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_CONJUGATE_GRADIENT_H
11 #define EIGEN_CONJUGATE_GRADIENT_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 /** \internal Low-level conjugate gradient algorithm
18   * \param mat The matrix A
19   * \param rhs The right hand side vector b
20   * \param x On input and initial solution, on output the computed solution.
21   * \param precond A preconditioner being able to efficiently solve for an
22   *                approximation of Ax=b (regardless of b)
23   * \param iters On input the max number of iteration, on output the number of performed iterations.
24   * \param tol_error On input the tolerance error, on output an estimation of the relative error.
25   */
26 template<typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
27 EIGEN_DONT_INLINE
conjugate_gradient(const MatrixType & mat,const Rhs & rhs,Dest & x,const Preconditioner & precond,Index & iters,typename Dest::RealScalar & tol_error)28 void conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x,
29                         const Preconditioner& precond, Index& iters,
30                         typename Dest::RealScalar& tol_error)
31 {
32   using std::sqrt;
33   using std::abs;
34   typedef typename Dest::RealScalar RealScalar;
35   typedef typename Dest::Scalar Scalar;
36   typedef Matrix<Scalar,Dynamic,1> VectorType;
37 
38   RealScalar tol = tol_error;
39   Index maxIters = iters;
40 
41   Index n = mat.cols();
42 
43   VectorType residual = rhs - mat * x; //initial residual
44 
45   RealScalar rhsNorm2 = rhs.squaredNorm();
46   if(rhsNorm2 == 0)
47   {
48     x.setZero();
49     iters = 0;
50     tol_error = 0;
51     return;
52   }
53   RealScalar threshold = tol*tol*rhsNorm2;
54   RealScalar residualNorm2 = residual.squaredNorm();
55   if (residualNorm2 < threshold)
56   {
57     iters = 0;
58     tol_error = sqrt(residualNorm2 / rhsNorm2);
59     return;
60   }
61 
62   VectorType p(n);
63   p = precond.solve(residual);      // initial search direction
64 
65   VectorType z(n), tmp(n);
66   RealScalar absNew = numext::real(residual.dot(p));  // the square of the absolute value of r scaled by invM
67   Index i = 0;
68   while(i < maxIters)
69   {
70     tmp.noalias() = mat * p;                    // the bottleneck of the algorithm
71 
72     Scalar alpha = absNew / p.dot(tmp);         // the amount we travel on dir
73     x += alpha * p;                             // update solution
74     residual -= alpha * tmp;                    // update residual
75 
76     residualNorm2 = residual.squaredNorm();
77     if(residualNorm2 < threshold)
78       break;
79 
80     z = precond.solve(residual);                // approximately solve for "A z = residual"
81 
82     RealScalar absOld = absNew;
83     absNew = numext::real(residual.dot(z));     // update the absolute value of r
84     RealScalar beta = absNew / absOld;          // calculate the Gram-Schmidt value used to create the new search direction
85     p = z + beta * p;                           // update search direction
86     i++;
87   }
88   tol_error = sqrt(residualNorm2 / rhsNorm2);
89   iters = i;
90 }
91 
92 }
93 
94 template< typename _MatrixType, int _UpLo=Lower,
95           typename _Preconditioner = DiagonalPreconditioner<typename _MatrixType::Scalar> >
96 class ConjugateGradient;
97 
98 namespace internal {
99 
100 template< typename _MatrixType, int _UpLo, typename _Preconditioner>
101 struct traits<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
102 {
103   typedef _MatrixType MatrixType;
104   typedef _Preconditioner Preconditioner;
105 };
106 
107 }
108 
109 /** \ingroup IterativeLinearSolvers_Module
110   * \brief A conjugate gradient solver for sparse (or dense) self-adjoint problems
111   *
112   * This class allows to solve for A.x = b linear problems using an iterative conjugate gradient algorithm.
113   * The matrix A must be selfadjoint. The matrix A and the vectors x and b can be either dense or sparse.
114   *
115   * \tparam _MatrixType the type of the matrix A, can be a dense or a sparse matrix.
116   * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower,
117   *               \c Upper, or \c Lower|Upper in which the full matrix entries will be considered.
118   *               Default is \c Lower, best performance is \c Lower|Upper.
119   * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
120   *
121   * \implsparsesolverconcept
122   *
123   * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
124   * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
125   * and NumTraits<Scalar>::epsilon() for the tolerance.
126   *
127   * The tolerance corresponds to the relative residual error: |Ax-b|/|b|
128   *
129   * \b Performance: Even though the default value of \c _UpLo is \c Lower, significantly higher performance is
130   * achieved when using a complete matrix and \b Lower|Upper as the \a _UpLo template parameter. Moreover, in this
131   * case multi-threading can be exploited if the user code is compiled with OpenMP enabled.
