Approximate Constraint Satisfaction requires Large LP Relaxations David Steurer Cornell Siu On Chan James R. Lee Prasad Raghavendra MSR Washington Berkeley TCS+ Seminar, December 2013
Max Cut, Traveling Salesman, best-known (approximation) algorithms for Sparsest Cut, Steiner Tree, โฆ many combinatorial optimization problems: common core = linear / semidefinite programming (LP/SDP) LP / SDP relaxations particular kind of reduction from hard problem to LP/SDP running time: polynomial in size of relaxation what guarantees are possible for approximation and running time?
๐ฆ ๐ = 1 ๐ฆ ๐ = โ1 example: basic LP relaxation for Max Cut Max Cut: Given a graph, find bipartition ๐ฆ โ ยฑ1 ๐ that cuts as many edges as possible 1 intended solution ๐น ๐ ๐๐ maximize ๐๐โ๐น ๐ ๐ฆ ๐๐ = 1, if ๐ฆ ๐ โ ๐ฆ ๐ , otherwise. ๐ ๐๐ โ ๐ ๐๐ โ ๐ ๐๐ โค 0 subject to 0, ๐ ๐๐ + ๐ ๐๐ + ๐ ๐๐ โค 2 integer linear program ๐ ๐๐ โ 0,1 ๐ ๐๐ โ 0,1 (relax integrality constraint) ๐ ๐ 3 inequalities approximation guarantee depend only on instances size optimal value of instance (but not instance itself) vs. optimal value of LP relaxation
challenges many possible relaxations for same problem small difference syntactically ๏ big difference for guarantees goal: identify โrightโ polynomial -size relaxation hierarchies = systematic ways to generate relaxations best-known: Sherali-Adams (LP), sum-of-squares/Lasserre (SDP); best possible? goal: compare hierarchies and general LP relaxations often : more complicated/larger relaxations ๏ better approximation P โ NP predicts limits of this approach; can we confirm them? goal: understand computational power of relaxations Rule out that poly-size LP relaxations show ๐ = ๐๐ ?
hierarchies [Lovรกsz โ Schrijver, Sherali โ Adams, Parrilo / Lasserre] great variety (sometimes different ways to apply same hierarchy) current champions : Sherali โ Adams (LP) & sum-of-squares / Lasserre (SDP) connections to proof complexity (Nullstellensatz and Positivstellensatz refutations) lower bounds [Mathieu โ Fernandez de la Vega Charikar โ Makarychev โ Makarychev] Sherali-Adams requires size 2 ๐ ฮฉ 1 to beat ratio ยฝ for Max Cut [Grigoriev, Schoenebeck] sum-of-squares requires size 2 ฮฉ ๐ to beat ratio 7 8 for Max 3-Sat upper bounds [Goemans-Williamson, implicit: many algorithms (e.g., Max Cut and Sparsest Cut) Arora-Rao-Vazirani] [Chlamtac, Arora-Barak-S., explicit: Coloring, Unique Games, Max Bisection Barak-Raghavendra-S., Raghavendra-Tan]
lower bounds for general LP formulations (extended formulations) characterization ; symmetric formulations for TSP & matching [Yannakakisโ88 ] [Fiorini โ Massar โ Pokutta general, exact formulations for TSP & Clique โ Tiwary โ de Wolfโ12] [Braun โ Fiorini โ Pokutta โS.โ12 approximate formulations for Clique Braverman โMoitraโ13] general, exact formulation for maximum matching [ Rothvoรโ13] geometric idea: complicated polytopes can be projections of simple polytopes
