lean4-htt/doc/monads/laws.lean

322 lines
11 KiB
Text
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

/-!
# Monad Laws
In the previous sections you learned how to use [Functors](functors.lean.md),
[Applicatives](applicatives.lean.md), and [Monads](monads.lean.md), and you played with some useful
instances including [Option](monads.lean.md), [IO](monads.lean.md), [Reader](readers.lean.md),
[State](states.lean.md) and [Except](except.lean.md) and you learned about composition using [Monad
Transformers](transformers.lean.md).
So far, you've learned the concrete details you need in order to _use_ monads in your Lean programs.
But there's still one more important concept you need if you want to _create_ new functors,
applicatives and monads. Namely, the notion of _structural "laws"_ -- rules that these type
classes should follow in order to meet other programmers' expectations about your code.
## Life without Laws
Remember Lean represents each of these abstract structures by a type class. Each of these type classes
has one or two main functions. So, as long as you implement those functions and it type checks, you
have a new functor, applicative, or monad, right?
Well not quite. Yes, your program will compile and you'll be able to use the instances. But this
doesn't mean your instances follow the mathematical constructs. If they don't, your instances won't
fulfill other programmers' expectations. Each type class has its own "laws". For instance, suppose
you have the following Point Functor:
-/
structure Point (α : Type) where
x : α
y : α
deriving Repr, BEq
def Point.map (f : α → β) (s : Point α) : Point β :=
{ x := f s.y, -- an example of something weird
y := f s.x }
instance : Functor Point where
map := Point.map
#eval (·+2) <$> (Point.mk 1 2) -- { x := 4, y := 3 }
/-!
This Point does something weird, when you `map` it because it transposes the `x` and `y` coordinates
which is not what other people would expect from a `map` function. In fact, it breaks the rules
as you will see below.
## What are the Functor laws?
Functors have two laws: the _identity_ law, and the _composition_ law. These laws express behaviors that
your functor instances should follow. If they don't, other programmers will be very confused at the
effect your instances have on their program.
The identity law says that if you "map" the identity function (`id`) over your functor, the
resulting functor should be the same. A succinct way of showing this on a `List` functor is:
-/
def list1 := [1,2,3]
#eval id <$> list1 == list1 -- true
/-!
Now let's try the same test on the `Point` functor:
-/
def p1 : Point Nat := (Point.mk 1 2)
#eval id <$> p1 == p1 -- false
/-!
Oh, and look while the `List` is behaving well, the `Point` functor fails this identity test.
The _composition_ law says that if you "map" two functions in succession over a functor, this
should be the same as "composing" the functions and simply mapping that one super-function over the
functor. In Lean you can compose two functions using `Function.comp f g` (or the syntax `f ∘ g`,
which you can type in VS code using `\o `) and you will get the same results from both of these
showing that the composition law holds for `List Nat`:
-/
def double (x : Nat) := x + x
def square (x : Nat) := x * x
#eval double <$> (square <$> list1) -- [2, 8, 18]
#eval (double <$> (square <$> list1)) == ((double ∘ square) <$> list1) -- true
-- ok, what about the Point class?
#eval double <$> (square <$> p1) -- { x := 2, y := 8 }
#eval (double ∘ square) <$> p1 -- { x := 8, y := 2 }
#eval double <$> (square <$> p1) == (double ∘ square) <$> p1 -- false
/-!
Note that composition also fails on the bad `Point` because the x/y transpose.
As you can see this bad `Point` implementation violates both of the functor laws. In this case it
would not be a true functor. Its behavior would confuse any other programmers trying to use it. You
should take care to make sure that your instances make sense. Once you get a feel for these type
classes, the likelihood is that the instances you'll create will follow the laws.
You can also write a bad functor that passes one law but not the other like this:
-/
def bad_option_map {α β : Type u} : (α → β) → Option α → Option β
| _, _ => none
instance : Functor Option where
map := bad_option_map
def t1 : Option Nat := some 10
#eval id <$> t1 == t1 -- false
#eval double <$> (square <$> t1) == (double ∘ square) <$> t1 -- true
/-!
This fails the id law but obeys the composition law. Hopefully this explains the value of these
laws, and you don't need to see any more bad examples!
## What are the Applicative Laws?
While functors have two laws, applicatives have four laws:
- Identity
- Homomorphism
- Interchange
- Composition
### Identity
`pure id <*> v = v`
Applying the identity function through an applicative structure should not change the underlying
values or structure. For example:
-/
instance : Applicative List where
pure := List.pure
seq f x := List.bind f fun y => Functor.map y (x ())
#eval pure id <*> [1, 2, 3] -- [1, 2, 3]
/-!
The `pure id` statement here is wrapping the identity function in an applicative structure
so that you can apply that over the container `[1, 2, 3]` using the Applicative `seq` operation
which has the notation `<*>`.
To prove this for all values `v` and any applicative `m` you can write this theorem:
-/
example [Applicative m] [LawfulApplicative m] (v : m α) :
pure id <*> v = v :=
by simp -- Goals accomplished 🎉
/-!
### Homomorphism
`pure f <*> pure x = pure (f x)`
Suppose you wrap a function and an object in `pure`. You can then apply the wrapped function over the
wrapped object. Of course, you could also apply the normal function over the normal object, and then
wrap it in `pure`. The homomorphism law states these results should be the same.
