• 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) 2012 Giacomo Po <gpo@ucla.edu>
5 // Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 
12 #ifndef EIGEN_MINRES_H_
13 #define EIGEN_MINRES_H_
14 
15 
16 namespace Eigen {
17 
18     namespace internal {
19 
20         /** \internal Low-level MINRES algorithm
21          * \param mat The matrix A
22          * \param rhs The right hand side vector b
23          * \param x On input and initial solution, on output the computed solution.
24          * \param precond A right preconditioner being able to efficiently solve for an
25          *                approximation of Ax=b (regardless of b)
26          * \param iters On input the max number of iteration, on output the number of performed iterations.
27          * \param tol_error On input the tolerance error, on output an estimation of the relative error.
28          */
29         template<typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
30         EIGEN_DONT_INLINE
minres(const MatrixType & mat,const Rhs & rhs,Dest & x,const Preconditioner & precond,Index & iters,typename Dest::RealScalar & tol_error)31         void minres(const MatrixType& mat, const Rhs& rhs, Dest& x,
32                     const Preconditioner& precond, Index& iters,
33                     typename Dest::RealScalar& tol_error)
34         {
35             using std::sqrt;
36             typedef typename Dest::RealScalar RealScalar;
37             typedef typename Dest::Scalar Scalar;
38             typedef Matrix<Scalar,Dynamic,1> VectorType;
39 
40             // Check for zero rhs
41             const RealScalar rhsNorm2(rhs.squaredNorm());
42             if(rhsNorm2 == 0)
43             {
44                 x.setZero();
45                 iters = 0;
46                 tol_error = 0;
47                 return;
48             }
49 
50             // initialize
51             const Index maxIters(iters);  // initialize maxIters to iters
52             const Index N(mat.cols());    // the size of the matrix
53             const RealScalar threshold2(tol_error*tol_error*rhsNorm2); // convergence threshold (compared to residualNorm2)
54 
55             // Initialize preconditioned Lanczos
56             VectorType v_old(N); // will be initialized inside loop
57             VectorType v( VectorType::Zero(N) ); //initialize v
58             VectorType v_new(rhs-mat*x); //initialize v_new
59             RealScalar residualNorm2(v_new.squaredNorm());
60             VectorType w(N); // will be initialized inside loop
61             VectorType w_new(precond.solve(v_new)); // initialize w_new
62 //            RealScalar beta; // will be initialized inside loop
63             RealScalar beta_new2(v_new.dot(w_new));
64             eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE");
65             RealScalar beta_new(sqrt(beta_new2));
66             const RealScalar beta_one(beta_new);
67             v_new /= beta_new;
68             w_new /= beta_new;
69             // Initialize other variables
70             RealScalar c(1.0); // the cosine of the Givens rotation
71             RealScalar c_old(1.0);
72             RealScalar s(0.0); // the sine of the Givens rotation
73             RealScalar s_old(0.0); // the sine of the Givens rotation
74             VectorType p_oold(N); // will be initialized in loop
75             VectorType p_old(VectorType::Zero(N)); // initialize p_old=0
76             VectorType p(p_old); // initialize p=0
77             RealScalar eta(1.0);
78 
79             iters = 0; // reset iters
80             while ( iters < maxIters )
81             {
82                 // Preconditioned Lanczos
83                 /* Note that there are 4 variants on the Lanczos algorithm. These are
84                  * described in Paige, C. C. (1972). Computational variants of
85                  * the Lanczos method for the eigenproblem. IMA Journal of Applied
86                  * Mathematics, 10(3), 373–381. The current implementation corresponds
87                  * to the case A(2,7) in the paper. It also corresponds to
88                  * algorithm 6.14 in Y. Saad, Iterative Methods for Sparse Linear
89                  * Systems, 2003 p.173. For the preconditioned version see
90                  * A. Greenbaum, Iterative Methods for Solving Linear Systems, SIAM (1987).
