plt-redex: capture-avoiding substitution for free? - racket

Every time I define a language in PLT redex, I need to manually define a (capture-avoiding) substitution function. For example, this model isn't finished because subst isn't defined:
#lang racket/base
(require redex/reduction-semantics)
(define-language Λ
[V ::= x (λ x M)]
[M ::= (M M) V]
[C ::= hole (V C) (C M)]
[x ::= variable-not-otherwise-mentioned])
(define -->β
(reduction-relation Λ
[--> (in-hole C ((λ x M) V))
(in-hole C (subst M x V))]))
But the definition of subst is obvious. Can PLT redex handle substitution automatically?

Yes! Just describe your language's binding structure with a #:binding-forms declaration.
Here's a similar model with capture-avoiding substitution via the substitute function:
#lang racket/base
(require redex/reduction-semantics)
(define-language Λ
[V ::= x (λ x M)]
[M ::= (M M) V]
[C ::= hole (V C) (C M)]
[x ::= variable-not-otherwise-mentioned]
#:binding-forms
(λ x M #:refers-to x)) ;; "term M refers to the variable x"
(define -->β
(reduction-relation Λ
[--> (in-hole C ((λ x M) V))
(in-hole C (substitute M x V))]))
(apply-reduction-relation -->β
(term ((λ x (λ y x)) y)))
;; '((λ y«2» y))
Alphabetic equivalence comes for free too, see alpha-equivalent?
(Thank you Paul Stansifer!)

Related

The Little Typer. I don't understand the meaning of The Initial Second Commandment of λ

I have tried following examples, but no matter y occurred or not,
The function f returns the same value as (λ(y)(f y)) after application.
I would like to do is to define a function that is not the same (-> Y X) as (λ (y)(f y)) when y occurred in y as a counter example, but I don't know how.
Do I misunderstand the meaning of The Initial Second Commandment of λ?
;;y does not occurs
(claim f (-> Nat Nat))
(define f
(λ(y)
0))
;; both return (the Nat 0)
(f 5)
((the (-> Nat Nat)
(λ(y)
(f y)))
5)
;; y occurs
(claim g (-> Nat Nat))
(define g
(λ(y)
y))
;;both return (the Nat 5)
(g 5)
((the (-> Nat Nat)
(λ(y)
(g y)))
5)
In order to create an example illustrating the importance of the caveat "...as long as y does not occur in f", we need to create a function f in which a name y occurs free. The provision that y is free in f is critical. This is also why it is difficult to create such an example: (top-level) functions cannot contain free variables. However, functions that are interior to other functions can. This is the key.
Here is function g that contains another function inside of it:
(claim g (-> Nat
(-> Nat
Nat)))
(define g
(lambda (y)
(lambda (x) ;; Call this inner
y))) ;; function "f"
(I've chosen to write the claim in this way to emphasize that we are thinking about a function inside of a function.)
To get our bearings, this simple function g expects two Nat arguments, and returns the first.
Let's call the inner function f. Note that f contains a free variable y (for this reason, f is meaningless outside of g). Let's substitute (lambda (y) (f y)) for f:
(claim g1 (-> Nat
(-> Nat
Nat)))
(define g1
(lambda (y)
(lambda (y) ;; Here we've replaced "f"
((lambda (x) ;; with an eta-expanded
y) ;; version, introducing
y)))) ;; the name "y"
We can eliminate the application to produce the following expression:
g1
---------------- SAME AS
(lambda (y)
(lambda (y)
((lambda (x)
y)
y)))
---------------- SAME AS
(lambda (y)
(lambda (y)
y))
---------------- SAME AS
(lambda (y)
(lambda (y1)
y1))
In the last step, I've renamed the second y to y1 to illustrate that the variable in the body of the inner function refers to the closer binding site, and not the farther one.
To recap, we started with a function g that "takes two (curried) arguments and returns the first". We then introduced a faulty eta-expansion around the inner function. As a result, we ended up with a function g1 that "takes two (curried) arguments and returns the second". Clearly not equivalent to the original function g.
So this commandment is about variable capture, which is the price we pay for working with names. I hope that helps!
IMPORTANT NOTE:
Due to the way that Pie checks types, you will need to introduce an annotation in the body of g if you want to try this example out:
(claim g1 (-> Nat
(-> Nat
Nat)))
(define g1
(lambda (y)
(lambda (y)
((the (-> Nat Nat)
(lambda (x)
y))
y))))

