Download presentation
Presentation is loading. Please wait.
Published byPiers Briggs Modified over 9 years ago
1
Inapproximability from different hardness assumptions Prahladh Harsha TIFR 2011 School on Approximability
2
Hardness of approximation Worst case hardness PCP theorem, Hardness of Label cover Unique Games Conjecture Average case hardness Feige’s Random-3-SAT assumption
3
Label Cover Label Cover (LC) G – Bipartite graph 1, 2 – labels (projection) constraint per edge c e: 1 2 Edge e is satisfied if c e (σ 1 )=σ 2 Goal: Find an assignment to vertices that satisfies the most edges Gap(α,β)-LC: Distinguish between instances At least α fraction of constraints satisfied At most β fraction of constraints satisfied 11 22
4
PCP Theorem Label Cover (LC) G – Bipartite graph 1, 2 – labels (projection) constraint per edge c e: 1 2 Edge e is satisfied if c e (σ 1 )=σ 2 PCP Theorem […., AS’92, ALMSS’92] Gap(1,0.9999)-LC is NP-hard 11 22
5
Stronger form of PCP Theorem Label Cover (LC) G – Bipartite graph 1, 2 – labels (projection) constraint per edge c e: 1 2 Edge e is satisfied if c e (σ 1 )=σ 2 PCP Theorem + Repetition Theorem [Raz’95] For every constant δ there exists alphabets 1, 2, Gap(1, δ )-LC is NP-hard (starting point for all tight hardness of approximation reductions) 11 22
6
Proving tight hardness results 11 22 Outer Verifier Dictatorship Test [Fourier Analysis] + (composition) Inner Verifier
7
Proving tight hardness results 11 22 + direction reduction SETCOVER Lattice probs. Fourier analysis MAX3LIN, MAX3SAT, CLIQUE Fourier analysis MAXCUT, VERTEX-COVER (with unique constraints) Unique Games Conjecture
8
Stronger form of PCP Theorem Label Cover (LC) G – Bipartite graph 1, 2 – labels (projection) constraint per edge c e: 1 2 Edge e is satisfied if c e (σ 1 )=σ 2 PCP Theorem + Repetition Theorem [Raz’95] For every constant δ there exists alphabets 1, 2, Gap(1, δ )-LC is NP-hard (mother of all tight hardness of approximation reductions) 11 22
9
Label Cover Constructions
10
Low Degree Test (LDT) [RS’92] Given function f:F m F (F – field), check if f is the evaluation of a low-degree polynomial without reading all of F f:F m F Use fact that restriction of low-degree polynomial to a line is still low-degree
11
Label Cover for LDT Points table f:F m F Lines table f lines Constraint: f lines (l)(x) = f(x) Large Alphabet Size
12
Label Cover -- LDT [AS’97, RS’97] Completeness: If f:F m F is a low-degree polynomial, there exists lines table f lines such that Pr[f lines (l)(x) = f(x)] = 1 Soundness: If f:F m F is “far” from being low-degree polynomial, there for all lines table f lines we have Pr[f lines (l)(x) = f(x)] ≤ δ
13
Label Cover for NP Encode problem in NP using polynomials to lift the label cover for LDT to all of NP Label Cover for NP [RS’97, AS’97]: For every alphabet and error δ =1/log|, Gap(1, δ )-LC is NP-hard, if | n polylog n Caveat: Large Alphabet Size Renders result “useless” for hardness results
14
Alphabet Reduction [MR’08, DH’09] Alphabet Reduction: Label Cover instance with large alphabet size Label Cover instance with small alphabet size Idea: Recurse!! [in the style of AS’92] Use an “Inner” Label Cover to reduce alphabet of outer label cover
15
Alphabet Reduction
17
Alphabet Size Reduced However 3-partite graph instead of bipartite Idea: [2-query composition DH’09] Combine leftmost and rightmost components by identifying nodes in left partition (combine all left-neighbours of a right vertex)
18
Label Cover for NP Performing alphabet reduction repeatedly: Label Cover for NP [MR’08, DH’09]: For every alphabet and error δ =1/log|, Gap(1, δ )-LC is NP-hard. Advantages: Sub-constant error achievable Nearly linear sized reduction
19
Label Cover variants Some hardness reductions require more structure of the label cover instance [KH’04] Hardness of Balanced homogenous linear equations MAXBISECTION Mixing property Bipartite graph is a good sampler Smoothness
20
Open Questions Sliding Scale Conjecture [BGLR’93] For every alphabet and error δ =1/poly|, Gap(1, δ )-LC is NP-hard (current results only obtain δ =1/log| Obtain polynomial sized mixing and smooth PCPs (current constructions require subexponential sized proofs)
21
Average Case Hardness
22
Assumptions Inapproximability results based on Worst case hardness assumptions (so far) Average case hardness assumptions Cryptographic assumptions Random 3SAT hardness (Feige)
23
Random 3SAT n variables x 1, x 2, …., x n m = Cn clauses (x i v x j v x k ) Chosen randomly and independently C – small, satisfiable w.h.p C – large, unsatisfiable w.h.p
24
Random 3SAT – large C For large C Typical – unsatisfiable In fact at most (7/8 + δ ) clauses satisfiable Rare – satisfiable Rare even for (1- δ ) clauses to be satisfied
25
Random 3SAT – large C Proofs of unsatisfiability (coNP proof) When C > √n, can find short of proof of unsatisfiability w.h.p. For large constant < C < √n, though unsatisfiable, current techniques do not prove unsatisfiability
26
Feige’s Random 3SAT hypothesis For all 0< δ <1/8, there exists a large constant C, and there does NOT exist a polynomial time algorithm that INPUT: Random 3CNF formula with n variables and Cn clauses OUTPUT: “typical” or “rare” on most inputs (> 50%) output typical but never outputs “typical” on a rare instance (i.e, (1- δ ) satisfiable )
27
MAX3SAT approximability Feige’s Random 3SAT hypothesis MAX3SAT is inapproximable (within polynomial time) to a factor better than (7/8 + δ ), for all δ >0.
