An O(log n/ log log n)-approximation Algorithm for the

Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Asymmetric Traveling Salesman Problem

Arash Asadpour∗ Michel X. Goemans† Aleksander Madry

,
‡
Shayan Oveis Gharan§
¶
Amin Saberi

Abstract theorem [24] shows that if the tree chosen in the first step is
We consider the Asymmetric Traveling Salesman problem “thin” then the cost of the Eulerian augmentation is within a
for costs satisfying the triangle inequality. We derive a ran- factor of the “thinness” of the (asymmetric) Held-Karp linear
domized algorithm which delivers a solution within a factor programming (LP) relaxation value (OPTHK ) [17]. This flow
O(log n/ log log n) of the optimum with high probability. argument works irrespectively of the actual directions of
the (directed) arcs correponding to the (undirected) edges
1 Introduction of the tree. Roughly speaking, a thin tree with respect to
the optimum solution x ∗ of the Held-Karp relaxation is a
In the Asymmetric Traveling Salesman problem (ATSP) we
spanning tree that, for every cut, contains a small multiple
are given a set V of n points and a cost function c : V × V →
(the thinness) of the corresponding value of x ∗ in this cut
R+ . The goal is to find a minimum cost tour that visits every
when the direction of the arcs are disregarded.
vertex at least once. Since we can replace every arc (u, v) in
A key step of our algorithm is to find a thin tree of small
the tour with the shortest path from u to v, we can assume c
cost compared to the LP relaxation value OPTHK . For this
satisfies the triangle inequality.
purpose, we consider the distribution with maximum en-
When the costs are symmetric, i.e. when for every
tropy among all those with marginal probabilities obtained
u, v ∈ V, c(u, v) = c(v, u), there is a factor 3/2 approximation
from the symmetrized LP solution (scaled by 1 − 1/n). From
algorithm due to Christofides [8]. This algorithm first finds a
the optimality conditions of a convex programming formula-
minimum cost spanning tree T on V, then finds the minimum
tion, we derive that this maximum entropy distribution corre-
cost Eulerian augmentation of that tree, and finally shortcuts
sponds to sampling a tree T with probability proportional to
the corresponding Eulerian walk into a tour.
e∈T λe for appropriately defined λe ’s for e ∈ E. We develop
Q
In this paper, we give an O(log n/ log log n) approxima-
a simple iterative algorithm for approximately computing
tion algorithm for the general asymmetric version. This fac-
these λe ’s efficiently. An important property of this scheme
tor finally breaks the Θ(log n) barrier from Frieze et al. [12]
is that the events correponding to edges being present in the
and subsequent improvements [3, 16, 11]. Our approach
sampled tree are negatively correlated. This means that the
for ATSP has similarities with Christofides’ algorithm; we
well-known Chernoff bound for the independent setting still
first construct a spanning tree with special properties. Then
holds, see Panconesi and Srinivasan [23].The proof of the
we find a minimum cost Eulerian augmentation of this tree,
O(log n/ log log n) thinness of the sampled tree is based on
and finally, shortcut the resulting Eulerian walk. For undi-
this tail bound.
rected graphs, being Eulerian means being connected and
The high level description of our algorithm can be found
having even degrees, while for directed graphs it means be-
in Figure 1. The proof of our main Theorem 6.3 also gives a
ing (strongly) connected and having the indegree of every
more formal overview of the algorithm.
vertex equal to its outdegree.
A simple flow argument using Hoffman’s circulation
2 Notation
∗ Stanford University, Department of Management Science and Engi- Before describing our approximation algorithm for ATSP in
neering. [email protected]. details, we need to introduce some notation. Throughout this
† MIT, Department of Mathematics. Supported by NSF contract CCF-
paper, we use a = (u, v) to denote the arc (directed edge)
0829878 and by ONR grant [email protected].
‡ MIT, Computer Science and Artificial Intelligence Laboratory. Sup-
from u to v and e = {u, v} for an undirected edge. Also we
ported by Fulbright Science and Technology Award, by NSF contract CCF-
use A (resp. E) for the set of arcs (resp. edges) in a directed
0829878, and by ONR grant N00014-05-1-0148. [email protected]. (resp. undirected) graph.
§ Stanford University, Department of Management Science and Engi- For a given function f : A → R, the cost of f is defined
neering. [email protected].
¶ Stanford University, Department of Management Science and Engi-

neering. [email protected].

379 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 INPUT: A set V consisting of n points and a cost function c : V × V → R+ satisfying the triangle inequality.
OUTPUT: O( logloglogn n )-approximation for the Asymmetric Traveling Salesman Problem on V.

ALGORITHM:
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

1. Solve the Held-Karp LP relaxation of the ATSP instance to get an optimum extreme point solution x ∗ . [See
LP (3.1).] Define z ∗ by (3.5); z ∗ can be interpreted as the marginal probabilities on the edges of a probability
distribution on spanning trees.
2. Sample Θ(log n) spanning trees T j ’s from a distribution p̃(·) that approximates the maximum entropy distribution
among all the distributions that approximately preserve the marginal probabilities imposed by z ∗ . Let T ∗ be the
tree with minimum (undirected) cost among all the sampled trees. [See Sections 4 and 5.]
3. Orient each edge of T ∗ so as to minimize its cost. Find a minimum cost integral circulation that contains the
oriented tree T~ ∗ . Shortcut this multigraph and output the resulting Hamiltonian cycle.

Figure 1: An O(log n/ log log n)-approximation algorithm for the ATSP.

as follows: X This relaxation is known as the Held-Karp relaxation [17]

