Problem with numpy and pysparse floating point view

Question

I started using numpy with the pysparse package, which interacts with UMFPACK, however there is a problem with numpy floating point results. By the way, it is a lanczos eigenvalue solver for structural problems.

When I do the same operations in MATLAB, I get different results, well, the results are of the order of 1e-6,1e-8 and with the MATLAB representation, I get the correct eigenvalues. The results of NumPy and PySparse are also not so far, at least at the order level, but using them to create a tridiagonal matrix on which to find eigenvalues ​​is a source of problems. I could not understand what was going wrong, but the problem is the floating point representation, but how to fix this if possible? I tried to use "Float64" as my data type, but this does not change the results of the problem. For instance,

q = ones(n, dtype = 'Float64') 

Another one, what is the most mature sparse package for python and what interfaces are provided, if any? As said, PySparse seemed beautiful to me at first glance ...

Answer 1

float64 is the default data type in Numpy. You can try using float128 for greater accuracy, but warn that some functions (and basically everything on Windows) will force it to float64 anyway.

I would recommend using scipy.sparse for your sparse eigenvector problems. I tried PySparse and scipy.sparse, and I would conclude that although PySparse is easier to use, scipy.sparse is more mature.

A rare linear algebra is described here: http://docs.scipy.org/doc/scipy/reference/sparse.linalg.html