Path: utzoo!attcan!uunet!mstan!amull
From: amull@Morgan.COM (Andrew P. Mullhaupt)
Newsgroups: comp.lang.apl
Subject: Re: Syntax for extended matrix operations
Summary: Clarification
Keywords: APL primitives, beyond quad-divide
Message-ID: <566@s5.Morgan.COM>
Date: 3 Dec 89 07:55:57 GMT
References: <539@s5.Morgan.COM> <6458@tank.uchicago.edu> <558@s5.Morgan.COM> <6510@tank.uchicago.edu>
Organization: Morgan Stanley & Co. NY, NY
Lines: 30


In response to your response (which for some reason won't include in
this response (?!)).

We are writing a protable APL interpreter. That's why I want to know
if there are good matrix primitives for modern numerical linear
algebra. I think the only hope to get over the sluggishness of
'one at a time' operations in APL is to devise elegant primitives
for the new language. The AP approach seems better at first, but
then you get the 'quality of implementation' and 'differential
portability' problems. Since we are the originators of our own APL
(with most of ISO plus some Dictionary and even a few other things)
we have to ask the question of what should be in it? Iverson's
first book on APL included data types and primitives not yet
included in APL, so it seems clear that rethinking the primitives
has been done before. My proposed diagonal primitive turns out to
be close to something proposed at SIGAPL 1981. 

We must keep in mind that APL became a language before the advent
of widespread acceptance of orthogonal matrix factorizations for
solving least squares problems, so in some sense quad divide is
not far off the state of the art circa 1965. Why should APL remain
interested in a primitive which can be almost always dominated in
usefulness and efficiency by a different set? Any new primitives
would completely embrace the current functionality of quad divide
and machine updating existing code would be nearly trivial, so
what's the big deal?

Later,
Andrew Mullhaupt