Effective-Resistance-Reducing Flows,
Spectrally Thin Trees, and ATSP
Nima Anari UC Berkeley
Asymmetric TSP (ATSP)
Given a list of cities and their pairwise βdistancesβ, satisfying the triangle inequality,
Find the shortest tour that
visits all cities exactly once.
Linear Programming Relaxation
[Held-Karpβ72]
πππ
β
π
,
π
π
(
π
,
π
)
π₯
π
,
π
ΒΏ
π
.
π‘
.
β
π
π₯
π
,
π
=
β
π
π₯
π
,
π
βπβπ
β
πβπ
,
πβπ
π₯
π
,
π
β₯
1
ΒΏ
0
β€
π₯
π
,
π
ΒΏ
ΒΏ
Previous Works
Approximation Algorithms
β’ log(n) [Frieze-Galbiati-Maffioliβ82]
β’ .999 log(n) [BlΓ€serβ02]
β’ 0.842 log(n) [Kaplan-Lewenstein-Shafrir-Sviridenkoβ05]
β’ 0.666 log(n) [Feige-Singhβ07]
β’ O(logn/loglogn) [Asadpour-Goemans-Madry-O-Saberiβ09]
β’ O(1) (planar/bd genus) [O-Saberiβ10,Erickson-Sidiropoulosβ13]
Integrality Gap
Main Result
For any cost function, the integrality gap of the LP relaxation is polyloglog(n).
Plan of the Talk
ATSP
ATSP Thin Spanning
Thin Spanning Trees
Def: Given a k-edge-connected graph . A spanning tree is -thin w.r.t. G if
Kn 2/n-thin tree
One-sided unweighted cut-sparsifier No lower-bound on
One-sided unweighted cut-sparsifier No lower-bound on
β π βπ ,
|
π(
π , π)
|
β€ πΌ β β¨πΈ(
π ,π)
β¨ΒΏExercise: Show that (k-dim cube) has O(1/k) thin tree
From Thin Trees to ATSP
[AGMOSβ09]: If for any -connected graph , then the integrality gap of LP is .
Furthermore, finding the tree algorithmically gives -approximation algorithm for ATSP.
Previous Works: Randomized Rounding
Thm: Any k-connected graph G has a thin tree
Pf. Sample each edge of G, indep, w.p. .
By Kargerβs cut counting argument, the sampled graph is -thin w.h.p.
Main Result
Any -edge-connected graph has an -thin tree.
Any -edge-connected graph has an -thin tree.
For any cost function, the integrality gap of the LP Relaxation is polyloglog(n).
In Pursuit of Thin Trees
Graph Laplacian
Let
For let
Laplacian Quadratic Form:
πΏπΊ=
[
2 β 1 β1 0
β1 2 0 β1
β1 0 2 β1 0 β 1 β1 2
]
Spectrally Thin Spanning Trees
Def: A spanning tree is -spectrally thin w.r.t. G if
Why?
β’ Generalizes (combinatorial) thinness.
β -spectral thinness implies -combinatorial thinness
β’ Testable in polynomial time.
Lem: The spectral thinness of any T is at least
Pf. If T is spectrally thin, then any subgraph of T is -spectrally thin, so
is the spectral thinness of .
A Necessary Condition for Spectral Thinness
A k-con Graph with no Spectrally Thin Tree
For any spanning tree, T,
Reff (π)β 1β π
2
π
k edges k edges
A Sufficient Condition for Spectral Thinness
[Marcus-Spielman-Srivastavaβ13,Harvey-Olverβ14]: Any G has an -spectrally thin tree.
Main Idea
An Example
An Observation
An Application of [MSSβ13]: If for any cut ,
Main Idea
Find a ``graphββ D s.t. and
i.e.,
Find a ``graphββ D s.t. and
i.e.,
D+G has a spectrally thin tree and
any spectrally thin tree of G+D is (comb) thin in G. D+G has a spectrally thin tree and
An Impossibility Theorem
Proof Overview
k-connected graph G for
k-connected graph G for
,
F is -connected, ,
F is -connected,
G has -comb thin tree G has -comb
thin tree
A General. of [MSSβ13] Main
Tech Thm
has -spectrally thin tree
has -spectrally thin tree
Note we may have Note we may have D is not
a graph D is not
Thm: Given a set of vectors s.t.
If then there is a basis s.t.
β¦β¦β¦...β¦,there are disjoint bases,
Thin Basis Problem
β
β
π βπ
π£
ππ£
ππβ
β€
π
(
π
)
.
d Linearly independent set of vectors
Proof Overview
k-connected graph G for
k-connected graph G for
,
F is -connected, ,
F is -connected,
G has -comb thin tree G has -comb
thin tree
A General. of [MSSβ13] Main
Tech Thm
has -spectrally thin tree
A Weaker Goal: Satisfying Degree Cuts
Thm: Given a k-connected graph, , s.t., for all v,
for
Let then by Markov Ineq,
, for all v.
A Convex Program for Optimum D
πππ
max
π£βπ
πΌ
πβΌπΏ(π£)Reff
π·(
π
)
ΒΏ
π
.
π‘
.
π·
βΌ
πΆπΏ
πΊΒΏ
ΒΏ
Has exp. many constraints: Has exp. many constraints:
Recall convexity of matrix inv Recall convexity of matrix inv
If write ,
then optimum is If write ,
then optimum is
Main Result
Any -edge-connected graph has an -thin tree. Any -edge-connected graph has an -thin tree.
For any cost function, the integrality gap of the LP Relaxation is polyloglog(n).
Conclusion
Main Idea:
β’ Symmetrize L2 structure of G while preserving its L1 structure
Tools:
β’ Interlacing polynomials/Real Stable polynomials β’ Convex optimization
β’ Graph partitioning
Future Works/Open Problems
β’ Algorithmic proof of [MSSβ13] and our extension.
β’ Existence of C/k thin trees and constant factor
approximation algorithms for ATSP.
β’ Subsequent work: Svensson designed a 27-app