A Tutorial Introduction to Proof Complexity
Download
Report
Transcript A Tutorial Introduction to Proof Complexity
A Tutorial Introduction to Proof
Complexity
Paul Beame
University of Washington
Proof Systems and Their
Complexity
2
NP, proofs, and proof systems
L∊NP: there is a polynomial time computable V s.t.
x ∊ L ⇔ y . |y|≤|x|O(1). V(x,y)
V is a verifier for L
y is a proof that x ∊ L
Defn: A proof system* for L is a polytime (verifier) V s.t.
x ∊ L ⇔ y . V(x,y)
*Cook-Reckhow
defn: polytime computable map f from Σ* onto L≠∅
(f(x,y)=x if V(x,y) and = a fixed x0 ∊ L otherwise)
3
Complexity of a proof system
Defn: A proof system for L is a polytime (verifier) V s.t.
x ∊ L ⇔ y . V(x,y)
Defn: The complexity of proof system V is smallest fnctn
S: ℕ →ℕ s.t. x ∊ L ⇔ y . |y|≤ S(|x|). V(x,y)
Defn: V is polynomially bounded iff S(n) is nO(1)
Cor: L∊NP iff L has a polynomially bounded proof system
4
Propositional proof systems
Defn: A propositional proof system is a proof system V
for the set TAUT of propositional logic tautologies
F ∊ TAUT ⇔ proof P . V(F,P)
direction is usually called soundness
direction is usually called completeness
Alt Defn: A propositional proof system is a proof system
V for the set UNSAT of propositional logic contradictions
Since TAUT and UNSAT are coNP-complete…
Thm: NP=coNP iff there is a polynomially bounded
propositional proof system
5
Proof complexity vs search
Proof complexity only measures how large proofs
must be, not how easy they are to find
Lower bounds on proof complexity imply lower bounds on
nondeterministic algorithms
Stronger than lower bounds on deterministic search
Any complete SAT solver yields a propositional proof
system
Proof that F∊UNSAT is transcript of failed search on input F
Size of proof ≈ running time of SAT solver on input F
6
p-simulation
Defn: Proof system U polynomially simulates proof
system V iff
they prove the same language L
proofs in V can be efficiently converted into proofs in U
Defn: U and V are polynomially equivalent iff they
polynomially simulate each other
7
Proof systems using CNF input
By same trick [Tseitin 68, Cook 71] that reduces SAT to CNFSAT, can
assume w.l.o.g. that propositional proof systems are for the
language CNF-UNSAT
Given propositional formula F produce CNF formula F’
Add an extra variable yG corresponding to each
sub-formula G of F
F’ has clauses expressing the fact that yG takes on the
value of G determined by the inputs to F
Add clause yF to express that F is must be true
Thm: F’ ∊ CNF-UNSAT iff F ∊ UNSAT
8
Clauses
if F = G H, include clauses
yG yF
yH yF
yF yG yH
if F = G H, include clauses
yF yG
yF yH
yG yH yF
if F = G, include clauses
yF yG
yF yG
9
Sample propositional proof systems
Truth tables
proof is a fully filled out truth table
easy to verify that it is filled out correctly and all truth
assignments yield T
Axiom/Inference systems
inference rules: e.g. modus ponens A, (A B) | B
axioms: e.g. excluded middle | (A A)
axioms & inference rules are schemas
can make consistent substitution of arbitrary formulas for
variables in schema
e.g. excluded middle yields ((xy) (xy))
10
Resolution
Refutation system using CNF clauses only
Start with original input clauses of CNF F
Resolution rule
(A x), (B x) | (A B)
Goal: derive empty clause ⊥
11
Frege Systems
Finite, implicationally complete set R of axioms/inference rules
Refutation version:
Proof of unsatisfiability of F: sequence F1,…,Fr of formulas
(called lines) s.t.
F1 = F
each Fj follows from an axiom in R or follows from
previous ones via an inference rule in R
Fr = ⊥ trivial falsehood, e.g. (x x)
Positive version:
Start with nothing, end with tautology F
12
All Frege systems are p-equivalent
Key idea: Every use of a rule of one system can be
derived in the other system in a constant # of steps
13
𝓒-Frege proof systems
Many circuit complexity classes 𝓒 are defined as follows:
𝓒 = {f: f is computed by polynomial-size circuits
with structural property P𝓒}
Define 𝓒-Frege to be the p-equivalence class of Frege-style
proof systems s.t.
