verifying the four colour theorem georges gonthier microsoft research cambridge

Verifying theFour Colour Theorem

Georges GonthierMicrosoft Research Cambridge

150 years of history…

1852 Conjecture (Guthrie → DeMorgan) 1878 Publication (Cayley) 1879 First proof (Kempe) 1880 Second proof (Tait) 1890 Rebuttal (Heawood) 1891 Second rebuttal (Petersen) 1913 Reducibility, connexity (Birkhoff) 1922 Up to 25 regions (Franklin) 1969 Discharging (Heesch) 1976 Computer proof (Appel & Haken) 1995 Streamlining (Robertson & al.) 2004 Self checking proof (Gonthier)

So what about it ?

It shows software can be as reliable as math.

It’s been done by applying computer science to mathematics.

The art of computer proving is maturing.

Outline

The Four Colour Theoremwhat it sayshow it’s proved

Formal proofsproving with programsthinking with programs

Computer proofshow it’s done

The Theorem

Every simple planar map can be colored with only four colors

disjoint subsets of R x Ropen and connected

∃good covering map with at most four regions

adjacent regions covered with different colors

have a common border point that is not a corner

touches more than two regions

Record simple_map (m : map) : Prop := SimpleMap {

simple_map_proper :> proper_map m; map_open : ∀z, open (m z); map_connected : ∀z, connected (m z) }.

Record coloring (m, k : map) : Prop := Coloring { coloring_proper :> proper_map k; coloring_inmap : subregion (inmap k) (inmap m); coloring_covers : covers m k; coloring_adj : ∀z1 z2, k z1 z2 → adjacent m z1 z2 → m z1 z2

}.

Definition map_colorable n m := ∃2 k, coloring m k & size_at_most n k.

Theorem four_color : ∀m, simple_map m → map_colorable 4 m.

The formal theorem

Definition covers (m, m' : map) := ∀z : point, subregion (m z) (m' z).

Definition size_at_most n m := ∃ f, ∀ z, inmap m z → ∃2 i, i < n & m (f

i).

Definition border m z1 z2 := intersect (closure (m z1)) (closure (m

z2)).

Definition corner_map (m : map) z : map :=

fun z1 z2 ⇒ m z1 z2 ∧ closure (m z1) z.

Definition not_corner m z := size_at_most 2 (corner_map m z).

Definition adjacent m z1 z2 := meet (not_corner m) (border m z1 z2).

Outline

The Four Colour Theoremwhat it sayshow it’s proved

Formal proofsproving with programsthinking with programs

Computer proofshow it’s done

Colouring by induction

Configuration

reduciblering

Cubic maps

Euler formula

R + N – E = 22E/R = 6 - 12/Rthe world’s a football

#sides

Small rings

Any non-trivial ring < 6 defines reducible configurations (Birkhoff, 1913)

If there are no small rings, then all 2-neighborhoods are double rings

look for configurations in 2-neighborhoods!

The proof in a nutshell

Find a set of configurations such that:

(A) reducibility: Each one can be coloured to match any planar ring colouring.

(B) unavoidability: At least one appears in any planar map.

Verify that the combinatorics fit the topology (graph theory + analysis).

10,000 cases1,000,000,000 cases

Progress in verification1976 A & H

IBM 370reducibility

2 books1000’s of

figures

1995 RSST

C programreducibility

unavoidability

35-pagepaper

2005 MSR

Gallinareducibility

unavoidabilitygraph theory

topologydata structures

…

35 lines ofdefinitions

?

?

Outline

The Four Colour Theoremwhat it sayshow it’s proved

Formal proofsproving with programsthinking with programs

Computer proofshow it’s done

The Poincaré principle How do you prove: Given 2 1+ (1+ 0)≝ 4 1+ (1+ (1+ (1+ 0)))≝ n + m ≝ if n is 1+ n’ then 1+ (n’ + m) else m

(a recursive program)

2 + 2 = 4 ?

a: 0 + 2 = 2 (neutral left)

b: (1 + 0) + 2 = 1 + (0 + 2) (associativity)

c: 2 + 2 = 1 + ((1 + 0) + 2) (def, associativity)

d: 2 + 2 = 1 + (1 + (0 + 2)) (replace b in c)

e: (replace a in d, def)a: (def, calculation)

Reflecting reducibility

SetupVariable cf : config.

Definition cfreducible : Prop := …

Definition check_reducible : bool := …

Lemma check_reducible_valid : check_reducible -> cfreducible.

UsageLemma cfred232 : cfreducible (Config 11 33 37 H 2 H 13 Y 5 H 10

H 1 H 1 Y 3 H 11 Y 4 H 9 H 1 Y 3 H 9 Y 6 Y 1 Y 1 Y 3 Y 1 Y Y 1 Y).

Proof. apply check_reducible_valid; by compute. Qed.20,000,000 cases

Border colouring (Tait 1880)

Chord flips (Kempe 1879)

(

)0

)1

•

(

•

•

chromogram

colouring

Model checking colourings

1

Ξ

0

Λ

0

Λi

Ξi Λ

Ξ

-1

Λi+1

Yes

No

Ξi+1

Ξ, Λ

restrict Λ

decrement Ξ

any ?

NoNo

kempe

Chromogram twiddling

Fixpoint gram_neg_rec (n : nat) (w : chromogram) {struct w}

: chromogram := match w, n with | Adds Gpush w', _ => Adds Gpush (gram_neg_rec (S n) w') | Adds Gskip w', _ => Adds Gskip (gram_neg_rec n w') | Adds s w', S n => Adds s (gram_neg_rec n w') | Adds Gpop0 w', O => Adds Gpop1 w' | Adds Gpop1 w', O => Adds Gpop0 w' | seq0, _ => w end.

