Singh Kunal Talwar MSR NE McGill MSR SV Chvátal Gomory Rounding and Integrality Gaps TexPoint fonts used in EMF Read the TexPoint manual before you delete this box A A A ID: 474437
Download Presentation The PPT/PDF document "Mohit" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Mohit Singh Kunal TalwarMSR NE, McGill MSR SV
Chvátal Gomory Rounding and Integrality Gaps
TexPoint fonts used in EMF.
Read the TexPoint manual before you delete this box.:
A
A
A
A
A
A
A
A
A
A
ASlide2
Approximation Algorithm Design
Cleverly define
Lower Bound
on
Optimum
Think hard
Show that
Write natural linear program Think hardShow that
Write natural linear program Add cleverly designed constraints to get Think hardShow that
Slide3
Would like to establish limits on what we can hope to doAPX-hardness: usually good evidence. Often unavailableLP Gaps show limits of specific LP
Show gap between and
When no good algorithms known
Write natural linear program
Think hard
Show
that
Write natural linear program
Add cleverly designed constraints to get Think hardShow that
Slide4
Ways to automate getting tighter relaxationsLovász-SchrijverSherali-AdamsLassere
Chvátal-GomoryOften (at least retrospectively), improve LP/SDP gapsMatching, MaxCut, Sparsest Cut, Unique Label Cover[
Arora-Bollobás-Lovász
2002
]
Can we establish limits for these procedures?Cut Generating Procedures
Write natural linear program
Add cleverly designed constraints to get
Think hardShow that Slide5
[Arora-Bollobás-Lovász
2002] Vertex Cover: Large Class of LPs has integrality gap
.Implies gaps for LS, SA.
GMTT07,DK07,S08,CMM09,MS09,T09,RS09,KS09,CL10
LS/SA/
Lassere/LS+ gaps for several problemsMaxCut
Unique Label CoverSparsest CutCSPLINMatching…
Gap for LS, SA etc.Slide6
Hypergraph matching in k-uniform hypergraphs
rounds of CG bring gap down to
[
Chan Lau 10
] SA gap is at least
even after
rounds.
This talk: What about Chvátal-Gomory Slide7
Gaps remain large for many rounds of CGVertex Cover
: Gap ) after
rounds
MaxCut
:
Gap
after
roundsUnique Label Cover:
Gap after
rounds-: Gap after rounds
Same as SA gaps.
This talk: What about Chvátal-GomorySlide8
[Gomory 1958]
For a polyhedron Let
where
Let
is polyhedron obtained after j rounds of CG
Defining
Chvátal-Gomory CutsSlide9
-uniform hypergraph:
Each edge
with
Goal: find largest subset of disjoint edges
s.t.
Hypergraph MatchingSlide10
Graph maximum matching
SA takes
rounds to get within
CG gets to integer hull in 1 round
APX hard
-
inapproximable
)approximation
[
Chan Lau 10] Gap after rounds of SAThere is a poly size LP with gap
Hypergraph
matchingSlide11
is an
intersecting family if
for all
[Chan Lau 10]
LP +
intersecting
has gap at most
Intersecting
familySlide12
is an
intersecting family if
for all
Fix
valid for
valid for
valid for
valid for
So
valid for
Intersecting family via CG
Slide13
Extremal combinatorics resultFor any intersecting family in a k-regular hypergraph
, there is one of size
Implies that
Integrality gap of
is bounded by
I.e. for
hypergraph
matching,
round CG is nearly a factor of two better than
round SA. Small families sufficeSlide14
Max Cut LPSlide15
[Charikar
Makarychev Makarychev 09]
for any subset
s.t.
a distribution
overs solutions
such that
is integral for any
Max Cut SA gap
Survives
rounds of SA
Slide16
Observation: for any constraint
in
,
are integers and
(can add arbitrary positive multiple of
to remove negative coefficients and get stronger constraint)
Main idea: show that
Max Cut CG
i
s feasible for
Slide17
Base case: k=0. InspectionInduction Step. Need to show
in
holds for
Case 1:
Case 2:
Proof by inductionSlide18
Let
for some
Recall
a distribution
overs solutions
s.t.
is integral for any
For each
,
For
, set For
, set
, for arbitrary fixed For , set New integral. Agrees with
on
.
. Therefore done.
:
Slide19
By definition,
-1 valid for
Therefore done.
:
Slide20
Similar proofs for unique games, CSPs, VCCG hierarchy often not much better than SANoticeably better for Hypergraph matching
What other problems show large gap between clever LP and LS/SA? Does CG capture them?Conclusions