JAMA没有计算正确的特征向量

Question

I have the following matrix, which is symmetric and real (It is a hamilton-operator): (Matlab friendly)

[63.000000, -1.732051, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000; -1.732051, 61.000000, -2.000000, 0.000000, -1.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000; 0.000000, -2.000000, 61.000000, -1.732051, 0.000000, -1.414214, 0.000000, 0.000000, 0.000000, 0.000000; 0.000000, 0.000000, -1.732051, 63.000000, 0.000000, 0.000000, -1.732051, 0.000000, 0.000000, 0.000000; 0.000000, -1.000000, 0.000000, 0.000000, 61.000000, -1.414214, 0.000000, 0.000000, 0.000000, 0.000000; 0.000000, 0.000000, -1.414214, 0.000000, -1.414214, 60.000000, -1.414214, -1.414214, 0.000000, 0.000000; 0.000000, 0.000000, 0.000000, -1.732051, 0.000000, -1.414214, 61.000000, 0.000000, -2.000000, 0.000000; 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, -1.414214, 0.000000, 61.000000, -1.000000, 0.000000; 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, -2.000000, -1.000000, 61.000000, -1.732051; 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, -1.732051, 63.000000]

If i make a JAMA-matrix with these values, and execute the eigenvalue decomposition, V*D*transpose(V) does NOT equal the hamilton. Do any of you guys know what goes wrong? The eigenvalues are consistent with MATLAB, but the eigenvectors are not.

Here is a class to test it

public class TestJama {


public static void main(String[] args) {

    double[][] m = new double[][] {
            {63.000000, -1.732051, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000},
            { -1.732051, 61.000000, -2.000000, 0.000000, -1.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000},
            { 0.000000, -2.000000, 61.000000, -1.732051, 0.000000, -1.414214, 0.000000, 0.000000, 0.000000, 0.000000},
            { 0.000000, 0.000000, -1.732051, 63.000000, 0.000000, 0.000000, -1.732051, 0.000000, 0.000000, 0.000000},
            { 0.000000, -1.000000, 0.000000, 0.000000, 61.000000, -1.414214, 0.000000, 0.000000, 0.000000, 0.000000},
            { 0.000000, 0.000000, -1.414214, 0.000000, -1.414214, 60.000000, -1.414214, -1.414214, 0.000000, 0.000000},
            { 0.000000, 0.000000, 0.000000, -1.732051, 0.000000, -1.414214, 61.000000, 0.000000, -2.000000, 0.000000},
            { 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, -1.414214, 0.000000, 61.000000, -1.000000, 0.000000},
            { 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, -2.000000, -1.000000, 61.000000, -1.732051},
            { 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, -1.732051, 63.000000}
    };


    Matrix hamilton = new Matrix(m);
    System.out.println(jamaToString(hamilton));

    EigenvalueDecomposition e = hamilton.eig();
    System.out.println(jamaToString(e.getD()));
    System.out.println(jamaToString(e.getV()));

    Matrix recomb = e.getV().times(e.getD()).times(e.getV().transpose());
    System.out.println(jamaToString(recomb));

    System.out.println(hamilton.equals(recomb));

}

private static String jamaToString(Matrix m) {
    StringBuilder b = new StringBuilder();
    b.append("[");
    for(int i=0; i<m.getRowDimension(); i++) {

        for(int j=0; j<m.getColumnDimension(); j++) {
            b.append(m.get(i, j));
            if(j<m.getColumnDimension() - 1) b.append(",");
        }
        if(i<m.getRowDimension() - 1) b.append(";");
    }
    b.append("]");
    return b.toString();
}
}

EDIT: The result (V*D*transpose(V)) yields

63.1093 -0.6379 0.3400 -0.6576 0.0938 -0.0437 -0.6056 -0.5066 0.3463 0.5039 -0.6379 61.3082 -0.2368 -1.7813 -0.0851 0.7949 -0.1337 0.7668 -0.0422 -2.4329 0.3400 -0.2368 60.1481 1.3323 -0.4099 -1.8834 -0.5780 0.7516 0.0946 0.1705 -0.6576 -1.7813 1.3323 61.2455 0.0972 -0.8075 -0.9004 0.0242 0.3963 -1.2527 0.0938 -0.0851 -0.4099 0.0972 60.3086 -0.1899 0.0394 0.1987 -0.0484 -0.1495 -0.0437 0.7949 -1.8834 -0.8075 -0.1899 61.7941 0.3741 0.8237 0.7772 0.7557 -0.6056 -0.1337 -0.5780 -0.9004 0.0394 0.3741 60.6415 0.6351 0.7099 0.3349 -0.5066 0.7668 0.7516 0.0242 0.1987 0.8237 0.6351 62.8108 1.3507 1.3002 0.3463 -0.0422 0.0946 0.3963 -0.0484 0.7772 0.7099 1.3507 63.3270 0.1244 0.5039 -2.4329 0.1705 -1.2527 -0.1495 0.7557 0.3349 1.3002 0.1244 60.3069

Answer 1

If it is only about the value of the matrix: It is correct.

EDIT: The above statement was written before the resulting matrix was inserted into the original question. Of course, this matrix is not correct. According to the comments, it seems that this wrong matrix was caused by some bug in an older JAMA version. With JAMA 1.0.3 it worked. The remaining part of this answer is still valid and relevant for the question nevertheless:

There are two problems here. First of all, the equals method of the JAMA Matrix class is not overridden to compare the contents of the matrices. And the default implementation of equals compares for identity of the referred objects. So even a trivial comparison like this

Matrix A = new Matrix(new double[][]{{1.0}});
Matrix B = new Matrix(new double[][]{{1.0}});
System.out.println(A.equals(B));

would print false .

The second problem is rather simple (and common) : Computations with double values are not infinitely precise. You can find many questions about this on StackOverflow, but might want to have a look at Wikipedia about Floating Point Accuracy Problems . (Some people would recommend the article about What Every Computer Scientist Should Know About Floating-Point Arithmetic , but it's rather involved...).

The basic problem can easily reproduced with this small example:

double x = 0.1;
double y = 0;
for (int i=0; i<10; i++)
{
    y += x;
}
System.out.println(y+" == 1.0: "+(y==1.0));

The result of this computation is not y==1.0 , but y==0.99999999999999... instead.

So even if the equals method was overridden to perform an element-wise comparison, the comparison would still yield false due to the floating point error.

One way of alleviating this problem is to check whether some values are "equal up to a small epsilon". This may still be tricky, because the magnitude of the values to be compared will make a difference here, but one solution could be to use the following method to compare matrices for epsilon-equality:

private static boolean epsilonEqual(Matrix a, Matrix b)
{
    int ra = a.getRowDimension();
    int rb = b.getRowDimension();
    if (ra != rb)
    {
        return false;
    }
    int ca = a.getColumnDimension();
    int cb = b.getColumnDimension();
    if (ca != cb)
    {
        return false;
    }
    for (int c=0; c<ca; c++)
    {
        for (int r=0; r<ra; r++)
        {
            double ea = a.get(r, c);
            double eb = b.get(r, c);
            if (!epsilonEqual(ea, eb))
            {
                return false;
            }
        }
    }
    return true;
}

private static boolean epsilonEqual(double x, double y)
{
  final double epsilon = 1e-8;
  return Math.abs(x - y) <= epsilon;
}