Agree with Nick; surely Mata will have the full precision available to
the combination of hardware and OS. Although I have no inside
information, I expect what this statement means (coupled with the
example displayed on the web site) is that matrix languages (e.g. GAUSS
or MATLAB) are generally hideously slow when operating on an
element-by-element level as one would in FORTRAN. Those matrix
languages are very fast when you use their built-in matrix operations.
Version 8 of Stata works the same way -- if you wrote an ado-file to
multiply matrices, it would be much slower than saying
mat c = a*b
I believe what Mata's authors are pointing out is that their approach
(using compiled-on-the-fly code) is so efficient that even if you
directly translate FORTRAN element-by-element code to Mata, it will run
like a bandit. This provides a nice intermediate path between using
ado-file code to do heavy computation (which can be very slow) and
going to the trouble of writing a C plugin. In Stata pre v9, the only
way to speed up something like gllamm is to recode some or all of it in
C, which makes maintenance and enhancement of the code more difficult.
Once such routines can be written in Mata -- a higher-level language
than C -- it should be possible to get more speed (and all avaiilable
accuracy) without sacrificing readability (or having to figure out how
to keep track of those %%^$% pointers!)
Again, just my interpretation of what I think we will see in Mata. I
have seen nothing more of it than what is on the website.
Kit Baum, Boston College Economics
http://ideas.repec.org/e/pba1.html
On Apr 13, 2005, at 2:33 AM, Renzo wrote:
At the page
http://www.stata.com/stata9/mata.html
it says "Everybody knows that matrix languages evaluate matrix
expressions,
such as b=invsym(X'X)*X'y, and Mata is no exception. Because of Mata's
design, however, it is fast enough to work at the element level".
This looks like a disclaimer of some sort about some loss of
precision, but
I got quite understand the implications of this.
Can anybody clarify?
How different is it (if any) from the way the same calculation was
done in
Stata8?
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