For example:
-/
def x := 1
def f := (· + 2)
#eval pure f <*> pure x = (pure (f x) : List Nat) -- true
/-!
You should see a distinct pattern here. The overriding theme of almost all these laws is that these
`Applicative` types should behave like normal containers. The `Applicative` functions should not
have any side effects. All they should do is facilitate the wrapping, unwrapping, and transformation
of data contained in the container resulting in a new container that has the same structure.
### Interchange
`u <*> pure y = pure (. y) <*> u`.
This law is is a little more complicated, so don't sweat it too much. It states that the order that
you wrap things shouldn't matter. One the left, you apply any applicative `u` over a pure wrapped
object. On the right, you first wrap a function applying the object as an argument. Note that `(·
y)` is short hand for: `fun f => f y`. Then you apply this to the first applicative `u`. These
should be the same.
For example:
-/
def y := 4
def g : List (Nat → Nat) := [(· + 2)]
#eval g <*> pure y = pure (· y) <*> g -- true
/-!
You can prove this with the following theorem:
-/
example [Applicative m] [LawfulApplicative m] (u : m (α → β)) (y : α) :
u <*> pure y = pure (· y) <*> u :=
by simp [pure_seq] -- Goals accomplished 🎉
/-!
### Composition:
`u <*> v <*> w = u <*> (v <*> w)`
This final applicative law mimics the second functor law. It is a composition law. It states that
function composition holds across applications within the applicative:
For example:
-/
def u := [1, 2]
def v := [3, 4]
def w := [5, 6]
#eval pure (·+·+·) <*> u <*> v <*> w
-- [9, 10, 10, 11, 10, 11, 11, 12]
#eval let grouping := pure (·+·) <*> v <*> w
pure (·+·) <*> u <*> grouping
-- [9, 10, 10, 11, 10, 11, 11, 12]
/-!
To test composition you see the separate grouping `(v <*> w)` then that can be used in the outer
sequence `u <*> grouping` to get the same final result `[9, 10, 10, 11, 10, 11, 11, 12]`.
## What are the Monad Laws?
Monads have three laws:
- Left Identity
- Right Identity
- Associativity
### Left Identity
Identity laws for monads specify that `pure` by itself shouldn't really change anything about the
structure or its values.
Left identity is `x >>= pure = x` and is demonstrated by the following examples on a monadic `List`:
-/
instance : Monad List where
pure := List.pure
bind := List.bind
def a := ["apple", "orange"]
#eval a >>= pure -- ["apple", "orange"]
#eval a >>= pure = a -- true
/-!
### Right Identity
Right identity is `pure x >>= f = f x` and is demonstrated by the following example:
-/
def h (x : Nat) : Option Nat := some (x + 1)
def z := 5
#eval pure z >>= h -- some 6
#eval h z -- some 6
#eval pure z >>= h = h z -- true
/-!
So in this example, with this specific `z` and `h`, you see that the rule holds true.
### Associativity
The associativity law is written as:
```lean,ignore
x >>= f >>= g = x >>= (λ x => f x >>= g)
```
where `(x : m α)` and `(f : α → m β)` and `(g : β → m γ)`.
The associativity law is difficult to parse like some of the applicative laws, but what it is saying
is that if you change the grouping of `bind` operations, you should still get the same result.
This law has a parallel structure to the other composition laws.
You can see this in action in the following rewrite of `runOptionFuncsBind` from [monads](monads.lean.md):
-/
def optionFunc1 : String -> Option Nat
| "" => none
| str => some str.length
def optionFunc2 (i : Nat) : Option Float :=
if i % 2 == 0 then none else some (i.toFloat * 3.14159)
def optionFunc3 (f : Float) : Option (List Nat) :=
if f > 15.0 then none else some [f.floor.toUInt32.toNat, f.ceil.toUInt32.toNat]
def runOptionFuncsBind (input : String) : Option (List Nat) :=
optionFunc1 input >>= optionFunc2 >>= optionFunc3
def runOptionFuncsBindGrouped (input : String) : Option (List Nat) :=
optionFunc1 input >>= (λ x => optionFunc2 x >>= optionFunc3)
#eval runOptionFuncsBind "big" -- some [9, 10]
#eval runOptionFuncsBindGrouped "big" -- some [9, 10]
/-!
Notice here we had to insert a `λ` function just like the definition says: `(λ x => f x >>= g)`.
This is because unlike applicatives, you can't resolve the structure of later operations without the
results of earlier operations quite as well because of the extra context monads provide. But you can
still group their later operations into composite functions taking their inputs from earlier on, and
the result should be the same.
## Summary
While these laws may be a bit difficult to understand just by looking at them, the good news is that
most of the instances you'll make will naturally follow the laws so long as you keep it simple, so
you shouldn't have to worry about them too much.
There are two main ideas from all the laws:
1. Applying the identity or pure function should not change the underlying values or structure.
1. It should not matter what order you group operations in. Another way to state this is function
composition should hold across your structures.
Following these laws will ensure other programmers are not confused by the behavior of your
new functors, applicatives and monads.
-/