SciPy's eigh - How to Lose an Afternoon

So the SciPy package for Python is full of lots of goodness.  Be very careful using the eigh() routine to get eigenvalues, though.  There's a "gotcha" buried in it.

The syntax is:

(eigenvalues, eigenvectors) = eigh(matrix)

This returns an array of eigenvalues and a 2D array of eigenvectors (each eigenvector consists of many components).

And there's the gotcha.  Let's say you want the nth eigenvalue and eigenvector.  I would write:

eigenvalues[n]
eigenvectors[n]

And I would be horribly wrong.  The eigenvectors and eigenvalues do share an index, but the index on the eigenvector is SECOND column:

eigenvectors[:,n]

Seriously WTF.

Update: It turns out that this is due to eigh() being part of LAPACK, which is a non-Python library (Fortran?) that has a different internal storage model for 2D arrays.  So possibly less WTF, but still dangerous as all hell.

Comments

[WeBru]

In my hands that's not how it behaves, but maybe it was fixed since you tried it?

[WeBru]

Correction: I was using the numpy.linalg.eigh. It works like you expect. I see in the scipy.linalg.eigh that it works the WTF way... but if you are used to SVD this way also makes sense.

[WeBru]

Actually, in my hands, the scipy.linalg.eigh documentation is wrong and the behavior is like you expected. But numpy.linalg.svd is the WTF way, but I think it's traditional for SVD.