[Numpy-discussion] Maturing the Matrix class in NumPy

Alan G Isaac aisaac@american....
Fri Jun 5 16:14:39 CDT 2009

On 6/5/2009 3:49 PM Stéfan van der Walt apparently wrote:
> If the Matrix class is to remain, we need to take the steps
> necessary to integrate it into NumPy properly.

I think this requires a list of current problems.
Many of the problems for NumPy have been addressed over time.
I believe the remaining problems center more on SciPy rather than NumPy.
This requires that users report difficulties.

For example, Jason Rennie says he ran into problems with
scipy.optimize.fmin_cg, although I do not recall him reporting
these (I do recall an optimization problem he reported using
ndarrays).  Has he filed a bug report detailing his problem?

> To get going we'll need a list of changes required (i.e. "in an ideal
> world, how would matrices work?").

The key anomaly concerning matrices comes with indexing.
See the introduction here: http://www.scipy.org/MatrixIndexing

However changing this for the current matrix object was rejected
in the last (exhausting) go round.

> There should be a set protocol for
> all numpy functions that guarantees compatibility with ndarrays,
> matrices and other derived classes.

My impression was that this was resolved as follows:
handle all ndarray based objects as arrays (using asarray)
in any NumPy function, but return the subclass when possible.
(E.g., using asmatrix, return a matrix output for a matrix input.)
This seems fine to me.

> Being one of the most vocal proponents of the Matrix class, would you
> be prepared to develop your Matrix Proposal at
> http://scipy.org/NewMatrixSpec further?

I consider my proposal to have the following status: rejected.
I consider the core reason to be: creates a backwards incompatibility.
That was a very long and exhausting discussion that was productive
in laying out the issues, but I do not think we can progress in that

The existing matrix object is very usable.
It's primary problem is some indexing anomalies,
and not everyone saw those as problems.
In terms of NumPy functions, I think the asarray/asmatrix
protocol fits the bill.  (Altho perhaps I am overlooking
something as a user that is obvious to a developer.)


More information about the Numpy-discussion mailing list