> I was glad to see formula coming up over your side of the coding
> divide;  we've also been working on it a little in nipy.

> I think we're beginning to get the idea, and it seems like it's a very
> good idea, so it would be very good to share some work in increasing
> understanding.

> First - Jonathan Taylor - who wrote this code - has commented on the
> design a bit on the NIPY mailing list - we've put his comments into
> our version of the code in:

> http://github.com/nipy/nipy/blob/master/nipy/modalities/fmri/formula.py
> http://github.com/nipy/nipy/blob/master/nipy/modalities/fmri/aliased.py

> The original comments are here:

> http://mail.scipy.org/pipermail/nipy-devel/2009-December/002387.html

>> here is my *current* opinion and information on it, one example is below

>> - it looks useful when working with categorical data
>>   (I don't think it's very useful for non-categorical data, as we
>> where working on in the last year.)
>> - it's very difficult to understand because of all the indirections
>>   a class produces a class which returns another class, and how do I
>> get the data in and out?  ...

> I think here you mean the aliased functions.   That stuff is - as you
> know - Sympy code, and after looking at it a while (with Ondrej Certik
> last week for example), I think we concluded that the indirection
> stuff could not be avoided because of the way Sympy makes new
> functions.   However, with a nice set of docstrings and examples, I
> honestly think that won't matter because the concept of the
> AliasedFunction is actually very simple, and goes like this:

> An AliasedFunction is a symbolic function that carries with it its own
> implementation as a numerical function.  When we call 'lambidify' for
> expressions containing this function, the function's own
> implementation is used.

>> - the only usage examples, I have seen are in the test file

> We could certainly do with some more examples, but there are some you
> may not have found in our tree:

> http://github.com/nipy/nipy/tree/master/examples/formula/

>> - in order to use it, we need to create examples and documentation for it
>> - I'm not sure everything is correct, in spite of passing the tests,
>>   or maybe I don't get expected results because I don't know how to use it
>> - because of all the indirection, usage inside statistical models makes them
>>   also difficult to understand
>> - it doesn't do the data handling, which has been more the focus of Skipper and
>>   Vincent, the "namespace" in formulas need to be filled with actual data

>> For now, I don't have the time to struggle with it for several days, but maybe
>> during summer, if I don't think distributions, time series analysis and lot's
>> of other things are more fun.

> Maybe more fun with some help?  We're very keen to work on this (at
> least, I know I am).

> Thanks for thinking and working on this.   I do think it could be very useful,

> Matthew