Presentation is loading. Please wait.

Presentation is loading. Please wait.

“(More) Consequences of Falsifying SETH

Similar presentations


Presentation on theme: "“(More) Consequences of Falsifying SETH"— Presentation transcript:

1 “(More) Consequences of Falsifying SETH
and the Orthogonal Vectors Conjecture” or “A Sample of Fine-grained Complexity” Jesper Nederlof joint work with Karl Bringmann, Amir Abboud, Holger Dell (Slides partly by Karl Bringmann and Holger Dell) Dauer: …

2 Classic Complexity Theory
Polynomial time Not in polynomial time unless P=NP Classic Complexity Theory Integer Linear Programming Satisfiability Maximum Matching Travelling Salesperson Longest common subsequence Clique Linear Programming

3 Fine-grained Complexity
Multivariate Complexity k O(nk) Fixed-parameter tractable (FPT) O(2kn) Complexity in P n! O*(2√n) O*(1.34n) O*(2n) 2O(n) Exponential Time Algorithms O(n) O(n2) ... n 𝑂 ∗ (⋅) suppresses factors poly in input size

4 Fine-Grained Complexity
OV, 𝑛 2 APSP, 𝑛 3 3SUM, 𝑛 2 SAT, 2 𝑛 SETH

5 Fine-Grained Complexity
Problem 𝒅-SAT: Given formula in 𝑑-CNF with 𝑛 variables, is it satisfiable? 𝑥 1 ∨¬ 𝑥 2 ∨ 𝑥 3 ∧ 𝑥 2 ∨¬ 𝑥 3 ∨ 𝑥 5 ∧(¬ 𝑥 1 ∨ 𝑥 4 ∨ 𝑥 5 ) Strong Exponential Time Hypothesis: SAT, 2 𝑛 [IP’01] SETH ∀𝜀>0 ∃𝑑: 𝑑-SAT has no 𝑂 ∗ ( 2 1−𝜀 𝑛 )-time algorithm Example Implications: [CDLMNOPSW‘16] Hitting Set, Set Splitting, NAE-SAT: 𝑂 ∗ ( 2 𝑛 ) but not 𝑂 ∗ ( (2−𝜀) 𝑛 ) Independent Set: 𝑂 ∗ ( 2 𝑡𝑤 ) but not 𝑂 ∗ ( (2−𝜀) 𝑡𝑤 ) [LMS‘11] Subset Sum: 𝑂 (𝑛+𝑡) but not 𝑡 1−𝜀 2 𝑜(𝑛) [B’17, ABHS’18+]

6 Fine-Grained Complexity
OV, 𝑛 2 Problem Orthogonal Vectors: Given sets 𝐴,𝐵⊆ {0,1} 𝑑 of size 𝑛, are any 𝑎∈𝐴,𝑏∈𝐵 orthogonal? ( 𝑖 𝑎 𝑖 ⋅ 𝑏 𝑖 =0) OV-Hypothesis: (moderate dimension) SAT, 2 𝑛 SETH ∀𝜀,𝛿>0: OV in 𝑑= 𝑛 𝛿 has no 𝑂( 𝑛 2−𝜀 )-time algorithm Example Implications: [BI’15,ABVW’15,BK’15,VWR’13,B’14,BI’16,BGL’17] No 𝑂( 𝑛 2−𝜀 ) algorithm for: Edit Distance, LCS, Diameter-2, Frechet distance, RegExp Matching, ... No 𝑂(𝑚⋅ 𝑛 2−𝜀 ) algorithm for All Pairs Maxflow [KT’17] Dynamic graph algorithms, ...

7 Fine-Grained Complexity
OV, 𝑛 2 APSP, 𝑛 3 Problem All Pairs Shortest Paths: Given graph 𝐺 with distance (weight) per edge compute distance between any two vertices SAT, 2 𝑛 SETH APSP-Hypothesis: ∀𝜀>0: APSP has no 𝑂( 𝑛 3−𝜀 ) algorithm Example Implications: [VWW’10,AGVW’15] No 𝑂( 𝑛 3−𝜀 ) algorithm for: Negative Triangle, Shortest Cycle, Radius, ...

