• 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 Desire Nuentsa Wakam <desire.nuentsa_wakam@inria.fr>
5 // Copyright (C) 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 #include "sparse.h"
10 #include <Eigen/SparseQR>
11 
12 template<typename MatrixType,typename DenseMat>
generate_sparse_rectangular_problem(MatrixType & A,DenseMat & dA,int maxRows=300,int maxCols=150)13 int generate_sparse_rectangular_problem(MatrixType& A, DenseMat& dA, int maxRows = 300, int maxCols = 150)
14 {
15   eigen_assert(maxRows >= maxCols);
16   typedef typename MatrixType::Scalar Scalar;
17   int rows = internal::random<int>(1,maxRows);
18   int cols = internal::random<int>(1,maxCols);
19   double density = (std::max)(8./(rows*cols), 0.01);
20 
21   A.resize(rows,cols);
22   dA.resize(rows,cols);
23   initSparse<Scalar>(density, dA, A,ForceNonZeroDiag);
24   A.makeCompressed();
25   int nop = internal::random<int>(0, internal::random<double>(0,1) > 0.5 ? cols/2 : 0);
26   for(int k=0; k<nop; ++k)
27   {
28     int j0 = internal::random<int>(0,cols-1);
29     int j1 = internal::random<int>(0,cols-1);
30     Scalar s = internal::random<Scalar>();
31     A.col(j0)  = s * A.col(j1);
32     dA.col(j0) = s * dA.col(j1);
33   }
34 
35 //   if(rows<cols) {
36 //     A.conservativeResize(cols,cols);
37 //     dA.conservativeResize(cols,cols);
38 //     dA.bottomRows(cols-rows).setZero();
39 //   }
40 
41   return rows;
42 }
43 
test_sparseqr_scalar()44 template<typename Scalar> void test_sparseqr_scalar()
45 {
46   typedef typename NumTraits<Scalar>::Real RealScalar;
47   typedef SparseMatrix<Scalar,ColMajor> MatrixType;
48   typedef Matrix<Scalar,Dynamic,Dynamic> DenseMat;
49   typedef Matrix<Scalar,Dynamic,1> DenseVector;
50   MatrixType A;
51   DenseMat dA;
52   DenseVector refX,x,b;
53   SparseQR<MatrixType, COLAMDOrdering<int> > solver;
54   generate_sparse_rectangular_problem(A,dA);
55 
56   b = dA * DenseVector::Random(A.cols());
57   solver.compute(A);
58 
59   // Q should be MxM
60   VERIFY_IS_EQUAL(solver.matrixQ().rows(), A.rows());
61   VERIFY_IS_EQUAL(solver.matrixQ().cols(), A.rows());
62 
63   // R should be MxN
64   VERIFY_IS_EQUAL(solver.matrixR().rows(), A.rows());
65   VERIFY_IS_EQUAL(solver.matrixR().cols(), A.cols());
66 
67   // Q and R can be multiplied
68   DenseMat recoveredA = solver.matrixQ()
69                       * DenseMat(solver.matrixR().template triangularView<Upper>())
70                       * solver.colsPermutation().transpose();
71   VERIFY_IS_EQUAL(recoveredA.rows(), A.rows());
72   VERIFY_IS_EQUAL(recoveredA.cols(), A.cols());
73 
74   // and in the full rank case the original matrix is recovered
75   if (solver.rank() == A.cols())
76   {
77       VERIFY_IS_APPROX(A, recoveredA);
78   }
79 
80   if(internal::random<float>(0,1)>0.5f)
81     solver.factorize(A);  // this checks that calling analyzePattern is not needed if the pattern do not change.
82   if (solver.info() != Success)
83   {
84     std::cerr << "sparse QR factorization failed\n";
85     exit(0);
86     return;
87   }
88   x = solver.solve(b);
89   if (solver.info() != Success)
90   {
91     std::cerr << "sparse QR factorization failed\n";
92     exit(0);
93     return;
94   }
95 
96   // Compare with a dense QR solver
97   ColPivHouseholderQR<DenseMat> dqr(dA);
98   refX = dqr.solve(b);
99 
100   bool rank_deficient = A.cols()>A.rows() || dqr.rank()<A.cols();
101   if(rank_deficient)
102   {
103     // rank deficient problem -> we might have to increase the threshold
104     // to get a correct solution.
105     RealScalar th = RealScalar(20)*dA.colwise().norm().maxCoeff()*(A.rows()+A.cols()) * NumTraits<RealScalar>::epsilon();
106     for(Index k=0; (k<16) && !test_isApprox(A*x,b); ++k)
107     {
108       th *= RealScalar(10);
109       solver.setPivotThreshold(th);
110       solver.compute(A);
111       x = solver.solve(b);
112     }
113   }
114 
115   VERIFY_IS_APPROX(A * x, b);
116 
117   // For rank deficient problem, the estimated rank might
118   // be slightly off, so let's only raise a warning in such cases.
119   if(rank_deficient) ++g_test_level;
120   VERIFY_IS_EQUAL(solver.rank(), dqr.rank());
121   if(rank_deficient) --g_test_level;
122 
123   if(solver.rank()==A.cols()) // full rank
124     VERIFY_IS_APPROX(x, refX);
125 //   else
126 //     VERIFY((dA * refX - b).norm() * 2 > (A * x - b).norm() );
127 
128   // Compute explicitly the matrix Q
129   MatrixType Q, QtQ, idM;
130   Q = solver.matrixQ();
131   //Check  ||Q' * Q - I ||
132   QtQ = Q * Q.adjoint();
133   idM.resize(Q.rows(), Q.rows()); idM.setIdentity();
134   VERIFY(idM.isApprox(QtQ));
135 
136   // Q to dense
137   DenseMat dQ;
138   dQ = solver.matrixQ();
139   VERIFY_IS_APPROX(Q, dQ);
140 }
EIGEN_DECLARE_TEST(sparseqr)141 EIGEN_DECLARE_TEST(sparseqr)
142 {
143   for(int i=0; i<g_repeat; ++i)
144   {
145     CALL_SUBTEST_1(test_sparseqr_scalar<double>());
146     CALL_SUBTEST_2(test_sparseqr_scalar<std::complex<double> >());
147   }
148 }
149 
150