[Numpy-discussion] Toward release 1.0 of NumPy
tim.hochberg at cox.net
Thu Apr 13 16:37:04 CDT 2006
Charles R Harris wrote:
> On 4/13/06, *Tim Hochberg* <tim.hochberg at cox.net
> <mailto:tim.hochberg at cox.net>> wrote:
> Charles R Harris wrote:
> > Tim,
> > On 4/13/06, *Tim Hochberg* <tim.hochberg at cox.net
> <mailto:tim.hochberg at cox.net>
> > <mailto: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
> > 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
> > indices.
> > Anyway, it sounds like you are proposing that the tensor (outer)
> > product of two matrices be reshaped to run over two indices. It
> > that likewise the tensor (outer) product of two vectors should be
> > reshaped to run over one index ( i.e. flat). That would do the
> I'm not proposing anything. I don't care at all what kron does. I
> 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
> but I'm not sure so correct me if I'm wrong. Here's what kron does as
> >>> 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]])
> Good at first look. Lets see a simpler version... Nevermind, seems
> numpy isn't working on this machine (X86_64, fc5 64 bit) at the
> moment, maybe I need to check out a clean version.
> >>> kron(a,a) => 8x1
> array([1, 1, 1, 1, 1, 1, 1, 1])
> Looks broken. a should be an operator (matrix), so either it should
> be (2,1) or (1,2).
Since a is an array here, a is shape (2,). Let's repeat this
excercise using matrices, which are always rank-2 and see if they make
>>> kron(m, m)
matrix([[1, 1, 1, 1],
[1, 1, 1, 1]])
That looks OK.
> In the first case, the return should have shape (4,2), in the latter
> (2,4). Should probably raise an error as the result strikes me as
> ambiguous. But I have to admit I am not sure what the point of this
> particular construction is.
> >>> kron(a, a)
> 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
>>> kron(m, m)
matrix([[1, 1, 1, 1]])
>>> kron(m[:,0], m[:,0])
> See above. this could be (1,4) or (4,1), depending.
All of these look like they're probably right without thinking about it
> >>> b.shape
> (2, 2, 2)
> >>> kron(b,b).shape
> (4, 4, 2, 2)
> I think this is doing transpose(outer(b,b), axis=(0,2,1,3)) and
> reshaping the first 4 indices into 2. Again, I am not sure what the
> point is for these operators. Now another way to get all this
> functionality is to have a contraction function or method with a list
> of axis. For instance, consider the matrices A(i,j) and B(k,l)
> operating on x(j) and y(l) like A(i,j)x(j) and B(k,l)y(l), then the
> outer product of all of these is
> with the summation convention on the indices j and l. The result
> should be the same as kron(A,B)*kron(x,y) up to a permutation of rows
> and columes. It is just a question of which basis is used and how the
> elements are indexed.
> So, it looks like the 2d x 2d product obeys Alan's definition. The
> products are probably all broken.
Here's my best guess as to what is going on:
1. There is a relatively large group of people who use Kronecker
product as Alan does (probably the matrix as opposed to tensor math
folks). I'm guessing it's a large group since they manage to write the
definitions at both mathworld and planetmath.
2. kron was meant to implement this.
2.5 People who need the other meaning of kron can just use outer, so
no real conflict.
3. The implementation was either inappropriately generalized or it
was assumed that all inputs would be matrices (and hence rank-2).
Assuming 3. is correct, and I'd like to hear from people if they think
that the behaviour in the non rank-2 cases is sensible, the next
question is whether the behaviour in the rank-2 cases makes sense. It
seem to, but I'm not a user of kron. If both of the preceeding are true,
it seems like a complete fix entails the following two things:
1. Forbid arguments that are not rank-2. This allows all matrices,
which is really the main target here I think.
2. Fix the return type issue. I have a fix for this ready to commit,
but I want to figure out the first part as well.
More information about the Numpy-discussion