c( f ) := c(a) f (a). and its optimum value, which we denote by OPTHK , can
a∈A be computed in polynomial-time (either by the ellipsoid
For a set S ⊆ A, we define algorithm or by reformulating it as an LP with polynomially-
X bounded size). Observe that (3.3) implies that any feasible
f (S ) := f (a). solution x to the Held-Karp relaxation satisfies
a∈S
(3.4) x (δ+ (U)) = x (δ− (U)),
We use the same notation for a function defined on the edge
set E of an undirected graph. For U ⊆ V, we also define the for any U ⊂ V.
following sets of arcs: Let x ∗ denote an optimum solution to this LP (3.1); thus
c(x ) = OPTHK . We can assume that x ∗ is an extreme point
∗
δ+ (U) := {a = (u, v) ∈ A : u ∈ U, v < U},
of the corresponding polytope. We first make this solution
δ− (U) := δ+ (V\U) symmetric and slightly scale it down by setting
A(U) := {a = (u, v) ∈ A : u ∈ U, v ∈ U}.
n−1 ∗
(3.5) z ∗{u,v} := (xuv + xvu
∗
).
Similarly, for an undirected graph G = (V, E), δ(U) denotes n
the set of edges with exactly one endpoint in U, and E(U) Let A denote the support of x ∗ , i.e. A = {(u, v) : x∗ > 0},
uv
denotes the edges entirely within U, i.e. E(U) = {{u, v} ∈ E : and E the support of z ∗ . For every edge e = {u, v} of E, we
u ∈ U, v ∈ U}. can define its cost as min{c(a) : a ∈ {(u, v), (v, u)} ∩ A}; with
Throughout the paper, log denotes the natural logarithm. the risk of overloading the notation, we denote this new cost
of this edge e by c(e). This implies that c(z ∗ ) < c(x ∗ ).
3 The Held-Karp Relaxation The main purpose of the scaling factor in the definition
Given an instance of ATSP corresponding to the cost func- (3.5) is to obtain a vector z ∗ which belongs to the spanning
tion c : V×V → R+ , we can obtain a lower bound on the opti- tree polytope P of the graph (V, E), i.e. z ∗ can be viewed
mum value by considering the following linear programming as a convex combination of incidence vectors of spanning
relaxation defined on the complete bidirected graph with ver- trees, see Lemma 3.1. In fact, z ∗ even belongs to the relative
tex set V: interior of P.
X
(3.1) min c(a)xa L 3.1. The vector z ∗ defined by (3.5) belongs to
a relint(P), the relative interior of the spanning tree polytope
(3.2) s.t. x (δ+ (U)) ≥ 1 ∀U ⊂ V, P.
(3.3) x (δ+ (v)) = x (δ− (v)) = 1 ∀v ∈ V,
Proof. From Edmonds’ characterization of the base poly-
xa ≥ 0 ∀a. tope of a matroid [10], it follows that the spanning tree

380 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 polytope P is defined by the following inequalities (see [24, In Section 4.1, through a convex program, we formally
Corollary 50.7c]): define the maximum entropy distribution over the spanning
trees with respect to the marginal probabilities given by z (an
(3.6) P = {z ∈ RE : z (E) = |V| − 1 arbitrary point in the relative interior of the spanning tree
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(3.7) z (E(U)) ≤ |U| − 1, ∀U ⊂ V polytope). From the optimality conditions for this convex
(3.8) ze ≥ 0 ∀e ∈ E.} program and its dual, we show that this distribution generates
a λ-random spanning tree for some vector λ ∈ R|E| , where
The relative interior of P corresponds to those z ∈ P any tree T is output with probability proportional to e∈T λe .
Q

satisfying all inequalities (3.7) and (3.8) strictly. Section 4.3 explains the main implication of such distri-
∗
Clearly, z satisfies (3.6) since: butions. The events corresponding to the edges of G being
included in a sampled λ-random tree are negatively corre-
∀v ∈ V, x ∗ (δ+ (v)) = 1 ⇒ x ∗ (A) = n = |V| lated. This enables us to use Chernoff bounds on such events.
⇒ z (E) = n − 1 = |V| − 1.
∗ We use these tail bounds to establish the thinness of a sam-
pled tree. (Roughly speaking, a tree is said to be thin if the
Consider any set U ⊂ V. We have number of its edges in each cut is not much higher than its
X expected value; see Section 5 for a formal definition of thin-
∗ + ∗ +
x (δ (v)) = |U| = x (A(U)) + x (δ (U))
∗ ness.)
v∈U It is possible to approximately find the γe ’s efficiently.
≥ x (A(U)) + 1.
∗ In fact, we have a rather simple and iterative algorithm that,
after a polynomial number of iterations, finds approximate
Since x ∗ satisfies (3.2) and (3.3), we have λ̃e ’s with new marginal probabilities z̃e , where for all edges e,
z̃e ≤ (1 + ε)ze . We postpone the description of this algorithm
n−1 ∗ and its analysis to Section 7. Instead, in Section 4.2 we
z (E(U)) =
∗
x (A(U)) < x (A(U)) ≤ |U| − 1,
∗
n show how to efficiently sample a tree from such a distribution
showing that z satisfies (3.7) strictly. Since E is the support given any vector λ.
∗

of z ∗ , (3.8) is also satisfied strictly by z ∗ . This shows that z ∗

is in the relative interior of P. 4.1 Maximum Entropy Distribution. Let T be the col-
lection of all the spanning trees of G = (V, E). The maximum
One implication of being in the relative interior of P entropy distribution p∗ (·) with respect to given marginal
is that z ∗ can be expressed as a convex combination of probabilities z is the optimum solution of the following con-
spanning trees such that the coefficient corresponding to any vex program CP:
spanning tree is positive. X
∗
Regarding the size of the extreme point x , it is known (4.9) inf p(T ) log p(T )
T ∈T
that its support A has at most 3n − 4 arcs (see [14, Theorem X
15]). In addition, we know that it can be expressed as the s.t. p(T ) = ze ∀e ∈ E,
unique solution of an invertible system with only 0 − 1 T 3e
coefficients, and therefore, we have that every entry xa∗ is p(T ) ≥ 0 ∀T ∈ T .
rational with integral numerator and denominator bounded
by 2O(n log n) . In particular, z∗min = mine∈E z∗e > 2−O(n log n) . This convex program is feasible whenever z belongs to
the spanning tree polytope P defined on G = (V, E). As
4 Maximum Entropy Sampling and Concentration the objective function is bounded and the feasible region is
Bounds compact (closed and bounded), the inf is attained and there
∗ exists an optimum solution p∗ (·). Furthermore, since the
Our aim in this section is to round z , as a point in the
objective function is strictly convex, this maximum entropy
relative interior of the spanning tree polytope, to a spanning
distribution p∗ (·) is unique. Let OPTCP denote the optimum
tree. Suppose that we are looking for a distribution over
value of this convex program CP.
spanning trees of G that preserves the marginal probabilities
The value p∗ (T ) determines the probability of sampling
imposed by z , i.e. PrT [e ∈ T ] = ze for every edge e ∈ E.
∗ ∗
any tree T in the maximum entropy rounding scheme. Note
There are plenty such distributions. The approach we use is
that it is implicit in the constraints of this convex program
based on sampling from the distribution that maximizes the
that, for any feasible solution p(.), we have T p(T ) = 1
P
entropy among all marginal preserving distributions. Such a
since
maximum entropy rounding scheme has been used in [2] for
sampling a random matching in a bipartite graph with given
X XX X
n−1= ze = p(T ) = (n − 1) p(T ).
marginal probabilities. e∈E e∈E T 3e T