28
MAX3AND INPUT Boolean formula on n variables x 1, x 2, …., x n m ANDs of 3 literals (x i ∧ x j ∧ x k ) OUTPUT Assignment OBJECTIVE maximizes number of ANDs being satisfied
29
MAX3AND - approximability Feige’s Random 3SAT hypothesis Not possible to approximate better than a factor (1/2 + δ ) In particular, can’t distinguish between MAX3AND instances > (1/4 - δ ) satisfiable < (1/8 + δ ) satisfiable (even for random MAX3AND instances)
30
Why MAX3AND? Gives inapproximability results MAX-COMPLETE BIPARITITE GRAPH MINBISECTION DENSE k-SUBGRAPH 2-CATALOG SEGMENTATION (not approximable beyond a particular constant) Previously, no known inapproximability results
31
Max Complete Bipartite Graph INPUT n x n bipartite graph OUTPUT k x k complete bipartite subgraph OBJECTIVE Maximize k Feige’s Hypothesis implies can’t approximate better than (1/2 + δ ) (reduction from hardness of randomMAX3AND)
32
Reduction from MAX3AND Given a random MAX3AND with m ANDs construct bipartite graph Vertices on each side m ANDS Edges If two ANDS can be satisfied simulatenously
33
Reduction from MAX3AND (contd) MAX3AND instance – (1/4 - δ ) satsifiable Corresponding vertices from a kxk complete bipartite graph with k = (1/4 - δ )m random MAX3AND instance – (1/8+ δ ) satisfiable Easy to check that for a random MAX3AND instance, whp every (1/8 + δ )m ANDs involve at least (n+1) literals Any kxk bipartite graph with k > (1/8+ δ )m involves a variable and its negation and hence not complete
34
Hardness of random MAX3AND Algorithm for random 3SAT (refuting Feige’s hypothesis) Idea: View input random 3CNF formula as a 3AND formula and use algorithm for random MAX3AND 3CNF - (7/8 + δ )-satisfiable 3AND formula – (1/8 + δ )-satisfiable 3CNF – (1- δ ) satisfiable 3AND formula – (1/4 - δ )-satisfiable Not exactly true, instead will use simple checks and SDPs to detect non-typical behavior
35
Hardness of MAX3AND (contd) Input: random 3SAT instance Algorithm 1. If any literal does not occur (3C/2± δ ) times, output “rare” 2. Construct graphs G 12, G 23, G 31 whose vertices are all 2n literals and edges as follows: G 12 :(x i, x j ) is there is a clause of the form (x i, v x j v x k ) 3. Run MAXCUT SDP on all 3 graphs, if SDP outputs larger than (1/2+ δ )m, output “rare” 4. Run MAX3AND algorithm on instance If output > (1/4- δ ), output “rare” If output < (1/8+ δ ), output “typical”
36
Hardness of MAX3AND (contd) Typical Instances ( < (7/8+ δ )-satisfiable) Easy to check algorithm outputs “typical” on most typical instances Rare instance (> (1- δ )-satisfiable) Literal-occurrence and SDP checks ensure that when viewed as a NOT-ALL-EQUAL-SAT instance, no assignment satisfies > (3/4+ δ ) clauses Hence, at least (1/4- δ )-clauses are satisfied as ANDs for which the algorithm outputs “rare”
37
Random 3SAT assumption Feige’s hypothesis MAX3AND inapproximable to better than ½ (even on random instances) Inapproximability results COMPLETE-BIPARTITE-GRAPH, MINBISECTION, DENSEST k-SUBGRAPH, ….
38
Other assumptions? Maximal number of equations satisfiable in a random linear system [Ale’03] Implies Feige’s hypothesis Inapproximability of nearest-codeword- problem to within n 1-δ Hard to distinguish low-rigidity matrices and random matrices
39
Quasirandom PCPs [Kho’04] Suffices to having following quasi- randomess of 3SAT For any set of half of the variables, (1/8±δ)- fraction of clauses have all 3 variables from this set Khot constructed PCPs with this quasirandom property leading to inapproximability results for earlier problems (based on worst case hardness)
40
Quasirandom PCPs PCPs which exhibit very different query behaviour on YES and NO instance PCP verifier makes d queries NO instances: For any set of half the proof locations, the probability that all the d queries are in the set ≈ 2 -d YES instances: There is a set of half the proof location, which the verifier queries more frequently ( > 2 -(d-1) )
41
Open Problems (Dis)prove Feige’s hypothesis Connections between average complexity and approximation complexity
42
THANK YOU
Similar presentations
© 2025 SlidePlayer.com Inc.
All rights reserved.