Solving and Verifying the Boolean Pythagorean

Triples Problem via Cube-and-Conquer

Marijn J.H. Heule

Joint work with

Oliver Kullmann and Victor W. Marek

1/40
Introduction

Solving Framework for Hard Problems

Encode
Transform
Split
Solve
Validate

Results

Conclusions and Future Work

2/40
Introduction

3/40
Schur’s Theorem [Schur 1917]
Can the set of natural numbers {1, 2, . . . } be partitioned into
k parts with no part containing a, b, and c such that a+b = c?
Otherwise, what is the smallest finite counter-example?

4/40
Schur’s Theorem [Schur 1917]
Can the set of natural numbers {1, 2, . . . } be partitioned into
k parts with no part containing a, b, and c such that a+b = c?
Otherwise, what is the smallest finite counter-example?
{1} {1} {1, 4} {1, 4}
→ → → → ×
{} {2} {2} {2, 3}
1+4 = 5
init 1+1 = 2 2+2 = 4 1+3 = 4
2+3 = 5

4/40
Schur’s Theorem [Schur 1917]
Can the set of natural numbers {1, 2, . . . } be partitioned into
k parts with no part containing a, b, and c such that a+b = c?
Otherwise, what is the smallest finite counter-example?
{1} {1} {1, 4} {1, 4}
→ → → → ×
{} {2} {2} {2, 3}
1+4 = 5
init 1+1 = 2 2+2 = 4 1+3 = 4
2+3 = 5
Theorem (Schur’s Theorem)
For each k > 0, there exists a number S(k), known as Schur
number k, such that [1, S(k)] can be partitioned into k parts
with no part containing a, b, and c such that a + b = c, while
this is impossible for [1, S(k)+1].
S(1) = 1, S(2) = 4, S(3) = 13, S(4) = 44 [Baumert 1965],
160 ≤ S(5) ≤ 315 [Exoo 1994, Fredricksen 1979].
4/40
Schur’s Theorem on Squares

Can the set of squares {12 , 22 , 32 , . . . } be partitioned into k

parts with no part containing a, b, and c such that a + b = c?

5/40
Schur’s Theorem on Squares

Can the set of squares {12 , 22 , 32 , . . . } be partitioned into k

parts with no part containing a, b, and c such that a + b = c?

The case k = 2 is already very difficult to determine:

{12 , 22 , 32 , 42 , 62 , 72 , 82 , 92 , 112 , 122 , 132 , 142 , 162 , 172 , 182 , 192 , . . . }
{52 , 102 , 152 , 202 , . . . }

5/40
Schur’s Theorem on Squares

Can the set of squares {12 , 22 , 32 , . . . } be partitioned into k

parts with no part containing a, b, and c such that a + b = c?

The case k = 2 is already very difficult to determine:

{12 , 22 , 32 , 42 , 62 , 72 , 82 , 92 , 112 , 122 , 132 , 142 , 162 , 172 , 182 , 192 , . . . }
{52 , 102 , 152 , 202 , . . . }

Partitioning the first thousand squares is easy (even manually).

A computer program can partition the first several thousands

squares ({12 , . . . , 76642 }) [Cooper and Overstreet 2015].

Can the infinite set of squares be partitioned into two parts?

5/40
Pythagorean Triples Problem [Erdős and Graham 1980]
The Boolean Pythagorean Triples problem is a reformulation
of Schur’s theorem on squares restricted to two parts:

Can the set of natural numbers {1, 2, 3, . . . } be partitioned

into two parts such that no part contains a Pythagorean triple
(a, b, c ∈ N with a2 + b 2 = c 2 )?

A partition into two parts is encoded using Boolean variables

xi with i ∈ {1, 2, 3, . . . } such that xi = 1 (= 0) means that i
occurs in the first (second) part. For each Pythagorean triple
(a, b, c) two clauses are added: (xa ∨ xb ∨ xc ) ∧ (x̄a ∨ x̄b ∨ x̄c ).