381 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 We now want to show that, if we assume that the vector z for any p ≥ 0 and any γ, where we have implicitly redefined
is in the relative interior of the spanning tree polytope of L due to our change of variables from δ to γ. Therefore, p∗
(V, E) then p∗ (T ) > 0 for every T ∈ T and p∗ (T ) admits a is the unique minimizer of L(p, γ∗ ) and from (4.11), we have
simple exponential formula. Note that the vector z ∗ obtained that
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

from the LP relaxation of the ATSP indeed satisfies this

p∗ (T ) = eγ (T ) .
∗
assumption. (4.13)
For this purpose, we write the Lagrange dual to the
Summarizing, the following theorem holds.
convex program CP, see for example [22]. For every e ∈
E, we associate a Lagrange multiplier δe to the constraint T 4.1. Given a vector z in the relative interior of the
corresponding to the marginal probability ze , and define the spanning tree polytope P on G = (V, E), there exist γe∗ for
Lagrange function by all e ∈ E such that if we sample a spanning tree T of G
according to p∗ (T ) := eγ (T ) then Pr[e ∈ T ] = ze for every
∗
 
X X X
L(p, δ) = δe  p(T ) − ze  .

p(T ) log p(T ) −  e ∈ E.
T ∈T e∈E T 3e
It is worth noting that the requirement that z is in the
This can also be written as: relative interior of the spanning tree polytope (as opposed
X X 

X  to being just in this polytope) is necessary (the fact that

L(p, δ) = δe ze +  p(T ) log p(T ) − p(T )
 δe  .
being in the spanning tree polytope was not sufficient had
e∈E T ∈T e∈T been observed before, see Exercise 4.19 in [21]). Let G be a
triangle and z be the vector ( 21 , 12 , 1). In this case, z is in the
The Lagrange dual to CP is now
polytope (but not in its relative interior) and there are no γe∗ ’s
(4.10) sup inf L(p, δ). that would satisfy the statement of the theorem (however, one
δ p≥0 can get arbitrarily close to ze for all e ∈ E).
The inner infimum in this dual is easy to solve. As the In Section 7 we show how to efficiently find γ̃’s that
contributions of the p(T )’s are separable, we have that, for approximately satisfy the conditions of Theorem 4.1. More
every T ∈ T , p(T ) must minimize the convex function formally, we prove the following theorem whose result we
use in the rest of the paper. For our ATSP application, zmin is
p(T ) log p(T ) − p(T )δ(T ), 2−O(n log n) and ε can be chosen to be 0.2.

where, as usual, δ(T ) = e∈T δe . Taking derivatives, we T 4.2. Given z in the spanning tree polytope of G =
P
derive that (V, E) and some ε > 0, values γ̃e for all e ∈ E can be found,
0 = 1 + log p(T ) − δ(T ), so that if we define the exponential family distribution
or 1 X
p̃(T ) := exp( γ̃e )
P
(4.11) p(T ) = eδ(T )−1 . e∈T

for all T ∈ T where

Thus, X X
inf L(p, δ) = δe ze − eδ(T )−1 .
X X
p≥0 P := exp( γ̃e )
e∈E T ∈T
T ∈T e∈T
Using the change of variables γe = δe − n−1 1
for e ∈ E, the
then, for every edge e ∈ E,
Lagrange dual (4.10) can therefore be rewritten as
X

X X
 z̃e := p̃(T ) ≤ (1 + ε)ze ,
γ(T ) 
sup 1 + z e γe − e  .
 
(4.12)  T ∈T :T 3e
γ e∈E T ∈T
i.e. the marginals are approximately preserved. Further-
Our assumption that the vector z is in the relative more, the running time is polynomial in n = |V|, − log zmin
interior of the spanning tree polytope is a Slater condition and 1/ε.
and, together with convexity, implies that the sup in (4.12)
is attained by some vector γ∗ , and the Lagrange dual value The distributions over trees considered in Theorem 4.2
equals the optimum value OPTCP of our convex program. are closely related to the notion of λ -random (spanning)
Furthermore, we have that the (unique) primal optimum trees. Given λe ≥ 0 for e ∈ E, a λ -random tree T of G
solution p∗ and any dual optimum solution γ∗ must satisfy is a tree T chosen from the set of all spanning trees of G
with probability proportional to e∈T λe . The notion of λ -
Q
L(p, γ∗ ) ≥ L(p∗ , γ∗ ) ≥ L(p∗ , γ), random trees has been extensively studied (see e.g. Ch.4 of

382 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 [21]) - note that in case of all λe being equal, a λ -random G0 and values of λe ’s, the probability pG0 [λ , f ] that some
tree is just a uniform spanning tree of G. Many of the results edge f is in a λ -random tree of G0 . This is well-known.
for uniform spanning trees carry over to λ -random spanning For this purpose, one can evaluate T ∈T e∈T λe for both G0
P Q
trees in a graph G since, for rational λe ’s, a λ -random and G /{ f } (in which edge f is contracted) using Kirchhoff’s
0
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

spanning tree in G corresponds to a uniform spanning tree matrix tree theorem (see [4]). The matrix tree theorem states
that T ∈T e∈T λe for any graph G is equal to the absolute
P Q
in a multigraph obtained from G by letting the multiplicity
of edge e be proportional to λe . value of any cofactor of the weighted Laplacian L where
Observe that a tree T sampled from an exponential
e = (i, j) ∈ E