universality result for LP relaxations of Max CSPs [this talk] general polynomial-size LP relaxations are no more powerful than polynomial-size Sherali-Adams relaxations also holds for almost quasi-polynomial size concrete consequences unconditional lower bound in powerful computational model confirm non-trivial prediction of P โ NP: poly-size LP relaxations cannot achieve 0.99 approximation for Max Cut, Max 3-Sat, or Max 2-Sat (NP-hard approximations) approximability and UGC: poly-size LP relaxation cannot refute Unique Games Conjecture (cannot improve current Max CSP approximations) separation of LP relaxation and SDP relaxation: poly-size LP relaxations are strictly weaker than SDP relaxations for Max Cut and Max 2Sat
universality result for LP relaxations of Max CSPs [this talk] general polynomial-size LP relaxations are no more powerful than polynomial-size Sherali-Adams relaxations also holds for almost quasi-polynomial size for concreteness: focus on Max Cut notation: cut ๐ป ๐ฆ = fraction of edges that bipartition ๐ฆ cuts in ๐ป Max Cut ๐ = Max Cut instances / graphs on ๐ vertices compare: general ๐ 1โ๐ ๐ -size LP relaxation for Max Cu t ๐ vs. ๐ ๐ -size Sherali-Adams relaxations for Max Cu t ๐
general LP relaxation for ๐๐๐ฒ ๐๐ฏ๐ฎ ๐จ example linearization 1 ๐น ๐ ๐ป ๐ = ๐ ๐๐ ๐๐โ๐น linearization ๐ ๐ฆ ๐๐ = 1, if ๐ฆ ๐ โ ๐ฆ ๐ , otherwise. 0, ๐ป โฆ ๐ ๐ป : โ ๐ โ โ linear such that ๐ ๐ป ๐ x = cut ๐ป ๐ฆ ๐ฆ โฆ ๐ ๐ฆ โ โ ๐ ๐ ๐ฆ . polytope of size R ๐ ๐ ๐ โ โ ๐ , at most ๐ facets, ๐ โ ๐ ๐ ๐ฆ ๐ฆโ ยฑ1 ๐ โ ๐ ๐ same polytope for all instances of size ๐ makes sense because solution space for Max Cut depends only on ๐
computing with size- ๐บ LP relaxation ๐ input computation output maximize ๐ ๐ป ๐ value โ ๐ป graph G = max ๐โ๐ ๐ ๐ป ๐ subject to ๐ โ ๐ ๐ on n vertices always upper-bounds Opt G poly (๐) -time computation how far in the worst-case? approximation ratio ๐ฝ ๐, ๐ก -approximation โ ๐ป โค ๐ฝ โ Opt ๐ป Opt ๐ป โค ๐ก โ โ ๐ป โค ๐ for all ๐ป โ Max Cut ๐ for all ๐ป โ Max Cut ๐ general computational model โ how to prove lower bounds?
geometric characterization (ร la Yannakakisโ88) every size- R LP relaxation ๐ for Max Cu ๐ฎ ๐ corresponds to nonnegative functions ๐ ๐ , โฆ , ๐ ๐บ : ยฑ1 ๐ โ โ โฅ0 such that ๐ โ cut ๐ป = ๐ ๐ ๐ ๐ โ ๐ป โค ๐ iff and ๐ 1 , โฆ , ๐ ๐ โฅ 0 ๐ for all ๐ป โ Max Cut ๐ certifies cut ๐ป โค ๐ over ยฑ1 ๐ canonical linear program of size ๐ example 2 ๐ standard basis functions correspond to exact 2 ๐ -size LP relaxation for Max Cu t ๐
