Computational Complexity Theory

Lecture 4: NP-complete problems,

Search versus Decision

Indian Institute of Science

Recap: Cook-Levin Theorem
 Definition. A boolean formula is in Conjunctive Normal
Form (CNF) if it is an AND of OR of literals.
e.g. ϕ = (x1 ∨ x2) ∧ (x3 ∨ ¬x2 )

 Definition. Let SAT be the language consisting of all

satisfiable CNF formulae.

 Theorem. (Cook-Levin) SAT is NP-complete.

Recap: Cook-Levin Theorem
 Let L ∈ NP. We intend to come up with a polynomial
time computable function f: x ϕx s.t.,
x∈L ϕx ∈ SAT

 Language L has a poly-time verifier M such that

x∈L ∃u ∈{0,1}p(|x|) s.t. M(x, u) = 1
 Recap: Cook-Levin Theorem
Output of ψ

T(n) qaccept o/p 1 …. ….

.
.
.
2 qstart u1 0 …. ….

1 qstart u1 1 …. a cell ….
….
Input u-variables of ψ Observe: ψ(u) = 1 iff N(u) = 1
Recap: Cook-Levin Theorem
 Let L ∈ NP. We intend to come up with a polynomial
time computable function f: x ϕx s.t.,
x∈L ϕx ∈ SAT

 Language L has a poly-time verifier M such that

x∈L ∃u ∈{0,1}p(|x|) s.t. ψx(u) = 1

ψx is a poly(|x|)-size circuit
Recap: Cook-Levin Theorem
 Let L ∈ NP. We intend to come up with a polynomial
time computable function f: x ϕx s.t.,
x∈L ϕx ∈ SAT

 Language L has a poly-time verifier M such that

x∈L ψx(u) is satisfiable

 Important note: A satisfying assignment u for ψx

trivially gives a certificate u such that M(x, u) = 1.
Recap: Cook-Levin Theorem
 Let L ∈ NP. We intend to come up with a polynomial
time computable function f: x ϕx s.t.,
x∈L ϕx ∈ SAT

 Language L has a poly-time verifier M such that

x∈L ψx(u) is satisfiable

a poly-size circuit but not a poly-size CNF

Recap: Cook-Levin Theorem
 From circuit to CNF. From circuit ψ construct a CNF
ϕ by introducing some extra variables v such that

ψ(u) = 1 iff ϕ(u, v) is satisfiable.

 Language L has a poly-time verifier M such that

x∈L ∃u ∈{0,1}p(|x|) s.t. ϕx(u, v) is satisfiable
Recap: Cook-Levin Theorem
 From circuit to CNF. From circuit ψ construct a CNF
ϕ by introducing some extra variables v such that

ψ(u) = 1 iff ϕ(u, v) is satisfiable.

 Language L has a poly-time verifier M such that

x∈L ϕx(u, v) is satisfiable

 Important note: A satisfying assignment (u, v) for ϕx

trivially gives a certificate u such that M(x, u) = 1.
Recap: Cook-Levin Theorem
 Definition. A CNF is a called a kCNF if every clause has
at most k literals.
e.g. a 2CNF ϕ = (x1 ∨ x2) ∧ (x3 ∨ ¬x2 )

 Definition. kSAT is the language consisting of all

satisfiable kCNFs.

 Cook-Levin. There’s some constant k such that kSAT is

NP-complete.
Recap: 3SAT is NP-complete
 Definition. A CNF is a called a kCNF if every clause has
at most k literals.
e.g. a 2CNF ϕ = (x1 ∨ x2) ∧ (x3 ∨ ¬x2 )

 Definition. kSAT is the language consisting of all

satisfiable kCNFs.

 Theorem. 3SAT is NP-complete.

