Re: Moving code from octave-forge to octave [Was: polyderiv problem?]

[Michael Creel]

On Thursday 10 February 2005 16:00, you wrote:
> On  9-Feb-2005, Michael Creel <address@hidden> wrote:
> | On Wednesday 09 February 2005 16:50, David Bateman wrote:
> | > The only
> | > remaining patch to consider is M-v-ops-2.1.31.patch that allows the
> | > extension of ./, .*, etc to allow Matrix by vector arguments.
> |
> | I sent this message to David, but I meant to send it to all, so here it
> | is:
> |
> | This is a behavior that I have found very useful in languages like Gauss
> | and Ox.
>
> Do these languages have special syntax for these operations?
No, .* gives the functionality that the function I hacked together provides. 
The same .* will do ordinary element-by-element multiplication.
>
> Although it has been discussed before, and I have probably even
> suggested implementing it in the past, I now hesitate to overload
> operators like .* for this purpose because (as David mentioned) it
> means that a dimensioning error can lead to bad results with no
> warning.  As another example, you might mean to multiply elements of
> M1 by elements of M2, but M2 accidentally becomes a vector, so
> suddenly you are scaling columns instead of seeing an error about
> non-conforming operands.  That seems bad to me.  But it would be
> useful to have this capability implemented efficiently, even if we do
> not have special operators for it.
I agree, this opens the door for unpredictable results due to errors.

>
> | Maybe it would be possible to allow it only for matrices and vectors, but
> | not NDArrays?
>
> There is probably some way to generalize to N-d arrays when the
> dimension of one operand is one less than the dimension of the other.
>
> | Or maybe a function would do the job. I attach a quck hack for
> | consideration. Michael
> | m_x_v.m
> | function result = m_x_v(m,v)
> |  if !ismatrix(m)
> |   error("m_x_v: first arg must be a matrix");
> |  endif
> |  if !isvector(v)
> |   error("m_x_v: second arg must be a vector");
> |  endif
> |
> |  [rm, cm] = size(m);
> |  [rv, cv] = size(v);
> |
> |  if (rm == rv)
> |   v = kron(v, ones(1,cm));
> |   result = m .* v;
> |  elseif (cm == cv)
> |   v = kron(v, ones(rm, 1));
> |   result = m .* v;
> |  else
> |   error("m_x_v: dimension of vector must match one of the dimensions of
> | the matr ix");
> |  endif
> | endfunction
>
> This kind of approach will work, but it will use a lot of memory, so
> it would be better to do this in a built-in function where we could
> use loops and not sacrifice speed or memory efficiency.

The thing I like about this is that a user must explicitly call a function by 
name to get the behavior, rather than type ".*" which may or may not give the 
desired results. The function gives an error message if the second argument 
is not a vector. As to memory and speed, one could write an .oct version if 
desired. The function is just to point out that the behavior can be obtained 
easily without changing .*

Michael