CSRMatrix::forNonZeroElementsInRowOrder inconsistencies

[bernlu]

Creating a 2x2 laplacian CSRMatrix from a graph with self-loops is for some reason not equal to a matrix created with the same values by hand (note that the laplacian matrix is actually not corrent, see #1120 , and this issue may be related to that). Additionally, the resulting laplacian matrix is not symmetric when checked with MatrixTools::isSymmetric.

Example code:

    Graph G(2);
    G.addEdge(0, 1);

    std::cout << CSRMatrix::laplacianMatrix(G) << std::endl;
    // 1, -1,
    // -1, 1,

    G.addEdge(1, 1);
    auto lp = CSRMatrix::laplacianMatrix(G);

    std::cout << lp << std::endl;
    // 1, -1,
    // -1, -1,

    std::cout << MatrixTools::isSymmetric(lp) << std::endl;
    // 0   -- expect 1

    CSRMatrix mat(2);
    mat.setValue(0, 0, 1);
    mat.setValue(1, 0, -1);
    mat.setValue(0, 1, -1);
    mat.setValue(1, 1, -1);

    std::cout << mat << std::endl;
    // 1, -1,
    // -1, -1,

    std::cout << MatrixTools::isSymmetric(mat) << std::endl;
    // 1

    std::cout << (lp == mat) << std::endl;
    // 0   -- expect 1

Both isSymmetric and operator== use CSRMatrix::forNonZeroElementsInOrder which seems to sometimes produce wrong values for diagonal entries. For the laplacian, the lambda is called with (1,1,1) even though the value should be -1 instead of 1.