Theorem (Main result via parallel SAT solving)

[1, 7824] can be partitioned into two parts, such that no part
contains a Pythagorean triple. This is impossible for [1, 7825].
6/40
An Extreme Solution (a valid partition of [1, 7824])

7/40
Main Contribution

We present a framework that combines, for the first time, all

pieces to produce verifiable SAT results for very hard problems.

The status quo of using combinatorial solvers and years of

computation is arguably intolerable for mathematicians:
I Kouril and Paul [2008] computed the sixth van der

Waerden number (W (2, 6) = 1132) using dedicated

hardware without producing a proof.
I McKay’s and Radziszowski’s big result [1995] in Ramsey

Theory (R(4, 5) = 25) still cannot be reproduced.

We demonstrate our framework on the Pythagorean triples

problem, potentially the hardest problem solved with SAT yet.

8/40
 Solving Framework for
Hard-Combinatorial Problems

9/40
Overview of Solving Framework

transformed
encoder
formula

cubes
1: encode 2: transform 3: split 4: solve

original transform tautology cube

formula proof proof proofs

5: validate

10/40
Overview of Solving Framework: Phase 1

transformed
encoder
formula

cubes
1: encode 2: transform 3: split 4: solve

original transform tautology cube

formula proof proof proofs

5: validate

11/40
Phase 1: Encode

Input: encoder program

Output: the “original” CNF formula
Goal: make the translation to SAT as simple as possible

for ( int a = 1; a <= n ; a ++)

for ( int b = a ; b <= n ; b ++) {
int c = sqrt ( a * a + b * b );
if (( c <= n ) && (( a * a + b * b ) == ( c * c ))) {
addClause ( a , b , c );
addClause ( -a , -b , -c ); } }

F7824 has 6492 (occurring) variables and 18930 clauses, and

F7825 has 6494 (occurring) variables and 18944 clauses.

12/40
Overview of Solving Framework: Phase 2

transformed
encoder
formula

cubes
1: encode 2: transform 3: split 4: solve

original transform tautology cube

formula proof proof proofs

5: validate

13/40
Phase 2: Transform

Input: original CNF formula

Output: transformed CNF formula and a transformation proof
Goal: optimize the formula regarding the later (solving) phases

We applied two transformations:

I Pythagorean Triple Elimination removes Pythagorean

Triples that contain an element that does not occur in any

other Pythagorean Triple, e.g. 32 + 42 = 52 . (till fixpoint)
I Symmetry breaking places the number most frequently

occurring in Pythagorean triples (2520) in the first part.

All transformation (pre-processing) techniques can be

expressed using RAT steps [Järvisalo, Heule, and Biere 2012].

14/40
Phase 2: Blocked Clauses [Kullmann’99]
Definition (Blocking literal)
A literal l in a clause C of a CNF F blocks C w.r.t. F if
¯
for every clause D ∈ Fl¯, the resolvent (C \ {l}) ∪ (D \ {l})
obtained from resolving C and D on l is a tautology.
With respect to a fixed CNF and its clauses we have:
Definition (Blocked clause)
A clause is blocked if it contains a literal that blocks it.

15/40
Phase 2: Blocked Clauses [Kullmann’99]
Definition (Blocking literal)
A literal l in a clause C of a CNF F blocks C w.r.t. F if
¯
for every clause D ∈ Fl¯, the resolvent (C \ {l}) ∪ (D \ {l})
obtained from resolving C and D on l is a tautology.
With respect to a fixed CNF and its clauses we have:
Definition (Blocked clause)
A clause is blocked if it contains a literal that blocks it.
Example (Blocking literals and blocked clauses)
Consider the formula (a ∨ b) ∧ (a ∨ b̄ ∨ c̄) ∧ (ā ∨ c).
First clause is not blocked.
Second clause is blocked by both a and c̄.
Third clause is blocked by c.

