diff options
author | Ohad Kammar <ohad.kammar@ed.ac.uk> | 2022-08-07 23:35:19 +0100 |
---|---|---|
committer | Ohad Kammar <ohad.kammar@ed.ac.uk> | 2022-08-07 23:35:19 +0100 |
commit | 85b191cd3134102c7205c1771806ee16f73b3a89 (patch) | |
tree | 9ca3e343244e8a1e5fc15f4874e220f4c852651a /doc/sources/Tutorial.md | |
parent | bc260cdfa7b309820594b7caf14f62be8343358e (diff) |
Write about equivalence relations
Diffstat (limited to 'doc/sources/Tutorial.md')
-rw-r--r-- | doc/sources/Tutorial.md | 230 |
1 files changed, 229 insertions, 1 deletions
diff --git a/doc/sources/Tutorial.md b/doc/sources/Tutorial.md index f143d45..85b09c8 100644 --- a/doc/sources/Tutorial.md +++ b/doc/sources/Tutorial.md @@ -1,5 +1,17 @@ ```idris hide module Tutorial + +import Data.Setoid +import Syntax.PreorderReasoning +import Syntax.PreorderReasoning.Setoid +import Data.List.Elem +import Data.List +import Data.Nat +import Frex +import Frexlet.Monoid.Commutative +import Notation.Additive +import Frexlet.Monoid.Notation.Additive +import Frexlet.Monoid.Commutative.Nat ``` # Tutorial: Setoids @@ -15,7 +27,223 @@ tutorial you will: + types with an equality relation that carries additional information -## Basic interface +If you want to see the source-code behind this tutorial, check the +[source-code](sources/Tutorial.md) out. + +## Equivalence relations + +A _relation_ over a type `ty` in Idris is any two-argument type-valued function: +``` +namespace Control.Relation + Rel : Type -> Type + Rel ty = ty -> ty -> Type +``` +This definition and its associated interfaces ship with idris's standard +library. Given a relation `rel : Rel ty` and `x,y : ty`, we can form +```x `rel` y : Type```: the type of ways in which `x` and `y` can be related. + +For example, two lists _overlap_ when they have a common element: +```idris +record Overlap {0 a : Type} (xs,ys : List a) where + constructor Overlapping + common : a + lhsPos : common `Elem` xs + rhsPos : common `Elem` ys + +``` +Lists can overlap in exactly one position: +```idris +Ex1 : Overlap [1,2,3] [6,7,2,8] +Ex1 = Overlapping + { common = 2 + , lhsPos = There Here + , rhsPos = There (There Here) + } +``` +But they can overlap in several ways: ```idris +Ex2a , +Ex2b , +Ex2c : Overlap [1,2,3] [2,3,2] +Ex2a = Overlapping + { common = 3 + , lhsPos = There (There Here) + , rhsPos = There Here + } +Ex2b = Overlapping + { common = 2 + , lhsPos = There Here + , rhsPos = Here + } +Ex2c = Overlapping + { common = 2 + , lhsPos = There Here + , rhsPos = There (There Here) + } +``` +We can think of a relation `rel : Rel ty` as the type of edges in a directed +graph between vertices in `ty`: + +* edges have a direction: the type `rel x y` is different to `rel y x` + +* multiple different edges between the same vertices `e1, e2 : rel x y` +* self-loops between the same vertex are allowed `loop : rel x x`. + +An _equivalence relation_ is a relation that's: + +* _reflexive_: we guarantee a specific way in which every element is related to + itself; + +* _symmetric_: we can reverse an edge between two edges; and + +* _transitive_: we can compose paths of related elements into a single edge. + +```idris +record Equivalence (A : Type) where + constructor MkEquivalence + 0 relation: Rel A + reflexive : (x : A) -> relation x x + symmetric : (x, y : A) -> relation x y -> relation y x + transitive: (x, y, z : A) -> relation x y -> relation y z + -> relation x z +``` +```idris hide +%hide Tutorial.Equivalence +``` +We equip the built-in relation `Equal` with the structure of an equivalence +relation, using the constructor `Refl` and the stdlib functions `sym`, and +`trans`: +```idris +EqualityEquivalence : Equivalence a +EqualityEquivalence = MkEquivalence + { relation = (===) + , reflexive = \x => Refl + , symmetric = \x,y,x_eq_y => sym x_eq_y + , transitive = \x,y,z,x_eq_y,y_eq_z => trans x_eq_y y_eq_z + } +``` + +We'll use the following relation on pairs of natural numbers as a running +example. We can represent an integer as the difference between a pair of natural +numbers: +```idris +infix 8 .-. + +record INT where + constructor (.