This walks through most of A State Monad Tutorial, which is addressed to a Haskell-using audience. But we convert it to OCaml. See our page on comparing OCaml and Haskell.

Some of what we do here will make use of our monad library for OCaml. See:

As we discussed in ?week9 TODO LINK, a State monad is implemented with the type:

store -> ('a * store)

It's common practice to encapsulate this in some way, so that the interpreter knows the difference between arbitrary functions from a blah to a pair of something and a blah and the values that you've specially designated as being State monadic values.

The most lightweight way encapsulate it would be just to add a data constructor to the type. In the same way that the 'a option type has the None and Some data constructors, we give our 'a state type a State data constructor:

(* we assume that the store type has already been declared *)
type 'a state = State of (store -> ('a * store))

Then a function expecting an 'a state will look for a value with the structure State ... rather than just one with the structure ....

To take a State (s -> (a,s)) and get at the s -> (a,s) it wraps, you use the same techniques you use to take an Some int and get at the int it wraps:

let xx = State (fun s -> (1, s)) in
let State unwrapped_xx = xx in


let xx = State (fun s -> (1, s)) in
match xx with
  | State unwrapped_xx -> ...

There are two heavierweight ways to encapsulate the type of the State monad. One is used by our ?monad library TODO LINK --- the type is hidden from the outside user, and only gets exposed by the run function. But the result of run xx is not itself recognized as a monadic value any longer. You can't replace xx in:

SomeMonad.(xx >>= ...)

with run xx. So you should only apply run when you've finished building up a monadic value. (That's why it's called run.)

Of course you can do this:

let xx = SomeMonad.(...) in
let intermediate_result = xx 0 in
let yy = SomeMonad.(xx >>= ...) in
let final_result = yy 0 in

The other heavyweight way to encapsulate the type of a monad is to use records. See here and here for some introduction to these. TODO LINKS We don't use this design in our OCaml monad library, but the Haskell monad libraries do, and it would be good for you to get acquainted with it so that you can see how to ignore it when you come across it in Haskell-based literature. (Or you might want to learn Haskell, who knows?)

We'll briefly illustrate this technique in OCaml code, for uniformity. See the translation page TODO LINKS about how this looks in Haskell.

To use the record technique, instead of saying;

type 'a state = State of (store -> ('a * store))

you'd say:

type 'a state = { state : store -> ('a * store) }

and instead of saying:

let xx = State (fun s -> (1, s)) in
let State unwrapped_xx = xx in

you'd say:

let xx = { state = fun s -> (1, s) } in
let unwrapped_xx = xx.state in

That's basically it. As with the other two techniques, the type of xx is not the same as store -> ('a * store), but the relevant code will know how to convert between them.

The main benefit of these techniques is that it gives you better type-checking: it makes sure that you're only using your monadic values in the hygenic ways you're supposed to. Perhaps you don't care about that. Well, then, if you want to write all your own monadic code, you can proceed as you like. If you ever want to use other people's code, though, or read papers or web posts about monads, you will encounter one or more of these techniques, and so you need to get comfortable enough with them not to let them confuse you.

OK, back to our walk-through of "A State Monad Tutorial". What shall we use for a store? Instead of a plain int, let's suppose our store is a structure of two values: a running total, and a count of how many times the store has been modified. We'll implement this with a record. Hence:

type store' = { total : int; modifications: int }

State monads employing this store will then have three salient values at any point in the computation: the total and modifications field in the store, and also the 'a payload that is currently wrapped in the monadic box.