family distribution p(·) as given in Theorem 4.2 is λ -random −λ
 Pe


Li, j =  λ i= j

for λe := eγe for all e ∈ E. As a result, we can use the tools  e∈δ({i}) e

developed for λ -random trees to obtain an efficient sampling
 0 otherwise.
procedure, see Section 4.2, and to derive sharp concentration
An alternative approach for computing pG0 [λ , f ] is to use the
bounds for the distribution p(·), see Section 4.3.
fact (see e.g. Ch. 4 of [21]) that pG0 [λ , f ] is equal to λ f times
the effective resistance of f in G0 treated as an electrical
4.2 Sampling a λ -Random Tree .
circuit with conductances of edges given by λ . The effective
There is a host of results (see [15, 20, 9, 1, 5, 25, 19] and
resistance can be expressed by an explicit linear-algebraic
references therein) on obtaining polynomial-time algorithms
formula whose computation boils down to inverting a certain
for generating a uniform spanning tree, i.e. a λ -random tree
matrix that can be easily derived from the Laplacian of G0
for the case of all λe ’s being equal. Almost all of them can
(see e.g. section 2.4 of [13] for details).
be easily modified to allow arbitrary λ ; however, not all of
them still guarantee a polynomial running time for general
4.3 Negative Correlation and a Concentration Bound.
λe ’s. We use for example an iterative approach similar to
We derive now the following concentration bound. As
[20].
discussed in the next section, this bound is instrumental in
The idea is to order the edges e1 , . . . , em of G arbitrarily
establishing the thinness of a sampled tree.
and process them one by one, deciding probabilistically
whether to add a given edge to the final tree or to discard it. T 4.3. For each edge e, let Xe be an indicator random
More precisely, when we process the j-th edge e j , we decide variable associated with the event [e ∈ T ], where T is a λ -
to add it to a final spanning tree T with probability p j being random tree. Also, for any subset C of the edges of G, define
the probability that e j is in a λ -random tree conditioned on X(C) = e∈C Xe . Then we have
P
the decisions that were made for edges e1 , . . . , e j−1 in earlier
iterations. Clearly, this procedure generates a λ -random tree, eδ
!E[X(C)]
and its running time is polynomial as long as the computation Pr[X(C) ≥ (1 + δ)E[X(C)]] ≤ .
(1 + δ)1+δ
of the probabilities p j can be done in polynomial time.
To compute these probabilities efficiently we note that, Usually, when we want to obtain such concentration
by definition, p1 = ze1 . Now, if we choose to include e1 in bounds, we prove that the variables {Xe }e are independent
the tree then: and use the Chernoff bound. Unfortunately, in our case, the
variables {Xe }e are not independent. However, it is well-
e∈T 0 λe
P Q
T 0 3e1 ,e2
p2 = Pr[e2 ∈ T |e1 ∈ T ] = known that they are negatively correlated, i.e. for any subset
e∈T 0 λe
P Q
0
F ⊆ E, Pr[∀e∈F Xe = 1] ≤ e∈F Pr[Xe = 1], see e.g. ch. 4 of
T 3e1
Q
e∈T 0 \e1 λe
P Q
T 0 3e ,e [21]1 .
= P 1 2Q .
T 0 3e1 e∈T 0 \e1 λe
L 4.1. The random variables {Xe }e are negatively cor-
As one can see, the probability that e2 ∈ T conditioned on related.
the event that e1 ∈ T is equal to the probability that e2 is in a
λ -random tree of a graph obtained from G by contracting the Once we have established negative correlation between
edge e1 . Similarly, if we choose to discard e1 , the probability the Xe ’s, Theorem 4.3 follows directly from the result of
p2 is equal to the probability that e2 is in a λ -random tree of Panconesi and Srinivasan [23] that the upper tail part of the
a graph obtained from G by removing e1 . In general, p j is
equal to the probability that e j is included in a λ -random tree 1 Lyons and Peres prove this fact only in the case of T being a uniform

of a graph obtained from G by contracting all edges that we spanning tree i.e. when all λe ’s are equal, but Section 4.1 of [21] contains a
have already decided to add to the tree, and deleting all edges justification why this proof implies this property also in the case of arbitrary
λe ’s. As mentioned before, for rational λe ’s, the main idea is to replace
that we have already decided to discard. each edge e with Cλe edges (for an appropriate choice of C) and consider a
Therefore, the only thing we need in order to compute uniform spanning tree in the corresponding multigraph. The irrational case
the p j ’s is to be able to compute efficiently for a given graph follows from a limit argument.

383 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 Chernoff bound requires only negative correlation (or even a from above by
weaker notion, see [23]) and not the full independence of the
random variables. Pr[|T ∩ δ(U)| > (1 + δ)E[|T ∩ δ(U)|]]
eδ
!z̃ (δ(U))
Since an earlier version of this paper, other ways of
≤
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

producing negatively correlated probability distributions on (1 + δ)1+δ

trees (or, more generally, matroid bases) satisfying some  e (1+δ)z̃ (δ(U))
given marginals z have been proposed [7, 26]. Chekuri and ≤
1+δ
Vondrak [7] use a randomized variant of pipage rounding  e βz ∗ (δ(U))
while Zenklusen uses a randomized selection based on the =
1+δ
basis exchange property of matroid bases. Both approaches !β z ∗ (δ(U))
 e(1 + ε) 

can also be used in the framework developed in this paper. 
≤ 
β

5 The Thinness Property ∗
≤ n−4(1−1/e)z (δ(U)) ,
In this section, we focus on the exponential family distribu-
tion p̃(·) that we obtain by applying the algorithm of The- where, in the last inequality, we have used that
orem 4.2 to z ∗ . We show that the tree sampled from the !β 
 e(1 + ε) 

log n
distribution p̃(·) is almost surely “thin”. First we define the log  = 4 [1 + log(1 + ε) − log(4)
β log log n