Proof sketch: (x1 ∨ x2 ∨ x3 ∨ ¬x4 ) is satisfiable iff (x1
∨ x2 ∨ z) ∧ ( x3 ∨ ¬x4 ∨ ¬z) is satisfiable.
More NP-complete problems
NP-complete problems: Examples
 Independent Set

 Clique And many many other

natural problems!
 Vertex Cover

 0/1 Integer Programming

 Max-Cut (NP-hard)
Example 1: Independent Set
 INDSET := {(G, k): G has independent set of size k}

 Goal: Design a poly-time reduction f s.t.

x ∈ 3SAT f(x) ∈ INDSET

 Reduction from 3SAT: Recall, a reduction is just an

efficient algorithm that takes input a 3CNF ϕ and
outputs a (G, k) tuple s.t
ϕ ∈ 3SAT (G, k) ∈ INDSET
Example 1: Independent Set
 Reduction: Let ϕ be a 3CNF with m clauses and n
variables. Assume, every clause has exactly 3 literals.
Example 1: Independent Set
 Reduction: Let ϕ be a 3CNF with m clauses and n
variables. Assume, every clause has exactly 3 literals.
A vertex stands for a partial
assignment of the variables in
Ci that satisfies the clause

For every clause Ci form a complete

graph (cluster) on 7 vertices
Example 1: Independent Set
 Reduction: Let ϕ be a 3CNF with m clauses and n
variables. Assume, every clause has exactly 3 literals.

Ci

Add an edge between two

vertices in two different clusters if
C1 the partial assignments they stand Cm
for are incompatible.
Example 1: Independent Set
 Reduction: Let ϕ be a 3CNF with m clauses and n
variables. Assume, every clause has exactly 3 literals.

Ci

C1 Cm
Graph G on 7m vertices
Example 1: Independent Set
 Reduction: Let ϕ be a 3CNF with m clauses and n
variables. Assume, every clause has exactly 3 literals.

Ci

C1 Cm

 Obs: ϕ is satisfiable iff G has an ind set of size m.

Example 2: Clique
 CLIQUE := {(H, k): H has a clique of size k}

 Goal: Design a poly-time reduction f s.t.

x ∈ INDSET f(x) ∈ CLIQUE

 Reduction from INDSET: The reduction algorithm

computes G from G
(G, k) ∈ INDSET (G, k) ∈ CLIQUE
Example 3: Vertex Cover
 VCover := {(H, k): H has a vertex cover of size k}

 Goal: Design a poly-time reduction f s.t.

x ∈ INDSET f(x) ∈ VCover

 Reduction from INDSET: Let n be the number of

vertices in G. The reduction algorithm maps (G, k) to
(G, n-k).
(G, k) ∈ INDSET (G, n-k) ∈ VCover
Example 4: 0/1 Integer Programming
 0/1 IProg := Set of satisfiable 0/1 integer programs
 A 0/1 integer program is a set of linear inequalities
with rational coefficients and the variables are
allowed to take only 0/1 values.

 Reduction from 3SAT: A clause is mapped to a linear

inequality as follows

x1 ∨ x2 ∨ x3 x1 + (1- x2) + x3 ≥ 1
Example 5: Max Cut
 MaxCut : Given a graph find a cut with the max size.
 A cut of G = (V, E) is a tuple (U, V\U), U ⊆ V. Size of a
cut (U, V\U) is the number of edges from U to V\U.

 MinVCover: Given H, find a Vcover with the min size.

 Obs: From MinVCover(H), we can readily check if (H,

k) ∈ VCover, for any k.
Example 5: Max Cut
 MaxCut : Given a graph find a cut with the max size.
 A cut of G = (V, E) is a tuple (U, V\U), U ⊆ V. Size of a
cut (U, V\U) is the number of edges from U to V\U.

 Goal: A poly-time reduction from VCover to

MaxCut. f
(H, k) G s.t.

Size of a MaxCut(G) = 2.|E(H)| - |MinVCover(H)|

Example 5: Max Cut
 MaxCut : Given a graph find a cut with the max size.
 A cut of G = (V, E) is a tuple (U, V\U), U ⊆ V. Size of a
cut (U, V\U) is the number of edges from U to V\U.