geometric characterization (ร la Yannakakisโ88) every size- R LP relaxation ๐ for Max Cu ๐ฎ ๐ corresponds to nonnegative functions ๐ ๐ , โฆ , ๐ ๐บ : ยฑ1 ๐ โ โ โฅ0 such that ๐ โ cut ๐ป = ๐ ๐ ๐ ๐ โ ๐ป โค ๐ iff and ๐ 1 , โฆ , ๐ ๐ โฅ 0 ๐ for all ๐ป โ Max Cut ๐ intuition: all inequalities for functions on ยฑ1 n with local proofs connection to Sherali-Adams hierarchy ๐ ๐ -size Sherali-Adams relaxation for Max Cut ๐ ๐ -junta = function on ยฑ1 ๐ exactly corresponds to depends on โค d coordinates nonnegative combinations of nonnegative ๐ -juntas on ยฑ1 ๐ )
geometric characterization (ร la Yannakakisโ88) every size- R LP relaxation ๐ for Max Cu ๐ฎ ๐ corresponds to nonnegative functions ๐ ๐ , โฆ , ๐ ๐บ : ยฑ1 ๐ โ โ โฅ0 such that ๐ โ cut ๐ป = ๐ ๐ ๐ ๐ โ ๐ป โค ๐ iff and ๐ 1 , โฆ , ๐ ๐ โฅ 0 ๐ for all ๐ป โ Max Cut ๐ cone ๐ 1 , โฆ , ๐ ๐ ๐ โ cut ๐ป ๐ ๐ ๐ ๐ = ๐ ๐ โฅ 0 ๐ to rule out (c,s)-approx. by size-R LP relaxation, show: for every size- ๐ nonnegative cone, exists ๐ป โ Max Cut ๐ with Opt ๐ป โค ๐ก but ๐ โ cut ๐ป outside of cone
lower-bound for Sherali โ Adams relaxations of size ๐ ๐ lower-bounds for size- ๐ ๐ nonneg. cones with restricted functions ๐ ๐ -juntas ๐ -juntas non-spiky general lower-bound for general LP relaxations of size ๐ 1โ๐ ๐
from ๐ -juntas to ๐ ๐ป -juntas let ๐ 1 , โฆ , ๐ ๐ be nonneg. ๐ ๐ -juntas on ยฑ1 ๐ for ๐ = ๐ 1โ10๐ ๐ want: subset ๐ โ ๐ of size ๐ โ ๐ ๐ ๐พ ๐ ๐/2 ๐พ 1 ๐พ 2 where functions behave like ๐ -juntas S ๐พ 3 ๐พ 4 let ๐พ 1 , โฆ , ๐พ ๐ be junta-coordinates of ๐ 1 , โฆ , ๐ ๐ [n] claim : there exists subset ๐ โ [๐] of size ๐ = ๐ ๐ such that ๐พ ๐ โฉ ๐ โค ๐ for all ๐ โ ๐ proof : choose ๐ at random ๐ ๐ = ๐ โ 1โ2๐ ๐ โ ๐ โฉ ๐พ ๐ > ๐ โค ๐ โ ๐พ ๐ ๏ can afford union bound over ๐ junta sets ๐พ 1 , โฆ , ๐พ ๐
lower-bound for Sherali โ Adams relaxations of size ๐ ๐ lower-bounds for size- ๐ ๐ nonneg. cones with restricted functions ๐ ๐ -juntas ๐ -juntas non-spiky general lower-bound for general LP relaxations of size ๐ 1โ๐ ๐
from ๐ ๐ป -juntas to non-spiky functions let ๐ be a nonnegative function on ยฑ1 ๐ with ๐ฝ๐ = 1 non-spiky: max ๐ โค 2 ๐ข small low-degree junta structure lemma: Fourier coefficients can approximate ๐ by nonnegative ๐ ๐ -junta ๐โฒ , error ๐ = ๐ โ ๐โฒ satisfies ๐ ๐ 2 โค ๐ข๐/๐ ๐ for ๐ < ๐ proof : nonnegative function ๐ ๏ probability distribution over ยฑ1 ๐ , +1/-1 rand. variables ๐ 1 , โฆ , ๐ ๐ (dependent) non-spiky ๏ entropy ๐ผ ๐ 1 , โฆ , ๐ ๐ โฅ ๐ โ ๐ข want: ๐พ โ [๐] of size ๐ ๐ such that โ๐ โ ๐พ . ๐ ๐ โฃ ๐ ๐พ โ uniform, that is, ๐ข๐ ( ๐ < ๐ ) S โ ๐ผ ๐ ๐ ๐ ๐พ โค ๐พ for ๐พ = ๐ ๐ construction: start with ๐พ = โ ; as long as bad ๐ exists, update ๐พ โ ๐พ โช ๐ ๐ข ๐๐ข ๐พ = ๐ ๐ analysis : total entropy defect โค ๐ข ๏ stop after ๐พ iterations ๏ ๐พ โค
Recommend
More recommend