thinness property.
− log log n + log log log n]
!
log log log n
D 5.1. We say that a tree T is α-thin if for each set ≤ −4 log n 1 −
U ⊂ V, log log n
!
1
|T ∩ δ(U)| ≤ αz ∗ (δ(U)). ≤ −4 1 − log n ≤ −2.5 log n,
e
Also we say that T is (α, s)-thin if it is α-thin and moreover, since e(1 + ε) < 4 and log log log n
≤ 1
for all n ≥ 5 (even for
log log n e
n ≥ 3).
c(T ) ≤ sOPTHK .
We are ready to prove the main theorem of this section.
We first prove that if we focus on a particular cut then T 5.1. Let n ≥ 5 and ε = 0.2. Let T 1 , . . . , T d2 log ne
the “α-thinness” property holds for it with overwhelming be d2 log ne independent samples from a distribution p̃(·) as
probability where α ∼ logloglogn n . given in Theorem 4.2. Let T ∗ be the tree among these sam-
ples that minimizes c(T j ). Then, T ∗ is (4 log n/ log log n, 2)-
L 5.1. If T is a spanning tree sampled from distribution thin with high probability.
p̃(·) for ε = 0.2 in a graph G with n ≥ 5 vertices then for any
set U ⊂ V, Here, high probability means probability at least 1 − 1/n, but
probability 1 − 1/nk can be achieved by sampling 2k log n
∗ trees.
Pr[|T ∩ δ(U)| > βz ∗ (δ(U))] ≤ n−2.5z (δ(U)) ,
Proof. We start by showing that for any 1 ≤ j ≤ d2 log ne,
where β = 4 log n/ log log n. T j is β-thin with high probability for β = 4 log n/ log log n.
From Lemma 5.1 we know that the probability of some
particular cut δ(U) violating the β-thinness of T j is at most
Proof. Note that by definition, for all edges e ∈ E, z̃e ≤ (1 + ∗
n−2.5z (δ(U)) . Now, we use a result of Karger [18] that shows
ε)ze , where ε = 0.2 is the desired accuracy of approximation
∗
that there are at most n2l cuts of size at most l times the
of z ∗ by z̃ as in Theorem 4.2. Hence,
minimum cut value for any half-integer l ≥ 1. Since, by
the definitions of the Held-Karp relaxation and of z ∗ , we
E[|T ∩ δ(U)|] = z̃ (δ(U)) ≤ (1 + ε)z (δ(U)).
∗
know that z ∗ (δ(U)) ≥ 2(1 − 1/n), it means there is at most nl
cuts δ(U) with z ∗ (δ(U)) ≤ l(1 − 1/n) for any integer l ≥ 2.
Applying Theorem 4.3 with Therefore, by applying the union bound (and n ≥ 5), we
derive that the probability that there exists some cut δ(U)
z ∗ (δ(U)) β with |T j ∩ δ(U)| > βz ∗ (δ(U)) is at most
1+δ=β ≥ ,
z̃ (δ(U)) 1+ε ∞
X
ni n−2.5(i−1)(1−1/n) ,
we derive that Pr[|T ∩ δ(U)| > βz ∗ (δ(U))] can be bounded i=3

384 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 where each term is an upper bound on the probability that undirected cost function, we have c(T~ ∗ ) = c(T ∗ ). We then
there exists a violating cut of size within [(i−1)(1−1/n), i(1− find a minimum cost augmentation of T~ ∗ into an Eulerian
1/n)]. Fo n ≥ 5, this simplifies to: directed graph; this can be formulated as a minimum cost
∞ ∞
circulation problem with integral lower capacities (and no or
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

X X 1 infinite upper capacities). Indeed, set

nni −2.5(i−1)(1−1/n)
≤ n−i+2
= ,
n − 1
1 a ∈ T~ ∗
i=3 i=3
(
l(a) =
Thus, indeed, T j is a β-thin spanning tree with high proba- 0 a < T~ ∗ ,
bility. and consider the minimum cost circulation problem
Now, the expected cost of T j is
min{c( f ) : f is a circulation and f (a) ≥ l(a) ∀a ∈ A}.
X n−1X ∗
E[c(T j )] ≤ z̃e ≤ (1 + ε) x ≤ (1 + ε)OPTHK . An optimum circulation f ∗ can be computed in polynomial
e∈E
n a∈A a
time and can be assumed to be integral, see e.g. [24,
∗
So, by Markov inequality we have that for any j, the Corollary 12.2a]. This integral circulation f corresponds to
~ ∗
probability that c(T j ) > 2OPTHK is at most (1 + ε)/2. Thus, a directed (multi)graph H which contains T . Every vertex
with probability at most ( 2 ) 1+ε 2 log n
< n for ε = 0.2, we have in H has an indegree equal to its outdegree. Therefore, every
1

c(T ∗ ) > 2OPTHK . This concludes the proof of the theorem. cut has the same number of arcs in both directions. As H is
weakly connected (as it contains T~ ∗ ), it is strongly connected
6 Transforming a Thin Spanning Tree into an Eulerian and thus, H is an Eulerian directed multigraph. We can
Walk extract an Eulerian walk of H and shortcut it to obtain our
Hamiltonian cycle of cost at most c( f ∗ ) since the costs satisfy
As the final step of the algorithm, we show how one can find
the triangle inequality.
an Eulerian walk with small cost using a thin tree. After
To complete the proof of Theorem 6.1, it remains to
finding such a walk, one can use the metric property to
show that c( f ∗ ) ≤ (2α + s)c(x ∗ ). For this purpose, we define
convert this walk into an Hamiltonian cycle of no greater
1 + 2αxa∗ a ∈ T~ ∗
(
cost (by shortcutting). In particular, the following theorem
u(a) =
justifies the definition of thin spanning trees. 2αxa∗ a < T~ ∗ .
T 6.1. Assume that we are given an (α, s)-thin span- We claim that there exists a circulation g satisfying l(a) ≤
ning tree T ∗ with respect to the LP relaxation x ∗ . Then g(a) ≤ u(a) for every a ∈ A. To prove this claim,
we can find a Hamiltonian cycle of cost no more than we use Hoffman’s circulation theorem 6.2. Indeed, by
(2α + s)c(x ∗ ) = (2α + s)OPTHK in polynomial time. construction, l(a) ≤ u(a) for every a ∈ A; furthermore,
Lemma 6.1 below shows that, for every U ⊂ V, we have
Before proceeding to the proof of Theorem 6.1, we l(δ− (U)) ≤ u(δ+ (U)). Thus the existence of the circulation g
recall some basic network flow results related to circulations. is established. Furthermore,
A function f : A → R is called a circulation if f (δ+ (v)) =
f (δ− (v)) for each vertex v ∈ V. Hoffman’s circulation c( f ∗ ) ≤ c(g) ≤ c(u) = c(T~ ∗ ) + 2αc(x ∗ ) ≤ (2α + s)c(x ∗ ),
theorem [24, Theorem 11.2] gives a necessary and sufficient establishing the bound on the cost of f ∗ . This completes the
condition for the existence of a circulation subject to lower proof of Theorem 6.1.
and upper capacities on arcs.
L 6.1. For the capacities l and u as constructed in the
T 6.2. (H’  ) Given proof of Theorem 6.1, the following holds for any subset
lower and upper capacities l, u : A → R, there exists a U ⊂ V:
circulation f satisfying l(a) ≤ f (a) ≤ u(a) for all a ∈ A if l(δ− (U)) ≤ u(δ+ (U)).
and only if
Proof. Irrespective of the orientation of T ∗ into T~ ∗ , the
1. l(a) ≤ u(a) for all a ∈ A and number of arcs of T~ ∗ in δ− (U) is at most αz ∗ (δ(U)) by
definition of α-thinness. Thus
2. for all subsets U ⊂ V, we have l(δ− (U)) ≤ u(δ+ (U)).
l(δ− (U)) ≤ αz ∗ (δ(U)) < 2αx ∗ (δ− (U)),
Furthermore, if l and u are integer-valued, f can be chosen
to be integer-valued. due to (3.4) and (3.5). On the other hand, we have
u(δ+ (U)) ≥ 2αx ∗ (δ+ (U)) = 2αx ∗ (δ− (U)) ≥ l(δ− (U)),
Proof. [Theorem 6.1] We first orient each edge {u, v} of T ∗
to arg min{c(a) : a ∈ {(u, v), (v, u)} ∩ A}, and denote the re- where we have used the fact that x ∗ itself is a circulation (see
sulting directed tree by T~ ∗ . Observe that by definition of our (3.4)). The lemma follows.