each line has structural property P𝓒
it has a finite set of axioms/inference rules that is complete
for circuits with property P𝓒
14
Some circuit classes 𝓒
P/poly - polysize circuits
NC1 - polysize formulas = O(log n)-depth fan-in 2 circuits
Clauses
k-DNF - k-DNF formulas
AC0 - constant-depth unbounded fan-in polysize
circuits using AND/OR/NOT gates
TC0 - threshold gates instead
15
Examples
Frege = NC1-Frege
NC1 (logarithmic depth fan-in 2) circuits can be expanded
into trees (formulas) of polynomial size
Formulas can always be re-balanced so they have
logarithmic depth (which are automatically polynomial size)
Resolution = Clauses-Frege
Every line is a clause
Res(k) = k-DNF-Frege
Every line is a k-DNF formula
16
Extended Frege Proofs
Like Frege proofs plus extra extension steps
Each extension step defines a new propositional variable to
stand for an arbitrary formula on the current set of variables
Using extension variables, each line represents a circuit in
the original variables
Extended-Frege = P/poly-Frege
Equivalent to Substitution-Frege in which each inferred
formula immediately is available as an axiom schema
Equivalent to Extended-Resolution which adds extension
clauses that define new variable y≡C using clauses y C
and z y for each z in C.
Idea: SAT to CNF-SAT conversion also works for circuits
17
The DAG of a proof
F1
Axioms/inputs
are sources
F5
F2
F3
F10
F8
F7
F6
F11
F4
F12
F9
F13
Sink labelled by tautology
(or ⊥ for a refutation)
Inference rule
associated with
each node
18
Proof structure
For any axiom/inference system can consider
Tree-like proofs
Restricted version in which the proof DAG must be a
tree. (Source formulas may be repeated.)
General (DAG-like) proofs
No restriction on the proof DAG
Other systems can be static (in which the entire proof
string can only be seen as one large inference)
E.g. Truth tables
Can sometimes use a similar idea for all three kinds of
proofs. Proof size: static ≥ tree-like ≥ general
19
Note: Special structure for resolution proofs
So far
Resolution
Proof structure is an unrestricted DAG
Tree-like resolution
Proof structure is a tree
Also
Regular resolution
A clause that is derived by resolving on variable x can
never lead to a clause in which variable x is re-introduced
Generalizes tree resolution which has this property wlog.
20
Davis-Putnam-Logemann-Loveland
(DPLL) Procedure
A family of complete SAT solvers
a collection of algorithms for finding SAT
assignments/proofs
Its traces form proofs of unsatifiability
≡ tree-like resolution refutations
21
Simple DPLL Algorithm
DPLL(F)
While (F contains a clause of size 1)
unit
set variable to make that clause true
propagation
simplify all clauses using this assignment
If F = ∅ then
output current truth assignment and HALT
If F does not contain ⊥ (empty clause) then
Choose unset literal x
Run DPLL(Fx0)
Run DPLL(Fx1 )
22
DPLL Refutation Trace
a
Clauses
0
1. a b c
2. ac
3. b
4. a d
5. d b
1
b
0
1
c
0
1
b
0
d
3
1
2
1
0
4
3
1
5
23
DPLL Refutation Trace
a
Clauses
0
1. a b c
2. ac
3. b
4. a d
5. d b
1
b
0
1
c
0
b
0
d
3
1
b
1
0
1
2
4
a b c
ac
a d
3
1
b
5
d b
24
Tree Resolution
a
Clauses
1. a b c
2. ac
3. b
4. a d
5. d b
b
c
b
d
3
3
b
b
1
2
4
a b c
ac
a d
5
d b
25
Tree Resolution
a
Clauses
1. a b c
2. ac
3. b
4. a d
5. d b
b
c
b
d
3
3
b
b
1
2
4
a b c
ac
a d
5
d b
26
Tree Resolution
a
Clauses
1. a b c
2. ac
3. b
4. a d
5. a b
b
b
d:a b
3
c: a b
3
b
b
1
2
4
a b c
ac
a d
5
d b
27
Tree Resolution
a
Clauses
1. a b c
2. ac
3. b
4. a d
5. a b
b
b
d:a b
3
c: a b
3
b
b
1
2
4
a b c
ac
a d
5
d b
28
Tree Resolution
a
Clauses
1. a b c
2. ac
3. b
4. a d
5. a b
b: a
b: a
c: a b
d:a b
3
3
b
b
1
2
4
a b c
ac
a d
5
d b
29
Tree Resolution
a
Clauses
1. a b c
2. ac
3. b
4. a d
5. a b
b: a
b: a
c: a b
d:a b
3
3
b
b
1
2
4
a b c
ac
a d
5
d b
30
Tree Resolution Refutation
a:⊥
Clauses
1. a b c
2. ac
3. b
4. a d
5. a b
b: a
b: a
c: a b
d:a b
3
3
b
b
1
2
4
a b c
ac
a d
5
d b
31
Conflict-Directed Clause-Learning (CDCL)
Complete SAT solvers that extend DPLL with clauses
inferred via special forms of resolution inference
Traces on unsatisfiable formulas yield resolution
refutations.
