[SciPy-User] linear algebra: quadratic forms without linalg.inv

Sturla Molden sturla@molden...
Sun Nov 1 21:03:14 CST 2009


josef.pktd@gmail.com skrev:
> In econometrics (including statsmodels) we have a lot of quadratic
> forms that are usually calculate with a matrix inverse. 
That is a sign of numerical incompetence.

You see this often in statistics as well, people who think matrix 
inverse is the way to calculate mahalanobis distances, when you should 
really use a Cholesky.

As for LU, I'd rather use an SVD as it is numerically more stabile. 
Using LU, you are betting on singular values not being tiny. With SVD 
you can solve an ill-conditioned system by zeroing tiny singular values. 
With LU you just get astronomic rounding errors.

Sturla


More information about the SciPy-User mailing list