[Numpy-discussion] A new matrix class
Sun May 11 15:06:22 CDT 2008
On Sun, May 11, 2008 at 12:44 PM, Charles R Harris
> On Sun, May 11, 2008 at 1:01 PM, Keith Goodman <firstname.lastname@example.org> wrote:
>> The most basic, and the most contentious, design decision of a new
>> matrix class is matrix indexing. There seems to be two camps:
>> 1. The matrix class should be more like the array class. In particular
>> x[0,:] should return a 1d array or a 1d array like object that
>> contains the orientation (row or column) as an attribute and x
>> should return a 1d array. (Is x.sum(1) also a 1d array like object?)
>> 2. A matrix is a matrix: all operations on a matrix, including
>> indexing, should return a matrix or a scalar.
>> Does that describe the two approaches to matrix indexing? Are there
>> other approaches?
> Pretty well, I think. The thing about 2) is that ndarray routines break if
> they can't treat arrays as nested sequences, i.e. scalar indexing needs to
> return an array of one less dimension. So the matrix class shouldn't
> subclass ndarray in that case, but rather should use an ndarray as a
> component. More code to write, but such is life.
> 3) Everything is an array. I think Matlab treats scalars as 1x1 arrays.
So #3 is behave in a similar way to octave/matlab? In octave/matlab
1x1 matrices are handled in special ways, e.g. (nxm) * (1x1) is
allowed. From the perspective of the user a 1x1 matrix is a scalar.
More information about the Numpy-discussion