[Numpy-discussion] Toward release 1.0 of NumPy

[Tim Hochberg]

Charles R Harris wrote:

> Tim,
>
> On 4/13/06, *Tim Hochberg* <tim.hochberg at cox.net 
> <mailto:tim.hochberg at cox.net>> wrote:
>
>     Alan G Isaac wrote:
>
>     >On Thu, 13 Apr 2006, Charles R Harris apparently wrote:
>     >
>     >
>     >>The Kronecker product (aka Tensor product) of two
>     >>matrices isn't a matrix.
>     >>
>     >>
>     >
>     >That is an unusual way to describe things in
>     >the world of econometrics.  Here is a more
>     >common way:
>     >http://planetmath.org/encyclopedia/KroneckerProduct.html
>     <http://planetmath.org/encyclopedia/KroneckerProduct.html>
>     >I share Sven's expectation.
>     >
>     >
>     mathworld also agrees with you. As does the documentation (as best
>     as I
>     can tell) and the actual output of kron. I think Charles must be
>     thinking of the tensor product instead. 
>
>
> It *is* the tensor product, A \tensor B, but it is not the most 
> general tensor with four indices just as a bivector is not the most 
> general tensor with two indices. Numerically, kron chooses to 
> represent the tensor product of two vector spaces a, b with dimensions 
> n,m respectively as the direct sum of n copies of b, and the  tensor 
> product of two operators takes the given form. More generally, the B 
> matrix in each spot could be replaced with an arbitrary matrix of the 
> correct dimensions and you would recover the general tensor with four 
> indices.
>
> Anyway, it sounds like you are proposing that the tensor (outer) 
> product of two matrices be reshaped to run over two indices. It seems 
> that likewise the tensor (outer) product of two vectors should be 
> reshaped to run over one index ( i.e. flat). That would do the trick.

I'm not proposing anything. I don't care at all what kron does. I just 
want to fix the return type if that's feasible so that people stop 
complaining about it. As far as I can tell, kron already returns a 
flattened tensor product of some sort. I believe the general tensor 
product that you are talking about is already covered by multiply.outer, 
but I'm not sure so correct me if I'm wrong. Here's what kron does as 
present:

 >>> a
array([[1, 1],
       [1, 1]])
 >>> kron(a,a) # => 4x4 matrix
array([[1, 1, 1, 1],
       [1, 1, 1, 1],
       [1, 1, 1, 1],
       [1, 1, 1, 1]])
 >>> kron(a,a[0]) => 8x1
array([1, 1, 1, 1, 1, 1, 1, 1])
 >>> kron(a[0], a[0])
Traceback (most recent call last):
  File "<stdin>", line 1, in ?
  File "C:\Python24\Lib\site-packages\numpy\lib\shape_base.py", line 
577, in kron
    result = concatenate(concatenate(o, axis=1), axis=1)
ValueError: 0-d arrays can't be concatenated
 >>>  b.shape
(2, 2, 2)
 >>> kron(b,b).shape
(4, 4, 2, 2)

So, it looks like the 2d x 2d product obeys Alan's definition. The other 
products are probably all broken.

Regards,

-tim