8 Fine-Grained Complexity
OV, 𝑛 2 APSP, 𝑛 3 3SUM, 𝑛 2 Problem 3SUM: Given set 𝑍 of 𝑛 integers, are there 𝑎,𝑏,𝑐∈𝑍 with 𝑎+𝑏=𝑐? SAT, 2 𝑛 SETH 3SUM-Hypothesis: ∀𝜀>0: 3SUM has no 𝑂( 𝑛 2−𝜀 ) algorithm Example Implications: [GO’95,P’10,AVWW’14] No 𝑂( 𝑛 2−𝜀 ) algorithm for: Colinearity, Conv3SUM, ... No 𝑂( 𝑛 3−𝜀 ) algorithm for ExactWeightTriangle [P’10,VW’09] Dynamic Problems, Listing Triangles, ...

9 Fine-Grained Complexity
OV, 𝑛 2 APSP, 𝑛 3 3SUM, 𝑛 2 Success of fine-grained complexity: SAT, 2 𝑛 These hypotheses explain why we cannot improve running times of fastest algorithms for many problems SETH Yield reasons to stop searching for faster algorithms Non-matching lower bounds suggest faster algorithms → same as for NP-hardness, but not as elegant….. Why are these hypotheses worth introducing, besides mentioned implications?

10 Reason 1: Decades of Effort
OV, 𝑛 2 Many people tried Fastest known algorithms for 𝑑-SAT: 𝑂( 2 1−𝑑/𝑘 𝑛 ) for some 𝑐>0 [PPSZ’98] SAT, 2 𝑛 SETH Not for lack of trying: [MS’85,S’92,K’99,PPZ’97,S’99,DGHKKPRS’02, HSSW’02,BS’03,MS’11,CSTT’13,H’14,…] OV is at least as hard as 𝑑-SAT

11 Reason 2: Restricted Algorithms
OV, 𝑛 2 Hypotheses hold for algorithms of some specific form 𝑎∨𝑏 𝑐∨¬𝑏 𝑎∨𝑐 SAT, 2 𝑛 Standard framework for refuting a 𝑑-SAT instance: Resolution SETH SETH holds for regular resolution and tree-like resolution [BI’13,BT’16] A weak version of SETH holds for `witness compression’ algorithms [PP’10,D’13] OVH holds for formulas and branching programs [KW’18+]

12 Reason 3: Circuit Lower Bounds
OV, 𝑛 2 Falsifying these hypotheses is difficult, since it would imply difficult-to-prove statements If SETH fails then: [W’13,JMV’15] SAT, 2 𝑛 SETH NEXP is not contained in linear-size VSP-circuits is very likely, but seems difficult to prove

13 Reason 4: Implications of falsifying hypotheses
OV, 𝑛 2 Falsifying these hypotheses implies other `amazing’ algorithms If SETH fails then there are 𝑂 ∗ 2−𝜀 𝑛 -time algorithms for: Hitting Set, Set Splitting, NAE-SAT, SAT on restricted circuit classes… SAT, 2 𝑛 [CDLMNOPSW‘16] SETH add TC 0 If OVH fails then: [GIKW‘17] First-order model checking with 𝑘≥3 quantifiers is in time 𝑂( 𝑚 𝑘−1−𝜀 ) on structures of total size 𝑚. e.g. ∃𝑢∃𝑣∃𝑤:𝐸 𝑢,𝑣 ∧𝐸 𝑣,𝑤 ∧𝐸(𝑢,𝑤) add a weighted, dense problem „does 𝐺 contain a triangle?“ ∀𝑢∀𝑣∃𝑤:𝐸 𝑢,𝑤 ∧𝐸 𝑤,𝑣 „does 𝐺 have diameter=2?“

14 Our Contributions ⟹ OV, 𝑛 2
[ABDN‘18] OV, 𝑛 2 We find more consequences of falsifying SETH/OVH If SETH fails then: SAT, 2 𝑛 there are 𝑂 ∗ 2−𝜀 𝑛 -time algorithms for sparse- TC 0 -SAT SETH If OVH fails then: there are 𝑂 ∗ 2−𝜀 𝑛 -time algorithms for sparse- TC 1 -SAT there are 𝑂 𝑛 1−𝜀 𝑘 -time algorithms for weighted 𝑘-Clique (even in hypergraphs)

