Part of the "Recursive types and folds" series (link)

Understanding Folds

Recursion vs. iteration

This post is the fourth in a series.

In the previous post, I introduced “folds”, a way of creating top-down iterative functions for recursive types.

In this post, we’ll spend some time understanding folds in more detail.

Series contents

Here’s the contents of this series:


Iteration vs. recursion

We now have three different functions – cata, fold and foldback.

So what exactly are the differences between them? We’ve seen that something that doesn’t work with fold will work with foldBack, but is there an easy way to remember the difference?

One way to differentiate the three approaches is by remembering this:

  • fold is top-down iteration
  • cata is bottom-up recursion
  • foldBack is bottom-up iteration

What does this mean?

Well, for in fold, the accumulator was initialized at the top level, and was passed down to each lower level until the lowest and last level was reached.

In code terms, each level did this:

accumulatorFromHigherLevel, combined with
  stuffFromThisLevel
    => stuffToSendDownToNextLowerLevel

In an imperative language, this is exactly a “for loop” with a mutable variable storing the accumulator.

var accumulator = initialValue
foreach level in levels do
{
  accumulator, combined with
    stuffFromThisLevel
      => update accumulator
}

So, this kind of top-to-bottom folding can be thought of as iteration (and in fact, the F# compiler will turn a tail-recursive function like this into an iteration behind the scenes).

On the other hand, in cata, the accumulator started at the bottom level, and was passed up to each higher level until the top level was reached.

In code terms, each level did this:

accumulatorFromLowerLevel, combined with
  stuffFromThisLevel
    => stuffToSendUpToNextHigherLevel

This is exactly a recursive loop:

let recurse (head::tail) =
    if atBottomLevel then
       return something
    else    // if not at bottom level
       let accumulatorFromLowerLevel = recurse tail
       return stuffFromThisLevel, combined with
          accumulatorFromLowerLevel

Finally, foldback can be thought of as “reverse iteration”. The accumulator is threaded through all the levels, but starting at the bottom rather than at the top. It has the benefits of cata in that the inner values are calculated first and passed back up, but because it is iterative, there cannot be a stack overflow.

Many of the concepts we have discussed so far become clear when expressed in terms of iteration vs. recursion. For example:

  • The iterative versions (fold and foldback) have no stack, and cannot cause a stack overflow.
  • The “total cost” function needed no inner data, and so the top-down iterative version (fold) worked without problems.
  • The “description” function though, needed inner text for correct formatting, and so the recursive version (cata) or bottom up iteration (foldback) was more suitable.

Fold example: File system domain

In the last post, we described some rules for creating folds. Let’s see if we can apply these rules to create a fold in another domain, the “File System” domain from the second post in the series.

As a reminder, here is the crude “file system” domain from that post:

type FileSystemItem =
    | File of File
    | Directory of Directory
and File = {name:string; fileSize:int}
and Directory = {name:string; dirSize:int; subitems:FileSystemItem list}

Note that each directory contains a list of subitems, so this is not a linear structure like Gift, but a tree-like structure. Our implementation of fold will have to take this into account.

Here are some sample values:

let readme = File {name="readme.txt"; fileSize=1}
let config = File {name="config.xml"; fileSize=2}
let build  = File {name="build.bat"; fileSize=3}
let src = Directory {name="src"; dirSize=10; subitems=[readme; config; build]}
let bin = Directory {name="bin"; dirSize=10; subitems=[]}
let root = Directory {name="root"; dirSize=5; subitems=[src; bin]}

We want to create a fold, foldFS, say. So, following the rules, let’s add an extra accumulator parameter acc and pass it to the File case:

let rec foldFS fFile fDir acc item :'r =
    let recurse = foldFS fFile fDir
    match item with
    | File file ->
        fFile acc file
    | Directory dir ->
        // to do

The Directory case is trickier. We are not supposed to know about the subitems, so that means that the only data we can use is the name, dirSize, and the accumulator passed in from a higher level. These are combined to make a new accumulator.

| Directory dir ->
    let newAcc = fDir acc (dir.name,dir.dirSize)
    // to do

NOTE: I’m keeping the name and dirSize as a tuple for grouping purposes, but of course you could pass them in as separate parameters.

Now we need to pass this new accumulator down to each subitem in turn, but each subitem will return a new accumulator of its own, so we need to use the following approach:

  • Take the newly created accumulator and pass it to the first subitem.
  • Take the output of that (another accumulator) and pass it to the second subitem.
  • Take the output of that (another accumulator) and pass it to the third subitem.
  • And so on. The output of the last subitem is the final result.