See tomorrow’s tutorial by Sam Buss to see their
generality
32
Proof systems based on other
representations of clauses
Algebraic:
Clause 𝑪 = (𝒙 ¬𝒚 𝒛) becomes equation
𝒇𝑪 ≡ 𝟏 − 𝒙 𝒚 𝟏 − 𝒛 = 𝟎 plus 𝒙𝟐 − 𝒙 = 𝟎 or …
𝒇𝑪 ≡ 𝒙′ 𝒚𝒛′ = 𝟎 plus equations 𝒙 + 𝒙’ − 𝟏 = 𝟎 etc.
Derive 𝟏 = 𝟎
0-1-inequalities:
Clause 𝑪 = (𝒙 ¬𝒚 𝒛) becomes inequalities
𝒙 + 𝟏 − 𝒚 + 𝒛 ≥ 𝟏 plus 𝒙 ≤ 𝟏 and 𝒙 ≥ 𝟎 etc.
Derive 𝟎 ≥ 𝟏
33
Hilbert’s Nullstellensatz
System of polynomials
𝒇𝟏(𝒙𝟏, … , 𝒙𝒏) = 𝟎, … , 𝒇𝒎(𝒙𝟏, … , 𝒙𝒏) = 𝟎
over field 𝑲 has no solution in any extension field of 𝑲
there exist polynomials 𝒈𝟏(𝒙𝟏, … , 𝒙𝒏), … , 𝒈𝒎(𝒙𝟏, … , 𝒙𝒏)
in 𝑲[𝒙𝟏, … , 𝒙𝒏] s.t. 𝒏𝒊=𝟏 𝒇𝒊 𝒈𝒊 ≡ 𝟏
34
Static: Nullstellensatz proof system
Clause 𝑪 = (𝒙𝟏 𝒙𝟐 𝒙𝟑) becomes equation
𝒇𝑪 ≡ (𝟏 − 𝒙𝟏)𝒙𝟐(𝟏 − 𝒙𝟑) = 𝟎
Add equations 𝒇𝒊 ≡ 𝒙𝟐𝒊 − 𝒙𝒊 = 𝟎 for each variable
Guarantees that only solutions are in {𝟎, 𝟏}𝒏
A refutation is polynomials 𝒈𝑪 and 𝒈𝒊 proving
unsatisfiability: i.e. such that
𝑪 𝒇𝑪 𝒈𝒄 + 𝒊 𝒇𝒊 𝒈𝒊 ≡ 𝟏
35
Polynomial Calculus with Resolution (PCR)
Axiom/inference system similar to static Nullstellensatz except:
Variables 𝒙𝒊 and 𝒙𝒊 ′ with 𝒙𝒊 + 𝒙′𝒊 = 𝟏, 𝒙𝟐𝒊 = 𝒙𝒊 , 𝒙′𝒊 𝟐 = 𝒙′𝒊
Clause 𝑪 = (𝒙¬𝒚 𝒛) becomes 𝒇𝑪 = 𝒙′ 𝒚𝒛′ = 𝟎
Given polynomials 𝑹 and 𝑺 can infer
𝒂𝑹 + 𝒃𝑺 for any 𝒂, 𝒃 in 𝑲
𝒙𝒊𝑹
Derive constant polynomial 𝟏
Degree = max degree of any polynomial in the proof
Size = # of monomials in proof
Can find proof of degree 𝒅 in time 𝒏𝐎
basis-like algorithm (linear algebra)
𝒅
using Groebner
36
PCR p-simulates resolution
Resolution
PCR
(abcd) (abcr)
(bcdr)
Given 𝒂’𝒃’𝒄’𝒅 and 𝒂𝒃’𝒄’𝒓
derive 𝒂’𝒃’𝒄’𝒅 𝒓 + (𝒂𝒃’𝒄’𝒓)𝒅
= 𝒂’ + 𝒂 𝒃’𝒄’𝒅𝒓
= 𝒃’𝒄’𝒅𝒓
37
Proof systems based on other
representations of clauses
Algebraic:
Clause 𝑪 = (𝒙 ¬𝒚 𝒛) becomes equation
𝒇𝑪 ≡ 𝟏 − 𝒙 𝒚 𝟏 − 𝒛 = 𝟎 plus 𝒙𝟐 −𝒙 = 𝟎 or …
𝒇𝑪 ≡ 𝒙′ 𝒚𝒛′ = 𝟎 plus equations 𝒙 + 𝒙’ − 𝟏 = 𝟎 etc.
Derive 𝟏 = 𝟎
0-1-inequalities:
Clause 𝑪 = (𝒙 ¬𝒚 𝒛) becomes inequality
𝒙 + 𝟏 − 𝒚 + 𝒛 ≥ 𝟏 plus 𝒙 ≤ 𝟏 and 𝒙 ≥ 𝟎 etc.