15 Circuit-SAT

16 Circuit-SAT ⟺ ⟺ ⟹ ∧ 𝒞 = class of circuits 𝐶 𝒞-SAT:
Given a circuit 𝐶∈𝒞 on 𝑛 variables, is 𝐶 satisfiable? 𝜔(𝒞) = inf{ 𝑤 | 𝒞-SAT is in time 𝑂 ∗ ( 2 𝑤⋅𝑛 ) } SETH: lim 𝑑→∞ 𝜔 𝑑−CNF =1 𝑥 1 𝑥 2 𝑥 𝑛 sparsification lemma lim 𝑑,𝑐→∞ 𝜔 𝑐−sparse 𝑑−CNF =1 [IPZ’01] lim 𝑐,𝛿→∞ 𝜔 𝑐−sparse depth−d ∨,∧,¬,𝑇𝐻𝑅 −circuit =1 𝑐−sparse depth−δ ∨,∧,¬,𝑇𝐻𝑅 −circuit ..sparse TC 0 [ABDN‘18] OVH fails: lim 𝑐,𝛿→∞ 𝜔 𝑐−sparse depth− 𝛿 log 𝑛 ∨,∧,¬,𝑇𝐻𝑅 −circuit <1 [ABDN‘18] .. sparse TC 1 Proof techniques: Refined Cook-Levin; Depth reduction by Valiant

17 Weighted k-Clique

18 Weighted k-Clique ⟹ ⟹ OV, 𝑛 2 APSP, 𝑛 3 SAT, 2 𝑛 Neg-𝑘-Clique, 𝑛 𝑘
1 -2 4 -1 3 -2 [VWW’10] 2 3 1 -1 SAT, 2 𝑛 Neg-𝑘-Clique, 𝑛 𝑘 SETH -5

19 Weighted k-Clique ⟹ ⟹ OV, 𝑛 2 APSP, 𝑛 3 SAT, 2 𝑛 Neg-𝑘-Clique, 𝑛 𝑘
1 -2 4 -1 3 -2 [VWW’10] 2 3 1 -1 SAT, 2 𝑛 Neg-𝑘-Clique, 𝑛 𝑘 SETH -5 Problem Negative-𝒌-Clique: Given edge-weighted graph 𝐺 is there a k-Clique with negative total edge-weight? Neg-𝒌-Clique-Hypothesis: ∀𝜀>0,𝑘≥3: Neg-𝑘-Clique has no 𝑂( 𝑛 𝑘−𝜀 ) algorithm

20 ⟹ Weighted k-Clique ⟹ ⟹ OV, 𝑛 2 APSP, 𝑛 3 SAT, 2 𝑛 Neg-𝑘-Clique, 𝑛 𝑘
1 -2 4 -1 3 -2 [VWW’10] 2 [ABDN‘18] 3 1 -1 SAT, 2 𝑛 Neg-𝑘-Clique, 𝑛 𝑘 SETH -5 relates two fine-grained hypotheses falsifying OV implies amazing algorithm for Neg-𝑘-Clique Proof technique: chain of tight reductions Neg-𝑘-Clique -> Exact-𝑘-Clique-> Clique in Hypergraphs -> OV

21 Concluding Remarks Fine-grained complexity is a currently very popular subject Hypotheses are very productive (often lead to faster algorithms if no connection to hypotheses can be made) Two of these hypotheses are SETH ( 𝑂 ∗ (2 𝑛 ) is best for 𝑑-CNF-SAT) and OVH ( 𝑂 𝑛 2 poly 𝑑 is best for OV) This work*: If SETH fails then: there are 𝑂 ∗ 2−𝜀 𝑛 -time algorithms for sparse- TC 0 -SAT If OVH fails then: there are 𝑂 ∗ 2−𝜀 𝑛 -time algorithms for sparse- TC 1 -SAT there are 𝑂 ∗ 𝑛 1−𝜀 𝑘 -time algorithms for weighted 𝑘-Clique *Amir Abboud and Karl Bringmann and Holger Dell and JN: More Consequences of Falsifying SETH and the Orthogonal Vectors Conjecture. In the conference proceedings of STOC 2018.


Download ppt "“(More) Consequences of Falsifying SETH"

Similar presentations


Ads by Google