--- layout: post title: Isomorphisms for Proof Reductions date: 2016-12-25 comments: true external-url: author: Vikraman Choudhury and Niki Vazou published: true categories: reflection demo: Iso.hs --- [Previously][refinement-reflection] we saw how Refinement Reflection can be used to write and prove **in Haskell** theorems **about Haskell** functions and have such proofs machine checked by Liquid Haskell. As a limitation, Liquid Haskell offers no proof generation techniques: The user needs to manually provide all the proofs. Today we will see how proof generation can be simplified by data type isomorphisms. As an example, a user defined `Peano` data type enjoyes all the arithmetic properties of natural numbers since `Peano` and natural numbers are provably **isomorphic**.


Recursion


Recursive Paper and Pencil Proofs. "Drawing Hands" by Escher.


First, we define `Peano` numbers as a data type and the function `leqPeano` that compares two peano numbers. \begin{code} {-@ data Peano [toNat] = Z | S Peano @-} data Peano = Z | S Peano deriving (Eq) {-@ axiomatize leqPeano @-} leqPeano :: Peano -> Peano -> Bool leqPeano Z _ = True leqPeano _ Z = False leqPeano (S n) (S m) = leqPeano n m \end{code} We can use Refinement Reflection to provide an explicit proof that comparison on peano numbers is *total*, that is, for every two numbers `n` and `m` either `leqPeano n m` or `leqPeano m n` always holds. \begin{code} {-@ leqNTotal :: n:Peano -> m:Peano -> {(leqPeano n m) || (leqPeano m n)} / [toNat n + toNat m] @-} leqNTotal :: Peano -> Peano -> Proof leqNTotal Z m = leqPeano Z m *** QED leqNTotal n Z = leqPeano Z n *** QED leqNTotal (S n) (S m) = (leqPeano (S n) (S m) || leqPeano (S m) (S n)) ==. (leqPeano n m || leqPeano (S m) (S n)) ? (leqNTotal n m) ==. (leqPeano n m || leqPeano m n) ? (leqNTotal m n) *** QED \end{code} The proof proceeds by induction on the sum of `n` and `m`. Liquid Haskell captures this generalized induction by ensuring that the value `toNat n + toNat m` is decreasing where `toNat` maps Peano to Natural numbers. \begin{code} {-@ measure toNat @-} {-@ toNat :: Peano -> Nat @-} toNat :: Peano -> Int toNat Z = 0 toNat (S n) = 1 + toNat n \end{code} Note, that the type `Nat` is just a refinement for the Haskell's integers \begin{spec} type Nat = {v:Int | 0 <= v} \end{spec} First, we'll express verified typeclasses using Haskell records, but with LiquidHaskell refinements to express laws. As an example, let's define `VerifiedOrd`, which is the same as the `Ord` typeclass, but with total order laws added. \begin{code} {-@ data VerifiedOrd a = VerifiedOrd { leq :: a -> a -> Bool , refl :: x:a -> { leq x x } , antisym :: x:a -> y:a -> { leq x y && leq y x ==> x == y } , trans :: x:a -> y:a -> z:a -> { leq x y && leq y z ==> leq x z } , total :: x:a -> y:a -> { leq x y || leq y x } } @-} data VerifiedOrd a = VerifiedOrd { leq :: a -> a -> Bool , refl :: a -> Proof , antisym :: a -> a -> Proof , trans :: a -> a -> a -> Proof , total :: a -> a -> Proof } \end{code} The `leq` function represents the `<=` operator in `Ord`, and `refl`, `antisym`, `trans`, `total`, express the reflexivity, antisymmetry, transitivity and totality properties respectively, that a total order requires. Notice how the refinements express the laws, and the actual code is simply a function that returns `Proof` or `()`. Now, let's see how to define some simple verified instances. We need to instantiate `leq` for our type, "reflect" it into the logic, and prove all the required