[Matrix-SIG] Bug in convolve.

[Tim Hochberg]

While working on JNumeric I found what appears to be a bug in convolve.
Here's an illustration:

CNumeric:
>>> from Numeric import *
>>> a = arange(5)
>>> convolve(a,a,2)
array([ 0,  4, 11, 20, 30, 20, 11,  4,  0])
>>> convolve(a,a[::-1],2)
array([ 0,  0,  1,  4, 10, 20, 25, 24, 16])
>>> convolve(a[::-1],a,2)
array([16, 24, 25, 20, 10,  4,  1,  0,  0])

JNumeric:
>>> from Numeric import *
>>> a = arange(5)
>>> convolve(a,a,2)
array([ 0,  0,  1,  4, 10, 20, 25, 24, 16])
>>> convolve(a, a[::-1], 2)
array([ 0,  4, 11, 20, 30, 20, 11,  4,  0])
>>> convolve(a[::-1], a, 2)
array([ 0,  4, 11, 20, 30, 20, 11,  4,  0])

One of these is obviously wrong. Since convolve(a, b) = convolve(b,a) should
hold, the culprit appears to be CNumeric's convolve. It appears that:

convolve_CNumeric(a, b, mode) = convolve_JNumeric(a, b[::-1], mode)

This means a quick fix could be implemented in Python by defining:

def convolve(a, b, mode=0):
    return convolve(a, b[::-1], mode)

in Numeric.

____
 /im (tim.hochberg@ieee.org)