Crash in signal.lfilter on nd input with wrong shaped zi

[ewmoore]

import numpy as np
from scipy import signal

y, zf = signal.lfilter([1,1], [1.], np.arange(12).reshape((4,3)), 0, [[1]])

On Windows 7, python 2.7, scipy 0.15.1 from anaconda64 this crashes.

Running on the same on Linux (python 2.7, scipy 0.15.1, from anaconda64) i get:

In [7]: y
Out[7]:
array([[  1.,   1.,   2.],
       [  3.,   5.,   7.],
       [  9.,  11.,  13.],
       [ 15.,  17.,  19.]])

In [8]: zf
Out[8]: array([[ 9.]])

Which still seems wrong, shouldn't this fail and complain about a zi having the wrong shape? Even if it would take a (1,1) shaped zi shouldn't the produced zf be (4,1) in this case? AFAICT, the handling of initial conditions for multidimensional x is not quite right. For instance, there is also this issue:

In [16]: signal.lfilter([1, -1], [1.], np.arange(10).reshape(2,5), 1, [[1],[1],[1],[1]])
Out[16]: 
(array([[ 1.,  1.,  1.,  1.,  1.],
        [ 6.,  1.,  1.,  1.,  1.]]), array([[ -4.00000000e+00],
        [ -9.00000000e+00],
        [  2.52787376e-45],
        [  2.64228486e-58]]))

Which gives the same uninitialized junk on both Windows and Linux. This makes me think the code that does not crash on Linux is just getting lucky and not crashing even though it is still stepping past the array bounds.

[rgommers]

that doesn't look good.

signal.lfilter([1, -1], [1.], np.arange(10).reshape(2,5), 1, [[1],[1],[1],[1]])

segfaults for me on 32-bit linux

[WarrenWeckesser]

Looks like the shape of the zi argument should be validated before lfilter makes the call to sigtools._linear_filter(b, a, x, axis, zi). For example, if x has shape (m, n, p) and axis=1, then zi should have shape (m, r, p), where r = max(len(a), len(b))-1.

[ewmoore]

Yes, I think that's exactly the problem. I'll have a fix shortly.