Derive 𝟎 ≥ 𝟏
38
Cutting Planes [Gomory 59, Chvatal 73]:
addition:
a1x1 + ... + anxn A
b1x1 + ... + bnxn B
(a1+b1)x1+...+(an+bn)xn A+B
multiplication by positive integer:
a1x1 + ... + anxn A
ca1x1 + ... + canxn cA
Division by positive integer:
ca1x1 + ... + canxn B
a1x1 + ... + anxn B/c
39
Cutting planes geometry
40
Cutting planes geometry
41
Cutting Planes p-simulates resolution
Resolution
Cutting
Planes
(abcd) (abcr)
(bcdr)
a + b + c + (1-d) 1
(1-a) + b + c + (1-r) 1
(1-d) 0
(1-r) 0
2b + 2c + 2(1-d) + 2(1-r) 1 Addition
b + c + (1-d) + (1-r) 1
Division
42
Semi-algebraic proof systems
Positivstellensatz
Start with same polynomial equalities 𝒇𝒋 = 𝟎 as
Nullstellensatz over ℝ
Write 𝒈𝒋 and 𝒉𝒌 such that
𝒋 𝒇𝒋 𝒈𝒋 = 𝟏 +
𝟐
𝒉
𝒌 𝒌
OR, start with linear inequalities 𝑳𝒋 ≥ 𝟎, derive 𝟎 ≥ 𝟏
Use higher-degree consequences including use of 𝒙𝟐 = 𝒙.
Possibly use 𝒑𝟐 (𝒙) ≥ 𝟎 for polynomials 𝒑(𝒙). (SDP)
Sherali-Adams static, no SDP
Lasserre/Sum-of-Squares static, SDP
Lovasz-Schrijver restricted dynamic, variants with SDP
43
ZFC
Some Proof System Relationships
P/poly-Frege
Frege
TC0-Frege
AC0-Frege
Positivstellensatz
Cutting Planes
Lasserre/SOS
Res(k)
Polynomial Calculus/PCR
Resolution
Nullstellensatz
DPLL
Truth Tables
44
Why all these proof systems?
Proof systems formalize different types of reasoning
Why even include the weaker systems within a given type of
reasoning?
many weaker proof systems have better associated proof
search strategies, e.g. DPLL, Polynomial Calculus.
Natural correspondence with circuit complexity classes
analyze systems working upwards in proof strength to gain
insight for techniques
Some proof systems correspond to our best algorithms for
NP-hard optimization problems
e.g. semi-definite programming and Lasserre/SOS
45
Lower Bound Methods
46
Width vs size in resolution proofs
Defn: If 𝑭 is a set of clauses
𝑹𝒆𝒔 (𝑭) = length of shortest resolution refutation of 𝑭
= ∞ if 𝑭 is satisfiable
𝒘(𝑭) = length of longest clause in 𝑭
𝒘𝒊𝒅𝒕𝒉(𝑭) = 𝐦𝐢𝐧𝑷 {𝒘(𝑷): 𝑷 is a resolution refutation of 𝑭}
Thm [Ben-Sasson,Wigderson]: Every resolution refutation of 𝑭 of
size 𝑺 can be converted to one of width 𝟐𝒏 ln 𝑺 + 𝒘(𝑭)
Corollary: 𝑹𝒆𝒔 𝑭 ≥ 𝒆𝒙𝒑(
𝒘𝒊𝒅𝒕𝒉 𝑭 – 𝒘 𝑭 𝟐
)
𝟐𝒏
47
Width vs size in resolution proofs
Thm [Ben-Sasson,Wigderson]: Every resolution refutation of 𝑭 of
size 𝑺 can be converted to one of width 𝟐𝒏 ln 𝑺 + 𝒘(𝑭)
Pf: Let 𝑾 = 𝟐𝒏 ln 𝑺. A clause 𝑪 is large iff 𝒘 𝑪 ≥ 𝑾.
By induction on 𝒏 and 𝒌: If (𝟏 − 𝑾/𝟐𝒏)𝒌 𝑺 ≤ 𝟏 then any 𝑭
with ≤ 𝑺 large clauses has refutation of width 𝒌 + 𝒘(𝑭)
Choose literal 𝒙 most frequently occurring in large clauses
and set it to 𝟏, satisfying 𝑾𝑺/𝟐𝒏 large clauses
Result is a proof of 𝑭𝒙𝟏 with 𝑺 𝟐(𝟏 − 𝑾/𝟐𝒏) large clauses
𝑾
−
𝑾 𝑺proof
By
𝒌𝒆−−ln
𝟏 +𝑺𝒘(𝑭)
𝟐𝒏 𝑺 =
𝒙𝟏 has
𝟏induction
− 𝑾 ∕ 𝑭𝟐𝒏
≤ 𝒆of width
𝑺=1
So can derive 𝒙 from 𝑭 in width 𝒌 + 𝒘(𝑭)
Also by induction, proof of 𝑭𝒙𝟎 of width 𝒌 + 𝒘(𝑭)
48
Width vs size in resolution proofs
Thm [Ben-Sasson,Wigderson]: Every resolution refutation of 𝑭 of
size 𝑺 can be converted to one of width 𝟐𝒏 ln 𝑺 + 𝒘(𝑭)
Pf: Let 𝑾 = 𝟐𝒏 ln 𝑺. A clause 𝑪 is large iff 𝒘 𝑪 ≥ 𝑾.