properties. For primitive types like `Int` or `Double`, it's "trivial" because we trust the SMT solver which already knows about integers. \begin{code} {-@ axiomatize leqInt @-} leqInt :: Int -> Int -> Bool leqInt x y = x <= y {-@ leqIntRefl :: x:Int -> { leqInt x x } @-} leqIntRefl :: Int -> Proof leqIntRefl x = leqInt x x ==. x <= x *** QED {-@ leqIntAntisym :: x:Int -> y:Int -> { leqInt x y && leqInt y x ==> x == y } @-} leqIntAntisym :: Int -> Int -> Proof leqIntAntisym x y = (leqInt x y && leqInt y x) ==. (x <= y && y <= x) ==. x == y *** QED {-@ leqIntTrans :: x:Int -> y:Int -> z:Int -> { leqInt x y && leqInt y z ==> leqInt x z } @-} leqIntTrans :: Int -> Int -> Int -> Proof leqIntTrans x y z = (leqInt x y && leqInt y z) ==. (x <= y && y <= z) ==. x <= z ==. leqInt x z *** QED {-@ leqIntTotal :: x:Int -> y:Int -> { leqInt x y || leqInt y x } @-} leqIntTotal :: Int -> Int -> Proof leqIntTotal x y = (leqInt x y || leqInt y x) ==. (x <= y || y <= x) *** QED vordInt :: VerifiedOrd Int vordInt = VerifiedOrd leqInt leqIntRefl leqIntAntisym leqIntTrans leqIntTotal \end{code} How about a complex datatype? Let's consider an inductive datatype, the peano natural numbers. Not surprisingly, the proofs follow by induction. For conciseness, we only show totality and elide the rest. Writing down proofs for more complex datatypes is tedious, it requires case analysis for each constructor and using the induction hypothesis. However, we can decompose Haskell datatypes into sums and products, and we can build up compound proofs using isomorphisms! To that end, we design some machinery to express isomorphisms, and prove that laws are preserved under isomorphic images. \begin{code} {-@ data Iso a b = Iso { to :: a -> b , from :: b -> a , tof :: y:b -> { to (from y) == y } , fot :: x:a -> { from (to x) == x } } @-} data Iso a b = Iso { to :: a -> b , from :: b -> a , tof :: b -> Proof , fot :: a -> Proof } \end{code} An isomorphism between types `a` and `b` is a pair of functions `to :: a -> b` and `from :: b -> a` that are mutually inverse to each other. We now claim that total order laws are preserved under `Iso`; this amounts to building up a `VerifiedOrd b`, given a `VerifiedOrd a` and an `Iso a b`. \begin{code} {-@ axiomatize leqFrom @-} leqFrom :: (a -> a -> Bool) -> (b -> a) -> (b -> b -> Bool) leqFrom leqa from x y = leqa (from x) (from y) {-@ leqFromRefl :: leqa:(a -> a -> Bool) -> leqaRefl:(x:a -> { leqa x x }) -> from:(b -> a) -> x:b -> { leqFrom leqa from x x } @-} leqFromRefl :: (a -> a -> Bool) -> (a -> Proof) -> (b -> a) -> b -> Proof leqFromRefl leqa leqaRefl from x = leqFrom leqa from x x ==. leqa (from x) (from x) ==. True ? leqaRefl (from x) *** QED {-@ leqFromAntisym :: leqa:(a -> a -> Bool) -> leqaAntisym:(x:a -> y:a -> { leqa x y && leqa y x ==> x == y }) -> to:(a -> b) -> from:(b -> a) -> tof:(y:b -> { to (from y) == y }) -> x:b -> y:b -> { leqFrom leqa from x y && leqFrom leqa from y x ==> x == y } @-} leqFromAntisym :: (Eq a, Eq b) => (a -> a -> Bool) -> (a -> a -> Proof) -> (a -> b) -> (b -> a) -> (b -> Proof) -> b -> b -> Proof leqFromAntisym leqa leqaAntisym to from tof x y = (leqFrom leqa from x y && leqFrom leqa from y x) ==. (leqa (from x) (from y) && leqa (from y) (from x)) ==. from x == from y ? leqaAntisym (from x) (from y) ==. to (from x) == to (from y) ==. x == to (from y) ? tof x ==. x == y ? tof y *** QED {-@ leqFromTrans :: leqa:(a -> a -> Bool) -> leqaTrans:(x:a -> y:a -> z:a -> { (leqa x y) && (leqa y z) ==> (leqa x z) }) -> from:(b -> a) -> x:b -> y:b -> z:b -> { leqFrom leqa from x y && leqFrom leqa from y z ==> leqFrom leqa from x z } @-} leqFromTrans :: (a -> a -> Bool) -> (a -> a -> a -> Proof) -> (b -> a) -> b -> b -> b -> Proof leqFromTrans leqa leqaTrans from x y z = (leqFrom leqa from x y && leqFrom leqa from y z) ==. (leqa (from x) (from y) && leqa (from y) (from z)) ==. leqa (from x) (from z) ? leqaTrans (from x) (from y) (from z) ==. leqFrom leqa from x z *** QED {-@ leqFromTotal :: leqa:(a -> a -> Bool) -> leqaTotal:(x:a -> y:a -> { (leqa x y) || (leqa y x) }) -> from:(b -> a) -> x:b -> y:b -> { leqFrom leqa from x y || leqFrom leqa from y x } @-} leqFromTotal :: (a -> a -> Bool) -> (a -> a -> Proof) -> (b -> a) -> b -> b -> Proof leqFromTotal leqa leqaTotal from x y = (leqFrom leqa from x y || leqFrom leqa from y x) ==. (leqa (from x) (from y) || leqa (from y) (from x)) ==. True ? leqaTotal (from x) (from y) ==. leqFrom leqa from y x *** QED vordIso :: (Eq a, Eq b) => Iso a b -> VerifiedOrd a -> VerifiedOrd b vordIso (Iso to from tof fot) (VerifiedOrd leqa leqaRefl leqaAntisym leqaTrans leqaTotal) = VerifiedOrd (leqFrom leqa from) (leqFromRefl leqa leqaRefl from) (leqFromAntisym leqa leqaAntisym to from tof) (leqFromTrans leqa leqaTrans from) (leqFromTotal leqa leqaTotal from) \end{code} We can now write a `VerifiedOrd` for `Peano` by mapping it onto `type Nat = { n:Int | 0 <= n }`, which is easier to prove properties about because it's just an integer! \begin{code} {-@ type Nat = { n:Int | 0 <= n } @-} type Nat = Int {-@ axiomatize fromNat @-} {-@ fromNat :: Nat -> Peano @-} fromNat :: Nat -> Peano fromNat n | n == 0 = Z | otherwise = S (fromNat (n - 1)) {-@ toFrom :: x:Nat -> { toNat (fromNat x) == x } @-} toFrom :: Nat -> Proof toFrom n | n == 0 = toNat (fromNat 0) ==. toNat Z ==. 0 *** QED | n > 0 = toNat (fromNat n) ==. toNat (S (fromNat (n - 1))) ==. 1 + toNat (fromNat (n - 1)) ==. 1 + (n - 1) ? toFrom (n - 1) ==. n *** QED {-@ fromTo :: x:Peano -> { fromNat (toNat x) == x } @-} fromTo :: Peano -> Proof fromTo Z = fromNat (toNat Z) ==. fromNat 0 ==. Z *** QED fromTo (S n) = fromNat (toNat (S n)) ==. fromNat (1 + toNat n) ==. S (fromNat ((1 + toNat n) - 1)) ==. S (fromNat (toNat n)) ==. S n ? fromTo n *** QED {-@ isoNatPeano :: Iso Nat Peano @-} isoNatPeano :: Iso Nat Peano isoNatPeano = Iso fromNat toNat fromTo toFrom vordNat :: VerifiedOrd Nat vordNat = VerifiedOrd leqInt leqIntRefl leqIntAntisym leqIntTrans leqIntTotal vordPeano :: VerifiedOrd Peano vordPeano = vordIso isoNatPeano vordNat \end{code} Similarly, we can break down compound datatypes into sums and products, just like what happens in `GHC.Generics`, and use the isomorphism to write down `VerifiedOrd` instances. We do however need to provide instances of `VerifiedOrd` for sums and products. \begin{code} vordProd :: VerifiedOrd a -> VerifiedOrd b -> VerifiedOrd (a, b) vordSum :: VerifiedOrd a -> VerifiedOrd b -> VerifiedOrd (Either a b) \end{code} \begin{comment} \begin{code} vordSum = undefined vordProd = undefined \end{code} \end{comment} Our [library](https://github.com/iu-parfunc/verified-instances) explores this idea to build some verified typeclasses, such as `VerifiedOrd` and `VerifiedMonoid`. We also provide combinators to build verified instances by using isomorphisms. Also, using a [reflection hack](https://www.schoolofhaskell.com/user/thoughtpolice/using-reflection), we can reify these "verified" terms to typeclass dictionaries at runtime, to call legacy functions which require `Ord` constraints and so on.