 Goal: A poly-time reduction from VCover to

MaxCut. f
(H, k) G s.t.

Thus, checking if (H, k) ∈ VCover reduces to finding

MaxCut(G).
Example 5: Max Cut
f
 The reduction: (H, k) G

w
u u
degH(u) – 1 edges
between u and w

H G

 G is formed by adding a new vertex w and adding

degH(u) – 1 edges between every u ∈ V(H) and w.
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Let SG(U) = no. of edges in G with exactly one end

vertex incident on a vertex in U.
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Let SG(U) = no. of edges going out of U in G.

Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Let SG(U) = size of the cut (U,V\U + w).

Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Let SH(U) = no. of edges in H with exactly one end

vertex incident on a vertex in U.
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Then SG(U) = SH(U) + Σ (degH(u) – 1)

u∈U

= SH(U) + ΣdegH(u) – |U|

u∈U
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Then SG(U) = SH(U) + Σ (degH(u) – 1)

u∈U

= SH(U) + ΣdegH(u) – |U|

u∈U
Obs: Twice the number of
edges in H with at least one
end vertex in U.
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Then SG(U) = SH(U) + Σ (degH(u) – 1)

u∈U

= SH(U) + ΣdegH(u) – |U|

u∈U

= 2.|EU(H)| - |U|
Example 5: Max Cut
 Claim: |MaxCut(G)| = 2.|E(H)| - |MinVCover(H)|
 Proof: Let V(H) = V. Then V(G) = V + w.
Suppose (U, V\U + w) is a cut in G.

 Then SG(U) = 2.|EU(H)| - |U| … Eqn (1)

 Proposition: If (U, V\U + w) is a max cut in G then U

is a vertex cover in H.
…proof of the claim follows from the above proposition
Example 5: Max Cut
 Proof of the Proposition: Suppose U is not a vertex
cover
degH(u)-1 edges
V\U + w
u
w
U v
Example 5: Max Cut
 Proof of the Proposition: Suppose U is not a vertex
cover
V\U + w
w
U v

Gain: degH(u)-1 + 1 edges

Loss: At most degH(u)-1 edges, these are the edges going from U to u
Net gain: At least 1 edge. Hence the cut is not a max cut.
Search versus Decision
Search version of NP problems
 Recall: A language L ⊆ {0,1}* is in NP if
➢ There’s a poly-time verifier M such that
➢ x∈L iff there’s a poly-size certificate u s.t M(x,u) = 1

• Search version of L: Given an input x ∈ {0,1}*, find a u

∈{0,1}p(|x|) such that M(x,u) = 1, if such a u exists.
Search version of NP problems
 Recall: A language L ⊆ {0,1}* is in NP if
➢ There’s a poly-time verifier M such that
➢ x∈L iff there’s a poly-size certificate u s.t M(x,u) = 1

• Search version of L: Given an input x ∈ {0,1}*, find a u

∈{0,1}p(|x|) such that M(x,u) = 1, if such a u exists.

• Example: Given a 3CNF ϕ, find a satisfying assignment

for ϕ if such an assignment exists.
Decision versus Search
 Is the search version of an NP-problem more difficult
than the corresponding decision version?
Decision versus Search
 Is the search version of an NP-problem more difficult
than the corresponding decision version?

 Theorem. Let L ⊆ {0,1}* be NP-complete. Then, the

search version of L can be solved in poly-time if and
only if the decision version can be solved in poly-time.
Decision versus Search
 Is the search version of an NP-problem more difficult
than the corresponding decision version?

 Theorem. Let L ⊆ {0,1}* be NP-complete. Then, the

search version of L can be solved in poly-time if and
only if the decision version can be solved in poly-time.

 Proof. (search decision) Obvious.

Decision versus Search
 Is the search version of an NP-problem more difficult
than the corresponding decision version?