91                  */
92                 const RealScalar beta(beta_new);
93                 v_old = v; // update: at first time step, this makes v_old = 0 so value of beta doesn't matter
94 //                const VectorType v_old(v); // NOT SURE IF CREATING v_old EVERY ITERATION IS EFFICIENT
95                 v = v_new; // update
96                 w = w_new; // update
97 //                const VectorType w(w_new); // NOT SURE IF CREATING w EVERY ITERATION IS EFFICIENT
98                 v_new.noalias() = mat*w - beta*v_old; // compute v_new
99                 const RealScalar alpha = v_new.dot(w);
100                 v_new -= alpha*v; // overwrite v_new
101                 w_new = precond.solve(v_new); // overwrite w_new
102                 beta_new2 = v_new.dot(w_new); // compute beta_new
103                 eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE");
104                 beta_new = sqrt(beta_new2); // compute beta_new
105                 v_new /= beta_new; // overwrite v_new for next iteration
106                 w_new /= beta_new; // overwrite w_new for next iteration
107 
108                 // Givens rotation
109                 const RealScalar r2 =s*alpha+c*c_old*beta; // s, s_old, c and c_old are still from previous iteration
110                 const RealScalar r3 =s_old*beta; // s, s_old, c and c_old are still from previous iteration
111                 const RealScalar r1_hat=c*alpha-c_old*s*beta;
112                 const RealScalar r1 =sqrt( std::pow(r1_hat,2) + std::pow(beta_new,2) );
113                 c_old = c; // store for next iteration
114                 s_old = s; // store for next iteration
115                 c=r1_hat/r1; // new cosine
116                 s=beta_new/r1; // new sine
117 
118                 // Update solution
119                 p_oold = p_old;
120 //                const VectorType p_oold(p_old); // NOT SURE IF CREATING p_oold EVERY ITERATION IS EFFICIENT
121                 p_old = p;
122                 p.noalias()=(w-r2*p_old-r3*p_oold) /r1; // IS NOALIAS REQUIRED?
123                 x += beta_one*c*eta*p;
124 
125                 /* Update the squared residual. Note that this is the estimated residual.
126                 The real residual |Ax-b|^2 may be slightly larger */
127                 residualNorm2 *= s*s;
128 
129                 if ( residualNorm2 < threshold2)
130                 {
131                     break;
132                 }
133 
134                 eta=-s*eta; // update eta
135                 iters++; // increment iteration number (for output purposes)
136             }
137 
138             /* Compute error. Note that this is the estimated error. The real
139              error |Ax-b|/|b| may be slightly larger */
140             tol_error = std::sqrt(residualNorm2 / rhsNorm2);
141         }
142 
143     }
144 
145     template< typename _MatrixType, int _UpLo=Lower,
146     typename _Preconditioner = IdentityPreconditioner>
147     class MINRES;
148 
149     namespace internal {
150 
151         template< typename _MatrixType, int _UpLo, typename _Preconditioner>
152         struct traits<MINRES<_MatrixType,_UpLo,_Preconditioner> >
153         {
154             typedef _MatrixType MatrixType;
155             typedef _Preconditioner Preconditioner;
156         };
157 
158     }
159 
160     /** \ingroup IterativeLinearSolvers_Module
161      * \brief A minimal residual solver for sparse symmetric problems
162      *
163      * This class allows to solve for A.x = b sparse linear problems using the MINRES algorithm
164      * of Paige and Saunders (1975). The sparse matrix A must be symmetric (possibly indefinite).
165      * The vectors x and b can be either dense or sparse.
166      *
167      * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
168      * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower,
169      *               Upper, or Lower|Upper in which the full matrix entries will be considered. Default is Lower.
170      * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
171      *
172      * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
173      * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
174      * and NumTraits<Scalar>::epsilon() for the tolerance.
175      *
176      * This class can be used as the direct solver classes. Here is a typical usage example:
177      * \code
178      * int n = 10000;
179      * VectorXd x(n), b(n);
180      * SparseMatrix<double> A(n,n);
181      * // fill A and b
182      * MINRES<SparseMatrix<double> > mr;
183      * mr.compute(A);
184      * x = mr.solve(b);
185      * std::cout << "#iterations:     " << mr.iterations() << std::endl;
186      * std::cout << "estimated error: " << mr.error()      << std::endl;
187      * // update b, and solve again
188      * x = mr.solve(b);
189      * \endcode
190      *
191      * By default the iterations start with x=0 as an initial guess of the solution.