-.) + pos, neg : Nat + +record SameDiff (x, y : INT) where + constructor Check + same : (x.pos + y.neg === y.pos + x.neg) +``` +The `SameDiff x y` relation is equivalent to mathematical equation that states +that the difference between the positive and negative parts is identical: +$$x_{pos} - x_{neg} = y_{pos} - y_{neg}$$ +But, unlike the last equation which requires us to define integers and +subtraction, its equivalent `(.same)` is expressed using only addition, and +so addition on `Nat` is enough. + +The relation `SameDiff` is an equivalence relation. The proofs are +straightforward, and a good opportunity to practice Idris's equational +reasoning combinators from `Syntax.PreorderReasoning`: +```idris +SameDiffEquivalence : Equivalence INT +SameDiffEquivalence = MkEquivalence + { relation = SameDiff + , reflexive = \x => Check $ Calc $ + |~ x.pos + x.neg + ~~ x.pos + x.neg ...(Refl) +``` +This equational proof represents the single-step equational proof: + +"Calculate: + +1. $x_{pos} + x_{neg}$ +2. $= x_{pos} + x_{neg}$ (by reflexivity)" + +The mnemonic behind the ASCII-art is that the first step in the proof +starts with a logical-judgement symbol $\vdash$, each step continues with an +equality sign $=$, and justified by a thought bubble `(...)`. + +```idris + , symmetric = \x,y,x_eq_y => Check $ Calc $ + |~ y.pos + x.neg + ~~ x.pos + y.neg ..<(x_eq_y.same) +``` + In this proof, we were given the proof `x_eq_y.same : x.pos + y.neg = y.pos + x.neg` + and so we appealed to the symmetric equation. The mnemonic here is that the + last bubble in the thought bubble `(...)` is replace with a left-pointing arrow, + reversing the reasoning step. +```idris + , transitive = \x,y,z,x_eq_y,y_eq_z => Check $ plusRightCancel _ _ y.pos + $ Calc $ + |~ x.pos + z.neg + y.pos + ~~ x.pos + (y.pos + z.neg) ...(solve 3 Monoid.Commutative.Free.Free + {a = Nat.Additive} $ + (X 0 .+. X 1) .+. X 2 + =-= X 0 .+. (X 2 .+. X 1)) + ~~ x.pos + (z.pos + y.neg) ...(cong (x.pos +) $ y_eq_z.same) + ~~ (x.pos + y.neg) + z.pos ...(solve 3 Monoid.Commutative.Free.Free + {a = Nat.Additive} $ + X 0 .+. (X 1 .+. X 2) + =-= (X 0 .+. X 2) .+. X 1) + ~~ (y.pos + x.neg) + z.pos ...(cong (+ z.pos) ?h2) + ~~ z.pos + x.neg + y.pos ...(solve 3 Monoid.Commutative.Free.Free + {a = Nat.Additive} $ + (X 0 .+. X 1) .+. X 2 + =-= (X 2 .+. X 1) .+. X 0) + } +``` +This proof is a lot more involved: + +1. We appeal to the cancellation property of addition: + $a + c = b + c \Rightarrow a = b$ +2. We rearrange the term, bringing the appropriate part of `y` into contact with + the appropriate part of `z` and `x` to transform the term. + +Here we use the idris library [`Frex`](http://www.github.com/frex-project/idris-frex) +that can perform such routine +rearrangements for common algebraic structures. In this case, we use the +commutative monoid simplifier from `Frex`. +If you want to read more about `Frex`, check the +[paper](https://www.denotational.co.uk/drafts/allais-brady-corbyn-kammar-yallop-frex-dependently-typed-algebraic-simplification.pdf) out. + + +Idris's `Control.Relation` defines interfaces for properties like reflexivity +and transitivity. While the +setoid package doesn't use them, we'll use them in a few examples. + +The `Overlap` relation from Examples 1 and 2 is symmetric: +```idris +Symmetric (List a) Overlap where + symmetric xs_overlaps_ys = Overlapping + { common = xs_overlaps_ys.common + , lhsPos = xs_overlaps_ys.rhsPos + , rhsPos = xs_overlaps_ys.lhsPos + } +``` +However, `Overlap` is neither reflexive nor transitive: + +* The empty list doesn't overlap with itself: +```idris +Ex3 : Not (Overlap [] []) +Ex3 nil_overlaps_nil = case nil_overlaps_nil.lhsPos of + _ impossible +``` + +* Two lists may overlap with a middle list, but on different elements. For example: +```idris +Ex4 : ( Overlap [1] [1,2] + , Overlap [1,2] [2] + , Not (Overlap [1] [2])) +Ex4 = + ( Overlapping 1 Here Here + , Overlapping 2 (There Here) Here + , \ one_overlaps_two => case one_overlaps_two.lhsPos of + There _ impossible + ) ``` +The outer lists agree on `1` and `2`, respectively, but they can't overlap on +on the first element of either, which exhausts all possibilities of overlap. |