在Eigen中乘以稀疏子矩阵

Question

I am trying to multiply sub-matrices such as blocks, columns, and rows of Eigen::SparseMatrix. However, whenever multiple sub-matrices are involved my program crashes (and gdb with it). I am working with Eigen 3.2.1.

Here an example:

const unsigned m = 3, d = 1;
SparseMatrix<double> H(3*m,3*m);
H.setIdentity();
SparseMatrix<double> G(m,d);
G.coeffRef(0,0) = 1;

// this works
SparseMatrix<double> H_00 = H.block(0,0,m,m);
double val = SparseMatrix<double>(G.col(0).transpose() * H_00 * G.col(0)).coeffRef(0,0);

// this crashes
val = SparseMatrix<double>(G.col(0).transpose() * H.block(0,0,m,m) * G.col(0)).coeffRef(0,0);

Is there a way of avoiding the costly ( m and d >> 1000) construction of H_00?

Answer 1

I checked out the 3.2.1 version of Eigen and this appears to be related to this Eigen bug: http://eigen.tuxfamily.org/bz/show_bug.cgi?id=875

What happens is that a function (named SparseMatrixBase::nonZeros() ) internally is called recursively Unfortunately, this was not properly fixed in the 3.2 branch (with version 3.2.10 your example does not compile). It does work with the latest 3.3 version of Eigen, however (3.3rc1 at the moment) -- which answers your question: Yes, upgrade to Eigen 3.3rc1 (or newer)