How to Fix Eigen::NaturalOrdering<int> >::MatrixType' has no member named 'isCompressed'

Question

I want to use SparseQR or SparseLU method for solving linear systems with sparse matrices. I am using Eigen Library for that purpose. The problem is that this methods gives error an the code doesnt work. I am programing in Eclipse C++, with MinGW compiler

I have used the bicgstab method from the Eigen Library and it is working, but now with the Sparse methods errors appear.

This is the code I want to use for solving linear systems, just changing the method name and one of the parameters it works for other methods, but not for sparse methods

SparseQR<MatrixXd, NaturalOrdering<int>> solver;
solver.analyzePattern(A);
solver.factorize(A);
x = solver.solve(y);

x,A and y are of the class MatrixXd.

This code works for the method bicgstab, and the syntaxis is the same. The problem is that the following error appears:

src\/src/SparseQR/SparseQR.h:381:66: error: 'const class
Eigen::SparseQR<Eigen::Matrix<double, -0x000000001,
-0x000000001>,Eigen::NaturalOrdering<int> >::MatrixType' has no member named 'isCompressed'

src\/src/SparseQR/SparseQR.h:381:66: error: 'const class
Eigen::SparseQR<Eigen::Matrix<double, -0x000000001,
-0x000000001>,Eigen::NaturalOrdering<int> >::MatrixType' has no member named 'outerIndexPtr'

and I dont know hot fix it.

If you have any idea of a possible solution I would really appreciate it. Thanks for your time

Answer 1

As said in the comments, SparseQR only works with sparse matrices. If for some reason your input matrix A is passed as a dense matrix, you could convert it using A.sparseView(), but you should really prefer storing it as a (compressed) sparse matrix in the first place. Also notice that the template parameter of SparseQR must be a SparseMatrix:

void foo(Eigen::MatrixXd const &A, Eigen::VectorXd const &y, Eigen::VectorXd& x)
{
    Eigen::SparseQR<Eigen::SparseMatrix<double>, Eigen::NaturalOrdering<int> > solver;
    solver.compute(A.sparseView());
    x = solver.solve(y);
}

If your matrix A in fact is dense (or "not very sparse") it is usually much more efficient to just use a dense solver, e.g.

void foo(Eigen::MatrixXd const &A, Eigen::VectorXd const &y, Eigen::VectorXd& x)
{
    Eigen::FullPivHouseholderQR<Eigen::MatrixXd> solver;
    solver.compute(A);
    x = solver.solve(y);
}