Extending a reduction relation

While taking a look at PLT redex, I wanted to play with simplification rules; so I defined this minimal language for booleans:
(define-language B0
(b T F (not b)))
I wanted to simplify a chain of (not (not ...)) so I extended the language to deal with contexts and defined a reduction relation to simplify the not:
(define-extended-language B1 B0
(C (not C) hole)
(BV T F))
(define red0
(reduction-relation
B1
(--> (in-hole C (not T)) (in-hole C F))
(--> (in-hole C (not F)) (in-hole C T))))
Now I wanted to extend my language to boolean equations and to allow not-simplification at each side of the equation, so I defined:
(define-extended-language B2 B1
(E (= C b) (= b C)))
hoping that:
(define red1
(extend-reduction-relation red0 B2))
will do the thing.
But no: red1 can reduce (not (not (not F))))) but not (= (not T) F)))
Am I doing something really silly here?
The problem with red1 is that it only contains the rules of red0 which use the limited context C. To make it work as expected you could either add the old rules modified to use E or make somehow the final extended context have the name C. One not very tedious approach could be:
(define-language L)
(define R
(reduction-relation L
(--> (not T) F)
(--> (not F) T)))
(define-language LB
(b T F (not b))
(C (compatible-closure-context b)))
(define RB (context-closure R LB C))
(define-extended-language LBE LB
(e (= b b))
(C .... (compatible-closure-context e #:wrt b)))
(define RBE (extend-reduction-relation RB LBE))
Note that this doesn't work in some older versions.
Two sources of useful information are this tutorial and of course the redex reference.

Redex Does Not Match

A common way of defining semantics is (for example):
return v if [some other condition]
otherwise, return error
For example, consider
(define-language simple-dispatch
(e ::= v (+ e e))
(v ::= number string)
(res ::= e err)
(E ::= hole (+ E e) (+ v E)))
We could then define the reduction relation
(define s-> (reduction-relation simple-dispatch
#:domain res
(--> (in-hole E (+ number_1 number_2))
(in-hole E ,(+ number_1 number_2)))
(--> (in-hole E (+ any any))
err)))
This is the natural way to do this, because it avoids having to write individual matchers for each of the 3 failure cases (number string, string number, string string). However, it then creates the problem that running it like this:
(apply-reduction-relation s-> (term (+ 2 2)))
Shows (correctly) that it lets you reduce both to an error or to the number 4. Is there a way to make an "except" pattern that avoids having to check all of the constituent cases?
What you want to use here is a combination of side-condition and redex-match?. Extending your reduction-relation gives:
(define s-> (reduction-relation simple-dispatch
#:domain res
(--> (in-hole E (+ number_1 number_2))
(in-hole E ,(+ (term number_1) (term number_2))))
(--> (in-hole E (+ any_1 any_2))
err
(side-condition
(not (redex-match? simple-dispatch
(+ number number)
(term (+ any_1 any_2))))))))
This just says you can take the second rule so long as the first one is not true, which is what the papers are saying implicitly, and just didn't draw out explicitly in the figure. (Note that you can use side-condition/hidden to get it to not draw the side condition when rendering the figure).
You can use this method to scale up to any number of patterns you want to disallow.

