[Numpy-discussion] Further matrix oddities: no inner product
Charles R Harris
charlesr.harris@gmail....
Tue Mar 27 22:01:04 CDT 2007
In mathematics, and inner product is a sesquilinear form on pairs of
vectors, so at the least it should return a scalar. In numpy inner is a sum
over the last indices. OK, so we have
In [10]: inner(ones(2),ones(2))
Out[10]: 2.0
This doesn't work as an inner product for column vectors, which would be the
usual textbook convention, but that's alright, it's not a 'real' inner
product. But what happens when matrices are involved?
In [11]: inner(mat(ones(2)),mat(ones(2)))
Out[11]: array([[ 2.]])
Hmm, we get an array, not a scalar. Maybe we can cheat
In [12]: mat(ones(2))*mat(ones(2)).T
Out[12]: matrix([[ 2.]])
What about vdot (conjugate of the mathematical convention, i.e., the Dirac
convention)
In [17]: vdot(mat(ones(2)),mat(ones(2)))
---------------------------------------------------------------------------
exceptions.ValueError Traceback (most recent
call last)
/home/charris/<ipython console>
ValueError: vectors have different lengths
In [18]: vdot(mat(ones(2)),mat(ones(2)).T)
---------------------------------------------------------------------------
exceptions.ValueError Traceback (most recent
call last)
/home/charris/<ipython console>
ValueError: vectors have different lengths
Nope, vdot doesn't work for row and column vectors. So there is *no* builtin
inner product that works for matrices. I wonder if we should have one, and
if so, what it should be called. I think that vdot should probably be
modified to do the job. There is also the question of whether or not v.T *
v should be a scalar when v is a column vector. I believe that construction
is commonly used in matrix algebra as an alias for the inner product,
although strictly speaking it uses the mapping between a vector space and
its dual that the inner product provides.
Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.scipy.org/pipermail/numpy-discussion/attachments/20070327/894308fa/attachment-0001.html
More information about the Numpy-discussion
mailing list