385 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 T 6.3. For a suitable choice of parameters, the al- • Compute δ such that if we define γ0 as γe0 = γe − δ,
gorithm given in Figure 1 finds a (2 + 8 log n/ log log n)- and γ0f = γ f for all f ∈ E \ {e}, then qe (γ0 ) =
approximate solution to the Asymmetric Traveling Salesman (1 + ε/2)ze
Problem with high probability and in time polynomial in the • Set γ ← γ0
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

size of the input.

3. Output γ̃ := γ.
Proof. The algorithm start by finding an optimal extreme-
point solution x ∗ to the Held-Karp LP relaxation of ATSP Clearly, if the above procedure terminates then the re-
of value OPTHK . Next, using the algorithm of Theorem sulting γ̃ satisfies the requirement of Theorem 4.2. There-
4.2 on z ∗ (which is defined by (3.5)) with ε = 0.2, we fore, what we need to show is that this algorithm terminates
obtain γ̃e ’sP that define the exponential family distribution in time polynomial in n, − log zmin and 1/ε, and that each
p̃(T ) := e e∈T γ̃e . Since x ∗ was an extreme point, we know iteration can be implemented in polynomial time.
that z∗min ≥ e−O(n log n) ; thus, the algorithm of Theorem 4.2 We start by bounding the number of iterations - we will
indeed runs in polynomial time. show that it is O( 1ε |E|2 [|V| log(|V|) − log(εzmin )]). In the
Next, we use the polynomial time sampling procedure next lemma, we derive an equation for δ, and prove that for
described in section 4.2 to sample Θ(log n) trees T j from the f , e the probabilities q f (·) do not decrease as a result of
distribution p̃(·), and take T ∗ to be the one among them that decreasing γe .
minimizes the cost c(T j ). By Theorem 5.1, we know that T ∗ L 7.1. If for some δ ≥ 0 and an edge e, we define γ0 by
is (β, 2)-thin with high probability. γe0 = γe − δ and γ0f = γ f for all f , e, then
Now, we use Theorem 6.1 to obtain, in polynomial time,
a (2+8 log n/ log log n)-approximation of our ATSP instance. 1. for all f ∈ E \ {e}, q f (γ0 ) ≥ q f (γ),
δ
 
0) − 1 = e
1 1
The proof also shows that the integrality gap of the Held- 2. qe (γ0 ) satisfies qe (γ qe (γ) − 1 .
Karp relaxation for the Asymmetric TSP is bounded above
In particular, in the main loop of the algorithm, since
by 2 + 8 log n/ log log n. The best known lower bound on the
qe (γ) > (1 + ε)ze and we want qe (γ0 ) = (1 + ε/2)ze , we get
integrality gap is only 2, as shown in [6]. Closing this gap
δ = log (1−qqe (γ)(1−(1+ε/2)ze )
e (γ))(1+ε/2)ze
> log (1+ε/2)
(1+ε)
> 4ε for ε ≤ 1 (for larger
is a challenging open question, and this possibly could be
values of ε, we can simply decrease ε to 1).
answered using thinner spanning trees.
Proof. Let us consider some f ∈ E \ {e}. We have
C 6.1. If there always exists a (C1 , C2 )-thin span- P 0
ning tree where C1 and C2 are constants, the integrality gap T ∈T : f ∈T exp(γ (T ))
q f (γ ) =
0
P 0
of the ATSP Held-Karp linear programming relaxation is a T ∈T exp(γ (T ))
γ (T )
constant. + T :e<T, f ∈T eγ (T )
P 0 P 0
T :e∈T, f ∈T e
= P γ0 (T ) +
P γ0 (T )
T 3e e T :e<T e
7 A Combinatorial Algorithm for Approximately
e−δ T :e∈T, f ∈T eγ(T ) + T :e<T, f ∈T eγ(T )
P P
Solving CP
=
e−δ T 3e eγ(T ) + T :e<T eγ(T )
P P
In this section, we provide a combinatorial algorithm to
efficiently find γ̃e ’s that approximately preserve the marginal e−δ a + b
=
probabilities given by z and therefore prove Theorem 4.2. e−δ c + d
Given a vector γ, for each edge e, define qe (γ) :=
P
exp(γ(T )) with a, b, c, d appropriately defined. The same expression
γ(T = γ
P
PT 3e , where ) f ∈T f . For notational conve-
T exp(γ(T )) holds for q f (γ) with the e−δ factors removed. But, for general
nience, we have dropped the fact that T ∈ T in these summa-
a, b, c, d ≥ 0, if ac ≤ a+b xa+b a+b
c+d then xc+d ≥ c+d for x ≤ 1. Since
tions; this shouldn’t lead to any confusion. Restated, qe (γ) is
the probability that edge e will be included in a spanning tree γ(T )
a+b
P
a T ∈T :e∈T, f ∈T e
T that is chosen with probability proportional to exp(γ(T )). = P ≤ q f (γ) =
c T ∈T :e∈T e
γ(T ) c+d
We compute γ̃ using the following simple algorithm.
Start with all γe equal, and as long as the marginal qe (γ) for by negative correlation (since a/c represents the conditional
some edge e is more than (1+ε)ze , we decrease appropriately probability that f is present given that e is present), we get
γe in order to decrease qe (γ) to (1 + ε/2)ze . More formally, that q f (γ0 ) ≥ q f (γ) for δ ≥ 0.
here is a description of the algorithm. Now, we proceed to deriving the equation for δ. By
definition of qe (γ), we have
1. Set γ = ~0. P γ(T )
1 T :e<T e
2. While there exists an edge e with qe (γ) > (1 + ε)ze : − 1 = P γ(T )
.
qe (γ) T 3e e

