[SciPy-User] How to calculate Yulewalk with scipy.optimize.leastsq

Sturla Molden sturla@molden...
Fri Jan 20 06:45:31 CST 2012

Den 18.01.2012 08:50, skrev Fabrice Silva:
> Note that talkbox seems to have some stuff on Yule-Walker
> http://www.ar.media.kyoto-u.ac.jp/members/david/softwares/talkbox/talkbox_doc/index.html
> in python for educational purpose, and C for performance.

No need to use C for performance here.

Computing the autocovariance for Yule-Walker can be vectorized with 
np.dot, which lets BLAS do the work. Something like this:

def covmtx_yulewalker(x,p):
     ''' autocorrelation method '''
     x = np.ascontiguousarray(x)
     n = x.shape[0]
     Rxx = np.zeros(p+1)
     for k in range(0,p+1):
         Rxx[k] = np.dot(x[:n-k],x[k:])/(n-k-1.0)
     return Rxx

Later on, in the code Josef posted, the next bulk of the computation is 
done by LAPACK (linalg.lstsq).

With NumPy linked against optimized BLAS and LAPACK libraries (e.g. MKL, 
ACML, GotoBLAS2, Cray libsci), doing this in C might actually end up 
being slower. Don't waste your time on C before (1) NumPy is proven to 
be too slow and (2) you have good reasons to believe that C will be 
substantially faster. (NumPy users familiar with MATLAB make the latter 
assumption far too often.)


More information about the SciPy-User mailing list