Re: curry on windows (was: encapsulated search)

From: Claus Reinke <claus.reinke_at_talk21.com>
Date: Sat, 12 May 2007 01:38:26 +0100

> Claus Reinke wrote:
>> assuming standard (normal-order) definitions of normal form (nf),
>> head nf (hnf), weak hnf (whnf), the typical reduction to whnf
>> reduces less than it could, but any reduction beyond hnf, towards
>> full nf, implies reductions that might not be needed in context.
>> what is the rationale for this unusual choice?
>
> Could you provide a reference to a "standard" definition of
> normal form? The current handling of normal form comes
> from the area of term rewriting (which is somehow the basis of Curry)
> where a term is in normal form if it does not contain a redex.
> This implies reductions at arbitrary places.

i can try;-) first, some context (even if it is probably already
known here), then, some references.

the handling of normal forms in programming calculi differs only
in that it tends to be relative to a set of rewrite rules and possibly
rewrite strategies, so for a set of rewrite rules X, there are
X-normal-forms, where no more X-rules are applicable anywhere,
and there are XS-normal-forms, where no X-rules are applicable
according to strategy S.

another difference is that in term rewriting, the rewrite system itself
is the program, whereas in programming calculi, the rewrite system
is the meta-program that interprets object-level programs.

what i was referring to are beta-normal forms for lambda calculus
under normal-order reduction strategy, reducing until no beta-redices
are left (nf), reducing until no head redices are left (hnf), or not
reducing under lambdas at all (whnf).

normalising (or standard) reduction strategies are those that are
guaranteed to produce a normal form if one exists. for lambda-
calculus, that is normal-order reduction (also known as leftmost
outermost reduction).

because it always reduces the leftmost outermost redex next,
there is always at most one normal-order redex (or head redex);
normal-order reduction will first reach weak hnf, then hnf, then
normal form, provided that all of these exist.

any lambda-abstraction is in whnf, no matter what beta-redices
might occur in the body - whnf originated in implementations,
and was long considered of no theoretical interest.

a head normal form has no more beta-redices in head position,
so an hnf consist of a prefix of lambdas, followed by a variable
in head position, followed by a postfix of arbitrary expression
parameters (pre- and postfix possibly empty):

    hnf = \v1..vn->v e1..em -- n,m >= 0

there may be redices in the parameters, but until the variable in
head position is substituted, we can not know whether any of
those parameters will be needed. so it is, in general, not safe to
go any further - hnf is the most useful and sadly the least used
of those three normal forms.

a beta-normal form has no beta-redices left anywhere, so
taking nf as the target for reduction runs the danger of attempting
to reduce redices that wouldn't be needed in context. it mostly
makes sense if one wants to see every subterm fully evaluated.

as for references:

for normal-order reduction, normal form and head normal form,
i'd suspect Barendregt's book, which is "the" standard reference
on lambda-calculus anyway (i don't have my copy here, but i
think head normal form is defined in the context of solvability
of lambda-terms):

    Barendregt, H. P.
    The Lambda Calculus: Its Syntax and Semantics.
    North Holland, Amsterdam (1984).

the origin of weak head normal form is more difficult to pin down,
as it goes back to implementations simply not reducing under
lambdas, but perhaps the first to consider the resulting normal
forms as more than an accident was Plotkin, who defined a
call-by-value calculus with values as '\var-> <expression>',
to correspond to Landin's ISWIM languages and SECD
machine:

    Gordon Plotkin
    Call-by-Name, Call-by Value and the Lambda Calculus
    Reproduced from Theoretical Computer Science , Vol. 1,
    pp. 125-159, Copyright (1975)
    http://homepages.inf.ed.ac.uk/gdp/publications/

completing the triangle of call-by-name, call-by-value, and
call-by-need, and closest to the deterministic cores of Haskell
and Curry, is

  A Call-By-Need Lambda Calculus (1995)
  Zena M. Ariola, Matthias Felleisen, John Maraist,
  Martin Odersky, Philip Wadler
  Proceedings of 22nd Annual ACM SIGACT-SIGPLAN
  Symposium on Principles of Programming Languages (POPL)
  http://citeseer.ist.psu.edu/132534.html

and they take whnf as their notion of values..

as we have seen, head normal forms under normal-order and
call-by-need reductions don't quite coincide, due to observable
sharing, and reducing all the way to normal form might lose
such sharing.

hth,
claus

> On the other hand, the notion of normal form is not really defined
> in the Curry report so that it is more or less a question of
> the various implementations what they show the user at the top level.

true (and haskell isn't any more precise in its report, it even tries
to ignore sharing). it is just the inconsistencies i find confusing.


_______________________________________________
curry mailing list
curry_at_lists.RWTH-Aachen.DE
http://MailMan.RWTH-Aachen.DE/mailman/listinfo/curry
Received on Mo Mai 14 2007 - 08:49:55 CEST

This archive was generated by hypermail 2.3.0 : Do Jun 20 2024 - 07:15:09 CEST