By induction on 𝒏 and 𝒌: If (𝟏 − 𝑾/𝟐𝒏)𝒌 𝑺 ≤ 𝟏 then any 𝑭
with ≤ 𝑺 large clauses has refutation of width 𝒌 + 𝒘(𝑭)
So can derive 𝒙 from 𝑭 in width 𝒌 + 𝒘(𝑭)
Also by induction, refutation of 𝑭𝒙𝟎 of width 𝒌 + 𝒘(𝑭)
Width 𝒌 + 𝒘(𝑭) refutation of 𝑭:
Derive 𝒙 from 𝑭
Resolve 𝒙 with all clauses of 𝑭 to yield 𝑭𝒙𝟎
Refute 𝑭𝒙𝟎
49
Width vs size in tree-resolution proofs
Defn: If 𝑭 is a set of clauses
𝑹𝒆𝒔𝑡𝑟𝑒𝑒 (𝑭) = length of shortest tree-resolution
refutation of 𝑭
= ∞ if 𝑭 is satisfiable
𝒘(𝑭) = length of longest clause in 𝑭
𝒘𝒊𝒅𝒕𝒉 (𝑭) = 𝐦𝐢𝐧𝑷 {𝒘(𝑷): 𝑷 is a resolution refutation of 𝑭}
Thm [Ben-Sasson,Wigderson]: Every DPLL/tree-resolution
refutation of 𝑭 of size 𝑺 can be converted to one of width
log2 𝑺 + 𝒘(𝑭)
Corollary: 𝑹𝒆𝒔𝑡𝑟𝑒𝑒 (𝑭) ≥ 𝟐𝒘𝒊𝒅𝒕𝒉(𝑭) – 𝒘(𝑭)
50
Proof: Width of Tree-like Resolution
⊥
𝒙
𝒙
S/2
S
51
Width of Tree-like Resolution
⊥
𝒙
𝑭 proves 𝒙 in
size at most S/2
𝒙
S/2
S
52
Width of Tree-like Resolution
⊥
𝒙
𝑭 proves 𝒙 in
size at most S/2
𝑭 proves 𝒙 in
size less than S
𝒙
S/2
S
53
Width of Tree-like Resolution
⊥
𝒙
Fx1 proves ⊥ in
size at most 𝑺/𝟐
Fx0 proves ⊥ in
size less than 𝑺
𝒙
S/2
S
54
Width of Tree-like Resolution
⊥
𝒙
Fx1 proves ⊥ in
size at most 𝑺/𝟐
𝒙
S/2
S
By induction Fx1 proves ⊥ in width at most
log2(𝑺/𝟐) + 𝒘(𝑭) = log2𝑺 + 𝒘(𝑭) − 𝟏
55
Width of Tree-like Resolution
⊥
𝒙
Fx1 proves ⊥ in
size at most 𝑺/𝟐
𝒙
𝒘
S
By induction Fx1 proves ⊥ in width at most
log2(𝑺/𝟐) + 𝒘(𝑭) = log2𝑺 + 𝒘(𝑭) − 𝟏
𝑭 proves in 𝒙 in width at most 𝒘 = log𝟐𝑺 + 𝒘(𝑭)
56
Width of Tree-like Resolution
⊥
𝒙
𝒙
𝒘
Fx0 proves ⊥ in
size < S
S
By induction Fx0 proves ⊥ in width
at most log2 𝑺 + 𝒘(𝑭)
57
Conclusion: Width of Tree-like Resolution
New Refutation:
1. Derive 𝒙 from 𝑭 in width 𝒘
2. Resolve 𝒙 with clauses
of 𝑭 containing 𝒙 to derive 𝑭𝒙𝟎
3. Prove ⊥ in width 𝒘 from 𝑭𝒙𝟎
𝒘 = log𝟐𝑺 + 𝒘(𝑭)
⊥
𝒘
𝒙
𝒙
𝒘
𝒙
𝒙
𝒘 𝒘 𝒘
58
Using width-size relationships
Cor: [Ben-SassonWigderson]:
𝑹𝒆𝒔
𝒘𝒊𝒅𝒕𝒉 𝑭 – 𝒘 𝑭 𝟐
𝑭 ≥ 𝒆𝒙𝒑
𝟐𝒏
(𝑭) ≥ 𝟐𝒘𝒊𝒅𝒕𝒉(𝑭) – 𝒘(𝑭)
𝑹𝒆𝒔𝑡𝑟𝑒𝑒
Cor: [BP 96]: If there is a resolution refutation of size
𝑺 then can find one in time/size 𝒏𝑶( 𝒏 ln 𝑺 )
Derive all clauses of width at most 𝑾 = 𝟐𝒏 ln 𝑺.