15/40
Phase 2: Blocked Clauses [Kullmann’99]
Definition (Blocking literal)
A literal l in a clause C of a CNF F blocks C w.r.t. F if
¯
for every clause D ∈ Fl¯, the resolvent (C \ {l}) ∪ (D \ {l})
obtained from resolving C and D on l is a tautology.
With respect to a fixed CNF and its clauses we have:
Definition (Blocked clause)
A clause is blocked if it contains a literal that blocks it.
Example (Blocking literals and blocked clauses)
Consider the formula (a ∨ b) ∧ (a ∨ b̄ ∨ c̄) ∧ (ā ∨ c).
First clause is not blocked.
Second clause is blocked by both a and c̄.
Third clause is blocked by c.
Proposition
Removal of an arbitrary blocked clause preserves satisfiability.
15/40
Phase 2: Blocked Clause Elimination (BCE)
Definition (BCE)
While a clause C in a formula F is blocked, remove C from F .
Example (BCE)
Consider (a ∨ b) ∧ (a ∨ b̄ ∨ c̄) ∧ (ā ∨ c).
After removing either (a ∨ b̄ ∨ c̄) or (ā ∨ c), the clause
(a ∨ b) becomes blocked (no clause with either b̄ or ā).
An extreme case in which BCE removes all clauses!
Example (Pythagorean Triples)
The clauses (x3 ∨ x4 ∨ x5 ) and (x̄3 ∨ x̄4 ∨ x̄5 ) are blocked in
F7824 and F7825 (actually in any Fn ).
BCE (F7824 ) has 3740 variables and 14652 clauses, and
BCE (F7825 ) has 3745 variables and 14672 clauses.
BCE can simulate many high-level reasoning techniques.
[Järvisalo, Biere, and Heule 2010]
16/40
Overview of Solving Framework: Phase 3

transformed
encoder
formula

cubes
1: encode 2: transform 3: split 4: solve

original transform tautology cube

formula proof proof proofs

5: validate

17/40
Phase 3: Split
Input: transformed formula t
x5
f
Output: cubes and tautology proof x3 x2
f t t f
Goal: partition input in as x7 x3
many subproblems such that t f t f
total wallclock time is minimal x6
f t
Two layers of splitting F7824 :
I The top level split partitions

the transformed formula into a 5 -3 0

a 5 3 7 0
exactly a million subproblems; a 5 3 -7 0
I Each subproblem is a -5 2 0
partitioned into tens of a -5 -2 3 -6 0
thousands of subsubproblems. a -5 -2 3 6 0
a -5 -2 -3 0
Total time: 35,000 CPU hours

18/40
Cube-and-Conquer [Heule, Kullmann, Wieringa, and Biere 2011]
There exists two main SAT solving paradigms:
I Look-ahead aims to construct a small binary search-tree

using (expensive) global heuristics.

I Conflict-driven clause-learning (CDCL) aims to find a

short refutation using (cheap) local heuristics.

19/40
Cube-and-Conquer [Heule, Kullmann, Wieringa, and Biere 2011]
There exists two main SAT solving paradigms:
I Look-ahead aims to construct a small binary search-tree

using (expensive) global heuristics.

I Conflict-driven clause-learning (CDCL) aims to find a

short refutation using (cheap) local heuristics.

Combining look-ahead and CDCL, called cube-and-conquer,
does not work out of the box. Crucial details are:
I Partition a given formula into many (millions) of

subproblems. When just a few subproblems are created,

say only 32, the performance could actually decrease.
I Use heuristics to create equally hard subproblems, i.e., not

simply using the depth of the search-tree.

Cube-and-conquer solves many hard-combinatorial problems
significantly faster than both pure CDCL and pure look-ahead.
19/40
Details regarding the splitting heuristics
Splitting based on look-aheads:
I Count the number of assigned variables (large formulas)

I Count the number of clause reductions (medium formulas)

I Sum of the weighted new binary clauses (3-SAT)

20/40
Details regarding the splitting heuristics
Splitting based on look-aheads:
I Count the number of assigned variables (large formulas)

I Count the number of clause reductions (medium formulas)

I Sum of the weighted new binary clauses (3-SAT)

Equations (init, average, and update) for 3-SAT heuristics h4 :

1 X

h0 (x) = h0 (x̄) = 1 µi = hi (x) + hi (x̄)
2n
x∈var(F )

