Re: Patching cov.m on current stable

[Ismael Núñez-Riboni]

> We've discussed cov incompatibilities before.  Please look at this
> thread:
>
> http://octave.1599824.n4.nabble.com/covariance-matrix-td1607376.html#a1607378

I see now that Octave's cov function has advantages over the Matlab 
version, which seems to me (without having a Matlab license to try on) 
that can only be mimicked in Matlab with a loop... I agree thus that 
Octave's cov function is better than the Matlab version. However, I 
disagree with the behavior being more intuitive than in Matlab. Contrary 
to the comments in that link, I see the actual behavior of Octave's cov 
function as counter-intuitive but I must say I'm strongly biased by 
almost 20 years of Matlab use. I think Matlab's behavior seems more 
logical for observational scientists (I'm myself an observational 
physicist) and engineers (it might be non intiutive for mathematicians 
or theoretical physicists, though): Normally you have two sets or time 
series X,Y and in matlab you get besides the covariance between the two 
vectors also their auto-covariances, i.e., you obtain the 2x2 covariance 
matrix. In octave you only get the covariance between the vectors unless 
you join them into a Nx2 matrix. I don't think that ex matlab users will 
easily get to find out that, even though the help of the Octave's 
function is clear enough and correct... to find out the mismatch with 
matlab's version you really have to read the help text carefully or get 
an error message when you are porting your code to Octave (my case).

I can suggest thus to write a warning in the help of the functions which 
are supposed to work differently than in Matlab, something like:

Warning: please note the behavior of cov is different than in matlab!

And/or to detail the behavior of the function with a couple of examples 
to drag attention to a different behavior. In Matlab's help this 
information:

"cov(X) is the covariance matrix. diag(cov(X)) is a vector of variances 
for each column, and sqrt(diag(cov(X))) is a vector of standard deviations"

is completely redundant, you can live without it, but it helps a lot to 
many users to understand the behavior of the function. Why not to 
include something similar in Octave's cov function to clearly state that 
its behavior is different than in Matlab? I can suggest to add to 
octave's cov help:

For two column vectors X and Y, cov([X,Y]) yields the covariance matrix 
between X and Y, while cov(X,Y) yields the scalar covariance.

I stress it: I am aware that this information is redundant with what 
already is written in the cov's help and most probably a new user to 
Octave will not need this explanation, but an ex matlab user could 
really get into troubles without it...

Finally: I guess you are aware that with Octave's cov function you are 
breaking compatibility with Matlab, even though the Octave's version of 
cov is better... But I guess that the decision of sticking to Matlab 
compatibility or improving functionality is not easy... For me, the more 
and more I use Octave I kind of think I don't need Matlab's 
compatibility anymore, but unfortunately I still exchange a lot of code 
with Matlab's users... Why not, in the case, to make "second version 
functions" (like with a "2" at the end, like cov2, or an "i" of 
"improved", like cov_i) for those functions incompatible with Matlab but 
which should exist because they are better?

Cheers, Ismael.