiag(j) = RealScalar(1);
176         }
177       }
178       Base::m_isInitialized = true;
179       return *this;
180     }
181 
182     template<typename MatType>
compute(const MatType & mat)183     LeastSquareDiagonalPreconditioner& compute(const MatType& mat)
184     {
185       return factorize(mat);
186     }
187 
info()188     ComputationInfo info() { return Success; }
189 
190   protected:
191 };
192 
193 /** \ingroup IterativeLinearSolvers_Module
194   * \brief A naive preconditioner which approximates any matrix as the identity matrix
195   *
196   * \implsparsesolverconcept
197   *
198   * \sa class DiagonalPreconditioner
199   */
200 class IdentityPreconditioner
201 {
202   public:
203 
IdentityPreconditioner()204     IdentityPreconditioner() {}
205 
206     template<typename MatrixType>
IdentityPreconditioner(const MatrixType &)207     explicit IdentityPreconditioner(const MatrixType& ) {}
208 
209     template<typename MatrixType>
analyzePattern(const MatrixType &)210     IdentityPreconditioner& analyzePattern(const MatrixType& ) { return *this; }
211 
212     template<typename MatrixType>
factorize(const MatrixType &)213     IdentityPreconditioner& factorize(const MatrixType& ) { return *this; }
214 
215     template<typename MatrixType>
compute(const MatrixType &)216     IdentityPreconditioner& compute(const MatrixType& ) { return *this; }
217 
218     template<typename Rhs>
solve(const Rhs & b)219     inline const Rhs& solve(const Rhs& b) const { return b; }
220 
info()221     ComputationInfo info() { return Success; }
222 };
223 
224 } // end namespace Eigen
225 
226 #endif // EIGEN_BASIC_PRECONDITIONERS_H
227