 Theorem. Let L ⊆ {0,1}* be NP-complete. Then, the

search version of L can be solved in poly-time if and
only if the decision version can be solved in poly-time.

 Proof. (decision search) We’ll prove this for

L = SAT first.
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn)
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

ϕ(0,…,xn)
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn)
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn)

SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

ϕ(1,0,…,xn)
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

A( ϕ(1,0,..) ) = Y ϕ(1,0,…,xn)
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

A( ϕ(1,0,..) ) = Y ϕ(1,0,…,xn)

ϕ(1,0,0,…,xn)
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

A( ϕ(1,0,..) ) = Y ϕ(1,0,…,xn)

A( ϕ(1,0,0...) ) = N ϕ(1,0,0,…,xn)
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

A( ϕ(1,0,..) ) = Y ϕ(1,0,…,xn)

A( ϕ(1,0,0...) ) = N ϕ(1,0,0,…,xn) ϕ(1,0,1,…,xn)

SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

A( ϕ(1,0,..) ) = Y ϕ(1,0,…,xn)

A( ϕ(1,0,0...) ) = N ϕ(1,0,0,…,xn) ϕ(1,0,1,…,xn) A( ϕ(1,0,0...) ) = Y

SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

A( ϕ(1,0,..) ) = Y ϕ(1,0,…,xn)

A( ϕ(1,0,0...) ) = N ϕ(1,0,0,…,xn) ϕ(1,0,1,…,xn) A( ϕ(1,0,0...) ) = Y

.
.
.
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

ϕ(x1,…,xn) A(ϕ) = Y

A( ϕ(0,..) ) = N ϕ(0,…,xn) ϕ(1,…,xn) A( ϕ(1,..) ) = Y

A( ϕ(1,0,..) ) = Y ϕ(1,0,…,xn)

A( ϕ(1,0,0...) ) = N ϕ(1,0,0,…,xn) ϕ(1,0,1,…,xn) A( ϕ(1,0,0...) ) = Y

.
.
.
SAT is downward self-reducible
 Proof. (decision search) Let L = SAT, and A be a
poly-time algorithm to decide if ϕ(x1,…,xn) is satisfiable.

 We can find a satisfying assignment of ϕ with at most 2n

calls to A.
Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.
Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.

SAT ≤p L L ≤p SAT
Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.

SAT ≤p L L ≤p SAT

x ϕx
Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.

SAT ≤p L L ≤p SAT

x ϕx

From Cook-Levin theorem, we

can find a certificate of x from
a satisfying assignment of ϕx.
Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.

SAT ≤p L L ≤p SAT

x ϕx

How to find a certificate of ϕx using algorithm B ?

Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.

SAT ≤p L L ≤p SAT

x ϕx

How to find a certificate of ϕx using algorithm B ?

… we know how to using A, a poly-time decider for SAT

Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.

SAT ≤p L L ≤p SAT
ϕ f(ϕ) x ϕx

How to find a certificate of ϕx using algorithm B ?

… we know how to using A, a poly-time decider for SAT

Decision ≡ Search for NPC problems
 Proof. (decision search) Let L be NP-complete, and
B be a poly-time algorithm to decide if x∈L.

SAT ≤p L L ≤p SAT
ϕ f(ϕ) x ϕx

How to find a certificate of ϕx using algorithm B ?

Take A(ϕ) = B( f(ϕ) )

Decision versus Search
 Is search equivalent to decision for every NP problem?
Decision versus Search
 Is search equivalent to decision for every NP problem?

Probably not!
Decision versus Search
 Is search equivalent to decision for every NP problem?

n
 Let EE = ∪ DTIME (2c.2 ) and
c≥0
Doubly exponential
n analogues of P and NP
NEE = ∪ NTIME (2 )
c.2
c≥0
Decision versus Search
 Is search equivalent to decision for every NP problem?

 Theorem. (Bellare-Goldwasser) If EE ≠ NEE then there’s a

language in NP for which search does not reduce to
decision.