132   * See \ref TopicMultiThreading for details.
133   *
134   * This class can be used as the direct solver classes. Here is a typical usage example:
135     \code
136     int n = 10000;
137     VectorXd x(n), b(n);
138     SparseMatrix<double> A(n,n);
139     // fill A and b
140     ConjugateGradient<SparseMatrix<double>, Lower|Upper> cg;
141     cg.compute(A);
142     x = cg.solve(b);
143     std::cout << "#iterations:     " << cg.iterations() << std::endl;
144     std::cout << "estimated error: " << cg.error()      << std::endl;
145     // update b, and solve again
146     x = cg.solve(b);
147     \endcode
148   *
149   * By default the iterations start with x=0 as an initial guess of the solution.
150   * One can control the start using the solveWithGuess() method.
151   *
152   * ConjugateGradient can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
153   *
154   * \sa class LeastSquaresConjugateGradient, class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
155   */
156 template< typename _MatrixType, int _UpLo, typename _Preconditioner>
157 class ConjugateGradient : public IterativeSolverBase<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
158 {
159   typedef IterativeSolverBase<ConjugateGradient> Base;
160   using Base::matrix;
161   using Base::m_error;
162   using Base::m_iterations;
163   using Base::m_info;
164   using Base::m_isInitialized;
165 public:
166   typedef _MatrixType MatrixType;
167   typedef typename MatrixType::Scalar Scalar;
168   typedef typename MatrixType::RealScalar RealScalar;
169   typedef _Preconditioner Preconditioner;
170 
171   enum {
172     UpLo = _UpLo
173   };
174 
175 public:
176 
177   /** Default constructor. */
178   ConjugateGradient() : Base() {}
179 
180   /** Initialize the solver with matrix \a A for further \c Ax=b solving.
181     *
182     * This constructor is a shortcut for the default constructor followed
183     * by a call to compute().
184     *
185     * \warning this class stores a reference to the matrix A as well as some
186     * precomputed values that depend on it. Therefore, if \a A is changed
187     * this class becomes invalid. Call compute() to update it with the new
188     * matrix A, or modify a copy of A.
189     */
190   template<typename MatrixDerived>
191   explicit ConjugateGradient(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
192 
193   ~ConjugateGradient() {}
194 
195   /** \internal */
196   template<typename Rhs,typename Dest>
197   void _solve_with_guess_impl(const Rhs& b, Dest& x) const
198   {
199     typedef typename Base::MatrixWrapper MatrixWrapper;
200     typedef typename Base::ActualMatrixType ActualMatrixType;
201     enum {
202       TransposeInput  =   (!MatrixWrapper::MatrixFree)
203                       &&  (UpLo==(Lower|Upper))
204                       &&  (!MatrixType::IsRowMajor)
205                       &&  (!NumTraits<Scalar>::IsComplex)
206     };
207     typedef typename internal::conditional<TransposeInput,Transpose<const ActualMatrixType>, ActualMatrixType const&>::type RowMajorWrapper;
208     EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(MatrixWrapper::MatrixFree,UpLo==(Lower|Upper)),MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY);
209     typedef typename internal::conditional<UpLo==(Lower|Upper),
210                                            RowMajorWrapper,
211                                            typename MatrixWrapper::template ConstSelfAdjointViewReturnType<UpLo>::Type
212                                           >::type SelfAdjointWrapper;
213     m_iterations = Base::maxIterations();
214     m_error = Base::m_tolerance;
215 
216     for(Index j=0; j<b.cols(); ++j)
217     {
218       m_iterations = Base::maxIterations();
219       m_error = Base::m_tolerance;
220 
221       typename Dest::ColXpr xj(x,j);
222       RowMajorWrapper row_mat(matrix());
223       internal::conjugate_gradient(SelfAdjointWrapper(row_mat), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error);
224     }
225 
226     m_isInitialized = true;
227     m_info = m_error <= Base::m_tolerance ? Success : NoConvergence;
228   }
229 
230   /** \internal */
231   using Base::_solve_impl;
232   template<typename Rhs,typename Dest>
233   void _solve_impl(const MatrixBase<Rhs>& b, Dest& x) const
234   {
235     x.setZero();
236     _solve_with_guess_impl(b.derived(),x);
237   }
238 
239 protected:
240 
241 };
242 
243 } // end namespace Eigen
244 
245 #endif // EIGEN_CONJUGATE_GRADIENT_H
246