386 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 Hence, iterations. Let γ be the vector of γe ’s computed at such an
iteration. For brevity, let us define qe := qe (γ) for all edges
P γ0 (T )
1 T :e<T e e.
−1 = γ0 (T )
qe (γ0 ) We prove first that there exists some e∗ ∈ E such that
P
T 3e e
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

ετ
P
T :e<T e
γ(T ) γe∗ < − 4m . Indeed, there are m edges, and by Lemma 7.1
= we know that in each iteration we decrease γe of one of these
e−δ T 3e eγ(T )
P
! edges by at least ε/4. Thus, we know that, after more than τ
1
= eδ −1 . iterations, there exists e∗ for which γe∗ is as desired.
qe (γ) Note that we never decrease γe for edges e with qe (·)
smaller than (1 + ε)ze , and Lemma 7.1 shows that reducing
Before bounding the number of iterations, we collect
γ f of edge f , e can only increase qe (·). Therefore, we
some basic results regarding spanning trees which we need
know that all the edges with γe being negative must satisfy
for the proof of the number of iterations.
qe ≥ (1 + ε/2)ze . In other words, all edges e such that
L 7.2. Let G = (V, E) be a graph with weights γe for qe < (1+ε/2)ze satisfy γe = 0. Finally, by a simple averaging
argument, we know that e qe = n − 1 < (1 + ε/2)(n − 1) =
P
e ∈ E. Let Q ⊂ E be such that for all f ∈ Q, e ∈ E \ Q,
we have γ f > γe + ∆ for some ∆ ≥ 0. Let r be the size of a (1 + ε/2) e ze . Hence, there exists at least one edge f ∗ with
P

maximum spanning forest of Q. Then q f ∗ < (1 + ε/2)z f ∗ and thus having γ f ∗ = 0.

We proceed now to exhibiting a set Q such that:
1. For any T ∈ T , we have |T ∩ Q| ≤ r.
(I): ∅ , Q ⊂ E, and
Define T= := {T ∈ T : |T ∩ Q| = r} and T< := {T ∈ T :
ετ
|T ∩ Q| < r}. (II): for all e ∈ E \ Q and f ∈ Q, γe + 4m2
< γf .

2. Any spanning tree T ∈ T= can be generated by taking We construct Q as follows. We set threshold values Γi =
ετi
the union of any spanning forest F (of cardinality r) − 4m 2 , for i ≥ 0, and define Qi = {e ∈ E |γe ≥ Γi }. Let Q = Q j

of the graph (V, Q) and a spanning tree (of cardinality where j is the first index such that Q j = Q j+1 . Clearly, by
n−r −1) of the graph G/Q in which the edges of Q have construction of Q, property (II) is satisfied. Also, Q is non-
been contracted. empty since f ∗ ∈ Q0 ⊆ Q j = Q. Finally, by the pigeonhole
principle, since we have m different edges, we know that
ετ
3. Let T max be a maximum spanning tree of G with respect j < m. Thus, for each e ∈ Q we have γe > Γm = − 4m .
to the weights γ(·), i.e. T max = arg maxT ∈T γ(T ). Then, This means that e < Q and thus Q has property (I).
∗

for any T ∈ T< , we have γ(T ) < γ(T max ) − ∆. Observe that Q satisfies the hypothesis of Lemma 7.2
ετ
with ∆ = 4m 2 . Thus, for any T ∈ T< , we have

Proof. These properties easily follow from the matroidal

ετ
properties of spanning trees. To prove 3., consider any T ∈ (7.14) γ(T max ) > γ(T ) + ,
T< . Since |T ∩ Q| < r, there exists an edge f ∈ (T max ∩ Q) \ T 4m2
such that (T ∩ Q) ∪ { f } is a forest of G. Therefore, the where T max and r are as defined in Lemma 7.2.
unique circuit in T ∪ { f } contains an edge e < Q. Thus Let G b be the graph G/Q obtained by contracting all
T = T ∪ { f } \ {e} is a spanning tree. Our assumption on
0 the edges in Q. So, G b consists only of edges not in Q
Q implies that (some of them can be self-loops). Let T̂ be the set of all
spanning trees of b , and for any given edge e < Q, let
G
γ(T max ) ≥ γ(T 0 ) = γ(T ) − γe + γ f > γ(T ) + ∆,
P
exp(γ(T̂ ))
q̂ := PT̂ ∈T̂ ,T̂ 3e
e be the probability that edge e is included
exp(γ(T̂ ))
T̂ ∈T̂

which yields the desired inequality. in a random spanning tree T̂ of G b , where each tree T̂ is
chosen with probability proportional to eγ(T̂ ) . Since spanning
We proceed to bounding the number of iterations. b have n − r − 1 edges, we have
trees of G
X
L 7.3. The algorithm executes at most (7.15) q̂e = n − r − 1.
O( 1ε |E|2 [|V| log(|V|) − log(εzmin )]) iterations of the main e∈E\Q

loop. On the other hand, since z satisfies z(E) = n − 1 and

z(Q) ≤ r (by definition of r, see Lemma 7.2, part 1.), we
Proof. Let n = |V| and m = |E|. Assume for the sake of
have that z(E \ Q) ≥ n − r − 1. Therefore, (7.15) implies that
contradiction that the algorithm executes more than
there must exist ê < Q such that q̂ê ≤ zê .
4 Our final step is to show that for any e < Q, qe <
τ := m2 [n log n − log(εzmin )] + εzmin
ε q̂e 2 . Note that once we establish this, we know that

387 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 qê < q̂ê + εz2min ≤ (1 + 2ε )zê , and thus it must be the case To complete the analysis of the algorithm, we need to
that γê = 0. But this contradicts the fact that ê < Q, as by argue that each iteration can be implemented in polynomial
construction all e with γe = 0 must be in Q. Thus, we obtain time. First, for any given vector γ, we can compute effi-
P P
a contradiction that concludes the proof of the Lemma. ciently the sums T exp(γ(T )) and T 3e exp(γ(T )) for any
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

