What are the differences between 'let' or 'letrec' and 'define' for creating local bindings?

Question

I don't understand what the differences are between (sorry for the contrived example):

(define average
  (lambda (elems)

   (define length
     (lambda (xs)
      (if (null? xs)
          0
          (+ 1 (length (cdr xs))))))

   (define sum
     (lambda (xs)
      (if (null? xs)
          0
          (+ (car xs) (sum (cdr xs))))))

   (define total (sum elems))

   (define count (length elems))

   (/ total count)))

and

(define average
  (lambda (elems)
   (letrec ((length
              (lambda (xs)
               (if (null? xs)
                   0
                   (+ 1 (length (cdr xs))))))

            (sum
              (lambda (xs)
               (if (null? xs)
                   0
                   (+ (car xs) (sum (cdr xs))))))

            (total (sum elems))

            (count (length elems)))

     (/ total count))))

As far as I can tell, they both create a new scope, and in that scope create 4 local variables that refer to each other and to themselves, and evaluate and return a body.

Am I missing something here, or is letrec synonymous with scoped defines?

I know this may be implementation dependent; I'm trying to get an understanding of the fundamentals of Lisps.

Answer 1

You are correct that there are parallels between the define and letrec versions of your code. However, the devil is in the details. In R5RS, internal define has letrec semantics. In R6RS, internal define has letrec* semantics.

What's the difference? Your code has actually just highlighted this difference. As Zhehao's answer mentions, your definition of total and count inside the same letrec as the length and sum is incorrect: length and sum are not guaranteed to be bound by the time you're evaluating the values of (length elems) and (sum elems), since the binding of those variables is not guaranteed to be left-to-right.

letrec* is similar to letrec, but with a left-to-right guarantee. So if you changed your letrec to letrec*, it'd be okay.

Now, back to my initial comment: because R5RS's internal define uses letrec semantics, even your define version of the code would be incorrect under an R5RS implementation, but it would be okay under an R6RS implementation, which has letrec* semantics.