fixed point combinator in lisp

;; compute the max of a list of integers
(define Y
(lambda (w)
((lambda (f)
(f f))
(lambda (f)
(w (lambda (x)
((f f) x)))))))
((Y
(lambda (max)
(lambda (l)
(cond ((null? l) -1)
((> (car l) (max (cdr l))) (car l))
(else (max (cdr l)))))))
'(1 2 3 4 5))
I wish to understand this construction. Can somebody give a clear and simple explanation for this code?
For example, supposing that I forget the formula of Y. How can I remember it , and reproduce it long after I work with it ?
Here's some related answers (by me):
Y combinator discussion in "The Little Schemer"
Unable to get implementation of Y combinator working
In Scheme, how do you use lambda to create a recursive function?
Basically, with Y defined as λr.(λh.h h) (λg.r (λx.(g g) x)), an application Y r reduces as
Y r
(λw.(λh.h h) (λg.w (λx.(g g) x))) r
(λh.h h) (λg.r (λx.(g g) x))
h h
;where
h = (λg.r (λx.(g g) x)) <----\
|
(λg.r (λx.(g g) x)) h |
r (λx.(g g) x) <-------------- | ----------\
;where | |
g = h -----/ |
;so that |
(g g) = (h h) = r (λx.(g g) x) ------/
So r must expect two arguments - first representing the recursive function to be called, and second - an actual argument:
r = λf (λx. ....x.....(f y)...... )
so that (Y r) x reduces as
(r (λx.(g g) x)) x
(r f) x
;where
f = (λx.(g g) x)
f y = (λx.(g g) x) y = (g g) y = (r f) y ; f is "fixed point" of r
The definiton f = (λx.(g g) x) means, when f y is called, (g g) y will be called, at which point g will be self-applied, r "pulled" from inside g and the result of (r f) called with y argument. I.e. any call (f y) in the body of lambda expression resulting from (r f) application, is translated back to (r f) y i.e. invocation of same body with a new argument y.
The important implementational detail is whether it is the same function body, or its copy, but the semantics are the same - we are able to enter the same function body with a new argument value.
The essence of Y combinator is replication through reference and self-application: we refer to the same thing through same name, twice; and thus we arrange for it to receive itself as an argument.
When there's no referencing, as in pure lambda calculus, and parameters receive textual copies of arguments - i.e. reduction is done by textual rewriting - this still works, because same copies get replicated and passed around, being fed as argument to self so it is available on the next iteration, if need be.
But it is much more efficient when shared referencing is available (all uses of same name refer to same thing). Under environment model of evaluation creation of self-referential function is simple as
(let ((fact #f))
(set! fact
(lambda (n) (if (< 2 n) 1
(* n (fact (- n 1))))))
fact)
In fact the definition in your answer is that of applicative-order Y combinator. With normal-order, eta-reduction can be applied without causing infinite looping, to get Ynorm = (λw.(λh.h h) (λg.w (g g))) which is canonically written as
Ynorm = (λf.(λx.f (x x)) (λx.f (x x)))
indeed
Ynorm g
= (λx.g (x x)) (λx.g (x x))
= g ((λx.g (x x)) (λx.g (x x)))

Function and binary operation

I want to create a function : OP from i=n to p = f(i)
OP is a binary operator
Here is my function
(defun sigmaOP (f o n p)
(loop for i from n to p do
(let (val (o (val (funcall f i))))
)
val
)
f is a function
o is the operator
n is the begining and p the end
And to call I use
(sigmaOP (lambda (x) (* 2 x)) '+ 1 3)
But it doesn't work
The o argument isn't consider as operator.
This function work if I remove o and instead there is + or *,...
Thanks
An operator is a function to call, too, so you need to precede o with funcall, or use it as a parameter in a call e.g. to reduce. Are you thinking of something like this?
(defun sigmaOP (f o n p)
(reduce o
(loop for i from n to p
collect (funcall f i))))
Call:
(sigmaOP (lambda (x) (* 2 x)) #'+ 1 3)