1c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath// This file is part of Eigen, a lightweight C++ template library
2c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath// for linear algebra.
3c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath//
4c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath// Copyright (C) 2012 Desire NUENTSA WAKAM <desire.nuentsa_wakam@inria.fr
5c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath//
6c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath// This Source Code Form is subject to the terms of the Mozilla
7c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath// Public License v. 2.0. If a copy of the MPL was not distributed
8c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
107faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez#ifndef EIGEN_ITERSCALING_H
117faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez#define EIGEN_ITERSCALING_H
12c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath/**
13c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * \ingroup IterativeSolvers_Module
14c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * \brief iterative scaling algorithm to equilibrate rows and column norms in matrices
15c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  *
16c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * This class can be used as a preprocessing tool to accelerate the convergence of iterative methods
17c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  *
18c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * This feature is  useful to limit the pivoting amount during LU/ILU factorization
19c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * The  scaling strategy as presented here preserves the symmetry of the problem
20c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * NOTE It is assumed that the matrix does not have empty row or column,
21c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  *
22c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * Example with key steps
23c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * \code
24c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * VectorXd x(n), b(n);
25c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * SparseMatrix<double> A;
26c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * // fill A and b;
277faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez  * IterScaling<SparseMatrix<double> > scal;
28c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * // Compute the left and right scaling vectors. The matrix is equilibrated at output
29c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * scal.computeRef(A);
30c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * // Scale the right hand side
31c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * b = scal.LeftScaling().cwiseProduct(b);
32c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * // Now, solve the equilibrated linear system with any available solver
33c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  *
34c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * // Scale back the computed solution
35c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * x = scal.RightScaling().cwiseProduct(x);
36c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * \endcode
37c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  *
38c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * \tparam _MatrixType the type of the matrix. It should be a real square sparsematrix
39c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  *
40c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * References : D. Ruiz and B. Ucar, A Symmetry Preserving Algorithm for Matrix Scaling, INRIA Research report RR-7552
41c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  *
42c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  * \sa \ref IncompleteLUT
43c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  */
447faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandeznamespace Eigen {
45c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamathusing std::abs;
46c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamathtemplate<typename _MatrixType>
477faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandezclass IterScaling
48c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath{
49c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  public:
50c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    typedef _MatrixType MatrixType;
51c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    typedef typename MatrixType::Scalar Scalar;
52c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    typedef typename MatrixType::Index Index;
53c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
54c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  public:
557faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez    IterScaling() { init(); }
56c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
577faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez    IterScaling(const MatrixType& matrix)
58c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    {
59c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      init();
60c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      compute(matrix);
61c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    }
62c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
637faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez    ~IterScaling() { }
64c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
65c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    /**
66c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     * Compute the left and right diagonal matrices to scale the input matrix @p mat
67c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     *
68c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     * FIXME This algorithm will be modified such that the diagonal elements are permuted on the diagonal.
69c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     *
70c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     * \sa LeftScaling() RightScaling()
71c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     */
72c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    void compute (const MatrixType& mat)
73c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    {
74c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      int m = mat.rows();
75c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      int n = mat.cols();
767faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez      eigen_assert((m>0 && m == n) && "Please give a non - empty matrix");
77c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_left.resize(m);
78c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_right.resize(n);
79c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_left.setOnes();
80c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_right.setOnes();
81c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_matrix = mat;
82c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      VectorXd Dr, Dc, DrRes, DcRes; // Temporary Left and right scaling vectors
83c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      Dr.resize(m); Dc.resize(n);
84c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      DrRes.resize(m); DcRes.resize(n);
85c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      double EpsRow = 1.0, EpsCol = 1.0;
86c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      int its = 0;
87c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      do
88c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      { // Iterate until the infinite norm of each row and column is approximately 1
89c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        // Get the maximum value in each row and column
90c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        Dr.setZero(); Dc.setZero();
91c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        for (int k=0; k<m_matrix.outerSize(); ++k)
92c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        {
93c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          for (typename MatrixType::InnerIterator it(m_matrix, k); it; ++it)
94c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          {
95c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath            if ( Dr(it.row()) < abs(it.value()) )
96c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath              Dr(it.row()) = abs(it.value());
97c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
98c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath            if ( Dc(it.col()) < abs(it.value()) )
99c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath              Dc(it.col()) = abs(it.value());
100c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          }
101c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        }
102c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        for (int i = 0; i < m; ++i)
103c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        {
104c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          Dr(i) = std::sqrt(Dr(i));
105c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          Dc(i) = std::sqrt(Dc(i));
106c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        }
107c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        // Save the scaling factors
108c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        for (int i = 0; i < m; ++i)
109c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        {
110c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          m_left(i) /= Dr(i);
111c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          m_right(i) /= Dc(i);
112c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        }
113c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        // Scale the rows and the columns of the matrix
114c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        DrRes.setZero(); DcRes.setZero();
115c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        for (int k=0; k<m_matrix.outerSize(); ++k)
116c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        {
117c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          for (typename MatrixType::InnerIterator it(m_matrix, k); it; ++it)
118c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          {
119c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath            it.valueRef() = it.value()/( Dr(it.row()) * Dc(it.col()) );
120c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath            // Accumulate the norms of the row and column vectors
121c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath            if ( DrRes(it.row()) < abs(it.value()) )
122c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath              DrRes(it.row()) = abs(it.value());
123c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
124c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath            if ( DcRes(it.col()) < abs(it.value()) )
125c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath              DcRes(it.col()) = abs(it.value());
126c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath          }
127c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        }
128c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        DrRes.array() = (1-DrRes.array()).abs();
129c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        EpsRow = DrRes.maxCoeff();
130c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        DcRes.array() = (1-DcRes.array()).abs();
131c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        EpsCol = DcRes.maxCoeff();
132c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath        its++;
133c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      }while ( (EpsRow >m_tol || EpsCol > m_tol) && (its < m_maxits) );
134c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_isInitialized = true;
135c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    }
136c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    /** Compute the left and right vectors to scale the vectors
137c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     * the input matrix is scaled with the computed vectors at output
138c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     *
139c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     * \sa compute()
140c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     */
141c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    void computeRef (MatrixType& mat)
142c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    {
143c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      compute (mat);
144c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      mat = m_matrix;
145c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    }
146c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    /** Get the vector to scale the rows of the matrix
147c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     */
148c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    VectorXd& LeftScaling()
149c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    {
150c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      return m_left;
151c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    }
152c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
153c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    /** Get the vector to scale the columns of the matrix
154c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     */
155c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    VectorXd& RightScaling()
156c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    {
157c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      return m_right;
158c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    }
159c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
160c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    /** Set the tolerance for the convergence of the iterative scaling algorithm
161c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath     */
162c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    void setTolerance(double tol)
163c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    {
164c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_tol = tol;
165c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    }
166c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
167c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath  protected:
168c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
169c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    void init()
170c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    {
171c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_tol = 1e-10;
172c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_maxits = 5;
173c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath      m_isInitialized = false;
174c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    }
175c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath
176c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    MatrixType m_matrix;
177c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    mutable ComputationInfo m_info;
178c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    bool m_isInitialized;
179c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    VectorXd m_left; // Left scaling vector
180c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    VectorXd m_right; // m_right scaling vector
181c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    double m_tol;
182c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath    int m_maxits; // Maximum number of iterations allowed
183c981c48f5bc9aefeffc0bcb0cc3934c2fae179ddNarayan Kamath};
1847faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez}
1857faaa9f3f0df9d23790277834d426c3d992ac3baCarlos Hernandez#endif
186