On Thu, Apr 28, 2011 at 11:32 AM, Jake Mannix <[email protected]> wrote:
>
> You're probably correct here. I thought there was a unit test in
> LanczosSolverTest
> which compared eigenvalues to that of using the old COLT Eigendecomposition
> class, like "testEigenvalue()" or something like that? I may have written
> that test
> wrong, and then convinced myself I had to make the test pass... :\
>
Arg:
lines 38-48 of TestLanczosSolver:
Matrix m = randomHierarchicalSymmetricMatrix(size);
int desiredRank = 80;
Vector initialVector = new DenseVector(size);
initialVector.assign(1d / Math.sqrt(size));
LanczosSolver solver = new LanczosSolver();
LanczosState state = new LanczosState(m, size, desiredRank,
initialVector);
// set initial vector?
solver.solve(state, desiredRank);
EigenvalueDecomposition decomposition = new EigenvalueDecomposition(m);
...
LanczosSolver.solve(LanczosState, int) -> .solve(LanczosState, int, false)
ie defaults to assuming *not* symmetric (so uses timesSquared()), but then
what was passed in was a symmetric matrix.
Of course, it's also not really checking (approximate) equality of the
eigenvalues,
even though it's printing it out. That test should also be checking that
the
COLT-derived eigenvalue is not too different from the Lanczos one.
-jake
