Home | History | Annotate | Download | only in IterativeSolvers
      1 // This file is part of Eigen, a lightweight C++ template library
      2 // for linear algebra.
      3 //
      4 // Copyright (C) 2012 Giacomo Po <gpo (at) ucla.edu>
      5 // Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud (at) 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
     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), 373381. 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