X  h (ȳ ) h (z̄)  X hi (ȳ )

i i
hi+1 (x) = max(α, min(β, · +γ )).
µi µi µi
(x∨y ∨z)∈F (x∨y )∈F

Rnd: α = 0.01, β = 25, γ = 3.3 Ptn: α = 8, β = 550, γ = 25

20/40
Overview of Solving Framework: Phase 4

transformed
encoder
formula

cubes
1: encode 2: transform 3: split 4: solve

original transform tautology cube

formula proof proof proofs

5: validate

21/40
Phase 4: Solve
Input: transformed formula and cubes
Output: cube proofs (or satisfying assignment)
Goal: solve (with proof logging) all cubes as fast as possible

Let ϕi be the i th cube with i ∈ [1, 1000000].

We first solved all F7824 ∧ ϕi , total runtime was 13, 000 CPU
hours (less than a day on the cluster). One cube is satisfiable.

The backbone of a formula is the set of literals that are

assigned to true in all solutions. The backbone of F7824 after
symmetry breaking consists of 2304 literals, including
I x5180 and x5865 , while 51802 + 58652 = 78252

I x̄625 and x̄7800 , while 6252 + 78002 = 78252

22/40
Phase 4: Short history of CDCL improvements
Conflict-driven clause learning (CDCL) has been the dominant
SAT solving paradigm and improved significantly in 20 years.
I Invented by Marques-Silva and Sakallah [1997];

I Dedicated data-structure and variable selection heuristics

made CDCL really competitive [MoskewiczMZZM 2001];

I Efficient implementation [Een and Sörensson 2003];

I New value selection and rapid restarts make CDCL ”local

search for UNSAT” [Pipatsrisawat and Darwiche 2007];

I An alternative restart implementation makes ultra rapid

restarts optimal [van der Tak, Ramos, and Heule 2011].

Crucial for our framework: reusing the heuristics and learnt

clauses while solving similar (sub)problems, known as
Incremental SAT Solving [Een and Sörensson 2003]
23/40
Overview of Solving Framework: Phase 5

transformed
encoder
formula

cubes
1: encode 2: transform 3: split 4: solve

original transform tautology cube

formula proof proof proofs

5: validate

24/40
Phase 5: Motivation for validating unsatisfiability proofs
Satisfiability solvers are used in amazing ways...
I Hardware and software verification (Intel and Microsoft)

I Hard-Combinatorial problems:
I van der Waerden numbers
[Dransfield, Marek, and Truszczynski, 2004; Kouril and Paul, 2008]
I Gardens of Eden in Conway’s Game of Life
[Hartman, Heule, Kwekkeboom, and Noels, 2013]
I Erdős Discrepancy Problem [Konev and Lisitsa, 2014]

..., but satisfiability solvers have errors and only return yes/no.
I Documented bugs in SAT, SMT, and QBF solvers

[Brummayer and Biere, 2009; Brummayer et al., 2010]

I Implementation errors often imply conceptual errors
I Mathematical results require a stronger justification than a
simple yes/no by a solver. UNSAT must be checkable.

25/40
Clausal Proof System [Järvisalo, Heule, and Biere 2012]

Learn: add a clause

* Preserve satisfiability

Unsatisfiable
* Learn empty clause
init
F
Satisfiable
* Forget last clause

Forget: remove a clause

* Preserve unsatisfiablity

26/40
Ideal Properties of a Proof System for SAT Solvers
Resolution Proofs
Zhang and Malik, 2003
Van Gelder, 2008; Biere, 2008

Easy to Emit Clausal Proofs

Goldberg and Novikov, 2003
Van Gelder, 2008

Compact Clausal proofs + clause deletion

Heule, Hunt, Jr., and Wetzler [STVR 2014]

Optimized clausal proof checker

Checked Efficiently Heule, Hunt, Jr., and Wetzler [FMCAD 2013]

Clausal RAT proofs

Expressive Heule, Hunt, Jr., and Wetzler [CADE 2013]

DRAT proofs (RAT + deletion)