It remains to prove that for any e < Q, qe < q̂e + εz2min . edge e - this will enable us to compute all qe (γ)’s. This can
We have that be done using Kirchhoff’s matrix tree theorem (see [4]), as
P γ(T )
discussed in Section 4.2 (with λe = eγe ). Observe that we can
T ∈T :e∈T e bound all entries of the weighted Laplacian matrix in terms
qe = P γ(T )
T ∈T e of the input size since the proof of Lemma 7.3 actually shows
γ(T ) ετ
+ T ∈T< :e∈T eγ(T ) ≤ γe ≤ 0 for all e ∈ E and any iteration of the algo-
P P
T ∈T= :e∈T e that − 4|E|
= P γ(T ) rithm. Therefore, we can compute these cofactors efficiently,
T ∈T e
P γ(T ) P γ(T ) in time polynomial in n, − log zmin and 1/ε. Finally, δ can be
T ∈T :e∈T e T ∈T :e∈T e
≤ P = γ(T ) + P < γ(T ) computed efficiently from Lemma 7.1.
T ∈T= e T ∈T e
γ(T )
eγ(T )
P
T ∈T :e∈T e Acknowledgments. We would like to thank the reviewers
X
(7.16) ≤ P = γ(T ) + ,
T ∈T= e T ∈T :e∈T
eγ(Tmax ) for many insightful comments.
<

the first inequality following from replacing T with T= in the References

first denominator, and the second inequality following from
considering only one term in the second denominator. Using
(7.14) and the fact that the number of spanning trees is at [1] D. J. Aldous. A random walk construction of uniform
most nn−2 , the second term is bounded by: spanning trees and uniform labelled trees. SIAM Journal on
Discrete Mathematics, 3(4):450–465, 1990.
X eγ(T ) n−2 −ετ/4m2 [2] A. Asadpour and A. Saberi. Maximum entropy selection: a
(7.17) ≤ n e
T ∈T< :e∈T
eγ(Tmax ) randomized rounding method. submitted.
[3] M. Bläser. A new approximation algorithm for the asymmet-
1 n −ετ/4m2 ric TSP with traingle inequality. In ACM-SIAM Symposium
< ne
2 on Discrete Algorithms, pages 638–645, 2002.
εzmin [4] B. Bollobas. Modern Graph Theory. Springer, 2002.
= ,
2 [5] A. Broder. Generating random spanning trees. In 30th
Annual Symposium on Foundations of Computer Science,
by definition of τ. To handle the first term of (7.16), we can pages 442–447, 1989.
use part 2. of Lemma 7.2 and factorize: [6] M. Charikar, M. Goemans, and H. Karloff. On the integrality
   ratio for the asymmetric traveling salesman problem. Mathe-
X
γ(T )
 X
γ(
  X
γ(T 0 
 matics of Operations Research, 31:245–252, 2006.
e =  T̂ )
e   e )  ,
  
[7] C. Chekuri and J. Vondrak. Randomized pipage
T ∈T= 0 T ∈F
T̂ ∈T̂ rounding for matroid polytopes and applications, 2009.
http://arxiv.org/abs/0909.4348v1.
where F is the set of all spanning forests of (V, Q). Similarly, [8] N. Christofides. Worst case analysis of a new heuristic for the
we can write traveling salesman problem. Report 388, Graduate School of
   Industrial Administration, Carnegie-Mellon University, Pitts-
burgh, PA, 1976.
X  X   X
γ(T ) γ( γ(T 0 

e =  T̂ )
e   e )  .
  

 [9] C. J. Colbourn, W. J. Myrvold, and E. Neufeld. Two algo-
T ∈T= ,T 3e T̂ ∈T̂ ,T̂ 3e T 0 ∈F
rithms for unranking arborescences. Journal of Algorithms,
20(2):268–281, 1996.
As a result, we have that the first term of (7.16) reduces to: [10] J. Edmonds. Matroids and the greedy algorithm. Mathemati-
P
γ(T̂ ) P

γ(T 0 )
 cal Programming, pages 127–136, 1971.
γ(T̂ )
T̂ ∈T̂ e T 0 ∈F e
P
T̂ ∈T̂ e [11] U. Feige and M. Singh. Improved approximation ratios for
= = q̂e .
γ(T̂ ) traveling salesman tours and paths in directed graphs. In 10th
P  P
γ(T̂ ) γ(T 0 )
P
T̂ ∈T̂ ,T̂ 3e e


T̂ ∈T̂ ,T̂ 3e e T 0 ∈F e
International Workshop, APPROX, pages 104–118, 2007.
[12] A. M. Frieze, G. Galbiati, and F. Maffioli. On the worst-case
Together with (7.16) and (7.17), this gives
performance of some algorithms for the asymmetric traveling
εzmin salesman problem. Networks, 12:23–39, 1982.
qe ≤ q̂e + , [13] A. Ghosh, S. Boyd, and A. Saberi. Minimizing effective
2
resistance of a graph. SIAM Rev., 50(1):37–66, 2008.
which completes the proof. [14] M. X. Goemans. Minimum bounded degree spanning trees.

388 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.
 In 47th Annual Symposium on Foundations of Computer [20] V. G. Kulkarni. Generating random combinatorial objects.
Science, pages 273–282, 2006. Journal of Algorithms, 11:185–207, 1990.
[15] A. Guenoche. Random spanning tree. Journal of Algorithms, [21] R. Lyons and Y. Peres. Probability on Trees and Net-
4:214–220, 1983. works. 2009. In preparation. Current version available at
Downloaded 12/24/14 to 134.225.1.226. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

[16] N. S. H. Kaplan, M. Lewenstein and M. Sviridenko. Approx- http://mypage.iu.edu/˜rdlyons/.

imation algorithms for asymmetric TSP by decomposing di- [22] A. Nemirovski. Lectures on modern convex optimization,
rected regular multigraphs. J. ACM, pages 602–626, 2005. 2005.
[17] M. Held and R. Karp. The traveling salesman problem and [23] A. Panconesi and A. Srinivasan. Randomized distributed
minimum spanning trees. Operations Research, 18:1138– edge coloring via an extension of the Chernoff-Hoeffding
1162, 1970. bounds. SIAM J. Comput., 26:350–368, 1997.
[18] D. R. Karger. Global min-cuts in RNC, and other ram- [24] A. Schrijver. Combinatorial Optimization, volume 2 of
ifications of a simple min-cut algorithm. In 4th annual Algorithms and Combinatorics. Springer, 2003.
ACM-SIAM Symposium on Discrete algorithms, pages 21–30, [25] D. B. Wilson. Generating random spanning trees more
Philadelphia, PA, USA, 1993. Society for Industrial and Ap- quickly than the cover time. In 28th Annual ACM Symposium
plied Mathematics. on the Theory of Computing, pages 296–303. ACM, 1996.
[19] J. Kelner and A. Madry. Faster generation of random span- [26] R. Zenklusen, July 2009. Personal communication.
ning trees. In 50th Annual Symposium on Foundations of
Computer Science, 2009.

389 Copyright © by SIAM.

Unauthorized reproduction of this article is prohibited.