Definition gram_neg := gram_neg_rec 0.

Correctness proof

Lemma match_gram_neg : forall b0 et w, matchg (Seq b0) et (gram_neg w) = matchg (Seq (¬ b0)) et w.Proof.move => b0 et w; set sb : bitseq := seq0.have Esb: forall b : bool, Adds b sb = add_last sb b by

done.rewrite /gram_neg -[0]/(size sb) 2!{}Esb.elim: w et sb => [|s w IHw] et lb; first by case lb.case Ds: s; (case: et => [|e et]; first by case lb); first [ by case: e (IHw et (Adds _ lb)) => /= | by case: e; case: lb => [|b lb]; rewrite /= ?if_negb ?IHw ].Qed.

Outline

The Four Colour Theoremwhat it sayshow it’s proved

Formal proofsproving with programsthinking with programs

Computer proofshow it’s done

Formalizing maps

n

f

node

edge

dart

e

Rings and disks

rest

disk

Pasting configurations

disk

paste

rest

contour cycle

full map

Folkloric proofs?

(3.3) Let K be a configuration appearing in a triangulation T, and let S be the free completion of K. Then there is a projection ɸ of S into T such that ɸ (x) = x for all x ∈ V(G(K)) ∪ E(G(K)) ∪ F(G(K)).

This is a “folklore” theorem, and we omit its [lengthy] proof…

Definition phi x := if ac x then h x else if ac (edge x) then edge (h (edge x)) else if ac (node x) then face (edge (h (node x))) else edge (node (node (h (node (edge x))))).

Unavoidable pattern

u

u l

u r

u l

u r

u r

u l

u

u

h

h

h l

h l

h r

h r

f0r

f1r

f2r

f1l

f1l

f2l

f0l

f0r

hubspoke

spoke

spokehat

fanfan

fan

left step

right step

subpart

unreachable dart

spoke

TopologyEuler: (n + 1)2 + n2 + 1 – 2n(n+1) = 2

Outline

The Four Colour Theoremwhat it sayshow it’s proved

Formal proofsproving with programsthinking with programs

Computer proofshow it’s done

1 subgoal

d : dataSet e : rel d============================ forall (p : seq d) (x : d), cycle p -> p x -> e x (next p x)

1 subgoal d : dataSet e : rel d y0 : d p : seq d x : d Hp : path y0 (add_last p y0)============================ setU1 y0 p x -> e x (next_at x y0 y0 p)

2 subgoals d : dataSet e : rel d y0 : d x : d y : d============================ e y y0 && true -> (y =d x) || false -> e x (if y =d x then y0 else x)

subgoal 2 is: e y y’ && path y’ (add_last p y0) -> or3b (y =d x) (y’ =d x) (p x) -> e x (if y =d x then y’ else next_at x y0 y’ p)

2 subgoals d : dataSet e : rel d y0 : d x : d y : d Hy : e y y0 Dy : y =d x============================ e x (if y =d x then y0 else x)

subgoal 2 is: e y y’ && path y’ (add_last p y0) -> or3b (y =d x) (y' =d x) (p x) -> e x (if y =d x then y’ else next_at x y0 y’ p)

1 subgoal d : dataSet e : rel d y0 : d x : d y’ : d p : seq d IHp : forall x0 : d, path x0 (add_last p y0) -> setU1 x0 p x -> e x (next_at x y0 x0 p) y : d ============================ e y y’ && path y’ (add_last p y0) -> or3b (y =d x) (y’ =d x) (p x) -> e x (if y =d x then y’ else next_at x y0 y’ p)

1 subgoal d : dataSet e : rel d y0 : d x : d y’ : d p : seq d IHp : forall x0 : d, path x0 (add_last p y0) -> setU1 x0 p x -> e x (next_at x y0 x0 p) y : d Hy : e y y’ Hp : path y’ (add_last p y0) ============================ or3b (y =d x) (y’ =d x) (p x) -> e x (if y =d x then y’ else next_at x y0 y’ p)

1 subgoal d : dataSet e : rel d y0 : d x : d y’ : d p : seq d IHp : forall x0 : d, path x0 (add_last p y0) -> setU1 x0 p x -> e x (next_at x y0 x0 p) y : d Hy : e y y’ Hp : path y’ (add_last p y0) ============================ or3b false (y’ =d x) (p x) -> e x (next_at x y0 y’ p)

Subtree proved!

Proof workbench

Lemma next_cycle : ∀p x, cycle p → p x → e x (next p x).Proof.move⇒ [ | y0 p] x //= Hp.elim: p {1 3 5}y0 Hp ⇒ [ | y' p IHp] y /=; rewrite eqdsym /setU1. - rewrite andbT orbF ⇒ Hy Dy. by rewrite Dy -(eqP Dy).move/andP⇒ [Hy Hp].case: (y =P x) ⇒ [←| _] //.exact: IHp Hp.Qed.

proof script

goal

context

other goal

e

e

e

e

e

e

e

x

next p x

cycle pp

recursion assumption

Five tactic summary

move: x y Hxy … => z t Hzt …bookkeeping : context goal moves

case: x Hxy … => [a | b c] …decomposition, induction (with elim)

rewrite /= orbF –[x]addn0 …partial evaluation, simplification

apply: lemma… => [x y Hxy | n | …] …match goal (backward chaining)

have label: conjecture by …set goal (forward chaining)

Some conclusions

Proof assistants can give real proofs of real theorems

Machine formalization can lead to new mathematical insights

Formally proving programs can be easier than formally proving theorems