That approach is already available to us though. It’s exactly what List.fold does! So here’s the code for the Directory case:

| Directory dir ->
    let newAcc = fDir acc (dir.name,dir.dirSize)
    dir.subitems |> List.fold recurse newAcc

And here’s the entire foldFS function:

let rec foldFS fFile fDir acc item :'r =
    let recurse = foldFS fFile fDir
    match item with
    | File file ->
        fFile acc file
    | Directory dir ->
        let newAcc = fDir acc (dir.name,dir.dirSize)
        dir.subitems |> List.fold recurse newAcc

With this in place, we can rewrite the same two functions we implemented in the last post.

First, the totalSize function, which just sums up all the sizes:

let totalSize fileSystemItem =
    let fFile acc (file:File) =
        acc + file.fileSize
    let fDir acc (name,size) =
        acc + size
    foldFS fFile fDir 0 fileSystemItem

And if we test it we get the same results as before:

readme |> totalSize  // 1
src |> totalSize     // 16 = 10 + (1 + 2 + 3)
root |> totalSize    // 31 = 5 + 16 + 10

File system domain: largestFile example

We can also reimplement the “what is the largest file in the tree?” function.

As before it will return a File option, because the tree might be empty. This means that the accumulator will be a File option too.

This time it is the File case handler which is tricky:

  • If the accumulator being passed in is None, then this current file becomes the new accumulator.
  • If the accumulator being passed in is Some file, then compare the size of that file with this file. Whichever is bigger becomes the new accumulator.

Here’s the code:

let fFile (largestSoFarOpt:File option) (file:File) =
    match largestSoFarOpt with
    | None ->
        Some file
    | Some largestSoFar ->
        if largestSoFar.fileSize > file.fileSize then
            Some largestSoFar
        else
            Some file

On the other hand, the Directory handler is trivial – just pass the “largest so far” accumulator down to the next level

let fDir largestSoFarOpt (name,size) =
    largestSoFarOpt

Here’s the complete implementation:

let largestFile fileSystemItem =
    let fFile (largestSoFarOpt:File option) (file:File) =
        match largestSoFarOpt with
        | None ->
            Some file
        | Some largestSoFar ->
            if largestSoFar.fileSize > file.fileSize then
                Some largestSoFar
            else
                Some file

    let fDir largestSoFarOpt (name,size) =
        largestSoFarOpt

    // call the fold
    foldFS fFile fDir None fileSystemItem

And if we test it we get the same results as before:

readme |> largestFile
// Some {name = "readme.txt"; fileSize = 1}

src |> largestFile
// Some {name = "build.bat"; fileSize = 3}

bin |> largestFile
// None

root |> largestFile
// Some {name = "build.bat"; fileSize = 3}

It is interesting to compare this implementation with the recursive version in the second post. I think that this one is easier to implement, myself.

Tree traversal types

The various fold functions discussed so far correspond to various kinds of tree traversals:

  • A fold function (as implemented here) is more properly called a “pre-order depth-first” tree traversal.
  • A foldback function would be a “post-order depth-first” tree traversal.
  • A cata function would not be a “traversal” at all, because each internal node deals with a list of all the subresults at once.

By tweaking the logic, you can make other variants.

For a description of the various kinds of tree traversals, see Wikipedia.

Do we need foldback?

Do we need to implement a foldback function for the FileSystem domain?

I don’t think so. If we need access to the inner data, we can just use the original “naive” catamorphism implementation in the previous post.

But, hey wait, didn’t I say at the beginning that we had to watch out for stack overflows?

Yes, if the recursive type is deeply nested. But consider a file system with only two subdirectories per directory. How many directories would there be if there were 64 nested levels? (Hint: you may be familiar with a similar problem. Something to do with grains on a chessboard).

We saw earlier that the stack overflow issue only occurs with more than 1000 nested levels, and that level of nesting generally only occurs with linear recursive types, not trees like the FileSystem domain.

Common questions about “fold”

At this point you might be getting overwhelmed! All these different implementations with different advantages and disadvantages.

So let’s take a short break and address some common questions.

What’s the difference between “left fold” and “right fold”

There is often quite a lot of confusion around the terminology of folds: “left” vs. “right”, “forward” vs. “backwards”, etc.

  • A left fold or forward fold is what I have just called fold – the top-down iterative approach.
  • A right fold or backward fold is what I have called foldBack – the bottom-up iterative approach.

These terms, though, really only apply to linear recursive structures like Gift. When it comes to more complex tree-like structures, these distinctions are too simple, because there are many ways to traverse them: breadth-first, depth-first, pre-order and post-order, and so on.

Which type of fold function should I use?

Here are some guidelines:

  • If your recursive type is not going to be too deeply nested (less than 100 levels deep, say), then the naive cata catamorphism we described in the first post is fine. It’s really easy to implement mechanically – just replace the main recursive type with 'r.
  • If your recursive type is going to be deeply nested and you want to prevent stack overflows, use the iterative fold.
  • If you are using an iterative fold but you need to have access to the inner data, pass a continuation function as an accumulator.
  • Finally, the iterative approach is generally faster and uses less memory than the recursive approach (but that advantage is lost if you pass around too many nested continuations).