Cor: If there is a tree-resolution refutation of size 𝑺
then can find one in time/size 𝒏𝑶(𝐥𝐨𝐠 𝑺) = 𝑺𝑶(𝐥𝐨𝐠 𝒏) .
59
Notes
Relationship between width and size is roughly
optimal for general resolution
[Bonet, et al 99] Ordering tautologies with constant
input clause size and polynomial-size proofs that
require width 𝛀( 𝒏)
DPLL/tree-resolution can require exponentially larger
proofs than general resolution [BEGJ 98],[BW 98].
Polynomial versus 𝟐Ω(𝒏/log 𝒏) size
Uses graph pebbling and width-based lower bound
60
Bounding width: boundary expansion
𝑭 - a set of clauses
𝜹𝑭 - boundary of 𝑭 = set of variables appearing
in exactly one clause of 𝑭
𝒔(𝑭) - minimum size non-empty subset
of 𝑯 ⊆ 𝑭 with 𝜹𝑯 = ∅
𝒆(𝑭) - boundary expansion of 𝑭
= min 𝜹𝑯 : 𝑯 ⊆ 𝑭,
𝒔(𝑭)
𝟑
≤ 𝑯 ≤
𝟐𝒔(𝑭)
𝟑
61
Width vs boundary expansion
Lemma [ChvatalSzemeredi] 𝒘𝒊𝒅𝒕𝒉(𝑭) ≥ 𝒆(𝑭).
≥ 𝒔(𝑭)
⊥
62
Width vs boundary expansion
Lemma [ChvatalSzemeredi] 𝒘𝒊𝒅𝒕𝒉(𝑭) ≥ 𝒆(𝑭).
≥ 𝒔(𝑭)
Start at ⊥
Go to parent with
more sources until
≤ 𝟐𝒔(𝑭)/𝟑 sources
⊥
63
Width vs boundary expansion
Lemma [ChvatalSzemeredi] 𝒘𝒊𝒅𝒕𝒉(𝑭) ≥ 𝒆(𝑭).
≥ 𝒔(𝑭)
Start at ⊥
Go to parent with
more sources until
≤ 𝟐𝒔(𝑭)/𝟑 sources
⊥
64
Width vs boundary expansion
Lemma [ChvatalSzemeredi] 𝒘𝒊𝒅𝒕𝒉(𝑭) ≥ 𝒆(𝑭).
≥ 𝒔(𝑭)
Start at ⊥
Go to parent with
more sources until
≤ 𝟐𝒔(𝑭)/𝟑 sources
⊥
65
Width vs boundary expansion
Lemma [ChvatalSzemeredi] 𝒘𝒊𝒅𝒕𝒉(𝑭) ≥ 𝒆(𝑭).
𝒔(𝑭)/𝟑 to 𝟐𝒔(𝑭)/𝟑
𝑯
contains 𝛅𝑯
literals
𝛅𝑯 ≥ 𝒆(𝑭)
⊥
66
Boundary expansion implies large size
Corollary [Ben-SassonWigderson]:
𝑹𝒆𝒔 𝑭 ≥ 𝒆𝒙𝒑
𝒆 𝑭 –𝒘 𝑭 𝟐
𝟐𝒏
𝑹𝒆𝒔𝑡𝑟𝑒𝑒(𝑭) ≥ 𝟐𝒆(𝑭) – 𝒘(𝑭)
So…to prove strong 𝟐𝛀 𝒏 resolution lower
bound for 𝑭 with small clauses
prove that 𝒆(𝑭) is 𝛀 𝒏
67
PCR degree bounds imply size bounds
Thm [Clegg, Edmonds, Impagliazzo] : Every PCR proof of
𝑭 of size 𝑺 can be converted to one of degree
𝒅 ≤ 𝟐𝒏 ln 𝑺 + 𝒘(𝑭)
[CEI] showed this for polynomial calculus before
BenSasson-Wigderson’s argument for resolution.
Suffices to prove lower bounds on PCR degree to get
strong lower bounds.
