Tail Recursivity in F# : Inversions with Quicksort - recursion

Question

Hi i have some difficulty in understanding tail-recursivity. I know thats it's important to avoid infinite loops and also for memory usage. I've seen some examples on simple functions like Fibonacci in "Expert in F#", but I don't think i've seen code when the result is something different than just a number.
What would be the accumulator then ? i'm not sure...
Here is a recursive function that I've written. It counts the number of inversions in an array, using the quicksort algorithm. [it's taken from an exercise of the Coursera MOOC Algo I by Stanford]
I'd be grateful if somebody could explain how to make that tail recursive.
[Also, i've translated that code from imperative code, as i had written that in R before, so the style is not functional at all...]
another question: is the syntax correct, A being a (mutable) array, i've written let A = .... everywhere ?
is A <- .... better / the same ?
open System.IO
open System
let X = [|57; 97; 17; 31; 54; 98; 87; 27; 89; 81; 18; 70; 3; 34; 63; 100; 46; 30; 99;
10; 33; 65; 96; 38; 48; 80; 95; 6; 16; 19; 56; 61; 1; 47; 12; 73; 49; 41;
37; 40; 59; 67; 93; 26; 75; 44; 58; 66; 8; 55; 94; 74; 83; 7; 15; 86; 42;
50; 5; 22; 90; 13; 69; 53; 43; 24; 92; 51; 23; 39; 78; 85; 4; 25; 52; 36;
60; 68; 9; 64; 79; 14; 45; 2; 77; 84; 11; 71; 35; 72; 28; 76; 82; 88; 32;
21; 20; 91; 62; 29|]
// not tail recursive. answer = 488
let N = X.Length
let mutable count = 0
let swap (A:int[]) a b =
let tmp = A.[a]
A.[a] <- A.[b]
A.[b] <- tmp
A
let rec quicksortNT (A:int[]) =
let L = A.Length
match L with
| 1 -> A
| 2 -> count <- count + 1
if (A.[0]<A.[1]) then A
else [|A.[1];A.[0]|]
| x -> let p = x
let pval = A.[p-1]
let A = swap A 0 (p-1)
let mutable i = 1
for j in 1 .. (x-1) do
if (A.[j]<pval) then let A = swap A i j
i <- i+1
// end of for loop
// putting back pivot at its right place
let A = swap A 0 (i-1)
let l1 = i-1
let l2 = x-i
if (l1=0) then
let A = Array.append [|A.[0]|] (quicksortNT A.[1..p-1])
count <- count + (l2-1)
A
elif (l2=0) then
let A = Array.append (quicksortNT A.[0..p-2]) [|A.[p-1]|]
count <- count + (l2-1)
A
else
let A = Array.append ( Array.append (quicksortNT A.[0..(i-2)]) [|A.[i-1]|] ) (quicksortNT A.[i..p-1])
count <- count + (l1-1)+(l2-1)
A
let Y = quicksortNT X
for i in 1..N do printfn "%d" Y.[i-1]
printfn "count = %d" count
Console.ReadKey() |> ignore
Thank you very much for your help

Answer 1

As I said in my comment: you do inplace-swapping so it makes no sense to recreate and return arrays.
But as you ask about tail-recursive solutions look at this version using lists and continuation-passing-style to make the algorithm tail-recursive:
let quicksort values =
let rec qsort xs cont =
match xs with
| [] -> cont xs
| (x::xs) ->
let lower = List.filter (fun y -> y <= x) xs
let upper = List.filter (fun y -> y > x) xs
qsort lower (fun lowerSorted ->
qsort upper (fun upperSorted -> cont (lowerSorted # x :: upperSorted)))
qsort values id
remarks:
you can think of it like this:
first partition the input into upper and lower parts
then start with sorting (recursively) the lower part, when you are done with this continue by...
... take lowerSorted and sort the upper part as well and continue with ...
... take both sorted parts, join them and pass them to the outer continuation
the outermost continuation should of course just be the id function
some will argue that this is not quicksort as it does not sort inplace!
maybe it's hard to see but it's tail-recursive as the very last call is to qsort and it's result will be the result of the current call
I used List because the pattern-matching is so much nicer - but you can adopt this to your version with arrays as well
in those cases (as here) where you have multiple recursive calls I always find cont-passing solutions to be easier to write and more natural - but accumulators could be used as well (but it will get messy as you need to pass where you are too)
this will not take less memory than the version without the cont-passing at all - it just will be placed on the heap instead of the stack (you usually have way more heap available ;) ) - so it's a bit like cheating
that's why the imperative algorithm is still way better performance-wise - so a usual compromise is to (for example) copy the array, use the inplace-algorithm on the copy and then return the copy - this way the algorithm behaves as if it's pure on the outside

Answer 2

The whole point to quicksort's swapping partition procedure is that it can mutate the same array; you just pass it the low and the high index of the array's range it has to process.
So make a nested function and pass it just the 2 indices. To make it tail recursive, add the third parameter, list-of-ranges-to-process; when that becomes empty, you're done. Wikibook says you mutate arrays with A.[i] <- A.[j].
A nested function can access its parent function's argument directly, because it is in scope. So, make swap nested too:
let rec quicksort (A:int[]) =
let swap a b =
let tmp = A.[a]
A.[a] <- A.[b]
A.[b] <- tmp
let todo = ... (* empty list *)
let rec partition low high =
.... (* run the swapping loop,
find the two new pairs of indices,
put one into TODO and call *)
partition new_low new_high
let L = A.Length
match L with
| 1 -> (* do nothing A *)
| 2 -> count <- count + 1
if (A.[0]<A.[1]) then (* do nothing A *)
else (* [|A.[1];A.[0]|] *) swap 1 0
| x -> ....
partition 0 L
So partition will be tail recursive, working inside the environment set up for it by quicksort.
(disclaimer: I don't know F# and have never used it, but I know Haskell and Scheme, to some degree).