Here's a monadic value that encodes the operation of incrementing the store's total and wrapping the value that was the former total:

let increment_store : store' -> (int * store') = fun s ->
  let value = in
  let s' = { total = succ; modifications = succ s.modifications } in
  (value, s')

If we wanted to work with one of the encapsulation techniques described above, we'd have to proceed a bit differently. Here is how to do it with the first, lightweight technique:

let increment_store' : 'a state = State (fun s ->
  let value = in
  let s' = { total = succ; modifications = succ s.modifications } in
  (value, s'))

Here is how you'd have to do it using our OCaml/Juli8 monad library:

(* have Juli8 loaded, as explained elsewhere *)
# module S = Monad.State(struct type store = store' end);;
# let increment_store'' : 'a S.t =
    S.(get >>= fun cur ->
         let value = in
         let s' = { total = succ; modifications = succ cur.modifications } in
         put s' >> mid value);;

Let's try it out:

# let s0 = { total = 42; modifications = 3 };;
# increment_store s0;;
- : int * store' = (42, {total = 43; modifications = 4})

Or if you used the OCaml/Juli8 monad library:

# increment_store'' s0;;
- : int * = (42, {total = 43; modifications = 4})


Can you write a monadic value that instead of incrementing each of the total and modifications fields in the store, doubles the total field and increments the modifications field?

What about a value that increments each of total and modifications twice? Well, you could custom-write that, as with the previous question. But we already have the tools to express it easily, using our existing increment_store value:

increment_store >>= fun value -> increment_store >> mid value

That ensures that the value we get at the end is the value returned by the first application of increment_store, that is, the contents of the total field in the store before we started modifying the store at all.

You should start to see here how chaining monadic values together gives us a kind of programming language. Of course, it's a cumbersome programming language. It'd be much easier to write, directly in OCaml:

let value =
in (value, { total = + 2; modifications = s0.modifications + 2})

or, using pattern-matching on the record (you don't have to specify every field in the record):

let { total = value; _ } = s0
in (value, { total = + 2; modifications = s0.modifications + 2})

But the point of learning how to do this monadically is that (1) monads show us how to embed more sophisticated programming techniques, such as imperative state and continuations, into frameworks that don't natively possess them (such as the set-theoretic metalanguage of Groenendijk, Stokhof and Veltman's paper); (2) becoming familiar with monads will enable you to see patterns you'd otherwise miss, and implement some seemingly complex computations using the same simple patterns (same-fringe is an example); and finally, of course (3) monads are delicious.

Keep in mind that the final result of a mbind chain doesn't have to be the same type as the starting value:

increment_store >>= fun value -> increment_store >> mid (string_of_int value)


mid 1 >> mid "blah"

The store keeps the same type throughout the computation, but the type of the wrapped value can change.

What are the special-purpose operations that the Monad.State module defines for us?

  • get is a monadic value that passes through the existing store unchanged, and also wraps that same store as its boxed payload. You use it like this:

    ... >>= fun _ -> get >>= fun cur -> ... here we can use cur ...

    As we've said, that's equivalent to:

    ... >> get >>= fun cur -> ...

    You can also get the current store at the start of the computation:

    get >> = fun cur -> ...
  • gets selector is like get, but it additionally applies the selector function to the store before depositing it in the box. If your store is structured, you can use this to only extract a piece of the structure:

    ... >> gets (fun cur -> >>= fun total -> ...

    For more complex structured stores, consider using the Monad.Ref variant of the State monad in the OCaml library.

  • put new_store replaces the existing store with new_store. Use it like this:

    ... >> put new_store >> fun () -> ...

    As that code snippet suggests, the boxed payload after the application of modify new_store is just (). If you want to preserve the existing payload but replace the store, do this:

    ... >>= fun value -> put new_store >> mid value >>= ...
  • Finally, modify modifier applies modifier to whatever the existing store is, and substitutes that as the new store. As with put, the boxed payload afterwards is ().

Here's an example from "A State Monad Tutorial":

increment_store >> get >>= fun cur ->
    State (fun s -> ((), { total = / 2; modifications = succ s.modifications })) >>
    increment_store >> mid

Or, as you'd have to write it using our OCaml monad library:

increment_store'' >> get >>= fun cur ->
    put { total = / 2; modifications = succ cur.modifications } >>
    increment_store'' >> mid

The last topic covered in "A State Monad Tutorial" is the use of do-notation to work with monads in Haskell. We discuss that on our translation page.