68
Hard examples
69
Counting
Pigeonhole principle 𝑷𝑯𝑷𝒎
𝒏
No 1-1 function from 𝒎 to 𝒏 for 𝒎 > 𝒏
If 𝒎 = 𝒏 + 𝟏 write as 𝑷𝑯𝑷𝒏
70
Pigeonhole propositional formulas
Variables
Complete bipartite graph of
variables 𝑷𝒊𝒋 representing 𝒇(𝒊) = 𝒋
Clauses
𝒇 is total: (𝑷𝒊𝟏 𝑷𝒊𝟐 … 𝑷𝒊𝒏) for 𝒊 = 𝟏, … , 𝒎
𝒇 is 1-1: (𝑷𝒊𝒋 𝑷𝒌𝒋) for 𝟏 𝒊 𝒌 𝒎, 𝒋 = 𝟏, … , 𝒏
𝑷𝑯𝑷(𝑮): for graph 𝑮 contained in 𝒎 × [𝒏]
𝒇 is total: 𝒋: 𝒊,𝒋 ∈𝑬(𝑮) 𝑷𝒊𝒋 for 𝒊 ∈ 𝒎
𝒇 is 1-1: 𝑷𝒊𝒋 𝑷𝒌𝒋 for 𝒊, 𝒋 ≠ 𝒌, 𝒋 ∈ 𝑬(𝑮), 𝒋 ∈ 𝒏
71
Resolution and 𝑷𝑯𝑷𝒏
Thm [Haken 84, BP 96] Any resolution proof of 𝑷𝑯𝑷𝒏 requires size
at least 𝟐𝒏/𝟐𝟎 .
Follows from more direct methods
Can’t use width-size bound directly since 𝒘(𝑷𝑯𝑷𝒏) is large
(total clauses have size 𝒏).
Instead use constant-degree bipartite graph 𝑮 on 𝒏 + 𝟏 × 𝒏
Prove boundary expansion lower bound for 𝑷𝑯𝑷(𝑮)
≡ expansion lower bound on 𝑮
Implies 𝟐𝛀
𝒏
lower bound for 𝑷𝑯𝑷𝒏
72
More counting
Counting 𝒎𝒐𝒅 𝟐
𝑪𝒐𝒖𝒏𝒕𝟐𝒏+𝟏
𝟐
no matching on an odd size set
variable for each edge
Counting 𝒎𝒐𝒅 𝒓
no perfect 𝒓-partition if 𝒓 doesn’t divide 𝒏
𝑪𝒐𝒖𝒏𝒕𝒏𝒓 for 𝒏 ≢ 𝟎 (𝒎𝒐𝒅 𝒓)
73
Counting and algebraic/inequality proofs
Cutting-planes can count:
Efficient refutations of 𝑷𝑯𝑷𝒏 and 𝑪𝒐𝒖𝒏𝒕𝒏𝒓
Algebraic proofs over field of characteristic 𝒑 can only
count mod 𝒑
PCR proofs require 𝛀 𝒏 degree/exponential size
to refute 𝑷𝑯𝑷𝒏 and 𝑪𝒐𝒖𝒏𝒕𝒏𝒓 for most 𝒏 and 𝑟.
74
Tseitin formulas - odd-charged graphs
Given a low degree graph 𝑮(𝑽, 𝑬) with 0-1 charge
𝝌 𝒗 on each node 𝒗 s.t. total charge is odd
One variable x𝒆 per edge 𝒆𝑬
Clauses saying parity of edges touching 𝒗 is 𝝌 𝒗
Needs 𝟐𝒅𝒆𝒈 𝒗 −𝟏 clauses
If degree is large, add extension variables to compute
parity at each vertex
0
Unsatisfiable
1
1
0
Sum of degrees is even
0
1
75
Expander graphs
Defn: Let 𝑮 = (𝑽, 𝑬) be a graph. Let 𝑬(𝑺, 𝑺) ⊆ 𝑬 be those edges
with one endpoint in 𝑺 and one outside 𝑺.
Graph 𝑮 has expansion 𝜺 if |𝑬 𝑺, 𝑺 | ≥ 𝜺|𝑺| for all subsets 𝑺 of
size at most 𝒏/𝟐 vertices.
Fact: Random constant-degree regular graphs almost always
have constant expansion 𝜺 > 𝟎.
Explicit constructions also with many applications in complexity
Originally considered for regular resolution lower bounds [Galil]
76
Tseitin formulas using mod 2 reasoning
Assume wlog # vertices 𝒏 is odd and all charges are 𝟏.
Input parity equations are just
𝒗∈𝒆 𝒙𝒆 = 𝟏 mod 𝟐 for each 𝒗 ∈ 𝑽.
Can add equations mod 2 to get 𝟎 = 𝟏.
Adding equations for all vertices in 𝑺 ⊆ 𝑽 yields
𝒗∈𝑬(𝑺,𝑺) 𝒙𝒆 = |𝑺| mod 𝟐
77
Tseitin formulas and mod 2 equations
Adding equations for all vertices in 𝑺 ⊆ 𝑽
yields
𝒗∈𝑬(𝑺,𝑺) 𝒙𝒆 = |𝑺| mod 𝟐
78
Tseitin formulas and mod 2 equations
Adding equations for all vertices in 𝑺 ⊆ 𝑽
yields
𝒗∈𝑬(𝑺,𝑺) 𝒙𝒆 = |𝑺| mod 𝟐
79
Tseitin formulas and mod 2 equations
Adding equations for all vertices in 𝑺 ⊆ 𝑽
yields
𝒗∈𝑬(𝑺,𝑺) 𝒙𝒆 = |𝑺| mod 𝟐
80
Tseitin formulas and mod 2 equations
81
Tseitin formulas and mod 2 equations
Adding equations for all vertices in 𝑺 ⊆ 𝑽
yields
𝒗∈𝑬(𝑺,𝑺) 𝒙𝒆 = |𝑺| mod 𝟐
82
Tseitin formulas using mod 2 reasoning
Assume wlog # vertices 𝒏 is odd and all charges are 𝟏.
