Yes, I understand. But the beginning of this thread suggests that the two functions might be easily confused and mahalanobis does not look all that useful to warrant keeping it alive. So in my opinion it would be better to have mahalanobis deprecated in core and not keep it alive indefinitely in the statistics package.

Thanks, Carnë,
I think it would be quite confusing to have both mahal and mahalanobis in the statistics package. For compatibility reasons mahal is the better option of the two, so it is good as it is now.

Lachlan, thank you for the explanation about what the function ought to be doing. You don't have to do anything for now, I will take care of adding mahal to Octave Forge.

When I submitted this bug report, I sent an email to Peter John Acklam asking for permission to use his implementation which also has some extra features. I will compare them and take care of the tests.

About the current one, I guess it would move to the statistics package too. Arno Onken will weight on that. For me it makes more sense to have only one function that does this.

Thanks, Rik,

Is the  mahal.m  that I sent satisfactory?  Is there anything else I need to do?

Out of interest, what was the rationale for deprecating mahalanobis?  As I pointed out, it is a different function from mahal.

Since it isn't a common function, and isn't in Matlab, it is likely that nobody was using it, but I don't understand the motivation for removing it.  If there was an existing log10 but not log, would you remove log10 when log was implemented?  This is much worse than that, because  log10  can easily be calculated from log,  but  mahalanobis can't (as?) easily be calculated from  mahal.

Greetings Carnë,

The function seem to do what it claims to do: it calculates the (single) distance between the points in X and the points in Y, assuming both are drawn from a Gaussian whose mean and variance is a fit to [X;Y].

This function has both a different name and different behaviour from Matlab's mahal function; was the issues that you thought it was mahal?

Attached are both an implementation of a Matlab-compatible  mahal  function, and a patch to Octave's  mahalanobis  function to improve its efficiency.  I'm still not familiar with Octave's test mechanism, so could somebody please check that the new test of  mahal  passes.

(file #34832, file #34833)

The function mahalanobis doesn't seem to work at all (note that it should return one value for each row of X).

octave> X = rand (10, 2);
octave> Y = rand (4, 2);
octave> mahalanobis (X, Y)
ans =  2.6666

I'm unsure since when it stopped working but the 7 tests it has, only check that it returns an error when input is complete garbage.

In a separate note, this function actually belongs to the statistics package named mahal.