Wetzler, Heule, and Hunt, Jr. [SAT 2014]

27/40
Ideal Properties of a Proof System for SAT Solvers
Resolution Proofs
Zhang and Malik, 2003
Van Gelder, 2008; Biere, 2008
Easy to Emit
Clausal Proofs
Goldberg and Novikov, 2003
Van Gelder, 2008
Compact
Clausal proofs + clause deletion
Heule, Hunt, Jr., and Wetzler [STVR 2014]

Checked Efficiently
Optimized clausal proof checker
Heule, Hunt, Jr., and Wetzler [FMCAD 2013]

Expressive
Clausal RAT proofs
Heule, Hunt, Jr., and Wetzler [CADE 2013]

Verified
DRAT proofs (RAT + deletion)
Wetzler, Heule, and Hunt, Jr. [SAT 2014]

27/40
Phase 5: Validate Pythagorean Triples Proofs.

original transform tautology cube

formula proof proof proofs

5: validate

We check the proofs with the DRAT-trim checker, which has

been used to validate the UNSAT results of the international
SAT Competitions since 2013.

Recently it was shown how to validate DRAT proofs in

parallel [Heule and Biere 2015].

The size of the merged proof is almost 200 terabyte and has
been validated in 16,000 CPU hours.
28/40
 Results
on the million subproblems

29/40
Histogram of Frequency of Cube Sizes

30/40
Runtime of Cube and Conquer Averaged per Cube Size

31/40
Scatterplot: Cube versus Conquer Runtimes

32/40
Scatterplot: Validation versus Conquer Runtimes

33/40
Scatterplot: CDCL versus Cube-and-Conquer Runtimes

34/40
Scatterplot: Look-ahead versus Cube-and-Conquer
Runtimes

35/40
Conclusions and Future Work

36/40
An Extreme Solution (a valid partition of [1, 7824])

37/40
Conclusions

Theorem (Main result)

[1, 7824] can be partitioned into two parts, such that no part
contains a Pythagorean triple. This is impossible for [1, 7825].

We solved and verified the theorem via SAT solving:

I Cube-and-conquer facilitated massive parallel solving.

I A new look-ahead heuristic was developed to substantially

reduce the search space.

I The proof is huge (200 terabyte), but can be compressed

to 68 gigabyte (13,000 CPU hours to decompress) and be

validated in 16,000 CPU hours.

38/40
 Heule’s Contributions to Solving Framework

[JAR 2012] transformed

encoder
[JAIR 2015] [LPAR 2010] formula
[HVC 2012] [JSAT 2011]
[ICGI 2010]
cubes
1: encode 2: transform 3: split 4: solve
[FAIA 2009]
[LPAR 2015] [HVC 2011] [CADE 2012]

original transform tautology cube

formula proof proof proofs

[SAT 2014] [FMCAD 2013]

5: validate [APPA 2014]
[STVR 2014]
[ITP 2013]

39/40
Future Directions
Apply our solving framework to other challenges in Ramsey
Theory and elsewhere:
I Existing results for which no proof was produced, for

example W(2,6) = 1132 [Kouril and Paul 2008].

I Century-old open problems appear solvable now, e.g. S(5).

Look-ahead heuristics are crucial and we had to develop

dedicated heuristics to solve the Pythagorean triples problem.
I Develop powerful heuristics that work out of the box.

I Alternatively, add heuristic-tuning techniques to the tool

chain [Hoos 2012].

Develop a mechanically-verified, fast clausal proof checker.

40/40
Future Directions
Apply our solving framework to other challenges in Ramsey
Theory and elsewhere:
I Existing results for which no proof was produced, for

example W(2,6) = 1132 [Kouril and Paul 2008].

I Century-old open problems appear solvable now, e.g. S(5).

Look-ahead heuristics are crucial and we had to develop

dedicated heuristics to solve the Pythagorean triples problem.
I Develop powerful heuristics that work out of the box.

I Alternatively, add heuristic-tuning techniques to the tool

chain [Hoos 2012].

Develop a mechanically-verified, fast clausal proof checker.

Thanks!
40/40