192      * One can control the start using the solveWithGuess() method.
193      *
194      * MINRES can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
195      *
196      * \sa class ConjugateGradient, BiCGSTAB, SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
197      */
198     template< typename _MatrixType, int _UpLo, typename _Preconditioner>
199     class MINRES : public IterativeSolverBase<MINRES<_MatrixType,_UpLo,_Preconditioner> >
200     {
201 
202         typedef IterativeSolverBase<MINRES> Base;
203         using Base::matrix;
204         using Base::m_error;
205         using Base::m_iterations;
206         using Base::m_info;
207         using Base::m_isInitialized;
208     public:
209         using Base::_solve_impl;
210         typedef _MatrixType MatrixType;
211         typedef typename MatrixType::Scalar Scalar;
212         typedef typename MatrixType::RealScalar RealScalar;
213         typedef _Preconditioner Preconditioner;
214 
215         enum {UpLo = _UpLo};
216 
217     public:
218 
219         /** Default constructor. */
220         MINRES() : Base() {}
221 
222         /** Initialize the solver with matrix \a A for further \c Ax=b solving.
223          *
224          * This constructor is a shortcut for the default constructor followed
225          * by a call to compute().
226          *
227          * \warning this class stores a reference to the matrix A as well as some
228          * precomputed values that depend on it. Therefore, if \a A is changed
229          * this class becomes invalid. Call compute() to update it with the new
230          * matrix A, or modify a copy of A.
231          */
232         template<typename MatrixDerived>
233         explicit MINRES(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
234 
235         /** Destructor. */
236         ~MINRES(){}
237 
238         /** \internal */
239         template<typename Rhs,typename Dest>
240         void _solve_with_guess_impl(const Rhs& b, Dest& x) const
241         {
242             typedef typename Base::MatrixWrapper MatrixWrapper;
243             typedef typename Base::ActualMatrixType ActualMatrixType;
244             enum {
245               TransposeInput  =   (!MatrixWrapper::MatrixFree)
246                               &&  (UpLo==(Lower|Upper))
247                               &&  (!MatrixType::IsRowMajor)
248                               &&  (!NumTraits<Scalar>::IsComplex)
249             };
250             typedef typename internal::conditional<TransposeInput,Transpose<const ActualMatrixType>, ActualMatrixType const&>::type RowMajorWrapper;
251             EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(MatrixWrapper::MatrixFree,UpLo==(Lower|Upper)),MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY);
252             typedef typename internal::conditional<UpLo==(Lower|Upper),
253                                                   RowMajorWrapper,
254                                                   typename MatrixWrapper::template ConstSelfAdjointViewReturnType<UpLo>::Type
255                                             >::type SelfAdjointWrapper;
256 
257             m_iterations = Base::maxIterations();
258             m_error = Base::m_tolerance;
259             RowMajorWrapper row_mat(matrix());
260             for(int j=0; j<b.cols(); ++j)
261             {
262                 m_iterations = Base::maxIterations();
263                 m_error = Base::m_tolerance;
264 
265                 typename Dest::ColXpr xj(x,j);
266                 internal::minres(SelfAdjointWrapper(row_mat), b.col(j), xj,
267                                  Base::m_preconditioner, m_iterations, m_error);
268             }
269 
270             m_isInitialized = true;
271             m_info = m_error <= Base::m_tolerance ? Success : NoConvergence;
272         }
273 
274         /** \internal */
275         template<typename Rhs,typename Dest>
276         void _solve_impl(const Rhs& b, MatrixBase<Dest> &x) const
277         {
278             x.setZero();
279             _solve_with_guess_impl(b,x.derived());
280         }
281 
282     protected:
283 
284     };
285 
286 } // end namespace Eigen
287 
288 #endif // EIGEN_MINRES_H
289 
290