Input parity equations are just
𝒗∈𝒆 𝒙𝒆 = 𝟏 mod 𝟐 for each 𝒗 ∈ 𝑽.
Can add equations mod 2 to get 𝟎 = 𝟏.
Adding equations for all vertices in 𝑺 ⊆ 𝑽 yields
𝒗∈𝑬(𝑺,𝑺) 𝒙𝒆 = |𝑺| mod 𝟐
Graph 𝑮 with expansion 𝜺 needs equation with at
least 𝜺𝒏/𝟐 variables to derive a contradiction.
We relate this to degree in PCR…
83
Parity equations and PCR
Given equations of form
𝒙𝟏 + 𝒙𝟐 + 𝒙𝟑 = 𝟎 (mod 𝟐)
Represent in the “Fourier basis” over {𝟏, −𝟏}
Polynomial equation 𝒚𝒊𝟐 − 𝟏 = 𝟎 for each variable
𝒚𝒊 = 𝟏 − 𝟐𝒙𝒊 = (−𝟏)𝒙𝒊
Polynomial equation 𝒚𝟏 𝒚𝟐 𝒚𝟑 − 𝟏 = 𝟎
would be 𝒚𝟏 𝒚𝟐 𝒚𝟑 + 𝟏 = 𝟎 if RHS were 𝟏
Thm: Since transformation is linear and invertible it
preserves degrees of PCR proofs if field’s characteristic
is not 2.
84
Tseitin formulas in Fourier basis
Variables are in {𝟏, −𝟏}
𝒚𝟐𝒆 = 𝟏 for every 𝒆 ∈ 𝑬
Parity of edges touching 𝒗 equals charge at 𝒗
𝝌(𝒗) for every 𝒗 ∈ 𝑽
𝒚
=
(−𝟏)
𝒗∈𝒆 𝒆
Degree of input polynomials equals degree of graph
Thm: There is a constant-degree graph 𝑮 s.t. a Tseitin formula for
𝑮 with all charges 𝟏 and odd # of vertices requires
degree 𝛀(𝒏) to refute in Nullstellensatz [Grigoriev]
degree 𝛀(𝒏) to refute in PCR [BussGrigorievImpagliazzoPitassi]
85
Proof idea: binomial equations
Every inferred polynomial from 𝑻𝒔𝒆𝒊𝒕𝒊𝒏(𝑮) over Fourier
basis has two terms with ±𝟏 coefficients
Every monomial corresponds to a parity of a subset of edges
So … each equivalence corresponds to a parity equation:
e.g. 𝒙𝟏 + 𝒙𝟐 + 𝒙𝟑 = 𝟏 + 𝒙𝟐 + 𝒙𝟒 mod 𝟐
Equivalently… 𝒙𝟏 + 𝒙𝟑 + 𝒙𝟒 = 𝟏 mod 𝟐
# of variables is at most 𝟐 × degree of PCR proof.
Degree of the PCR proof of 𝑻𝒔𝒆𝒊𝒕𝒊𝒏(𝑮)
𝟏
≥ min # variables in mod 2 equations proof of 𝑻𝒔𝒆𝒊𝒕𝒊𝒏(𝑮)
≥
𝟐
𝟏
𝜺𝒏/𝟐 =
𝟐
𝜺𝒏/𝟒
86
Random k-CNF formulas
𝒎 = 𝒓𝒏 clauses chosen independently at random
𝒏
𝒌
from all 𝟐
clauses of size 𝒌.
𝒌
Exponentially hard for PCR
Idea [BenSasson-Impagliazzo]: If characteristic isn’t 2 then
lower bound holds even if each clause is strengthened to a
parity
[Alekhnovich-Razborov]: Same bound holds for
characteristic 2
87
DPLL on random 3-CNF
Proof complexity
shows 𝟐𝛀(𝒏 /𝒓) time
# of DPLL
backtracks
is required for
unsatisfiable formulas
for 𝒓 > 𝒓∗𝟑
[B,Karp,Saks,Pitassi 98]
[Ben-Sasson 02]
1
0
4.267
ratio of clauses to variables
𝒓
88
ZFC
Much more…
P/poly-Frege
Frege
TC0-Frege
AC0-Frege
Positivstellensatz
Cutting Planes
Lasserre/SOS
Res(k)
Polynomial Calculus/PCR
Resolution
Nullstellensatz
DPLL
Truth Tables
89
Thanks for listening!
Questions…?
90