Another way to think about it is to look at your “combiner” function. At each step, you are combining data from the different levels:

level1 data [combined with] level2 data [combined with] level3 data [combined with] level4 data

If your combiner function is “left associative” like this:

(((level1 combineWith level2) combineWith level3) combineWith level4)

then use the iterative approach, but if your combiner function is “right associative” like this:

(level1 combineWith (level2 combineWith (level3 combineWith level4)))

then use the cata or foldback approach.

And if your combiner function doesn’t care (like addition, for example), use whichever one is more convenient.

How can I know whether code is tail-recursive or not?

It’s not always obvious whether an implementation is tail-recursive or not. The easiest way to be sure is to look at the very last expression for each case.

If the call to “recurse” is the very last expression, then it is tail-recursive. If there is any other work after that, then it is not tail-recursive.

See for yourself with the three implementations that we have discussed.

First, here’s the code for the WithACard case in the original cataGift implementation:

| WithACard (gift,message) ->
    fCard (recurse gift,message)
//         ~~~~~~~  <= Call to recurse is not last expression.
//                     Tail-recursive? No!

The cataGift implementation is not tail-recursive.

Here’s the code from the foldGift implementation:

| WithACard (innerGift,message) ->
    let newAcc = fCard acc message
    recurse newAcc innerGift
//  ~~~~~~~  <= Call to recurse is last expression.
//              Tail-recursive? Yes!

The foldGift implementation is tail-recursive.

And here’s the code from the foldbackGift implementation:

| WithACard (innerGift,message) ->
    let newGenerator innerVal =
        let newInnerVal = fCard innerVal message
        generator newInnerVal
    recurse newGenerator innerGift
//  ~~~~~~~  <= Call to recurse is last expression.
//              Tail-recursive? Yes!

The foldbackGift implementation is also tail-recursive.

How do I short-circuit a fold?

In a language like C#, you can exit a iterative loop early using break like this:

foreach (var elem in collection)
{
    // do something

    if ( x == "error")
    {
        break;
    }
}

So how do you do the same thing with a fold?

The short answer is, you can’t! A fold is designed to visit all elements in turn. The Visitor Pattern has the same constraint.

There are three workarounds.

The first one is to not use fold at all and create your own recursive function that terminates on the required condition:

In this example, the loop exits when the sum is larger than 100:

let rec firstSumBiggerThan100 sumSoFar listOfInts =
    match listOfInts with
    | [] ->
        sumSoFar // exhausted all the ints!
    | head::tail ->
        let newSumSoFar = head + sumSoFar
        if newSumSoFar > 100 then
            newSumSoFar
        else
            firstSumBiggerThan100 newSumSoFar tail

// test
[30;40;50;60] |> firstSumBiggerThan100 0  // 120
[1..3..100] |> firstSumBiggerThan100 0  // 117

The second approach is to use fold but to add some kind of “ignore” flag to the accumulator that is passed around. Once this flag is set, the remaining iterations do nothing.

Here’s an example of calculating the sum, but the accumulator is actually a tuple with an ignoreFlag in addition to the sumSoFar:

let firstSumBiggerThan100 listOfInts =

    let folder accumulator i =
        let (ignoreFlag,sumSoFar) = accumulator
        if not ignoreFlag then
            let newSumSoFar = i + sumSoFar
            let newIgnoreFlag  = newSumSoFar > 100
            (newIgnoreFlag, newSumSoFar)
        else
            // pass the accumulator along
            accumulator

    let initialAcc = (false,0)

    listOfInts
    |> List.fold folder initialAcc  // use fold
    |> snd // get the sumSoFar

/// test
[30;40;50;60] |> firstSumBiggerThan100  // 120
[1..3..100] |> firstSumBiggerThan100  // 117

The third version is a variant of the second – create a special value to signal that the remaining data should be ignored, but wrap it in a computation expression so that it looks more natural.

This approach is documented on Tomas Petricek’s blog and the code looks like this:

let firstSumBiggerThan100 listOfInts =
    let mutable sumSoFar = 0
    imperative {
        for x in listOfInts do
            sumSoFar <- x + sumSoFar
            if sumSoFar > 100 then do! break
    }
    sumSoFar

Summary

The goal of this post was to help you understand folds better, and to show how they could be applied to a tree structure like the file system. I hope it was helpful!

Up to this point in the series all the examples have been very concrete; we have implemented custom folds for each domain we have encountered. Can we be a bit more generic and build some reusable fold implementations?

In the next post we’ll look at generic recursive types, and how to work with them.

The source code for this post is available at this gist.

Comments

blog comments powered by Disqus